LCOV - code coverage report
Current view: top level - lib/thread - thread.c (source / functions) Hit Total Coverage
Test: ut_cov_unit.info Lines: 844 1314 64.2 %
Date: 2024-11-05 10:06:02 Functions: 111 163 68.1 %

          Line data    Source code
       1             : /*   SPDX-License-Identifier: BSD-3-Clause
       2             :  *   Copyright (C) 2016 Intel Corporation.
       3             :  *   All rights reserved.
       4             :  *   Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
       5             :  */
       6             : 
       7             : #include "spdk/stdinc.h"
       8             : 
       9             : #include "spdk/env.h"
      10             : #include "spdk/likely.h"
      11             : #include "spdk/queue.h"
      12             : #include "spdk/string.h"
      13             : #include "spdk/thread.h"
      14             : #include "spdk/trace.h"
      15             : #include "spdk/util.h"
      16             : #include "spdk/fd_group.h"
      17             : 
      18             : #include "spdk/log.h"
      19             : #include "spdk_internal/thread.h"
      20             : #include "spdk_internal/usdt.h"
      21             : #include "thread_internal.h"
      22             : 
      23             : #include "spdk_internal/trace_defs.h"
      24             : 
      25             : #ifdef __linux__
      26             : #include <sys/timerfd.h>
      27             : #include <sys/eventfd.h>
      28             : #endif
      29             : 
      30             : #ifdef SPDK_HAVE_EXECINFO_H
      31             : #include <execinfo.h>
      32             : #endif
      33             : 
      34             : #define SPDK_MSG_BATCH_SIZE             8
      35             : #define SPDK_MAX_DEVICE_NAME_LEN        256
      36             : #define SPDK_THREAD_EXIT_TIMEOUT_SEC    5
      37             : #define SPDK_MAX_POLLER_NAME_LEN        256
      38             : #define SPDK_MAX_THREAD_NAME_LEN        256
      39             : 
      40             : static struct spdk_thread *g_app_thread;
      41             : 
      42             : struct spdk_interrupt {
      43             :         int                     efd;
      44             :         struct spdk_thread      *thread;
      45             :         spdk_interrupt_fn       fn;
      46             :         void                    *arg;
      47             :         char                    name[SPDK_MAX_POLLER_NAME_LEN + 1];
      48             : };
      49             : 
      50             : enum spdk_poller_state {
      51             :         /* The poller is registered with a thread but not currently executing its fn. */
      52             :         SPDK_POLLER_STATE_WAITING,
      53             : 
      54             :         /* The poller is currently running its fn. */
      55             :         SPDK_POLLER_STATE_RUNNING,
      56             : 
      57             :         /* The poller was unregistered during the execution of its fn. */
      58             :         SPDK_POLLER_STATE_UNREGISTERED,
      59             : 
      60             :         /* The poller is in the process of being paused.  It will be paused
      61             :          * during the next time it's supposed to be executed.
      62             :          */
      63             :         SPDK_POLLER_STATE_PAUSING,
      64             : 
      65             :         /* The poller is registered but currently paused.  It's on the
      66             :          * paused_pollers list.
      67             :          */
      68             :         SPDK_POLLER_STATE_PAUSED,
      69             : };
      70             : 
      71             : struct spdk_poller {
      72             :         TAILQ_ENTRY(spdk_poller)        tailq;
      73             :         RB_ENTRY(spdk_poller)           node;
      74             : 
      75             :         /* Current state of the poller; should only be accessed from the poller's thread. */
      76             :         enum spdk_poller_state          state;
      77             : 
      78             :         uint64_t                        period_ticks;
      79             :         uint64_t                        next_run_tick;
      80             :         uint64_t                        run_count;
      81             :         uint64_t                        busy_count;
      82             :         uint64_t                        id;
      83             :         spdk_poller_fn                  fn;
      84             :         void                            *arg;
      85             :         struct spdk_thread              *thread;
      86             :         struct spdk_interrupt           *intr;
      87             :         spdk_poller_set_interrupt_mode_cb set_intr_cb_fn;
      88             :         void                            *set_intr_cb_arg;
      89             : 
      90             :         char                            name[SPDK_MAX_POLLER_NAME_LEN + 1];
      91             : };
      92             : 
      93             : enum spdk_thread_state {
      94             :         /* The thread is processing poller and message by spdk_thread_poll(). */
      95             :         SPDK_THREAD_STATE_RUNNING,
      96             : 
      97             :         /* The thread is in the process of termination. It reaps unregistering
      98             :          * poller are releasing I/O channel.
      99             :          */
     100             :         SPDK_THREAD_STATE_EXITING,
     101             : 
     102             :         /* The thread is exited. It is ready to call spdk_thread_destroy(). */
     103             :         SPDK_THREAD_STATE_EXITED,
     104             : };
     105             : 
     106             : struct spdk_thread {
     107             :         uint64_t                        tsc_last;
     108             :         struct spdk_thread_stats        stats;
     109             :         /*
     110             :          * Contains pollers actively running on this thread.  Pollers
     111             :          *  are run round-robin. The thread takes one poller from the head
     112             :          *  of the ring, executes it, then puts it back at the tail of
     113             :          *  the ring.
     114             :          */
     115             :         TAILQ_HEAD(active_pollers_head, spdk_poller)    active_pollers;
     116             :         /**
     117             :          * Contains pollers running on this thread with a periodic timer.
     118             :          */
     119             :         RB_HEAD(timed_pollers_tree, spdk_poller)        timed_pollers;
     120             :         struct spdk_poller                              *first_timed_poller;
     121             :         /*
     122             :          * Contains paused pollers.  Pollers on this queue are waiting until
     123             :          * they are resumed (in which case they're put onto the active/timer
     124             :          * queues) or unregistered.
     125             :          */
     126             :         TAILQ_HEAD(paused_pollers_head, spdk_poller)    paused_pollers;
     127             :         struct spdk_ring                *messages;
     128             :         int                             msg_fd;
     129             :         SLIST_HEAD(, spdk_msg)          msg_cache;
     130             :         size_t                          msg_cache_count;
     131             :         spdk_msg_fn                     critical_msg;
     132             :         uint64_t                        id;
     133             :         uint64_t                        next_poller_id;
     134             :         enum spdk_thread_state          state;
     135             :         int                             pending_unregister_count;
     136             :         uint32_t                        for_each_count;
     137             : 
     138             :         RB_HEAD(io_channel_tree, spdk_io_channel)       io_channels;
     139             :         TAILQ_ENTRY(spdk_thread)                        tailq;
     140             : 
     141             :         char                            name[SPDK_MAX_THREAD_NAME_LEN + 1];
     142             :         struct spdk_cpuset              cpumask;
     143             :         uint64_t                        exit_timeout_tsc;
     144             : 
     145             :         int32_t                         lock_count;
     146             : 
     147             :         /* spdk_thread is bound to current CPU core. */
     148             :         bool                            is_bound;
     149             : 
     150             :         /* Indicates whether this spdk_thread currently runs in interrupt. */
     151             :         bool                            in_interrupt;
     152             :         bool                            poller_unregistered;
     153             :         struct spdk_fd_group            *fgrp;
     154             : 
     155             :         uint16_t                        trace_id;
     156             : 
     157             :         uint8_t                         reserved[6];
     158             : 
     159             :         /* User context allocated at the end */
     160             :         uint8_t                         ctx[0];
     161             : };
     162             : 
     163             : /*
     164             :  * Assert that spdk_thread struct is 8 byte aligned to ensure
     165             :  * the user ctx is also 8-byte aligned.
     166             :  */
     167             : SPDK_STATIC_ASSERT((sizeof(struct spdk_thread)) % 8 == 0, "Incorrect size");
     168             : 
     169             : static pthread_mutex_t g_devlist_mutex = PTHREAD_MUTEX_INITIALIZER;
     170             : 
     171             : static spdk_new_thread_fn g_new_thread_fn = NULL;
     172             : static spdk_thread_op_fn g_thread_op_fn = NULL;
     173             : static spdk_thread_op_supported_fn g_thread_op_supported_fn;
     174             : static size_t g_ctx_sz = 0;
     175             : /* Monotonic increasing ID is set to each created thread beginning at 1. Once the
     176             :  * ID exceeds UINT64_MAX, further thread creation is not allowed and restarting
     177             :  * SPDK application is required.
     178             :  */
     179             : static uint64_t g_thread_id = 1;
     180             : 
     181             : enum spin_error {
     182             :         SPIN_ERR_NONE,
     183             :         /* Trying to use an SPDK lock while not on an SPDK thread */
     184             :         SPIN_ERR_NOT_SPDK_THREAD,
     185             :         /* Trying to lock a lock already held by this SPDK thread */
     186             :         SPIN_ERR_DEADLOCK,
     187             :         /* Trying to unlock a lock not held by this SPDK thread */
     188             :         SPIN_ERR_WRONG_THREAD,
     189             :         /* pthread_spin_*() returned an error */
     190             :         SPIN_ERR_PTHREAD,
     191             :         /* Trying to destroy a lock that is held */
     192             :         SPIN_ERR_LOCK_HELD,
     193             :         /* lock_count is invalid */
     194             :         SPIN_ERR_LOCK_COUNT,
     195             :         /*
     196             :          * An spdk_thread may migrate to another pthread. A spinlock held across migration leads to
     197             :          * undefined behavior. A spinlock held when an SPDK thread goes off CPU would lead to
     198             :          * deadlock when another SPDK thread on the same pthread tries to take that lock.
     199             :          */
     200             :         SPIN_ERR_HOLD_DURING_SWITCH,
     201             :         /* Trying to use a lock that was destroyed (but not re-initialized) */
     202             :         SPIN_ERR_DESTROYED,
     203             :         /* Trying to use a lock that is not initialized */
     204             :         SPIN_ERR_NOT_INITIALIZED,
     205             : 
     206             :         /* Must be last, not an actual error code */
     207             :         SPIN_ERR_LAST
     208             : };
     209             : 
     210             : static const char *spin_error_strings[] = {
     211             :         [SPIN_ERR_NONE]                 = "No error",
     212             :         [SPIN_ERR_NOT_SPDK_THREAD]      = "Not an SPDK thread",
     213             :         [SPIN_ERR_DEADLOCK]             = "Deadlock detected",
     214             :         [SPIN_ERR_WRONG_THREAD]         = "Unlock on wrong SPDK thread",
     215             :         [SPIN_ERR_PTHREAD]              = "Error from pthread_spinlock",
     216             :         [SPIN_ERR_LOCK_HELD]            = "Destroying a held spinlock",
     217             :         [SPIN_ERR_LOCK_COUNT]           = "Lock count is invalid",
     218             :         [SPIN_ERR_HOLD_DURING_SWITCH]   = "Lock(s) held while SPDK thread going off CPU",
     219             :         [SPIN_ERR_DESTROYED]            = "Lock has been destroyed",
     220             :         [SPIN_ERR_NOT_INITIALIZED]      = "Lock has not been initialized",
     221             : };
     222             : 
     223             : #define SPIN_ERROR_STRING(err) (err < 0 || err >= SPDK_COUNTOF(spin_error_strings)) \
     224             :                                 ? "Unknown error" : spin_error_strings[err]
     225             : 
     226             : static void
     227           0 : __posix_abort(enum spin_error err)
     228             : {
     229           0 :         abort();
     230             : }
     231             : 
     232             : typedef void (*spin_abort)(enum spin_error err);
     233             : spin_abort g_spin_abort_fn = __posix_abort;
     234             : 
     235             : #define SPIN_ASSERT_IMPL(cond, err, extra_log, ret) \
     236             :         do { \
     237             :                 if (spdk_unlikely(!(cond))) { \
     238             :                         SPDK_ERRLOG("unrecoverable spinlock error %d: %s (%s)\n", err, \
     239             :                                     SPIN_ERROR_STRING(err), #cond); \
     240             :                         extra_log; \
     241             :                         g_spin_abort_fn(err); \
     242             :                         ret; \
     243             :                 } \
     244             :         } while (0)
     245             : #define SPIN_ASSERT_LOG_STACKS(cond, err, lock) \
     246             :         SPIN_ASSERT_IMPL(cond, err, sspin_stacks_print(sspin), return)
     247             : #define SPIN_ASSERT_RETURN(cond, err, ret)      SPIN_ASSERT_IMPL(cond, err, , return ret)
     248             : #define SPIN_ASSERT(cond, err)                  SPIN_ASSERT_IMPL(cond, err, ,)
     249             : 
     250             : struct io_device {
     251             :         void                            *io_device;
     252             :         char                            name[SPDK_MAX_DEVICE_NAME_LEN + 1];
     253             :         spdk_io_channel_create_cb       create_cb;
     254             :         spdk_io_channel_destroy_cb      destroy_cb;
     255             :         spdk_io_device_unregister_cb    unregister_cb;
     256             :         struct spdk_thread              *unregister_thread;
     257             :         uint32_t                        ctx_size;
     258             :         uint32_t                        for_each_count;
     259             :         RB_ENTRY(io_device)             node;
     260             : 
     261             :         uint32_t                        refcnt;
     262             : 
     263             :         bool                            pending_unregister;
     264             :         bool                            unregistered;
     265             : };
     266             : 
     267             : static RB_HEAD(io_device_tree, io_device) g_io_devices = RB_INITIALIZER(g_io_devices);
     268             : 
     269             : static int
     270       14324 : io_device_cmp(struct io_device *dev1, struct io_device *dev2)
     271             : {
     272       14324 :         return (dev1->io_device < dev2->io_device ? -1 : dev1->io_device > dev2->io_device);
     273             : }
     274             : 
     275       24947 : RB_GENERATE_STATIC(io_device_tree, io_device, node, io_device_cmp);
     276             : 
     277             : static int
     278        9278 : io_channel_cmp(struct spdk_io_channel *ch1, struct spdk_io_channel *ch2)
     279             : {
     280        9278 :         return (ch1->dev < ch2->dev ? -1 : ch1->dev > ch2->dev);
     281             : }
     282             : 
     283       25836 : RB_GENERATE_STATIC(io_channel_tree, spdk_io_channel, node, io_channel_cmp);
     284             : 
     285             : struct spdk_msg {
     286             :         spdk_msg_fn             fn;
     287             :         void                    *arg;
     288             : 
     289             :         SLIST_ENTRY(spdk_msg)   link;
     290             : };
     291             : 
     292             : static struct spdk_mempool *g_spdk_msg_mempool = NULL;
     293             : 
     294             : static TAILQ_HEAD(, spdk_thread) g_threads = TAILQ_HEAD_INITIALIZER(g_threads);
     295             : static uint32_t g_thread_count = 0;
     296             : 
     297             : static __thread struct spdk_thread *tls_thread = NULL;
     298             : 
     299             : static void
     300           0 : thread_trace(void)
     301             : {
     302           0 :         struct spdk_trace_tpoint_opts opts[] = {
     303             :                 {
     304             :                         "THREAD_IOCH_GET", TRACE_THREAD_IOCH_GET,
     305             :                         OWNER_TYPE_NONE, OBJECT_NONE, 0,
     306             :                         {{ "refcnt", SPDK_TRACE_ARG_TYPE_INT, 4 }}
     307             :                 },
     308             :                 {
     309             :                         "THREAD_IOCH_PUT", TRACE_THREAD_IOCH_PUT,
     310             :                         OWNER_TYPE_NONE, OBJECT_NONE, 0,
     311             :                         {{ "refcnt", SPDK_TRACE_ARG_TYPE_INT, 4 }}
     312             :                 }
     313             :         };
     314             : 
     315           0 :         spdk_trace_register_owner_type(OWNER_TYPE_THREAD, 't');
     316           0 :         spdk_trace_register_description_ext(opts, SPDK_COUNTOF(opts));
     317           0 : }
     318          49 : SPDK_TRACE_REGISTER_FN(thread_trace, "thread", TRACE_GROUP_THREAD)
     319             : 
     320             : /*
     321             :  * If this compare function returns zero when two next_run_ticks are equal,
     322             :  * the macro RB_INSERT() returns a pointer to the element with the same
     323             :  * next_run_tick.
     324             :  *
     325             :  * Fortunately, the macro RB_REMOVE() takes not a key but a pointer to the element
     326             :  * to remove as a parameter.
     327             :  *
     328             :  * Hence we allow RB_INSERT() to insert elements with the same keys on the right
     329             :  * side by returning 1 when two next_run_ticks are equal.
