Date: Tue, 8 Mar 2016 16:11:59 +0000 (UTC) From: Alexander Motin <mav@FreeBSD.org> To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-vendor@freebsd.org Subject: svn commit: r296505 - vendor-sys/illumos/dist/uts/common/fs/zfs vendor-sys/illumos/dist/uts/common/fs/zfs/sys vendor/illumos/dist/cmd/zinject Message-ID: <201603081611.u28GBxl2029770@repo.freebsd.org>
next in thread | raw e-mail | index | archive | help
Author: mav Date: Tue Mar 8 16:11:59 2016 New Revision: 296505 URL: https://svnweb.freebsd.org/changeset/base/296505 Log: 6531 Provide mechanism to artificially limit disk performance Reviewed by: Paul Dagnelie <pcd@delphix.com> Reviewed by: Matthew Ahrens <mahrens@delphix.com> Reviewed by: George Wilson <george.wilson@delphix.com> Approved by: Dan McDonald <danmcd@omniti.com> Author: Prakash Surya <prakash.surya@delphix.com> illumos/illumos-gate@97e81309571898df9fdd94aab1216dfcf23e060b Modified: vendor-sys/illumos/dist/uts/common/fs/zfs/sys/zfs_ioctl.h vendor-sys/illumos/dist/uts/common/fs/zfs/sys/zio.h vendor-sys/illumos/dist/uts/common/fs/zfs/vdev_disk.c vendor-sys/illumos/dist/uts/common/fs/zfs/vdev_file.c vendor-sys/illumos/dist/uts/common/fs/zfs/vdev_queue.c vendor-sys/illumos/dist/uts/common/fs/zfs/zio.c vendor-sys/illumos/dist/uts/common/fs/zfs/zio_inject.c Changes in other areas also in this revision: Modified: vendor/illumos/dist/cmd/zinject/zinject.c Modified: vendor-sys/illumos/dist/uts/common/fs/zfs/sys/zfs_ioctl.h ============================================================================== --- vendor-sys/illumos/dist/uts/common/fs/zfs/sys/zfs_ioctl.h Tue Mar 8 15:55:43 2016 (r296504) +++ vendor-sys/illumos/dist/uts/common/fs/zfs/sys/zfs_ioctl.h Tue Mar 8 16:11:59 2016 (r296505) @@ -292,6 +292,7 @@ typedef struct zinject_record { uint32_t zi_iotype; int32_t zi_duration; uint64_t zi_timer; + uint64_t zi_nlanes; uint32_t zi_cmd; uint32_t zi_pad; } zinject_record_t; Modified: vendor-sys/illumos/dist/uts/common/fs/zfs/sys/zio.h ============================================================================== --- vendor-sys/illumos/dist/uts/common/fs/zfs/sys/zio.h Tue Mar 8 15:55:43 2016 (r296504) +++ vendor-sys/illumos/dist/uts/common/fs/zfs/sys/zio.h Tue Mar 8 16:11:59 2016 (r296505) @@ -419,6 +419,7 @@ struct zio { uint64_t io_offset; hrtime_t io_timestamp; + hrtime_t io_target_timestamp; avl_node_t io_queue_node; avl_node_t io_offset_node; @@ -506,6 +507,8 @@ extern int zio_wait(zio_t *zio); extern void zio_nowait(zio_t *zio); extern void zio_execute(zio_t *zio); extern void zio_interrupt(zio_t *zio); +extern void zio_delay_init(zio_t *zio); +extern void zio_delay_interrupt(zio_t *zio); extern zio_t *zio_walk_parents(zio_t *cio); extern zio_t *zio_walk_children(zio_t *pio); @@ -567,7 +570,7 @@ extern int zio_handle_fault_injection(zi extern int zio_handle_device_injection(vdev_t *vd, zio_t *zio, int error); extern int zio_handle_label_injection(zio_t *zio, int error); extern void zio_handle_ignored_writes(zio_t *zio); -extern uint64_t zio_handle_io_delay(zio_t *zio); +extern hrtime_t zio_handle_io_delay(zio_t *zio); /* * Checksum ereport functions Modified: vendor-sys/illumos/dist/uts/common/fs/zfs/vdev_disk.c ============================================================================== --- vendor-sys/illumos/dist/uts/common/fs/zfs/vdev_disk.c Tue Mar 8 15:55:43 2016 (r296504) +++ vendor-sys/illumos/dist/uts/common/fs/zfs/vdev_disk.c Tue Mar 8 16:11:59 2016 (r296505) @@ -20,7 +20,7 @@ */ /* * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. - * Copyright (c) 2012, 2014 by Delphix. All rights reserved. + * Copyright (c) 2012, 2015 by Delphix. All rights reserved. * Copyright 2013 Nexenta Systems, Inc. All