LCOV - code coverage report
Current view: top level - lib/thread - thread.c (source / functions) Hit Total Coverage
Test: ut_cov_unit.info Lines: 814 1302 62.5 %
Date: 2024-07-15 14:06:47 Functions: 104 160 65.0 %

          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             :         /* User context allocated at the end */
     156             :         uint8_t                         ctx[0];
     157             : };
     158             : 
     159             : static pthread_mutex_t g_devlist_mutex = PTHREAD_MUTEX_INITIALIZER;
     160             : 
     161             : static spdk_new_thread_fn g_new_thread_fn = NULL;
     162             : static spdk_thread_op_fn g_thread_op_fn = NULL;
     163             : static spdk_thread_op_supported_fn g_thread_op_supported_fn;
     164             : static size_t g_ctx_sz = 0;
     165             : /* Monotonic increasing ID is set to each created thread beginning at 1. Once the
     166             :  * ID exceeds UINT64_MAX, further thread creation is not allowed and restarting
     167             :  * SPDK application is required.
     168             :  */
     169             : static uint64_t g_thread_id = 1;
     170             : 
     171             : enum spin_error {
     172             :         SPIN_ERR_NONE,
     173             :         /* Trying to use an SPDK lock while not on an SPDK thread */
     174             :         SPIN_ERR_NOT_SPDK_THREAD,
     175             :         /* Trying to lock a lock already held by this SPDK thread */
     176             :         SPIN_ERR_DEADLOCK,
     177             :         /* Trying to unlock a lock not held by this SPDK thread */
     178             :         SPIN_ERR_WRONG_THREAD,
     179             :         /* pthread_spin_*() returned an error */
     180             :         SPIN_ERR_PTHREAD,
     181             :         /* Trying to destroy a lock that is held */
     182             :         SPIN_ERR_LOCK_HELD,
     183             :         /* lock_count is invalid */
     184             :         SPIN_ERR_LOCK_COUNT,
     185             :         /*
     186             :          * An spdk_thread may migrate to another pthread. A spinlock held across migration leads to
     187             :          * undefined behavior. A spinlock held when an SPDK thread goes off CPU would lead to
     188             :          * deadlock when another SPDK thread on the same pthread tries to take that lock.
     189             :          */
     190             :         SPIN_ERR_HOLD_DURING_SWITCH,
     191             :         /* Trying to use a lock that was destroyed (but not re-initialized) */
     192             :         SPIN_ERR_DESTROYED,
     193             :         /* Trying to use a lock that is not initialized */
     194             :         SPIN_ERR_NOT_INITIALIZED,
     195             : 
     196             :         /* Must be last, not an actual error code */
     197             :         SPIN_ERR_LAST
     198             : };
     199             : 
     200             : static const char *spin_error_strings[] = {
     201             :         [SPIN_ERR_NONE]                 = "No error",
     202             :         [SPIN_ERR_NOT_SPDK_THREAD]      = "Not an SPDK thread",
     203             :         [SPIN_ERR_DEADLOCK]             = "Deadlock detected",
     204             :         [SPIN_ERR_WRONG_THREAD]         = "Unlock on wrong SPDK thread",
     205             :         [SPIN_ERR_PTHREAD]              = "Error from pthread_spinlock",
     206             :         [SPIN_ERR_LOCK_HELD]            = "Destroying a held spinlock",
     207             :         [SPIN_ERR_LOCK_COUNT]           = "Lock count is invalid",
     208             :         [SPIN_ERR_HOLD_DURING_SWITCH]   = "Lock(s) held while SPDK thread going off CPU",
     209             :         [SPIN_ERR_DESTROYED]            = "Lock has been destroyed",
     210             :         [SPIN_ERR_NOT_INITIALIZED]      = "Lock has not been initialized",
     211             : };
     212             : 
     213             : #define SPIN_ERROR_STRING(err) (err < 0 || err >= SPDK_COUNTOF(spin_error_strings)) \
     214             :                                 ? "Unknown error" : spin_error_strings[err]
     215             : 
     216             : static void
     217           0 : __posix_abort(enum spin_error err)
     218             : {
     219           0 :         abort();
     220             : }
     221             : 
     222             : typedef void (*spin_abort)(enum spin_error err);
     223             : spin_abort g_spin_abort_fn = __posix_abort;
     224             : 
     225             : #define SPIN_ASSERT_IMPL(cond, err, extra_log, ret) \
     226             :         do { \
     227             :                 if (spdk_unlikely(!(cond))) { \
     228             :                         SPDK_ERRLOG("unrecoverable spinlock error %d: %s (%s)\n", err, \
     229             :                                     SPIN_ERROR_STRING(err), #cond); \
     230             :                         extra_log; \
     231             :                         g_spin_abort_fn(err); \
     232             :                         ret; \
     233             :                 } \
     234             :         } while (0)
     235             : #define SPIN_ASSERT_LOG_STACKS(cond, err, lock) \
     236             :         SPIN_ASSERT_IMPL(cond, err, sspin_stacks_print(sspin), return)
     237             : #define SPIN_ASSERT_RETURN(cond, err, ret)      SPIN_ASSERT_IMPL(cond, err, , return ret)
     238             : #define SPIN_ASSERT(cond, err)                  SPIN_ASSERT_IMPL(cond, err, ,)
     239             : 
     240             : struct io_device {
     241             :         void                            *io_device;
     242             :         char                            name[SPDK_MAX_DEVICE_NAME_LEN + 1];
     243             :         spdk_io_channel_create_cb       create_cb;
     244             :         spdk_io_channel_destroy_cb      destroy_cb;
     245             :         spdk_io_device_unregister_cb    unregister_cb;
     246             :         struct spdk_thread              *unregister_thread;
     247             :         uint32_t                        ctx_size;
     248             :         uint32_t                        for_each_count;
     249             :         RB_ENTRY(io_device)             node;
     250             : 
     251             :         uint32_t                        refcnt;
     252             : 
     253             :         bool                            pending_unregister;
     254             :         bool                            unregistered;
     255             : };
     256             : 
     257             : static RB_HEAD(io_device_tree, io_device) g_io_devices = RB_INITIALIZER(g_io_devices);
     258             : 
     259             : static int
     260        9210 : io_device_cmp(struct io_device *dev1, struct io_device *dev2)
     261             : {
     262        9210 :         return (dev1->io_device < dev2->io_device ? -1 : dev1->io_device > dev2->io_device);
     263             : }
     264             : 
     265       15133 : RB_GENERATE_STATIC(io_device_tree, io_device, node, io_device_cmp);
     266             : 
     267             : static int
     268        7113 : io_channel_cmp(struct spdk_io_channel *ch1, struct spdk_io_channel *ch2)
     269             : {
     270        7113 :         return (ch1->dev < ch2->dev ? -1 : ch1->dev > ch2->dev);
     271             : }
     272             : 
     273       16303 : RB_GENERATE_STATIC(io_channel_tree, spdk_io_channel, node, io_channel_cmp);
     274             : 
     275             : struct spdk_msg {
     276             :         spdk_msg_fn             fn;
     277             :         void                    *arg;
     278             : 
     279             :         SLIST_ENTRY(spdk_msg)   link;
     280             : };
     281             : 
     282             : static struct spdk_mempool *g_spdk_msg_mempool = NULL;
     283             : 
     284             : static TAILQ_HEAD(, spdk_thread) g_threads = TAILQ_HEAD_INITIALIZER(g_threads);
     285             : static uint32_t g_thread_count = 0;
     286             : 
     287             : static __thread struct spdk_thread *tls_thread = NULL;
     288             : 
     289          38 : SPDK_TRACE_REGISTER_FN(thread_trace, "thread", TRACE_GROUP_THREAD)
     290             : {
     291           0 :         struct spdk_trace_tpoint_opts opts[] = {
     292             :                 {
     293             :                         "THREAD_IOCH_GET", TRACE_THREAD_IOCH_GET,
     294             :                         OWNER_TYPE_NONE, OBJECT_NONE, 0,
     295             :                         {{ "refcnt", SPDK_TRACE_ARG_TYPE_INT, 4 }}
     296             :                 },
     297             :                 {
     298             :                         "THREAD_IOCH_PUT", TRACE_THREAD_IOCH_PUT,
     299             :                         OWNER_TYPE_NONE, OBJECT_NONE, 0,
     300             :                         {{ "refcnt", SPDK_TRACE_ARG_TYPE_INT, 4 }}
     301             :                 }
     302             :         };
     303             : 
     304           0 :         spdk_trace_register_description_ext(opts, SPDK_COUNTOF(opts));
     305           0 : }
     306             : 
     307             : /*
     308             :  * If this compare function returns zero when two next_run_ticks are equal,
     309             :  * the macro RB_INSERT() returns a pointer to the element with the same
     310             :  * next_run_tick.
     311             :  *
     312             :  * Fortunately, the macro RB_REMOVE() takes not a key but a pointer to the element
     313             :  * to remove as a parameter.
     314             :  *
     315             :  * Hence we allow RB_INSERT() to insert elements with the same keys on the right
     316             :  * side by returning 1 when two next_run_ticks are equal.
     317             :  */
     318             : static inline int
     319         580 : timed_poller_compare(struct spdk_poller *poller1, struct spdk_poller *poller2)
     320             : {
     321         580 :         if (poller1->next_run_tick < poller2->next_run_tick) {
     322         261 :                 return -1;
     323             :         } else {
     324         319 :                 return 1;
     325             :         }
     326             : }
     327             : 
     328        3581 : RB_GENERATE_STATIC(timed_pollers_tree, spdk_poller, node, timed_poller_compare);
     329             : 
     330             : static inline struct spdk_thread *
     331      599697 : _get_thread(void)
     332             : {
     333      599697 :         return tls_thread;
     334             : }
     335             : 
     336             : static int
     337          80 : _thread_lib_init(size_t ctx_sz, size_t msg_mempool_sz)
     338             : {
     339          80 :         char mempool_name[SPDK_MAX_MEMZONE_NAME_LEN];
     340             : 
     341          80 :         g_ctx_sz = ctx_sz;
     342             : 
     343          80 :         snprintf(mempool_name, sizeof(mempool_name), "msgpool_%d", getpid());
     344          80 :         g_spdk_msg_mempool = spdk_mempool_create(mempool_name, msg_mempool_sz,
     345             :                              sizeof(struct spdk_msg),
     346             :                              0, /* No cache. We do our own. */
     347             :                              SPDK_ENV_SOCKET_ID_ANY);
     348             : 
     349          80 :         SPDK_DEBUGLOG(thread, "spdk_msg_mempool was created with size: %zu\n",
     350             :                       msg_mempool_sz);
     351             : 
     352          80 :         if (!g_spdk_msg_mempool) {
     353           0 :                 SPDK_ERRLOG("spdk_msg_mempool creation failed\n");
     354           0 :                 return -ENOMEM;
     355             :         }
     356             : 
     357          80 :         return 0;
     358             : }
     359             : 
     360             : static void thread_interrupt_destroy(struct spdk_thread *thread);
     361             : static int thread_interrupt_create(struct spdk_thread *thread);
     362             : 
     363             : static void
     364         161 : _free_thread(struct spdk_thread *thread)
     365             : {
     366             :         struct spdk_io_channel *ch;
     367             :         struct spdk_msg *msg;
     368             :         struct spdk_poller *poller, *ptmp;
     369             : 
     370         161 :         RB_FOREACH(ch, io_channel_tree, &thread->io_channels) {
     371           0 :                 SPDK_ERRLOG("thread %s still has channel for io_device %s\n",
     372             :                             thread->name, ch->dev->name);
     373             :         }
     374             : 
     375         161 :         TAILQ_FOREACH_SAFE(poller, &thread->active_pollers, tailq, ptmp) {
     376           0 :                 if (poller->state != SPDK_POLLER_STATE_UNREGISTERED) {
     377           0 :                         SPDK_WARNLOG("active_poller %s still registered at thread exit\n",
     378             :                                      poller->name);
     379             :                 }
     380           0 :                 TAILQ_REMOVE(&thread->active_pollers, poller, tailq);
     381           0 :                 free(poller);
     382             :         }
     383             : 
     384         183 :         RB_FOREACH_SAFE(poller, timed_pollers_tree, &thread->timed_pollers, ptmp) {
     385          22 :                 if (poller->state != SPDK_POLLER_STATE_UNREGISTERED) {
     386           0 :                         SPDK_WARNLOG("timed_poller %s still registered at thread exit\n",
     387             :                                      poller->name);
     388             :                 }
     389          22 :                 RB_REMOVE(timed_pollers_tree, &thread->timed_pollers, poller);
     390          22 :                 free(poller);
     391             :         }
     392             : 
     393         161 :         TAILQ_FOREACH_SAFE(poller, &thread->paused_pollers, tailq, ptmp) {
     394           0 :                 SPDK_WARNLOG("paused_poller %s still registered at thread exit\n", poller->name);
     395           0 :                 TAILQ_REMOVE(&thread->paused_pollers, poller, tailq);
     396           0 :                 free(poller);
     397             :         }
     398             : 
     399         161 :         pthread_mutex_lock(&g_devlist_mutex);
     400         161 :         assert(g_thread_count > 0);
     401         161 :         g_thread_count--;
     402         161 :         TAILQ_REMOVE(&g_threads, thread, tailq);
     403         161 :         pthread_mutex_unlock(&g_devlist_mutex);
     404             : 
     405         161 :         msg = SLIST_FIRST(&thread->msg_cache);
     406      164872 :         while (msg != NULL) {
     407      164711 :                 SLIST_REMOVE_HEAD(&thread->msg_cache, link);
     408             : 
     409      164711 :                 assert(thread->msg_cache_count > 0);
     410      164711 :                 thread->msg_cache_count--;
     411      164711 :                 spdk_mempool_put(g_spdk_msg_mempool, msg);
     412             : 
     413      164711 :                 msg = SLIST_FIRST(&thread->msg_cache);
     414             :         }
     415             : 
     416         161 :         assert(thread->msg_cache_count == 0);
     417             : 
     418         161 :         if (spdk_interrupt_mode_is_enabled()) {
     419           0 :                 thread_interrupt_destroy(thread);
     420             :         }
     421             : 
     422         161 :         spdk_ring_free(thread->messages);
     423         161 :         free(thread);
     424         161 : }
     425             : 
     426             : int
     427          70 : spdk_thread_lib_init(spdk_new_thread_fn new_thread_fn, size_t ctx_sz)
     428             : {
     429          70 :         assert(g_new_thread_fn == NULL);
     430          70 :         assert(g_thread_op_fn == NULL);
     431             : 
     432          70 :         if (new_thread_fn == NULL) {
     433          68 :                 SPDK_INFOLOG(thread, "new_thread_fn was not specified at spdk_thread_lib_init\n");
     434             :         } else {
     435           2 :                 g_new_thread_fn = new_thread_fn;
     436             :         }
     437             : 
     438          70 :         return _thread_lib_init(ctx_sz, SPDK_DEFAULT_MSG_MEMPOOL_SIZE);
     439             : }
     440             : 
     441             : int
     442          10 : spdk_thread_lib_init_ext(spdk_thread_op_fn thread_op_fn,
     443             :                          spdk_thread_op_supported_fn thread_op_supported_fn,
     444             :                          size_t ctx_sz, size_t msg_mempool_sz)
     445             : {
     446          10 :         assert(g_new_thread_fn == NULL);
     447          10 :         assert(g_thread_op_fn == NULL);
     448          10 :         assert(g_thread_op_supported_fn == NULL);
     449             : 
     450          10 :         if ((thread_op_fn != NULL) != (thread_op_supported_fn != NULL)) {
     451           0 :                 SPDK_ERRLOG("Both must be defined or undefined together.\n");
     452           0 :                 return -EINVAL;
     453             :         }
     454             : 
     455          10 :         if (thread_op_fn == NULL && thread_op_supported_fn == NULL) {
     456           0 :                 SPDK_INFOLOG(thread, "thread_op_fn and thread_op_supported_fn were not specified\n");
     457             :         } else {
     458          10 :                 g_thread_op_fn = thread_op_fn;
     459          10 :                 g_thread_op_supported_fn = thread_op_supported_fn;
     460             :         }
     461             : 
     462          10 :         return _thread_lib_init(ctx_sz, msg_mempool_sz);
     463             : }
     464             : 
     465             : void
     466          79 : spdk_thread_lib_fini(void)
     467             : {
     468             :         struct io_device *dev;
     469             : 
     470          80 :         RB_FOREACH(dev, io_device_tree, &g_io_devices) {
     471           1 :                 SPDK_ERRLOG("io_device %s not unregistered\n", dev->name);
     472             :         }
     473             : 
     474          79 :         g_new_thread_fn = NULL;
     475          79 :         g_thread_op_fn = NULL;
     476          79 :         g_thread_op_supported_fn = NULL;
     477          79 :         g_ctx_sz = 0;
     478          79 :         if (g_app_thread != NULL) {
     479          76 :                 _free_thread(g_app_thread);
     480          76 :                 g_app_thread = NULL;
     481             :         }
     482             : 
     483          79 :         if (g_spdk_msg_mempool) {
     484          79 :                 spdk_mempool_free(g_spdk_msg_mempool);
     485          79 :                 g_spdk_msg_mempool = NULL;
     486             :         }
     487          79 : }
     488             : 
     489             : struct spdk_thread *
     490         196 : spdk_thread_create(const char *name, const struct spdk_cpuset *cpumask)
     491             : {
     492         196 :         struct spdk_thread *thread, *null_thread;
     493         196 :         struct spdk_msg *msgs[SPDK_MSG_MEMPOOL_CACHE_SIZE];
     494         196 :         int rc = 0, i;
     495             : 
     496         196 :         thread = calloc(1, sizeof(*thread) + g_ctx_sz);
     497         196 :         if (!thread) {
     498           0 :                 SPDK_ERRLOG("Unable to allocate memory for thread\n");
     499           0 :                 return NULL;
     500             :         }
     501             : 
