+ timevaladd(&target, &period);
+
+ if (need_timer == FALSE || timevalcmp(&target, &min_target, <)) {
+ min_target = target;
+ need_timer = TRUE;
+ }
+ }
+ if (timevalcmp(&info->throttle_min_timer_deadline, &now, >)) {
+ if (timevalcmp(&info->throttle_min_timer_deadline, &min_target, >))
+ min_target = info->throttle_min_timer_deadline;
+ }
+
+ if (info->throttle_timer_active) {
+ if (thread_call_cancel(info->throttle_timer_call) == FALSE) {
+ /*
+ * couldn't kill the timer because it's already
+ * been dispatched, so don't try to start a new
+ * one... once we drop the lock, the timer will
+ * proceed and eventually re-run this function
+ */
+ need_timer = FALSE;
+ } else
+ info->throttle_timer_active = 0;
+ }
+ if (need_timer == TRUE) {
+ /*
+ * This is defined as an int (32-bit) rather than a 64-bit
+ * value because it would need a really big period in the
+ * order of ~500 days to overflow this. So, we let this be
+ * 32-bit which allows us to use the clock_interval_to_deadline()
+ * routine.
+ */
+ int target_msecs;
+
+ if (info->throttle_timer_ref == 0) {
+ /*
+ * take a reference for the timer
+ */
+ throttle_info_ref(info);
+
+ info->throttle_timer_ref = 1;
+ }
+ elapsed = min_target;
+ timevalsub(&elapsed, &now);
+ target_msecs = elapsed.tv_sec * 1000 + elapsed.tv_usec / 1000;
+
+ if (target_msecs <= 0) {
+ /*
+ * we may have computed a deadline slightly in the past
+ * due to various factors... if so, just set the timer
+ * to go off in the near future (we don't need to be precise)
+ */
+ target_msecs = 1;
+ }
+ clock_interval_to_deadline(target_msecs, 1000000, &deadline);
+
+ thread_call_enter_delayed(info->throttle_timer_call, deadline);
+ info->throttle_timer_active = 1;
+ }
+ }
+ return (throttle_level);
+}
+
+
+static void
+throttle_timer(struct _throttle_io_info_t *info)
+{
+ uthread_t ut, utlist;
+ struct timeval elapsed;
+ struct timeval now;
+ uint64_t elapsed_msecs;
+ int throttle_level;
+ int level;
+ int wake_level;
+ caddr_t wake_address = NULL;
+ boolean_t update_io_count = FALSE;
+ boolean_t need_wakeup = FALSE;
+ boolean_t need_release = FALSE;
+
+ ut = NULL;
+ lck_mtx_lock(&info->throttle_lock);
+
+ info->throttle_timer_active = 0;
+ microuptime(&now);
+
+ elapsed = now;
+ timevalsub(&elapsed, &info->throttle_start_IO_period_timestamp[THROTTLE_LEVEL_THROTTLED]);
+ elapsed_msecs = (uint64_t)elapsed.tv_sec * (uint64_t)1000 + (elapsed.tv_usec / 1000);
+
+ if (elapsed_msecs >= (uint64_t)info->throttle_io_periods[THROTTLE_LEVEL_THROTTLED]) {
+
+ wake_level = info->throttle_next_wake_level;
+
+ for (level = THROTTLE_LEVEL_START; level < THROTTLE_LEVEL_END; level++) {
+
+ elapsed = now;
+ timevalsub(&elapsed, &info->throttle_start_IO_period_timestamp[wake_level]);
+ elapsed_msecs = (uint64_t)elapsed.tv_sec * (uint64_t)1000 + (elapsed.tv_usec / 1000);
+
+ if (elapsed_msecs >= (uint64_t)info->throttle_io_periods[wake_level] && !TAILQ_EMPTY(&info->throttle_uthlist[wake_level])) {
+ /*
+ * we're closing out the current IO period...
+ * if we have a waiting thread, wake it up
+ * after we have reset the I/O window info
+ */
+ need_wakeup = TRUE;
+ update_io_count = TRUE;
+
+ info->throttle_next_wake_level = wake_level - 1;
+
+ if (info->throttle_next_wake_level == THROTTLE_LEVEL_START)
+ info->throttle_next_wake_level = THROTTLE_LEVEL_END;
+
+ break;
+ }
+ wake_level--;
+
+ if (wake_level == THROTTLE_LEVEL_START)
+ wake_level = THROTTLE_LEVEL_END;
+ }
+ }
+ if (need_wakeup == TRUE) {
+ if (!TAILQ_EMPTY(&info->throttle_uthlist[wake_level])) {
+
+ ut = (uthread_t)TAILQ_FIRST(&info->throttle_uthlist[wake_level]);
+ TAILQ_REMOVE(&info->throttle_uthlist[wake_level], ut, uu_throttlelist);
+ ut->uu_on_throttlelist = THROTTLE_LEVEL_NONE;
+ ut->uu_is_throttled = FALSE;
+
+ wake_address = (caddr_t)&ut->uu_on_throttlelist;
+ }
+ } else
+ wake_level = THROTTLE_LEVEL_START;
+
+ throttle_level = throttle_timer_start(info, update_io_count, wake_level);
+
+ if (wake_address != NULL)
+ wakeup(wake_address);
+
+ for (level = THROTTLE_LEVEL_THROTTLED; level <= throttle_level; level++) {
+
+ TAILQ_FOREACH_SAFE(ut, &info->throttle_uthlist[level], uu_throttlelist, utlist) {
+
+ TAILQ_REMOVE(&info->throttle_uthlist[level], ut, uu_throttlelist);
+ ut->uu_on_throttlelist = THROTTLE_LEVEL_NONE;
+ ut->uu_is_throttled = FALSE;
+
+ wakeup(&ut->uu_on_throttlelist);
+ }
+ }
+ if (info->throttle_timer_active == 0 && info->throttle_timer_ref) {
+ info->throttle_timer_ref = 0;
+ need_release = TRUE;
+ }
+ lck_mtx_unlock(&info->throttle_lock);
+
+ if (need_release == TRUE)
+ throttle_info_rel(info);
+}
+
+
+static int
+throttle_add_to_list(struct _throttle_io_info_t *info, uthread_t ut, int mylevel, boolean_t insert_tail)
+{
+ boolean_t start_timer = FALSE;
+ int level = THROTTLE_LEVEL_START;
+
+ if (TAILQ_EMPTY(&info->throttle_uthlist[mylevel])) {
+ info->throttle_start_IO_period_timestamp[mylevel] = info->throttle_last_IO_timestamp[mylevel];
+ start_timer = TRUE;
+ }
+
+ if (insert_tail == TRUE)
+ TAILQ_INSERT_TAIL(&info->throttle_uthlist[mylevel], ut, uu_throttlelist);
+ else
+ TAILQ_INSERT_HEAD(&info->throttle_uthlist[mylevel], ut, uu_throttlelist);
+
+ ut->uu_on_throttlelist = mylevel;
+
+ if (start_timer == TRUE) {
+ /* we may need to start or rearm the timer */
+ level = throttle_timer_start(info, FALSE, THROTTLE_LEVEL_START);
+
+ if (level == THROTTLE_LEVEL_END) {
+ if (ut->uu_on_throttlelist >= THROTTLE_LEVEL_THROTTLED) {
+ TAILQ_REMOVE(&info->throttle_uthlist[ut->uu_on_throttlelist], ut, uu_throttlelist);
+
+ ut->uu_on_throttlelist = THROTTLE_LEVEL_NONE;
+ }
+ }
+ }
+ return (level);
+}
+
+static void
+throttle_init_throttle_window(void)
+{
+ int throttle_window_size;
+
+ /*
+ * The hierarchy of throttle window values is as follows:
+ * - Global defaults
+ * - Device tree properties
+ * - Boot-args
+ * All values are specified in msecs.
