X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/b0d623f7f2ae71ed96e60569f61f9a9a27016e80..4d15aeb193b2c68f1d38666c317f8d3734f5f083:/bsd/miscfs/specfs/spec_vnops.c?ds=inline diff --git a/bsd/miscfs/specfs/spec_vnops.c b/bsd/miscfs/specfs/spec_vnops.c index cbd0de6d9..f698e5d68 100644 --- a/bsd/miscfs/specfs/spec_vnops.c +++ b/bsd/miscfs/specfs/spec_vnops.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2000-2006 Apple Computer, Inc. All rights reserved. + * Copyright (c) 2000-2016 Apple Computer, Inc. All rights reserved. * * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ * @@ -69,8 +69,9 @@ #include #include #include -#include #include +#include +#include #include #include #include @@ -80,17 +81,29 @@ #include #include #include +#include #include #include +#include +#include +#include +#include +#include +#include + +#include + #include /* XXX following three prototypes should be in a header file somewhere */ extern dev_t chrtoblk(dev_t dev); -extern int iskmemdev(dev_t dev); +extern boolean_t iskmemdev(dev_t dev); extern int bpfkqfilter(dev_t dev, struct knote *kn); extern int ptsd_kqfilter(dev_t dev, struct knote *kn); +extern int ignore_is_ssd; + struct vnode *speclisth[SPECHSZ]; /* symbolic sleep message strings for devices */ @@ -150,6 +163,82 @@ struct vnodeopv_desc spec_vnodeop_opv_desc = static void set_blocksize(vnode_t, dev_t); +#define LOWPRI_TIER1_WINDOW_MSECS 25 +#define LOWPRI_TIER2_WINDOW_MSECS 100 +#define LOWPRI_TIER3_WINDOW_MSECS 500 + +#define LOWPRI_TIER1_IO_PERIOD_MSECS 40 +#define LOWPRI_TIER2_IO_PERIOD_MSECS 85 +#define LOWPRI_TIER3_IO_PERIOD_MSECS 200 + +#define LOWPRI_TIER1_IO_PERIOD_SSD_MSECS 5 +#define LOWPRI_TIER2_IO_PERIOD_SSD_MSECS 15 +#define LOWPRI_TIER3_IO_PERIOD_SSD_MSECS 25 + + +int throttle_windows_msecs[THROTTLE_LEVEL_END + 1] = { + 0, + LOWPRI_TIER1_WINDOW_MSECS, + LOWPRI_TIER2_WINDOW_MSECS, + LOWPRI_TIER3_WINDOW_MSECS, +}; + +int throttle_io_period_msecs[THROTTLE_LEVEL_END + 1] = { + 0, + LOWPRI_TIER1_IO_PERIOD_MSECS, + LOWPRI_TIER2_IO_PERIOD_MSECS, + LOWPRI_TIER3_IO_PERIOD_MSECS, +}; + +int throttle_io_period_ssd_msecs[THROTTLE_LEVEL_END + 1] = { + 0, + LOWPRI_TIER1_IO_PERIOD_SSD_MSECS, + LOWPRI_TIER2_IO_PERIOD_SSD_MSECS, + LOWPRI_TIER3_IO_PERIOD_SSD_MSECS, +}; + + +int throttled_count[THROTTLE_LEVEL_END + 1]; + +struct _throttle_io_info_t { + lck_mtx_t throttle_lock; + + struct timeval throttle_last_write_timestamp; + struct timeval throttle_min_timer_deadline; + struct timeval throttle_window_start_timestamp[THROTTLE_LEVEL_END + 1]; /* window starts at both the beginning and completion of an I/O */ + struct timeval throttle_last_IO_timestamp[THROTTLE_LEVEL_END + 1]; + pid_t throttle_last_IO_pid[THROTTLE_LEVEL_END + 1]; + struct timeval throttle_start_IO_period_timestamp[THROTTLE_LEVEL_END + 1]; + int32_t throttle_inflight_count[THROTTLE_LEVEL_END + 1]; + + TAILQ_HEAD( , uthread) throttle_uthlist[THROTTLE_LEVEL_END + 1]; /* Lists of throttled uthreads */ + int throttle_next_wake_level; + + thread_call_t throttle_timer_call; + int32_t throttle_timer_ref; + int32_t throttle_timer_active; + + int32_t throttle_io_count; + int32_t throttle_io_count_begin; + int *throttle_io_periods; + uint32_t throttle_io_period_num; + + int32_t throttle_refcnt; + int32_t throttle_alloc; + int32_t throttle_disabled; + int32_t throttle_is_fusion_with_priority; +}; + +struct _throttle_io_info_t _throttle_io_info[LOWPRI_MAX_NUM_DEV]; + + +int lowpri_throttle_enabled = 1; + + +static void throttle_info_end_io_internal(struct _throttle_io_info_t *info, int throttle_level); +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); +static int throttle_get_thread_throttle_level(uthread_t ut); +static int throttle_get_thread_throttle_level_internal(uthread_t ut, int io_tier); /* * Trivial lookup routine that always fails. @@ -229,25 +318,62 @@ spec_open(struct vnop_open_args *ap) */ if (securelevel >= 2 && isdisk(dev, VCHR)) return (EPERM); + + /* Never allow writing to /dev/mem or /dev/kmem */ + if (iskmemdev(dev)) + return (EPERM); /* - * When running in secure mode, do not allow opens - * for writing of /dev/mem, /dev/kmem, or character - * devices whose corresponding block devices are - * currently mounted. + * When running in secure mode, do not allow opens for + * writing of character devices whose corresponding block + * devices are currently mounted. */ if (securelevel >= 1) { if ((bdev = chrtoblk(dev)) != NODEV && check_mountedon(bdev, VBLK, &error)) return (error); - if (iskmemdev(dev)) - return (EPERM); } } - if (cdevsw[maj].d_type == D_TTY) { - vnode_lock(vp); - vp->v_flag |= VISTTY; - vnode_unlock(vp); - } + + devsw_lock(dev, S_IFCHR); error = (*cdevsw[maj].d_open)(dev, ap->a_mode, S_IFCHR, p); + + if (error == 0) { + vp->v_specinfo->si_opencount++; + } + + devsw_unlock(dev, S_IFCHR); + + if (error == 0 && cdevsw[maj].d_type == D_DISK && !vp->v_un.vu_specinfo->si_initted) { + int