     502         196 :         if (cpumask) {
     503          22 :                 spdk_cpuset_copy(&thread->cpumask, cpumask);
     504             :         } else {
     505         174 :                 spdk_cpuset_negate(&thread->cpumask);
     506             :         }
     507             : 
     508         196 :         RB_INIT(&thread->io_channels);
     509         196 :         TAILQ_INIT(&thread->active_pollers);
     510         196 :         RB_INIT(&thread->timed_pollers);
     511         196 :         TAILQ_INIT(&thread->paused_pollers);
     512         196 :         SLIST_INIT(&thread->msg_cache);
     513         196 :         thread->msg_cache_count = 0;
     514             : 
     515         196 :         thread->tsc_last = spdk_get_ticks();
     516             : 
     517             :         /* Monotonic increasing ID is set to each created poller beginning at 1. Once the
     518             :          * ID exceeds UINT64_MAX a warning message is logged
     519             :          */
     520         196 :         thread->next_poller_id = 1;
     521             : 
     522         196 :         thread->messages = spdk_ring_create(SPDK_RING_TYPE_MP_SC, 65536, SPDK_ENV_SOCKET_ID_ANY);
     523         196 :         if (!thread->messages) {
     524           0 :                 SPDK_ERRLOG("Unable to allocate memory for message ring\n");
     525           0 :                 free(thread);
     526           0 :                 return NULL;
     527             :         }
     528             : 
     529             :         /* Fill the local message pool cache. */
     530         196 :         rc = spdk_mempool_get_bulk(g_spdk_msg_mempool, (void **)msgs, SPDK_MSG_MEMPOOL_CACHE_SIZE);
     531         196 :         if (rc == 0) {
     532             :                 /* If we can't populate the cache it's ok. The cache will get filled
     533             :                  * up organically as messages are passed to the thread. */
     534      200900 :                 for (i = 0; i < SPDK_MSG_MEMPOOL_CACHE_SIZE; i++) {
     535      200704 :                         SLIST_INSERT_HEAD(&thread->msg_cache, msgs[i], link);
     536      200704 :                         thread->msg_cache_count++;
     537             :                 }
     538             :         }
     539             : 
     540         196 :         if (name) {
     541          41 :                 snprintf(thread->name, sizeof(thread->name), "%s", name);
     542             :         } else {
     543         155 :                 snprintf(thread->name, sizeof(thread->name), "%p", thread);
     544             :         }
     545             : 
     546         196 :         pthread_mutex_lock(&g_devlist_mutex);
     547         196 :         if (g_thread_id == 0) {
     548           0 :                 SPDK_ERRLOG("Thread ID rolled over. Further thread creation is not allowed.\n");
     549           0 :                 pthread_mutex_unlock(&g_devlist_mutex);
     550           0 :                 _free_thread(thread);
     551           0 :                 return NULL;
     552             :         }
     553         196 :         thread->id = g_thread_id++;
     554         196 :         TAILQ_INSERT_TAIL(&g_threads, thread, tailq);
     555         196 :         g_thread_count++;
     556         196 :         pthread_mutex_unlock(&g_devlist_mutex);
     557             : 
     558         196 :         SPDK_DEBUGLOG(thread, "Allocating new thread (%" PRIu64 ", %s)\n",
     559             :                       thread->id, thread->name);
     560             : 
     561         196 :         if (spdk_interrupt_mode_is_enabled()) {
     562           0 :                 thread->in_interrupt = true;
     563           0 :                 rc = thread_interrupt_create(thread);
     564           0 :                 if (rc != 0) {
     565           0 :                         _free_thread(thread);
     566           0 :                         return NULL;
     567             :                 }
     568             :         }
     569             : 
     570         196 :         if (g_new_thread_fn) {
     571           4 :                 rc = g_new_thread_fn(thread);
     572         192 :         } else if (g_thread_op_supported_fn && g_thread_op_supported_fn(SPDK_THREAD_OP_NEW)) {
     573          15 :                 rc = g_thread_op_fn(thread, SPDK_THREAD_OP_NEW);
     574             :         }
     575             : 
     576         196 :         if (rc != 0) {
     577           2 :                 _free_thread(thread);
     578           2 :                 return NULL;
     579             :         }
     580             : 
     581         194 :         thread->state = SPDK_THREAD_STATE_RUNNING;
     582             : 
     583             :         /* If this is the first thread, save it as the app thread.  Use an atomic
     584             :          * compare + exchange to guard against crazy users who might try to
     585             :          * call spdk_thread_create() simultaneously on multiple threads.
     586             :          */
     587         194 :         null_thread = NULL;
     588         194 :         __atomic_compare_exchange_n(&g_app_thread, &null_thread, thread, false,
     589             :                                     __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST);
     590             : 
     591         194 :         return thread;
     592             : }
     593             : 
     594             : struct spdk_thread *
     595        1526 : spdk_thread_get_app_thread(void)
     596             : {
     597        1526 :         return g_app_thread;
     598             : }
     599             : 
     600             : bool
     601         147 : spdk_thread_is_app_thread(struct spdk_thread *thread)
     602             : {
     603         147 :         if (thread == NULL) {
     604         146 :                 thread = _get_thread();
     605             :         }
     606             : 
     607         147 :         return g_app_thread == thread;
     608             : }
     609             : 
     610             : void
     611           1 : spdk_thread_bind(struct spdk_thread *thread, bool bind)
     612             : {
     613           1 :         thread->is_bound = bind;
     614           1 : }
     615             : 
     616             : bool
     617          10 : spdk_thread_is_bound(struct spdk_thread *thread)
     618             : {
     619          10 :         return thread->is_bound;
     620             : }
     621             : 
     622             : void
     623      310879 : spdk_set_thread(struct spdk_thread *thread)
     624             : {
     625      310879 :         tls_thread = thread;
     626      310879 : }
     627             : 
     628             : static void
     629         197 : thread_exit(struct spdk_thread *thread, uint64_t now)
     630             : {
     631             :         struct spdk_poller *poller;
     632             :         struct spdk_io_channel *ch;
     633             : 
     634         197 :         if (now >= thread->exit_timeout_tsc) {
     635           1 :                 SPDK_ERRLOG("thread %s got timeout, and move it to the exited state forcefully\n",
     636             :                             thread->name);
     637           1 :                 goto exited;
     638             :         }
     639             : 
     640         196 :         if (spdk_ring_count(thread->messages) > 0) {
     641           2 :                 SPDK_INFOLOG(thread, "thread %s still has messages\n", thread->name);
     642           2 :                 return;
     643             :         }
     644             : 
     645         194 :         if (thread->for_each_count > 0) {
     646           4 :                 SPDK_INFOLOG(thread, "thread %s is still executing %u for_each_channels/threads\n",
     647             :                              thread->name, thread->for_each_count);
     648           4 :                 return;
     649             :         }
     650             : 
     651         190 :         TAILQ_FOREACH(poller, &thread->active_pollers, tailq) {
     652           2 :                 if (poller->state != SPDK_POLLER_STATE_UNREGISTERED) {
     653           2 :                         SPDK_INFOLOG(thread,
     654             :                                      "thread %s still has active poller %s\n",
     655             :                                      thread->name, poller->name);
     656           2 :                         return;
     657             :                 }
     658             :         }
     659             : 
     660         230 :         RB_FOREACH(poller, timed_pollers_tree, &thread->timed_pollers) {
     661          42 :                 if (poller->state != SPDK_POLLER_STATE_UNREGISTERED) {
     662           0 :                         SPDK_INFOLOG(thread,
     663             :                                      "thread %s still has active timed poller %s\n",
     664             :                                      thread->name, poller->name);
     665           0 :                         return;
     666             :                 }
     667             :         }
     668             : 
     669         188 :         TAILQ_FOREACH(poller, &thread->paused_pollers, tailq) {
     670           0 :                 SPDK_INFOLOG(thread,
     671             :                              "thread %s still has paused poller %s\n",
     672             :                              thread->name, poller->name);
     673           0 :                 return;
     674             :         }
     675             : 
     676         188 :         RB_FOREACH(ch, io_channel_tree, &thread->io_channels) {
     677           4 :                 SPDK_INFOLOG(thread,
     678             :                              "thread %s still has channel for io_device %s\n",
     679             :                              thread->name, ch->dev->name);
     680           4 :                 return;
     681             :         }
     682             : 
     683         184 :         if (thread->pending_unregister_count > 0) {
     684           2 :                 SPDK_INFOLOG(thread,
     685             :                              "thread %s is still unregistering io_devices\n",
     686             :                              thread->name);
     687           2 :                 return;
     688             :         }
     689             : 
     690         182 : exited:
     691         183 :         thread->state = SPDK_THREAD_STATE_EXITED;
     692         183 :         if (spdk_unlikely(thread->in_interrupt)) {
     693           0 :                 g_thread_op_fn(thread, SPDK_THREAD_OP_RESCHED);
     694             :         }
     695             : }
     696             : 
     697             : static void _thread_exit(void *ctx);
     698             : 
     699             : int
     700         193 : spdk_thread_exit(struct spdk_thread *thread)
     701             : {
     702         193 :         SPDK_DEBUGLOG(thread, "Exit thread %s\n", thread->name);
     703             : 
     704         193 :         assert(tls_thread == thread);
     705             : 
     706         193 :         if (thread->state >= SPDK_THREAD_STATE_EXITING) {
     707           9 :                 SPDK_INFOLOG(thread,
     708             :                              "thread %s is already exiting\n",
     709             :                              thread->name);
     710           9 :                 return 0;
     711             :         }
     712             : 
     713         184 :         thread->exit_timeout_tsc = spdk_get_ticks() + (spdk_get_ticks_hz() *
     714             :                                    SPDK_THREAD_EXIT_TIMEOUT_SEC);
     715         184 :         thread->state = SPDK_THREAD_STATE_EXITING;
     716             : 
     717         184 :         if (spdk_interrupt_mode_is_enabled()) {
     718           0 :                 spdk_thread_send_msg(thread, _thread_exit, thread);
     719             :         }
     720             : 
     721         184 :         return 0;
     722             : }
     723             : 
     724             : bool
     725           0 : spdk_thread_is_running(struct spdk_thread *thread)
     726             : {
     727           0 :         return thread->state == SPDK_THREAD_STATE_RUNNING;
     728             : }
     729             : 
     730             : bool
     731         367 : spdk_thread_is_exited(struct spdk_thread *thread)
     732             : {
     733         367 :         return thread->state == SPDK_THREAD_STATE_EXITED;
     734             : }
     735             : 
     736             : void
     737         163 : spdk_thread_destroy(struct spdk_thread *thread)
     738             : {
     739         163 :         assert(thread != NULL);
     740         163 :         SPDK_DEBUGLOG(thread, "Destroy thread %s\n", thread->name);
     741             : 
     742         163 :         assert(thread->state == SPDK_THREAD_STATE_EXITED);
     743             : 
     744         163 :         if (tls_thread == thread) {
     745         157 :                 tls_thread = NULL;
     746             :         }
     747             : 
     748             :         /* To be safe, do not free the app thread until spdk_thread_lib_fini(). */
     749         163 :         if (thread != g_app_thread) {
     750          83 :                 _free_thread(thread);
     751             :         }
     752         163 : }
     753             : 
     754             : void *
     755          47 : spdk_thread_get_ctx(struct spdk_thread *thread)
     756             : {
     757          47 :         if (g_ctx_sz > 0) {
     758          47 :                 return thread->ctx;
     759             :         }
     760             : 
     761           0 :         return NULL;
     762             : }
     763             : 
     764             : struct spdk_cpuset *
     765          44 : spdk_thread_get_cpumask(struct spdk_thread *thread)
     766             : {
     767          44 :         return &thread->cpumask;
     768             : }
     769             : 
     770             : int
     771           2 : spdk_thread_set_cpumask(struct spdk_cpuset *cpumask)
     772             : {
     773             :         struct spdk_thread *thread;
     774             : 
     775           2 :         if (!g_thread_op_supported_fn || !g_thread_op_supported_fn(SPDK_THREAD_OP_RESCHED)) {
     776           0 :                 SPDK_ERRLOG("Framework does not support reschedule operation.\n");
     777           0 :                 assert(false);
     778             :                 return -ENOTSUP;
     779             :         }
     780             : 
     781           2 :         thread = spdk_get_thread();
     782           2 :         if (!thread) {
     783           0 :                 SPDK_ERRLOG("Called from non-SPDK thread\n");
     784           0 :                 assert(false);
     785             :                 return -EINVAL;
     786             :         }
     787             : 
     788           2 :         spdk_cpuset_copy(&thread->cpumask, cpumask);
     789             : 
     790             :         /* Invoke framework's reschedule operation. If this function is called multiple times
     791             :          * in a single spdk_thread_poll() context, the last cpumask will be used in the
     792             :          * reschedule operation.
     793             :          */
     794           2 :         g_thread_op_fn(thread, SPDK_THREAD_OP_RESCHED);
     795             : 
     796           2 :         return 0;
     797             : }
     798             : 
     799             : struct spdk_thread *
     800         202 : spdk_thread_get_from_ctx(void *ctx)
     801             : {
     802         202 :         if (ctx == NULL) {
     803           0 :                 assert(false);
     804             :                 return NULL;
     805             :         }
     806             : 
     807         202 :         assert(g_ctx_sz > 0);
     808             : 
     809         202 :         return SPDK_CONTAINEROF(ctx, struct spdk_thread, ctx);
     810             : }
     811             : 
     812             : static inline uint32_t
     813      298948 : msg_queue_run_batch(struct spdk_thread *thread, uint32_t max_msgs)
     814             : {
     815             :         unsigned count, i;
     816      298948 :         void *messages[SPDK_MSG_BATCH_SIZE];
     817      298948 :         uint64_t notify = 1;
     818             :         int rc;
     819             : 
     820             : #ifdef DEBUG
     821             :         /*
     822             :          * spdk_ring_dequeue() fills messages and returns how many entries it wrote,
     823             :          * so we will never actually read uninitialized data from events, but just to be sure
     824             :          * (and to silence a static analyzer false positive), initialize the array to NULL pointers.
     825             :          */
     826      298948 :         memset(messages, 0, sizeof(messages));
     827             : #endif
     828             : 
     829      298948 :         if (max_msgs > 0) {
     830         281 :                 max_msgs = spdk_min(max_msgs, SPDK_MSG_BATCH_SIZE);
     831             :         } else {
     832      298667 :                 max_msgs = SPDK_MSG_BATCH_SIZE;
     833             :         }
     834             : 
     835      298948 :         count = spdk_ring_dequeue(thread->messages, messages, max_msgs);
     836      298948 :         if (spdk_unlikely(thread->in_interrupt) &&
     837           0 :             spdk_ring_count(thread->messages) != 0) {
     838           0 :                 rc = write(thread->msg_fd, &notify, sizeof(notify));
     839           0 :                 if (rc < 0) {
     840           0 :                         SPDK_ERRLOG("failed to notify msg_queue: %s.\n", spdk_strerror(errno));
     841             :                 }
     842             :         }
     843      298948 :         if (count == 0) {
     844      223863 :                 return 0;
     845             :         }
     846             : 
     847      151592 :         for (i = 0; i < count; i++) {
     848       76507 :                 struct spdk_msg *msg = messages[i];
     849             : 
     850       76507 :                 assert(msg != NULL);
     851             : 
     852             :                 SPDK_DTRACE_PROBE2(msg_exec, msg->fn, msg->arg);
     853             : 
     854       76507 :                 msg->fn(msg->arg);
     855             : 
     856       76507 :                 SPIN_ASSERT(thread->lock_count == 0, SPIN_ERR_HOLD_DURING_SWITCH);
     857             : 
     858       76507 :                 if (thread->msg_cache_count < SPDK_MSG_MEMPOOL_CACHE_SIZE) {
     859             :                         /* Insert the messages at the head. We want to re-use the hot
     860             :                          * ones. */
     861       76361 :                         SLIST_INSERT_HEAD(&thread->msg_cache, msg, link);
     862       76361 :                         thread->msg_cache_count++;
     863             :                 } else {
     864         146 :                         spdk_mempool_put(g_spdk_msg_mempool, msg);
     865             :                 }
     866             :         }
     867             : 
     868       75085 :         return count;
     869             : }
     870             : 
     871             : static void
     872         468 : poller_insert_timer(struct spdk_thread *thread, struct spdk_poller *poller, uint64_t now)
     873             : {
     874             :         struct spdk_poller *tmp __attribute__((unused));
     875             : 
     876         468 :         poller->next_run_tick = now + poller->period_ticks;
     877             : 
     878             :         /*
     879             :          * Insert poller in the thread's timed_pollers tree by next scheduled run time
     880             :          * as its key.