     330             :  */
     331             : static inline int
     332         909 : timed_poller_compare(struct spdk_poller *poller1, struct spdk_poller *poller2)
     333             : {
     334         909 :         if (poller1->next_run_tick < poller2->next_run_tick) {
     335         502 :                 return -1;
     336             :         } else {
     337         407 :                 return 1;
     338             :         }
     339             : }
     340             : 
     341        4994 : RB_GENERATE_STATIC(timed_pollers_tree, spdk_poller, node, timed_poller_compare);
     342             : 
     343             : static inline struct spdk_thread *
     344      775886 : _get_thread(void)
     345             : {
     346      775886 :         return tls_thread;
     347             : }
     348             : 
     349             : static int
     350          95 : _thread_lib_init(size_t ctx_sz, size_t msg_mempool_sz)
     351             : {
     352          95 :         char mempool_name[SPDK_MAX_MEMZONE_NAME_LEN];
     353             : 
     354          95 :         g_ctx_sz = ctx_sz;
     355             : 
     356          95 :         snprintf(mempool_name, sizeof(mempool_name), "msgpool_%d", getpid());
     357          95 :         g_spdk_msg_mempool = spdk_mempool_create(mempool_name, msg_mempool_sz,
     358             :                              sizeof(struct spdk_msg),
     359             :                              0, /* No cache. We do our own. */
     360             :                              SPDK_ENV_NUMA_ID_ANY);
     361             : 
     362          95 :         SPDK_DEBUGLOG(thread, "spdk_msg_mempool was created with size: %zu\n",
     363             :                       msg_mempool_sz);
     364             : 
     365          95 :         if (!g_spdk_msg_mempool) {
     366           0 :                 SPDK_ERRLOG("spdk_msg_mempool creation failed\n");
     367           0 :                 return -ENOMEM;
     368             :         }
     369             : 
     370          95 :         return 0;
     371             : }
     372             : 
     373             : static void thread_interrupt_destroy(struct spdk_thread *thread);
     374             : static int thread_interrupt_create(struct spdk_thread *thread);
     375             : 
     376             : static void
     377         175 : _free_thread(struct spdk_thread *thread)
     378             : {
     379             :         struct spdk_io_channel *ch;
     380             :         struct spdk_msg *msg;
     381             :         struct spdk_poller *poller, *ptmp;
     382             : 
     383         175 :         RB_FOREACH(ch, io_channel_tree, &thread->io_channels) {
     384           0 :                 SPDK_ERRLOG("thread %s still has channel for io_device %s\n",
     385             :                             thread->name, ch->dev->name);
     386             :         }
     387             : 
     388         175 :         TAILQ_FOREACH_SAFE(poller, &thread->active_pollers, tailq, ptmp) {
     389           0 :                 if (poller->state != SPDK_POLLER_STATE_UNREGISTERED) {
     390           0 :                         SPDK_WARNLOG("active_poller %s still registered at thread exit\n",
     391             :                                      poller->name);
     392             :                 }
     393           0 :                 TAILQ_REMOVE(&thread->active_pollers, poller, tailq);
     394           0 :                 free(poller);
     395             :         }
     396             : 
     397         202 :         RB_FOREACH_SAFE(poller, timed_pollers_tree, &thread->timed_pollers, ptmp) {
     398          27 :                 if (poller->state != SPDK_POLLER_STATE_UNREGISTERED) {
     399           0 :                         SPDK_WARNLOG("timed_poller %s still registered at thread exit\n",
     400             :                                      poller->name);
     401             :                 }
     402          27 :                 RB_REMOVE(timed_pollers_tree, &thread->timed_pollers, poller);
     403          27 :                 free(poller);
     404             :         }
     405             : 
     406         175 :         TAILQ_FOREACH_SAFE(poller, &thread->paused_pollers, tailq, ptmp) {
     407           0 :                 SPDK_WARNLOG("paused_poller %s still registered at thread exit\n", poller->name);
     408           0 :                 TAILQ_REMOVE(&thread->paused_pollers, poller, tailq);
     409           0 :                 free(poller);
     410             :         }
     411             : 
     412         175 :         pthread_mutex_lock(&g_devlist_mutex);
     413         175 :         assert(g_thread_count > 0);
     414         175 :         g_thread_count--;
     415         175 :         TAILQ_REMOVE(&g_threads, thread, tailq);
     416         175 :         pthread_mutex_unlock(&g_devlist_mutex);
     417             : 
     418         175 :         msg = SLIST_FIRST(&thread->msg_cache);
     419      179229 :         while (msg != NULL) {
     420      179054 :                 SLIST_REMOVE_HEAD(&thread->msg_cache, link);
     421             : 
     422      179054 :                 assert(thread->msg_cache_count > 0);
     423      179054 :                 thread->msg_cache_count--;
     424      179054 :                 spdk_mempool_put(g_spdk_msg_mempool, msg);
     425             : 
     426      179054 :                 msg = SLIST_FIRST(&thread->msg_cache);
     427             :         }
     428             : 
     429         175 :         assert(thread->msg_cache_count == 0);
     430             : 
     431         175 :         if (spdk_interrupt_mode_is_enabled()) {
     432           0 :                 thread_interrupt_destroy(thread);
     433             :         }
     434             : 
     435         175 :         spdk_ring_free(thread->messages);
     436         175 :         free(thread);
     437         175 : }
     438             : 
     439             : int
     440          83 : spdk_thread_lib_init(spdk_new_thread_fn new_thread_fn, size_t ctx_sz)
     441             : {
     442          83 :         assert(g_new_thread_fn == NULL);
     443          83 :         assert(g_thread_op_fn == NULL);
     444             : 
     445          83 :         if (new_thread_fn == NULL) {
     446          81 :                 SPDK_INFOLOG(thread, "new_thread_fn was not specified at spdk_thread_lib_init\n");
     447             :         } else {
     448           2 :                 g_new_thread_fn = new_thread_fn;
     449             :         }
     450             : 
     451          83 :         return _thread_lib_init(ctx_sz, SPDK_DEFAULT_MSG_MEMPOOL_SIZE);
     452             : }
     453             : 
     454             : int
     455          12 : spdk_thread_lib_init_ext(spdk_thread_op_fn thread_op_fn,
     456             :                          spdk_thread_op_supported_fn thread_op_supported_fn,
     457             :                          size_t ctx_sz, size_t msg_mempool_sz)
     458             : {
     459          12 :         assert(g_new_thread_fn == NULL);
     460          12 :         assert(g_thread_op_fn == NULL);
     461          12 :         assert(g_thread_op_supported_fn == NULL);
     462             : 
     463          12 :         if ((thread_op_fn != NULL) != (thread_op_supported_fn != NULL)) {
     464           0 :                 SPDK_ERRLOG("Both must be defined or undefined together.\n");
     465           0 :                 return -EINVAL;
     466             :         }
     467             : 
     468          12 :         if (thread_op_fn == NULL && thread_op_supported_fn == NULL) {
     469           0 :                 SPDK_INFOLOG(thread, "thread_op_fn and thread_op_supported_fn were not specified\n");
     470             :         } else {
     471          12 :                 g_thread_op_fn = thread_op_fn;
     472          12 :                 g_thread_op_supported_fn = thread_op_supported_fn;
     473             :         }
     474             : 
     475          12 :         return _thread_lib_init(ctx_sz, msg_mempool_sz);
     476             : }
     477             : 
     478             : void
     479          94 : spdk_thread_lib_fini(void)
     480             : {
     481             :         struct io_device *dev;
     482             : 
     483          95 :         RB_FOREACH(dev, io_device_tree, &g_io_devices) {
     484           1 :                 SPDK_ERRLOG("io_device %s not unregistered\n", dev->name);
     485             :         }
     486             : 
     487          94 :         g_new_thread_fn = NULL;
     488          94 :         g_thread_op_fn = NULL;
     489          94 :         g_thread_op_supported_fn = NULL;
     490          94 :         g_ctx_sz = 0;
     491          94 :         if (g_app_thread != NULL) {
     492          90 :                 _free_thread(g_app_thread);
     493          90 :                 g_app_thread = NULL;
     494             :         }
     495             : 
     496          94 :         if (g_spdk_msg_mempool) {
     497          94 :                 spdk_mempool_free(g_spdk_msg_mempool);
     498          94 :                 g_spdk_msg_mempool = NULL;
     499             :         }
     500          94 : }
     501             : 
     502             : struct spdk_thread *
     503         211 : spdk_thread_create(const char *name, const struct spdk_cpuset *cpumask)
     504             : {
     505         211 :         struct spdk_thread *thread, *null_thread;
     506         211 :         size_t size = SPDK_ALIGN_CEIL(sizeof(*thread) + g_ctx_sz, SPDK_CACHE_LINE_SIZE);
     507         211 :         struct spdk_msg *msgs[SPDK_MSG_MEMPOOL_CACHE_SIZE];
     508         211 :         int rc = 0, i;
     509             : 
     510             :         /* Since this spdk_thread object will be used by another core, ensure that it won't share a
     511             :          * cache line with any other object allocated on this core */
     512         211 :         rc = posix_memalign((void **)&thread, SPDK_CACHE_LINE_SIZE, size);
     513         211 :         if (rc != 0) {
     514           0 :                 SPDK_ERRLOG("Unable to allocate memory for thread\n");
     515           0 :                 return NULL;
     516             :         }
     517         211 :         memset(thread, 0, size);
     518             : 
     519         211 :         if (cpumask) {
     520          23 :                 spdk_cpuset_copy(&thread->cpumask, cpumask);
     521             :         } else {
     522         188 :                 spdk_cpuset_negate(&thread->cpumask);
     523             :         }
     524             : 
     525         211 :         RB_INIT(&thread->io_channels);
     526         211 :         TAILQ_INIT(&thread->active_pollers);
     527         211 :         RB_INIT(&thread->timed_pollers);
     528         211 :         TAILQ_INIT(&thread->paused_pollers);
     529         211 :         SLIST_INIT(&thread->msg_cache);
     530         211 :         thread->msg_cache_count = 0;
     531             : 
     532         211 :         thread->tsc_last = spdk_get_ticks();
     533             : 
     534             :         /* Monotonic increasing ID is set to each created poller beginning at 1. Once the
     535             :          * ID exceeds UINT64_MAX a warning message is logged
     536             :          */
     537         211 :         thread->next_poller_id = 1;
     538             : 
     539         211 :         thread->messages = spdk_ring_create(SPDK_RING_TYPE_MP_SC, 65536, SPDK_ENV_NUMA_ID_ANY);
     540         211 :         if (!thread->messages) {
     541           0 :                 SPDK_ERRLOG("Unable to allocate memory for message ring\n");
     542           0 :                 free(thread);
     543           0 :                 return NULL;
     544             :         }
     545             : 
     546             :         /* Fill the local message pool cache. */
     547         211 :         rc = spdk_mempool_get_bulk(g_spdk_msg_mempool, (void **)msgs, SPDK_MSG_MEMPOOL_CACHE_SIZE);
     548         211 :         if (rc == 0) {
     549             :                 /* If we can't populate the cache it's ok. The cache will get filled
     550             :                  * up organically as messages are passed to the thread. */
     551      216275 :                 for (i = 0; i < SPDK_MSG_MEMPOOL_CACHE_SIZE; i++) {
     552      216064 :                         SLIST_INSERT_HEAD(&thread->msg_cache, msgs[i], link);
     553      216064 :                         thread->msg_cache_count++;
     554             :                 }
     555             :         }
     556             : 
     557         211 :         if (name) {
     558          42 :                 snprintf(thread->name, sizeof(thread->name), "%s", name);
     559             :         } else {
     560         169 :                 snprintf(thread->name, sizeof(thread->name), "%p", thread);
     561             :         }
     562             : 
     563         211 :         thread->trace_id = spdk_trace_register_owner(OWNER_TYPE_THREAD, thread->name);
     564             : 
     565         211 :         pthread_mutex_lock(&g_devlist_mutex);
     566         211 :         if (g_thread_id == 0) {
     567           0 :                 SPDK_ERRLOG("Thread ID rolled over. Further thread creation is not allowed.\n");
     568           0 :                 pthread_mutex_unlock(&g_devlist_mutex);
     569           0 :                 _free_thread(thread);
     570           0 :                 return NULL;
     571             :         }
     572         211 :         thread->id = g_thread_id++;
     573         211 :         TAILQ_INSERT_TAIL(&g_threads, thread, tailq);
     574         211 :         g_thread_count++;
     575         211 :         pthread_mutex_unlock(&g_devlist_mutex);
     576             : 
     577         211 :         SPDK_DEBUGLOG(thread, "Allocating new thread (%" PRIu64 ", %s)\n",
     578             :                       thread->id, thread->name);
     579             : 
     580         211 :         if (spdk_interrupt_mode_is_enabled()) {
     581           0 :                 thread->in_interrupt = true;
     582           0 :                 rc = thread_interrupt_create(thread);
     583           0 :                 if (rc != 0) {
     584           0 :                         _free_thread(thread);
     585           0 :                         return NULL;
     586             :                 }
     587             :         }
     588             : 
     589         211 :         if (g_new_thread_fn) {
     590           5 :                 rc = g_new_thread_fn(thread);
     591         206 :         } else if (g_thread_op_supported_fn && g_thread_op_supported_fn(SPDK_THREAD_OP_NEW)) {
     592          16 :                 rc = g_thread_op_fn(thread, SPDK_THREAD_OP_NEW);
     593             :         }
     594             : 
     595         211 :         if (rc != 0) {
     596           2 :                 _free_thread(thread);
     597           2 :                 return NULL;
     598             :         }
     599             : 
     600         209 :         thread->state = SPDK_THREAD_STATE_RUNNING;
     601             : 
     602             :         /* If this is the first thread, save it as the app thread.  Use an atomic
     603             :          * compare + exchange to guard against crazy users who might try to
     604             :          * call spdk_thread_create() simultaneously on multiple threads.
     605             :          */
     606         209 :         null_thread = NULL;
     607         209 :         __atomic_compare_exchange_n(&g_app_thread, &null_thread, thread, false,
     608             :                                     __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST);
     609             : 
     610         209 :         return thread;
     611             : }
     612             : 
     613             : struct spdk_thread *
     614        1619 : spdk_thread_get_app_thread(void)
     615             : {
     616        1619 :         return g_app_thread;
     617             : }
     618             : 
     619             : bool
     620         146 : spdk_thread_is_app_thread(struct spdk_thread *thread)
     621             : {
     622         146 :         if (thread == NULL) {
     623         145 :                 thread = _get_thread();
     624             :         }
     625             : 
     626         146 :         return g_app_thread == thread;
     627             : }
     628             : 
     629             : void
     630           1 : spdk_thread_bind(struct spdk_thread *thread, bool bind)
     631             : {
     632           1 :         thread->is_bound = bind;
     633           1 : }
     634             : 
     635             : bool
     636          10 : spdk_thread_is_bound(struct spdk_thread *thread)
     637             : {
     638          10 :         return thread->is_bound;
     639             : }
     640             : 
     641             : void
     642      355086 : spdk_set_thread(struct spdk_thread *thread)
     643             : {
     644      355086 :         tls_thread = thread;
     645      355086 : }
     646             : 
     647             : static void
     648         213 : thread_exit(struct spdk_thread *thread, uint64_t now)
     649             : {
     650             :         struct spdk_poller *poller;
     651             :         struct spdk_io_channel *ch;
     652             : 
     653         213 :         if (now >= thread->exit_timeout_tsc) {
     654           1 :                 SPDK_ERRLOG("thread %s got timeout, and move it to the exited state forcefully\n",
     655             :                             thread->name);
     656           1 :                 goto exited;
     657             :         }
     658             : 
     659         212 :         if (spdk_ring_count(thread->messages) > 0) {
     660           2 :                 SPDK_INFOLOG(thread, "thread %s still has messages\n", thread->name);
     661           2 :                 return;
     662             :         }
     663             : 
     664         210 :         if (thread->for_each_count > 0) {
     665           4 :                 SPDK_INFOLOG(thread, "thread %s is still executing %u for_each_channels/threads\n",
     666             :                              thread->name, thread->for_each_count);
     667           4 :                 return;
     668             :         }
     669             : 
     670         206 :         TAILQ_FOREACH(poller, &thread->active_pollers, tailq) {
     671           2 :                 if (poller->state != SPDK_POLLER_STATE_UNREGISTERED) {
     672           2 :                         SPDK_INFOLOG(thread,
     673             :                                      "thread %s still has active poller %s\n",
     674             :                                      thread->name, poller->name);
     675           2 :                         return;
     676             :                 }
     677             :         }
     678             : 
     679         251 :         RB_FOREACH(poller, timed_pollers_tree, &thread->timed_pollers) {
     680          47 :                 if (poller->state != SPDK_POLLER_STATE_UNREGISTERED) {
     681           0 :                         SPDK_INFOLOG(thread,
     682             :                                      "thread %s still has active timed poller %s\n",
     683             :                                      thread->name, poller->name);
     684           0 :                         return;
     685             :                 }
     686             :         }
     687             : 
     688         204 :         TAILQ_FOREACH(poller, &thread->paused_pollers, tailq) {
     689           0 :                 SPDK_INFOLOG(thread,
     690             :                              "thread %s still has paused poller %s\n",
     691             :                              thread->name, poller->name);
     692           0 :                 return;
     693             :         }
     694             : 
     695         204 :         RB_FOREACH(ch, io_channel_tree, &thread->io_channels) {
     696           6 :                 SPDK_INFOLOG(thread,
     697             :                              "thread %s still has channel for io_device %s\n",
     698             :                              thread->name, ch->dev->name);
     699           6 :                 return;
     700             :         }
     701             : 
     702         198 :         if (thread->pending_unregister_count > 0) {
     703           2 :                 SPDK_INFOLOG(thread,
     704             :                              "thread %s is still unregistering io_devices\n",
     705             :                              thread->name);
     706           2 :                 return;
     707             :         }
     708             : 
     709         196 : exited:
     710         197 :         thread->state = SPDK_THREAD_STATE_EXITED;
     711         197 :         if (spdk_unlikely(thread->in_interrupt)) {
     712           0 :                 g_thread_op_fn(thread, SPDK_THREAD_OP_RESCHED);
     713             :         }
     714             : }
     715             : 
     716             : static void _thread_exit(void *ctx);
     717             : 
     718             : int
     719         208 : spdk_thread_exit(struct spdk_thread *thread)
     720             : {
     721         208 :         SPDK_DEBUGLOG(thread, "Exit thread %s\n", thread->name);
     722             : 
     723         208 :         assert(tls_thread == thread);
     724             : 
     725         208 :         if (thread->state >= SPDK_THREAD_STATE_EXITING) {
     726           9 :                 SPDK_INFOLOG(thread,
     727             :                              "thread %s is already exiting\n",
     728             :                              thread->name);
     729           9 :                 return 0;
     730             :         }
     731             : 
     732         199 :         thread->exit_timeout_tsc = spdk_get_ticks() + (spdk_get_ticks_hz() *
     733             :                                    SPDK_THREAD_EXIT_TIMEOUT_SEC);
     734         199 :         thread->state = SPDK_THREAD_STATE_EXITING;
     735             : 
     736         199 :         if (spdk_interrupt_mode_is_enabled()) {
     737           0 :                 spdk_thread_send_msg(thread, _thread_exit, thread);
     738             :         }
     739             : 
     740         199 :         return 0;
     741             : }
     742             : 
     743             : bool
     744           0 : spdk_thread_is_running(struct spdk_thread *thread)
     745             : {
     746           0 :         return thread->state == SPDK_THREAD_STATE_RUNNING;
     747             : }
     748             : 
     749             : bool
     750         395 : spdk_thread_is_exited(struct spdk_thread *thread)
     751             : {
     752         395 :         return thread->state == SPDK_THREAD_STATE_EXITED;
     753             : }
     754             : 
     755             : void
     756         178 : spdk_thread_destroy(struct spdk_thread *thread)
     757             : {
     758         178 :         assert(thread != NULL);
     759         178 :         SPDK_DEBUGLOG(thread, "Destroy thread %s\n", thread->name);
     760             : 
     761         178 :         assert(thread->state == SPDK_THREAD_STATE_EXITED);
     762             : 
     763         178 :         if (tls_thread == thread) {
     764         171 :                 tls_thread = NULL;
     765             :         }
     766             : 
     767             :         /* To be safe, do not free the app thread until spdk_thread_lib_fini(). */
     768         178 :         if (thread != g_app_thread) {
     769          83 :                 _free_thread(thread);
     770             :         }
     771         178 : }
     772             : 
     773             : void *
     774          49 : spdk_thread_get_ctx(struct spdk_thread *thread)
     775             : {
     776          49 :         if (g_ctx_sz > 0) {
     777          49 :                 return thread->ctx;
     778             :         }
     779             : 
     780           0 :         return NULL;
     781             : }
     782             : 
     783             : struct spdk_cpuset *
     784          46 : spdk_thread_get_cpumask(struct spdk_thread *thread)
     785             : {
     786          46 :         return &thread->cpumask;
     787             : }
     788             : 
     789             : int
     790           2 : spdk_thread_set_cpumask(struct spdk_cpuset *cpumask)
     791             : {
     792             :         struct spdk_thread *thread;
     793             : 
     794           2 :         if (!g_thread_op_supported_fn || !g_thread_op_supported_fn(SPDK_THREAD_OP_RESCHED)) {
     795           0 :                 SPDK_ERRLOG("Framework does not support reschedule operation.\n");
     796           0 :                 assert(false);
     797             :                 return -ENOTSUP;
     798             :         }
     799             : 
     800           2 :         thread = spdk_get_thread();
     801           2 :         if (!thread) {
     802           0 :                 SPDK_ERRLOG("Called from non-SPDK thread\n");
     803           0 :                 assert(false);
     804             :                 return -EINVAL;
     805             :         }
     806             : 
     807           2 :         spdk_cpuset_copy(&thread->cpumask, cpumask);
     808             : 
     809             :         /* Invoke framework's reschedule operation. If this function is called multiple times
     810             :          * in a single spdk_thread_poll() context, the last cpumask will be used in the
     811             :          * reschedule operation.