+ */
+
+ /* Override global values with device-tree properties */
+ if (PE_get_default("kern.io_throttle_window_tier1", &throttle_window_size, sizeof(throttle_window_size)))
+ throttle_windows_msecs[THROTTLE_LEVEL_TIER1] = throttle_window_size;
+
+ if (PE_get_default("kern.io_throttle_window_tier2", &throttle_window_size, sizeof(throttle_window_size)))
+ throttle_windows_msecs[THROTTLE_LEVEL_TIER2] = throttle_window_size;
+
+ if (PE_get_default("kern.io_throttle_window_tier3", &throttle_window_size, sizeof(throttle_window_size)))
+ throttle_windows_msecs[THROTTLE_LEVEL_TIER3] = throttle_window_size;
+
+ /* Override with boot-args */
+ if (PE_parse_boot_argn("io_throttle_window_tier1", &throttle_window_size, sizeof(throttle_window_size)))
+ throttle_windows_msecs[THROTTLE_LEVEL_TIER1] = throttle_window_size;
+
+ if (PE_parse_boot_argn("io_throttle_window_tier2", &throttle_window_size, sizeof(throttle_window_size)))
+ throttle_windows_msecs[THROTTLE_LEVEL_TIER2] = throttle_window_size;
+
+ if (PE_parse_boot_argn("io_throttle_window_tier3", &throttle_window_size, sizeof(throttle_window_size)))
+ throttle_windows_msecs[THROTTLE_LEVEL_TIER3] = throttle_window_size;
+}
+
+static void
+throttle_init_throttle_period(struct _throttle_io_info_t *info, boolean_t isssd)
+{
+ int throttle_period_size;
+
+ /*
+ * The hierarchy of throttle period values is as follows:
+ * - Global defaults
+ * - Device tree properties
+ * - Boot-args
+ * All values are specified in msecs.
+ */
+
+ /* Assign global defaults */
+ if ((isssd == TRUE) && (info->throttle_is_fusion_with_priority == 0))
+ info->throttle_io_periods = &throttle_io_period_ssd_msecs[0];
+ else
+ info->throttle_io_periods = &throttle_io_period_msecs[0];
+
+ /* Override global values with device-tree properties */
+ if (PE_get_default("kern.io_throttle_period_tier1", &throttle_period_size, sizeof(throttle_period_size)))
+ info->throttle_io_periods[THROTTLE_LEVEL_TIER1] = throttle_period_size;
+
+ if (PE_get_default("kern.io_throttle_period_tier2", &throttle_period_size, sizeof(throttle_period_size)))
+ info->throttle_io_periods[THROTTLE_LEVEL_TIER2] = throttle_period_size;
+
+ if (PE_get_default("kern.io_throttle_period_tier3", &throttle_period_size, sizeof(throttle_period_size)))
+ info->throttle_io_periods[THROTTLE_LEVEL_TIER3] = throttle_period_size;
+
+ /* Override with boot-args */
+ if (PE_parse_boot_argn("io_throttle_period_tier1", &throttle_period_size, sizeof(throttle_period_size)))
+ info->throttle_io_periods[THROTTLE_LEVEL_TIER1] = throttle_period_size;
+
+ if (PE_parse_boot_argn("io_throttle_period_tier2", &throttle_period_size, sizeof(throttle_period_size)))
+ info->throttle_io_periods[THROTTLE_LEVEL_TIER2] = throttle_period_size;
+
+ if (PE_parse_boot_argn("io_throttle_period_tier3", &throttle_period_size, sizeof(throttle_period_size)))
+ info->throttle_io_periods[THROTTLE_LEVEL_TIER3] = throttle_period_size;
+
+}
+
+#if CONFIG_IOSCHED
+extern void vm_io_reprioritize_init(void);
+int iosched_enabled = 1;
+#endif
+
+void
+throttle_init(void)
+{
+ struct _throttle_io_info_t *info;
+ int i;
+ int level;
+#if CONFIG_IOSCHED
+ int iosched;
+#endif
+ /*
+ * allocate lock group attribute and group
+ */
+ throttle_lock_grp_attr = lck_grp_attr_alloc_init();
+ throttle_lock_grp = lck_grp_alloc_init("throttle I/O", throttle_lock_grp_attr);
+
+ /* Update throttle parameters based on device tree configuration */
+ throttle_init_throttle_window();
+
+ /*
+ * allocate the lock attribute
+ */
+ throttle_lock_attr = lck_attr_alloc_init();
+
+ for (i = 0; i < LOWPRI_MAX_NUM_DEV; i++) {
+ info = &_throttle_io_info[i];
+
+ lck_mtx_init(&info->throttle_lock, throttle_lock_grp, throttle_lock_attr);
+ info->throttle_timer_call = thread_call_allocate((thread_call_func_t)throttle_timer, (thread_call_param_t)info);
+
+ for (level = 0; level <= THROTTLE_LEVEL_END; level++) {
+ TAILQ_INIT(&info->throttle_uthlist[level]);
+ info->throttle_last_IO_pid[level] = 0;
+ info->throttle_inflight_count[level] = 0;
+ }
+ info->throttle_next_wake_level = THROTTLE_LEVEL_END;
+ info->throttle_disabled = 0;
+ info->throttle_is_fusion_with_priority = 0;
+ }
+#if CONFIG_IOSCHED
+ if (PE_parse_boot_argn("iosched", &iosched, sizeof(iosched))) {
+ iosched_enabled = iosched;
+ }
+ if (iosched_enabled) {
+ /* Initialize I/O Reprioritization mechanism */
+ vm_io_reprioritize_init();
+ }
+#endif
+}
+
+void
+sys_override_io_throttle(int flag)
+{
+ if (flag == THROTTLE_IO_ENABLE)
+ lowpri_throttle_enabled = 1;
+
+ if (flag == THROTTLE_IO_DISABLE)
+ lowpri_throttle_enabled = 0;
+}
+
+int rethrottle_wakeups = 0;
+
+/*
+ * the uu_rethrottle_lock is used to synchronize this function
+ * with "throttle_lowpri_io" which is where a throttled thread
+ * will block... that function will grab this lock before beginning
+ * it's decision making process concerning the need to block, and
+ * hold it through the assert_wait. When that thread is awakened
+ * for any reason (timer or rethrottle), it will reacquire the
+ * uu_rethrottle_lock before determining if it really is ok for
+ * it to now run. This is the point at which the thread could
+ * enter a different throttling queue and reblock or return from
+ * the throttle w/o having waited out it's entire throttle if
+ * the rethrottle has now moved it out of any currently
+ * active throttle window.
+ *
+ *
+ * NOTES:
+ * 1 - This may be called with the task lock held.
+ * 2 - This may be called with preemption and interrupts disabled
+ * in the kqueue wakeup path so we can't take the throttle_lock which is a mutex
+ * 3 - This cannot safely dereference uu_throttle_info, as it may
+ * get deallocated out from under us
+ */
+
+void
+rethrottle_thread(uthread_t ut)
+{
+ /*
+ * If uthread doesn't have throttle state, then there's no chance
+ * of it needing a rethrottle.