     881             :          */
     882         468 :         tmp = RB_INSERT(timed_pollers_tree, &thread->timed_pollers, poller);
     883         468 :         assert(tmp == NULL);
     884             : 
     885             :         /* Update the cache only if it is empty or the inserted poller is earlier than it.
     886             :          * RB_MIN() is not necessary here because all pollers, which has exactly the same
     887             :          * next_run_tick as the existing poller, are inserted on the right side.
     888             :          */
     889         468 :         if (thread->first_timed_poller == NULL ||
     890         351 :             poller->next_run_tick < thread->first_timed_poller->next_run_tick) {
     891         244 :                 thread->first_timed_poller = poller;
     892             :         }
     893         468 : }
     894             : 
     895             : static inline void
     896           0 : poller_remove_timer(struct spdk_thread *thread, struct spdk_poller *poller)
     897             : {
     898             :         struct spdk_poller *tmp __attribute__((unused));
     899             : 
     900           0 :         tmp = RB_REMOVE(timed_pollers_tree, &thread->timed_pollers, poller);
     901           0 :         assert(tmp != NULL);
     902             : 
     903             :         /* This function is not used in any case that is performance critical.
     904             :          * Update the cache simply by RB_MIN() if it needs to be changed.
     905             :          */
     906           0 :         if (thread->first_timed_poller == poller) {
     907           0 :                 thread->first_timed_poller = RB_MIN(timed_pollers_tree, &thread->timed_pollers);
     908             :         }
     909           0 : }
     910             : 
     911             : static void
     912         772 : thread_insert_poller(struct spdk_thread *thread, struct spdk_poller *poller)
     913             : {
     914         772 :         if (poller->period_ticks) {
     915         334 :                 poller_insert_timer(thread, poller, spdk_get_ticks());
     916             :         } else {
     917         438 :                 TAILQ_INSERT_TAIL(&thread->active_pollers, poller, tailq);
     918             :         }
     919         772 : }
     920             : 
     921             : static inline void
     922      298948 : thread_update_stats(struct spdk_thread *thread, uint64_t end,
     923             :                     uint64_t start, int rc)
     924             : {
     925      298948 :         if (rc == 0) {
     926             :                 /* Poller status idle */
     927      223438 :                 thread->stats.idle_tsc += end - start;
     928       75510 :         } else if (rc > 0) {
     929             :                 /* Poller status busy */
     930       75510 :                 thread->stats.busy_tsc += end - start;
     931             :         }
     932             :         /* Store end time to use it as start time of the next spdk_thread_poll(). */
     933      298948 :         thread->tsc_last = end;
     934      298948 : }
     935             : 
     936             : static inline int
     937        1635 : thread_execute_poller(struct spdk_thread *thread, struct spdk_poller *poller)
     938             : {
     939             :         int rc;
     940             : 
     941        1635 :         switch (poller->state) {
     942          95 :         case SPDK_POLLER_STATE_UNREGISTERED:
     943          95 :                 TAILQ_REMOVE(&thread->active_pollers, poller, tailq);
     944          95 :                 free(poller);
     945          95 :                 return 0;
     946           4 :         case SPDK_POLLER_STATE_PAUSING:
     947           4 :                 TAILQ_REMOVE(&thread->active_pollers, poller, tailq);
     948           4 :                 TAILQ_INSERT_TAIL(&thread->paused_pollers, poller, tailq);
     949           4 :                 poller->state = SPDK_POLLER_STATE_PAUSED;
     950           4 :                 return 0;
     951        1536 :         case SPDK_POLLER_STATE_WAITING:
     952        1536 :                 break;
     953           0 :         default:
     954           0 :                 assert(false);
     955             :                 break;
     956             :         }
     957             : 
     958        1536 :         poller->state = SPDK_POLLER_STATE_RUNNING;
     959        1536 :         rc = poller->fn(poller->arg);
     960             : 
     961        1536 :         SPIN_ASSERT(thread->lock_count == 0, SPIN_ERR_HOLD_DURING_SWITCH);
     962             : 
     963        1536 :         poller->run_count++;
     964        1536 :         if (rc > 0) {
     965         434 :                 poller->busy_count++;
     966             :         }
     967             : 
     968             : #ifdef DEBUG
     969        1536 :         if (rc == -1) {
     970          19 :                 SPDK_DEBUGLOG(thread, "Poller %s returned -1\n", poller->name);
     971             :         }
     972             : #endif
     973             : 
     974        1536 :         switch (poller->state) {
     975         341 :         case SPDK_POLLER_STATE_UNREGISTERED:
     976         341 :                 TAILQ_REMOVE(&thread->active_pollers, poller, tailq);
     977         341 :                 free(poller);
     978         341 :                 break;
     979           6 :         case SPDK_POLLER_STATE_PAUSING:
     980           6 :                 TAILQ_REMOVE(&thread->active_pollers, poller, tailq);
     981           6 :                 TAILQ_INSERT_TAIL(&thread->paused_pollers, poller, tailq);
     982           6 :                 poller->state = SPDK_POLLER_STATE_PAUSED;
     983           6 :                 break;
     984           0 :         case SPDK_POLLER_STATE_PAUSED:
     985             :         case SPDK_POLLER_STATE_WAITING:
     986           0 :                 break;
     987        1189 :         case SPDK_POLLER_STATE_RUNNING:
     988        1189 :                 poller->state = SPDK_POLLER_STATE_WAITING;
     989        1189 :                 break;
     990           0 :         default:
     991           0 :                 assert(false);
     992             :                 break;
     993             :         }
     994             : 
     995        1536 :         return rc;
     996             : }
     997             : 
     998             : static inline int
     999         426 : thread_execute_timed_poller(struct spdk_thread *thread, struct spdk_poller *poller,
    1000             :                             uint64_t now)
    1001             : {
    1002             :         int rc;
    1003             : 
    1004         426 :         switch (poller->state) {
    1005         133 :         case SPDK_POLLER_STATE_UNREGISTERED:
    1006         133 :                 free(poller);
    1007         133 :                 return 0;
    1008          13 :         case SPDK_POLLER_STATE_PAUSING:
    1009          13 :                 TAILQ_INSERT_TAIL(&thread->paused_pollers, poller, tailq);
    1010          13 :                 poller->state = SPDK_POLLER_STATE_PAUSED;
    1011          13 :                 return 0;
    1012         280 :         case SPDK_POLLER_STATE_WAITING:
    1013         280 :                 break;
    1014           0 :         default:
    1015           0 :                 assert(false);
    1016             :                 break;
    1017             :         }
    1018             : 
    1019         280 :         poller->state = SPDK_POLLER_STATE_RUNNING;
    1020         280 :         rc = poller->fn(poller->arg);
    1021             : 
    1022         280 :         SPIN_ASSERT(thread->lock_count == 0, SPIN_ERR_HOLD_DURING_SWITCH);
    1023             : 
    1024         280 :         poller->run_count++;
    1025         280 :         if (rc > 0) {
    1026         200 :                 poller->busy_count++;
    1027             :         }
    1028             : 
    1029             : #ifdef DEBUG
    1030         280 :         if (rc == -1) {
    1031           5 :                 SPDK_DEBUGLOG(thread, "Timed poller %s returned -1\n", poller->name);
    1032             :         }
    1033             : #endif
    1034             : 
    1035         280 :         switch (poller->state) {
    1036         142 :         case SPDK_POLLER_STATE_UNREGISTERED:
    1037         142 :                 free(poller);
    1038         142 :                 break;
    1039           4 :         case SPDK_POLLER_STATE_PAUSING:
    1040           4 :                 TAILQ_INSERT_TAIL(&thread->paused_pollers, poller, tailq);
    1041           4 :                 poller->state = SPDK_POLLER_STATE_PAUSED;
    1042           4 :                 break;
    1043           0 :         case SPDK_POLLER_STATE_PAUSED:
    1044           0 :                 break;
    1045         134 :         case SPDK_POLLER_STATE_RUNNING:
    1046         134 :                 poller->state = SPDK_POLLER_STATE_WAITING;
    1047             :         /* fallthrough */
    1048         134 :         case SPDK_POLLER_STATE_WAITING:
    1049         134 :                 poller_insert_timer(thread, poller, now);
    1050         134 :                 break;
    1051           0 :         default:
    1052           0 :                 assert(false);
    1053             :                 break;
    1054             :         }
    1055             : 
    1056         280 :         return rc;
    1057             : }
    1058             : 
    1059             : static int
    1060      298948 : thread_poll(struct spdk_thread *thread, uint32_t max_msgs, uint64_t now)
    1061             : {
    1062             :         uint32_t msg_count;
    1063             :         struct spdk_poller *poller, *tmp;
    1064             :         spdk_msg_fn critical_msg;
    1065      298948 :         int rc = 0;
    1066             : 
    1067      298948 :         thread->tsc_last = now;
    1068             : 
    1069      298948 :         critical_msg = thread->critical_msg;
    1070      298948 :         if (spdk_unlikely(critical_msg != NULL)) {
    1071           0 :                 critical_msg(NULL);
    1072           0 :                 thread->critical_msg = NULL;
    1073           0 :                 rc = 1;
    1074             :         }
    1075             : 
    1076      298948 :         msg_count = msg_queue_run_batch(thread, max_msgs);
    1077      298948 :         if (msg_count) {
    1078       75085 :                 rc = 1;
    1079             :         }
    1080             : 
    1081      300583 :         TAILQ_FOREACH_REVERSE_SAFE(poller, &thread->active_pollers,
    1082             :                                    active_pollers_head, tailq, tmp) {
    1083             :                 int poller_rc;
    1084             : 
    1085        1635 :                 poller_rc = thread_execute_poller(thread, poller);
    1086        1635 :                 if (poller_rc > rc) {
    1087         243 :                         rc = poller_rc;
    1088             :                 }
    1089             :         }
    1090             : 
    1091      298948 :         poller = thread->first_timed_poller;
    1092      299374 :         while (poller != NULL) {
    1093        3010 :                 int timer_rc = 0;
    1094             : 
    1095        3010 :                 if (now < poller->next_run_tick) {
    1096        2584 :                         break;
    1097             :                 }
    1098             : 
    1099         426 :                 tmp = RB_NEXT(timed_pollers_tree, &thread->timed_pollers, poller);
    1100         426 :                 RB_REMOVE(timed_pollers_tree, &thread->timed_pollers, poller);
    1101             : 
    1102             :                 /* Update the cache to the next timed poller in the list
    1103             :                  * only if the current poller is still the closest, otherwise,
    1104             :                  * do nothing because the cache has been already updated.
    1105             :                  */
    1106         426 :                 if (thread->first_timed_poller == poller) {
    1107         426 :                         thread->first_timed_poller = tmp;
    1108             :                 }
    1109             : 
    1110         426 :                 timer_rc = thread_execute_timed_poller(thread, poller, now);
    1111         426 :                 if (timer_rc > rc) {
    1112         182 :                         rc = timer_rc;
    1113             :                 }
    1114             : 
    1115         426 :                 poller = tmp;
    1116             :         }
    1117             : 
    1118      298948 :         return rc;
    1119             : }
    1120             : 
    1121             : static void
    1122           0 : _thread_remove_pollers(void *ctx)
    1123             : {
    1124           0 :         struct spdk_thread *thread = ctx;
    1125             :         struct spdk_poller *poller, *tmp;
    1126             : 
    1127           0 :         TAILQ_FOREACH_REVERSE_SAFE(poller, &thread->active_pollers,
    1128             :                                    active_pollers_head, tailq, tmp) {
    1129           0 :                 if (poller->state == SPDK_POLLER_STATE_UNREGISTERED) {
    1130           0 :                         TAILQ_REMOVE(&thread->active_pollers, poller, tailq);
    1131           0 :                         free(poller);
    1132             :                 }
    1133             :         }
    1134             : 
    1135           0 :         RB_FOREACH_SAFE(poller, timed_pollers_tree, &thread->timed_pollers, tmp) {
    1136           0 :                 if (poller->state == SPDK_POLLER_STATE_UNREGISTERED) {
    1137           0 :                         poller_remove_timer(thread, poller);
    1138           0 :                         free(poller);
    1139             :                 }
    1140             :         }
    1141             : 
    1142           0 :         thread->poller_unregistered = false;
    1143           0 : }
    1144             : 
    1145             : static void
    1146           0 : _thread_exit(void *ctx)
    1147             : {
    1148           0 :         struct spdk_thread *thread = ctx;
    1149             : 
    1150           0 :         assert(thread->state == SPDK_THREAD_STATE_EXITING);
    1151             : 
    1152           0 :         thread_exit(thread, spdk_get_ticks());
    1153             : 
    1154           0 :         if (thread->state != SPDK_THREAD_STATE_EXITED) {
    1155           0 :                 spdk_thread_send_msg(thread, _thread_exit, thread);
    1156             :         }
    1157           0 : }
    1158             : 
    1159             : int
    1160      298948 : spdk_thread_poll(struct spdk_thread *thread, uint32_t max_msgs, uint64_t now)
    1161             : {
    1162             :         struct spdk_thread *orig_thread;
    1163             :         int rc;
    1164             : 
    1165      298948 :         orig_thread = _get_thread();
    1166      298948 :         tls_thread = thread;
    1167             : 
    1168      298948 :         if (now == 0) {
    1169      294618 :                 now = spdk_get_ticks();
    1170             :         }
    1171             : 
    1172      298948 :         if (spdk_likely(!thread->in_interrupt)) {
    1173      298948 :                 rc = thread_poll(thread, max_msgs, now);
    1174      298948 :                 if (spdk_unlikely(thread->in_interrupt)) {
    1175             :                         /* The thread transitioned to interrupt mode during the above poll.
    1176             :                          * Poll it one more time in case that during the transition time
    1177             :                          * there is msg received without notification.
    1178             :                          */
    1179           0 :                         rc = thread_poll(thread, max_msgs, now);
    1180             :                 }
    1181             : 
    1182      298948 :                 if (spdk_unlikely(thread->state == SPDK_THREAD_STATE_EXITING)) {
    1183         197 :                         thread_exit(thread, now);
    1184             :                 }
    1185             :         } else {
    1186             :                 /* Non-block wait on thread's fd_group */
    1187           0 :                 rc = spdk_fd_group_wait(thread->fgrp, 0);
    1188             :         }
    1189             : 
    1190      298948 :         thread_update_stats(thread, spdk_get_ticks(), now, rc);
    1191             : 
    1192      298948 :         tls_thread = orig_thread;
    1193             : 
    1194      298948 :         return rc;
    1195             : }
    1196             : 
    1197             : uint64_t
    1198           0 : spdk_thread_next_poller_expiration(struct spdk_thread *thread)
    1199             : {
    1200             :         struct spdk_poller *poller;
    1201             : 
    1202           0 :         poller = thread->first_timed_poller;
    1203           0 :         if (poller) {
    1204           0 :                 return poller->next_run_tick;
    1205             :         }
    1206             : 
    1207           0 :         return 0;
    1208             : }
    1209             : 
    1210             : int
    1211           0 : spdk_thread_has_active_pollers(struct spdk_thread *thread)
    1212             : {
    1213           0 :         return !TAILQ_EMPTY(&thread->active_pollers);
    1214             : }
    1215             : 
    1216             : static bool
    1217          24 : thread_has_unpaused_pollers(struct spdk_thread *thread)
    1218             : {
    1219          24 :         if (TAILQ_EMPTY(&thread->active_pollers) &&
    1220          24 :             RB_EMPTY(&thread->timed_pollers)) {
    1221          24 :                 return false;
    1222             :         }
    1223             : 
    1224           0 :         return true;
    1225             : }
    1226             : 
    1227             : bool
    1228           2 : spdk_thread_has_pollers(struct spdk_thread *thread)
    1229             : {
    1230           2 :         if (!thread_has_unpaused_pollers(thread) &&
    1231           2 :             TAILQ_EMPTY(&thread->paused_pollers)) {
    1232           2 :                 return false;
    1233             :         }
    1234             : 
    1235           0 :         return true;
    1236             : }
    1237             : 
    1238             : bool
    1239          22 : spdk_thread_is_idle(struct spdk_thread *thread)
    1240             : {
    1241          44 :         if (spdk_ring_count(thread->messages) ||
    1242          22 :             thread_has_unpaused_pollers(thread) ||
    1243          22 :             thread->critical_msg != NULL) {
    1244           0 :                 return false;
    1245             :         }
    1246             : 
    1247          22 :         return true;
    1248             : }
    1249             : 
    1250             : uint32_t
    1251          13 : spdk_thread_get_count(void)
    1252             : {
    1253             :         /*
    1254             :          * Return cached value of the current thread count.  We could acquire the
    1255             :          *  lock and iterate through the TAILQ of threads to count them, but that
    1256             :          *  count could still be invalidated after we release the lock.