     812             :          */
     813           2 :         g_thread_op_fn(thread, SPDK_THREAD_OP_RESCHED);
     814             : 
     815           2 :         return 0;
     816             : }
     817             : 
     818             : struct spdk_thread *
     819         191 : spdk_thread_get_from_ctx(void *ctx)
     820             : {
     821         191 :         if (ctx == NULL) {
     822           0 :                 assert(false);
     823             :                 return NULL;
     824             :         }
     825             : 
     826         191 :         assert(g_ctx_sz > 0);
     827             : 
     828         191 :         return SPDK_CONTAINEROF(ctx, struct spdk_thread, ctx);
     829             : }
     830             : 
     831             : static inline uint32_t
     832      337929 : msg_queue_run_batch(struct spdk_thread *thread, uint32_t max_msgs)
     833             : {
     834             :         unsigned count, i;
     835      337929 :         void *messages[SPDK_MSG_BATCH_SIZE];
     836      337929 :         uint64_t notify = 1;
     837             :         int rc;
     838             : 
     839             : #ifdef DEBUG
     840             :         /*
     841             :          * spdk_ring_dequeue() fills messages and returns how many entries it wrote,
     842             :          * so we will never actually read uninitialized data from events, but just to be sure
     843             :          * (and to silence a static analyzer false positive), initialize the array to NULL pointers.
     844             :          */
     845      337929 :         memset(messages, 0, sizeof(messages));
     846             : #endif
     847             : 
     848      337929 :         if (max_msgs > 0) {
     849         336 :                 max_msgs = spdk_min(max_msgs, SPDK_MSG_BATCH_SIZE);
     850             :         } else {
     851      337593 :                 max_msgs = SPDK_MSG_BATCH_SIZE;
     852             :         }
     853             : 
     854      337929 :         count = spdk_ring_dequeue(thread->messages, messages, max_msgs);
     855      337929 :         if (spdk_unlikely(thread->in_interrupt) &&
     856           0 :             spdk_ring_count(thread->messages) != 0) {
     857           0 :                 rc = write(thread->msg_fd, &notify, sizeof(notify));
     858           0 :                 if (rc < 0) {
     859           0 :                         SPDK_ERRLOG("failed to notify msg_queue: %s.\n", spdk_strerror(errno));
     860             :                 }
     861             :         }
     862      337929 :         if (count == 0) {
     863      201413 :                 return 0;
     864             :         }
     865             : 
     866      278145 :         for (i = 0; i < count; i++) {
     867      141629 :                 struct spdk_msg *msg = messages[i];
     868             : 
     869      141629 :                 assert(msg != NULL);
     870             : 
     871             :                 SPDK_DTRACE_PROBE2(msg_exec, msg->fn, msg->arg);
     872             : 
     873      141629 :                 msg->fn(msg->arg);
     874             : 
     875      141629 :                 SPIN_ASSERT(thread->lock_count == 0, SPIN_ERR_HOLD_DURING_SWITCH);
     876             : 
     877      141629 :                 if (thread->msg_cache_count < SPDK_MSG_MEMPOOL_CACHE_SIZE) {
     878             :                         /* Insert the messages at the head. We want to re-use the hot
     879             :                          * ones. */
     880      141492 :                         SLIST_INSERT_HEAD(&thread->msg_cache, msg, link);
     881      141492 :                         thread->msg_cache_count++;
     882             :                 } else {
     883         137 :                         spdk_mempool_put(g_spdk_msg_mempool, msg);
     884             :                 }
     885             :         }
     886             : 
     887      136516 :         return count;
     888             : }
     889             : 
     890             : static void
     891         546 : poller_insert_timer(struct spdk_thread *thread, struct spdk_poller *poller, uint64_t now)
     892             : {
     893             :         struct spdk_poller *tmp __attribute__((unused));
     894             : 
     895         546 :         poller->next_run_tick = now + poller->period_ticks;
     896             : 
     897             :         /*
     898             :          * Insert poller in the thread's timed_pollers tree by next scheduled run time
     899             :          * as its key.
     900             :          */
     901         546 :         tmp = RB_INSERT(timed_pollers_tree, &thread->timed_pollers, poller);
     902         546 :         assert(tmp == NULL);
     903             : 
     904             :         /* Update the cache only if it is empty or the inserted poller is earlier than it.
     905             :          * RB_MIN() is not necessary here because all pollers, which has exactly the same
     906             :          * next_run_tick as the existing poller, are inserted on the right side.
     907             :          */
     908         546 :         if (thread->first_timed_poller == NULL ||
     909         458 :             poller->next_run_tick < thread->first_timed_poller->next_run_tick) {
     910         254 :                 thread->first_timed_poller = poller;
     911             :         }
     912         546 : }
     913             : 
     914             : static inline void
     915           0 : poller_remove_timer(struct spdk_thread *thread, struct spdk_poller *poller)
     916             : {
     917             :         struct spdk_poller *tmp __attribute__((unused));
     918             : 
     919           0 :         tmp = RB_REMOVE(timed_pollers_tree, &thread->timed_pollers, poller);
     920           0 :         assert(tmp != NULL);
     921             : 
     922             :         /* This function is not used in any case that is performance critical.
     923             :          * Update the cache simply by RB_MIN() if it needs to be changed.
     924             :          */
     925           0 :         if (thread->first_timed_poller == poller) {
     926           0 :                 thread->first_timed_poller = RB_MIN(timed_pollers_tree, &thread->timed_pollers);
     927             :         }
     928           0 : }
     929             : 
     930             : static void
     931         882 : thread_insert_poller(struct spdk_thread *thread, struct spdk_poller *poller)
     932             : {
     933         882 :         if (poller->period_ticks) {
     934         367 :                 poller_insert_timer(thread, poller, spdk_get_ticks());
     935             :         } else {
     936         515 :                 TAILQ_INSERT_TAIL(&thread->active_pollers, poller, tailq);
     937             :         }
     938         882 : }
     939             : 
     940             : static inline void
     941      337929 : thread_update_stats(struct spdk_thread *thread, uint64_t end,
     942             :                     uint64_t start, int rc)
     943             : {
     944      337929 :         if (rc == 0) {
     945             :                 /* Poller status idle */
     946      200873 :                 thread->stats.idle_tsc += end - start;
     947      137056 :         } else if (rc > 0) {
     948             :                 /* Poller status busy */
     949      137056 :                 thread->stats.busy_tsc += end - start;
     950             :         }
     951             :         /* Store end time to use it as start time of the next spdk_thread_poll(). */
     952      337929 :         thread->tsc_last = end;
     953      337929 : }
     954             : 
     955             : static inline int
     956        4378 : thread_execute_poller(struct spdk_thread *thread, struct spdk_poller *poller)
     957             : {
     958             :         int rc;
     959             : 
     960        4378 :         switch (poller->state) {
     961         109 :         case SPDK_POLLER_STATE_UNREGISTERED:
     962         109 :                 TAILQ_REMOVE(&thread->active_pollers, poller, tailq);
     963         109 :                 free(poller);
     964         109 :                 return 0;
     965           5 :         case SPDK_POLLER_STATE_PAUSING:
     966           5 :                 TAILQ_REMOVE(&thread->active_pollers, poller, tailq);
     967           5 :                 TAILQ_INSERT_TAIL(&thread->paused_pollers, poller, tailq);
     968           5 :                 poller->state = SPDK_POLLER_STATE_PAUSED;
     969           5 :                 return 0;
     970        4264 :         case SPDK_POLLER_STATE_WAITING:
     971        4264 :                 break;
     972           0 :         default:
     973           0 :                 assert(false);
     974             :                 break;
     975             :         }
     976             : 
     977        4264 :         poller->state = SPDK_POLLER_STATE_RUNNING;
     978        4264 :         rc = poller->fn(poller->arg);
     979             : 
     980        4264 :         SPIN_ASSERT(thread->lock_count == 0, SPIN_ERR_HOLD_DURING_SWITCH);
     981             : 
     982        4264 :         poller->run_count++;
     983        4264 :         if (rc > 0) {
     984         564 :                 poller->busy_count++;
     985             :         }
     986             : 
     987             : #ifdef DEBUG
     988        4264 :         if (rc == -1) {
     989          13 :                 SPDK_DEBUGLOG(thread, "Poller %s returned -1\n", poller->name);
     990             :         }
     991             : #endif
     992             : 
     993        4264 :         switch (poller->state) {
     994         403 :         case SPDK_POLLER_STATE_UNREGISTERED:
     995         403 :                 TAILQ_REMOVE(&thread->active_pollers, poller, tailq);
     996         403 :                 free(poller);
     997         403 :                 break;
     998           7 :         case SPDK_POLLER_STATE_PAUSING:
     999           7 :                 TAILQ_REMOVE(&thread->active_pollers, poller, tailq);
    1000           7 :                 TAILQ_INSERT_TAIL(&thread->paused_pollers, poller, tailq);
    1001           7 :                 poller->state = SPDK_POLLER_STATE_PAUSED;
    1002           7 :                 break;
    1003           0 :         case SPDK_POLLER_STATE_PAUSED:
    1004             :         case SPDK_POLLER_STATE_WAITING:
    1005           0 :                 break;
    1006        3854 :         case SPDK_POLLER_STATE_RUNNING:
    1007        3854 :                 poller->state = SPDK_POLLER_STATE_WAITING;
    1008        3854 :                 break;
    1009           0 :         default:
    1010           0 :                 assert(false);
    1011             :                 break;
    1012             :         }
    1013             : 
    1014        4264 :         return rc;
    1015             : }
    1016             : 
    1017             : static inline int
    1018         499 : thread_execute_timed_poller(struct spdk_thread *thread, struct spdk_poller *poller,
    1019             :                             uint64_t now)
    1020             : {
    1021             :         int rc;
    1022             : 
    1023         499 :         switch (poller->state) {
    1024         156 :         case SPDK_POLLER_STATE_UNREGISTERED:
    1025         156 :                 free(poller);
    1026         156 :                 return 0;
    1027          13 :         case SPDK_POLLER_STATE_PAUSING:
    1028          13 :                 TAILQ_INSERT_TAIL(&thread->paused_pollers, poller, tailq);
    1029          13 :                 poller->state = SPDK_POLLER_STATE_PAUSED;
    1030          13 :                 return 0;
    1031         330 :         case SPDK_POLLER_STATE_WAITING:
    1032         330 :                 break;
    1033           0 :         default:
    1034           0 :                 assert(false);
    1035             :                 break;
    1036             :         }
    1037             : 
    1038         330 :         poller->state = SPDK_POLLER_STATE_RUNNING;
    1039         330 :         rc = poller->fn(poller->arg);
    1040             : 
    1041         330 :         SPIN_ASSERT(thread->lock_count == 0, SPIN_ERR_HOLD_DURING_SWITCH);
    1042             : 
    1043         330 :         poller->run_count++;
    1044         330 :         if (rc > 0) {
    1045         237 :                 poller->busy_count++;
    1046             :         }
    1047             : 
    1048             : #ifdef DEBUG
    1049         330 :         if (rc == -1) {
    1050           5 :                 SPDK_DEBUGLOG(thread, "Timed poller %s returned -1\n", poller->name);
    1051             :         }
    1052             : #endif
    1053             : 
    1054         330 :         switch (poller->state) {
    1055         147 :         case SPDK_POLLER_STATE_UNREGISTERED:
    1056         147 :                 free(poller);
    1057         147 :                 break;
    1058           4 :         case SPDK_POLLER_STATE_PAUSING:
    1059           4 :                 TAILQ_INSERT_TAIL(&thread->paused_pollers, poller, tailq);
    1060           4 :                 poller->state = SPDK_POLLER_STATE_PAUSED;
    1061           4 :                 break;
    1062           0 :         case SPDK_POLLER_STATE_PAUSED:
    1063           0 :                 break;
    1064         179 :         case SPDK_POLLER_STATE_RUNNING:
    1065         179 :                 poller->state = SPDK_POLLER_STATE_WAITING;
    1066             :         /* fallthrough */
    1067         179 :         case SPDK_POLLER_STATE_WAITING:
    1068         179 :                 poller_insert_timer(thread, poller, now);
    1069         179 :                 break;
    1070           0 :         default:
    1071           0 :                 assert(false);
    1072             :                 break;
    1073             :         }
    1074             : 
    1075         330 :         return rc;
    1076             : }
    1077             : 
    1078             : static int
    1079      337929 : thread_poll(struct spdk_thread *thread, uint32_t max_msgs, uint64_t now)
    1080             : {
    1081             :         uint32_t msg_count;
    1082             :         struct spdk_poller *poller, *tmp;
    1083             :         spdk_msg_fn critical_msg;
    1084      337929 :         int rc = 0;
    1085             : 
    1086      337929 :         thread->tsc_last = now;
    1087             : 
    1088      337929 :         critical_msg = thread->critical_msg;
    1089      337929 :         if (spdk_unlikely(critical_msg != NULL)) {
    1090           0 :                 critical_msg(NULL);
    1091           0 :                 thread->critical_msg = NULL;
    1092           0 :                 rc = 1;
    1093             :         }
    1094             : 
    1095      337929 :         msg_count = msg_queue_run_batch(thread, max_msgs);
    1096      337929 :         if (msg_count) {
    1097      136516 :                 rc = 1;
    1098             :         }
    1099             : 
    1100      342307 :         TAILQ_FOREACH_REVERSE_SAFE(poller, &thread->active_pollers,
    1101             :                                    active_pollers_head, tailq, tmp) {
    1102             :                 int poller_rc;
    1103             : 
    1104        4378 :                 poller_rc = thread_execute_poller(thread, poller);
    1105        4378 :                 if (poller_rc > rc) {
    1106         342 :                         rc = poller_rc;
    1107             :                 }
    1108             :         }
    1109             : 
    1110      337929 :         poller = thread->first_timed_poller;
    1111      338428 :         while (poller != NULL) {
    1112        3353 :                 int timer_rc = 0;
    1113             : 
    1114        3353 :                 if (now < poller->next_run_tick) {
    1115        2854 :                         break;
    1116             :                 }
    1117             : 
    1118         499 :                 tmp = RB_NEXT(timed_pollers_tree, &thread->timed_pollers, poller);
    1119         499 :                 RB_REMOVE(timed_pollers_tree, &thread->timed_pollers, poller);
    1120             : 
    1121             :                 /* Update the cache to the next timed poller in the list
    1122             :                  * only if the current poller is still the closest, otherwise,
    1123             :                  * do nothing because the cache has been already updated.
    1124             :                  */
    1125         499 :                 if (thread->first_timed_poller == poller) {
    1126         499 :                         thread->first_timed_poller = tmp;
    1127             :                 }
    1128             : 
    1129         499 :                 timer_rc = thread_execute_timed_poller(thread, poller, now);
    1130         499 :                 if (timer_rc > rc) {
    1131         198 :                         rc = timer_rc;
    1132             :                 }
    1133             : 
    1134         499 :                 poller = tmp;
    1135             :         }
    1136             : 
    1137      337929 :         return rc;
    1138             : }
    1139             : 
    1140             : static void
    1141           0 : _thread_remove_pollers(void *ctx)
    1142             : {
    1143           0 :         struct spdk_thread *thread = ctx;
    1144             :         struct spdk_poller *poller, *tmp;
    1145             : 
    1146           0 :         TAILQ_FOREACH_REVERSE_SAFE(poller, &thread->active_pollers,
    1147             :                                    active_pollers_head, tailq, tmp) {
    1148           0 :                 if (poller->state == SPDK_POLLER_STATE_UNREGISTERED) {
    1149           0 :                         TAILQ_REMOVE(&thread->active_pollers, poller, tailq);
    1150           0 :                         free(poller);
    1151             :                 }
    1152             :         }
    1153             : 
    1154           0 :         RB_FOREACH_SAFE(poller, timed_pollers_tree, &thread->timed_pollers, tmp) {
    1155           0 :                 if (poller->state == SPDK_POLLER_STATE_UNREGISTERED) {
    1156           0 :                         poller_remove_timer(thread, poller);
    1157           0 :                         free(poller);
    1158             :                 }
    1159             :         }
    1160             : 
    1161           0 :         thread->poller_unregistered = false;
    1162           0 : }
    1163             : 
    1164             : static void
    1165           0 : _thread_exit(void *ctx)
    1166             : {
    1167           0 :         struct spdk_thread *thread = ctx;
    1168             : 
    1169           0 :         assert(thread->state == SPDK_THREAD_STATE_EXITING);
    1170             : 
    1171           0 :         thread_exit(thread, spdk_get_ticks());
    1172             : 
    1173           0 :         if (thread->state != SPDK_THREAD_STATE_EXITED) {
    1174           0 :                 spdk_thread_send_msg(thread, _thread_exit, thread);
    1175             :         }
    1176           0 : }
    1177             : 
    1178             : int
    1179      337929 : spdk_thread_poll(struct spdk_thread *thread, uint32_t max_msgs, uint64_t now)
    1180             : {
    1181             :         struct spdk_thread *orig_thread;
    1182             :         int rc;
    1183             : 
    1184      337929 :         orig_thread = _get_thread();
    1185      337929 :         tls_thread = thread;
    1186             : 
    1187      337929 :         if (now == 0) {
    1188      330845 :                 now = spdk_get_ticks();
    1189             :         }
    1190             : 
    1191      337929 :         if (spdk_likely(!thread->in_interrupt)) {
    1192      337929 :                 rc = thread_poll(thread, max_msgs, now);
    1193      337929 :                 if (spdk_unlikely(thread->in_interrupt)) {
    1194             :                         /* The thread transitioned to interrupt mode during the above poll.