rights reserved. * Copyright (c) 2013 Joyent, Inc. All rights reserved. */ @@ -691,7 +691,7 @@ vdev_disk_io_intr(buf_t *bp) kmem_free(vb, sizeof (vdev_buf_t)); - zio_interrupt(zio); + zio_delay_interrupt(zio); } static void @@ -797,6 +797,7 @@ vdev_disk_io_start(zio_t *zio) } ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); + zio->io_target_timestamp = zio_handle_io_delay(zio); vb = kmem_alloc(sizeof (vdev_buf_t), KM_SLEEP); Modified: vendor-sys/illumos/dist/uts/common/fs/zfs/vdev_file.c ============================================================================== --- vendor-sys/illumos/dist/uts/common/fs/zfs/vdev_file.c Tue Mar 8 15:55:43 2016 (r296504) +++ vendor-sys/illumos/dist/uts/common/fs/zfs/vdev_file.c Tue Mar 8 16:11:59 2016 (r296505) @@ -20,7 +20,7 @@ */ /* * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. - * Copyright (c) 2011, 2014 by Delphix. All rights reserved. + * Copyright (c) 2011, 2015 by Delphix. All rights reserved. */ #include <sys/zfs_context.h> @@ -158,7 +158,7 @@ vdev_file_io_intr(buf_t *bp) zio->io_error = SET_ERROR(ENOSPC); kmem_free(vb, sizeof (vdev_buf_t)); - zio_interrupt(zio); + zio_delay_interrupt(zio); } static void @@ -212,6 +212,7 @@ vdev_file_io_start(zio_t *zio) } ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); + zio->io_target_timestamp = zio_handle_io_delay(zio); vb = kmem_alloc(sizeof (vdev_buf_t), KM_SLEEP); Modified: vendor-sys/illumos/dist/uts/common/fs/zfs/vdev_queue.c ============================================================================== --- vendor-sys/illumos/dist/uts/common/fs/zfs/vdev_queue.c Tue Mar 8 15:55:43 2016 (r296504) +++ vendor-sys/illumos/dist/uts/common/fs/zfs/vdev_queue.c Tue Mar 8 16:11:59 2016 (r296505) @@ -729,9 +729,6 @@ vdev_queue_io_done(zio_t *zio) vdev_queue_t *vq = &zio->io_vd->vdev_queue; zio_t *nio; - if (zio_injection_enabled) - delay(SEC_TO_TICK(zio_handle_io_delay(zio))); - mutex_enter(&vq->vq_lock); vdev_queue_pending_remove(vq, zio); Modified: vendor-sys/illumos/dist/uts/common/fs/zfs/zio.c ============================================================================== --- vendor-sys/illumos/dist/uts/common/fs/zfs/zio.c Tue Mar 8 15:55:43 2016 (r296504) +++ vendor-sys/illumos/dist/uts/common/fs/zfs/zio.c Tue Mar 8 16:11:59 2016 (r296505) @@ -1352,6 +1352,58 @@ zio_interrupt(zio_t *zio) zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); } +void +zio_delay_interrupt(zio_t *zio) +{ + /* + * The timeout_generic() function isn't defined in userspace, so + * rather than trying to implement the function, the zio delay + * functionality has been disabled for userspace builds. + */ + +#ifdef _KERNEL + /* + * If io_target_timestamp is zero, then no delay has been registered + * for this IO, thus jump to the end of this function and "skip" the + * delay; issuing it directly to the zio layer. + */ + if (zio->io_target_timestamp != 0) { + hrtime_t now = gethrtime(); + + if (now >= zio->io_target_timestamp) { + /* + * This IO has already taken longer than the target + * delay to complete, so we don't want to delay it + * any longer; we "miss" the delay and issue it + * directly to the zio layer. This is likely due to + * the target latency being set to a value less than + * the underlying hardware can satisfy (e.g. delay + * set to 1ms, but the disks take 10ms to complete an + * IO request). + */ + + DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, + hrtime_t, now); + + zio_interrupt(zio); + } else { + hrtime_t diff = zio->io_target_timestamp - now; + + DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, + hrtime_t, now, hrtime_t, diff); + + (void) timeout_generic(CALLOUT_NORMAL, + (void (*)(void *))zio_interrupt, zio, diff, 1, 