    1257             :          */
    1258          13 :         return g_thread_count;
    1259             : }
    1260             : 
    1261             : struct spdk_thread *
    1262      219889 : spdk_get_thread(void)
    1263             : {
    1264      219889 :         return _get_thread();
    1265             : }
    1266             : 
    1267             : const char *
    1268           4 : spdk_thread_get_name(const struct spdk_thread *thread)
    1269             : {
    1270           4 :         return thread->name;
    1271             : }
    1272             : 
    1273             : uint64_t
    1274          15 : spdk_thread_get_id(const struct spdk_thread *thread)
    1275             : {
    1276          15 :         return thread->id;
    1277             : }
    1278             : 
    1279             : struct spdk_thread *
    1280          13 : spdk_thread_get_by_id(uint64_t id)
    1281             : {
    1282             :         struct spdk_thread *thread;
    1283             : 
    1284          13 :         if (id == 0 || id >= g_thread_id) {
    1285           0 :                 SPDK_ERRLOG("invalid thread id: %" PRIu64 ".\n", id);
    1286           0 :                 return NULL;
    1287             :         }
    1288          13 :         pthread_mutex_lock(&g_devlist_mutex);
    1289          28 :         TAILQ_FOREACH(thread, &g_threads, tailq) {
    1290          28 :                 if (thread->id == id) {
    1291          13 :                         break;
    1292             :                 }
    1293             :         }
    1294          13 :         pthread_mutex_unlock(&g_devlist_mutex);
    1295          13 :         return thread;
    1296             : }
    1297             : 
    1298             : int
    1299          56 : spdk_thread_get_stats(struct spdk_thread_stats *stats)
    1300             : {
    1301             :         struct spdk_thread *thread;
    1302             : 
    1303          56 :         thread = _get_thread();
    1304          56 :         if (!thread) {
    1305           0 :                 SPDK_ERRLOG("No thread allocated\n");
    1306           0 :                 return -EINVAL;
    1307             :         }
    1308             : 
    1309          56 :         if (stats == NULL) {
    1310           0 :                 return -EINVAL;
    1311             :         }
    1312             : 
    1313          56 :         *stats = thread->stats;
    1314             : 
    1315          56 :         return 0;
    1316             : }
    1317             : 
    1318             : uint64_t
    1319       75135 : spdk_thread_get_last_tsc(struct spdk_thread *thread)
    1320             : {
    1321       75135 :         if (thread == NULL) {
    1322           0 :                 thread = _get_thread();
    1323             :         }
    1324             : 
    1325       75135 :         return thread->tsc_last;
    1326             : }
    1327             : 
    1328             : static inline int
    1329       76516 : thread_send_msg_notification(const struct spdk_thread *target_thread)
    1330             : {
    1331       76516 :         uint64_t notify = 1;
    1332             :         int rc;
    1333             : 
    1334             :         /* Not necessary to do notification if interrupt facility is not enabled */
    1335       76516 :         if (spdk_likely(!spdk_interrupt_mode_is_enabled())) {
    1336       76516 :                 return 0;
    1337             :         }
    1338             : 
    1339             :         /* When each spdk_thread can switch between poll and interrupt mode dynamically,
    1340             :          * after sending thread msg, it is necessary to check whether target thread runs in
    1341             :          * interrupt mode and then decide whether do event notification.
    1342             :          */
    1343           0 :         if (spdk_unlikely(target_thread->in_interrupt)) {
    1344           0 :                 rc = write(target_thread->msg_fd, &notify, sizeof(notify));
    1345           0 :                 if (rc < 0) {
    1346           0 :                         SPDK_ERRLOG("failed to notify msg_queue: %s.\n", spdk_strerror(errno));
    1347           0 :                         return -EIO;
    1348             :                 }
    1349             :         }
    1350             : 
    1351           0 :         return 0;
    1352             : }
    1353             : 
    1354             : int
    1355       76516 : spdk_thread_send_msg(const struct spdk_thread *thread, spdk_msg_fn fn, void *ctx)
    1356             : {
    1357             :         struct spdk_thread *local_thread;
    1358       76516 :         struct spdk_msg *msg;
    1359             :         int rc;
    1360             : 
    1361       76516 :         assert(thread != NULL);
    1362             : 
    1363       76516 :         if (spdk_unlikely(thread->state == SPDK_THREAD_STATE_EXITED)) {
    1364           0 :                 SPDK_ERRLOG("Thread %s is marked as exited.\n", thread->name);
    1365           0 :                 return -EIO;
    1366             :         }
    1367             : 
    1368       76516 :         local_thread = _get_thread();
    1369             : 
    1370       76516 :         msg = NULL;
    1371       76516 :         if (local_thread != NULL) {
    1372       76514 :                 if (local_thread->msg_cache_count > 0) {
    1373       76514 :                         msg = SLIST_FIRST(&local_thread->msg_cache);
    1374       76514 :                         assert(msg != NULL);
    1375       76514 :                         SLIST_REMOVE_HEAD(&local_thread->msg_cache, link);
    1376       76514 :                         local_thread->msg_cache_count--;
    1377             :                 }
    1378             :         }
    1379             : 
    1380       76516 :         if (msg == NULL) {
    1381           2 :                 msg = spdk_mempool_get(g_spdk_msg_mempool);
    1382           2 :                 if (!msg) {
    1383           0 :                         SPDK_ERRLOG("msg could not be allocated\n");
    1384           0 :                         return -ENOMEM;
    1385             :                 }
    1386             :         }
    1387             : 
    1388       76516 :         msg->fn = fn;
    1389       76516 :         msg->arg = ctx;
    1390             : 
    1391       76516 :         rc = spdk_ring_enqueue(thread->messages, (void **)&msg, 1, NULL);
    1392       76516 :         if (rc != 1) {
    1393           0 :                 SPDK_ERRLOG("msg could not be enqueued\n");
    1394           0 :                 spdk_mempool_put(g_spdk_msg_mempool, msg);
    1395           0 :                 return -EIO;
    1396             :         }
    1397             : 
    1398       76516 :         return thread_send_msg_notification(thread);
    1399             : }
    1400             : 
    1401             : int
    1402           0 : spdk_thread_send_critical_msg(struct spdk_thread *thread, spdk_msg_fn fn)
    1403             : {
    1404           0 :         spdk_msg_fn expected = NULL;
    1405             : 
    1406           0 :         if (!__atomic_compare_exchange_n(&thread->critical_msg, &expected, fn, false, __ATOMIC_SEQ_CST,
    1407             :                                          __ATOMIC_SEQ_CST)) {
    1408           0 :                 return -EIO;
    1409             :         }
    1410             : 
    1411           0 :         return thread_send_msg_notification(thread);
    1412             : }
    1413             : 
    1414             : #ifdef __linux__
    1415             : static int
    1416           0 : interrupt_timerfd_process(void *arg)
    1417             : {
    1418           0 :         struct spdk_poller *poller = arg;
    1419           0 :         uint64_t exp;
    1420             :         int rc;
    1421             : 
    1422             :         /* clear the level of interval timer */
    1423           0 :         rc = read(poller->intr->efd, &exp, sizeof(exp));
    1424           0 :         if (rc < 0) {
    1425           0 :                 if (rc == -EAGAIN) {
    1426           0 :                         return 0;
    1427             :                 }
    1428             : 
    1429           0 :                 return rc;
    1430             :         }
    1431             : 
    1432             :         SPDK_DTRACE_PROBE2(timerfd_exec, poller->fn, poller->arg);
    1433             : 
    1434           0 :         return poller->fn(poller->arg);
    1435             : }
    1436             : 
    1437             : static int
    1438           0 : period_poller_interrupt_init(struct spdk_poller *poller)
    1439             : {
    1440             :         int timerfd;
    1441             : 
    1442           0 :         SPDK_DEBUGLOG(thread, "timerfd init for periodic poller %s\n", poller->name);
    1443           0 :         timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK | TFD_CLOEXEC);
    1444           0 :         if (timerfd < 0) {
    1445           0 :                 return -errno;
    1446             :         }
    1447             : 
    1448           0 :         poller->intr = spdk_interrupt_register(timerfd, interrupt_timerfd_process, poller, poller->name);
    1449           0 :         if (poller->intr == NULL) {
    1450           0 :                 close(timerfd);
    1451           0 :                 return -1;
    1452             :         }
    1453             : 
    1454           0 :         return 0;
    1455             : }
    1456             : 
    1457             : static void
    1458           0 : period_poller_set_interrupt_mode(struct spdk_poller *poller, void *cb_arg, bool interrupt_mode)
    1459             : {
    1460             :         int timerfd;
    1461           0 :         uint64_t now_tick = spdk_get_ticks();
    1462           0 :         uint64_t ticks = spdk_get_ticks_hz();
    1463             :         int ret;
    1464           0 :         struct itimerspec new_tv = {};
    1465           0 :         struct itimerspec old_tv = {};
    1466             : 
    1467           0 :         assert(poller->intr != NULL);
    1468           0 :         assert(poller->period_ticks != 0);
    1469             : 
    1470           0 :         timerfd = poller->intr->efd;
    1471             : 
    1472           0 :         assert(timerfd >= 0);
    1473             : 
    1474           0 :         SPDK_DEBUGLOG(thread, "timerfd set poller %s into %s mode\n", poller->name,
    1475             :                       interrupt_mode ? "interrupt" : "poll");
    1476             : 
    1477           0 :         if (interrupt_mode) {
    1478             :                 /* Set repeated timer expiration */
    1479           0 :                 new_tv.it_interval.tv_sec = poller->period_ticks / ticks;
    1480           0 :                 new_tv.it_interval.tv_nsec = poller->period_ticks % ticks * SPDK_SEC_TO_NSEC / ticks;
    1481             : 
    1482             :                 /* Update next timer expiration */
    1483           0 :                 if (poller->next_run_tick == 0) {
    1484           0 :                         poller->next_run_tick = now_tick + poller->period_ticks;
    1485           0 :                 } else if (poller->next_run_tick < now_tick) {
    1486           0 :                         poller->next_run_tick = now_tick;
    1487             :                 }
    1488             : 
    1489           0 :                 new_tv.it_value.tv_sec = (poller->next_run_tick - now_tick) / ticks;
    1490           0 :                 new_tv.it_value.tv_nsec = (poller->next_run_tick - now_tick) % ticks * SPDK_SEC_TO_NSEC / ticks;
    1491             : 
    1492           0 :                 ret = timerfd_settime(timerfd, 0, &new_tv, NULL);
    1493           0 :                 if (ret < 0) {
    1494           0 :                         SPDK_ERRLOG("Failed to arm timerfd: error(%d)\n", errno);
    1495           0 :                         assert(false);
    1496             :                 }
    1497             :         } else {
    1498             :                 /* Disarm the timer */
    1499           0 :                 ret = timerfd_settime(timerfd, 0, &new_tv, &old_tv);
    1500           0 :                 if (ret < 0) {
    1501             :                         /* timerfd_settime's failure indicates that the timerfd is in error */
    1502           0 :                         SPDK_ERRLOG("Failed to disarm timerfd: error(%d)\n", errno);
    1503           0 :                         assert(false);
    1504             :                 }
    1505             : 
    1506             :                 /* In order to reuse poller_insert_timer, fix now_tick, so next_run_tick would be
    1507             :                  * now_tick + ticks * old_tv.it_value.tv_sec + (ticks * old_tv.it_value.tv_nsec) / SPDK_SEC_TO_NSEC
    1508             :                  */
    1509           0 :                 now_tick = now_tick - poller->period_ticks + ticks * old_tv.it_value.tv_sec + \
    1510           0 :                            (ticks * old_tv.it_value.tv_nsec) / SPDK_SEC_TO_NSEC;
    1511           0 :                 poller_remove_timer(poller->thread, poller);
    1512           0 :                 poller_insert_timer(poller->thread, poller, now_tick);
    1513             :         }
    1514           0 : }
    1515             : 
    1516             : static void
    1517           0 : poller_interrupt_fini(struct spdk_poller *poller)
    1518             : {
    1519             :         int fd;
    1520             : 
    1521           0 :         SPDK_DEBUGLOG(thread, "interrupt fini for poller %s\n", poller->name);
    1522           0 :         assert(poller->intr != NULL);
    1523           0 :         fd = poller->intr->efd;
    1524           0 :         spdk_interrupt_unregister(&poller->intr);
    1525           0 :         close(fd);
    1526           0 : }
    1527             : 
    1528             : static int
    1529           0 : busy_poller_interrupt_init(struct spdk_poller *poller)
    1530             : {
    1531             :         int busy_efd;
    1532             : 
    1533           0 :         SPDK_DEBUGLOG(thread, "busy_efd init for busy poller %s\n", poller->name);
    1534           0 :         busy_efd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
    1535           0 :         if (busy_efd < 0) {
    1536           0 :                 SPDK_ERRLOG("Failed to create eventfd for Poller(%s).\n", poller->name);
    1537           0 :                 return -errno;
    1538             :         }
    1539             : 
    1540           0 :         poller->intr = spdk_interrupt_register(busy_efd, poller->fn, poller->arg, poller->name);
    1541           0 :         if (poller->intr == NULL) {
    1542           0 :                 close(busy_efd);
    1543           0 :                 return -1;
    1544             :         }
    1545             : 
    1546           0 :         return 0;
    1547             : }
    1548             : 
    1549             : static void
    1550           0 : busy_poller_set_interrupt_mode(struct spdk_poller *poller, void *cb_arg, bool interrupt_mode)
    1551             : {
    1552           0 :         int busy_efd = poller->intr->efd;
    1553           0 :         uint64_t notify = 1;
    1554             :         int rc __attribute__((unused));
    1555             : 
    1556           0 :         assert(busy_efd >= 0);
    1557             : 
    1558           0 :         if (interrupt_mode) {
    1559             :                 /* Write without read on eventfd will get it repeatedly triggered. */
    1560           0 :                 if (write(busy_efd, &notify, sizeof(notify)) < 0) {
    1561           0 :                         SPDK_ERRLOG("Failed to set busy wait for Poller(%s).\n", poller->name);
    1562             :                 }
    1563             :         } else {
    1564             :                 /* Read on eventfd will clear its level triggering. */
    1565           0 :                 rc = read(busy_efd, &notify, sizeof(notify));
    1566             :         }
    1567           0 : }
    1568             : 
    1569             : #else
    1570             : 
    1571             : static int
    1572             : period_poller_interrupt_init(struct spdk_poller *poller)
    1573             : {
    1574             :         return -ENOTSUP;
    1575             : }
    1576             : 
    1577             : static void
    1578             : period_poller_set_interrupt_mode(struct spdk_poller *poller, void *cb_arg, bool interrupt_mode)
    1579             : {
    1580             : }
    1581             : 
    1582             : static void
    1583             : poller_interrupt_fini(struct spdk_poller *poller)
    1584             : {
    1585             : }
    1586             : 
    1587             : static int
    1588             : busy_poller_interrupt_init(struct spdk_poller *poller)
    1589             : {
    1590             :         return -ENOTSUP;
    1591             : }
    1592             : 
    1593             : static void
    1594             : busy_poller_set_interrupt_mode(struct spdk_poller *poller, void *cb_arg, bool interrupt_mode)
    1595             : {
    1596             : }
    1597             : 
    1598             : #endif
    1599             : 
    1600             : void
    1601           0 : spdk_poller_register_interrupt(struct spdk_poller *poller,
    1602             :                                spdk_poller_set_interrupt_mode_cb cb_fn,
    1603             :                                void *cb_arg)
    1604             : {
    1605           0 :         assert(poller != NULL);
    1606           0 :         assert(cb_fn != NULL);
    1607           0 :         assert(spdk_get_thread() == poller->thread);
    1608             : 
    1609           0 :         if (!spdk_interrupt_mode_is_enabled()) {
    1610           0 :                 return;
    1611             :         }
    1612             : 
    1613             :         /* If this poller already had an interrupt, clean the old one up. */
    1614           0 :         if (poller->intr != NULL) {
    1615           0 :                 poller_interrupt_fini(poller);
    1616             :         }
    1617             : 
    1618           0 :         poller->set_intr_cb_fn = cb_fn;
    1619           0 :         poller->set_intr_cb_arg = cb_arg;
    1620             : 
    1621             :         /* Set poller into interrupt mode if thread is in interrupt. */
    1622           0 :         if (poller->thread->in_interrupt) {
    1623           0 :                 poller->set_intr_cb_fn(poller, poller->set_intr_cb_arg, true);
    1624             :         }
    1625             : }
    1626             : 
    1627             : static uint64_t
    1628         753 : convert_us_to_ticks(uint64_t us)
    1629             : {
    1630             :         uint64_t quotient, remainder, ticks;
    1631             : 
    1632         753 :         if (us) {
    1633         320 :                 quotient = us / SPDK_SEC_TO_USEC;
    1634         320 :                 remainder = us % SPDK_SEC_TO_USEC;
    1635         320 :                 ticks = spdk_get_ticks_hz();
    1636             : 
    1637         320 :                 return ticks * quotient + (ticks * remainder) / SPDK_SEC_TO_USEC;
    1638             :         } else {
    1639         433 :                 return 0;
    1640             :         }
    1641             : }
    1642             : 
    1643             : static struct spdk_poller *
    1644         753 : poller_register(spdk_poller_fn fn,
    1645             :                 void *arg,
    1646             :                 uint64_t period_microseconds,
    1647             :                 const char *name)
    1648             : {
    1649             :         struct spdk_thread *thread;
    1650             :         struct spdk_poller *poller;
    1651             : 
    1652         753 :         thread = spdk_get_thread();
    1653         753 :         if (!thread) {
    1654           0 :                 assert(false);
    1655             :                 return NULL;
    1656             :         }
    1657             : 
    1658         753 :         if (spdk_unlikely(thread->state == SPDK_THREAD_STATE_EXITED)) {
    1659           0 :                 SPDK_ERRLOG("thread %s is marked as exited\n", thread->name);
    1660           0 :                 return NULL;
    1661             :         }
    1662             : 
    1663         753 :         poller = calloc(1, sizeof(*poller));
    1664         753 :         if (poller == NULL) {
    1665           0 :                 SPDK_ERRLOG("Poller memory allocation failed\n");
    1666           0 :                 return NULL;
    1667             :         }
    1668             : 
    1669         753 :         if (name) {
    1670         702 :                 snprintf(poller->name, sizeof(poller->name), "%s", name);
    1671             :         } else {
    1672          51 :                 snprintf(poller->name, sizeof(poller->name), "%p", fn);
    1673             :         }
    1674             : 
    1675         753 :         poller->state = SPDK_POLLER_STATE_WAITING;
    1676         753 :         poller->fn = fn;
    1677         753 :         poller->arg = arg;
    1678         753 :         poller->thread = thread;
    1679         753 :         poller->intr = NULL;
    1680         753 :         if (thread->next_poller_id == 0) {
    1681           0 :                 SPDK_WARNLOG("Poller ID rolled over. Poller ID is duplicated.\n");
    1682           0 :                 thread->next_poller_id = 1;
    1683             :         }
    1684         753 :         poller->id = thread->next_poller_id++;
    1685             : 
    1686         753 :         poller->period_ticks = convert_us_to_ticks(period_microseconds);
    1687             : 
    1688         753 :         if (spdk_interrupt_mode_is_enabled()) {
    1689             :                 int rc;
    1690             : 
    1691           0 :                 if (period_microseconds) {
    1692           0 :                         rc = period_poller_interrupt_init(poller);
    1693           0 :                         if (rc < 0) {
    1694           0 :                                 SPDK_ERRLOG("Failed to register interruptfd for periodic poller: %s\n", spdk_strerror(-rc));
    1695           0 :                                 free(poller);
    1696           0 :                                 return NULL;
    1697             :                         }
    1698             : 
    1699           0 :                         poller->set_intr_cb_fn = period_poller_set_interrupt_mode;
    1700           0 :                         poller->set_intr_cb_arg = NULL;
    1701             : 
    1702             :                 } else {
    1703             :                         /* If the poller doesn't have a period, create interruptfd that's always
    1704             :                          * busy automatically when running in interrupt mode.