    1195             :                          * Poll it one more time in case that during the transition time
    1196             :                          * there is msg received without notification.
    1197             :                          */
    1198           0 :                         rc = thread_poll(thread, max_msgs, now);
    1199             :                 }
    1200             : 
    1201      337929 :                 if (spdk_unlikely(thread->state == SPDK_THREAD_STATE_EXITING)) {
    1202         213 :                         thread_exit(thread, now);
    1203             :                 }
    1204             :         } else {
    1205             :                 /* Non-block wait on thread's fd_group */
    1206           0 :                 rc = spdk_fd_group_wait(thread->fgrp, 0);
    1207             :         }
    1208             : 
    1209      337929 :         thread_update_stats(thread, spdk_get_ticks(), now, rc);
    1210             : 
    1211      337929 :         tls_thread = orig_thread;
    1212             : 
    1213      337929 :         return rc;
    1214             : }
    1215             : 
    1216             : uint64_t
    1217           0 : spdk_thread_next_poller_expiration(struct spdk_thread *thread)
    1218             : {
    1219             :         struct spdk_poller *poller;
    1220             : 
    1221           0 :         poller = thread->first_timed_poller;
    1222           0 :         if (poller) {
    1223           0 :                 return poller->next_run_tick;
    1224             :         }
    1225             : 
    1226           0 :         return 0;
    1227             : }
    1228             : 
    1229             : int
    1230           0 : spdk_thread_has_active_pollers(struct spdk_thread *thread)
    1231             : {
    1232           0 :         return !TAILQ_EMPTY(&thread->active_pollers);
    1233             : }
    1234             : 
    1235             : static bool
    1236          25 : thread_has_unpaused_pollers(struct spdk_thread *thread)
    1237             : {
    1238          25 :         if (TAILQ_EMPTY(&thread->active_pollers) &&
    1239          25 :             RB_EMPTY(&thread->timed_pollers)) {
    1240          25 :                 return false;
    1241             :         }
    1242             : 
    1243           0 :         return true;
    1244             : }
    1245             : 
    1246             : bool
    1247           2 : spdk_thread_has_pollers(struct spdk_thread *thread)
    1248             : {
    1249           2 :         if (!thread_has_unpaused_pollers(thread) &&
    1250           2 :             TAILQ_EMPTY(&thread->paused_pollers)) {
    1251           2 :                 return false;
    1252             :         }
    1253             : 
    1254           0 :         return true;
    1255             : }
    1256             : 
    1257             : bool
    1258          23 : spdk_thread_is_idle(struct spdk_thread *thread)
    1259             : {
    1260          46 :         if (spdk_ring_count(thread->messages) ||
    1261          23 :             thread_has_unpaused_pollers(thread) ||
    1262          23 :             thread->critical_msg != NULL) {
    1263           0 :                 return false;
    1264             :         }
    1265             : 
    1266          23 :         return true;
    1267             : }
    1268             : 
    1269             : uint32_t
    1270          15 : spdk_thread_get_count(void)
    1271             : {
    1272             :         /*
    1273             :          * Return cached value of the current thread count.  We could acquire the
    1274             :          *  lock and iterate through the TAILQ of threads to count them, but that
    1275             :          *  count could still be invalidated after we release the lock.
    1276             :          */
    1277          15 :         return g_thread_count;
    1278             : }
    1279             : 
    1280             : struct spdk_thread *
    1281      290299 : spdk_get_thread(void)
    1282             : {
    1283      290299 :         return _get_thread();
    1284             : }
    1285             : 
    1286             : const char *
    1287           4 : spdk_thread_get_name(const struct spdk_thread *thread)
    1288             : {
    1289           4 :         return thread->name;
    1290             : }
    1291             : 
    1292             : uint64_t
    1293          15 : spdk_thread_get_id(const struct spdk_thread *thread)
    1294             : {
    1295          15 :         return thread->id;
    1296             : }
    1297             : 
    1298             : struct spdk_thread *
    1299          13 : spdk_thread_get_by_id(uint64_t id)
    1300             : {
    1301             :         struct spdk_thread *thread;
    1302             : 
    1303          13 :         if (id == 0 || id >= g_thread_id) {
    1304           0 :                 SPDK_ERRLOG("invalid thread id: %" PRIu64 ".\n", id);
    1305           0 :                 return NULL;
    1306             :         }
    1307          13 :         pthread_mutex_lock(&g_devlist_mutex);
    1308          28 :         TAILQ_FOREACH(thread, &g_threads, tailq) {
    1309          28 :                 if (thread->id == id) {
    1310          13 :                         break;
    1311             :                 }
    1312             :         }
    1313          13 :         pthread_mutex_unlock(&g_devlist_mutex);
    1314          13 :         return thread;
    1315             : }
    1316             : 
    1317             : int
    1318          57 : spdk_thread_get_stats(struct spdk_thread_stats *stats)
    1319             : {
    1320             :         struct spdk_thread *thread;
    1321             : 
    1322          57 :         thread = _get_thread();
    1323          57 :         if (!thread) {
    1324           0 :                 SPDK_ERRLOG("No thread allocated\n");
    1325           0 :                 return -EINVAL;
    1326             :         }
    1327             : 
    1328          57 :         if (stats == NULL) {
    1329           0 :                 return -EINVAL;
    1330             :         }
    1331             : 
    1332          57 :         *stats = thread->stats;
    1333             : 
    1334          57 :         return 0;
    1335             : }
    1336             : 
    1337             : uint64_t
    1338      144008 : spdk_thread_get_last_tsc(struct spdk_thread *thread)
    1339             : {
    1340      144008 :         if (thread == NULL) {
    1341           0 :                 thread = _get_thread();
    1342             :         }
    1343             : 
    1344      144008 :         return thread->tsc_last;
    1345             : }
    1346             : 
    1347             : static inline int
    1348      141638 : thread_send_msg_notification(const struct spdk_thread *target_thread)
    1349             : {
    1350      141638 :         uint64_t notify = 1;
    1351             :         int rc;
    1352             : 
    1353             :         /* Not necessary to do notification if interrupt facility is not enabled */
    1354      141638 :         if (spdk_likely(!spdk_interrupt_mode_is_enabled())) {
    1355      141638 :                 return 0;
    1356             :         }
    1357             : 
    1358             :         /* When each spdk_thread can switch between poll and interrupt mode dynamically,
    1359             :          * after sending thread msg, it is necessary to check whether target thread runs in
    1360             :          * interrupt mode and then decide whether do event notification.
    1361             :          */
    1362           0 :         if (spdk_unlikely(target_thread->in_interrupt)) {
    1363           0 :                 rc = write(target_thread->msg_fd, &notify, sizeof(notify));
    1364           0 :                 if (rc < 0) {
    1365           0 :                         SPDK_ERRLOG("failed to notify msg_queue: %s.\n", spdk_strerror(errno));
    1366           0 :                         return -EIO;
    1367             :                 }
    1368             :         }
    1369             : 
    1370           0 :         return 0;
    1371             : }
    1372             : 
    1373             : int
    1374      141638 : spdk_thread_send_msg(const struct spdk_thread *thread, spdk_msg_fn fn, void *ctx)
    1375             : {
    1376             :         struct spdk_thread *local_thread;
    1377      141638 :         struct spdk_msg *msg;
    1378             :         int rc;
    1379             : 
    1380      141638 :         assert(thread != NULL);
    1381             : 
    1382      141638 :         if (spdk_unlikely(thread->state == SPDK_THREAD_STATE_EXITED)) {
    1383           0 :                 SPDK_ERRLOG("Thread %s is marked as exited.\n", thread->name);
    1384           0 :                 return -EIO;
    1385             :         }
    1386             : 
    1387      141638 :         local_thread = _get_thread();
    1388             : 
    1389      141638 :         msg = NULL;
    1390      141638 :         if (local_thread != NULL) {
    1391      141638 :                 if (local_thread->msg_cache_count > 0) {
    1392      141638 :                         msg = SLIST_FIRST(&local_thread->msg_cache);
    1393      141638 :                         assert(msg != NULL);
    1394      141638 :                         SLIST_REMOVE_HEAD(&local_thread->msg_cache, link);
    1395      141638 :                         local_thread->msg_cache_count--;
    1396             :                 }
    1397             :         }
    1398             : 
    1399      141638 :         if (msg == NULL) {
    1400           0 :                 msg = spdk_mempool_get(g_spdk_msg_mempool);
    1401           0 :                 if (!msg) {
    1402           0 :                         SPDK_ERRLOG("msg could not be allocated\n");
    1403           0 :                         return -ENOMEM;
    1404             :                 }
    1405             :         }
    1406             : 
    1407      141638 :         msg->fn = fn;
    1408      141638 :         msg->arg = ctx;
    1409             : 
    1410      141638 :         rc = spdk_ring_enqueue(thread->messages, (void **)&msg, 1, NULL);
    1411      141638 :         if (rc != 1) {
    1412           0 :                 SPDK_ERRLOG("msg could not be enqueued\n");
    1413           0 :                 spdk_mempool_put(g_spdk_msg_mempool, msg);
    1414           0 :                 return -EIO;
    1415             :         }
    1416             : 
    1417      141638 :         return thread_send_msg_notification(thread);
    1418             : }
    1419             : 
    1420             : int
    1421           0 : spdk_thread_send_critical_msg(struct spdk_thread *thread, spdk_msg_fn fn)
    1422             : {
    1423           0 :         spdk_msg_fn expected = NULL;
    1424             : 
    1425           0 :         if (!__atomic_compare_exchange_n(&thread->critical_msg, &expected, fn, false, __ATOMIC_SEQ_CST,
    1426             :                                          __ATOMIC_SEQ_CST)) {
    1427           0 :                 return -EIO;
    1428             :         }
    1429             : 
    1430           0 :         return thread_send_msg_notification(thread);
    1431             : }
    1432             : 
    1433             : #ifdef __linux__
    1434             : static int
    1435           0 : interrupt_timerfd_process(void *arg)
    1436             : {
    1437           0 :         struct spdk_poller *poller = arg;
    1438           0 :         uint64_t exp;
    1439             :         int rc;
    1440             : 
    1441             :         /* clear the level of interval timer */
    1442           0 :         rc = read(poller->intr->efd, &exp, sizeof(exp));
    1443           0 :         if (rc < 0) {
    1444           0 :                 if (rc == -EAGAIN) {
    1445           0 :                         return 0;
    1446             :                 }
    1447             : 
    1448           0 :                 return rc;
    1449             :         }
    1450             : 
    1451             :         SPDK_DTRACE_PROBE2(timerfd_exec, poller->fn, poller->arg);
    1452             : 
    1453           0 :         return poller->fn(poller->arg);
    1454             : }
    1455             : 
    1456             : static int
    1457           0 : period_poller_interrupt_init(struct spdk_poller *poller)
    1458             : {
    1459             :         int timerfd;
    1460             : 
    1461           0 :         SPDK_DEBUGLOG(thread, "timerfd init for periodic poller %s\n", poller->name);
    1462           0 :         timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK | TFD_CLOEXEC);
    1463           0 :         if (timerfd < 0) {
    1464           0 :                 return -errno;
    1465             :         }
    1466             : 
    1467           0 :         poller->intr = spdk_interrupt_register(timerfd, interrupt_timerfd_process, poller, poller->name);
    1468           0 :         if (poller->intr == NULL) {
    1469           0 :                 close(timerfd);
    1470           0 :                 return -1;
    1471             :         }
    1472             : 
    1473           0 :         return 0;
    1474             : }
    1475             : 
    1476             : static void
    1477           0 : period_poller_set_interrupt_mode(struct spdk_poller *poller, void *cb_arg, bool interrupt_mode)
    1478             : {
    1479             :         int timerfd;
    1480           0 :         uint64_t now_tick = spdk_get_ticks();
    1481           0 :         uint64_t ticks = spdk_get_ticks_hz();
    1482             :         int ret;
    1483           0 :         struct itimerspec new_tv = {};
    1484           0 :         struct itimerspec old_tv = {};
    1485             : 
    1486           0 :         assert(poller->intr != NULL);
    1487           0 :         assert(poller->period_ticks != 0);
    1488             : 
    1489           0 :         timerfd = poller->intr->efd;
    1490             : 
    1491           0 :         assert(timerfd >= 0);
    1492             : 
    1493           0 :         SPDK_DEBUGLOG(thread, "timerfd set poller %s into %s mode\n", poller->name,
    1494             :                       interrupt_mode ? "interrupt" : "poll");
    1495             : 
    1496           0 :         if (interrupt_mode) {
    1497             :                 /* Set repeated timer expiration */
    1498           0 :                 new_tv.it_interval.tv_sec = poller->period_ticks / ticks;
    1499           0 :                 new_tv.it_interval.tv_nsec = poller->period_ticks % ticks * SPDK_SEC_TO_NSEC / ticks;
    1500             : 
    1501             :                 /* Update next timer expiration */
    1502           0 :                 if (poller->next_run_tick == 0) {
    1503           0 :                         poller->next_run_tick = now_tick + poller->period_ticks;
    1504           0 :                 } else if (poller->next_run_tick < now_tick) {
    1505           0 :                         poller->next_run_tick = now_tick;
    1506             :                 }
    1507             : 
    1508           0 :                 new_tv.it_value.tv_sec = (poller->next_run_tick - now_tick) / ticks;
    1509           0 :                 new_tv.it_value.tv_nsec = (poller->next_run_tick - now_tick) % ticks * SPDK_SEC_TO_NSEC / ticks;
    1510             : 
    1511           0 :                 ret = timerfd_settime(timerfd, 0, &new_tv, NULL);
    1512           0 :                 if (ret < 0) {
    1513           0 :                         SPDK_ERRLOG("Failed to arm timerfd: error(%d)\n", errno);
    1514           0 :                         assert(false);
    1515             :                 }
    1516             :         } else {
    1517             :                 /* Disarm the timer */
    1518           0 :                 ret = timerfd_settime(timerfd, 0, &new_tv, &old_tv);
    1519           0 :                 if (ret < 0) {
    1520             :                         /* timerfd_settime's failure indicates that the timerfd is in error */
    1521           0 :                         SPDK_ERRLOG("Failed to disarm timerfd: error(%d)\n", errno);
    1522           0 :                         assert(false);
    1523             :                 }
    1524             : 
    1525             :                 /* In order to reuse poller_insert_timer, fix now_tick, so next_run_tick would be
    1526             :                  * now_tick + ticks * old_tv.it_value.tv_sec + (ticks * old_tv.it_value.tv_nsec) / SPDK_SEC_TO_NSEC
    1527             :                  */
    1528           0 :                 now_tick = now_tick - poller->period_ticks + ticks * old_tv.it_value.tv_sec + \
    1529           0 :                            (ticks * old_tv.it_value.tv_nsec) / SPDK_SEC_TO_NSEC;
    1530           0 :                 poller_remove_timer(poller->thread, poller);
    1531           0 :                 poller_insert_timer(poller->thread, poller, now_tick);
    1532             :         }
    1533           0 : }
    1534             : 
    1535             : static void
    1536           0 : poller_interrupt_fini(struct spdk_poller *poller)
    1537             : {
    1538             :         int fd;
    1539             : 
    1540           0 :         SPDK_DEBUGLOG(thread, "interrupt fini for poller %s\n", poller->name);
    1541           0 :         assert(poller->intr != NULL);
    1542           0 :         fd = poller->intr->efd;
    1543           0 :         spdk_interrupt_unregister(&poller->intr);
    1544           0 :         close(fd);
    1545           0 : }
    1546             : 
    1547             : static int
    1548           0 : busy_poller_interrupt_init(struct spdk_poller *poller)
    1549             : {