+ */
+ if (ut->uu_throttle_info == NULL)
+ return;
+
+ boolean_t s = ml_set_interrupts_enabled(FALSE);
+ lck_spin_lock(&ut->uu_rethrottle_lock);
+
+ if (ut->uu_is_throttled == FALSE)
+ ut->uu_was_rethrottled = TRUE;
+ else {
+ int my_new_level = throttle_get_thread_throttle_level(ut);
+
+ if (my_new_level != ut->uu_on_throttlelist) {
+ /*
+ * ut is currently blocked (as indicated by
+ * ut->uu_is_throttled == TRUE)
+ * and we're changing it's throttle level, so
+ * we need to wake it up.
+ */
+ ut->uu_is_throttled = FALSE;
+ wakeup(&ut->uu_on_throttlelist);
+
+ rethrottle_wakeups++;
+ KERNEL_DEBUG_CONSTANT((FSDBG_CODE(DBG_FSRW, 102)), thread_tid(ut->uu_thread), ut->uu_on_throttlelist, my_new_level, 0, 0);
+ }
+ }
+ lck_spin_unlock(&ut->uu_rethrottle_lock);
+ ml_set_interrupts_enabled(s);
+}
+
+
+/*
+ * KPI routine
+ *
+ * Create and take a reference on a throttle info structure and return a
+ * pointer for the file system to use when calling throttle_info_update.
+ * Calling file system must have a matching release for every create.
+ */
+void *
+throttle_info_create(void)
+{
+ struct _throttle_io_info_t *info;
+ int level;
+
+ MALLOC(info, struct _throttle_io_info_t *, sizeof(*info), M_TEMP, M_ZERO | M_WAITOK);
+ /* Should never happen but just in case */
+ if (info == NULL)
+ return NULL;
+ /* Mark that this one was allocated and needs to be freed */
+ DEBUG_ALLOC_THROTTLE_INFO("Creating info = %p\n", info, info );
+ info->throttle_alloc = TRUE;
+
+ lck_mtx_init(&info->throttle_lock, throttle_lock_grp, throttle_lock_attr);
+ info->throttle_timer_call = thread_call_allocate((thread_call_func_t)throttle_timer, (thread_call_param_t)info);
+
+ for (level = 0; level <= THROTTLE_LEVEL_END; level++) {
+ TAILQ_INIT(&info->throttle_uthlist[level]);
+ }
+ info->throttle_next_wake_level = THROTTLE_LEVEL_END;
+
+ /* Take a reference */
+ OSIncrementAtomic(&info->throttle_refcnt);
+ return info;
+}
+
+/*
+ * KPI routine
+ *
+ * Release the throttle info pointer if all the reference are gone. Should be
+ * called to release reference taken by throttle_info_create
+ */
+void
+throttle_info_release(void *throttle_info)
+{
+ DEBUG_ALLOC_THROTTLE_INFO("Releaseing info = %p\n",
+ (struct _throttle_io_info_t *)throttle_info,
+ (struct _throttle_io_info_t *)throttle_info);
+ if (throttle_info) /* Just to be careful */
+ throttle_info_rel(throttle_info);
+}
+
+/*
+ * KPI routine
+ *
+ * File Systems that create an info structure, need to call this routine in
+ * their mount routine (used by cluster code). File Systems that call this in
+ * their mount routines must call throttle_info_mount_rel in their unmount
+ * routines.
+ */
+void
+throttle_info_mount_ref(mount_t mp, void *throttle_info)
+{
+ if ((throttle_info == NULL) || (mp == NULL))
+ return;
+ throttle_info_ref(throttle_info);
+
+ /*
+ * We already have a reference release it before adding the new one
+ */
+ if (mp->mnt_throttle_info)
+ throttle_info_rel(mp->mnt_throttle_info);
+ mp->mnt_throttle_info = throttle_info;
+}
+
+/*
+ * Private KPI routine
+ *
+ * return a handle for accessing throttle_info given a throttle_mask. The
+ * handle must be released by throttle_info_rel_by_mask
+ */
+int
+throttle_info_ref_by_mask(uint64_t throttle_mask, throttle_info_handle_t *throttle_info_handle)
+{
+ int dev_index;
+ struct _throttle_io_info_t *info;
+
+ if (throttle_info_handle == NULL)
+ return EINVAL;
+
+ dev_index = num_trailing_0(throttle_mask);
+ info = &_throttle_io_info[dev_index];
+ throttle_info_ref(info);
+ *(struct _throttle_io_info_t**)throttle_info_handle = info;
+
+ return 0;
+}
+
+/*
+ * Private KPI routine
+ *
+ * release the handle obtained by throttle_info_ref_by_mask
+ */
+void
+throttle_info_rel_by_mask(throttle_info_handle_t throttle_info_handle)
+{
+ /*
+ * for now the handle is just a pointer to _throttle_io_info_t
+ */
+ throttle_info_rel((struct _throttle_io_info_t*)throttle_info_handle);
+}
+
+/*
+ * KPI routine
+ *
+ * File Systems that throttle_info_mount_ref, must call this routine in their
+ * umount routine.
+ */
+void
+throttle_info_mount_rel(mount_t mp)
+{
+ if (mp->mnt_throttle_info)
+ throttle_info_rel(mp->mnt_throttle_info);
+ mp->mnt_throttle_info = NULL;
+}
+
+/*
+ * Reset throttling periods for the given mount point
+ *
+ * private interface used by disk conditioner to reset
+ * throttling periods when 'is_ssd' status changes
+ */
+void
+throttle_info_mount_reset_period(mount_t mp, int isssd)
+{
+ struct _throttle_io_info_t *info;
+
+ if (mp == NULL)
+ info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1];
+ else if (mp->mnt_throttle_info == NULL)
+ info = &_throttle_io_info[mp->mnt_devbsdunit];
+ else
+ info = mp->mnt_throttle_info;
+
+ throttle_init_throttle_period(info, isssd);
+}
+
+void
+throttle_info_get_last_io_time(mount_t mp, struct timeval *tv)
+{
+ struct _throttle_io_info_t *info;
+
+ if (mp == NULL)
+ info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1];
+ else if (mp->mnt_throttle_info == NULL)
+ info = &_throttle_io_info[mp->mnt_devbsdunit];
+ else
+ info = mp->mnt_throttle_info;
+
+ *tv = info->throttle_last_write_timestamp;
+}
+
+void
+update_last_io_time(mount_t mp)
+{
+ struct _throttle_io_info_t *info;
+
+ if (mp == NULL)
+ info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1];
+ else if (mp->mnt_throttle_info == NULL)
+ info = &_throttle_io_info[mp->mnt_devbsdunit];
+ else
+ info = mp->mnt_throttle_info;
+
+ microuptime(&info->throttle_last_write_timestamp);
+ if (mp != NULL)
+ mp->mnt_last_write_completed_timestamp = info->throttle_last_write_timestamp;
+}
+
+int
+throttle_get_io_policy(uthread_t *ut)
+{
+ if (ut != NULL)
+ *ut = get_bsdthread_info(current_thread());
+
+ return (proc_get_effective_thread_policy(current_thread(), TASK_POLICY_IO));
+}
+
+int
+throttle_get_passive_io_policy(uthread_t *ut)
+{
+ if (ut != NULL)
+ *ut = get_bsdthread_info(current_thread());
+
+ return (proc_get_effective_thread_policy(current_thread(), TASK_POLICY_PASSIVE_IO));
+}
+
+
+static int
+throttle_get_thread_throttle_level(uthread_t ut)
+{
+ uthread_t *ut_p = (ut == NULL) ? &ut : NULL;
+ int io_tier = throttle_get_io_policy(ut_p);
+
+ return throttle_get_thread_throttle_level_internal(ut, io_tier);
+}
+
+/*
+ * Return a throttle level given an existing I/O tier (such as returned by throttle_get_io_policy)
+ */
+static int
+throttle_get_thread_throttle_level_internal(uthread_t ut, int io_tier) {
+ int thread_throttle_level = io_tier;
+ int user_idle_level;
+
+ assert(ut != NULL);
+
+ /* Bootcache misses should always be throttled */
+ if (ut->uu_throttle_bc == TRUE)
+ thread_throttle_level = THROTTLE_LEVEL_TIER3;
+
+ /*
+ * Issue tier3 I/O as tier2 when the user is idle
+ * to allow maintenance tasks to make more progress.