    1705             :                          */
    1706           0 :                         rc = busy_poller_interrupt_init(poller);
    1707           0 :                         if (rc > 0) {
    1708           0 :                                 SPDK_ERRLOG("Failed to register interruptfd for busy poller: %s\n", spdk_strerror(-rc));
    1709           0 :                                 free(poller);
    1710           0 :                                 return NULL;
    1711             :                         }
    1712             : 
    1713           0 :                         poller->set_intr_cb_fn = busy_poller_set_interrupt_mode;
    1714           0 :                         poller->set_intr_cb_arg = NULL;
    1715             :                 }
    1716             : 
    1717             :                 /* Set poller into interrupt mode if thread is in interrupt. */
    1718           0 :                 if (poller->thread->in_interrupt) {
    1719           0 :                         poller->set_intr_cb_fn(poller, poller->set_intr_cb_arg, true);
    1720             :                 }
    1721             :         }
    1722             : 
    1723         753 :         thread_insert_poller(thread, poller);
    1724             : 
    1725         753 :         return poller;
    1726             : }
    1727             : 
    1728             : struct spdk_poller *
    1729          51 : spdk_poller_register(spdk_poller_fn fn,
    1730             :                      void *arg,
    1731             :                      uint64_t period_microseconds)
    1732             : {
    1733          51 :         return poller_register(fn, arg, period_microseconds, NULL);
    1734             : }
    1735             : 
    1736             : struct spdk_poller *
    1737         702 : spdk_poller_register_named(spdk_poller_fn fn,
    1738             :                            void *arg,
    1739             :                            uint64_t period_microseconds,
    1740             :                            const char *name)
    1741             : {
    1742         702 :         return poller_register(fn, arg, period_microseconds, name);
    1743             : }
    1744             : 
    1745             : static void
    1746           0 : wrong_thread(const char *func, const char *name, struct spdk_thread *thread,
    1747             :              struct spdk_thread *curthread)
    1748             : {
    1749           0 :         if (thread == NULL) {
    1750           0 :                 SPDK_ERRLOG("%s(%s) called with NULL thread\n", func, name);
    1751           0 :                 abort();
    1752             :         }
    1753           0 :         SPDK_ERRLOG("%s(%s) called from wrong thread %s:%" PRIu64 " (should be "
    1754             :                     "%s:%" PRIu64 ")\n", func, name, curthread->name, curthread->id,
    1755             :                     thread->name, thread->id);
    1756           0 :         assert(false);
    1757             : }
    1758             : 
    1759             : void
    1760        1160 : spdk_poller_unregister(struct spdk_poller **ppoller)
    1761             : {
    1762             :         struct spdk_thread *thread;
    1763             :         struct spdk_poller *poller;
    1764             : 
    1765        1160 :         poller = *ppoller;
    1766        1160 :         if (poller == NULL) {
    1767         407 :                 return;
    1768             :         }
    1769             : 
    1770         753 :         *ppoller = NULL;
    1771             : 
    1772         753 :         thread = spdk_get_thread();
    1773         753 :         if (!thread) {
    1774           0 :                 assert(false);
    1775             :                 return;
    1776             :         }
    1777             : 
    1778         753 :         if (poller->thread != thread) {
    1779           0 :                 wrong_thread(__func__, poller->name, poller->thread, thread);
    1780           0 :                 return;
    1781             :         }
    1782             : 
    1783         753 :         if (spdk_interrupt_mode_is_enabled()) {
    1784             :                 /* Release the interrupt resource for period or busy poller */
    1785           0 :                 if (poller->intr != NULL) {
    1786           0 :                         poller_interrupt_fini(poller);
    1787             :                 }
    1788             : 
    1789             :                 /* If there is not already a pending poller removal, generate
    1790             :                  * a message to go process removals. */
    1791           0 :                 if (!thread->poller_unregistered) {
    1792           0 :                         thread->poller_unregistered = true;
    1793           0 :                         spdk_thread_send_msg(thread, _thread_remove_pollers, thread);
    1794             :                 }
    1795             :         }
    1796             : 
    1797             :         /* If the poller was paused, put it on the active_pollers list so that
    1798             :          * its unregistration can be processed by spdk_thread_poll().
    1799             :          */
    1800         753 :         if (poller->state == SPDK_POLLER_STATE_PAUSED) {
    1801           8 :                 TAILQ_REMOVE(&thread->paused_pollers, poller, tailq);
    1802           8 :                 TAILQ_INSERT_TAIL(&thread->active_pollers, poller, tailq);
    1803           8 :                 poller->period_ticks = 0;
    1804             :         }
    1805             : 
    1806             :         /* Simply set the state to unregistered. The poller will get cleaned up
    1807             :          * in a subsequent call to spdk_thread_poll().
    1808             :          */
    1809         753 :         poller->state = SPDK_POLLER_STATE_UNREGISTERED;
    1810             : }
    1811             : 
    1812             : void
    1813          38 : spdk_poller_pause(struct spdk_poller *poller)
    1814             : {
    1815             :         struct spdk_thread *thread;
    1816             : 
    1817          38 :         thread = spdk_get_thread();
    1818          38 :         if (!thread) {
    1819           0 :                 assert(false);
    1820             :                 return;
    1821             :         }
    1822             : 
    1823          38 :         if (poller->thread != thread) {
    1824           0 :                 wrong_thread(__func__, poller->name, poller->thread, thread);
    1825           0 :                 return;
    1826             :         }
    1827             : 
    1828             :         /* We just set its state to SPDK_POLLER_STATE_PAUSING and let
    1829             :          * spdk_thread_poll() move it. It allows a poller to be paused from
    1830             :          * another one's context without breaking the TAILQ_FOREACH_REVERSE_SAFE
    1831             :          * iteration, or from within itself without breaking the logic to always
    1832             :          * remove the closest timed poller in the TAILQ_FOREACH_SAFE iteration.
    1833             :          */
    1834          38 :         switch (poller->state) {
    1835           2 :         case SPDK_POLLER_STATE_PAUSED:
    1836             :         case SPDK_POLLER_STATE_PAUSING:
    1837           2 :                 break;
    1838          36 :         case SPDK_POLLER_STATE_RUNNING:
    1839             :         case SPDK_POLLER_STATE_WAITING:
    1840          36 :                 poller->state = SPDK_POLLER_STATE_PAUSING;
    1841          36 :                 break;
    1842           0 :         default:
    1843           0 :                 assert(false);
    1844             :                 break;
    1845             :         }
    1846             : }
    1847             : 
    1848             : void
    1849          28 : spdk_poller_resume(struct spdk_poller *poller)
    1850             : {
    1851             :         struct spdk_thread *thread;
    1852             : 
    1853          28 :         thread = spdk_get_thread();
    1854          28 :         if (!thread) {
    1855           0 :                 assert(false);
    1856             :                 return;
    1857             :         }
    1858             : 
    1859          28 :         if (poller->thread != thread) {
    1860           0 :                 wrong_thread(__func__, poller->name, poller->thread, thread);
    1861           0 :                 return;
    1862             :         }
    1863             : 
    1864             :         /* If a poller is paused it has to be removed from the paused pollers
    1865             :          * list and put on the active list or timer tree depending on its
    1866             :          * period_ticks.  If a poller is still in the process of being paused,
    1867             :          * we just need to flip its state back to waiting, as it's already on
    1868             :          * the appropriate list or tree.
    1869             :          */
    1870          28 :         switch (poller->state) {
    1871          19 :         case SPDK_POLLER_STATE_PAUSED:
    1872          19 :                 TAILQ_REMOVE(&thread->paused_pollers, poller, tailq);
    1873          19 :                 thread_insert_poller(thread, poller);
    1874             :         /* fallthrough */
    1875          27 :         case SPDK_POLLER_STATE_PAUSING:
    1876          27 :                 poller->state = SPDK_POLLER_STATE_WAITING;
    1877          27 :                 break;
    1878           1 :         case SPDK_POLLER_STATE_RUNNING:
    1879             :         case SPDK_POLLER_STATE_WAITING:
    1880           1 :                 break;
    1881           0 :         default:
    1882           0 :                 assert(false);
    1883             :                 break;
    1884             :         }
    1885             : }
    1886             : 
    1887             : const char *
    1888           0 : spdk_poller_get_name(struct spdk_poller *poller)
    1889             : {
    1890           0 :         return poller->name;
    1891             : }
    1892             : 
    1893             : uint64_t
    1894           0 : spdk_poller_get_id(struct spdk_poller *poller)
    1895             : {
    1896           0 :         return poller->id;
    1897             : }
    1898             : 
    1899             : const char *
    1900           0 : spdk_poller_get_state_str(struct spdk_poller *poller)
    1901             : {
    1902           0 :         switch (poller->state) {
    1903           0 :         case SPDK_POLLER_STATE_WAITING:
    1904           0 :                 return "waiting";
    1905           0 :         case SPDK_POLLER_STATE_RUNNING:
    1906           0 :                 return "running";
    1907           0 :         case SPDK_POLLER_STATE_UNREGISTERED:
    1908           0 :                 return "unregistered";
    1909           0 :         case SPDK_POLLER_STATE_PAUSING:
    1910           0 :                 return "pausing";
    1911           0 :         case SPDK_POLLER_STATE_PAUSED:
    1912           0 :                 return "paused";
    1913           0 :         default:
    1914           0 :                 return NULL;
    1915             :         }
    1916             : }
    1917             : 
    1918             : uint64_t
    1919           0 : spdk_poller_get_period_ticks(struct spdk_poller *poller)
    1920             : {
    1921           0 :         return poller->period_ticks;
    1922             : }
    1923             : 
    1924             : void
    1925           0 : spdk_poller_get_stats(struct spdk_poller *poller, struct spdk_poller_stats *stats)
    1926             : {
    1927           0 :         stats->run_count = poller->run_count;
    1928           0 :         stats->busy_count = poller->busy_count;
    1929           0 : }
    1930             : 
    1931             : struct spdk_poller *
    1932           0 : spdk_thread_get_first_active_poller(struct spdk_thread *thread)
    1933             : {
    1934           0 :         return TAILQ_FIRST(&thread->active_pollers);
    1935             : }
    1936             : 
    1937             : struct spdk_poller *
    1938           0 : spdk_thread_get_next_active_poller(struct spdk_poller *prev)
    1939             : {
    1940           0 :         return TAILQ_NEXT(prev, tailq);
    1941             : }
    1942             : 
    1943             : struct spdk_poller *
    1944           0 : spdk_thread_get_first_timed_poller(struct spdk_thread *thread)
    1945             : {
    1946           0 :         return RB_MIN(timed_pollers_tree, &thread->timed_pollers);
    1947             : }
    1948             : 
    1949             : struct spdk_poller *
    1950           0 : spdk_thread_get_next_timed_poller(struct spdk_poller *prev)
    1951             : {
    1952           0 :         return RB_NEXT(timed_pollers_tree, &thread->timed_pollers, prev);
    1953             : }
    1954             : 
    1955             : struct spdk_poller *
    1956           0 : spdk_thread_get_first_paused_poller(struct spdk_thread *thread)
    1957             : {
    1958           0 :         return TAILQ_FIRST(&thread->paused_pollers);
    1959             : }
    1960             : 
    1961             : struct spdk_poller *
    1962           0 : spdk_thread_get_next_paused_poller(struct spdk_poller *prev)
    1963             : {
    1964           0 :         return TAILQ_NEXT(prev, tailq);
    1965             : }
    1966             : 
    1967             : struct spdk_io_channel *
    1968           0 : spdk_thread_get_first_io_channel(struct spdk_thread *thread)
    1969             : {
    1970           0 :         return RB_MIN(io_channel_tree, &thread->io_channels);
    1971             : }
    1972             : 
    1973             : struct spdk_io_channel *
    1974           0 : spdk_thread_get_next_io_channel(struct spdk_io_channel *prev)
    1975             : {
    1976           0 :         return RB_NEXT(io_channel_tree, &thread->io_channels, prev);
    1977             : }
    1978             : 
    1979             : struct call_thread {
    1980             :         struct spdk_thread *cur_thread;
    1981             :         spdk_msg_fn fn;
    1982             :         void *ctx;
    1983             : 
    1984             :         struct spdk_thread *orig_thread;
    1985             :         spdk_msg_fn cpl;
    1986             : };
    1987             : 
    1988             : static void
    1989           2 : _back_to_orig_thread(void *ctx)
    1990             : {
    1991           2 :         struct call_thread *ct = ctx;
    1992             : 
    1993           2 :         assert(ct->orig_thread->for_each_count > 0);
    1994           2 :         ct->orig_thread->for_each_count--;
    1995             : 
    1996           2 :         ct->cpl(ct->ctx);
    1997           2 :         free(ctx);
    1998           2 : }
    1999             : 
    2000             : static void
    2001           6 : _on_thread(void *ctx)
    2002             : {
    2003           6 :         struct call_thread *ct = ctx;
    2004             :         int rc __attribute__((unused));
    2005             : 
    2006           6 :         ct->fn(ct->ctx);
    2007             : 
    2008           6 :         pthread_mutex_lock(&g_devlist_mutex);
    2009           6 :         ct->cur_thread = TAILQ_NEXT(ct->cur_thread, tailq);
    2010           6 :         while (ct->cur_thread && ct->cur_thread->state != SPDK_THREAD_STATE_RUNNING) {
    2011           0 :                 SPDK_DEBUGLOG(thread, "thread %s is not running but still not destroyed.\n",
    2012             :                               ct->cur_thread->name);
    2013           0 :                 ct->cur_thread = TAILQ_NEXT(ct->cur_thread, tailq);
    2014             :         }
    2015           6 :         pthread_mutex_unlock(&g_devlist_mutex);
    2016             : 
    2017           6 :         if (!ct->cur_thread) {
    2018           2 :                 SPDK_DEBUGLOG(thread, "Completed thread iteration\n");
    2019             : 
    2020           2 :                 rc = spdk_thread_send_msg(ct->orig_thread, _back_to_orig_thread, ctx);
    2021             :         } else {
    2022           4 :                 SPDK_DEBUGLOG(thread, "Continuing thread iteration to %s\n",
    2023             :                               ct->cur_thread->name);
    2024             : 
    2025           4 :                 rc = spdk_thread_send_msg(ct->cur_thread, _on_thread, ctx);
    2026             :         }
    2027           6 :         assert(rc == 0);
    2028           6 : }
    2029             : 
    2030             : void
    2031           2 : spdk_for_each_thread(spdk_msg_fn fn, void *ctx, spdk_msg_fn cpl)
    2032             : {
    2033             :         struct call_thread *ct;
    2034             :         struct spdk_thread *thread;
    2035             :         int rc __attribute__((unused));
    2036             : 
    2037           2 :         ct = calloc(1, sizeof(*ct));
    2038           2 :         if (!ct) {
    2039           0 :                 SPDK_ERRLOG("Unable to perform thread iteration\n");
    2040           0 :                 cpl(ctx);
    2041           0 :                 return;
    2042             :         }
    2043             : 
    2044           2 :         ct->fn = fn;
    2045           2 :         ct->ctx = ctx;
    2046           2 :         ct->cpl = cpl;
    2047             : 
    2048           2 :         thread = _get_thread();
    2049           2 :         if (!thread) {
    2050           0 :                 SPDK_ERRLOG("No thread allocated\n");
    2051           0 :                 free(ct);
    2052           0 :                 cpl(ctx);
    2053           0 :                 return;
    2054             :         }
    2055           2 :         ct->orig_thread = thread;
    2056             : 
    2057           2 :         ct->orig_thread->for_each_count++;
    2058             : 
    2059           2 :         pthread_mutex_lock(&g_devlist_mutex);
    2060           2 :         ct->cur_thread = TAILQ_FIRST(&g_threads);
    2061           2 :         pthread_mutex_unlock(&g_devlist_mutex);
    2062             : 
    2063           2 :         SPDK_DEBUGLOG(thread, "Starting thread iteration from %s\n",
    2064             :                       ct->orig_thread->name);