    1550             :         int busy_efd;
    1551             : 
    1552           0 :         SPDK_DEBUGLOG(thread, "busy_efd init for busy poller %s\n", poller->name);
    1553           0 :         busy_efd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
    1554           0 :         if (busy_efd < 0) {
    1555           0 :                 SPDK_ERRLOG("Failed to create eventfd for Poller(%s).\n", poller->name);
    1556           0 :                 return -errno;
    1557             :         }
    1558             : 
    1559           0 :         poller->intr = spdk_interrupt_register(busy_efd, poller->fn, poller->arg, poller->name);
    1560           0 :         if (poller->intr == NULL) {
    1561           0 :                 close(busy_efd);
    1562           0 :                 return -1;
    1563             :         }
    1564             : 
    1565           0 :         return 0;
    1566             : }
    1567             : 
    1568             : static void
    1569           0 : busy_poller_set_interrupt_mode(struct spdk_poller *poller, void *cb_arg, bool interrupt_mode)
    1570             : {
    1571           0 :         int busy_efd = poller->intr->efd;
    1572           0 :         uint64_t notify = 1;
    1573             :         int rc __attribute__((unused));
    1574             : 
    1575           0 :         assert(busy_efd >= 0);
    1576             : 
    1577           0 :         if (interrupt_mode) {
    1578             :                 /* Write without read on eventfd will get it repeatedly triggered. */
    1579           0 :                 if (write(busy_efd, &notify, sizeof(notify)) < 0) {
    1580           0 :                         SPDK_ERRLOG("Failed to set busy wait for Poller(%s).\n", poller->name);
    1581             :                 }
    1582             :         } else {
    1583             :                 /* Read on eventfd will clear its level triggering. */
    1584           0 :                 rc = read(busy_efd, &notify, sizeof(notify));
    1585             :         }
    1586           0 : }
    1587             : 
    1588             : #else
    1589             : 
    1590             : static int
    1591             : period_poller_interrupt_init(struct spdk_poller *poller)
    1592             : {
    1593             :         return -ENOTSUP;
    1594             : }
    1595             : 
    1596             : static void
    1597             : period_poller_set_interrupt_mode(struct spdk_poller *poller, void *cb_arg, bool interrupt_mode)
    1598             : {
    1599             : }
    1600             : 
    1601             : static void
    1602             : poller_interrupt_fini(struct spdk_poller *poller)
    1603             : {
    1604             : }
    1605             : 
    1606             : static int
    1607             : busy_poller_interrupt_init(struct spdk_poller *poller)
    1608             : {
    1609             :         return -ENOTSUP;
    1610             : }
    1611             : 
    1612             : static void
    1613             : busy_poller_set_interrupt_mode(struct spdk_poller *poller, void *cb_arg, bool interrupt_mode)
    1614             : {
    1615             : }
    1616             : 
    1617             : #endif
    1618             : 
    1619             : void
    1620           6 : spdk_poller_register_interrupt(struct spdk_poller *poller,
    1621             :                                spdk_poller_set_interrupt_mode_cb cb_fn,
    1622             :                                void *cb_arg)
    1623             : {
    1624           6 :         assert(poller != NULL);
    1625           6 :         assert(spdk_get_thread() == poller->thread);
    1626             : 
    1627           6 :         if (!spdk_interrupt_mode_is_enabled()) {
    1628           6 :                 return;
    1629             :         }
    1630             : 
    1631             :         /* If this poller already had an interrupt, clean the old one up. */
    1632           0 :         if (poller->intr != NULL) {
    1633           0 :                 poller_interrupt_fini(poller);
    1634             :         }
    1635             : 
    1636           0 :         poller->set_intr_cb_fn = cb_fn;
    1637           0 :         poller->set_intr_cb_arg = cb_arg;
    1638             : 
    1639             :         /* Set poller into interrupt mode if thread is in interrupt. */
    1640           0 :         if (poller->thread->in_interrupt && poller->set_intr_cb_fn) {
    1641           0 :                 poller->set_intr_cb_fn(poller, poller->set_intr_cb_arg, true);
    1642             :         }
    1643             : }
    1644             : 
    1645             : static uint64_t
    1646         862 : convert_us_to_ticks(uint64_t us)
    1647             : {
    1648             :         uint64_t quotient, remainder, ticks;
    1649             : 
    1650         862 :         if (us) {
    1651         353 :                 quotient = us / SPDK_SEC_TO_USEC;
    1652         353 :                 remainder = us % SPDK_SEC_TO_USEC;
    1653         353 :                 ticks = spdk_get_ticks_hz();
    1654             : 
    1655         353 :                 return ticks * quotient + (ticks * remainder) / SPDK_SEC_TO_USEC;
    1656             :         } else {
    1657         509 :                 return 0;
    1658             :         }
    1659             : }
    1660             : 
    1661             : static struct spdk_poller *
    1662         862 : poller_register(spdk_poller_fn fn,
    1663             :                 void *arg,
    1664             :                 uint64_t period_microseconds,
    1665             :                 const char *name)
    1666             : {
    1667             :         struct spdk_thread *thread;
    1668             :         struct spdk_poller *poller;
    1669             : 
    1670         862 :         thread = spdk_get_thread();
    1671         862 :         if (!thread) {
    1672           0 :                 assert(false);
    1673             :                 return NULL;
    1674             :         }
    1675             : 
    1676         862 :         if (spdk_unlikely(thread->state == SPDK_THREAD_STATE_EXITED)) {
    1677           0 :                 SPDK_ERRLOG("thread %s is marked as exited\n", thread->name);
    1678           0 :                 return NULL;
    1679             :         }
    1680             : 
    1681         862 :         poller = calloc(1, sizeof(*poller));
    1682         862 :         if (poller == NULL) {
    1683           0 :                 SPDK_ERRLOG("Poller memory allocation failed\n");
    1684           0 :                 return NULL;
    1685             :         }
    1686             : 
    1687         862 :         if (name) {
    1688         801 :                 snprintf(poller->name, sizeof(poller->name), "%s", name);
    1689             :         } else {
    1690          61 :                 snprintf(poller->name, sizeof(poller->name), "%p", fn);
    1691             :         }
    1692             : 
    1693         862 :         poller->state = SPDK_POLLER_STATE_WAITING;
    1694         862 :         poller->fn = fn;
    1695         862 :         poller->arg = arg;
    1696         862 :         poller->thread = thread;
    1697         862 :         poller->intr = NULL;
    1698         862 :         if (thread->next_poller_id == 0) {
    1699           0 :                 SPDK_WARNLOG("Poller ID rolled over. Poller ID is duplicated.\n");
    1700           0 :                 thread->next_poller_id = 1;
    1701             :         }
    1702         862 :         poller->id = thread->next_poller_id++;
    1703             : 
    1704         862 :         poller->period_ticks = convert_us_to_ticks(period_microseconds);
    1705             : 
    1706         862 :         if (spdk_interrupt_mode_is_enabled()) {
    1707             :                 int rc;
    1708             : 
    1709           0 :                 if (period_microseconds) {
    1710           0 :                         rc = period_poller_interrupt_init(poller);
    1711           0 :                         if (rc < 0) {
    1712           0 :                                 SPDK_ERRLOG("Failed to register interruptfd for periodic poller: %s\n", spdk_strerror(-rc));
    1713           0 :                                 free(poller);
    1714           0 :                                 return NULL;
    1715             :                         }
    1716             : 
    1717           0 :                         poller->set_intr_cb_fn = period_poller_set_interrupt_mode;
    1718           0 :                         poller->set_intr_cb_arg = NULL;
    1719             : 
    1720             :                 } else {
    1721             :                         /* If the poller doesn't have a period, create interruptfd that's always
    1722             :                          * busy automatically when running in interrupt mode.
    1723             :                          */
    1724           0 :                         rc = busy_poller_interrupt_init(poller);
    1725           0 :                         if (rc > 0) {
    1726           0 :                                 SPDK_ERRLOG("Failed to register interruptfd for busy poller: %s\n", spdk_strerror(-rc));
    1727           0 :                                 free(poller);
    1728           0 :                                 return NULL;
    1729             :                         }
    1730             : 
    1731           0 :                         poller->set_intr_cb_fn = busy_poller_set_interrupt_mode;
    1732           0 :                         poller->set_intr_cb_arg = NULL;
    1733             :                 }
    1734             : 
    1735             :                 /* Set poller into interrupt mode if thread is in interrupt. */
    1736           0 :                 if (poller->thread->in_interrupt) {
    1737           0 :                         poller->set_intr_cb_fn(poller, poller->set_intr_cb_arg, true);
    1738             :                 }
    1739             :         }
    1740             : 
    1741         862 :         thread_insert_poller(thread, poller);
    1742             : 
    1743         862 :         return poller;
    1744             : }
    1745             : 
    1746             : struct spdk_poller *
    1747          61 : spdk_poller_register(spdk_poller_fn fn,
    1748             :                      void *arg,
    1749             :                      uint64_t period_microseconds)
    1750             : {
    1751          61 :         return poller_register(fn, arg, period_microseconds, NULL);
    1752             : }
    1753             : 
    1754             : struct spdk_poller *
    1755         801 : spdk_poller_register_named(spdk_poller_fn fn,
    1756             :                            void *arg,
    1757             :                            uint64_t period_microseconds,
    1758             :                            const char *name)
    1759             : {
    1760         801 :         return poller_register(fn, arg, period_microseconds, name);
    1761             : }
    1762             : 
    1763             : static void
    1764           0 : wrong_thread(const char *func, const char *name, struct spdk_thread *thread,
    1765             :              struct spdk_thread *curthread)
    1766             : {
    1767           0 :         if (thread == NULL) {
    1768           0 :                 SPDK_ERRLOG("%s(%s) called with NULL thread\n", func, name);
    1769           0 :                 abort();
    1770             :         }
    1771           0 :         SPDK_ERRLOG("%s(%s) called from wrong thread %s:%" PRIu64 " (should be "
    1772             :                     "%s:%" PRIu64 ")\n", func, name, curthread->name, curthread->id,
    1773             :                     thread->name, thread->id);
    1774           0 :         assert(false);
    1775             : }
    1776             : 
    1777             : void
    1778        1300 : spdk_poller_unregister(struct spdk_poller **ppoller)
    1779             : {
    1780             :         struct spdk_thread *thread;
    1781             :         struct spdk_poller *poller;
    1782             : 
    1783        1300 :         poller = *ppoller;
    1784        1300 :         if (poller == NULL) {
    1785         438 :                 return;
    1786             :         }
    1787             : 
    1788         862 :         *ppoller = NULL;
    1789             : 
    1790         862 :         thread = spdk_get_thread();
    1791         862 :         if (!thread) {
    1792           0 :                 assert(false);
    1793             :                 return;
    1794             :         }
    1795             : 
    1796         862 :         if (poller->thread != thread) {
    1797           0 :                 wrong_thread(__func__, poller->name, poller->thread, thread);
    1798           0 :                 return;
    1799             :         }
    1800             : 
    1801         862 :         if (spdk_interrupt_mode_is_enabled()) {
    1802             :                 /* Release the interrupt resource for period or busy poller */
    1803           0 :                 if (poller->intr != NULL) {
    1804           0 :                         poller_interrupt_fini(poller);
    1805             :                 }
    1806             : 
    1807             :                 /* If there is not already a pending poller removal, generate
    1808             :                  * a message to go process removals. */
    1809           0 :                 if (!thread->poller_unregistered) {
    1810           0 :                         thread->poller_unregistered = true;
    1811           0 :                         spdk_thread_send_msg(thread, _thread_remove_pollers, thread);
    1812             :                 }
    1813             :         }
    1814             : 
    1815             :         /* If the poller was paused, put it on the active_pollers list so that
    1816             :          * its unregistration can be processed by spdk_thread_poll().
    1817             :          */
    1818         862 :         if (poller->state == SPDK_POLLER_STATE_PAUSED) {
    1819           9 :                 TAILQ_REMOVE(&thread->paused_pollers, poller, tailq);
    1820           9 :                 TAILQ_INSERT_TAIL(&thread->active_pollers, poller, tailq);
    1821           9 :                 poller->period_ticks = 0;
    1822             :         }
    1823             : 
    1824             :         /* Simply set the state to unregistered. The poller will get cleaned up
    1825             :          * in a subsequent call to spdk_thread_poll().
    1826             :          */
    1827         862 :         poller->state = SPDK_POLLER_STATE_UNREGISTERED;
    1828             : }
    1829             : 
    1830             : void
    1831          41 : spdk_poller_pause(struct spdk_poller *poller)
    1832             : {
    1833             :         struct spdk_thread *thread;
    1834             : 
    1835          41 :         thread = spdk_get_thread();
    1836          41 :         if (!thread) {
    1837           0 :                 assert(false);
    1838             :                 return;
    1839             :         }
    1840             : 
    1841          41 :         if (poller->thread != thread) {
    1842           0 :                 wrong_thread(__func__, poller->name, poller->thread, thread);
    1843           0 :                 return;
    1844             :         }
    1845             : 
    1846             :         /* We just set its state to SPDK_POLLER_STATE_PAUSING and let
    1847             :          * spdk_thread_poll() move it. It allows a poller to be paused from
    1848             :          * another one's context without breaking the TAILQ_FOREACH_REVERSE_SAFE
    1849             :          * iteration, or from within itself without breaking the logic to always
    1850             :          * remove the closest timed poller in the TAILQ_FOREACH_SAFE iteration.
    1851             :          */
    1852          41 :         switch (poller->state) {
    1853           2 :         case SPDK_POLLER_STATE_PAUSED:
    1854             :         case SPDK_POLLER_STATE_PAUSING:
    1855           2 :                 break;
    1856          39 :         case SPDK_POLLER_STATE_RUNNING:
    1857             :         case SPDK_POLLER_STATE_WAITING:
    1858          39 :                 poller->state = SPDK_POLLER_STATE_PAUSING;
    1859          39 :                 break;
    1860           0 :         default:
    1861           0 :                 assert(false);
    1862             :                 break;
    1863             :         }
    1864             : }
    1865             : 
    1866             : void
    1867          40 : spdk_poller_resume(struct spdk_poller *poller)
    1868             : {
    1869             :         struct spdk_thread *thread;
    1870             : 
    1871          40 :         thread = spdk_get_thread();
    1872          40 :         if (!thread) {
    1873           0 :                 assert(false);
    1874             :                 return;
    1875             :         }
    1876             : 
    1877          40 :         if (poller->thread != thread) {
    1878           0 :                 wrong_thread(__func__, poller->name, poller->thread, thread);
    1879           0 :                 return;
    1880             :         }
    1881             : 
    1882             :         /* If a poller is paused it has to be removed from the paused pollers
    1883             :          * list and put on the active list or timer tree depending on its
    1884             :          * period_ticks.  If a poller is still in the process of being paused,
    1885             :          * we just need to flip its state back to waiting, as it's already on
    1886             :          * the appropriate list or tree.