+ *
+ * Assume any positive idle level is enough... for now it's
+ * only ever 0 or 128 but this is not defined anywhere.
+ */
+ if (thread_throttle_level >= THROTTLE_LEVEL_TIER3) {
+ user_idle_level = timer_get_user_idle_level();
+ if (user_idle_level > 0) {
+ thread_throttle_level--;
+ }
+ }
+
+ return (thread_throttle_level);
+}
+
+/*
+ * I/O will be throttled if either of the following are true:
+ * - Higher tiers have in-flight I/O
+ * - The time delta since the last start/completion of a higher tier is within the throttle window interval
+ *
+ * In-flight I/O is bookended by throttle_info_update_internal/throttle_info_end_io_internal
+ */
+static int
+throttle_io_will_be_throttled_internal(void * throttle_info, int * mylevel, int * throttling_level)
+{
+ struct _throttle_io_info_t *info = throttle_info;
+ struct timeval elapsed;
+ struct timeval now;
+ uint64_t elapsed_msecs;
+ int thread_throttle_level;
+ int throttle_level;
+
+ if ((thread_throttle_level = throttle_get_thread_throttle_level(NULL)) < THROTTLE_LEVEL_THROTTLED)
+ return (THROTTLE_DISENGAGED);
+
+ microuptime(&now);
+
+ for (throttle_level = THROTTLE_LEVEL_START; throttle_level < thread_throttle_level; throttle_level++) {
+ if (info->throttle_inflight_count[throttle_level]) {
+ break;
+ }
+ elapsed = now;
+ timevalsub(&elapsed, &info->throttle_window_start_timestamp[throttle_level]);
+ elapsed_msecs = (uint64_t)elapsed.tv_sec * (uint64_t)1000 + (elapsed.tv_usec / 1000);
+
+ if (elapsed_msecs < (uint64_t)throttle_windows_msecs[thread_throttle_level])
+ break;
+ }
+ if (throttle_level >= thread_throttle_level) {
+ /*
+ * we're beyond all of the throttle windows
+ * that affect the throttle level of this thread,
+ * so go ahead and treat as normal I/O
+ */
+ return (THROTTLE_DISENGAGED);
+ }
+ if (mylevel)
+ *mylevel = thread_throttle_level;
+ if (throttling_level)
+ *throttling_level = throttle_level;
+
+ if (info->throttle_io_count != info->throttle_io_count_begin) {
+ /*
+ * we've already issued at least one throttleable I/O
+ * in the current I/O window, so avoid issuing another one
+ */
+ return (THROTTLE_NOW);
+ }
+ /*
+ * we're in the throttle window, so
+ * cut the I/O size back
+ */
+ return (THROTTLE_ENGAGED);
+}
+
+/*
+ * If we have a mount point and it has a throttle info pointer then
+ * use it to do the check, otherwise use the device unit number to find
+ * the correct throttle info array element.
+ */
+int
+throttle_io_will_be_throttled(__unused int lowpri_window_msecs, mount_t mp)
+{
+ struct _throttle_io_info_t *info;
+
+ /*
+ * Should we just return zero if no mount point
+ */
+ if (mp == NULL)
+ info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1];
+ else if (mp->mnt_throttle_info == NULL)
+ info = &_throttle_io_info[mp->mnt_devbsdunit];
+ else
+ info = mp->mnt_throttle_info;
+
+ if (info->throttle_is_fusion_with_priority) {
+ uthread_t ut = get_bsdthread_info(current_thread());
+ if (ut->uu_lowpri_window == 0)
+ return (THROTTLE_DISENGAGED);
+ }
+
+ if (info->throttle_disabled)
+ return (THROTTLE_DISENGAGED);
+ else
+ return throttle_io_will_be_throttled_internal(info, NULL, NULL);
+}
+
+/*
+ * Routine to increment I/O throttling counters maintained in the proc
+ */
+
+static void
+throttle_update_proc_stats(pid_t throttling_pid, int count)
+{
+ proc_t throttling_proc;
+ proc_t throttled_proc = current_proc();
+
+ /* The throttled_proc is always the current proc; so we are not concerned with refs */
+ OSAddAtomic64(count, &(throttled_proc->was_throttled));
+
+ /* The throttling pid might have exited by now */
+ throttling_proc = proc_find(throttling_pid);
+ if (throttling_proc != PROC_NULL) {
+ OSAddAtomic64(count, &(throttling_proc->did_throttle));
+ proc_rele(throttling_proc);
+ }
+}
+
+/*
+ * Block until woken up by the throttle timer or by a rethrottle call.
+ * As long as we hold the throttle_lock while querying the throttle tier, we're
+ * safe against seeing an old throttle tier after a rethrottle.