0); + } + + return; + } +#endif + + DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); + zio_interrupt(zio); +} + /* * Execute the I/O pipeline until one of the following occurs: * Modified: vendor-sys/illumos/dist/uts/common/fs/zfs/zio_inject.c ============================================================================== --- vendor-sys/illumos/dist/uts/common/fs/zfs/zio_inject.c Tue Mar 8 15:55:43 2016 (r296504) +++ vendor-sys/illumos/dist/uts/common/fs/zfs/zio_inject.c Tue Mar 8 16:11:59 2016 (r296505) @@ -20,7 +20,7 @@ */ /* * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. - * Copyright (c) 2012, 2014 by Delphix. All rights reserved. + * Copyright (c) 2012, 2015 by Delphix. All rights reserved. */ /* @@ -49,15 +49,53 @@ uint32_t zio_injection_enabled; +/* + * Data describing each zinject handler registered on the system, and + * contains the list node linking the handler in the global zinject + * handler list. + */ typedef struct inject_handler { int zi_id; spa_t *zi_spa; zinject_record_t zi_record; + uint64_t *zi_lanes; + int zi_next_lane; list_node_t zi_link; } inject_handler_t; +/* + * List of all zinject handlers registered on the system, protected by + * the inject_lock defined below. + */ static list_t inject_handlers; + +/* + * This protects insertion into, and traversal of, the inject handler + * list defined above; as well as the inject_delay_count. Any time a + * handler is inserted or removed from the list, this lock should be + * taken as a RW_WRITER; and any time traversal is done over the list + * (without modification to it) this lock should be taken as a RW_READER. + */ static krwlock_t inject_lock; + +/* + * This holds the number of zinject delay handlers that have been + * registered on the system. It is protected by the inject_lock defined + * above. Thus modifications to this count must be a RW_WRITER of the + * inject_lock, and reads of this count must be (at least) a RW_READER + * of the lock. + */ +static int inject_delay_count = 0; + +/* + * This lock is used only in zio_handle_io_delay(), refer to the comment + * in that function for more details. + */ +static kmutex_t inject_delay_mtx; + +/* + * Used to assign unique identifying numbers to each new zinject handler. + */ static int inject_next_id = 1; /* @@ -360,32 +398,164 @@ spa_handle_ignored_writes(spa_t *spa) rw_exit(&inject_lock); } -uint64_t +hrtime_t zio_handle_io_delay(zio_t *zio) { vdev_t *vd = zio->io_vd; - inject_handler_t *handler; - uint64_t seconds = 0; - - if (zio_injection_enabled == 0) - return (0); + inject_handler_t *min_handler = NULL; + hrtime_t min_target = 0; rw_enter(&inject_lock, RW_READER); - for (handler = list_head(&inject_handlers); handler != NULL; - handler = list_next(&inject_handlers, handler)) { + /* + * inject_delay_count is a subset of zio_injection_enabled that + * is only incremented for delay handlers. These checks are + * mainly added to remind the reader why we're not explicitly + * checking zio_injection_enabled like the other functions. + */ + IMPLY(inject_delay_count > 0, zio_injection_enabled > 0); + IMPLY(zio_injection_enabled == 0, inject_delay_count == 0); + + /* + * If there aren't any inject delay handlers registered, then we + * can short circuit and simply return 0 here. A value of zero + * informs zio_delay_interrupt() that this request should not be + * delayed. This short circuit keeps us from acquiring the + * inject_delay_mutex unnecessarily. + */ + if (inject_delay_count == 0) { + rw_exit(&inject_lock); + return (0); + } + + /* + * Each inject handler has a number of "lanes" associated with + * it. Each lane is able to handle requests independently of one + * another, and at a latency