    1887             :          */
    1888          40 :         switch (poller->state) {
    1889          20 :         case SPDK_POLLER_STATE_PAUSED:
    1890          20 :                 TAILQ_REMOVE(&thread->paused_pollers, poller, tailq);
    1891          20 :                 thread_insert_poller(thread, poller);
    1892             :         /* fallthrough */
    1893          29 :         case SPDK_POLLER_STATE_PAUSING:
    1894          29 :                 poller->state = SPDK_POLLER_STATE_WAITING;
    1895          29 :                 break;
    1896          11 :         case SPDK_POLLER_STATE_RUNNING:
    1897             :         case SPDK_POLLER_STATE_WAITING:
    1898          11 :                 break;
    1899           0 :         default:
    1900           0 :                 assert(false);
    1901             :                 break;
    1902             :         }
    1903             : }
    1904             : 
    1905             : const char *
    1906           2 : spdk_poller_get_name(struct spdk_poller *poller)
    1907             : {
    1908           2 :         return poller->name;
    1909             : }
    1910             : 
    1911             : uint64_t
    1912           3 : spdk_poller_get_id(struct spdk_poller *poller)
    1913             : {
    1914           3 :         return poller->id;
    1915             : }
    1916             : 
    1917             : const char *
    1918           9 : spdk_poller_get_state_str(struct spdk_poller *poller)
    1919             : {
    1920           9 :         switch (poller->state) {
    1921           4 :         case SPDK_POLLER_STATE_WAITING:
    1922           4 :                 return "waiting";
    1923           2 :         case SPDK_POLLER_STATE_RUNNING:
    1924           2 :                 return "running";
    1925           0 :         case SPDK_POLLER_STATE_UNREGISTERED:
    1926           0 :                 return "unregistered";
    1927           1 :         case SPDK_POLLER_STATE_PAUSING:
    1928           1 :                 return "pausing";
    1929           2 :         case SPDK_POLLER_STATE_PAUSED:
    1930           2 :                 return "paused";
    1931           0 :         default:
    1932           0 :                 return NULL;
    1933             :         }
    1934             : }
    1935             : 
    1936             : uint64_t
    1937           2 : spdk_poller_get_period_ticks(struct spdk_poller *poller)
    1938             : {
    1939           2 :         return poller->period_ticks;
    1940             : }
    1941             : 
    1942             : void
    1943           2 : spdk_poller_get_stats(struct spdk_poller *poller, struct spdk_poller_stats *stats)
    1944             : {
    1945           2 :         stats->run_count = poller->run_count;
    1946           2 :         stats->busy_count = poller->busy_count;
    1947           2 : }
    1948             : 
    1949             : struct spdk_poller *
    1950           0 : spdk_thread_get_first_active_poller(struct spdk_thread *thread)
    1951             : {
    1952           0 :         return TAILQ_FIRST(&thread->active_pollers);
    1953             : }
    1954             : 
    1955             : struct spdk_poller *
    1956           0 : spdk_thread_get_next_active_poller(struct spdk_poller *prev)
    1957             : {
    1958           0 :         return TAILQ_NEXT(prev, tailq);
    1959             : }
    1960             : 
    1961             : struct spdk_poller *
    1962           0 : spdk_thread_get_first_timed_poller(struct spdk_thread *thread)
    1963             : {
    1964           0 :         return RB_MIN(timed_pollers_tree, &thread->timed_pollers);
    1965             : }
    1966             : 
    1967             : struct spdk_poller *
    1968           0 : spdk_thread_get_next_timed_poller(struct spdk_poller *prev)
    1969             : {
    1970           0 :         return RB_NEXT(timed_pollers_tree, &thread->timed_pollers, prev);
    1971             : }
    1972             : 
    1973             : struct spdk_poller *
    1974           0 : spdk_thread_get_first_paused_poller(struct spdk_thread *thread)
    1975             : {
    1976           0 :         return TAILQ_FIRST(&thread->paused_pollers);
    1977             : }
    1978             : 
    1979             : struct spdk_poller *
    1980           0 : spdk_thread_get_next_paused_poller(struct spdk_poller *prev)
    1981             : {
    1982           0 :         return TAILQ_NEXT(prev, tailq);
    1983             : }
    1984             : 
    1985             : struct spdk_io_channel *
    1986           0 : spdk_thread_get_first_io_channel(struct spdk_thread *thread)
    1987             : {
    1988           0 :         return RB_MIN(io_channel_tree, &thread->io_channels);
    1989             : }
    1990             : 
    1991             : struct spdk_io_channel *
    1992           0 : spdk_thread_get_next_io_channel(struct spdk_io_channel *prev)
    1993             : {
    1994           0 :         return RB_NEXT(io_channel_tree, &thread->io_channels, prev);
    1995             : }
    1996             : 
    1997             : uint16_t
    1998           0 : spdk_thread_get_trace_id(struct spdk_thread *thread)
    1999             : {
    2000           0 :         return thread->trace_id;
    2001             : }
    2002             : 
    2003             : struct call_thread {
    2004             :         struct spdk_thread *cur_thread;
    2005             :         spdk_msg_fn fn;
    2006             :         void *ctx;
    2007             : 
    2008             :         struct spdk_thread *orig_thread;
    2009             :         spdk_msg_fn cpl;
    2010             : };
    2011             : 
    2012             : static void
    2013           2 : _back_to_orig_thread(void *ctx)
    2014             : {
    2015           2 :         struct call_thread *ct = ctx;
    2016             : 
    2017           2 :         assert(ct->orig_thread->for_each_count > 0);
    2018           2 :         ct->orig_thread->for_each_count--;
    2019             : 
    2020           2 :         ct->cpl(ct->ctx);
    2021           2 :         free(ctx);
    2022           2 : }
    2023             : 
    2024             : static void
    2025           6 : _on_thread(void *ctx)
    2026             : {
    2027           6 :         struct call_thread *ct = ctx;
    2028             :         int rc __attribute__((unused));
    2029             : 
    2030           6 :         ct->fn(ct->ctx);
    2031             : 
    2032           6 :         pthread_mutex_lock(&g_devlist_mutex);
    2033           6 :         ct->cur_thread = TAILQ_NEXT(ct->cur_thread, tailq);
    2034           6 :         while (ct->cur_thread && ct->cur_thread->state != SPDK_THREAD_STATE_RUNNING) {
    2035           0 :                 SPDK_DEBUGLOG(thread, "thread %s is not running but still not destroyed.\n",
    2036             :                               ct->cur_thread->name);
    2037           0 :                 ct->cur_thread = TAILQ_NEXT(ct->cur_thread, tailq);
    2038             :         }
    2039           6 :         pthread_mutex_unlock(&g_devlist_mutex);
    2040             : 
    2041           6 :         if (!ct->cur_thread) {
    2042           2 :                 SPDK_DEBUGLOG(thread, "Completed thread iteration\n");
    2043             : 
    2044           2 :                 rc = spdk_thread_send_msg(ct->orig_thread, _back_to_orig_thread, ctx);
    2045             :         } else {
    2046           4 :                 SPDK_DEBUGLOG(thread, "Continuing thread iteration to %s\n",
    2047             :                               ct->cur_thread->name);
    2048             : 
    2049           4 :                 rc = spdk_thread_send_msg(ct->cur_thread, _on_thread, ctx);
    2050             :         }
    2051           6 :         assert(rc == 0);
    2052           6 : }
    2053             : 
    2054             : void
    2055           2 : spdk_for_each_thread(spdk_msg_fn fn, void *ctx, spdk_msg_fn cpl)
    2056             : {
    2057             :         struct call_thread *ct;
    2058             :         struct spdk_thread *thread;
    2059             :         int rc __attribute__((unused));
    2060             : 
    2061           2 :         ct = calloc(1, sizeof(*ct));
    2062           2 :         if (!ct) {
    2063           0 :                 SPDK_ERRLOG("Unable to perform thread iteration\n");
    2064           0 :                 cpl(ctx);
    2065           0 :                 return;
    2066             :         }
    2067             : 
    2068           2 :         ct->fn = fn;
    2069           2 :         ct->ctx = ctx;
    2070           2 :         ct->cpl = cpl;
    2071             : 
    2072           2 :         thread = _get_thread();
    2073           2 :         if (!thread) {
    2074           0 :                 SPDK_ERRLOG("No thread allocated\n");
    2075           0 :                 free(ct);
    2076           0 :                 cpl(ctx);
    2077           0 :                 return;
    2078             :         }
    2079           2 :         ct->orig_thread = thread;
    2080             : 
    2081           2 :         ct->orig_thread->for_each_count++;
    2082             : 
    2083           2 :         pthread_mutex_lock(&g_devlist_mutex);
    2084           2 :         ct->cur_thread = TAILQ_FIRST(&g_threads);
    2085           2 :         pthread_mutex_unlock(&g_devlist_mutex);
    2086             : 
    2087           2 :         SPDK_DEBUGLOG(thread, "Starting thread iteration from %s\n",
    2088             :                       ct->orig_thread->name);
    2089             : 
    2090           2 :         rc = spdk_thread_send_msg(ct->cur_thread, _on_thread, ct);
    2091           2 :         assert(rc == 0);
    2092             : }
    2093             : 
    2094             : static inline void
    2095           0 : poller_set_interrupt_mode(struct spdk_poller *poller, bool interrupt_mode)
    2096             : {
    2097           0 :         if (poller->state == SPDK_POLLER_STATE_UNREGISTERED) {
    2098           0 :                 return;
    2099             :         }
    2100             : 
    2101           0 :         if (poller->set_intr_cb_fn) {
    2102           0 :                 poller->set_intr_cb_fn(poller, poller->set_intr_cb_arg, interrupt_mode);
    2103             :         }
    2104             : }
    2105             : 
    2106             : void
    2107           0 : spdk_thread_set_interrupt_mode(bool enable_interrupt)
    2108             : {
    2109           0 :         struct spdk_thread *thread = _get_thread();
    2110             :         struct spdk_poller *poller, *tmp;
    2111             : 
    2112           0 :         assert(thread);
    2113           0 :         assert(spdk_interrupt_mode_is_enabled());
    2114             : 
    2115           0 :         SPDK_NOTICELOG("Set spdk_thread (%s) to %s mode from %s mode.\n",
    2116             :                        thread->name,  enable_interrupt ? "intr" : "poll",
    2117             :                        thread->in_interrupt ? "intr" : "poll");
    2118             : 
    2119           0 :         if (thread->in_interrupt == enable_interrupt) {
    2120           0 :                 return;
    2121             :         }
    2122             : 
    2123             :         /* Set pollers to expected mode */
    2124           0 :         RB_FOREACH_SAFE(poller, timed_pollers_tree, &thread->timed_pollers, tmp) {
    2125           0 :                 poller_set_interrupt_mode(poller, enable_interrupt);
    2126             :         }
    2127           0 :         TAILQ_FOREACH_SAFE(poller, &thread->active_pollers, tailq, tmp) {
    2128           0 :                 poller_set_interrupt_mode(poller, enable_interrupt);
    2129             :         }
    2130             :         /* All paused pollers will go to work in interrupt mode */
    2131           0 :         TAILQ_FOREACH_SAFE(poller, &thread->paused_pollers, tailq, tmp) {
    2132           0 :                 poller_set_interrupt_mode(poller, enable_interrupt);
    2133             :         }
    2134             : 
    2135           0 :         thread->in_interrupt = enable_interrupt;
    2136           0 :         return;
    2137             : }
    2138             : 
    2139             : static struct io_device *
    2140        8178 : io_device_get(void *io_device)
    2141             : {
    2142        8178 :         struct io_device find = {};
    2143             : 
    2144        8178 :         find.io_device = io_device;
    2145        8178 :         return RB_FIND(io_device_tree, &g_io_devices, &find);
    2146             : }
    2147             : 
    2148             : void
    2149        2359 : spdk_io_device_register(void *io_device, spdk_io_channel_create_cb create_cb,
    2150             :                         spdk_io_channel_destroy_cb destroy_cb, uint32_t ctx_size,
    2151             :                         const char *name)
    2152             : {
    2153             :         struct io_device *dev, *tmp;
    2154             :         struct spdk_thread *thread;
    2155             : 
    2156        2359 :         assert(io_device != NULL);
    2157        2359 :         assert(create_cb != NULL);
    2158        2359 :         assert(destroy_cb != NULL);
    2159             : 
    2160        2359 :         thread = spdk_get_thread();
    2161        2359 :         if (!thread) {
    2162           0 :                 SPDK_ERRLOG("called from non-SPDK thread\n");
    2163           0 :                 assert(false);
    2164             :                 return;
    2165             :         }
    2166             : 
    2167        2359 :         dev = calloc(1, sizeof(struct io_device));
    2168        2359 :         if (dev == NULL) {
    2169           0 :                 SPDK_ERRLOG("could not allocate io_device\n");
    2170           0 :                 return;
    2171             :         }
    2172             : 
    2173        2359 :         dev->io_device = io_device;
    2174        2359 :         if (name) {
    2175        1451 :                 snprintf(dev->name, sizeof(dev->name), "%s", name);
    2176             :         } else {
    2177         908 :                 snprintf(dev->name, sizeof(dev->name), "%p", dev);
    2178             :         }
    2179        2359 :         dev->create_cb = create_cb;
    2180        2359 :         dev->destroy_cb = destroy_cb;
    2181        2359 :         dev->unregister_cb = NULL;
    2182        2359 :         dev->ctx_size = ctx_size;
    2183        2359 :         dev->for_each_count = 0;
    2184        2359 :         dev->unregistered = false;
    2185        2359 :         dev->refcnt = 0;
    2186             : 
    2187        2359 :         SPDK_DEBUGLOG(thread, "Registering io_device %s (%p) on thread %s\n",
    2188             :                       dev->name, dev->io_device, thread->name);
    2189             : 
    2190        2359 :         pthread_mutex_lock(&g_devlist_mutex);
    2191        2359 :         tmp = RB_INSERT(io_device_tree, &g_io_devices, dev);
    2192        2359 :         if (tmp != NULL) {
    2193           2 :                 SPDK_ERRLOG("io_device %p already registered (old:%s new:%s)\n",
    2194             :                             io_device, tmp->name, dev->name);
    2195           2 :                 free(dev);
    2196             :         }
    2197             : 
    2198        2359 :         pthread_mutex_unlock(&g_devlist_mutex);
    2199             : }
    2200             : 
    2201             : static void
    2202        2167 : _finish_unregister(void *arg)
    2203             : {
    2204        2167 :         struct io_device *dev = arg;
    2205             :         struct spdk_thread *thread;
    2206             : 
    2207        2167 :         thread = spdk_get_thread();
    2208        2167 :         assert(thread == dev->unregister_thread);
    2209             : 
    2210        2167 :         SPDK_DEBUGLOG(thread, "Finishing unregistration of io_device %s (%p) on thread %s\n",
    2211             :                       dev->name, dev->io_device, thread->name);
    2212             : 
    2213        2167 :         assert(thread->pending_unregister_count > 0);
    2214        2167 :         thread->pending_unregister_count--;
    2215             : 
    2216        2167 :         dev->unregister_cb(dev->io_device);
    2217        2167 :         free(dev);
    2218        2167 : }
    2219             : 
    2220             : static void
    2221        2356 : io_device_free(struct io_device *dev)
    2222             : {
    2223             :         int rc __attribute__((unused));
    2224             : 
    2225        2356 :         if (dev->unregister_cb == NULL) {
    2226         189 :                 free(dev);
    2227             :         } else {
    2228        2167 :                 assert(dev->unregister_thread != NULL);
    2229        2167 :                 SPDK_DEBUGLOG(thread, "io_device %s (%p) needs to unregister from thread %s\n",
    2230             :                               dev->name, dev->io_device, dev->unregister_thread->name);
    2231        2167 :                 rc = spdk_thread_send_msg(dev->unregister_thread, _finish_unregister, dev);
    2232        2167 :                 assert(rc == 0);
    2233             :         }
    2234        2356 : }
    2235             : 
    2236             : void
    2237        2358 : spdk_io_device_unregister(void *io_device, spdk_io_device_unregister_cb unregister_cb)
    2238             : {
    2239             :         struct io_device *dev;
    2240             :         uint32_t refcnt;
    2241             :         struct spdk_thread *thread;
    2242             : 
    2243        2358 :         thread = spdk_get_thread();
    2244        2358 :         if (!thread) {
    2245           0 :                 SPDK_ERRLOG("called from non-SPDK thread\n");
    2246           0 :                 assert(false);
    2247             :                 return;
    2248             :         }
    2249             : 
    2250        2358 :         pthread_mutex_lock(&g_devlist_mutex);
    2251        2358 :         dev = io_device_get(io_device);
    2252        2358 :         if (!dev) {
    2253           0 :                 SPDK_ERRLOG("io_device %p not found\n", io_device);
    2254           0 :                 assert(false);
    2255             :                 pthread_mutex_unlock(&g_devlist_mutex);
    2256             :                 return;
    2257             :         }
    2258             : 
    2259             :         /* The for_each_count check differentiates the user attempting to unregister the
    2260             :          * device a second time, from the internal call to this function that occurs
    2261             :          * after the for_each_count reaches 0.
    2262             :          */
    2263        2358 :         if (dev->pending_unregister && dev->for_each_count > 0) {
    2264           0 :                 SPDK_ERRLOG("io_device %p already has a pending unregister\n", io_device);
    2265           0 :                 assert(false);
    2266             :                 pthread_mutex_unlock(&g_devlist_mutex);
    2267             :                 return;
    2268             :         }
    2269             : 
    2270        2358 :         dev->unregister_cb = unregister_cb;
    2271        2358 :         dev->unregister_thread = thread;
    2272             : 
    2273        2358 :         if (dev->for_each_count > 0) {
    2274           1 :                 SPDK_WARNLOG("io_device %s (%p) has %u for_each calls outstanding\n",
    2275             :                              dev->name, io_device, dev->for_each_count);
    2276           1 :                 dev->pending_unregister = true;
    2277           1 :                 pthread_mutex_unlock(&g_devlist_mutex);
    2278           1 :                 return;
    2279             :         }
    2280             : 
    2281        2357 :         dev->unregistered = true;
    2282        2357 :         RB_REMOVE(io_device_tree, &g_io_devices, dev);
    2283        2357 :         refcnt = dev->refcnt;
    2284        2357 :         pthread_mutex_unlock(&g_devlist_mutex);
    2285             : 
    2286        2357 :         SPDK_DEBUGLOG(thread, "Unregistering io_device %s (%p) from thread %s\n",
    2287             :                       dev->name, dev->io_device, thread->name);
    2288             : 
    2289        2357 :         if (unregister_cb) {
    2290        2167 :                 thread->pending_unregister_count++;
    2291             :         }
    2292             : 
    2293        2357 :         if (refcnt > 0) {
    2294             :                 /* defer deletion */
    2295         799 :                 return;
    2296             :         }
    2297             : 
    2298        1558 :         io_device_free(dev);
    2299             : }
    2300             : 
    2301             : const char *
    2302           0 : spdk_io_device_get_name(struct io_device *dev)
    2303             : {
    2304           0 :         return dev->name;
    2305             : }
    2306             : 
    2307             : static struct spdk_io_channel *
    2308        9278 : thread_get_io_channel(struct spdk_thread *thread, struct io_device *dev)
    2309             : {
    2310        9278 :         struct spdk_io_channel find = {};
    2311             : 
    2312        9278 :         find.dev = dev;
    2313        9278 :         return RB_FIND(io_channel_tree, &thread->io_channels, &find);
    2314             : }
    2315             : 
    2316             : struct spdk_io_channel *
    2317        4253 : spdk_get_io_channel(void *io_device)
    2318             : {
    2319             :         struct spdk_io_channel *ch;
    2320             :         struct spdk_thread *thread;
    2321             :         struct io_device *dev;
    2322             :         int rc;
    2323             : 
    2324        4253 :         pthread_mutex_lock(&g_devlist_mutex);
    2325        4253 :         dev = io_device_get(io_device);
    2326        4253 :         if (dev == NULL) {
    2327           1 :                 SPDK_ERRLOG("could not find io_device %p\n", io_device);
    2328           1 :                 pthread_mutex_unlock(&g_devlist_mutex);
    2329           1 :                 return NULL;
    2330             :         }
    2331             : 
    2332        4252 :         thread = _get_thread();
    2333        4252 :         if (!thread) {
    2334           0 :                 SPDK_ERRLOG("No thread allocated\n");
    2335           0 :                 pthread_mutex_unlock(&g_devlist_mutex);
    2336           0 :                 return NULL;
    2337             :         }
    2338             : 
    2339        4252 :         if (spdk_unlikely(thread->state == SPDK_THREAD_STATE_EXITED)) {
    2340           0 :                 SPDK_ERRLOG("Thread %s is marked as exited\n", thread->name);
    2341           0 :                 pthread_mutex_unlock(&g_devlist_mutex);
    2342           0 :                 return NULL;
    2343             :         }
    2344             : 
    2345        4252 :         ch = thread_get_io_channel(thread, dev);
    2346        4252 :         if (ch != NULL) {
    2347        1258 :                 ch->ref++;
    2348             : 
    2349        1258 :                 SPDK_DEBUGLOG(thread, "Get io_channel %p for io_device %s (%p) on thread %s refcnt %u\n",
    2350             :                               ch, dev->name, dev->io_device, thread->name, ch->ref);
    2351             : 
    2352             :                 /*
    2353             :                  * An I/O channel already exists for this device on this
    2354             :                  *  thread, so return it.
