1992 lines · c
1// SPDX-License-Identifier: GPL-2.02/*3 * Copyright (c) 2010 Red Hat, Inc. All Rights Reserved.4 */5 6#include "xfs.h"7#include "xfs_fs.h"8#include "xfs_format.h"9#include "xfs_log_format.h"10#include "xfs_shared.h"11#include "xfs_trans_resv.h"12#include "xfs_mount.h"13#include "xfs_extent_busy.h"14#include "xfs_trans.h"15#include "xfs_trans_priv.h"16#include "xfs_log.h"17#include "xfs_log_priv.h"18#include "xfs_trace.h"19#include "xfs_discard.h"20 21/*22 * Allocate a new ticket. Failing to get a new ticket makes it really hard to23 * recover, so we don't allow failure here. Also, we allocate in a context that24 * we don't want to be issuing transactions from, so we need to tell the25 * allocation code this as well.26 *27 * We don't reserve any space for the ticket - we are going to steal whatever28 * space we require from transactions as they commit. To ensure we reserve all29 * the space required, we need to set the current reservation of the ticket to30 * zero so that we know to steal the initial transaction overhead from the31 * first transaction commit.32 */33static struct xlog_ticket *34xlog_cil_ticket_alloc(35 struct xlog *log)36{37 struct xlog_ticket *tic;38 39 tic = xlog_ticket_alloc(log, 0, 1, 0);40 41 /*42 * set the current reservation to zero so we know to steal the basic43 * transaction overhead reservation from the first transaction commit.44 */45 tic->t_curr_res = 0;46 tic->t_iclog_hdrs = 0;47 return tic;48}49 50static inline void51xlog_cil_set_iclog_hdr_count(struct xfs_cil *cil)52{53 struct xlog *log = cil->xc_log;54 55 atomic_set(&cil->xc_iclog_hdrs,56 (XLOG_CIL_BLOCKING_SPACE_LIMIT(log) /57 (log->l_iclog_size - log->l_iclog_hsize)));58}59 60/*61 * Check if the current log item was first committed in this sequence.62 * We can't rely on just the log item being in the CIL, we have to check63 * the recorded commit sequence number.64 *65 * Note: for this to be used in a non-racy manner, it has to be called with66 * CIL flushing locked out. As a result, it should only be used during the67 * transaction commit process when deciding what to format into the item.68 */69static bool70xlog_item_in_current_chkpt(71 struct xfs_cil *cil,72 struct xfs_log_item *lip)73{74 if (test_bit(XLOG_CIL_EMPTY, &cil->xc_flags))75 return false;76 77 /*78 * li_seq is written on the first commit of a log item to record the79 * first checkpoint it is written to. Hence if it is different to the80 * current sequence, we're in a new checkpoint.81 */82 return lip->li_seq == READ_ONCE(cil->xc_current_sequence);83}84 85bool86xfs_log_item_in_current_chkpt(87 struct xfs_log_item *lip)88{89 return xlog_item_in_current_chkpt(lip->li_log->l_cilp, lip);90}91 92/*93 * Unavoidable forward declaration - xlog_cil_push_work() calls94 * xlog_cil_ctx_alloc() itself.95 */96static void xlog_cil_push_work(struct work_struct *work);97 98static struct xfs_cil_ctx *99xlog_cil_ctx_alloc(void)100{101 struct xfs_cil_ctx *ctx;102 103 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL | __GFP_NOFAIL);104 INIT_LIST_HEAD(&ctx->committing);105 INIT_LIST_HEAD(&ctx->busy_extents.extent_list);106 INIT_LIST_HEAD(&ctx->log_items);107 INIT_LIST_HEAD(&ctx->lv_chain);108 INIT_WORK(&ctx->push_work, xlog_cil_push_work);109 return ctx;110}111 112/*113 * Aggregate the CIL per cpu structures into global counts, lists, etc and114 * clear the percpu state ready for the next context to use. This is called115 * from the push code with the context lock held exclusively, hence nothing else116 * will be accessing or modifying the per-cpu counters.117 */118static void119xlog_cil_push_pcp_aggregate(120 struct xfs_cil *cil,121 struct xfs_cil_ctx *ctx)122{123 struct xlog_cil_pcp *cilpcp;124 int cpu;125 126 for_each_cpu(cpu, &ctx->cil_pcpmask) {127 cilpcp = per_cpu_ptr(cil->xc_pcp, cpu);128 129 ctx->ticket->t_curr_res += cilpcp->space_reserved;130 cilpcp->space_reserved = 0;131 132 if (!list_empty(&cilpcp->busy_extents)) {133 list_splice_init(&cilpcp->busy_extents,134 &ctx->busy_extents.extent_list);135 }136 if (!list_empty(&cilpcp->log_items))137 list_splice_init(&cilpcp->log_items, &ctx->log_items);138 139 /*140 * We're in the middle of switching cil contexts. Reset the141 * counter we use to detect when the current context is nearing142 * full.143 */144 cilpcp->space_used = 0;145 }146}147 148/*149 * Aggregate the CIL per-cpu space used counters into the global atomic value.150 * This is called when the per-cpu counter aggregation will first pass the soft151 * limit threshold so we can switch to atomic counter aggregation for accurate152 * detection of hard limit traversal.153 */154static void155xlog_cil_insert_pcp_aggregate(156 struct xfs_cil *cil,157 struct xfs_cil_ctx *ctx)158{159 int cpu;160 int count = 0;161 162 /* Trigger atomic updates then aggregate only for the first caller */163 if (!test_and_clear_bit(XLOG_CIL_PCP_SPACE, &cil->xc_flags))164 return;165 166 /*167 * We can race with other cpus setting cil_pcpmask. However, we've168 * atomically cleared PCP_SPACE which forces other threads to add to169 * the global space used count. cil_pcpmask is a superset of cilpcp170 * structures that could have a nonzero space_used.171 */172 for_each_cpu(cpu, &ctx->cil_pcpmask) {173 struct xlog_cil_pcp *cilpcp = per_cpu_ptr(cil->xc_pcp, cpu);174 int old = READ_ONCE(cilpcp->space_used);175 176 while (!try_cmpxchg(&cilpcp->space_used, &old, 0))177 ;178 count += old;179 }180 atomic_add(count, &ctx->space_used);181}182 183static void184xlog_cil_ctx_switch(185 struct xfs_cil *cil,186 struct xfs_cil_ctx *ctx)187{188 xlog_cil_set_iclog_hdr_count(cil);189 set_bit(XLOG_CIL_EMPTY, &cil->xc_flags);190 set_bit(XLOG_CIL_PCP_SPACE, &cil->xc_flags);191 ctx->sequence = ++cil->xc_current_sequence;192 ctx->cil = cil;193 cil->xc_ctx = ctx;194}195 196/*197 * After the first stage of log recovery is done, we know where the head and198 * tail of the log are. We need this log initialisation done before we can199 * initialise the first CIL checkpoint context.200 *201 * Here we allocate a log ticket to track space usage during a CIL push. This202 * ticket is passed to xlog_write() directly so that we don't slowly leak log203 * space by failing to account for space used by log headers and additional204 * region headers for split regions.205 */206void207xlog_cil_init_post_recovery(208 struct xlog *log)209{210 log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log);211 log->l_cilp->xc_ctx->sequence = 1;212 xlog_cil_set_iclog_hdr_count(log->l_cilp);213}214 215static inline int216xlog_cil_iovec_space(217 uint niovecs)218{219 return round_up((sizeof(struct xfs_log_vec) +220 niovecs * sizeof(struct xfs_log_iovec)),221 sizeof(uint64_t));222}223 224/*225 * Allocate or pin log vector buffers for CIL insertion.226 *227 * The CIL currently uses disposable buffers for copying a snapshot of the228 * modified items into the log during a push. The biggest problem with this is229 * the requirement to allocate the disposable buffer during the commit if:230 * a) does not exist; or231 * b) it is too small232 *233 * If we do this allocation within xlog_cil_insert_format_items(), it is done234 * under the xc_ctx_lock, which means that a CIL push