defined by the inject handler + * record's zi_timer field. Thus if a handler in configured with + * a single lane with a 10ms latency, it will delay requests + * such that only a single request is completed every 10ms. So, + * if more than one request is attempted per each 10ms interval, + * the average latency of the requests will be greater than + * 10ms; but if only a single request is submitted each 10ms + * interval the average latency will be 10ms. + * + * We need to acquire this mutex to prevent multiple concurrent + * threads being assigned to the same lane of a given inject + * handler. The mutex allows us to perform the following two + * operations atomically: + * + * 1. determine the minimum handler and minimum target + * value of all the possible handlers + * 2. update that minimum handler's lane array + * + * Without atomicity, two (or more) threads could pick the same + * lane in step (1), and then conflict with each other in step + * (2). This could allow a single lane handler to process + * multiple requests simultaneously, which shouldn't be possible. + */ + mutex_enter(&inject_delay_mtx); + for (inject_handler_t *handler = list_head(&inject_handlers); + handler != NULL; handler = list_next(&inject_handlers, handler)) { if (handler->zi_record.zi_cmd != ZINJECT_DELAY_IO) continue; - if (vd->vdev_guid == handler->zi_record.zi_guid) { - seconds = handler->zi_record.zi_timer; - break; + if (vd->vdev_guid != handler->zi_record.zi_guid) + continue; + + /* + * Defensive; should never happen as the array allocation + * occurs prior to inserting this handler on the list. + */ + ASSERT3P(handler->zi_lanes, !=, NULL); + + /* + * This should never happen, the zinject command should + * prevent a user from setting an IO delay with zero lanes. + */ + ASSERT3U(handler->zi_record.zi_nlanes, !=, 0); + + ASSERT3U(handler->zi_record.zi_nlanes, >, + handler->zi_next_lane); + + /* + * We want to issue this IO to the lane that will become + * idle the soonest, so we compare the soonest this + * specific handler can complete the IO with all other + * handlers, to find the lowest value of all possible + * lanes. We then use this lane to submit the request. + * + * Since each handler has a constant value for its + * delay, we can just use the "next" lane for that + * handler; as it will always be the lane with the + * lowest value for that particular handler (i.e. the + * lane that will become idle the soonest). This saves a + * scan of each handler's lanes array. + * + * There's two cases to consider when determining when + * this specific IO request should complete. If this + * lane is idle, we want to "submit" the request now so + * it will complete after zi_timer milliseconds. Thus, + * we set the target to now + zi_timer. + * + * If the lane is busy, we want this request to complete + * zi_timer milliseconds after the lane becomes idle. + * Since the 'zi_lanes' array holds the time at which + * each lane will become idle, we use that value to + * determine when this request should complete. + */ + hrtime_t idle = handler->zi_record.zi_timer + gethrtime(); + hrtime_t busy = handler->zi_record.zi_timer + + handler->zi_lanes[handler->zi_next_lane]; + hrtime_t target = MAX(idle, busy); + + if (min_handler == NULL) { + min_handler = handler; + min_target = target; + continue; } + ASSERT3P(min_handler, !=, NULL); + ASSERT3U(min_target, !