    2355             :                  */
    2356        1258 :                 pthread_mutex_unlock(&g_devlist_mutex);
    2357        1258 :                 spdk_trace_record(TRACE_THREAD_IOCH_GET, 0, 0,
    2358             :                                   (uint64_t)spdk_io_channel_get_ctx(ch), ch->ref);
    2359        1258 :                 return ch;
    2360             :         }
    2361             : 
    2362        2994 :         ch = calloc(1, sizeof(*ch) + dev->ctx_size);
    2363        2994 :         if (ch == NULL) {
    2364           0 :                 SPDK_ERRLOG("could not calloc spdk_io_channel\n");
    2365           0 :                 pthread_mutex_unlock(&g_devlist_mutex);
    2366           0 :                 return NULL;
    2367             :         }
    2368             : 
    2369        2994 :         ch->dev = dev;
    2370        2994 :         ch->destroy_cb = dev->destroy_cb;
    2371        2994 :         ch->thread = thread;
    2372        2994 :         ch->ref = 1;
    2373        2994 :         ch->destroy_ref = 0;
    2374        2994 :         RB_INSERT(io_channel_tree, &thread->io_channels, ch);
    2375             : 
    2376        2994 :         SPDK_DEBUGLOG(thread, "Get io_channel %p for io_device %s (%p) on thread %s refcnt %u\n",
    2377             :                       ch, dev->name, dev->io_device, thread->name, ch->ref);
    2378             : 
    2379        2994 :         dev->refcnt++;
    2380             : 
    2381        2994 :         pthread_mutex_unlock(&g_devlist_mutex);
    2382             : 
    2383        2994 :         rc = dev->create_cb(io_device, (uint8_t *)ch + sizeof(*ch));
    2384        2994 :         if (rc != 0) {
    2385           3 :                 pthread_mutex_lock(&g_devlist_mutex);
    2386           3 :                 RB_REMOVE(io_channel_tree, &ch->thread->io_channels, ch);
    2387           3 :                 dev->refcnt--;
    2388           3 :                 free(ch);
    2389           3 :                 SPDK_ERRLOG("could not create io_channel for io_device %s (%p): %s (rc=%d)\n",
    2390             :                             dev->name, io_device, spdk_strerror(-rc), rc);
    2391           3 :                 pthread_mutex_unlock(&g_devlist_mutex);
    2392           3 :                 return NULL;
    2393             :         }
    2394             : 
    2395        2991 :         spdk_trace_record(TRACE_THREAD_IOCH_GET, 0, 0, (uint64_t)spdk_io_channel_get_ctx(ch), 1);
    2396        2991 :         return ch;
    2397             : }
    2398             : 
    2399             : static void
    2400        2993 : put_io_channel(void *arg)
    2401             : {
    2402        2993 :         struct spdk_io_channel *ch = arg;
    2403        2993 :         bool do_remove_dev = true;
    2404             :         struct spdk_thread *thread;
    2405             : 
    2406        2993 :         thread = spdk_get_thread();
    2407        2993 :         if (!thread) {
    2408           0 :                 SPDK_ERRLOG("called from non-SPDK thread\n");
    2409           0 :                 assert(false);
    2410             :                 return;
    2411             :         }
    2412             : 
    2413        2993 :         SPDK_DEBUGLOG(thread,
    2414             :                       "Releasing io_channel %p for io_device %s (%p) on thread %s\n",
    2415             :                       ch, ch->dev->name, ch->dev->io_device, thread->name);
    2416             : 
    2417        2993 :         assert(ch->thread == thread);
    2418             : 
    2419        2993 :         ch->destroy_ref--;
    2420             : 
    2421        2993 :         if (ch->ref > 0 || ch->destroy_ref > 0) {
    2422             :                 /*
    2423             :                  * Another reference to the associated io_device was requested
    2424             :                  *  after this message was sent but before it had a chance to
    2425             :                  *  execute.
    2426             :                  */
    2427           4 :                 return;
    2428             :         }
    2429             : 
    2430        2989 :         pthread_mutex_lock(&g_devlist_mutex);
    2431        2989 :         RB_REMOVE(io_channel_tree, &ch->thread->io_channels, ch);
    2432        2989 :         pthread_mutex_unlock(&g_devlist_mutex);
    2433             : 
    2434             :         /* Don't hold the devlist mutex while the destroy_cb is called. */
    2435        2989 :         ch->destroy_cb(ch->dev->io_device, spdk_io_channel_get_ctx(ch));
    2436             : 
    2437        2989 :         pthread_mutex_lock(&g_devlist_mutex);
    2438        2989 :         ch->dev->refcnt--;
    2439             : 
    2440        2989 :         if (!ch->dev->unregistered) {
    2441        2185 :                 do_remove_dev = false;
    2442             :         }
    2443             : 
    2444        2989 :         if (ch->dev->refcnt > 0) {
    2445         140 :                 do_remove_dev = false;
    2446             :         }
    2447             : 
    2448        2989 :         pthread_mutex_unlock(&g_devlist_mutex);
    2449             : 
    2450        2989 :         if (do_remove_dev) {
    2451         798 :                 io_device_free(ch->dev);
    2452             :         }
    2453        2989 :         free(ch);
    2454             : }
    2455             : 
    2456             : void
    2457        4248 : spdk_put_io_channel(struct spdk_io_channel *ch)
    2458             : {
    2459             :         struct spdk_thread *thread;
    2460             :         int rc __attribute__((unused));
    2461             : 
    2462        4248 :         spdk_trace_record(TRACE_THREAD_IOCH_PUT, 0, 0,
    2463             :                           (uint64_t)spdk_io_channel_get_ctx(ch), ch->ref);
    2464             : 
    2465        4248 :         thread = spdk_get_thread();
    2466        4248 :         if (!thread) {
    2467           0 :                 SPDK_ERRLOG("called from non-SPDK thread\n");
    2468           0 :                 assert(false);
    2469             :                 return;
    2470             :         }
    2471             : 
    2472        4248 :         if (ch->thread != thread) {
    2473           0 :                 wrong_thread(__func__, "ch", ch->thread, thread);
    2474           0 :                 return;
    2475             :         }
    2476             : 
    2477        4248 :         SPDK_DEBUGLOG(thread,
    2478             :                       "Putting io_channel %p for io_device %s (%p) on thread %s refcnt %u\n",
    2479             :                       ch, ch->dev->name, ch->dev->io_device, thread->name, ch->ref);
    2480             : 
    2481        4248 :         ch->ref--;
    2482             : 
    2483        4248 :         if (ch->ref == 0) {
    2484        2993 :                 ch->destroy_ref++;
    2485        2993 :                 rc = spdk_thread_send_msg(thread, put_io_channel, ch);
    2486        2993 :                 assert(rc == 0);
    2487             :         }
    2488             : }
    2489             : 
    2490             : struct spdk_io_channel *
    2491      656442 : spdk_io_channel_from_ctx(void *ctx)
    2492             : {
    2493      656442 :         return (struct spdk_io_channel *)((uint8_t *)ctx - sizeof(struct spdk_io_channel));
    2494             : }
    2495             : 
    2496             : struct spdk_thread *
    2497        2155 : spdk_io_channel_get_thread(struct spdk_io_channel *ch)
    2498             : {
    2499        2155 :         return ch->thread;
    2500             : }
    2501             : 
    2502             : void *
    2503      654010 : spdk_io_channel_get_io_device(struct spdk_io_channel *ch)
    2504             : {
    2505      654010 :         return ch->dev->io_device;
    2506             : }
    2507             : 
    2508             : const char *
    2509           0 : spdk_io_channel_get_io_device_name(struct spdk_io_channel *ch)
    2510             : {
    2511           0 :         return spdk_io_device_get_name(ch->dev);
    2512             : }
    2513             : 
    2514             : int
    2515           0 : spdk_io_channel_get_ref_count(struct spdk_io_channel *ch)
    2516             : {
    2517           0 :         return ch->ref;
    2518             : }
    2519             : 
    2520             : struct spdk_io_channel_iter {
    2521             :         void *io_device;
    2522             :         struct io_device *dev;
    2523             :         spdk_channel_msg fn;
    2524             :         int status;
    2525             :         void *ctx;
    2526             :         struct spdk_io_channel *ch;
    2527             : 
    2528             :         struct spdk_thread *cur_thread;
    2529             : 
    2530             :         struct spdk_thread *orig_thread;
    2531             :         spdk_channel_for_each_cpl cpl;
    2532             : };
    2533             : 
    2534             : void *
    2535         752 : spdk_io_channel_iter_get_io_device(struct spdk_io_channel_iter *i)
    2536             : {
    2537         752 :         return i->io_device;
    2538             : }
    2539             : 
    2540             : struct spdk_io_channel *
    2541        1054 : spdk_io_channel_iter_get_channel(struct spdk_io_channel_iter *i)
    2542             : {
    2543        1054 :         return i->ch;
    2544             : }
    2545             : 
    2546             : void *
    2547        2609 : spdk_io_channel_iter_get_ctx(struct spdk_io_channel_iter *i)
    2548             : {
    2549        2609 :         return i->ctx;
    2550             : }
    2551             : 
    2552             : static void
    2553        1564 : _call_completion(void *ctx)
    2554             : {
    2555        1564 :         struct spdk_io_channel_iter *i = ctx;
    2556             : 
    2557        1564 :         assert(i->orig_thread->for_each_count > 0);
    2558        1564 :         i->orig_thread->for_each_count--;
    2559             : 
    2560        1564 :         if (i->cpl != NULL) {
    2561        1564 :                 i->cpl(i, i->status);
    2562             :         }
    2563        1564 :         free(i);
    2564        1564 : }
    2565             : 
    2566             : static void
    2567        1620 : _call_channel(void *ctx)
    2568             : {
    2569        1620 :         struct spdk_io_channel_iter *i = ctx;
    2570             :         struct spdk_io_channel *ch;
    2571             : 
    2572             :         /*
    2573             :          * It is possible that the channel was deleted before this
    2574             :          *  message had a chance to execute.  If so, skip calling
    2575             :          *  the fn() on this thread.
    2576             :          */
    2577        1620 :         pthread_mutex_lock(&g_devlist_mutex);
    2578        1620 :         ch = thread_get_io_channel(i->cur_thread, i->dev);
    2579        1620 :         pthread_mutex_unlock(&g_devlist_mutex);
    2580             : 
    2581        1620 :         if (ch) {
    2582        1579 :                 i->fn(i);
    2583             :         } else {
    2584          41 :                 spdk_for_each_channel_continue(i, 0);
    2585             :         }
    2586        1620 : }
    2587             : 
    2588             : void
    2589        1564 : spdk_for_each_channel(void *io_device, spdk_channel_msg fn, void *ctx,
    2590             :                       spdk_channel_for_each_cpl cpl)
    2591             : {
    2592             :         struct spdk_thread *thread;
    2593             :         struct spdk_io_channel *ch;
    2594             :         struct spdk_io_channel_iter *i;
    2595             :         int rc __attribute__((unused));
    2596             : 
    2597        1564 :         i = calloc(1, sizeof(*i));
    2598        1564 :         if (!i) {
    2599           0 :                 SPDK_ERRLOG("Unable to allocate iterator\n");
    2600           0 :                 assert(false);
    2601             :                 return;
    2602             :         }
    2603             : 
    2604        1564 :         i->io_device = io_device;
    2605        1564 :         i->fn = fn;
    2606        1564 :         i->ctx = ctx;
    2607        1564 :         i->cpl = cpl;
    2608        1564 :         i->orig_thread = _get_thread();
    2609             : 
    2610        1564 :         i->orig_thread->for_each_count++;
    2611             : 
    2612        1564 :         pthread_mutex_lock(&g_devlist_mutex);
    2613        1564 :         i->dev = io_device_get(io_device);
    2614        1564 :         if (i->dev == NULL) {
    2615           0 :                 SPDK_ERRLOG("could not find io_device %p\n", io_device);
    2616           0 :                 assert(false);
    2617             :                 i->status = -ENODEV;
    2618             :                 goto end;
    2619             :         }
    2620             : 
    2621             :         /* Do not allow new for_each operations if we are already waiting to unregister
    2622             :          * the device for other for_each operations to complete.
    2623             :          */
    2624        1564 :         if (i->dev->pending_unregister) {
    2625           0 :                 SPDK_ERRLOG("io_device %p has a pending unregister\n", io_device);
    2626           0 :                 i->status = -ENODEV;
    2627           0 :                 goto end;
    2628             :         }
    2629             : 
    2630        1853 :         TAILQ_FOREACH(thread, &g_threads, tailq) {
    2631        1723 :                 ch = thread_get_io_channel(thread, i->dev);
    2632        1723 :                 if (ch != NULL) {
    2633        1434 :                         ch->dev->for_each_count++;
    2634        1434 :                         i->cur_thread = thread;
    2635        1434 :                         i->ch = ch;
    2636        1434 :                         pthread_mutex_unlock(&g_devlist_mutex);
    2637        1434 :                         rc = spdk_thread_send_msg(thread, _call_channel, i);
    2638        1434 :                         assert(rc == 0);
    2639        1434 :                         return;
    2640             :                 }
    2641             :         }
    2642             : 
    2643         130 : end:
    2644         130 :         pthread_mutex_unlock(&g_devlist_mutex);
    2645             : 
    2646         130 :         rc = spdk_thread_send_msg(i->orig_thread, _call_completion, i);
    2647         130 :         assert(rc == 0);
    2648             : }
    2649             : 
    2650             : static void
    2651           1 : __pending_unregister(void *arg)
    2652             : {
    2653           1 :         struct io_device *dev = arg;
    2654             : 
    2655           1 :         assert(dev->pending_unregister);
    2656           1 :         assert(dev->for_each_count == 0);
    2657           1 :         spdk_io_device_unregister(dev->io_device, dev->unregister_cb);
    2658           1 : }
    2659             : 
    2660             : void
    2661        1620 : spdk_for_each_channel_continue(struct spdk_io_channel_iter *i, int status)
    2662             : {
    2663             :         struct spdk_thread *thread;
    2664             :         struct spdk_io_channel *ch;
    2665             :         struct io_device *dev;
    2666             :         int rc __attribute__((unused));
    2667             : 
    2668        1620 :         assert(i->cur_thread == spdk_get_thread());
    2669             : 
    2670        1620 :         i->status = status;
    2671             : 
    2672        1620 :         pthread_mutex_lock(&g_devlist_mutex);
    2673        1620 :         dev = i->dev;
    2674        1620 :         if (status) {
    2675          11 :                 goto end;
    2676             :         }
    2677             : 
    2678        1609 :         thread = TAILQ_NEXT(i->cur_thread, tailq);
    2679        3106 :         while (thread) {
    2680        1683 :                 ch = thread_get_io_channel(thread, dev);
    2681        1683 :                 if (ch != NULL) {
    2682         186 :                         i->cur_thread = thread;
    2683         186 :                         i->ch = ch;
    2684         186 :                         pthread_mutex_unlock(&g_devlist_mutex);
    2685         186 :                         rc = spdk_thread_send_msg(thread, _call_channel, i);
    2686         186 :                         assert(rc == 0);
    2687         186 :                         return;
    2688             :                 }
    2689        1497 :                 thread = TAILQ_NEXT(thread, tailq);
    2690             :         }
    2691             : 
    2692        1423 : end:
    2693        1434 :         dev->for_each_count--;
    2694        1434 :         i->ch = NULL;
    2695        1434 :         pthread_mutex_unlock(&g_devlist_mutex);
    2696             : 
    2697        1434 :         rc = spdk_thread_send_msg(i->orig_thread, _call_completion, i);
    2698        1434 :         assert(rc == 0);
    2699             : 
    2700        1434 :         pthread_mutex_lock(&g_devlist_mutex);
    2701        1434 :         if (dev->pending_unregister && dev->for_each_count == 0) {
    2702           1 :                 rc = spdk_thread_send_msg(dev->unregister_thread, __pending_unregister, dev);
    2703           1 :                 assert(rc == 0);
    2704             :         }
    2705        1434 :         pthread_mutex_unlock(&g_devlist_mutex);
    2706             : }
    2707             : 
    2708             : static void
    2709           0 : thread_interrupt_destroy(struct spdk_thread *thread)
    2710             : {
    2711           0 :         struct spdk_fd_group *fgrp = thread->fgrp;
    2712             : 
    2713           0 :         SPDK_INFOLOG(thread, "destroy fgrp for thread (%s)\n", thread->name);
    2714             : 
    2715           0 :         if (thread->msg_fd < 0) {
    2716           0 :                 return;
    2717             :         }
    2718             : 
    2719           0 :         spdk_fd_group_remove(fgrp, thread->msg_fd);
    2720           0 :         close(thread->msg_fd);
    2721           0 :         thread->msg_fd = -1;
    2722             : 
    2723           0 :         spdk_fd_group_destroy(fgrp);
    2724           0 :         thread->fgrp = NULL;
    2725             : }
    2726             : 
    2727             : #ifdef __linux__
    2728             : static int
    2729           0 : thread_interrupt_msg_process(void *arg)
    2730             : {
    2731           0 :         struct spdk_thread *thread = arg;
    2732             :         struct spdk_thread *orig_thread;
    2733             :         uint32_t msg_count;
    2734             :         spdk_msg_fn critical_msg;
    2735           0 :         int rc = 0;
    2736           0 :         uint64_t notify = 1;
    2737             : 
    2738           0 :         assert(spdk_interrupt_mode_is_enabled());
    2739             : 
    2740           0 :         orig_thread = spdk_get_thread();
    2741           0 :         spdk_set_thread(thread);
    2742             : 
    2743             :         /* There may be race between msg_acknowledge and another producer's msg_notify,
    2744             :          * so msg_acknowledge should be applied ahead. And then check for self's msg_notify.
    2745             :          * This can avoid msg notification missing.
    2746             :          */
    2747           0 :         rc = read(thread->msg_fd, &notify, sizeof(notify));
    2748           0 :         if (rc < 0 && errno != EAGAIN) {
    2749           0 :                 SPDK_ERRLOG("failed to acknowledge msg event: %s.\n", spdk_strerror(errno));
    2750             :         }
    2751             : 
    2752           0 :         critical_msg = thread->critical_msg;
    2753           0 :         if (spdk_unlikely(critical_msg != NULL)) {
    2754           0 :                 critical_msg(NULL);
    2755           0 :                 thread->critical_msg = NULL;
    2756           0 :                 rc = 1;
    2757             :         }
    2758             : 
    2759           0 :         msg_count = msg_queue_run_batch(thread, 0);
    2760           0 :         if (msg_count) {
    2761           0 :                 rc = 1;
    2762             :         }
    2763             : 
    2764           0 :         SPIN_ASSERT(thread->lock_count == 0, SPIN_ERR_HOLD_DURING_SWITCH);
    2765           0 :         if (spdk_unlikely(!thread->in_interrupt)) {
    2766             :                 /* The thread transitioned to poll mode in a msg during the above processing.
    2767             :                  * Clear msg_fd since thread messages will be polled directly in poll mode.