cannot occur during235 * the memory allocation. This means that we have a potential deadlock situation236 * under low memory conditions when we have lots of dirty metadata pinned in237 * the CIL and we need a CIL commit to occur to free memory.238 *239 * To avoid this, we need to move the memory allocation outside the240 * xc_ctx_lock, but because the log vector buffers are disposable, that opens241 * up a TOCTOU race condition w.r.t. the CIL committing and removing the log242 * vector buffers between the check and the formatting of the item into the243 * log vector buffer within the xc_ctx_lock.244 *245 * Because the log vector buffer needs to be unchanged during the CIL push246 * process, we cannot share the buffer between the transaction commit (which247 * modifies the buffer) and the CIL push context that is writing the changes248 * into the log. This means skipping preallocation of buffer space is249 * unreliable, but we most definitely do not want to be allocating and freeing250 * buffers unnecessarily during commits when overwrites can be done safely.251 *252 * The simplest solution to this problem is to allocate a shadow buffer when a253 * log item is committed for the second time, and then to only use this buffer254 * if necessary. The buffer can remain attached to the log item until such time255 * it is needed, and this is the buffer that is reallocated to match the size of256 * the incoming modification. Then during the formatting of the item we can swap257 * the active buffer with the new one if we can't reuse the existing buffer. We258 * don't free the old buffer as it may be reused on the next modification if259 * it's size is right, otherwise we'll free and reallocate it at that point.260 *261 * This function builds a vector for the changes in each log item in the262 * transaction. It then works out the length of the buffer needed for each log263 * item, allocates them and attaches the vector to the log item in preparation264 * for the formatting step which occurs under the xc_ctx_lock.265 *266 * While this means the memory footprint goes up, it avoids the repeated267 * alloc/free pattern that repeated modifications of an item would otherwise268 * cause, and hence minimises the CPU overhead of such behaviour.269 */270static void271xlog_cil_alloc_shadow_bufs(272 struct xlog *log,273 struct xfs_trans *tp)274{275 struct xfs_log_item *lip;276 277 list_for_each_entry(lip, &tp->t_items, li_trans) {278 struct xfs_log_vec *lv;279 int niovecs = 0;280 int nbytes = 0;281 int buf_size;282 bool ordered = false;283 284 /* Skip items which aren't dirty in this transaction. */285 if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))286 continue;287 288 /* get number of vecs and size of data to be stored */289 lip->li_ops->iop_size(lip, &niovecs, &nbytes);290 291 /*292 * Ordered items need to be tracked but we do not wish to write293 * them. We need a logvec to track the object, but we do not294 * need an iovec or buffer to be allocated for copying data.295 */296 if (niovecs == XFS_LOG_VEC_ORDERED) {297 ordered = true;298 niovecs = 0;299 nbytes = 0;300 }301 302 /*303 * We 64-bit align the length of each iovec so that the start of304 * the next one is naturally aligned. We'll need to account for305 * that slack space here.306 *307 * We also add the xlog_op_header to each region when308 * formatting, but that's not accounted to the size of the item309 * at this point. Hence we'll need an addition number of bytes310 * for each vector to hold an opheader.311 *312 * Then round nbytes up to 64-bit alignment so that the initial313 * buffer alignment is easy to calculate and verify.314 */315 nbytes += niovecs *316 (sizeof(uint64_t) + sizeof(struct xlog_op_header));317 nbytes = round_up(nbytes, sizeof(uint64_t));318 319 /*320 * The data buffer needs to start 64-bit aligned, so round up321 * that space to ensure we can align it appropriately and not322 * overrun the buffer.323 */324 buf_size = nbytes + xlog_cil_iovec_space(niovecs);325 326 /*327 * if we have no shadow buffer, or it is too small, we need to328 * reallocate it.329 */330 if (!lip->li_lv_shadow ||331 buf_size > lip->li_lv_shadow->lv_size) {332 /*333 * We free and allocate here as a realloc would copy334 * unnecessary data. We don't use kvzalloc() for the335 * same reason - we don't need to zero the data area in336 * the buffer, only the log vector header and the iovec337 * storage.338 */339 kvfree(lip->li_lv_shadow);340 lv = xlog_kvmalloc(buf_size);341 342 memset(lv, 0, xlog_cil_iovec_space(niovecs));343 344 INIT_LIST_HEAD(&lv->lv_list);345 lv->lv_item = lip;346 lv->lv_size = buf_size;347 if (ordered)348 lv->lv_buf_len = XFS_LOG_VEC_ORDERED;349 else350 lv->lv_iovecp = (struct xfs_log_iovec *)&lv[1];351 lip->li_lv_shadow = lv;352 } else {353 /* same or smaller, optimise common overwrite case */354 lv = lip->li_lv_shadow;355 if (ordered)356 lv->lv_buf_len = XFS_LOG_VEC_ORDERED;357 else358 lv->lv_buf_len = 0;359 lv->lv_bytes = 0;360 }361 362 /* Ensure the lv is set up according to ->iop_size */363 lv->lv_niovecs = niovecs;364 365 /* The allocated data region lies beyond the iovec region */366 lv->lv_buf = (char *)lv + xlog_cil_iovec_space(niovecs);367 }368 369}370 371/*372 * Prepare the log item for insertion into the CIL. Calculate the difference in373 * log space it will consume, and if it is a new item pin it as well.374 */375STATIC void376xfs_cil_prepare_item(377 struct xlog *log,378 struct xfs_log_vec *lv,379 struct xfs_log_vec *old_lv,380 int *diff_len)381{382 /* Account for the new LV being passed in */383 if (lv->lv_buf_len != XFS_LOG_VEC_ORDERED)384 *diff_len += lv->lv_bytes;385 386 /*387 * If there is no old LV, this is the first time we've seen the item in388 * this CIL context and so we need to pin it. If we are replacing the389 * old_lv, then remove the space it accounts for and make it the shadow390 * buffer for later freeing. In both cases we are now switching to the391 * shadow buffer, so update the pointer to it appropriately.392 */393 if (!old_lv) {394 if (lv->lv_item->li_ops->iop_pin)395 lv->lv_item->li_ops->iop_pin(lv->lv_item);396 lv->lv_item->li_lv_shadow = NULL;397 } else if (old_lv != lv) {398 ASSERT(lv->lv_buf_len != XFS_LOG_VEC_ORDERED);399 400 *diff_len -= old_lv->lv_bytes;401 lv->lv_item->li_lv_shadow = old_lv;402 }403 404 /* attach new log vector to log item */405 lv->lv_item->li_lv = lv;406 407 /*408 * If this is the first time the item is being committed to the409 * CIL, store the sequence number on the log item so we can410 * tell in future commits whether this is the first checkpoint411 * the item is being committed into.412 */413 if (!lv->lv_item->li_seq)414 lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence;415}416 417/*418 * Format log item into a flat buffers419 *420 * For delayed logging, we need to hold a formatted buffer containing all the421 * changes on the log item. This enables us to relog the item in memory and422 * write it out asynchronously without needing to relock the object that was423 * modified at the time it gets written into the iclog.424 *425 * This function takes the prepared log vectors attached to each log item, and426 * formats the changes