+ */
+uint32_t
+throttle_lowpri_io(int sleep_amount)
+{
+ uthread_t ut;
+ struct _throttle_io_info_t *info;
+ int throttle_type = 0;
+ int mylevel = 0;
+ int throttling_level = THROTTLE_LEVEL_NONE;
+ int sleep_cnt = 0;
+ uint32_t throttle_io_period_num = 0;
+ boolean_t insert_tail = TRUE;
+ boolean_t s;
+
+ ut = get_bsdthread_info(current_thread());
+
+ if (ut->uu_lowpri_window == 0)
+ return (0);
+
+ info = ut->uu_throttle_info;
+
+ if (info == NULL) {
+ ut->uu_throttle_bc = FALSE;
+ ut->uu_lowpri_window = 0;
+ return (0);
+ }
+ lck_mtx_lock(&info->throttle_lock);
+ assert(ut->uu_on_throttlelist < THROTTLE_LEVEL_THROTTLED);
+
+ if (sleep_amount == 0)
+ goto done;
+
+ if (sleep_amount == 1 && ut->uu_throttle_bc == FALSE)
+ sleep_amount = 0;
+
+ throttle_io_period_num = info->throttle_io_period_num;
+
+ ut->uu_was_rethrottled = FALSE;
+
+ while ( (throttle_type = throttle_io_will_be_throttled_internal(info, &mylevel, &throttling_level)) ) {
+
+ if (throttle_type == THROTTLE_ENGAGED) {
+ if (sleep_amount == 0)
+ break;
+ if (info->throttle_io_period_num < throttle_io_period_num)
+ break;
+ if ((info->throttle_io_period_num - throttle_io_period_num) >= (uint32_t)sleep_amount)
+ break;
+ }
+ /*
+ * keep the same position in the list if "rethrottle_thread" changes our throttle level and
+ * then puts us back to the original level before we get a chance to run
+ */
+ if (ut->uu_on_throttlelist >= THROTTLE_LEVEL_THROTTLED && ut->uu_on_throttlelist != mylevel) {
+ /*
+ * must have been awakened via "rethrottle_thread" (the timer pulls us off the list)
+ * and we've changed our throttling level, so pull ourselves off of the appropriate list
+ * and make sure we get put on the tail of the new list since we're starting anew w/r to
+ * the throttling engine
+ */
+ TAILQ_REMOVE(&info->throttle_uthlist[ut->uu_on_throttlelist], ut, uu_throttlelist);
+ ut->uu_on_throttlelist = THROTTLE_LEVEL_NONE;
+ insert_tail = TRUE;
+ }
+ if (ut->uu_on_throttlelist < THROTTLE_LEVEL_THROTTLED) {
+ if (throttle_add_to_list(info, ut, mylevel, insert_tail) == THROTTLE_LEVEL_END)
+ goto done;
+ }
+ assert(throttling_level >= THROTTLE_LEVEL_START && throttling_level <= THROTTLE_LEVEL_END);
+
+ s = ml_set_interrupts_enabled(FALSE);
+ lck_spin_lock(&ut->uu_rethrottle_lock);
+
+ /*
+ * this is the critical section w/r to our interaction
+ * with "rethrottle_thread"
+ */
+ if (ut->uu_was_rethrottled == TRUE) {
+
+ lck_spin_unlock(&ut->uu_rethrottle_lock);
+ ml_set_interrupts_enabled(s);
+ lck_mtx_yield(&info->throttle_lock);
+
+ KERNEL_DEBUG_CONSTANT((FSDBG_CODE(DBG_FSRW, 103)), thread_tid(ut->uu_thread), ut->uu_on_throttlelist, 0, 0, 0);
+
+ ut->uu_was_rethrottled = FALSE;
+ continue;
+ }
+ KERNEL_DEBUG_CONSTANT((FSDBG_CODE(DBG_THROTTLE, PROCESS_THROTTLED)) | DBG_FUNC_NONE,
+ info->throttle_last_IO_pid[throttling_level], throttling_level, proc_selfpid(), mylevel, 0);
+
+ if (sleep_cnt == 0) {
+ KERNEL_DEBUG_CONSTANT((FSDBG_CODE(DBG_FSRW, 97)) | DBG_FUNC_START,
+ throttle_windows_msecs[mylevel], info->throttle_io_periods[mylevel], info->throttle_io_count, 0, 0);
+ throttled_count[mylevel]++;
+ }
+ ut->uu_wmesg = "throttle_lowpri_io";
+
+ assert_wait((caddr_t)&ut->uu_on_throttlelist, THREAD_UNINT);
+
+ ut->uu_is_throttled = TRUE;
+ lck_spin_unlock(&ut->uu_rethrottle_lock);
+ ml_set_interrupts_enabled(s);
+
+ lck_mtx_unlock(&info->throttle_lock);
+
+ thread_block(THREAD_CONTINUE_NULL);
+
+ ut->uu_wmesg = NULL;
+
+ ut->uu_is_throttled = FALSE;
+ ut->uu_was_rethrottled = FALSE;
+
+ lck_mtx_lock(&info->throttle_lock);
+
+ sleep_cnt++;
+
+ if (sleep_amount == 0)
+ insert_tail = FALSE;
+ else if (info->throttle_io_period_num < throttle_io_period_num ||
+ (info->throttle_io_period_num - throttle_io_period_num) >= (uint32_t)sleep_amount) {
+ insert_tail = FALSE;
+ sleep_amount = 0;
+ }
+ }
+done:
+ if (ut->uu_on_throttlelist >= THROTTLE_LEVEL_THROTTLED) {
+ TAILQ_REMOVE(&info->throttle_uthlist[ut->uu_on_throttlelist], ut, uu_throttlelist);
+ ut->uu_on_throttlelist = THROTTLE_LEVEL_NONE;
+ }
+ lck_mtx_unlock(&info->throttle_lock);
+
+ if (sleep_cnt) {
+ KERNEL_DEBUG_CONSTANT((FSDBG_CODE(DBG_FSRW, 97)) | DBG_FUNC_END,
+ throttle_windows_msecs[mylevel], info->throttle_io_periods[mylevel], info->throttle_io_count, 0, 0);
+ /*
+ * We update the stats for the last pid which opened a throttle window for the throttled thread.
+ * This might not be completely accurate since the multiple throttles seen by the lower tier pid
+ * might have been caused by various higher prio pids. However, updating these stats accurately
+ * means doing a proc_find while holding the throttle lock which leads to deadlock.
+ */
+ throttle_update_proc_stats(info->throttle_last_IO_pid[throttling_level], sleep_cnt);
+ }
+
+ ut->uu_throttle_info = NULL;
+ ut->uu_throttle_bc = FALSE;
+ ut->uu_lowpri_window = 0;
+
+ throttle_info_rel(info);
+
+ return (sleep_cnt);
+}
+
+/*
+ * KPI routine
+ *
+ * set a kernel thread's IO policy. policy can be:
+ * IOPOL_NORMAL, IOPOL_THROTTLE, IOPOL_PASSIVE, IOPOL_UTILITY, IOPOL_STANDARD
+ *
+ * explanations about these policies are in the man page of setiopolicy_np
+ */
+void throttle_set_thread_io_policy(int policy)
+{
+ proc_set_thread_policy(current_thread(), TASK_POLICY_INTERNAL, TASK_POLICY_IOPOL, policy);
+}
+
+int throttle_get_thread_effective_io_policy()
+{
+ return proc_get_effective_thread_policy(current_thread(), TASK_POLICY_IO);
+}
+
+void throttle_info_reset_window(uthread_t ut)
+{
+ struct _throttle_io_info_t *info;
+
+ if (ut == NULL)
+ ut = get_bsdthread_info(current_thread());
+
+ if ( (info = ut->uu_throttle_info) ) {
+ throttle_info_rel(info);
+
+ ut->uu_throttle_info = NULL;
+ ut->uu_lowpri_window = 0;
+ ut->uu_throttle_bc = FALSE;
+ }
+}
+
+static
+void throttle_info_set_initial_window(uthread_t ut, struct _throttle_io_info_t *info, boolean_t BC_throttle, boolean_t isssd)
+{
+ if (lowpri_throttle_enabled == 0 || info->throttle_disabled)
+ return;
+
+ if (info->throttle_io_periods == 0) {
+ throttle_init_throttle_period(info, isssd);
+ }
+ if (ut->uu_throttle_info == NULL) {
+
+ ut->uu_throttle_info = info;
+ throttle_info_ref(info);
+ DEBUG_ALLOC_THROTTLE_INFO("updating info = %p\n", info, info );
+
+ ut->uu_lowpri_window = 1;
+ ut->uu_throttle_bc = BC_throttle;
+ }
+}
+
+/*
+ * Update inflight IO count and throttling window
+ * Should be called when an IO is done
+ *
+ * Only affects IO that was sent through spec_strategy
+ */
+void throttle_info_end_io(buf_t bp) {
+ mount_t mp;
+ struct bufattr *bap;
+ struct _throttle_io_info_t *info;
+ int io_tier;
+
+ bap = &bp->b_attr;
+ if (!ISSET(bap->ba_flags, BA_STRATEGY_TRACKED_IO)) {
+ return;
+ }
+ CLR(bap->ba_flags, BA_STRATEGY_TRACKED_IO);
+
+ mp = buf_vnode(bp)->v_mount;
+ if (mp != NULL) {
+ info = &_throttle_io_info[mp->mnt_devbsdunit];
+ } else {
+ info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1];
+ }
+
+ io_tier = GET_BUFATTR_IO_TIER(bap);
+ if (ISSET(bap->ba_flags, BA_IO_TIER_UPGRADE)) {
+ io_tier--;
+ }
+
+ throttle_info_end_io_internal(info, io_tier);
+}
+
+/*
+ * Decrement inflight count initially incremented by throttle_info_update_internal
+ */
+static
+void throttle_info_end_io_internal(struct _throttle_io_info_t *info, int throttle_level) {
+ if (throttle_level == THROTTLE_LEVEL_NONE) {
+ return;
+ }
+
+ microuptime(&info->throttle_window_start_timestamp[throttle_level]);
+ OSDecrementAtomic(&info->throttle_inflight_count[throttle_level]);
+ assert(info->throttle_inflight_count[throttle_level] >= 0);
+}
+
+/*
+ * If inflight is TRUE and bap is NULL then the caller is responsible for calling
+ * throttle_info_end_io_internal to avoid leaking in-flight I/O.