    2065             : 
    2066           2 :         rc = spdk_thread_send_msg(ct->cur_thread, _on_thread, ct);
    2067           2 :         assert(rc == 0);
    2068             : }
    2069             : 
    2070             : static inline void
    2071           0 : poller_set_interrupt_mode(struct spdk_poller *poller, bool interrupt_mode)
    2072             : {
    2073           0 :         if (poller->state == SPDK_POLLER_STATE_UNREGISTERED) {
    2074           0 :                 return;
    2075             :         }
    2076             : 
    2077           0 :         if (!poller->set_intr_cb_fn) {
    2078           0 :                 SPDK_ERRLOG("Poller(%s) doesn't support set interrupt mode.\n", poller->name);
    2079           0 :                 assert(false);
    2080             :                 return;
    2081             :         }
    2082             : 
    2083           0 :         poller->set_intr_cb_fn(poller, poller->set_intr_cb_arg, interrupt_mode);
    2084             : }
    2085             : 
    2086             : void
    2087           0 : spdk_thread_set_interrupt_mode(bool enable_interrupt)
    2088             : {
    2089           0 :         struct spdk_thread *thread = _get_thread();
    2090             :         struct spdk_poller *poller, *tmp;
    2091             : 
    2092           0 :         assert(thread);
    2093           0 :         assert(spdk_interrupt_mode_is_enabled());
    2094             : 
    2095           0 :         SPDK_NOTICELOG("Set spdk_thread (%s) to %s mode from %s mode.\n",
    2096             :                        thread->name,  enable_interrupt ? "intr" : "poll",
    2097             :                        thread->in_interrupt ? "intr" : "poll");
    2098             : 
    2099           0 :         if (thread->in_interrupt == enable_interrupt) {
    2100           0 :                 return;
    2101             :         }
    2102             : 
    2103             :         /* Set pollers to expected mode */
    2104           0 :         RB_FOREACH_SAFE(poller, timed_pollers_tree, &thread->timed_pollers, tmp) {
    2105           0 :                 poller_set_interrupt_mode(poller, enable_interrupt);
    2106             :         }
    2107           0 :         TAILQ_FOREACH_SAFE(poller, &thread->active_pollers, tailq, tmp) {
    2108           0 :                 poller_set_interrupt_mode(poller, enable_interrupt);
    2109             :         }
    2110             :         /* All paused pollers will go to work in interrupt mode */
    2111           0 :         TAILQ_FOREACH_SAFE(poller, &thread->paused_pollers, tailq, tmp) {
    2112           0 :                 poller_set_interrupt_mode(poller, enable_interrupt);
    2113             :         }
    2114             : 
    2115           0 :         thread->in_interrupt = enable_interrupt;
    2116           0 :         return;
    2117             : }
    2118             : 
    2119             : static struct io_device *
    2120        5654 : io_device_get(void *io_device)
    2121             : {
    2122        5654 :         struct io_device find = {};
    2123             : 
    2124        5654 :         find.io_device = io_device;
    2125        5654 :         return RB_FIND(io_device_tree, &g_io_devices, &find);
    2126             : }
    2127             : 
    2128             : void
    2129        1511 : spdk_io_device_register(void *io_device, spdk_io_channel_create_cb create_cb,
    2130             :                         spdk_io_channel_destroy_cb destroy_cb, uint32_t ctx_size,
    2131             :                         const char *name)
    2132             : {
    2133             :         struct io_device *dev, *tmp;
    2134             :         struct spdk_thread *thread;
    2135             : 
    2136        1511 :         assert(io_device != NULL);
    2137        1511 :         assert(create_cb != NULL);
    2138        1511 :         assert(destroy_cb != NULL);
    2139             : 
    2140        1511 :         thread = spdk_get_thread();
    2141        1511 :         if (!thread) {
    2142           0 :                 SPDK_ERRLOG("called from non-SPDK thread\n");
    2143           0 :                 assert(false);
    2144             :                 return;
    2145             :         }
    2146             : 
    2147        1511 :         dev = calloc(1, sizeof(struct io_device));
    2148        1511 :         if (dev == NULL) {
    2149           0 :                 SPDK_ERRLOG("could not allocate io_device\n");
    2150           0 :                 return;
    2151             :         }
    2152             : 
    2153        1511 :         dev->io_device = io_device;
    2154        1511 :         if (name) {
    2155        1396 :                 snprintf(dev->name, sizeof(dev->name), "%s", name);
    2156             :         } else {
    2157         115 :                 snprintf(dev->name, sizeof(dev->name), "%p", dev);
    2158             :         }
    2159        1511 :         dev->create_cb = create_cb;
    2160        1511 :         dev->destroy_cb = destroy_cb;
    2161        1511 :         dev->unregister_cb = NULL;
    2162        1511 :         dev->ctx_size = ctx_size;
    2163        1511 :         dev->for_each_count = 0;
    2164        1511 :         dev->unregistered = false;
    2165        1511 :         dev->refcnt = 0;
    2166             : 
    2167        1511 :         SPDK_DEBUGLOG(thread, "Registering io_device %s (%p) on thread %s\n",
    2168             :                       dev->name, dev->io_device, thread->name);
    2169             : 
    2170        1511 :         pthread_mutex_lock(&g_devlist_mutex);
    2171        1511 :         tmp = RB_INSERT(io_device_tree, &g_io_devices, dev);
    2172        1511 :         if (tmp != NULL) {
    2173           2 :                 SPDK_ERRLOG("io_device %p already registered (old:%s new:%s)\n",
    2174             :                             io_device, tmp->name, dev->name);
    2175           2 :                 free(dev);
    2176             :         }
    2177             : 
    2178        1511 :         pthread_mutex_unlock(&g_devlist_mutex);
    2179             : }
    2180             : 
    2181             : static void
    2182        1325 : _finish_unregister(void *arg)
    2183             : {
    2184        1325 :         struct io_device *dev = arg;
    2185             :         struct spdk_thread *thread;
    2186             : 
    2187        1325 :         thread = spdk_get_thread();
    2188        1325 :         assert(thread == dev->unregister_thread);
    2189             : 
    2190        1325 :         SPDK_DEBUGLOG(thread, "Finishing unregistration of io_device %s (%p) on thread %s\n",
    2191             :                       dev->name, dev->io_device, thread->name);
    2192             : 
    2193        1325 :         assert(thread->pending_unregister_count > 0);
    2194        1325 :         thread->pending_unregister_count--;
    2195             : 
    2196        1325 :         dev->unregister_cb(dev->io_device);
    2197        1325 :         free(dev);
    2198        1325 : }
    2199             : 
    2200             : static void
    2201        1508 : io_device_free(struct io_device *dev)
    2202             : {
    2203             :         int rc __attribute__((unused));
    2204             : 
    2205        1508 :         if (dev->unregister_cb == NULL) {
    2206         183 :                 free(dev);
    2207             :         } else {
    2208        1325 :                 assert(dev->unregister_thread != NULL);
    2209        1325 :                 SPDK_DEBUGLOG(thread, "io_device %s (%p) needs to unregister from thread %s\n",
    2210             :                               dev->name, dev->io_device, dev->unregister_thread->name);
    2211        1325 :                 rc = spdk_thread_send_msg(dev->unregister_thread, _finish_unregister, dev);
    2212        1325 :                 assert(rc == 0);
    2213             :         }
    2214        1508 : }
    2215             : 
    2216             : void
    2217        1510 : spdk_io_device_unregister(void *io_device, spdk_io_device_unregister_cb unregister_cb)
    2218             : {
    2219             :         struct io_device *dev;
    2220             :         uint32_t refcnt;
    2221             :         struct spdk_thread *thread;
    2222             : 
    2223        1510 :         thread = spdk_get_thread();
    2224        1510 :         if (!thread) {
    2225           0 :                 SPDK_ERRLOG("called from non-SPDK thread\n");
    2226           0 :                 assert(false);
    2227             :                 return;
    2228             :         }
    2229             : 
    2230        1510 :         pthread_mutex_lock(&g_devlist_mutex);
    2231        1510 :         dev = io_device_get(io_device);
    2232        1510 :         if (!dev) {
    2233           0 :                 SPDK_ERRLOG("io_device %p not found\n", io_device);
    2234           0 :                 assert(false);
    2235             :                 pthread_mutex_unlock(&g_devlist_mutex);
    2236             :                 return;
    2237             :         }
    2238             : 
    2239             :         /* The for_each_count check differentiates the user attempting to unregister the
    2240             :          * device a second time, from the internal call to this function that occurs
    2241             :          * after the for_each_count reaches 0.
    2242             :          */
    2243        1510 :         if (dev->pending_unregister && dev->for_each_count > 0) {
    2244           0 :                 SPDK_ERRLOG("io_device %p already has a pending unregister\n", io_device);
    2245           0 :                 assert(false);
    2246             :                 pthread_mutex_unlock(&g_devlist_mutex);
    2247             :                 return;
    2248             :         }
    2249             : 
    2250        1510 :         dev->unregister_cb = unregister_cb;
    2251        1510 :         dev->unregister_thread = thread;
    2252             : 
    2253        1510 :         if (dev->for_each_count > 0) {
    2254           1 :                 SPDK_WARNLOG("io_device %s (%p) has %u for_each calls outstanding\n",
    2255             :                              dev->name, io_device, dev->for_each_count);
    2256           1 :                 dev->pending_unregister = true;
    2257           1 :                 pthread_mutex_unlock(&g_devlist_mutex);
    2258           1 :                 return;
    2259             :         }
    2260             : 
    2261        1509 :         dev->unregistered = true;
    2262        1509 :         RB_REMOVE(io_device_tree, &g_io_devices, dev);
    2263        1509 :         refcnt = dev->refcnt;
    2264        1509 :         pthread_mutex_unlock(&g_devlist_mutex);
    2265             : 
    2266        1509 :         SPDK_DEBUGLOG(thread, "Unregistering io_device %s (%p) from thread %s\n",
    2267             :                       dev->name, dev->io_device, thread->name);
    2268             : 
    2269        1509 :         if (unregister_cb) {
    2270        1325 :                 thread->pending_unregister_count++;
    2271             :         }
    2272             : 
    2273        1509 :         if (refcnt > 0) {
    2274             :                 /* defer deletion */
    2275         799 :                 return;
    2276             :         }
    2277             : 
    2278         710 :         io_device_free(dev);
    2279             : }
    2280             : 
    2281             : const char *
    2282           0 : spdk_io_device_get_name(struct io_device *dev)
    2283             : {
    2284           0 :         return dev->name;
    2285             : }
    2286             : 
    2287             : static struct spdk_io_channel *
    2288        7524 : thread_get_io_channel(struct spdk_thread *thread, struct io_device *dev)
    2289             : {
    2290        7524 :         struct spdk_io_channel find = {};
    2291             : 
    2292        7524 :         find.dev = dev;
    2293        7524 :         return RB_FIND(io_channel_tree, &thread->io_channels, &find);
    2294             : }
    2295             : 
    2296             : struct spdk_io_channel *
    2297        2605 : spdk_get_io_channel(void *io_device)
    2298             : {
    2299             :         struct spdk_io_channel *ch;
    2300             :         struct spdk_thread *thread;
    2301             :         struct io_device *dev;
    2302             :         int rc;
    2303             : 
    2304        2605 :         pthread_mutex_lock(&g_devlist_mutex);
    2305        2605 :         dev = io_device_get(io_device);
    2306        2605 :         if (dev == NULL) {
    2307           1 :                 SPDK_ERRLOG("could not find io_device %p\n", io_device);
    2308           1 :                 pthread_mutex_unlock(&g_devlist_mutex);
    2309           1 :                 return NULL;
    2310             :         }
    2311             : 
    2312        2604 :         thread = _get_thread();
    2313        2604 :         if (!thread) {
    2314           0 :                 SPDK_ERRLOG("No thread allocated\n");
    2315           0 :                 pthread_mutex_unlock(&g_devlist_mutex);
    2316           0 :                 return NULL;
    2317             :         }
    2318             : 
    2319        2604 :         if (spdk_unlikely(thread->state == SPDK_THREAD_STATE_EXITED)) {
    2320           0 :                 SPDK_ERRLOG("Thread %s is marked as exited\n", thread->name);
    2321           0 :                 pthread_mutex_unlock(&g_devlist_mutex);
    2322           0 :                 return NULL;
    2323             :         }
    2324             : 
    2325        2604 :         ch = thread_get_io_channel(thread, dev);
    2326        2604 :         if (ch != NULL) {
    2327        1135 :                 ch->ref++;
    2328             : 
    2329        1135 :                 SPDK_DEBUGLOG(thread, "Get io_channel %p for io_device %s (%p) on thread %s refcnt %u\n",
    2330             :                               ch, dev->name, dev->io_device, thread->name, ch->ref);
    2331             : 
    2332             :                 /*
    2333             :                  * An I/O channel already exists for this device on this
    2334             :                  *  thread, so return it.
    2335             :                  */
    2336        1135 :                 pthread_mutex_unlock(&g_devlist_mutex);
    2337        1135 :                 spdk_trace_record(TRACE_THREAD_IOCH_GET, 0, 0,
    2338             :                                   (uint64_t)spdk_io_channel_get_ctx(ch), ch->ref);
    2339        1135 :                 return ch;
    2340             :         }
    2341             : 
    2342        1469 :         ch = calloc(1, sizeof(*ch) + dev->ctx_size);
    2343        1469 :         if (ch == NULL) {
    2344           0 :                 SPDK_ERRLOG("could not calloc spdk_io_channel\n");
    2345           0 :                 pthread_mutex_unlock(&g_devlist_mutex);
    2346           0 :                 return NULL;
    2347             :         }
    2348             : 
    2349        1469 :         ch->dev = dev;
    2350        1469 :         ch->destroy_cb = dev->destroy_cb;
    2351        1469 :         ch->thread = thread;
    2352        1469 :         ch->ref = 1;
    2353        1469 :         ch->destroy_ref = 0;
    2354        1469 :         RB_INSERT(io_channel_tree, &thread->io_channels, ch);
    2355             : 
    2356        1469 :         SPDK_DEBUGLOG(thread, "Get io_channel %p for io_device %s (%p) on thread %s refcnt %u\n",
    2357             :                       ch, dev->name, dev->io_device, thread->name, ch->ref);
    2358             : 
    2359        1469 :         dev->refcnt++;
    2360             : 
    2361        1469 :         pthread_mutex_unlock(&g_devlist_mutex);
    2362             : 
    2363        1469 :         rc = dev->create_cb(io_device, (uint8_t *)ch + sizeof(*ch));
    2364        1469 :         if (rc != 0) {
    2365           3 :                 pthread_mutex_lock(&g_devlist_mutex);
    2366           3 :                 RB_REMOVE(io_channel_tree, &ch->thread->io_channels, ch);
    2367           3 :                 dev->refcnt--;
    2368           3 :                 free(ch);
    2369           3 :                 SPDK_ERRLOG("could not create io_channel for io_device %s (%p): %s (rc=%d)\n",
    2370             :                             dev->name, io_device, spdk_strerror(-rc), rc);
    2371           3 :                 pthread_mutex_unlock(&g_devlist_mutex);
    2372           3 :                 return NULL;
    2373             :         }
    2374             : 
    2375        1466 :         spdk_trace_record(TRACE_THREAD_IOCH_GET, 0, 0, (uint64_t)spdk_io_channel_get_ctx(ch), 1);
    2376        1466 :         return ch;
    2377             : }
    2378             : 
    2379             : static void
    2380        1469 : put_io_channel(void *arg)
    2381             : {
    2382        1469 :         struct spdk_io_channel *ch = arg;
    2383        1469 :         bool do_remove_dev = true;
    2384             :         struct spdk_thread *thread;
    2385             : 
    2386        1469 :         thread = spdk_get_thread();
    2387        1469 :         if (!thread) {
    2388           0 :                 SPDK_ERRLOG("called from non-SPDK thread\n");
    2389           0 :                 assert(false);
    2390             :                 return;
    2391             :         }
    2392             : 
    2393        1469 :         SPDK_DEBUGLOG(thread,
    2394             :                       "Releasing io_channel %p for io_device %s (%p) on thread %s\n",
    2395             :                       ch, ch->dev->name, ch->dev->io_device, thread->name);
    2396             : 
    2397        1469 :         assert(ch->thread == thread);
    2398             : 
    2399        1469 :         ch->destroy_ref--;
    2400             : 
    2401        1469 :         if (ch->ref > 0 || ch->destroy_ref > 0) {
    2402             :                 /*
    2403             :                  * Another reference to the associated io_device was requested
    2404             :                  *  after this message was sent but before it had a chance to
    2405             :                  *  execute.