into the log vector buffer. The buffer it uses is427 * dependent on the current state of the vector in the CIL - the shadow lv is428 * guaranteed to be large enough for the current modification, but we will only429 * use that if we can't reuse the existing lv. If we can't reuse the existing430 * lv, then simple swap it out for the shadow lv. We don't free it - that is431 * done lazily either by th enext modification or the freeing of the log item.432 *433 * We don't set up region headers during this process; we simply copy the434 * regions into the flat buffer. We can do this because we still have to do a435 * formatting step to write the regions into the iclog buffer. Writing the436 * ophdrs during the iclog write means that we can support splitting large437 * regions across iclog boundares without needing a change in the format of the438 * item/region encapsulation.439 *440 * Hence what we need to do now is change the rewrite the vector array to point441 * to the copied region inside the buffer we just allocated. This allows us to442 * format the regions into the iclog as though they are being formatted443 * directly out of the objects themselves.444 */445static void446xlog_cil_insert_format_items(447 struct xlog *log,448 struct xfs_trans *tp,449 int *diff_len)450{451 struct xfs_log_item *lip;452 453 /* Bail out if we didn't find a log item. */454 if (list_empty(&tp->t_items)) {455 ASSERT(0);456 return;457 }458 459 list_for_each_entry(lip, &tp->t_items, li_trans) {460 struct xfs_log_vec *lv;461 struct xfs_log_vec *old_lv = NULL;462 struct xfs_log_vec *shadow;463 bool ordered = false;464 465 /* Skip items which aren't dirty in this transaction. */466 if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))467 continue;468 469 /*470 * The formatting size information is already attached to471 * the shadow lv on the log item.472 */473 shadow = lip->li_lv_shadow;474 if (shadow->lv_buf_len == XFS_LOG_VEC_ORDERED)475 ordered = true;476 477 /* Skip items that do not have any vectors for writing */478 if (!shadow->lv_niovecs && !ordered)479 continue;480 481 /* compare to existing item size */482 old_lv = lip->li_lv;483 if (lip->li_lv && shadow->lv_size <= lip->li_lv->lv_size) {484 /* same or smaller, optimise common overwrite case */485 lv = lip->li_lv;486 487 if (ordered)488 goto insert;489 490 /*491 * set the item up as though it is a new insertion so492 * that the space reservation accounting is correct.493 */494 *diff_len -= lv->lv_bytes;495 496 /* Ensure the lv is set up according to ->iop_size */497 lv->lv_niovecs = shadow->lv_niovecs;498 499 /* reset the lv buffer information for new formatting */500 lv->lv_buf_len = 0;501 lv->lv_bytes = 0;502 lv->lv_buf = (char *)lv +503 xlog_cil_iovec_space(lv->lv_niovecs);504 } else {505 /* switch to shadow buffer! */506 lv = shadow;507 lv->lv_item = lip;508 if (ordered) {509 /* track as an ordered logvec */510 ASSERT(lip->li_lv == NULL);511 goto insert;512 }513 }514 515 ASSERT(IS_ALIGNED((unsigned long)lv->lv_buf, sizeof(uint64_t)));516 lip->li_ops->iop_format(lip, lv);517insert:518 xfs_cil_prepare_item(log, lv, old_lv, diff_len);519 }520}521 522/*523 * The use of lockless waitqueue_active() requires that the caller has524 * serialised itself against the wakeup call in xlog_cil_push_work(). That525 * can be done by either holding the push lock or the context lock.526 */527static inline bool528xlog_cil_over_hard_limit(529 struct xlog *log,530 int32_t space_used)531{532 if (waitqueue_active(&log->l_cilp->xc_push_wait))533 return true;534 if (space_used >= XLOG_CIL_BLOCKING_SPACE_LIMIT(log))535 return true;536 return false;537}538 539/*540 * Insert the log items into the CIL and calculate the difference in space541 * consumed by the item. Add the space to the checkpoint ticket and calculate542 * if the change requires additional log metadata. If it does, take that space543 * as well. Remove the amount of space we added to the checkpoint ticket from544 * the current transaction ticket so that the accounting works out correctly.545 */546static void547xlog_cil_insert_items(548 struct xlog *log,549 struct xfs_trans *tp,550 uint32_t released_space)551{552 struct xfs_cil *cil = log->l_cilp;553 struct xfs_cil_ctx *ctx = cil->xc_ctx;554 struct xfs_log_item *lip;555 int len = 0;556 int iovhdr_res = 0, split_res = 0, ctx_res = 0;557 int space_used;558 int order;559 unsigned int cpu_nr;560 struct xlog_cil_pcp *cilpcp;561 562 ASSERT(tp);563 564 /*565 * We can do this safely because the context can't checkpoint until we566 * are done so it doesn't matter exactly how we update the CIL.567 */568 xlog_cil_insert_format_items(log, tp, &len);569 570 /*571 * Subtract the space released by intent cancelation from the space we572 * consumed so that we remove it from the CIL space and add it back to573 * the current transaction reservation context.574 */575 len -= released_space;576 577 /*578 * Grab the per-cpu pointer for the CIL before we start any accounting.579 * That ensures that we are running with pre-emption disabled and so we580 * can't be scheduled away between split sample/update operations that581 * are done without outside locking to serialise them.582 */583 cpu_nr = get_cpu();584 cilpcp = this_cpu_ptr(cil->xc_pcp);585 586 /* Tell the future push that there was work added by this CPU. */587 if (!cpumask_test_cpu(cpu_nr, &ctx->cil_pcpmask))588 cpumask_test_and_set_cpu(cpu_nr, &ctx->cil_pcpmask);589 590 /*591 * We need to take the CIL checkpoint unit reservation on the first592 * commit into the CIL. Test the XLOG_CIL_EMPTY bit first so we don't593 * unnecessarily do an atomic op in the fast path here. We can clear the594 * XLOG_CIL_EMPTY bit as we are under the xc_ctx_lock here and that595 * needs to be held exclusively to reset the XLOG_CIL_EMPTY bit.596 */597 if (test_bit(XLOG_CIL_EMPTY, &cil->xc_flags) &&598 test_and_clear_bit(XLOG_CIL_EMPTY, &cil->xc_flags))599 ctx_res = ctx->ticket->t_unit_res;600 601 /*602 * Check if we need to steal iclog headers. atomic_read() is not a603 * locked atomic operation, so we can check the value before we do any604 * real atomic ops in the fast path. If we've already taken the CIL unit605 * reservation from this commit, we've already got one iclog header606 * space reserved so we have to account for that otherwise we risk607 * overrunning the reservation on this ticket.608 *609 * If the CIL is already at the hard limit, we might need more header610 * space that originally reserved. So steal more header space from every611 * commit that occurs once we are over the hard limit to ensure the CIL612 * push won't run out of reservation space.613 *614 * This can steal more than we need, but that's OK.615 *616 * The cil->xc_ctx_lock provides the serialisation necessary for safely617 * calling xlog_cil_over_hard_limit() in this context.618 */619 space_used = atomic_read(&ctx->space_used) + cilpcp->space_used + len;620 if (atomic_read(&cil->xc_iclog_hdrs) > 0 ||621 xlog_cil_over_hard_limit(log, space_used)) {622 split_res = log->l_iclog_hsize +623 sizeof(struct xlog_op_header);624 if (ctx_res)625 ctx_res += split_res * (tp->t_ticket->t_iclog_hdrs - 1);626 else627 ctx_res = split_res * tp->t_ticket->t_iclog_hdrs;628 