+ */
+static
+int throttle_info_update_internal(struct _throttle_io_info_t *info, uthread_t ut, int flags, boolean_t isssd, boolean_t inflight, struct bufattr *bap)
+{
+ int thread_throttle_level;
+
+ if (lowpri_throttle_enabled == 0 || info->throttle_disabled)
+ return THROTTLE_LEVEL_NONE;
+
+ if (ut == NULL)
+ ut = get_bsdthread_info(current_thread());
+
+ if (bap && inflight && !ut->uu_throttle_bc) {
+ thread_throttle_level = GET_BUFATTR_IO_TIER(bap);
+ if (ISSET(bap->ba_flags, BA_IO_TIER_UPGRADE)) {
+ thread_throttle_level--;
+ }
+ } else {
+ thread_throttle_level = throttle_get_thread_throttle_level(ut);
+ }
+
+ if (thread_throttle_level != THROTTLE_LEVEL_NONE) {
+ if(!ISSET(flags, B_PASSIVE)) {
+ info->throttle_last_IO_pid[thread_throttle_level] = proc_selfpid();
+ if (inflight && !ut->uu_throttle_bc) {
+ if (NULL != bap) {
+ SET(bap->ba_flags, BA_STRATEGY_TRACKED_IO);
+ }
+ OSIncrementAtomic(&info->throttle_inflight_count[thread_throttle_level]);
+ } else {
+ microuptime(&info->throttle_window_start_timestamp[thread_throttle_level]);
+ }
+ KERNEL_DEBUG_CONSTANT((FSDBG_CODE(DBG_THROTTLE, OPEN_THROTTLE_WINDOW)) | DBG_FUNC_NONE,
+ current_proc()->p_pid, thread_throttle_level, 0, 0, 0);
+ }
+ microuptime(&info->throttle_last_IO_timestamp[thread_throttle_level]);
+ }
+
+
+ if (thread_throttle_level >= THROTTLE_LEVEL_THROTTLED) {
+ /*
+ * I'd really like to do the IOSleep here, but
+ * we may be holding all kinds of filesystem related locks
+ * and the pages for this I/O marked 'busy'...
+ * we don't want to cause a normal task to block on
+ * one of these locks while we're throttling a task marked
+ * for low priority I/O... we'll mark the uthread and
+ * do the delay just before we return from the system
+ * call that triggered this I/O or from vnode_pagein
+ */
+ OSAddAtomic(1, &info->throttle_io_count);
+
+ throttle_info_set_initial_window(ut, info, FALSE, isssd);
+ }
+
+ return thread_throttle_level;
+}
+
+void *throttle_info_update_by_mount(mount_t mp)
+{
+ struct _throttle_io_info_t *info;
+ uthread_t ut;
+ boolean_t isssd = FALSE;
+
+ ut = get_bsdthread_info(current_thread());
+
+ if (mp != NULL) {
+ if (disk_conditioner_mount_is_ssd(mp))
+ isssd = TRUE;
+ info = &_throttle_io_info[mp->mnt_devbsdunit];
+ } else
+ info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1];
+
+ if (!ut->uu_lowpri_window)
+ throttle_info_set_initial_window(ut, info, FALSE, isssd);
+
+ return info;
+}
+
+
+/*
+ * KPI routine
+ *
+ * this is usually called before every I/O, used for throttled I/O
+ * book keeping. This routine has low overhead and does not sleep
+ */
+void throttle_info_update(void *throttle_info, int flags)
+{
+ if (throttle_info)
+ throttle_info_update_internal(throttle_info, NULL, flags, FALSE, FALSE, NULL);
+}
+
+/*
+ * KPI routine
+ *
+ * this is usually called before every I/O, used for throttled I/O
+ * book keeping. This routine has low overhead and does not sleep
+ */
+void throttle_info_update_by_mask(void *throttle_info_handle, int flags)
+{
+ void *throttle_info = throttle_info_handle;
+
+ /*
+ * for now we only use the lowest bit of the throttle mask, so the
+ * handle is the same as the throttle_info. Later if we store a
+ * set of throttle infos in the handle, we will want to loop through
+ * them and call throttle_info_update in a loop
+ */
+ throttle_info_update(throttle_info, flags);
+}
+/*
+ * KPI routine
+ *
+ * This routine marks the throttle info as disabled. Used for mount points which
+ * support I/O scheduling.
+ */
+
+void throttle_info_disable_throttle(int devno, boolean_t isfusion)
+{
+ struct _throttle_io_info_t *info;
+
+ if (devno < 0 || devno >= LOWPRI_MAX_NUM_DEV)
+ panic("Illegal devno (%d) passed into throttle_info_disable_throttle()", devno);
+
+ info = &_throttle_io_info[devno];
+ // don't disable software throttling on devices that are part of a fusion device
+ // and override the software throttle periods to use HDD periods
+ if (isfusion) {
+ info->throttle_is_fusion_with_priority = isfusion;
+ throttle_init_throttle_period(info, FALSE);
+ }
+ info->throttle_disabled = !info->throttle_is_fusion_with_priority;
+ return;
+}
+
+
+/*
+ * KPI routine (private)
+ * Called to determine if this IO is being throttled to this level so that it can be treated specially
+ */
+int throttle_info_io_will_be_throttled(void * throttle_info, int policy)
+{
+ struct _throttle_io_info_t *info = throttle_info;
+ struct timeval elapsed;
+ uint64_t elapsed_msecs;
+ int throttle_level;
+ int thread_throttle_level;
+
+ switch (policy) {
+
+ case IOPOL_THROTTLE:
+ thread_throttle_level = THROTTLE_LEVEL_TIER3;
+ break;
+ case IOPOL_UTILITY:
+ thread_throttle_level = THROTTLE_LEVEL_TIER2;
+ break;
+ case IOPOL_STANDARD:
+ thread_throttle_level = THROTTLE_LEVEL_TIER1;
+ break;
+ default:
+ thread_throttle_level = THROTTLE_LEVEL_TIER0;
+ break;
+ }
+ for (throttle_level = THROTTLE_LEVEL_START; throttle_level < thread_throttle_level; throttle_level++) {
+ if (info->throttle_inflight_count[throttle_level]) {
+ break;
+ }
+
+ microuptime(&elapsed);
+ timevalsub(&elapsed, &info->throttle_window_start_timestamp[throttle_level]);
+ elapsed_msecs = (uint64_t)elapsed.tv_sec * (uint64_t)1000 + (elapsed.tv_usec / 1000);
+
+ if (elapsed_msecs < (uint64_t)throttle_windows_msecs[thread_throttle_level])