isssd = 0; + uint64_t throttle_mask = 0; + uint32_t devbsdunit = 0; + + if (VNOP_IOCTL(vp, DKIOCGETTHROTTLEMASK, (caddr_t)&throttle_mask, 0, NULL) == 0) { + + if (throttle_mask != 0 && + VNOP_IOCTL(vp, DKIOCISSOLIDSTATE, (caddr_t)&isssd, 0, ap->a_context) == 0) { + /* + * as a reasonable approximation, only use the lowest bit of the mask + * to generate a disk unit number + */ + devbsdunit = num_trailing_0(throttle_mask); + + vnode_lock(vp); + + vp->v_un.vu_specinfo->si_isssd = isssd; + vp->v_un.vu_specinfo->si_devbsdunit = devbsdunit; + vp->v_un.vu_specinfo->si_throttle_mask = throttle_mask; + vp->v_un.vu_specinfo->si_throttleable = 1; + vp->v_un.vu_specinfo->si_initted = 1; + + vnode_unlock(vp); + } + } + if (vp->v_un.vu_specinfo->si_initted == 0) { + vnode_lock(vp); + vp->v_un.vu_specinfo->si_initted = 1; + vnode_unlock(vp); + } + } return (error); case VBLK: @@ -266,7 +392,14 @@ spec_open(struct vnop_open_args *ap) */ if ( (error = vfs_mountedon(vp)) ) return (error); + + devsw_lock(dev, S_IFBLK); error = (*bdevsw[maj].d_open)(dev, ap->a_mode, S_IFBLK, p); + if (!error) { + vp->v_specinfo->si_opencount++; + } + devsw_unlock(dev, S_IFBLK); + if (!error) { u_int64_t blkcnt; u_int32_t blksize; @@ -339,9 +472,22 @@ spec_read(struct vnop_read_args *ap) switch (vp->v_type) { case VCHR: + { + struct _throttle_io_info_t *throttle_info = NULL; + int thread_throttle_level; + if (cdevsw[major(vp->v_rdev)].d_type == D_DISK && vp->v_un.vu_specinfo->si_throttleable) { + throttle_info = &_throttle_io_info[vp->v_un.vu_specinfo->si_devbsdunit]; + thread_throttle_level = throttle_info_update_internal(throttle_info, NULL, 0, vp->v_un.vu_specinfo->si_isssd, TRUE, NULL); + } error = (*cdevsw[major(vp->v_rdev)].d_read) (vp->v_rdev, uio, ap->a_ioflag); + + if (throttle_info) { + throttle_info_end_io_internal(throttle_info, thread_throttle_level); + } + return (error); + } case VBLK: if (uio->uio_offset < 0) @@ -382,7 +528,7 @@ spec_read(struct vnop_read_args *ap) } n = min((unsigned)(n - on), uio_resid(uio)); - error = uiomove((char *)0 + buf_dataptr(bp) + on, n, uio); + error = uiomove((char *)buf_dataptr(bp) + on, n, uio); if (n + on == bsize) buf_markaged(bp); buf_brelse(bp); @@ -424,9 +570,25 @@ spec_write(struct vnop_write_args *ap) switch (vp->v_type) { case VCHR: + { + struct _throttle_io_info_t *throttle_info = NULL; + int thread_throttle_level; + if (cdevsw[major(vp->v_rdev)].d_type == D_DISK && vp->v_un.vu_specinfo->si_throttleable) { + throttle_info = &_throttle_io_info[vp->v_un.vu_specinfo->si_devbsdunit]; + + thread_throttle_level = throttle_info_update_internal(throttle_info, NULL, 0, vp->v_un.vu_specinfo->si_isssd, TRUE, NULL); + + microuptime(&throttle_info->throttle_last_write_timestamp); + } error = (*cdevsw[major(vp->v_rdev)].d_write) (vp->v_rdev, uio, ap->a_ioflag); + + if (throttle_info) { + throttle_info_end_io_internal(throttle_info, thread_throttle_level); + } + return (error); + } case VBLK: if (uio_resid(uio) == 0) @@ -484,7 +646,7 @@ spec_write(struct vnop_write_args *ap) } n = min(n, bsize - buf_resid(bp)); - error = uiomove((char *)0 + buf_dataptr(bp) + on, n, uio); + error = uiomove((char *)buf_dataptr(bp) + on, n, uio); if (error) { buf_brelse(bp); return (error); @@ -521,7 +683,7 @@ spec_ioctl(struct vnop_ioctl_args *ap) int retval = 0; KERNEL_DEBUG_CONSTANT(FSDBG_CODE(DBG_IOCTL, 0) | DBG_FUNC_START, - (unsigned int)dev, (unsigned int)ap->a_command, (unsigned int)ap->a_fflag, (unsigned int)ap->a_vp->v_type, 0); + dev, ap->a_command, ap->a_fflag, ap->a_vp->v_type, 0); switch (ap->a_vp->v_type) { @@ -531,8 +693,9 @@ spec_ioctl(struct vnop_ioctl_args *ap) break; case VBLK: - retval = (*bdevsw[major(dev)].d_ioctl)(dev, ap->a_command, ap->a_data, - ap->a_fflag, p); + retval = (*bdevsw[major(dev)].d_ioctl)(dev, ap->a_command, ap->a_data, ap->a_fflag, p); + if (!retval && ap->a_command == DKIOCSETBLOCKSIZE) + ap->a_vp->v_specsize = *(uint32_t *)ap->a_data; break; default: @@ -540,7 +703,7 @@ spec_ioctl(struct vnop_ioctl_args *ap) /* NOTREACHED */ } KERNEL_DEBUG_CONSTANT(FSDBG_CODE(DBG_IOCTL, 0) | DBG_FUNC_END, - (unsigned int)dev, (unsigned int)ap->a_command, (unsigned int)ap->a_fflag, retval, 0); + dev, ap->a_command, ap->a_fflag, retval, 0); return (retval); } @@ -562,27 +725,38 @@ spec_select(struct vnop_select_args *ap) } } +static int filt_specattach(struct knote *kn); + int spec_kqfilter(vnode_t vp, struct knote *kn) { dev_t dev; - int err = EINVAL; + assert(vnode_ischr(vp)); + + dev = vnode_specrdev(vp); + +#if NETWORKING /* - * For a few special kinds of devices, we can attach knotes. - * Each filter function must check whether the dev type matches it. + * Try a bpf device, as defined in bsd/net/bpf.c + * If it doesn't error out the attach, then it + * claimed it. Otherwise, fall through and try + * a regular spec attach. */ - dev = vnode_specrdev(vp); + int32_t tmp_flags = kn->kn_flags; + int64_t tmp_data = kn->kn_data; + int res; - if (vnode_istty(vp)) { - /* We can hook into the slave side of a tty */ - err = ptsd_kqfilter(dev, kn); - } else { - /* Try a bpf device, as defined in bsd/net/bpf.c */ - err = bpfkqfilter(dev, kn); + res = bpfkqfilter(dev, kn); + if ((kn->kn_flags & EV_ERROR) == 0) { + return res; } + kn->kn_flags = tmp_flags; + kn->kn_data = tmp_data; +#endif - return err; + /* Try to attach to other char special devices */ + return filt_specattach(kn); } /* @@ -607,34 +781,12 @@ spec_fsync(struct vnop_fsync_args *ap) return spec_fsync_internal(ap->a_vp, ap->a_waitfor, ap->a_context); } + /* * Just call the device strategy routine */ -extern int hard_throttle_on_root; -void IOSleep(int); - -// the low priority process may wait for at most LOWPRI_MAX_DELAY millisecond -#define LOWPRI_INITIAL_WINDOW_MSECS 100 -#define LOWPRI_WINDOW_MSECS_INC 50 -#define LOWPRI_MAX_WINDOW_MSECS 200 -#define LOWPRI_MAX_WAITING_MSECS 200 -#define LOWPRI_SLEEP_INTERVAL 5 - - -struct _throttle_io_info_t { - struct timeval last_normal_IO_timestamp; - struct timeval last_IO_timestamp; - SInt32 numthreads_throttling; - SInt32 refcnt; - SInt32 alloc; - -}; +void throttle_init(void); -struct _throttle_io_info_t _throttle_io_info[LOWPRI_MAX_NUM_DEV]; -int lowpri_IO_initial_window_msecs = LOWPRI_INITIAL_WINDOW_MSECS; -int lowpri_IO_window_msecs_inc = LOWPRI_WINDOW_MSECS_INC; -int lowpri_max_window_msecs = LOWPRI_MAX_WINDOW_MSECS; -int lowpri_max_waiting_msecs = LOWPRI_MAX_WAITING_MSECS; #if 0 #define DEBUG_ALLOC_THROTTLE_INFO(format, debug_info, args...) \ @@ -647,10 +799,48 @@ int lowpri_max_waiting_msecs = LOWPRI_MAX_WAITING_MSECS; #define DEBUG_ALLOC_THROTTLE_INFO(format, debug_info, args...) #endif -SYSCTL_INT(_debug, OID_AUTO, lowpri_IO_initial_window_msecs, CTLFLAG_RW, &lowpri_IO_initial_window_msecs, LOWPRI_INITIAL_WINDOW_MSECS, ""); -SYSCTL_INT(_debug, OID_AUTO, lowpri_IO_window_inc, CTLFLAG_RW, &lowpri_IO_window_msecs_inc, LOWPRI_INITIAL_WINDOW_MSECS, ""); -SYSCTL_INT(_debug, OID_AUTO, lowpri_max_window_msecs, CTLFLAG_RW, &lowpri_max_window_msecs, LOWPRI_INITIAL_WINDOW_MSECS, ""); -SYSCTL_INT(_debug, OID_AUTO, lowpri_max_waiting_msecs, CTLFLAG_RW, &lowpri_max_waiting_msecs, LOWPRI_INITIAL_WINDOW_MSECS, ""); + +SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier1_window_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_windows_msecs[THROTTLE_LEVEL_TIER1], 0, ""); +SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier2_window_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_windows_msecs[THROTTLE_LEVEL_TIER2], 0, ""); +SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier3_window_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_windows_msecs[THROTTLE_LEVEL_TIER3], 0, ""); + +SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier1_io_period_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_io_period_msecs[THROTTLE_LEVEL_TIER1], 0, ""); +SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier2_io_period_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_io_period_msecs[THROTTLE_LEVEL_TIER2], 0, ""); +SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier3_io_period_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_io_period_msecs[THROTTLE_LEVEL_TIER3], 0, ""); + +SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier1_io_period_ssd_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_io_period_ssd_msecs[THROTTLE_LEVEL_TIER1], 0, ""); +SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier2_io_period_ssd_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_io_period_ssd_msecs[THROTTLE_LEVEL_TIER2], 0, ""); +SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier3_io_period_ssd_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_io_period_ssd_msecs[THROTTLE_LEVEL_TIER3], 0, ""); + +SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_enabled, CTLFLAG_RW | CTLFLAG_LOCKED, &lowpri_throttle_enabled, 0, ""); + + +static lck_grp_t *throttle_lock_grp; +static lck_attr_t *throttle_lock_attr; +static lck_grp_attr_t *throttle_lock_grp_attr; + + +/* + * throttled I/O helper function + * convert the index of the lowest set bit to a device index + */ +int +num_trailing_0(uint64_t n) +{ + /* + * since in most cases the number of trailing 0s is very small, + * we simply counting sequentially from the lowest bit + */ + if (n == 0) + return sizeof(n) * 8; + int count = 0; + while (!ISSET(n, 1)) { + n >>= 1; + ++count; + } + return count; +} + /* * Release the reference and if the item was allocated and this is the last @@ -661,7 +851,7 @@ SYSCTL_INT(_debug, OID_AUTO, lowpri_max_waiting_msecs, CTLFLAG_RW, &lowpri_max_w static int throttle_info_rel(struct _throttle_io_info_t *info) { - SInt32 oldValue = OSDecrementAtomic(&info->refcnt); + SInt32 oldValue = OSDecrementAtomic(&info->throttle_refcnt); DEBUG_ALLOC_THROTTLE_INFO("refcnt = %d info = %p\n", info, (int)(oldValue -1), info ); @@ -674,13 +864,16 @@ throttle_info_rel(struct _throttle_io_info_t *info) * Once reference count is zero, no one else should be able to take a * reference */ - if ((info->refcnt == 0) && (info->alloc)) { - DEBUG_ALLOC_THROTTLE_INFO("Freeing info = %p\n", info, info ); + if ((info->throttle_refcnt == 0) && (info->throttle_alloc)) { + DEBUG_ALLOC_THROTTLE_INFO("Freeing info = %p\n", info); + + lck_mtx_destroy(&info->throttle_lock, throttle_lock_grp); FREE(info, M_TEMP); } return oldValue; } + /* * Just take a reference on the throttle info structure. * @@ -689,17 +882,503 @@ throttle_info_rel(struct _throttle_io_info_t *info) static SInt32 throttle_info_ref(struct _throttle_io_info_t *info) { - SInt32 oldValue = OSIncrementAtomic(&info->refcnt); + SInt32 oldValue = OSIncrementAtomic(&info->throttle_refcnt); DEBUG_ALLOC_THROTTLE_INFO("refcnt = %d info = %p\n", info, (int)(oldValue -1), info ); /* Allocated items should never have a reference of zero */ - if (info->alloc && (oldValue == 0)) + if (info->throttle_alloc && (oldValue == 0)) panic("Taking a reference without calling create throttle info!\n"); return oldValue; } +/* + * on entry the throttle_lock is held... + * this function is responsible for taking + * and dropping the reference on the info + * structure which will keep it from going + * away while the timer is running if it + * happens to have been dynamically allocated by + * a network fileystem kext which is now trying + * to free it + */ +static uint32_t +throttle_timer_start(struct _throttle_io_info_t *info, boolean_t update_io_count, int wakelevel) +{ + struct timeval elapsed; + struct timeval now; + struct timeval period; + uint64_t elapsed_msecs; + int throttle_level; + int level; + int msecs; + boolean_t throttled = FALSE; + boolean_t need_timer = FALSE; + + microuptime(&now); + + if (update_io_count == TRUE) { + info->throttle_io_count_begin = info->throttle_io_count; + info->throttle_io_period_num++; + + while (wakelevel >= THROTTLE_LEVEL_THROTTLED) + info->throttle_start_IO_period_timestamp[wakelevel--] = now; + + info->throttle_min_timer_deadline = now; + + msecs = info->throttle_io_periods[THROTTLE_LEVEL_THROTTLED]; + period.tv_sec = msecs / 1000; + period.tv_usec = (msecs % 1000) * 1000; + + timevaladd(&info->throttle_min_timer_deadline, &period); + } + for (throttle_level = THROTTLE_LEVEL_START; throttle_level < THROTTLE_LEVEL_END; throttle_level++) { + + 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); + + for (level = throttle_level + 1; level <= THROTTLE_LEVEL_END; level++) { + + if (!TAILQ_EMPTY(&info->throttle_uthlist[level])) { + + if (elapsed_msecs < (uint64_t)throttle_windows_msecs[level] || info->throttle_inflight_count[throttle_level]) { + /* + * we had an I/O occur at a higher priority tier within + * this tier's throttle window + */ + throttled = TRUE; + } + /* + * we assume that the windows are the same or longer + * as we drop through the throttling tiers... thus + * we can stop looking once we run into a tier with + * threads to schedule regardless of whether it's + * still in its throttling window or not + */ + break; + } + } + if (throttled == TRUE) + break; + } + if (throttled == TRUE) { + uint64_t deadline = 0; + struct timeval target; + struct timeval min_target; + + /* + * we've got at least one tier still in a throttled window + * so we need a timer running... compute the next deadline + * and schedule it + */ + for (level = throttle_level+1; level <= THROTTLE_LEVEL_END; level++) { + + if (TAILQ_EMPTY(&info->throttle_uthlist[level])) + continue; + + target = info->throttle_start_IO_period_timestamp[level]; + + msecs = info->throttle_io_periods[level]; + period.tv_sec = msecs / 1000; + period.tv_usec = (msecs % 1000) * 1000; + + 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 * @@ -711,6 +1390,7 @@ 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 */ @@ -718,9 +1398,18 @@ throttle_info_create(void) return NULL; /* Mark that this one was allocated and needs to be freed */ DEBUG_ALLOC_THROTTLE_INFO("Creating info = %p\n", info, info ); - info->alloc = TRUE; + 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->refcnt); + OSIncrementAtomic(&info->throttle_refcnt); return info; } @@ -754,20 +1443,60 @@ 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 */ + + /* + * 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; } /* - * KPI routine + * Private 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) + * 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); @@ -780,13 +1509,13 @@ 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]; + info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1]; else if (mp->mnt_throttle_info == NULL) - info = &_throttle_io_info[mp->mnt_devbsdunit]; + info = &_throttle_io_info[mp->mnt_devbsdunit]; else - info = mp->mnt_throttle_info; + info = mp->mnt_throttle_info; - *tv = info->last_IO_timestamp; + *tv = info->throttle_last_write_timestamp; } void @@ -795,30 +1524,135 @@ 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]; + info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1]; else if (mp->mnt_throttle_info == NULL) - info = &_throttle_io_info[mp->mnt_devbsdunit]; + info = &_throttle_io_info[mp->mnt_devbsdunit]; else - info = mp->mnt_throttle_info; + 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--; + } + } - microuptime(&info->last_IO_timestamp); + 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(int lowpri_window_msecs, void * throttle_info) +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; - int elapsed_msecs; + 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); - microuptime(&elapsed); - timevalsub(&elapsed, &info->last_normal_IO_timestamp); - elapsed_msecs = elapsed.tv_sec * 1000 + elapsed.tv_usec / 1000; + 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 (lowpri_window_msecs == -1) // use the max waiting time - lowpri_window_msecs = lowpri_max_waiting_msecs; + 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; - return elapsed_msecs < lowpri_window_msecs; + 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); } /* @@ -827,118 +1661,349 @@ throttle_io_will_be_throttled_internal(int lowpri_window_msecs, void * throttle_ * the correct throttle info array element. */ int -throttle_io_will_be_throttled(int lowpri_window_msecs, mount_t mp) +throttle_io_will_be_throttled(__unused int lowpri_window_msecs, mount_t mp) { - void *info; + struct _throttle_io_info_t *info; - /* Should we just return zero if no mount point */ + /* + * Should we just return zero if no mount point + */ if (mp == NULL) - info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1]; + info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1]; else if (mp->mnt_throttle_info == NULL) - info = &_throttle_io_info[mp->mnt_devbsdunit]; + info = &_throttle_io_info[mp->mnt_devbsdunit]; else - info = mp->mnt_throttle_info; - return throttle_io_will_be_throttled_internal(lowpri_window_msecs, info); + 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); } -void throttle_lowpri_io(boolean_t ok_to_sleep) +/* + * Routine to increment I/O throttling counters maintained in the proc + */ + +static void +throttle_update_proc_stats(pid_t throttling_pid, int count) { - int i; - int max_try_num; - struct uthread *ut; + 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) || (ut->uu_throttle_info == NULL)) - goto done; + if (ut->uu_lowpri_window == 0) + return (0); info = ut->uu_throttle_info; - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 97)) | DBG_FUNC_START, - ut->uu_lowpri_window, ok_to_sleep, 0, 0, 0); - if (ok_to_sleep == TRUE) { - max_try_num = lowpri_max_waiting_msecs / LOWPRI_SLEEP_INTERVAL * MAX(1, info->numthreads_throttling); + 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; - for (i=0; iuu_lowpri_window, info)) { - IOSleep(LOWPRI_SLEEP_INTERVAL); - DEBUG_ALLOC_THROTTLE_INFO("sleeping because of info = %p\n", info, info ); - } else { + 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; - } } - } - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 97)) | DBG_FUNC_END, - ut->uu_lowpri_window, i*5, 0, 0, 0); - SInt32 oldValue; - oldValue = OSDecrementAtomic(&info->numthreads_throttling); + /* + * 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"; - if (oldValue <= 0) { - panic("%s: numthreads negative", __func__); + 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: - ut->uu_lowpri_window = 0; - if (ut->uu_throttle_info) - throttle_info_rel(ut->uu_throttle_info); + 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); } -int throttle_get_io_policy(struct uthread **ut) +/* + * 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) { - int policy = IOPOL_DEFAULT; - proc_t p = current_proc(); + proc_set_thread_policy(current_thread(), TASK_POLICY_INTERNAL, TASK_POLICY_IOPOL, policy); +} - *ut = get_bsdthread_info(current_thread()); - - if (p != NULL) - policy = p->p_iopol_disk; +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); - if (*ut != NULL) { - // the I/O policy of the thread overrides that of the process - // unless the I/O policy of the thread is default - if ((*ut)->uu_iopol_disk != IOPOL_DEFAULT) - policy = (*ut)->uu_iopol_disk; + ut->uu_throttle_info = NULL; + ut->uu_lowpri_window = 0; + ut->uu_throttle_bc = FALSE; } - return policy; } -void throttle_info_update(void *throttle_info, int flags) +static +void throttle_info_set_initial_window(uthread_t ut, struct _throttle_io_info_t *info, boolean_t BC_throttle, boolean_t isssd) { - struct _throttle_io_info_t *info = throttle_info; - struct uthread *ut; - int policy; - int is_throttleable_io = 0; - int is_passive_io = 0; - SInt32 oldValue; - - if (!lowpri_IO_initial_window_msecs || (info == NULL)) + if (lowpri_throttle_enabled == 0 || info->throttle_disabled) return; - policy = throttle_get_io_policy(&ut); - switch (policy) { - case IOPOL_DEFAULT: - case IOPOL_NORMAL: - break; - case IOPOL_THROTTLE: - is_throttleable_io = 1; - break; - case IOPOL_PASSIVE: - is_passive_io = 1; - break; - default: - printf("unknown I/O policy %d", policy); - break; + if (info->throttle_io_periods == 0) { + throttle_init_throttle_period(info, isssd); } + if (ut->uu_throttle_info == NULL) { - if (!is_throttleable_io && ISSET(flags, B_PASSIVE)) - is_passive_io |= 1; + 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; + } +} - if (!is_throttleable_io) { - if (!is_passive_io){ - microuptime(&info->last_normal_IO_timestamp); +/* + * 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 if (ut) { + } 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 @@ -949,101 +2014,334 @@ void throttle_info_update(void *throttle_info, int flags) * do the delay just before we return from the system * call that triggered this I/O or from vnode_pagein */ - if (ut->uu_lowpri_window == 0) { - ut->uu_throttle_info = info; - throttle_info_ref(ut->uu_throttle_info); - DEBUG_ALLOC_THROTTLE_INFO("updating info = %p\n", info, info ); - - oldValue = OSIncrementAtomic(&info->numthreads_throttling); - if (oldValue < 0) { - panic("%s: numthreads negative", __func__); - } - ut->uu_lowpri_window = lowpri_IO_initial_window_msecs; - ut->uu_lowpri_window += lowpri_IO_window_msecs_inc * oldValue; - } else { - /* The thread sends I/Os to different devices within the same system call */ - if (ut->uu_throttle_info != info) { - struct _throttle_io_info_t *old_info = ut->uu_throttle_info; - - // keep track of the numthreads in the right device - OSDecrementAtomic(&old_info->numthreads_throttling); - OSIncrementAtomic(&info->numthreads_throttling); - - DEBUG_ALLOC_THROTTLE_INFO("switching from info = %p\n", old_info, old_info ); - DEBUG_ALLOC_THROTTLE_INFO("switching to info = %p\n", info, info ); - /* This thread no longer needs a reference on that throttle info */ - throttle_info_rel(ut->uu_throttle_info); - ut->uu_throttle_info = info; - /* Need to take a reference on this throttle info */ - throttle_info_ref(ut->uu_throttle_info); - } - int numthreads = MAX(1, info->numthreads_throttling); - ut->uu_lowpri_window += lowpri_IO_window_msecs_inc * numthreads; - if (ut->uu_lowpri_window > lowpri_max_window_msecs * numthreads) - ut->uu_lowpri_window = lowpri_max_window_msecs * numthreads; + 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 ((mp->mnt_kern_flag & MNTK_SSD) && !ignore_is_ssd) + 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; } int spec_strategy(struct vnop_strategy_args *ap) { - buf_t bp; + buf_t bp; int bflags; - int policy; + int io_tier; + int passive; dev_t bdev; uthread_t ut; - size_t devbsdunit; 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; + + proc_t curproc = current_proc(); bp = ap->a_bp; bdev = buf_device(bp); - bflags = buf_flags(bp); mp = buf_vnode(bp)->v_mount; + bap = &bp->b_attr; - if (kdebug_enable) { - int code = 0; + io_tier = throttle_get_io_policy(&ut); + passive = throttle_get_passive_io_policy(&ut); - if (bflags & B_READ) - code |= DKIO_READ; - if (bflags & B_ASYNC) - code |= DKIO_ASYNC; + /* + * 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 (bflags & B_META) - code |= DKIO_META; - else if (bflags & B_PAGEIO) - code |= DKIO_PAGING; + if (bp->b_flags & B_META) + bap->ba_flags |= BA_META; - KERNEL_DEBUG_CONSTANT(FSDBG_CODE(DBG_DKRW, code) | DBG_FUNC_NONE, - bp, bdev, (int)buf_blkno(bp), buf_count(bp), 0); - } - if (((bflags & (B_IOSTREAMING | B_PAGEIO | B_READ)) == (B_PAGEIO | B_READ)) && - mp && (mp->mnt_kern_flag & MNTK_ROOTDEV)) - hard_throttle_on_root = 1; +#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 (mp != NULL) - devbsdunit = mp->mnt_devbsdunit; - else - devbsdunit = LOWPRI_MAX_NUM_DEV - 1; + if ((curproc != NULL) && ((curproc->p_flag & P_DELAYIDLESLEEP) == P_DELAYIDLESLEEP)) + bap->ba_flags |= BA_DELAYIDLESLEEP; + + 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); - throttle_info_update(&_throttle_io_info[devbsdunit], bflags); - if ((policy = throttle_get_io_policy(&ut)) == IOPOL_THROTTLE) { - bp->b_flags |= B_THROTTLED_IO; + 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 ((mp->mnt_kern_flag & MNTK_SSD) && !ignore_is_ssd) + 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_io_info[devbsdunit].last_IO_timestamp); + 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); } - - (*bdevsw[major(bdev)].d_strategy)(bp); + /* + * 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); + + 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); } @@ -1066,8 +2364,7 @@ spec_close(struct vnop_close_args *ap) { struct vnode *vp = ap->a_vp; dev_t dev = vp->v_rdev; - int (*devclose)(dev_t, int, int, struct proc *); - int mode, error; + int error = 0; int flags = ap->a_fflag; struct proc *p = vfs_context_proc(ap->a_context); struct session *sessp; @@ -1082,44 +2379,66 @@ spec_close(struct vnop_close_args *ap) * 