atomic_sub(tp->t_ticket->t_iclog_hdrs, &cil->xc_iclog_hdrs);629 }630 cilpcp->space_reserved += ctx_res;631 632 /*633 * Accurately account when over the soft limit, otherwise fold the634 * percpu count into the global count if over the per-cpu threshold.635 */636 if (!test_bit(XLOG_CIL_PCP_SPACE, &cil->xc_flags)) {637 atomic_add(len, &ctx->space_used);638 } else if (cilpcp->space_used + len >639 (XLOG_CIL_SPACE_LIMIT(log) / num_online_cpus())) {640 space_used = atomic_add_return(cilpcp->space_used + len,641 &ctx->space_used);642 cilpcp->space_used = 0;643 644 /*645 * If we just transitioned over the soft limit, we need to646 * transition to the global atomic counter.647 */648 if (space_used >= XLOG_CIL_SPACE_LIMIT(log))649 xlog_cil_insert_pcp_aggregate(cil, ctx);650 } else {651 cilpcp->space_used += len;652 }653 /* attach the transaction to the CIL if it has any busy extents */654 if (!list_empty(&tp->t_busy))655 list_splice_init(&tp->t_busy, &cilpcp->busy_extents);656 657 /*658 * Now update the order of everything modified in the transaction659 * and insert items into the CIL if they aren't already there.660 * We do this here so we only need to take the CIL lock once during661 * the transaction commit.662 */663 order = atomic_inc_return(&ctx->order_id);664 list_for_each_entry(lip, &tp->t_items, li_trans) {665 /* Skip items which aren't dirty in this transaction. */666 if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))667 continue;668 669 lip->li_order_id = order;670 if (!list_empty(&lip->li_cil))671 continue;672 list_add_tail(&lip->li_cil, &cilpcp->log_items);673 }674 put_cpu();675 676 /*677 * If we've overrun the reservation, dump the tx details before we move678 * the log items. Shutdown is imminent...679 */680 tp->t_ticket->t_curr_res -= ctx_res + len;681 if (WARN_ON(tp->t_ticket->t_curr_res < 0)) {682 xfs_warn(log->l_mp, "Transaction log reservation overrun:");683 xfs_warn(log->l_mp,684 " log items: %d bytes (iov hdrs: %d bytes)",685 len, iovhdr_res);686 xfs_warn(log->l_mp, " split region headers: %d bytes",687 split_res);688 xfs_warn(log->l_mp, " ctx ticket: %d bytes", ctx_res);689 xlog_print_trans(tp);690 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);691 }692}693 694static inline void695xlog_cil_ail_insert_batch(696 struct xfs_ail *ailp,697 struct xfs_ail_cursor *cur,698 struct xfs_log_item **log_items,699 int nr_items,700 xfs_lsn_t commit_lsn)701{702 int i;703 704 spin_lock(&ailp->ail_lock);705 /* xfs_trans_ail_update_bulk drops ailp->ail_lock */706 xfs_trans_ail_update_bulk(ailp, cur, log_items, nr_items, commit_lsn);707 708 for (i = 0; i < nr_items; i++) {709 struct xfs_log_item *lip = log_items[i];710 711 if (lip->li_ops->iop_unpin)712 lip->li_ops->iop_unpin(lip, 0);713 }714}715 716/*717 * Take the checkpoint's log vector chain of items and insert the attached log718 * items into the AIL. This uses bulk insertion techniques to minimise AIL lock719 * traffic.720 *721 * The AIL tracks log items via the start record LSN of the checkpoint,722 * not the commit record LSN. This is because we can pipeline multiple723 * checkpoints, and so the start record of checkpoint N+1 can be724 * written before the commit record of checkpoint N. i.e:725 *726 * start N commit N727 * +-------------+------------+----------------+728 * start N+1 commit N+1729 *730 * The tail of the log cannot be moved to the LSN of commit N when all731 * the items of that checkpoint are written back, because then the732 * start record for N+1 is no longer in the active portion of the log733 * and recovery will fail/corrupt the filesystem.734 *735 * Hence when all the log items in checkpoint N are written back, the736 * tail of the log most now only move as far forwards as the start LSN737 * of checkpoint N+1.738 *739 * If we are called with the aborted flag set, it is because a log write during740 * a CIL checkpoint commit has failed. In this case, all the items in the741 * checkpoint have already gone through iop_committed and iop_committing, which742 * means that checkpoint commit abort handling is treated exactly the same as an743 * iclog write error even though we haven't started any IO yet. Hence in this744 * case all we need to do is iop_committed processing, followed by an745 * iop_unpin(aborted) call.746 *747 * The AIL cursor is used to optimise the insert process. If commit_lsn is not748 * at the end of the AIL, the insert cursor avoids the need to walk the AIL to749 * find the insertion point on every xfs_log_item_batch_insert() call. This750 * saves a lot of needless list walking and is a net win, even though it751 * slightly increases that amount of AIL lock traffic to set it up and tear it752 * down.753 */754static void755xlog_cil_ail_insert(756 struct xfs_cil_ctx *ctx,757 bool aborted)758{759#define LOG_ITEM_BATCH_SIZE 32760 struct xfs_ail *ailp = ctx->cil->xc_log->l_ailp;761 struct xfs_log_item *log_items[LOG_ITEM_BATCH_SIZE];762 struct xfs_log_vec *lv;763 struct xfs_ail_cursor cur;764 xfs_lsn_t old_head;765 int i = 0;766 767 /*768 * Update the AIL head LSN with the commit record LSN of this769 * checkpoint. As iclogs are always completed in order, this should770 * always be the same (as iclogs can contain multiple commit records) or771 * higher LSN than the current head. We do this before insertion of the772 * items so that log space checks during insertion will reflect the773 * space that this checkpoint has already consumed. We call774 * xfs_ail_update_finish() so that tail space and space-based wakeups775 * will be recalculated appropriately.776 */777 ASSERT(XFS_LSN_CMP(ctx->commit_lsn, ailp->ail_head_lsn) >= 0 ||778 aborted);779 spin_lock(&ailp->ail_lock);780 xfs_trans_ail_cursor_last(ailp, &cur, ctx->start_lsn);781 old_head = ailp->ail_head_lsn;782 ailp->ail_head_lsn = ctx->commit_lsn;783 /* xfs_ail_update_finish() drops the ail_lock */784 xfs_ail_update_finish(ailp, NULLCOMMITLSN);785 786 /*787 * We move the AIL head forwards to account for the space used in the788 * log before we remove that space from the grant heads. This prevents a789 * transient condition where reservation space appears to become790 * available on return, only for it to disappear again immediately as791 * the AIL head update accounts in the log tail space.792 */793 smp_wmb(); /* paired with smp_rmb in xlog_grant_space_left */794 xlog_grant_return_space(ailp->ail_log, old_head, ailp->ail_head_lsn);795 796 /* unpin all the log items */797 list_for_each_entry(lv, &ctx->lv_chain, lv_list) {798 struct xfs_log_item *lip = lv->lv_item;799 xfs_lsn_t item_lsn;800 801 if (aborted)802 set_bit(XFS_LI_ABORTED, &lip->li_flags);803 804 if (lip->li_ops->flags & XFS_ITEM_RELEASE_WHEN_COMMITTED) {805 lip->li_ops->iop_release(lip);806 continue;807 }808 809 if (lip->li_ops->iop_committed)810 item_lsn = lip->li_ops->iop_committed(lip,811 ctx->start_lsn);812 else813 item_lsn = ctx->start_lsn;814 815 /* item_lsn of -1 means the item needs no further processing */816 if (XFS_LSN_CMP(item_lsn, (xfs_lsn_t)-1) == 0)817 continue;818 819 /*820 * if we are aborting the operation, no point in inserting the821 * object into the AIL as we are in a shutdown situation.822 */823 if (aborted) {824 