+ break;
+ }
+ if (throttle_level >= thread_throttle_level) {
+ /*
+ * we're beyond all of the throttle windows
+ * so go ahead and treat as normal I/O
+ */
+ return (THROTTLE_DISENGAGED);
+ }
+ /*
+ * we're in the throttle window
+ */
+ return (THROTTLE_ENGAGED);
+}
+
+int throttle_lowpri_window(void)
+{
+ struct uthread *ut = get_bsdthread_info(current_thread());
+ return ut->uu_lowpri_window;
+}
+
+
+#if CONFIG_IOSCHED
+int upl_get_cached_tier(void *);
+#endif
+
+int
+spec_strategy(struct vnop_strategy_args *ap)
+{
+ buf_t bp;
+ int bflags;
+ int io_tier;
+ int passive;
+ dev_t bdev;
+ uthread_t ut;
+ mount_t mp;
+ struct bufattr *bap;
+ int strategy_ret;
+ struct _throttle_io_info_t *throttle_info;
+ boolean_t isssd = FALSE;
+ boolean_t inflight = FALSE;
+ boolean_t upgrade = FALSE;
+ int code = 0;
+
+#if !CONFIG_EMBEDDED
+ proc_t curproc = current_proc();
+#endif /* !CONFIG_EMBEDDED */
+
+ bp = ap->a_bp;
+ bdev = buf_device(bp);
+ mp = buf_vnode(bp)->v_mount;
+ bap = &bp->b_attr;
+
+#if CONFIG_IOSCHED
+ if (bp->b_flags & B_CLUSTER) {
+
+ io_tier = upl_get_cached_tier(bp->b_upl);
+
+ if (io_tier == -1)
+ io_tier = throttle_get_io_policy(&ut);
+#if DEVELOPMENT || DEBUG
+ else {
+ int my_io_tier = throttle_get_io_policy(&ut);
+
+ if (io_tier != my_io_tier)
+ KERNEL_DEBUG_CONSTANT((FSDBG_CODE(DBG_THROTTLE, IO_TIER_UPL_MISMATCH)) | DBG_FUNC_NONE, buf_kernel_addrperm_addr(bp), my_io_tier, io_tier, 0, 0);
+ }
+#endif
+ } else
+ io_tier = throttle_get_io_policy(&ut);
+#else
+ io_tier = throttle_get_io_policy(&ut);
+#endif
+ passive = throttle_get_passive_io_policy(&ut);
+
+ /*
+ * Mark if the I/O was upgraded by throttle_get_thread_throttle_level
+ * while preserving the original issued tier (throttle_get_io_policy
+ * does not return upgraded tiers)
+ */
+ if (mp && io_tier > throttle_get_thread_throttle_level_internal(ut, io_tier)) {
+#if CONFIG_IOSCHED
+ if (!(mp->mnt_ioflags & MNT_IOFLAGS_IOSCHED_SUPPORTED)) {
+ upgrade = TRUE;
+ }
+#else /* CONFIG_IOSCHED */
+ upgrade = TRUE;
+#endif /* CONFIG_IOSCHED */
+ }
+
+ if (bp->b_flags & B_META)
+ bap->ba_flags |= BA_META;
+
+#if CONFIG_IOSCHED
+ /*
+ * For I/O Scheduling, we currently do not have a way to track and expedite metadata I/Os.
+ * To ensure we dont get into priority inversions due to metadata I/Os, we use the following rules:
+ * For metadata reads, ceil all I/Os to IOSCHED_METADATA_TIER & mark them passive if the I/O tier was upgraded
+ * For metadata writes, unconditionally mark them as IOSCHED_METADATA_TIER and passive
+ */
+ if (bap->ba_flags & BA_META) {
+ if (mp && (mp->mnt_ioflags & MNT_IOFLAGS_IOSCHED_SUPPORTED)) {
+ if (bp->b_flags & B_READ) {
+ if (io_tier > IOSCHED_METADATA_TIER) {
+ io_tier = IOSCHED_METADATA_TIER;
+ passive = 1;
+ }
+ } else {
+ io_tier = IOSCHED_METADATA_TIER;
+ passive = 1;
+ }
+ }
+ }
+#endif /* CONFIG_IOSCHED */
+
+ SET_BUFATTR_IO_TIER(bap, io_tier);
+
+ if (passive) {
+ bp->b_flags |= B_PASSIVE;
+ bap->ba_flags |= BA_PASSIVE;
+ }
+
+#if !CONFIG_EMBEDDED
+ if ((curproc != NULL) && ((curproc->p_flag & P_DELAYIDLESLEEP) == P_DELAYIDLESLEEP))
+ bap->ba_flags |= BA_DELAYIDLESLEEP;
+#endif /* !CONFIG_EMBEDDED */
+
+ bflags = bp->b_flags;
+
+ if (((bflags & B_READ) == 0) && ((bflags & B_ASYNC) == 0))
+ bufattr_markquickcomplete(bap);
+
+ if (bflags & B_READ)
+ code |= DKIO_READ;
+ if (bflags & B_ASYNC)
+ code |= DKIO_ASYNC;
+
+ if (bap->ba_flags & BA_META)
+ code |= DKIO_META;
+ else if (bflags & B_PAGEIO)
+ code |= DKIO_PAGING;
+
+ if (io_tier != 0)
+ code |= DKIO_THROTTLE;
+
+ code |= ((io_tier << DKIO_TIER_SHIFT) & DKIO_TIER_MASK);
+
+ if (bflags & B_PASSIVE)
+ code |= DKIO_PASSIVE;
+
+ if (bap->ba_flags & BA_NOCACHE)
+ code |= DKIO_NOCACHE;
+
+ if (upgrade) {
+ code |= DKIO_TIER_UPGRADE;
+ SET(bap->ba_flags, BA_IO_TIER_UPGRADE);
+ }
+
+ if (kdebug_enable) {
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_COMMON, FSDBG_CODE(DBG_DKRW, code) | DBG_FUNC_NONE,
+ buf_kernel_addrperm_addr(bp), bdev, (int)buf_blkno(bp), buf_count(bp), 0);
+ }
+
+ thread_update_io_stats(current_thread(), buf_count(bp), code);
+
+ if (mp != NULL) {
+ if (disk_conditioner_mount_is_ssd(mp))
+ isssd = TRUE;
+ /*
+ * Partially initialized mounts don't have a final devbsdunit and should not be tracked.
+ * Verify that devbsdunit is initialized (non-zero) or that 0 is the correct initialized value
+ * (mnt_throttle_mask is initialized and num_trailing_0 would be 0)
+ */
+ if (mp->mnt_devbsdunit || (mp->mnt_throttle_mask != LOWPRI_MAX_NUM_DEV - 1 && mp->mnt_throttle_mask & 0x1)) {
+ inflight = TRUE;
+ }
+ throttle_info = &_throttle_io_info[mp->mnt_devbsdunit];
+
+ } else
+ throttle_info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1];
+
+ throttle_info_update_internal(throttle_info, ut, bflags, isssd, inflight, bap);
+
+ if ((bflags & B_READ) == 0) {
+ microuptime(&throttle_info->throttle_last_write_timestamp);
+
+ if (mp) {
+ mp->mnt_last_write_issued_timestamp = throttle_info->throttle_last_write_timestamp;
+ INCR_PENDING_IO(buf_count(bp), mp->mnt_pending_write_size);
+ }
+ } else if (mp) {
+ INCR_PENDING_IO(buf_count(bp), mp->mnt_pending_read_size);
+ }
+ /*
+ * The BootCache may give us special information about
+ * the IO, so it returns special values that we check
+ * for here.