=, 0); + + /* + * We don't yet increment the "next lane" variable since + * we still might find a lower value lane in another + * handler during any remaining iterations. Once we're + * sure we've selected the absolute minimum, we'll claim + * the lane and increment the handler's "next lane" + * field below. + */ + + if (target < min_target) { + min_handler = handler; + min_target = target; + } } + + /* + * 'min_handler' will be NULL if no IO delays are registered for + * this vdev, otherwise it will point to the handler containing + * the lane that will become idle the soonest. + */ + if (min_handler != NULL) { + ASSERT3U(min_target, !=, 0); + min_handler->zi_lanes[min_handler->zi_next_lane] = min_target; + + /* + * If we've used all possible lanes for this handler, + * loop back and start using the first lane again; + * otherwise, just increment the lane index. + */ + min_handler->zi_next_lane = (min_handler->zi_next_lane + 1) % + min_handler->zi_record.zi_nlanes; + } + + mutex_exit(&inject_delay_mtx); rw_exit(&inject_lock); - return (seconds); + + return (min_target); } /* @@ -409,6 +579,24 @@ zio_inject_fault(char *name, int flags, if ((error = spa_reset(name)) != 0) return (error); + if (record->zi_cmd == ZINJECT_DELAY_IO) { + /* + * A value of zero for the number of lanes or for the + * delay time doesn't make sense. + */ + if (record->zi_timer == 0 || record->zi_nlanes == 0) + return (SET_ERROR(EINVAL)); + + /* + * The number of lanes is directly mapped to the size of + * an array used by the handler. Thus, to ensure the + * user doesn't trigger an allocation that's "too large" + * we cap the number of lanes here. + */ + if (record->zi_nlanes >= UINT16_MAX) + return (SET_ERROR(EINVAL)); + } + if (!(flags & ZINJECT_NULL)) { /* * spa_inject_ref() will add an injection reference, which will @@ -420,11 +608,34 @@ zio_inject_fault(char *name, int flags, handler = kmem_alloc(sizeof (inject_handler_t), KM_SLEEP); + handler->zi_spa = spa; + handler->zi_record = *record; + + if (handler->zi_record.zi_cmd == ZINJECT_DELAY_IO) { + handler->zi_lanes = kmem_zalloc( + sizeof (*handler->zi_lanes) * + handler->zi_record.zi_nlanes, KM_SLEEP); + handler->zi_next_lane = 0; + } else { + handler->zi_lanes = NULL; + handler->zi_next_lane = 0; + } + rw_enter(&inject_lock, RW_WRITER); + /* + * We can't move this increment into the conditional + * above because we need to hold the RW_WRITER lock of + * inject_lock, and we don't want to hold that while + * allocating the handler's zi_lanes array. + */ + if (handler->zi_record.zi_cmd == ZINJECT_DELAY_IO) { + ASSERT3S(inject_delay_count, >=, 0); + inject_delay_count++; + ASSERT3S(inject_delay_count, >, 0); + } + *id = handler->zi_id = inject_next_id++; - handler->zi_spa = spa; - handler->zi_record = *record; list_insert_tail(&inject_handlers, handler); atomic_inc_32(&zio_injection_enabled); @@ -502,9 +713,23 @@ zio_clear_fault(int id) return (SET_ERROR(ENOENT)); } + if (handler->zi_record.zi_cmd == ZINJECT_DELAY_IO) { + ASSERT3S(inject_delay_count, >, 0); + inject_delay_count--; + ASSERT3S(inject_delay_count, >=, 0); + } + list_remove(&inject_handlers, handler); rw_exit(&inject_lock); + if (handler->zi_record.zi_cmd == ZINJECT_DELAY_IO) { + ASSERT3P(handler->zi_lanes, !=, NULL); + kmem_free(handler->zi_lanes, sizeof (*handler->zi_lanes) * + handler->zi_record.zi_nlanes); + } else { + ASSERT3P(handler->zi_lanes, ==, NULL); + } + spa_inject_delref(handler->zi_spa); kmem_free(handler, sizeof (inject_handler_t)); atomic_dec_32(&zio_injection_enabled); @@ -516,6 +741,7 @@ void zio_inject_init(void) { rw_init(&inject_lock, NULL, RW_DEFAULT, NULL); + mutex_init(&inject_delay_mtx, NULL, MUTEX_DEFAULT, NULL); list_create(&inject_handlers, sizeof (inject_handler_t), offsetof(inject_handler_t, zi_link)); } @@ -524,5 +750,6 @@ void zio_inject_fini(void) { list_destroy(&inject_handlers); + mutex_destroy(&inject_delay_mtx); rw_destroy(&inject_lock); }
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201603081611.u28GBxl2029770>