ASSERT(xlog_is_shutdown(ailp->ail_log));825 if (lip->li_ops->iop_unpin)826 lip->li_ops->iop_unpin(lip, 1);827 continue;828 }829 830 if (item_lsn != ctx->start_lsn) {831 832 /*833 * Not a bulk update option due to unusual item_lsn.834 * Push into AIL immediately, rechecking the lsn once835 * we have the ail lock. Then unpin the item. This does836 * not affect the AIL cursor the bulk insert path is837 * using.838 */839 spin_lock(&ailp->ail_lock);840 if (XFS_LSN_CMP(item_lsn, lip->li_lsn) > 0)841 xfs_trans_ail_update(ailp, lip, item_lsn);842 else843 spin_unlock(&ailp->ail_lock);844 if (lip->li_ops->iop_unpin)845 lip->li_ops->iop_unpin(lip, 0);846 continue;847 }848 849 /* Item is a candidate for bulk AIL insert. */850 log_items[i++] = lv->lv_item;851 if (i >= LOG_ITEM_BATCH_SIZE) {852 xlog_cil_ail_insert_batch(ailp, &cur, log_items,853 LOG_ITEM_BATCH_SIZE, ctx->start_lsn);854 i = 0;855 }856 }857 858 /* make sure we insert the remainder! */859 if (i)860 xlog_cil_ail_insert_batch(ailp, &cur, log_items, i,861 ctx->start_lsn);862 863 spin_lock(&ailp->ail_lock);864 xfs_trans_ail_cursor_done(&cur);865 spin_unlock(&ailp->ail_lock);866}867 868static void869xlog_cil_free_logvec(870 struct list_head *lv_chain)871{872 struct xfs_log_vec *lv;873 874 while (!list_empty(lv_chain)) {875 lv = list_first_entry(lv_chain, struct xfs_log_vec, lv_list);876 list_del_init(&lv->lv_list);877 kvfree(lv);878 }879}880 881/*882 * Mark all items committed and clear busy extents. We free the log vector883 * chains in a separate pass so that we unpin the log items as quickly as884 * possible.885 */886static void887xlog_cil_committed(888 struct xfs_cil_ctx *ctx)889{890 struct xfs_mount *mp = ctx->cil->xc_log->l_mp;891 bool abort = xlog_is_shutdown(ctx->cil->xc_log);892 893 /*894 * If the I/O failed, we're aborting the commit and already shutdown.895 * Wake any commit waiters before aborting the log items so we don't896 * block async log pushers on callbacks. Async log pushers explicitly do897 * not wait on log force completion because they may be holding locks898 * required to unpin items.899 */900 if (abort) {901 spin_lock(&ctx->cil->xc_push_lock);902 wake_up_all(&ctx->cil->xc_start_wait);903 wake_up_all(&ctx->cil->xc_commit_wait);904 spin_unlock(&ctx->cil->xc_push_lock);905 }906 907 xlog_cil_ail_insert(ctx, abort);908 909 xfs_extent_busy_sort(&ctx->busy_extents.extent_list);910 xfs_extent_busy_clear(mp, &ctx->busy_extents.extent_list,911 xfs_has_discard(mp) && !abort);912 913 spin_lock(&ctx->cil->xc_push_lock);914 list_del(&ctx->committing);915 spin_unlock(&ctx->cil->xc_push_lock);916 917 xlog_cil_free_logvec(&ctx->lv_chain);918 919 if (!list_empty(&ctx->busy_extents.extent_list)) {920 ctx->busy_extents.mount = mp;921 ctx->busy_extents.owner = ctx;922 xfs_discard_extents(mp, &ctx->busy_extents);923 return;924 }925 926 kfree(ctx);927}928 929void930xlog_cil_process_committed(931 struct list_head *list)932{933 struct xfs_cil_ctx *ctx;934 935 while ((ctx = list_first_entry_or_null(list,936 struct xfs_cil_ctx, iclog_entry))) {937 list_del(&ctx->iclog_entry);938 xlog_cil_committed(ctx);939 }940}941 942/*943* Record the LSN of the iclog we were just granted space to start writing into.944* If the context doesn't have a start_lsn recorded, then this iclog will945* contain the start record for the checkpoint. Otherwise this write contains946* the commit record for the checkpoint.947*/948void949xlog_cil_set_ctx_write_state(950 struct xfs_cil_ctx *ctx,951 struct xlog_in_core *iclog)952{953 struct xfs_cil *cil = ctx->cil;954 xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn);955 956 ASSERT(!ctx->commit_lsn);957 if (!ctx->start_lsn) {958 spin_lock(&cil->xc_push_lock);959 /*960 * The LSN we need to pass to the log items on transaction961 * commit is the LSN reported by the first log vector write, not962 * the commit lsn. If we use the commit record lsn then we can963 * move the grant write head beyond the tail LSN and overwrite964 * it.965 */966 ctx->start_lsn = lsn;967 wake_up_all(&cil->xc_start_wait);968 spin_unlock(&cil->xc_push_lock);969 970 /*971 * Make sure the metadata we are about to overwrite in the log972 * has been flushed to stable storage before this iclog is973 * issued.974 */975 spin_lock(&cil->xc_log->l_icloglock);976 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH;977 spin_unlock(&cil->xc_log->l_icloglock);978 return;979 }980 981 /*982 * Take a reference to the iclog for the context so that we still hold983 * it when xlog_write is done and has released it. This means the984 * context controls when the iclog is released for IO.985 */986 atomic_inc(&iclog->ic_refcnt);987 988 /*989 * xlog_state_get_iclog_space() guarantees there is enough space in the990 * iclog for an entire commit record, so we can attach the context991 * callbacks now. This needs to be done before we make the commit_lsn992 * visible to waiters so that checkpoints with commit records in the993 * same iclog order their IO completion callbacks in the same order that994 * the commit records appear in the iclog.995 */996 spin_lock(&cil->xc_log->l_icloglock);997 list_add_tail(&ctx->iclog_entry, &iclog->ic_callbacks);998 spin_unlock(&cil->xc_log->l_icloglock);999 1000 /*1001 * Now we can record the commit LSN and wake anyone waiting for this1002 * sequence to have the ordered commit record assigned to a physical1003 * location in the log.1004 */1005 spin_lock(&cil->xc_push_lock);1006 ctx->commit_iclog = iclog;1007 ctx->commit_lsn = lsn;1008 wake_up_all(&cil->xc_commit_wait);1009 spin_unlock(&cil->xc_push_lock);1010}1011 1012 1013/*1014 * Ensure that the order of log writes follows checkpoint sequence order. This1015 * relies on the context LSN being zero until the log write has guaranteed the1016 * LSN that the log write will start at via xlog_state_get_iclog_space().1017 */1018enum _record_type {1019 _START_RECORD,1020 _COMMIT_RECORD,1021};1022 1023static int1024xlog_cil_order_write(1025 struct xfs_cil *cil,1026 xfs_csn_t sequence,1027 enum _record_type record)1028{1029 struct xfs_cil_ctx *ctx;1030 1031restart:1032 spin_lock(&cil->xc_push_lock);1033 list_for_each_entry(ctx, &cil->xc_committing, committing) {1034 /*1035 * Avoid getting stuck in this loop because we were woken by the1036 * shutdown, but then went back to sleep once already in the1037 * shutdown state.1038 */1039 if (xlog_is_shutdown(cil->xc_log)) {1040 spin_unlock(&cil->xc_push_lock);1041 return -EIO;1042 }1043 1044 /*1045 * Higher sequences will wait for this one so skip them.1046 * Don't wait for our own sequence, either.1047 */1048 if (ctx->sequence >= sequence)1049 continue;1050 1051 /* Wait until the LSN for the record has been recorded. */1052 switch (record) {1053 case _START_RECORD:1054 if (!ctx->start_lsn) {1055 xlog_wait(&cil->xc_start_wait, &cil->xc_push_lock);1056 goto restart;1057 }1058 break;1059 case _COMMIT_RECORD:1060 if (!ctx->commit_lsn) {1061 xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);1062 goto restart;1063 }1064 break;1065 }1066 }1067 spin_unlock(&cil->xc_push_lock);1068 return 0;1069}1070 1071/*1072 * Write out the log vector change now attached to the CIL context. This will1073 * write a start record that needs to