+ *
+ * IO_SATISFIED_BY_CACHE
+ * The read has been satisfied by the boot cache. Don't
+ * throttle the thread unnecessarily.
+ *
+ * IO_SHOULD_BE_THROTTLED
+ * The boot cache is playing back a playlist and this IO
+ * cut through. Throttle it so we're not cutting through
+ * the boot cache too often.
+ *
+ * Note that typical strategy routines are defined with
+ * a void return so we'll get garbage here. In the
+ * unlikely case the garbage matches our special return
+ * value, it's not a big deal since we're only adjusting
+ * the throttling delay.
+ */
+#define IO_SATISFIED_BY_CACHE ((int)0xcafefeed)
+#define IO_SHOULD_BE_THROTTLED ((int)0xcafebeef)
+ typedef int strategy_fcn_ret_t(struct buf *bp);
+
+ strategy_ret = (*(strategy_fcn_ret_t*)bdevsw[major(bdev)].d_strategy)(bp);
+
+ // disk conditioner needs to track when this I/O actually starts
+ // which means track it after `strategy` which may include delays
+ // from inflight I/Os
+ microuptime(&bp->b_timestamp_tv);
+
+ if (IO_SATISFIED_BY_CACHE == strategy_ret) {
+ /*
+ * If this was a throttled IO satisfied by the boot cache,
+ * don't delay the thread.
+ */
+ throttle_info_reset_window(ut);
+
+ } else if (IO_SHOULD_BE_THROTTLED == strategy_ret) {
+ /*
+ * If the boot cache indicates this IO should be throttled,
+ * delay the thread.
+ */
+ throttle_info_set_initial_window(ut, throttle_info, TRUE, isssd);
+ }
+ return (0);
+}
+
+
+/*
+ * This is a noop, simply returning what one has been given.
+ */
+int
+spec_blockmap(__unused struct vnop_blockmap_args *ap)
+{
+ return (ENOTSUP);
+}
+
+
+/*
+ * Device close routine
+ */
+int
+spec_close(struct vnop_close_args *ap)
+{
+ struct vnode *vp = ap->a_vp;
+ dev_t dev = vp->v_rdev;
+ int error = 0;
+ int flags = ap->a_fflag;
+ struct proc *p = vfs_context_proc(ap->a_context);
+ struct session *sessp;
+
+ switch (vp->v_type) {
+
+ case VCHR:
+ /*
+ * Hack: a tty device that is a controlling terminal
+ * has a reference from the session structure.
+ * We cannot easily tell that a character device is
+ * a controlling terminal, unless it is the closing
+ * process' controlling terminal. In that case,
+ * if the reference count is 1 (this is the very
+ * last close)
+ */
+ sessp = proc_session(p);
+ devsw_lock(dev, S_IFCHR);
+ if (sessp != SESSION_NULL) {
+ if (vp == sessp->s_ttyvp && vcount(vp) == 1) {
+ struct tty *tp = TTY_NULL;
+
+ devsw_unlock(dev, S_IFCHR);
+ session_lock(sessp);
+ if (vp == sessp->s_ttyvp) {
+ tp = SESSION_TP(sessp);
+ sessp->s_ttyvp = NULL;
+ sessp->s_ttyvid = 0;
+ sessp->s_ttyp = TTY_NULL;
+ sessp->s_ttypgrpid = NO_PID;
+ }
+ session_unlock(sessp);
+
+ if (tp != TTY_NULL) {
+ /*
+ * We may have won a race with a proc_exit
+ * of the session leader, the winner
+ * clears the flag (even if not set)
+ */
+ tty_lock(tp);
+ ttyclrpgrphup(tp);
+ tty_unlock(tp);
+
+ ttyfree(tp);
+ }
+ devsw_lock(dev, S_IFCHR);
+ }
+ session_rele(sessp);
+ }
+
+ if (--vp->v_specinfo->si_opencount < 0)
+ panic("negative open count (c, %u, %u)", major(dev), minor(dev));
+
+ /*
+ * close on last reference or on vnode revoke call
+ */
+ if (vcount(vp) == 0 || (flags & IO_REVOKE) != 0)
+ error = cdevsw[major(dev)].d_close(dev, flags, S_IFCHR, p);
+
+ devsw_unlock(dev, S_IFCHR);
+ break;
+
+ case VBLK:
+ /*
+ * If there is more than one outstanding open, don't
+ * send the close to the device.
+ */
+ devsw_lock(dev, S_IFBLK);
+ if (vcount(vp) > 1) {
+ vp->v_specinfo->si_opencount--;
+ devsw_unlock(dev, S_IFBLK);
+ return (0);
+ }
+ devsw_unlock(dev, S_IFBLK);
+
+ /*
+ * On last close of a block device (that isn't mounted)
+ * we must invalidate any in core blocks, so that
+ * we can, for instance, change floppy disks.
+ */
+ if ((error = spec_fsync_internal(vp, MNT_WAIT, ap->a_context)))
+ return (error);
+
+ error = buf_invalidateblks(vp, BUF_WRITE_DATA, 0, 0);
+ if (error)
+ return (error);
+
+ devsw_lock(dev, S_IFBLK);
+
+ if (--vp->v_specinfo->si_opencount < 0)
+ panic("negative open count (b, %u, %u)", major(dev), minor(dev));
+
+ if (vcount(vp) == 0)
+ error = bdevsw[major(dev)].d_close(dev, flags, S_IFBLK, p);
+
+ devsw_unlock(dev, S_IFBLK);
+ break;
+
+ default:
+ panic("spec_close: not special");
+ return(EBADF);
+ }
+
+ return error;
+}
+
+/*
+ * Return POSIX pathconf information applicable to special devices.
+ */
+int
+spec_pathconf(struct vnop_pathconf_args *ap)
+{
+
+ switch (ap->a_name) {
+ case _PC_LINK_MAX:
+ *ap->a_retval = LINK_MAX;
+ return (0);
+ case _PC_MAX_CANON:
+ *ap->a_retval = MAX_CANON;
+ return (0);
+ case _PC_MAX_INPUT:
+ *ap->a_retval = MAX_INPUT;
+ return (0);
+ case _PC_PIPE_BUF:
+ *ap->a_retval = PIPE_BUF;
+ return (0);
+ case _PC_CHOWN_RESTRICTED:
+ *ap->a_retval = 200112; /* _POSIX_CHOWN_RESTRICTED */
+ return (0);
+ case _PC_VDISABLE:
+ *ap->a_retval = _POSIX_VDISABLE;
+ return (0);
+ default:
+ return (EINVAL);
+ }
+ /* NOTREACHED */
+}
+
+/*
+ * Special device failed operation
+ */
+int
+spec_ebadf(__unused void *dummy)
+{
+
+ return (EBADF);
+}
+
+/* Blktooff derives file offset from logical block number */
+int
+spec_blktooff(struct vnop_blktooff_args *ap)
+{
+ struct vnode *vp = ap->a_vp;
+
+ switch (vp->v_type) {
+ case VCHR:
+ *ap->a_offset = (off_t)-1; /* failure */
+ return (ENOTSUP);
+
+ case VBLK:
+ printf("spec_blktooff: not implemented for VBLK\n");
+ *ap->a_offset = (off_t)-1; /* failure */
+ return (ENOTSUP);
+
+ default:
+ panic("spec_blktooff type");
+ }
+ /* NOTREACHED */
+
+ return (0);
+}
projects / apple / xnu.git / blobdiff