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) + * last close) */ sessp = proc_session(p); + devsw_lock(dev, S_IFCHR); if (sessp != SESSION_NULL) { - if ((vcount(vp) == 1) && - (vp == sessp->s_ttyvp)) { + if (vp == sessp->s_ttyvp && vcount(vp) == 1) { + struct tty *tp = TTY_NULL; + + devsw_unlock(dev, S_IFCHR); session_lock(sessp); - sessp->s_ttyvp = NULL; - sessp->s_ttyvid = 0; - sessp->s_ttyp = TTY_NULL; - sessp->s_ttypgrpid = NO_PID; + 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); - vnode_rele(vp); + + 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); } - devclose = cdevsw[major(dev)].d_close; - mode = S_IFCHR; + 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 ((flags & IO_REVOKE) != 0) - break; - if (vcount(vp) > 0) - return (0); + 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: /* - * Since every use (buffer, vnode, swap, blockmap) - * holds a reference to the vnode, and because we mark - * any other vnodes that alias this device, when the - * sum of the reference counts on all the aliased - * vnodes descends to zero, we are on last close. + * If there is more than one outstanding open, don't + * send the close to the device. */ - if (vcount(vp) > 0) + 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) @@ -1133,8 +2452,15 @@ spec_close(struct vnop_close_args *ap) if (error) return (error); - devclose = bdevsw[major(dev)].d_close; - mode = S_IFBLK; + 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: @@ -1142,7 +2468,7 @@ spec_close(struct vnop_close_args *ap) return(EBADF); } - return ((*devclose)(dev, flags, mode, p)); + return error; } /* @@ -1234,3 +2560,298 @@ spec_offtoblk(struct vnop_offtoblk_args *ap) return (0); } + +static void filt_specdetach(struct knote *kn); +static int filt_spec(struct knote *kn, long hint); +static int filt_spectouch(struct knote *kn, struct kevent_internal_s *kev); +static int filt_specprocess(struct knote *kn, struct filt_process_s *data, struct kevent_internal_s *kev); +static unsigned filt_specpeek(struct knote *kn); + +struct filterops spec_filtops = { + .f_isfd = 1, + .f_attach = filt_specattach, + .f_detach = filt_specdetach, + .f_event = filt_spec, + .f_touch = filt_spectouch, + .f_process = filt_specprocess, + .f_peek = filt_specpeek +}; + +static int +filter_to_seltype(int16_t filter) +{ + switch (filter) { + case EVFILT_READ: + return FREAD; + case EVFILT_WRITE: + return FWRITE; + default: + panic("filt_to_seltype(): invalid filter %d\n", filter); + return 0; + } +} + +static int +filt_specattach(struct knote *kn) +{ + vnode_t vp; + dev_t dev; + + vp = (vnode_t)kn->kn_fp->f_fglob->fg_data; /* Already have iocount, and vnode is alive */ + + assert(vnode_ischr(vp)); + + dev = vnode_specrdev(vp); + + if (major(dev) > nchrdev) { + kn->kn_flags |= EV_ERROR; + kn->kn_data = ENXIO; + return 0; + } + + /* + * For a few special kinds of devices, we can attach knotes with + * no restrictions because their "select" vectors return the amount + * of data available. Others require an explicit NOTE_LOWAT with + * data of 1, indicating that the caller doesn't care about actual + * data counts, just an indication that the device has data. + */ + + if ((cdevsw_flags[major(dev)] & CDEVSW_SELECT_KQUEUE) == 0 && + ((kn->kn_sfflags & NOTE_LOWAT) == 0 || kn->kn_sdata != 1)) { + kn->kn_flags |= EV_ERROR; + kn->kn_data = EINVAL; + return 0; + } + + kn->kn_hook_data = 0; + + kn->kn_filtid = EVFILTID_SPEC; + kn->kn_hookid = vnode_vid(vp); + + knote_markstayactive(kn); + + return 0; +} + +static void +filt_specdetach(struct knote *kn) +{ + knote_clearstayactive(kn); + + /* + * This is potentially tricky: the device's selinfo waitq that was + * tricked into being part of this knote's waitq set may not be a part + * of any other set, and the device itself may have revoked the memory + * in which the waitq was held. We use the knote's kn_hook_data field + * to keep the ID of the waitq's prepost table object. This + * object keeps a pointer back to the waitq, and gives us a safe way + * to decouple the dereferencing of driver allocated memory: if the + * driver goes away (taking the waitq with it) then the prepost table + * object will be invalidated. The waitq details are handled in the + * waitq API invoked here. + */ + if (kn->kn_hook_data) { + waitq_unlink_by_prepost_id(kn->kn_hook_data, &(knote_get_kq(kn)->kq_wqs)); + kn->kn_hook_data = 0; + } +} + +static int +filt_spec(__unused struct knote *kn, __unused long hint) +{ + panic("filt_spec()"); + return 0; +} + + + +static int +filt_spectouch(struct knote *kn, struct kevent_internal_s *kev) +{ + kn->kn_sdata = kev->data; + kn->kn_sfflags = kev->fflags; + if ((kn->kn_status & KN_UDATA_SPECIFIC) == 0) + kn->kn_udata = kev->udata; + + /* stayqueued knotes don't need hints from touch */ + return 0; +} + +static int +filt_specprocess(struct knote *kn, struct filt_process_s *data, struct kevent_internal_s *kev) +{ +#pragma unused(data) + vnode_t vp; + uthread_t uth; + struct waitq_set *old_wqs; + vfs_context_t ctx; + int res; + int selres; + int error; + int use_offset; + dev_t dev; + uint64_t flags; + uint64_t rsvd, rsvd_arg; + uint64_t *rlptr = NULL; + + uth = get_bsdthread_info(current_thread()); + ctx = vfs_context_current(); + vp = (vnode_t)kn->kn_fp->f_fglob->fg_data; + + /* JMM - locking against touches? */ + + error = vnode_getwithvid(vp, kn->kn_hookid); + if (error != 0) { + kn->kn_flags |= (EV_EOF | EV_ONESHOT); + *kev = kn->kn_kevent; + return 1; + } + + dev = vnode_specrdev(vp); + flags = cdevsw_flags[major(dev)]; + use_offset = ((flags & CDEVSW_USE_OFFSET) != 0); + + /* + * This function may be called many times to link or re-link the + * underlying vnode to the kqueue. If we've already linked the two, + * we will have a valid kn_hook_data which ties us to the underlying + * device's waitq via a the waitq's prepost table object. However, + * devices can abort any select action by calling selthreadclear(). + * This is OK because the table object will be invalidated by the + * driver (through a call to selthreadclear), so any attempt to access + * the associated waitq will fail because the table object is invalid. + * + * Even if we've already registered, we need to pass a pointer + * to a reserved link structure. Otherwise, selrecord() will + * infer that we're in the second pass of select() and won't + * actually do anything! + */ + rsvd = rsvd_arg = waitq_link_reserve(NULL); + rlptr = (void *)&rsvd_arg; + + /* + * Trick selrecord() into hooking kqueue's wait queue set + * set into device's selinfo wait queue + */ + old_wqs = uth->uu_wqset; + uth->uu_wqset = &(knote_get_kq(kn)->kq_wqs); + selres = VNOP_SELECT(vp, filter_to_seltype(kn->kn_filter), + 0, rlptr, ctx); + uth->uu_wqset = old_wqs; + + /* + * make sure to cleanup the reserved link - this guards against + * drivers that may not actually call selrecord(). + */ + waitq_link_release(rsvd); + if (rsvd != rsvd_arg) { + /* the driver / handler called selrecord() */ + struct waitq *wq; + memcpy(&wq, rlptr, sizeof(void *)); + + /* + * The waitq_get_prepost_id() function will (potentially) + * allocate a prepost table object for the waitq and return + * the table object's ID to us. It will also set the + * waitq_prepost_id field within the waitq structure. + * + * We can just overwrite kn_hook_data because it's simply a + * table ID used to grab a reference when needed. + * + * We have a reference on the vnode, so we know that the + * device won't go away while we get this ID. + */ + kn->kn_hook_data = waitq_get_prepost_id(wq); + } + + if (use_offset) { + if (kn->kn_fp->f_fglob->fg_offset >= (uint32_t)selres) { + kn->kn_data = 0; + } else { + kn->kn_data = ((uint32_t)selres) - kn->kn_fp->f_fglob->fg_offset; + } + } else { + kn->kn_data = selres; + } + + vnode_put(vp); + + res = ((kn->kn_sfflags & NOTE_LOWAT) != 0) ? + (kn->kn_data >= kn->kn_sdata) : kn->kn_data; + + if (res) { + *kev = kn->kn_kevent; + if (kn->kn_flags & EV_CLEAR) { + kn->kn_fflags = 0; + kn->kn_data = 0; + } + } + + return res; +} + +static unsigned +filt_specpeek(struct knote *kn) +{ + vnode_t vp; + uthread_t uth; + struct waitq_set *old_wqs; + vfs_context_t ctx; + int error, selres; + uint64_t rsvd, rsvd_arg; + uint64_t *rlptr = NULL; + + uth = get_bsdthread_info(current_thread()); + ctx = vfs_context_current(); + vp = (vnode_t)kn->kn_fp->f_fglob->fg_data; + + error = vnode_getwithvid(vp, kn->kn_hookid); + if (error != 0) { + return 1; /* Just like VNOP_SELECT() on recycled vnode */ + } + + /* + * Even if we've already registered, we need to pass a pointer + * to a reserved link structure. Otherwise, selrecord() will + * infer that we're in the second pass of select() and won't + * actually do anything! + */ + rsvd = rsvd_arg = waitq_link_reserve(NULL); + rlptr = (void *)&rsvd_arg; + + old_wqs = uth->uu_wqset; + uth->uu_wqset = &(knote_get_kq(kn)->kq_wqs); + selres = VNOP_SELECT(vp, filter_to_seltype(kn->kn_filter), + 0, (void *)rlptr, ctx); + uth->uu_wqset = old_wqs; + + /* + * make sure to cleanup the reserved link - this guards against + * drivers that may not actually call selrecord() + */ + waitq_link_release(rsvd); + if (rsvd != rsvd_arg) { + /* the driver / handler called selrecord() */ + struct waitq *wq; + memcpy(&wq, rlptr, sizeof(void *)); + + /* + * The waitq_get_prepost_id() function will (potentially) + * allocate a prepost table object for the waitq and return + * the table object's ID to us. It will also set the + * waitq_prepost_id field within the waitq structure. + * + * We can just overwrite kn_hook_data because it's simply a + * table ID used to grab a reference when needed. + * + * We have a reference on the vnode, so we know that the + * device won't go away while we get this ID. + */ + kn->kn_hook_data = waitq_get_prepost_id(wq); + } + + vnode_put(vp); + return selres; +} +