    2768             :                  */
    2769           0 :                 rc = read(thread->msg_fd, &notify, sizeof(notify));
    2770           0 :                 if (rc < 0 && errno != EAGAIN) {
    2771           0 :                         SPDK_ERRLOG("failed to acknowledge msg queue: %s.\n", spdk_strerror(errno));
    2772             :                 }
    2773             :         }
    2774             : 
    2775           0 :         spdk_set_thread(orig_thread);
    2776           0 :         return rc;
    2777             : }
    2778             : 
    2779             : static int
    2780           0 : thread_interrupt_create(struct spdk_thread *thread)
    2781             : {
    2782             :         int rc;
    2783             : 
    2784           0 :         SPDK_INFOLOG(thread, "Create fgrp for thread (%s)\n", thread->name);
    2785             : 
    2786           0 :         rc = spdk_fd_group_create(&thread->fgrp);
    2787           0 :         if (rc) {
    2788           0 :                 return rc;
    2789             :         }
    2790             : 
    2791           0 :         thread->msg_fd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
    2792           0 :         if (thread->msg_fd < 0) {
    2793           0 :                 rc = -errno;
    2794           0 :                 spdk_fd_group_destroy(thread->fgrp);
    2795           0 :                 thread->fgrp = NULL;
    2796             : 
    2797           0 :                 return rc;
    2798             :         }
    2799             : 
    2800           0 :         return SPDK_FD_GROUP_ADD(thread->fgrp, thread->msg_fd,
    2801             :                                  thread_interrupt_msg_process, thread);
    2802             : }
    2803             : #else
    2804             : static int
    2805             : thread_interrupt_create(struct spdk_thread *thread)
    2806             : {
    2807             :         return -ENOTSUP;
    2808             : }
    2809             : #endif
    2810             : 
    2811             : static int
    2812           0 : _interrupt_wrapper(void *ctx)
    2813             : {
    2814           0 :         struct spdk_interrupt *intr = ctx;
    2815             :         struct spdk_thread *orig_thread, *thread;
    2816             :         int rc;
    2817             : 
    2818           0 :         orig_thread = spdk_get_thread();
    2819           0 :         thread = intr->thread;
    2820             : 
    2821           0 :         spdk_set_thread(thread);
    2822             : 
    2823             :         SPDK_DTRACE_PROBE4(interrupt_fd_process, intr->name, intr->efd,
    2824             :                            intr->fn, intr->arg);
    2825             : 
    2826           0 :         rc = intr->fn(intr->arg);
    2827             : 
    2828           0 :         SPIN_ASSERT(thread->lock_count == 0, SPIN_ERR_HOLD_DURING_SWITCH);
    2829             : 
    2830           0 :         spdk_set_thread(orig_thread);
    2831             : 
    2832           0 :         return rc;
    2833             : }
    2834             : 
    2835             : struct spdk_interrupt *
    2836           0 : spdk_interrupt_register(int efd, spdk_interrupt_fn fn,
    2837             :                         void *arg, const char *name)
    2838             : {
    2839           0 :         return spdk_interrupt_register_for_events(efd, SPDK_INTERRUPT_EVENT_IN, fn, arg, name);
    2840             : }
    2841             : 
    2842             : struct spdk_interrupt *
    2843           0 : spdk_interrupt_register_for_events(int efd, uint32_t events, spdk_interrupt_fn fn, void *arg,
    2844             :                                    const char *name)
    2845             : {
    2846           0 :         struct spdk_event_handler_opts opts = {};
    2847             : 
    2848           0 :         spdk_fd_group_get_default_event_handler_opts(&opts, sizeof(opts));
    2849           0 :         opts.events = events;
    2850           0 :         opts.fd_type = SPDK_FD_TYPE_DEFAULT;
    2851             : 
    2852           0 :         return spdk_interrupt_register_ext(efd, fn, arg, name, &opts);
    2853             : }
    2854             : 
    2855             : struct spdk_interrupt *
    2856           0 : spdk_interrupt_register_ext(int efd, spdk_interrupt_fn fn, void *arg, const char *name,
    2857             :                             struct spdk_event_handler_opts *opts)
    2858             : {
    2859             :         struct spdk_thread *thread;
    2860             :         struct spdk_interrupt *intr;
    2861             :         int ret;
    2862             : 
    2863           0 :         thread = spdk_get_thread();
    2864           0 :         if (!thread) {
    2865           0 :                 assert(false);
    2866             :                 return NULL;
    2867             :         }
    2868             : 
    2869           0 :         if (spdk_unlikely(thread->state != SPDK_THREAD_STATE_RUNNING)) {
    2870           0 :                 SPDK_ERRLOG("thread %s is marked as exited\n", thread->name);
    2871           0 :                 return NULL;
    2872             :         }
    2873             : 
    2874           0 :         intr = calloc(1, sizeof(*intr));
    2875           0 :         if (intr == NULL) {
    2876           0 :                 SPDK_ERRLOG("Interrupt handler allocation failed\n");
    2877           0 :                 return NULL;
    2878             :         }
    2879             : 
    2880           0 :         if (name) {
    2881           0 :                 snprintf(intr->name, sizeof(intr->name), "%s", name);
    2882             :         } else {
    2883           0 :                 snprintf(intr->name, sizeof(intr->name), "%p", fn);
    2884             :         }
    2885             : 
    2886           0 :         intr->efd = efd;
    2887           0 :         intr->thread = thread;
    2888           0 :         intr->fn = fn;
    2889           0 :         intr->arg = arg;
    2890             : 
    2891           0 :         ret = spdk_fd_group_add_ext(thread->fgrp, efd, _interrupt_wrapper, intr, intr->name, opts);
    2892             : 
    2893           0 :         if (ret != 0) {
    2894           0 :                 SPDK_ERRLOG("thread %s: failed to add fd %d: %s\n",
    2895             :                             thread->name, efd, spdk_strerror(-ret));
    2896           0 :                 free(intr);
    2897           0 :                 return NULL;
    2898             :         }
    2899             : 
    2900           0 :         return intr;
    2901             : }
    2902             : 
    2903             : void
    2904           1 : spdk_interrupt_unregister(struct spdk_interrupt **pintr)
    2905             : {
    2906             :         struct spdk_thread *thread;
    2907             :         struct spdk_interrupt *intr;
    2908             : 
    2909           1 :         intr = *pintr;
    2910           1 :         if (intr == NULL) {
    2911           1 :                 return;
    2912             :         }
    2913             : 
    2914           0 :         *pintr = NULL;
    2915             : 
    2916           0 :         thread = spdk_get_thread();
    2917           0 :         if (!thread) {
    2918           0 :                 assert(false);
    2919             :                 return;
    2920             :         }
    2921             : 
    2922           0 :         if (intr->thread != thread) {
    2923           0 :                 wrong_thread(__func__, intr->name, intr->thread, thread);
    2924           0 :                 return;
    2925             :         }
    2926             : 
    2927           0 :         spdk_fd_group_remove(thread->fgrp, intr->efd);
    2928           0 :         free(intr);
    2929             : }
    2930             : 
    2931             : int
    2932           0 : spdk_interrupt_set_event_types(struct spdk_interrupt *intr,
    2933             :                                enum spdk_interrupt_event_types event_types)
    2934             : {
    2935             :         struct spdk_thread *thread;
    2936             : 
    2937           0 :         thread = spdk_get_thread();
    2938           0 :         if (!thread) {
    2939           0 :                 assert(false);
    2940             :                 return -EINVAL;
    2941             :         }
    2942             : 
    2943           0 :         if (intr->thread != thread) {
    2944           0 :                 wrong_thread(__func__, intr->name, intr->thread, thread);
    2945           0 :                 return -EINVAL;
    2946             :         }
    2947             : 
    2948           0 :         return spdk_fd_group_event_modify(thread->fgrp, intr->efd, event_types);
    2949             : }
    2950             : 
    2951             : int
    2952           0 : spdk_thread_get_interrupt_fd(struct spdk_thread *thread)
    2953             : {
    2954           0 :         return spdk_fd_group_get_fd(thread->fgrp);
    2955             : }
    2956             : 
    2957             : struct spdk_fd_group *
    2958           0 : spdk_thread_get_interrupt_fd_group(struct spdk_thread *thread)
    2959             : {
    2960           0 :         return thread->fgrp;
    2961             : }
    2962             : 
    2963             : static bool g_interrupt_mode = false;
    2964             : 
    2965             : int
    2966           0 : spdk_interrupt_mode_enable(void)
    2967             : {
    2968             :         /* It must be called once prior to initializing the threading library.
    2969             :          * g_spdk_msg_mempool will be valid if thread library is initialized.
    2970             :          */
    2971           0 :         if (g_spdk_msg_mempool) {
    2972           0 :                 SPDK_ERRLOG("Failed due to threading library is already initialized.\n");
    2973           0 :                 return -1;
    2974             :         }
    2975             : 
    2976             : #ifdef __linux__
    2977           0 :         SPDK_NOTICELOG("Set SPDK running in interrupt mode.\n");
    2978           0 :         g_interrupt_mode = true;
    2979           0 :         return 0;
    2980             : #else
    2981             :         SPDK_ERRLOG("SPDK interrupt mode supports only Linux platform now.\n");
    2982             :         g_interrupt_mode = false;
    2983             :         return -ENOTSUP;
    2984             : #endif
    2985             : }
    2986             : 
    2987             : bool
    2988      144777 : spdk_interrupt_mode_is_enabled(void)
    2989             : {
    2990      144777 :         return g_interrupt_mode;
    2991             : }
    2992             : 
    2993             : #define SSPIN_DEBUG_STACK_FRAMES 16
    2994             : 
    2995             : struct sspin_stack {
    2996             :         void *addrs[SSPIN_DEBUG_STACK_FRAMES];
    2997             :         uint32_t depth;
    2998             : };
    2999             : 
    3000             : struct spdk_spinlock_internal {
    3001             :         struct sspin_stack init_stack;
    3002             :         struct sspin_stack lock_stack;
    3003             :         struct sspin_stack unlock_stack;
    3004             : };
    3005             : 
    3006             : static void
    3007        1273 : sspin_init_internal(struct spdk_spinlock *sspin)
    3008             : {
    3009             : #ifdef DEBUG
    3010        1273 :         sspin->internal = calloc(1, sizeof(*sspin->internal));
    3011             : #endif
    3012        1273 : }
    3013             : 
    3014             : static void
    3015        1262 : sspin_fini_internal(struct spdk_spinlock *sspin)
    3016             : {
    3017             : #ifdef DEBUG
    3018        1262 :         free(sspin->internal);
    3019        1262 :         sspin->internal = NULL;
    3020             : #endif
    3021        1262 : }
    3022             : 
    3023             : #if defined(DEBUG) && defined(SPDK_HAVE_EXECINFO_H)
    3024             : #define SSPIN_GET_STACK(sspin, which) \
    3025             :         do { \
    3026             :                 if (sspin->internal != NULL) { \
    3027             :                         struct sspin_stack *stack = &sspin->internal->which ## _stack; \
    3028             :                         stack->depth = backtrace(stack->addrs, SPDK_COUNTOF(stack->addrs)); \
    3029             :                 } \
    3030             :         } while (0)
    3031             : #else
    3032             : #define SSPIN_GET_STACK(sspin, which) do { } while (0)
    3033             : #endif
    3034             : 
    3035             : static void
    3036          15 : sspin_stack_print(const char *title, const struct sspin_stack *sspin_stack)
    3037             : {
    3038             : #ifdef SPDK_HAVE_EXECINFO_H
    3039             :         char **stack;
    3040             :         size_t i;
    3041             : 
    3042             :         stack = backtrace_symbols(sspin_stack->addrs, sspin_stack->depth);
    3043             :         if (stack == NULL) {
    3044             :                 SPDK_ERRLOG("Out of memory while allocate stack for %s\n", title);
    3045             :                 return;
    3046             :         }
    3047             :         SPDK_ERRLOG("  %s:\n", title);
    3048             :         for (i = 0; i < sspin_stack->depth; i++) {
    3049             :                 /*
    3050             :                  * This does not print line numbers. In gdb, use something like "list *0x444b6b" or
    3051             :                  * "list *sspin_stack->addrs[0]".  Or more conveniently, load the spdk gdb macros
    3052             :                  * and use use "print *sspin" or "print sspin->internal.lock_stack".  See
    3053             :                  * gdb_macros.md in the docs directory for details.
    3054             :                  */
    3055             :                 SPDK_ERRLOG("    #%" PRIu64 ": %s\n", i, stack[i]);
    3056             :         }
    3057             :         free(stack);
    3058             : #endif /* SPDK_HAVE_EXECINFO_H */
    3059          15 : }
    3060             : 
    3061             : static void
    3062           5 : sspin_stacks_print(const struct spdk_spinlock *sspin)
    3063             : {
    3064           5 :         if (sspin->internal == NULL) {
    3065           0 :                 return;
    3066             :         }
    3067           5 :         SPDK_ERRLOG("spinlock %p\n", sspin);
    3068           5 :         sspin_stack_print("Lock initialized at", &sspin->internal->init_stack);
    3069           5 :         sspin_stack_print("Last locked at", &sspin->internal->lock_stack);
    3070           5 :         sspin_stack_print("Last unlocked at", &sspin->internal->unlock_stack);
    3071             : }
    3072             : 
    3073             : void
    3074        1273 : spdk_spin_init(struct spdk_spinlock *sspin)
    3075             : {
    3076             :         int rc;
    3077             : 
    3078        1273 :         memset(sspin, 0, sizeof(*sspin));
    3079        1273 :         rc = pthread_spin_init(&sspin->spinlock, PTHREAD_PROCESS_PRIVATE);
    3080        1273 :         SPIN_ASSERT_LOG_STACKS(rc == 0, SPIN_ERR_PTHREAD, sspin);
    3081        1273 :         sspin_init_internal(sspin);
    3082             :         SSPIN_GET_STACK(sspin, init);
    3083        1273 :         sspin->initialized = true;
    3084             : }
    3085             : 
    3086             : void
    3087        1263 : spdk_spin_destroy(struct spdk_spinlock *sspin)
    3088             : {
    3089             :         int rc;
    3090             : 
    3091        1263 :         SPIN_ASSERT_LOG_STACKS(!sspin->destroyed, SPIN_ERR_DESTROYED, sspin);
    3092        1263 :         SPIN_ASSERT_LOG_STACKS(sspin->initialized, SPIN_ERR_NOT_INITIALIZED, sspin);
    3093        1263 :         SPIN_ASSERT_LOG_STACKS(sspin->thread == NULL, SPIN_ERR_LOCK_HELD, sspin);
    3094             : 
    3095        1262 :         rc = pthread_spin_destroy(&sspin->spinlock);
    3096        1262 :         SPIN_ASSERT_LOG_STACKS(rc == 0, SPIN_ERR_PTHREAD, sspin);
    3097             : 
    3098        1262 :         sspin_fini_internal(sspin);
    3099        1262 :         sspin->initialized = false;
    3100        1262 :         sspin->destroyed = true;
    3101             : }
    3102             : 
    3103             : void
    3104       27579 : spdk_spin_lock(struct spdk_spinlock *sspin)
    3105             : {
    3106       27579 :         struct spdk_thread *thread = spdk_get_thread();
    3107             :         int rc;
    3108             : 
    3109       27579 :         SPIN_ASSERT_LOG_STACKS(!sspin->destroyed, SPIN_ERR_DESTROYED, sspin);
    3110       27579 :         SPIN_ASSERT_LOG_STACKS(sspin->initialized, SPIN_ERR_NOT_INITIALIZED, sspin);
    3111       27579 :         SPIN_ASSERT_LOG_STACKS(thread != NULL, SPIN_ERR_NOT_SPDK_THREAD, sspin);
    3112       27578 :         SPIN_ASSERT_LOG_STACKS(thread != sspin->thread, SPIN_ERR_DEADLOCK, sspin);
    3113             : 
    3114       27577 :         rc = pthread_spin_lock(&sspin->spinlock);
    3115       27577 :         SPIN_ASSERT_LOG_STACKS(rc == 0, SPIN_ERR_PTHREAD, sspin);
    3116             : 
    3117       27577 :         sspin->thread = thread;
    3118       27577 :         sspin->thread->lock_count++;
    3119             : 
    3120             :         SSPIN_GET_STACK(sspin, lock);
    3121             : }
    3122             : 
    3123             : void
    3124       27579 : spdk_spin_unlock(struct spdk_spinlock *sspin)
    3125             : {
    3126       27579 :         struct spdk_thread *thread = spdk_get_thread();
    3127             :         int rc;
    3128             : 
    3129       27579 :         SPIN_ASSERT_LOG_STACKS(!sspin->destroyed, SPIN_ERR_DESTROYED, sspin);
    3130       27579 :         SPIN_ASSERT_LOG_STACKS(sspin->initialized, SPIN_ERR_NOT_INITIALIZED, sspin);
    3131       27579 :         SPIN_ASSERT_LOG_STACKS(thread != NULL, SPIN_ERR_NOT_SPDK_THREAD, sspin);
    3132       27579 :         SPIN_ASSERT_LOG_STACKS(thread == sspin->thread, SPIN_ERR_WRONG_THREAD, sspin);
    3133             : 
    3134       27577 :         SPIN_ASSERT_LOG_STACKS(thread->lock_count > 0, SPIN_ERR_LOCK_COUNT, sspin);
    3135       27577 :         thread->lock_count--;
    3136       27577 :         sspin->thread = NULL;
    3137             : 
    3138             :         SSPIN_GET_STACK(sspin, unlock);
    3139             : 
    3140       27577 :         rc = pthread_spin_unlock(&sspin->spinlock);
    3141       27577 :         SPIN_ASSERT_LOG_STACKS(rc == 0, SPIN_ERR_PTHREAD, sspin);
    3142             : }
    3143             : 
    3144             : bool
    3145       30698 : spdk_spin_held(struct spdk_spinlock *sspin)
    3146             : {
    3147       30698 :         struct spdk_thread *thread = spdk_get_thread();
    3148             : 
    3149       30698 :         SPIN_ASSERT_RETURN(thread != NULL, SPIN_ERR_NOT_SPDK_THREAD, false);
    3150             : 
    3151       30697 :         return sspin->thread == thread;
    3152             : }
    3153             : 
    3154          49 : SPDK_LOG_REGISTER_COMPONENT(thread)

Generated by: LCOV version 1.15