    2406             :                  */
    2407           4 :                 return;
    2408             :         }
    2409             : 
    2410        1465 :         pthread_mutex_lock(&g_devlist_mutex);
    2411        1465 :         RB_REMOVE(io_channel_tree, &ch->thread->io_channels, ch);
    2412        1465 :         pthread_mutex_unlock(&g_devlist_mutex);
    2413             : 
    2414             :         /* Don't hold the devlist mutex while the destroy_cb is called. */
    2415        1465 :         ch->destroy_cb(ch->dev->io_device, spdk_io_channel_get_ctx(ch));
    2416             : 
    2417        1465 :         pthread_mutex_lock(&g_devlist_mutex);
    2418        1465 :         ch->dev->refcnt--;
    2419             : 
    2420        1465 :         if (!ch->dev->unregistered) {
    2421         661 :                 do_remove_dev = false;
    2422             :         }
    2423             : 
    2424        1465 :         if (ch->dev->refcnt > 0) {
    2425         133 :                 do_remove_dev = false;
    2426             :         }
    2427             : 
    2428        1465 :         pthread_mutex_unlock(&g_devlist_mutex);
    2429             : 
    2430        1465 :         if (do_remove_dev) {
    2431         798 :                 io_device_free(ch->dev);
    2432             :         }
    2433        1465 :         free(ch);
    2434             : }
    2435             : 
    2436             : void
    2437        2601 : spdk_put_io_channel(struct spdk_io_channel *ch)
    2438             : {
    2439             :         struct spdk_thread *thread;
    2440             :         int rc __attribute__((unused));
    2441             : 
    2442        2601 :         spdk_trace_record(TRACE_THREAD_IOCH_PUT, 0, 0,
    2443             :                           (uint64_t)spdk_io_channel_get_ctx(ch), ch->ref);
    2444             : 
    2445        2601 :         thread = spdk_get_thread();
    2446        2601 :         if (!thread) {
    2447           0 :                 SPDK_ERRLOG("called from non-SPDK thread\n");
    2448           0 :                 assert(false);
    2449             :                 return;
    2450             :         }
    2451             : 
    2452        2601 :         if (ch->thread != thread) {
    2453           0 :                 wrong_thread(__func__, "ch", ch->thread, thread);
    2454           0 :                 return;
    2455             :         }
    2456             : 
    2457        2601 :         SPDK_DEBUGLOG(thread,
    2458             :                       "Putting io_channel %p for io_device %s (%p) on thread %s refcnt %u\n",
    2459             :                       ch, ch->dev->name, ch->dev->io_device, thread->name, ch->ref);
    2460             : 
    2461        2601 :         ch->ref--;
    2462             : 
    2463        2601 :         if (ch->ref == 0) {
    2464        1469 :                 ch->destroy_ref++;
    2465        1469 :                 rc = spdk_thread_send_msg(thread, put_io_channel, ch);
    2466        1469 :                 assert(rc == 0);
    2467             :         }
    2468             : }
    2469             : 
    2470             : struct spdk_io_channel *
    2471        2407 : spdk_io_channel_from_ctx(void *ctx)
    2472             : {
    2473        2407 :         return (struct spdk_io_channel *)((uint8_t *)ctx - sizeof(struct spdk_io_channel));
    2474             : }
    2475             : 
    2476             : struct spdk_thread *
    2477        2038 : spdk_io_channel_get_thread(struct spdk_io_channel *ch)
    2478             : {
    2479        2038 :         return ch->thread;
    2480             : }
    2481             : 
    2482             : void *
    2483          46 : spdk_io_channel_get_io_device(struct spdk_io_channel *ch)
    2484             : {
    2485          46 :         return ch->dev->io_device;
    2486             : }
    2487             : 
    2488             : const char *
    2489           0 : spdk_io_channel_get_io_device_name(struct spdk_io_channel *ch)
    2490             : {
    2491           0 :         return spdk_io_device_get_name(ch->dev);
    2492             : }
    2493             : 
    2494             : int
    2495           0 : spdk_io_channel_get_ref_count(struct spdk_io_channel *ch)
    2496             : {
    2497           0 :         return ch->ref;
    2498             : }
    2499             : 
    2500             : struct spdk_io_channel_iter {
    2501             :         void *io_device;
    2502             :         struct io_device *dev;
    2503             :         spdk_channel_msg fn;
    2504             :         int status;
    2505             :         void *ctx;
    2506             :         struct spdk_io_channel *ch;
    2507             : 
    2508             :         struct spdk_thread *cur_thread;
    2509             : 
    2510             :         struct spdk_thread *orig_thread;
    2511             :         spdk_channel_for_each_cpl cpl;
    2512             : };
    2513             : 
    2514             : void *
    2515         480 : spdk_io_channel_iter_get_io_device(struct spdk_io_channel_iter *i)
    2516             : {
    2517         480 :         return i->io_device;
    2518             : }
    2519             : 
    2520             : struct spdk_io_channel *
    2521        1014 : spdk_io_channel_iter_get_channel(struct spdk_io_channel_iter *i)
    2522             : {
    2523        1014 :         return i->ch;
    2524             : }
    2525             : 
    2526             : void *
    2527        2351 : spdk_io_channel_iter_get_ctx(struct spdk_io_channel_iter *i)
    2528             : {
    2529        2351 :         return i->ctx;
    2530             : }
    2531             : 
    2532             : static void
    2533        1536 : _call_completion(void *ctx)
    2534             : {
    2535        1536 :         struct spdk_io_channel_iter *i = ctx;
    2536             : 
    2537        1536 :         assert(i->orig_thread->for_each_count > 0);
    2538        1536 :         i->orig_thread->for_each_count--;
    2539             : 
    2540        1536 :         if (i->cpl != NULL) {
    2541        1536 :                 i->cpl(i, i->status);
    2542             :         }
    2543        1536 :         free(i);
    2544        1536 : }
    2545             : 
    2546             : static void
    2547        1571 : _call_channel(void *ctx)
    2548             : {
    2549        1571 :         struct spdk_io_channel_iter *i = ctx;
    2550             :         struct spdk_io_channel *ch;
    2551             : 
    2552             :         /*
    2553             :          * It is possible that the channel was deleted before this
    2554             :          *  message had a chance to execute.  If so, skip calling
    2555             :          *  the fn() on this thread.
    2556             :          */
    2557        1571 :         pthread_mutex_lock(&g_devlist_mutex);
    2558        1571 :         ch = thread_get_io_channel(i->cur_thread, i->dev);
    2559        1571 :         pthread_mutex_unlock(&g_devlist_mutex);
    2560             : 
    2561        1571 :         if (ch) {
    2562        1531 :                 i->fn(i);
    2563             :         } else {
    2564          40 :                 spdk_for_each_channel_continue(i, 0);
    2565             :         }
    2566        1571 : }
    2567             : 
    2568             : void
    2569        1536 : spdk_for_each_channel(void *io_device, spdk_channel_msg fn, void *ctx,
    2570             :                       spdk_channel_for_each_cpl cpl)
    2571             : {
    2572             :         struct spdk_thread *thread;
    2573             :         struct spdk_io_channel *ch;
    2574             :         struct spdk_io_channel_iter *i;
    2575             :         int rc __attribute__((unused));
    2576             : 
    2577        1536 :         i = calloc(1, sizeof(*i));
    2578        1536 :         if (!i) {
    2579           0 :                 SPDK_ERRLOG("Unable to allocate iterator\n");
    2580           0 :                 assert(false);
    2581             :                 return;
    2582             :         }
    2583             : 
    2584        1536 :         i->io_device = io_device;
    2585        1536 :         i->fn = fn;
    2586        1536 :         i->ctx = ctx;
    2587        1536 :         i->cpl = cpl;
    2588        1536 :         i->orig_thread = _get_thread();
    2589             : 
    2590        1536 :         i->orig_thread->for_each_count++;
    2591             : 
    2592        1536 :         pthread_mutex_lock(&g_devlist_mutex);
    2593        1536 :         i->dev = io_device_get(io_device);
    2594        1536 :         if (i->dev == NULL) {
    2595           0 :                 SPDK_ERRLOG("could not find io_device %p\n", io_device);
    2596           0 :                 assert(false);
    2597             :                 i->status = -ENODEV;
    2598             :                 goto end;
    2599             :         }
    2600             : 
    2601             :         /* Do not allow new for_each operations if we are already waiting to unregister
    2602             :          * the device for other for_each operations to complete.
    2603             :          */
    2604        1536 :         if (i->dev->pending_unregister) {
    2605           0 :                 SPDK_ERRLOG("io_device %p has a pending unregister\n", io_device);
    2606           0 :                 i->status = -ENODEV;
    2607           0 :                 goto end;
    2608             :         }
    2609             : 
    2610        1831 :         TAILQ_FOREACH(thread, &g_threads, tailq) {
    2611        1698 :                 ch = thread_get_io_channel(thread, i->dev);
    2612        1698 :                 if (ch != NULL) {
    2613        1403 :                         ch->dev->for_each_count++;
    2614        1403 :                         i->cur_thread = thread;
    2615        1403 :                         i->ch = ch;
    2616        1403 :                         pthread_mutex_unlock(&g_devlist_mutex);
    2617        1403 :                         rc = spdk_thread_send_msg(thread, _call_channel, i);
    2618        1403 :                         assert(rc == 0);
    2619        1403 :                         return;
    2620             :                 }
    2621             :         }
    2622             : 
    2623         133 : end:
    2624         133 :         pthread_mutex_unlock(&g_devlist_mutex);
    2625             : 
    2626         133 :         rc = spdk_thread_send_msg(i->orig_thread, _call_completion, i);
    2627         133 :         assert(rc == 0);
    2628             : }
    2629             : 
    2630             : static void
    2631           1 : __pending_unregister(void *arg)
    2632             : {
    2633           1 :         struct io_device *dev = arg;
    2634             : 
    2635           1 :         assert(dev->pending_unregister);
    2636           1 :         assert(dev->for_each_count == 0);
    2637           1 :         spdk_io_device_unregister(dev->io_device, dev->unregister_cb);
    2638           1 : }
    2639             : 
    2640             : void
    2641        1571 : spdk_for_each_channel_continue(struct spdk_io_channel_iter *i, int status)
    2642             : {
    2643             :         struct spdk_thread *thread;
    2644             :         struct spdk_io_channel *ch;
    2645             :         struct io_device *dev;
    2646             :         int rc __attribute__((unused));
    2647             : 
    2648        1571 :         assert(i->cur_thread == spdk_get_thread());
    2649             : 
    2650        1571 :         i->status = status;
    2651             : 
    2652        1571 :         pthread_mutex_lock(&g_devlist_mutex);
    2653        1571 :         dev = i->dev;
    2654        1571 :         if (status) {
    2655          11 :                 goto end;
    2656             :         }
    2657             : 
    2658        1560 :         thread = TAILQ_NEXT(i->cur_thread, tailq);
    2659        3043 :         while (thread) {
    2660        1651 :                 ch = thread_get_io_channel(thread, dev);
    2661        1651 :                 if (ch != NULL) {
    2662         168 :                         i->cur_thread = thread;
    2663         168 :                         i->ch = ch;
    2664         168 :                         pthread_mutex_unlock(&g_devlist_mutex);
    2665         168 :                         rc = spdk_thread_send_msg(thread, _call_channel, i);
    2666         168 :                         assert(rc == 0);
    2667         168 :                         return;
    2668             :                 }
    2669        1483 :                 thread = TAILQ_NEXT(thread, tailq);
    2670             :         }
    2671             : 
    2672        1392 : end:
    2673        1403 :         dev->for_each_count--;
    2674        1403 :         i->ch = NULL;
    2675        1403 :         pthread_mutex_unlock(&g_devlist_mutex);
    2676             : 
    2677        1403 :         rc = spdk_thread_send_msg(i->orig_thread, _call_completion, i);
    2678        1403 :         assert(rc == 0);
    2679             : 
    2680        1403 :         pthread_mutex_lock(&g_devlist_mutex);
    2681        1403 :         if (dev->pending_unregister && dev->for_each_count == 0) {
    2682           1 :                 rc = spdk_thread_send_msg(dev->unregister_thread, __pending_unregister, dev);
    2683           1 :                 assert(rc == 0);
    2684             :         }
    2685        1403 :         pthread_mutex_unlock(&g_devlist_mutex);
    2686             : }
    2687             : 
    2688             : static void
    2689           0 : thread_interrupt_destroy(struct spdk_thread *thread)
    2690             : {
    2691           0 :         struct spdk_fd_group *fgrp = thread->fgrp;
    2692             : 
    2693           0 :         SPDK_INFOLOG(thread, "destroy fgrp for thread (%s)\n", thread->name);
    2694             : 
    2695           0 :         if (thread->msg_fd < 0) {
    2696           0 :                 return;
    2697             :         }
    2698             : 
    2699           0 :         spdk_fd_group_remove(fgrp, thread->msg_fd);
    2700           0 :         close(thread->msg_fd);
    2701           0 :         thread->msg_fd = -1;
    2702             : 
    2703           0 :         spdk_fd_group_destroy(fgrp);
    2704           0 :         thread->fgrp = NULL;
    2705             : }
    2706             : 
    2707             : #ifdef __linux__
    2708             : static int
    2709           0 : thread_interrupt_msg_process(void *arg)
    2710             : {
    2711           0 :         struct spdk_thread *thread = arg;
    2712             :         struct spdk_thread *orig_thread;
    2713             :         uint32_t msg_count;
    2714             :         spdk_msg_fn critical_msg;
    2715           0 :         int rc = 0;
    2716           0 :         uint64_t notify = 1;
    2717             : 
    2718           0 :         assert(spdk_interrupt_mode_is_enabled());
    2719             : 
    2720           0 :         orig_thread = spdk_get_thread();
    2721           0 :         spdk_set_thread(thread);
    2722             : 
    2723             :         /* There may be race between msg_acknowledge and another producer's msg_notify,
    2724             :          * so msg_acknowledge should be applied ahead. And then check for self's msg_notify.
    2725             :          * This can avoid msg notification missing.
    2726             :          */
    2727           0 :         rc = read(thread->msg_fd, &notify, sizeof(notify));
    2728           0 :         if (rc < 0 && errno != EAGAIN) {
    2729           0 :                 SPDK_ERRLOG("failed to acknowledge msg event: %s.\n", spdk_strerror(errno));
    2730             :         }
    2731             : 
    2732           0 :         critical_msg = thread->critical_msg;
    2733           0 :         if (spdk_unlikely(critical_msg != NULL)) {
    2734           0 :                 critical_msg(NULL);
    2735           0 :                 thread->critical_msg = NULL;
    2736           0 :                 rc = 1;
    2737             :         }
    2738             : 
    2739           0 :         msg_count = msg_queue_run_batch(thread, 0);
    2740           0 :         if (msg_count) {
    2741           0 :                 rc = 1;
    2742             :         }
    2743             : 
    2744           0 :         SPIN_ASSERT(thread->lock_count == 0, SPIN_ERR_HOLD_DURING_SWITCH);
    2745           0 :         if (spdk_unlikely(!thread->in_interrupt)) {
    2746             :                 /* The thread transitioned to poll mode in a msg during the above processing.
    2747             :                  * Clear msg_fd since thread messages will be polled directly in poll mode.