be strictly ordered in ascending CIL1074 * sequence order so that log recovery will always use in-order start LSNs when1075 * replaying checkpoints.1076 */1077static int1078xlog_cil_write_chain(1079 struct xfs_cil_ctx *ctx,1080 uint32_t chain_len)1081{1082 struct xlog *log = ctx->cil->xc_log;1083 int error;1084 1085 error = xlog_cil_order_write(ctx->cil, ctx->sequence, _START_RECORD);1086 if (error)1087 return error;1088 return xlog_write(log, ctx, &ctx->lv_chain, ctx->ticket, chain_len);1089}1090 1091/*1092 * Write out the commit record of a checkpoint transaction to close off a1093 * running log write. These commit records are strictly ordered in ascending CIL1094 * sequence order so that log recovery will always replay the checkpoints in the1095 * correct order.1096 */1097static int1098xlog_cil_write_commit_record(1099 struct xfs_cil_ctx *ctx)1100{1101 struct xlog *log = ctx->cil->xc_log;1102 struct xlog_op_header ophdr = {1103 .oh_clientid = XFS_TRANSACTION,1104 .oh_tid = cpu_to_be32(ctx->ticket->t_tid),1105 .oh_flags = XLOG_COMMIT_TRANS,1106 };1107 struct xfs_log_iovec reg = {1108 .i_addr = &ophdr,1109 .i_len = sizeof(struct xlog_op_header),1110 .i_type = XLOG_REG_TYPE_COMMIT,1111 };1112 struct xfs_log_vec vec = {1113 .lv_niovecs = 1,1114 .lv_iovecp = ®,1115 };1116 int error;1117 LIST_HEAD(lv_chain);1118 list_add(&vec.lv_list, &lv_chain);1119 1120 if (xlog_is_shutdown(log))1121 return -EIO;1122 1123 error = xlog_cil_order_write(ctx->cil, ctx->sequence, _COMMIT_RECORD);1124 if (error)1125 return error;1126 1127 /* account for space used by record data */1128 ctx->ticket->t_curr_res -= reg.i_len;1129 error = xlog_write(log, ctx, &lv_chain, ctx->ticket, reg.i_len);1130 if (error)1131 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);1132 return error;1133}1134 1135struct xlog_cil_trans_hdr {1136 struct xlog_op_header oph[2];1137 struct xfs_trans_header thdr;1138 struct xfs_log_iovec lhdr[2];1139};1140 1141/*1142 * Build a checkpoint transaction header to begin the journal transaction. We1143 * need to account for the space used by the transaction header here as it is1144 * not accounted for in xlog_write().1145 *1146 * This is the only place we write a transaction header, so we also build the1147 * log opheaders that indicate the start of a log transaction and wrap the1148 * transaction header. We keep the start record in it's own log vector rather1149 * than compacting them into a single region as this ends up making the logic1150 * in xlog_write() for handling empty opheaders for start, commit and unmount1151 * records much simpler.1152 */1153static void1154xlog_cil_build_trans_hdr(1155 struct xfs_cil_ctx *ctx,1156 struct xlog_cil_trans_hdr *hdr,1157 struct xfs_log_vec *lvhdr,1158 int num_iovecs)1159{1160 struct xlog_ticket *tic = ctx->ticket;1161 __be32 tid = cpu_to_be32(tic->t_tid);1162 1163 memset(hdr, 0, sizeof(*hdr));1164 1165 /* Log start record */1166 hdr->oph[0].oh_tid = tid;1167 hdr->oph[0].oh_clientid = XFS_TRANSACTION;1168 hdr->oph[0].oh_flags = XLOG_START_TRANS;1169 1170 /* log iovec region pointer */1171 hdr->lhdr[0].i_addr = &hdr->oph[0];1172 hdr->lhdr[0].i_len = sizeof(struct xlog_op_header);1173 hdr->lhdr[0].i_type = XLOG_REG_TYPE_LRHEADER;1174 1175 /* log opheader */1176 hdr->oph[1].oh_tid = tid;1177 hdr->oph[1].oh_clientid = XFS_TRANSACTION;1178 hdr->oph[1].oh_len = cpu_to_be32(sizeof(struct xfs_trans_header));1179 1180 /* transaction header in host byte order format */1181 hdr->thdr.th_magic = XFS_TRANS_HEADER_MAGIC;1182 hdr->thdr.th_type = XFS_TRANS_CHECKPOINT;1183 hdr->thdr.th_tid = tic->t_tid;1184 hdr->thdr.th_num_items = num_iovecs;1185 1186 /* log iovec region pointer */1187 hdr->lhdr[1].i_addr = &hdr->oph[1];1188 hdr->lhdr[1].i_len = sizeof(struct xlog_op_header) +1189 sizeof(struct xfs_trans_header);1190 hdr->lhdr[1].i_type = XLOG_REG_TYPE_TRANSHDR;1191 1192 lvhdr->lv_niovecs = 2;1193 lvhdr->lv_iovecp = &hdr->lhdr[0];1194 lvhdr->lv_bytes = hdr->lhdr[0].i_len + hdr->lhdr[1].i_len;1195 1196 tic->t_curr_res -= lvhdr->lv_bytes;1197}1198 1199/*1200 * CIL item reordering compare function. We want to order in ascending ID order,1201 * but we want to leave items with the same ID in the order they were added to1202 * the list. This is important for operations like reflink where we log 4 order1203 * dependent intents in a single transaction when we overwrite an existing1204 * shared extent with a new shared extent. i.e. BUI(unmap), CUI(drop),1205 * CUI (inc), BUI(remap)...1206 */1207static int1208xlog_cil_order_cmp(1209 void *priv,1210 const struct list_head *a,1211 const struct list_head *b)1212{1213 struct xfs_log_vec *l1 = container_of(a, struct xfs_log_vec, lv_list);1214 struct xfs_log_vec *l2 = container_of(b, struct xfs_log_vec, lv_list);1215 1216 return l1->lv_order_id > l2->lv_order_id;1217}1218 1219/*1220 * Pull all the log vectors off the items in the CIL, and remove the items from1221 * the CIL. We don't need the CIL lock here because it's only needed on the1222 * transaction commit side which is currently locked out by the flush lock.1223 *1224 * If a log item is marked with a whiteout, we do not need to write it to the1225 * journal and so we just move them to the whiteout list for the caller to1226 * dispose of appropriately.1227 */1228static void1229xlog_cil_build_lv_chain(1230 struct xfs_cil_ctx *ctx,1231 struct list_head *whiteouts,1232 uint32_t *num_iovecs,1233 uint32_t *num_bytes)1234{1235 while (!list_empty(&ctx->log_items)) {1236 struct xfs_log_item *item;1237 struct xfs_log_vec *lv;1238 1239 item = list_first_entry(&ctx->log_items,1240 struct xfs_log_item, li_cil);1241 1242 if (test_bit(XFS_LI_WHITEOUT, &item->li_flags)) {1243 list_move(&item->li_cil, whiteouts);1244 trace_xfs_cil_whiteout_skip(item);1245 continue;1246 }1247 1248 lv = item->li_lv;1249 lv->lv_order_id = item->li_order_id;1250 1251 /* we don't write ordered log vectors */1252 if (lv->lv_buf_len != XFS_LOG_VEC_ORDERED)1253 *num_bytes += lv->lv_bytes;1254 *num_iovecs += lv->lv_niovecs;1255 list_add_tail(&lv->lv_list, &ctx->lv_chain);1256 1257 list_del_init(&item->li_cil);1258 item->li_order_id = 0;1259 item->li_lv = NULL;1260 }1261}1262 1263static void1264xlog_cil_cleanup_whiteouts(1265 struct list_head *whiteouts)1266{1267 while (!list_empty(whiteouts)) {1268 struct xfs_log_item *item = list_first_entry(whiteouts,1269 struct xfs_log_item, li_cil);1270 list_del_init(&item->li_cil);1271 trace_xfs_cil_whiteout_unpin(item);1272 item->li_ops->iop_unpin(item, 1);1273 }1274}1275 1276/*1277 * Push the Committed Item List to the log.1278 *1279 * If the current sequence is the same as xc_push_seq we need to do a flush. If1280 * xc_push_seq is less than the current sequence, then it has already been1281 * flushed and we don't need to do anything - the caller will wait for it to1282 * complete if necessary.1283 *1284 * xc_push_seq is checked unlocked against the sequence number for a match.1285 * Hence we can allow log forces to run racily and not issue pushes for the1286 * same sequence twice. If we get a race between multiple pushes for the same1287 * sequence they will block on the first one and then abort, hence avoiding1288 * needless pushes.1289 *1290 * This runs from a workqueue so it does not inherent any specific memory1291 * allocation context. However, we do not want to block on memory reclaim1292 * recursing back into the filesystem because this push may have been triggered1293 * by memory reclaim itself. Hence we really need to run under full GFP_NOFS1294 * contraints here.1295 */1296static void1297xlog_cil_push_work(1298 struct work_struct *work)1299{1300 unsigned int nofs_flags = memalloc_nofs_save();1301 struct xfs_cil_ctx *ctx =1302 container_of(work, struct xfs_cil_ctx, push_work);1303 struct xfs_cil *cil = ctx->cil;1304 struct xlog *log = cil->xc_log;1305 struct xfs_cil_ctx *new_ctx;1306 int num_iovecs = 0;1307 int num_bytes = 0;1308 int error = 0;1309 struct xlog_cil_trans_hdr thdr;1310 struct xfs_log_vec lvhdr = {};1311 xfs_csn_t push_seq;1312 bool push_commit_stable;1313 LIST_HEAD (whiteouts);1314 struct xlog_ticket *ticket;1315 1316 new_ctx = xlog_cil_ctx_alloc();1317 new_ctx->ticket = xlog_cil_ticket_alloc(log);1318 1319 down_write(&cil->xc_ctx_lock);1320 1321 spin_lock(&cil->xc_push_lock);1322 push_seq = cil->xc_push_seq;1323 ASSERT(push_seq <= ctx->sequence);1324 push_commit_stable = cil->xc_push_commit_stable;1325 cil->xc_push_commit_stable = false;1326 1327 /*1328 * As we are about to switch to a new, empty CIL context, we no longer1329 * need to throttle tasks on CIL space overruns. Wake any waiters that1330 * the hard push throttle may have caught so they can start committing1331 * to the new context. The ctx->xc_push_lock provides the serialisation1332 * necessary for safely using the lockless waitqueue_active() check in1333 * this context.1334 */1335 if (waitqueue_active(&cil->xc_push_wait))1336 wake_up_all(&cil->xc_push_wait);1337 1338 xlog_cil_push_pcp_aggregate(cil, ctx);1339 1340 /*1341 * Check if we've anything to push. If there is nothing, then we don't1342 * move on to a new sequence number and so we have to be able to push1343 * this sequence again later.1344 */1345 if (test_bit(XLOG_CIL_EMPTY, &cil->xc_flags)) {1346 cil->xc_push_seq = 0;1347 spin_unlock(&cil->xc_push_lock);1348 goto out_skip;1349 }1350 1351 1352 /* check for a previously pushed sequence */1353 if (push_seq < ctx->sequence) {1354 spin_unlock(&cil->xc_push_lock);1355 goto out_skip;1356 }1357 1358 /*1359 * We are now going to push this context, so add it to the committing1360 * list before we do anything else. This ensures that anyone waiting on1361 * this push can easily detect the difference between a "push in1362 * progress" and "CIL is empty, nothing to do".1363 *1364 * IOWs, a wait loop can now check for:1365 * the current sequence not being found on the committing list;1366 * an empty CIL; and1367 * an unchanged sequence number1368 * to detect a push that had nothing to do and therefore does not need1369 * waiting on. If the CIL is not empty, we get put on the committing1370 * list before emptying the CIL and bumping the sequence number. Hence1371 * an empty CIL and an unchanged sequence number means we jumped out1372 * above after doing nothing.1373 *1374 * Hence the waiter will either find the commit sequence on the1375 * committing list or the sequence number will be unchanged and the CIL1376 * still dirty. In that latter case, the push has not yet started, and1377 * so the waiter will have to continue trying to check the CIL1378 * committing list until it is found. In extreme cases of delay, the1379 * sequence may fully commit between the attempts the wait makes to wait1380 * on the commit sequence.1381 */1382 list_add(&ctx->committing, &cil->xc_committing);1383 spin_unlock(&cil->xc_push_lock);1384 1385 xlog_cil_build_lv_chain(ctx, &whiteouts, &num_iovecs, &num_bytes);1386 1387 /*1388 * Switch the contexts so we can drop the context lock and move out1389 * of a shared context. We can't just go straight to the commit record,1390 * though - we need to synchronise with previous and future commits so1391 * that the commit records are correctly ordered in the log to ensure1392 * that we process items during log IO completion in the correct order.1393 *1394 * For example, if we get an EFI in one checkpoint and the EFD in the1395 * next (e.g. due to log forces), we do not want the checkpoint with1396 * the EFD to be committed before the checkpoint with the EFI. Hence1397 * we must strictly order the commit records of the checkpoints so1398 * that: a) the checkpoint callbacks are attached to the iclogs in the1399 * correct order; and b) the checkpoints are replayed in correct order1400 * in log recovery.1401 *1402 * Hence we need to add this context to the committing context list so1403 * that higher sequences will wait for us to write out a commit record1404 * before they do.1405 *1406 * xfs_log_force_seq requires us to mirror the new sequence into the cil1407 * structure atomically with the addition of this sequence to the1408 * committing list. This also ensures that we can do unlocked checks1409 * against the current sequence in log forces without risking1410 * deferencing a freed context pointer.1411 */1412 spin_lock(&cil->xc_push_lock);1413 xlog_cil_ctx_switch(cil, new_ctx);1414 spin_unlock(&cil->xc_push_lock);1415 up_write(&cil->xc_ctx_lock);1416 1417 /*1418 * Sort the log vector chain before we add the transaction headers.1419 * This ensures we always have the transaction headers at the start1420 * of the chain.1421 */1422 list_sort(NULL, &ctx->lv_chain, xlog_cil_order_cmp);1423 1424 /*1425 * Build a checkpoint transaction header and write it to the log to1426 * begin the transaction. We need to account for the space used by the1427 * transaction header here as it is not accounted for in xlog_write().1428 * Add the lvhdr to the head of the lv chain we pass to xlog_write() so1429 * it gets written into the iclog first.1430 */1431 xlog_cil_build_trans_hdr(ctx, &thdr, &lvhdr, num_iovecs);1432 num_bytes += lvhdr.lv_bytes;1433 list_add(&lvhdr.lv_list, &ctx->lv_chain);1434 1435 /*1436 * Take the lvhdr back off the lv_chain immediately after calling1437 * xlog_cil_write_chain() as it should not be passed to log IO1438 * completion.1439 */1440 error = xlog_cil_write_chain(ctx, num_bytes);1441 list_del(&lvhdr.lv_list);1442 if (error)1443 goto out_abort_free_ticket;1444 1445 error = xlog_cil_write_commit_record(ctx);1446 if (error)1447 goto out_abort_free_ticket;1448 1449 /*1450 * Grab the ticket from the ctx so we can ungrant it after releasing the1451 * commit_iclog. The ctx may be freed by the time we return from1452 * releasing the commit_iclog (i.e. checkpoint has been completed and1453 * callback run) so we can't reference the ctx after the call to1454 * xlog_state_release_iclog().1455 */1456 ticket = ctx->ticket;1457 1458 /*1459 * If the checkpoint spans multiple iclogs, wait for all previous iclogs1460 * to complete before we submit the commit_iclog. We can't use state1461 * checks for this - ACTIVE can be either a past completed iclog or a1462 * future iclog being filled, while WANT_SYNC through SYNC_DONE can be a1463 * past or future iclog awaiting IO or ordered IO completion to be run.1464 * In the latter case, if it's a future iclog and we wait on it, the we1465 * will hang because it won't