    2748             :                  */
    2749           0 :                 rc = read(thread->msg_fd, &notify, sizeof(notify));
    2750           0 :                 if (rc < 0 && errno != EAGAIN) {
    2751           0 :                         SPDK_ERRLOG("failed to acknowledge msg queue: %s.\n", spdk_strerror(errno));
    2752             :                 }
    2753             :         }
    2754             : 
    2755           0 :         spdk_set_thread(orig_thread);
    2756           0 :         return rc;
    2757             : }
    2758             : 
    2759             : static int
    2760           0 : thread_interrupt_create(struct spdk_thread *thread)
    2761             : {
    2762             :         int rc;
    2763             : 
    2764           0 :         SPDK_INFOLOG(thread, "Create fgrp for thread (%s)\n", thread->name);
    2765             : 
    2766           0 :         rc = spdk_fd_group_create(&thread->fgrp);
    2767           0 :         if (rc) {
    2768           0 :                 return rc;
    2769             :         }
    2770             : 
    2771           0 :         thread->msg_fd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
    2772           0 :         if (thread->msg_fd < 0) {
    2773           0 :                 rc = -errno;
    2774           0 :                 spdk_fd_group_destroy(thread->fgrp);
    2775           0 :                 thread->fgrp = NULL;
    2776             : 
    2777           0 :                 return rc;
    2778             :         }
    2779             : 
    2780           0 :         return SPDK_FD_GROUP_ADD(thread->fgrp, thread->msg_fd,
    2781             :                                  thread_interrupt_msg_process, thread);
    2782             : }
    2783             : #else
    2784             : static int
    2785             : thread_interrupt_create(struct spdk_thread *thread)
    2786             : {
    2787             :         return -ENOTSUP;
    2788             : }
    2789             : #endif
    2790             : 
    2791             : static int
    2792           0 : _interrupt_wrapper(void *ctx)
    2793             : {
    2794           0 :         struct spdk_interrupt *intr = ctx;
    2795             :         struct spdk_thread *orig_thread, *thread;
    2796             :         int rc;
    2797             : 
    2798           0 :         orig_thread = spdk_get_thread();
    2799           0 :         thread = intr->thread;
    2800             : 
    2801           0 :         spdk_set_thread(thread);
    2802             : 
    2803             :         SPDK_DTRACE_PROBE4(interrupt_fd_process, intr->name, intr->efd,
    2804             :                            intr->fn, intr->arg);
    2805             : 
    2806           0 :         rc = intr->fn(intr->arg);
    2807             : 
    2808           0 :         SPIN_ASSERT(thread->lock_count == 0, SPIN_ERR_HOLD_DURING_SWITCH);
    2809             : 
    2810           0 :         spdk_set_thread(orig_thread);
    2811             : 
    2812           0 :         return rc;
    2813             : }
    2814             : 
    2815             : struct spdk_interrupt *
    2816           0 : spdk_interrupt_register(int efd, spdk_interrupt_fn fn,
    2817             :                         void *arg, const char *name)
    2818             : {
    2819             :         struct spdk_thread *thread;
    2820             :         struct spdk_interrupt *intr;
    2821             :         int ret;
    2822             : 
    2823           0 :         thread = spdk_get_thread();
    2824           0 :         if (!thread) {
    2825           0 :                 assert(false);
    2826             :                 return NULL;
    2827             :         }
    2828             : 
    2829           0 :         if (spdk_unlikely(thread->state != SPDK_THREAD_STATE_RUNNING)) {
    2830           0 :                 SPDK_ERRLOG("thread %s is marked as exited\n", thread->name);
    2831           0 :                 return NULL;
    2832             :         }
    2833             : 
    2834           0 :         intr = calloc(1, sizeof(*intr));
    2835           0 :         if (intr == NULL) {
    2836           0 :                 SPDK_ERRLOG("Interrupt handler allocation failed\n");
    2837           0 :                 return NULL;
    2838             :         }
    2839             : 
    2840           0 :         if (name) {
    2841           0 :                 snprintf(intr->name, sizeof(intr->name), "%s", name);
    2842             :         } else {
    2843           0 :                 snprintf(intr->name, sizeof(intr->name), "%p", fn);
    2844             :         }
    2845             : 
    2846           0 :         intr->efd = efd;
    2847           0 :         intr->thread = thread;
    2848           0 :         intr->fn = fn;
    2849           0 :         intr->arg = arg;
    2850             : 
    2851           0 :         ret = spdk_fd_group_add(thread->fgrp, efd, _interrupt_wrapper, intr, intr->name);
    2852             : 
    2853           0 :         if (ret != 0) {
    2854           0 :                 SPDK_ERRLOG("thread %s: failed to add fd %d: %s\n",
    2855             :                             thread->name, efd, spdk_strerror(-ret));
    2856           0 :                 free(intr);
    2857           0 :                 return NULL;
    2858             :         }
    2859             : 
    2860           0 :         return intr;
    2861             : }
    2862             : 
    2863             : void
    2864           0 : spdk_interrupt_unregister(struct spdk_interrupt **pintr)
    2865             : {
    2866             :         struct spdk_thread *thread;
    2867             :         struct spdk_interrupt *intr;
    2868             : 
    2869           0 :         intr = *pintr;
    2870           0 :         if (intr == NULL) {
    2871           0 :                 return;
    2872             :         }
    2873             : 
    2874           0 :         *pintr = NULL;
    2875             : 
    2876           0 :         thread = spdk_get_thread();
    2877           0 :         if (!thread) {
    2878           0 :                 assert(false);
    2879             :                 return;
    2880             :         }
    2881             : 
    2882           0 :         if (intr->thread != thread) {
    2883           0 :                 wrong_thread(__func__, intr->name, intr->thread, thread);
    2884           0 :                 return;
    2885             :         }
    2886             : 
    2887           0 :         spdk_fd_group_remove(thread->fgrp, intr->efd);
    2888           0 :         free(intr);
    2889             : }
    2890             : 
    2891             : int
    2892           0 : spdk_interrupt_set_event_types(struct spdk_interrupt *intr,
    2893             :                                enum spdk_interrupt_event_types event_types)
    2894             : {
    2895             :         struct spdk_thread *thread;
    2896             : 
    2897           0 :         thread = spdk_get_thread();
    2898           0 :         if (!thread) {
    2899           0 :                 assert(false);
    2900             :                 return -EINVAL;
    2901             :         }
    2902             : 
    2903           0 :         if (intr->thread != thread) {
    2904           0 :                 wrong_thread(__func__, intr->name, intr->thread, thread);
    2905           0 :                 return -EINVAL;
    2906             :         }
    2907             : 
    2908           0 :         return spdk_fd_group_event_modify(thread->fgrp, intr->efd, event_types);
    2909             : }
    2910             : 
    2911             : int
    2912           0 : spdk_thread_get_interrupt_fd(struct spdk_thread *thread)
    2913             : {
    2914           0 :         return spdk_fd_group_get_fd(thread->fgrp);
    2915             : }
    2916             : 
    2917             : struct spdk_fd_group *
    2918           0 : spdk_thread_get_interrupt_fd_group(struct spdk_thread *thread)
    2919             : {
    2920           0 :         return thread->fgrp;
    2921             : }
    2922             : 
    2923             : static bool g_interrupt_mode = false;
    2924             : 
    2925             : int
    2926           0 : spdk_interrupt_mode_enable(void)
    2927             : {
    2928             :         /* It must be called once prior to initializing the threading library.
    2929             :          * g_spdk_msg_mempool will be valid if thread library is initialized.
    2930             :          */
    2931           0 :         if (g_spdk_msg_mempool) {
    2932           0 :                 SPDK_ERRLOG("Failed due to threading library is already initialized.\n");
    2933           0 :                 return -1;
    2934             :         }
    2935             : 
    2936             : #ifdef __linux__
    2937           0 :         SPDK_NOTICELOG("Set SPDK running in interrupt mode.\n");
    2938           0 :         g_interrupt_mode = true;
    2939           0 :         return 0;
    2940             : #else
    2941             :         SPDK_ERRLOG("SPDK interrupt mode supports only Linux platform now.\n");
    2942             :         g_interrupt_mode = false;
    2943             :         return -ENOTSUP;
    2944             : #endif
    2945             : }
    2946             : 
    2947             : bool
    2948       78653 : spdk_interrupt_mode_is_enabled(void)
    2949             : {
    2950       78653 :         return g_interrupt_mode;
    2951             : }
    2952             : 
    2953             : #define SSPIN_DEBUG_STACK_FRAMES 16
    2954             : 
    2955             : struct sspin_stack {
    2956             :         void *addrs[SSPIN_DEBUG_STACK_FRAMES];
    2957             :         uint32_t depth;
    2958             : };
    2959             : 
    2960             : struct spdk_spinlock_internal {
    2961             :         struct sspin_stack init_stack;
    2962             :         struct sspin_stack lock_stack;
    2963             :         struct sspin_stack unlock_stack;
    2964             : };
    2965             : 
    2966             : static void
    2967        1191 : sspin_init_internal(struct spdk_spinlock *sspin)
    2968             : {
    2969             : #ifdef DEBUG
    2970        1191 :         sspin->internal = calloc(1, sizeof(*sspin->internal));
    2971             : #endif
    2972        1191 : }
    2973             : 
    2974             : static void
    2975        1181 : sspin_fini_internal(struct spdk_spinlock *sspin)
    2976             : {
    2977             : #ifdef DEBUG
    2978        1181 :         free(sspin->internal);
    2979        1181 :         sspin->internal = NULL;
    2980             : #endif
    2981        1181 : }
    2982             : 
    2983             : #if defined(DEBUG) && defined(SPDK_HAVE_EXECINFO_H)
    2984             : #define SSPIN_GET_STACK(sspin, which) \
    2985             :         do { \
    2986             :                 if (sspin->internal != NULL) { \
    2987             :                         struct sspin_stack *stack = &sspin->internal->which ## _stack; \
    2988             :                         stack->depth = backtrace(stack->addrs, SPDK_COUNTOF(stack->addrs)); \
    2989             :                 } \
    2990             :         } while (0)
    2991             : #else
    2992             : #define SSPIN_GET_STACK(sspin, which) do { } while (0)
    2993             : #endif
    2994             : 
    2995             : static void
    2996          15 : sspin_stack_print(const char *title, const struct sspin_stack *sspin_stack)
    2997             : {
    2998             : #ifdef SPDK_HAVE_EXECINFO_H
    2999             :         char **stack;
    3000             :         size_t i;
    3001             : 
    3002             :         stack = backtrace_symbols(sspin_stack->addrs, sspin_stack->depth);
    3003             :         if (stack == NULL) {
    3004             :                 SPDK_ERRLOG("Out of memory while allocate stack for %s\n", title);
    3005             :                 return;
    3006             :         }
    3007             :         SPDK_ERRLOG("  %s:\n", title);
    3008             :         for (i = 0; i < sspin_stack->depth; i++) {
    3009             :                 /*
    3010             :                  * This does not print line numbers. In gdb, use something like "list *0x444b6b" or
    3011             :                  * "list *sspin_stack->addrs[0]".  Or more conveniently, load the spdk gdb macros
    3012             :                  * and use use "print *sspin" or "print sspin->internal.lock_stack".  See
    3013             :                  * gdb_macros.md in the docs directory for details.
    3014             :                  */
    3015             :                 SPDK_ERRLOG("    #%" PRIu64 ": %s\n", i, stack[i]);
    3016             :         }
    3017             :         free(stack);
    3018             : #endif /* SPDK_HAVE_EXECINFO_H */
    3019          15 : }
    3020             : 
    3021             : static void
    3022           5 : sspin_stacks_print(const struct spdk_spinlock *sspin)
    3023             : {
    3024           5 :         if (sspin->internal == NULL) {
    3025           0 :                 return;
    3026             :         }
    3027           5 :         SPDK_ERRLOG("spinlock %p\n", sspin);
    3028           5 :         sspin_stack_print("Lock initalized at", &sspin->internal->init_stack);
    3029           5 :         sspin_stack_print("Last locked at", &sspin->internal->lock_stack);
    3030           5 :         sspin_stack_print("Last unlocked at", &sspin->internal->unlock_stack);
    3031             : }
    3032             : 
    3033             : void
    3034        1191 : spdk_spin_init(struct spdk_spinlock *sspin)
    3035             : {
    3036             :         int rc;
    3037             : 
    3038        1191 :         memset(sspin, 0, sizeof(*sspin));
    3039        1191 :         rc = pthread_spin_init(&sspin->spinlock, PTHREAD_PROCESS_PRIVATE);
    3040        1191 :         SPIN_ASSERT_LOG_STACKS(rc == 0, SPIN_ERR_PTHREAD, sspin);
    3041        1191 :         sspin_init_internal(sspin);
    3042             :         SSPIN_GET_STACK(sspin, init);
    3043        1191 :         sspin->initialized = true;
    3044             : }
    3045             : 
    3046             : void
    3047        1182 : spdk_spin_destroy(struct spdk_spinlock *sspin)
    3048             : {
    3049             :         int rc;
    3050             : 
    3051        1182 :         SPIN_ASSERT_LOG_STACKS(!sspin->destroyed, SPIN_ERR_DESTROYED, sspin);
    3052        1182 :         SPIN_ASSERT_LOG_STACKS(sspin->initialized, SPIN_ERR_NOT_INITIALIZED, sspin);
    3053        1182 :         SPIN_ASSERT_LOG_STACKS(sspin->thread == NULL, SPIN_ERR_LOCK_HELD, sspin);
    3054             : 
    3055        1181 :         rc = pthread_spin_destroy(&sspin->spinlock);
    3056        1181 :         SPIN_ASSERT_LOG_STACKS(rc == 0, SPIN_ERR_PTHREAD, sspin);
    3057             : 
    3058        1181 :         sspin_fini_internal(sspin);
    3059        1181 :         sspin->initialized = false;
    3060        1181 :         sspin->destroyed = true;
    3061             : }
    3062             : 
    3063             : void
    3064       27012 : spdk_spin_lock(struct spdk_spinlock *sspin)
    3065             : {
    3066       27012 :         struct spdk_thread *thread = spdk_get_thread();
    3067             :         int rc;
    3068             : 
    3069       27012 :         SPIN_ASSERT_LOG_STACKS(!sspin->destroyed, SPIN_ERR_DESTROYED, sspin);
    3070       27012 :         SPIN_ASSERT_LOG_STACKS(sspin->initialized, SPIN_ERR_NOT_INITIALIZED, sspin);
    3071       27012 :         SPIN_ASSERT_LOG_STACKS(thread != NULL, SPIN_ERR_NOT_SPDK_THREAD, sspin);
    3072       27011 :         SPIN_ASSERT_LOG_STACKS(thread != sspin->thread, SPIN_ERR_DEADLOCK, sspin);
    3073             : 
    3074       27010 :         rc = pthread_spin_lock(&sspin->spinlock);
    3075       27010 :         SPIN_ASSERT_LOG_STACKS(rc == 0, SPIN_ERR_PTHREAD, sspin);
    3076             : 
    3077       27010 :         sspin->thread = thread;
    3078       27010 :         sspin->thread->lock_count++;
    3079             : 
    3080             :         SSPIN_GET_STACK(sspin, lock);
    3081             : }
    3082             : 
    3083             : void
    3084       27012 : spdk_spin_unlock(struct spdk_spinlock *sspin)
    3085             : {
    3086       27012 :         struct spdk_thread *thread = spdk_get_thread();
    3087             :         int rc;
    3088             : 
    3089       27012 :         SPIN_ASSERT_LOG_STACKS(!sspin->destroyed, SPIN_ERR_DESTROYED, sspin);
    3090       27012 :         SPIN_ASSERT_LOG_STACKS(sspin->initialized, SPIN_ERR_NOT_INITIALIZED, sspin);
    3091       27012 :         SPIN_ASSERT_LOG_STACKS(thread != NULL, SPIN_ERR_NOT_SPDK_THREAD, sspin);
    3092       27012 :         SPIN_ASSERT_LOG_STACKS(thread == sspin->thread, SPIN_ERR_WRONG_THREAD, sspin);
    3093             : 
    3094       27010 :         SPIN_ASSERT_LOG_STACKS(thread->lock_count > 0, SPIN_ERR_LOCK_COUNT, sspin);
    3095       27010 :         thread->lock_count--;
    3096       27010 :         sspin->thread = NULL;
    3097             : 
    3098             :         SSPIN_GET_STACK(sspin, unlock);
    3099             : 
    3100       27010 :         rc = pthread_spin_unlock(&sspin->spinlock);
    3101       27010 :         SPIN_ASSERT_LOG_STACKS(rc == 0, SPIN_ERR_PTHREAD, sspin);
    3102             : }
    3103             : 
    3104             : bool
    3105       30632 : spdk_spin_held(struct spdk_spinlock *sspin)
    3106             : {
    3107       30632 :         struct spdk_thread *thread = spdk_get_thread();
    3108             : 
    3109       30632 :         SPIN_ASSERT_RETURN(thread != NULL, SPIN_ERR_NOT_SPDK_THREAD, false);
    3110             : 
    3111       30631 :         return sspin->thread == thread;
    3112             : }
    3113             : 
    3114          38 : SPDK_LOG_REGISTER_COMPONENT(thread)

Generated by: LCOV version 1.15