get processed through to ic_force_wait1466 * wakeup until this commit_iclog is written to disk. Hence we use the1467 * iclog header lsn and compare it to the commit lsn to determine if we1468 * need to wait on iclogs or not.1469 */1470 spin_lock(&log->l_icloglock);1471 if (ctx->start_lsn != ctx->commit_lsn) {1472 xfs_lsn_t plsn;1473 1474 plsn = be64_to_cpu(ctx->commit_iclog->ic_prev->ic_header.h_lsn);1475 if (plsn && XFS_LSN_CMP(plsn, ctx->commit_lsn) < 0) {1476 /*1477 * Waiting on ic_force_wait orders the completion of1478 * iclogs older than ic_prev. Hence we only need to wait1479 * on the most recent older iclog here.1480 */1481 xlog_wait_on_iclog(ctx->commit_iclog->ic_prev);1482 spin_lock(&log->l_icloglock);1483 }1484 1485 /*1486 * We need to issue a pre-flush so that the ordering for this1487 * checkpoint is correctly preserved down to stable storage.1488 */1489 ctx->commit_iclog->ic_flags |= XLOG_ICL_NEED_FLUSH;1490 }1491 1492 /*1493 * The commit iclog must be written to stable storage to guarantee1494 * journal IO vs metadata writeback IO is correctly ordered on stable1495 * storage.1496 *1497 * If the push caller needs the commit to be immediately stable and the1498 * commit_iclog is not yet marked as XLOG_STATE_WANT_SYNC to indicate it1499 * will be written when released, switch it's state to WANT_SYNC right1500 * now.1501 */1502 ctx->commit_iclog->ic_flags |= XLOG_ICL_NEED_FUA;1503 if (push_commit_stable &&1504 ctx->commit_iclog->ic_state == XLOG_STATE_ACTIVE)1505 xlog_state_switch_iclogs(log, ctx->commit_iclog, 0);1506 ticket = ctx->ticket;1507 xlog_state_release_iclog(log, ctx->commit_iclog, ticket);1508 1509 /* Not safe to reference ctx now! */1510 1511 spin_unlock(&log->l_icloglock);1512 xlog_cil_cleanup_whiteouts(&whiteouts);1513 xfs_log_ticket_ungrant(log, ticket);1514 memalloc_nofs_restore(nofs_flags);1515 return;1516 1517out_skip:1518 up_write(&cil->xc_ctx_lock);1519 xfs_log_ticket_put(new_ctx->ticket);1520 kfree(new_ctx);1521 memalloc_nofs_restore(nofs_flags);1522 return;1523 1524out_abort_free_ticket:1525 ASSERT(xlog_is_shutdown(log));1526 xlog_cil_cleanup_whiteouts(&whiteouts);1527 if (!ctx->commit_iclog) {1528 xfs_log_ticket_ungrant(log, ctx->ticket);1529 xlog_cil_committed(ctx);1530 memalloc_nofs_restore(nofs_flags);1531 return;1532 }1533 spin_lock(&log->l_icloglock);1534 ticket = ctx->ticket;1535 xlog_state_release_iclog(log, ctx->commit_iclog, ticket);1536 /* Not safe to reference ctx now! */1537 spin_unlock(&log->l_icloglock);1538 xfs_log_ticket_ungrant(log, ticket);1539 memalloc_nofs_restore(nofs_flags);1540}1541 1542/*1543 * We need to push CIL every so often so we don't cache more than we can fit in1544 * the log. The limit really is that a checkpoint can't be more than half the1545 * log (the current checkpoint is not allowed to overwrite the previous1546 * checkpoint), but commit latency and memory usage limit this to a smaller1547 * size.1548 */1549static void1550xlog_cil_push_background(1551 struct xlog *log)1552{1553 struct xfs_cil *cil = log->l_cilp;1554 int space_used = atomic_read(&cil->xc_ctx->space_used);1555 1556 /*1557 * The cil won't be empty because we are called while holding the1558 * context lock so whatever we added to the CIL will still be there.1559 */1560 ASSERT(!test_bit(XLOG_CIL_EMPTY, &cil->xc_flags));1561 1562 /*1563 * We are done if:1564 * - we haven't used up all the space available yet; or1565 * - we've already queued up a push; and1566 * - we're not over the hard limit; and1567 * - nothing has been over the hard limit.1568 *1569 * If so, we don't need to take the push lock as there's nothing to do.1570 */1571 if (space_used < XLOG_CIL_SPACE_LIMIT(log) ||1572 (cil->xc_push_seq == cil->xc_current_sequence &&1573 space_used < XLOG_CIL_BLOCKING_SPACE_LIMIT(log) &&1574 !waitqueue_active(&cil->xc_push_wait))) {1575 up_read(&cil->xc_ctx_lock);1576 return;1577 }1578 1579 spin_lock(&cil->xc_push_lock);1580 if (cil->xc_push_seq < cil->xc_current_sequence) {1581 cil->xc_push_seq = cil->xc_current_sequence;1582 queue_work(cil->xc_push_wq, &cil->xc_ctx->push_work);1583 }1584 1585 /*1586 * Drop the context lock now, we can't hold that if we need to sleep1587 * because we are over the blocking threshold. The push_lock is still1588 * held, so blocking threshold sleep/wakeup is still correctly1589 * serialised here.1590 */1591 up_read(&cil->xc_ctx_lock);1592 1593 /*1594 * If we are well over the space limit, throttle the work that is being1595 * done until the push work on this context has begun. Enforce the hard1596 * throttle on all transaction commits once it has been activated, even1597 * if the committing transactions have resulted in the space usage1598 * dipping back down under the hard limit.1599 *1600 * The ctx->xc_push_lock provides the serialisation necessary for safely1601 * calling xlog_cil_over_hard_limit() in this context.1602 */1603 if (xlog_cil_over_hard_limit(log, space_used)) {1604 trace_xfs_log_cil_wait(log, cil->xc_ctx->ticket);1605 ASSERT(space_used < log->l_logsize);1606 xlog_wait(&cil->xc_push_wait, &cil->xc_push_lock);1607 return;1608 }1609 1610 spin_unlock(&cil->xc_push_lock);1611 1612}1613 1614/*1615 * xlog_cil_push_now() is used to trigger an immediate CIL push to the sequence1616 * number that is passed. When it returns, the work will be queued for1617 * @push_seq, but it won't be completed.1618 *1619 * If the caller is performing a synchronous force, we will flush the workqueue1620 * to get previously queued work moving to minimise the wait time they will1621 * undergo waiting for all outstanding pushes to complete. The caller is1622 * expected to do the required waiting for push_seq to complete.1623 *1624 * If the caller is performing an async push, we need to ensure that the1625 * checkpoint is fully flushed out of the iclogs when we finish the push. If we1626 * don't do this, then the commit record may remain sitting in memory in an1627 * ACTIVE iclog. This then requires another full log force to push to disk,1628 * which defeats the purpose of having an async, non-blocking CIL force1629 * mechanism. Hence in this case we need to pass a flag to the push work to1630 * indicate it needs to flush the commit record itself.1631 */1632static void1633xlog_cil_push_now(1634 struct xlog *log,1635 xfs_lsn_t push_seq,1636 bool async)1637{1638 struct xfs_cil *cil = log->l_cilp;1639 1640 if (!cil)1641 return;1642 1643 ASSERT(push_seq && push_seq <= cil->xc_current_sequence);1644 1645 /* start on any pending background push to minimise wait time on it */1646 if (!async)1647 flush_workqueue(cil->xc_push_wq);1648 1649 spin_lock(&cil->xc_push_lock);1650 1651 /*1652 * If this is an async flush request, we always need to set the1653 * xc_push_commit_stable flag even if something else has already queued1654 * a push. The flush caller is asking for the CIL to be on stable1655 * storage when the next push completes, so regardless of who has queued1656 * the push, the flush requires stable semantics from it.1657 */1658 cil->xc_push_commit_stable = async;1659 1660 /*1661 * If the CIL is empty or we've already pushed the sequence then1662 * there's no more work that we need to do.1663 */1664 if (test_bit(XLOG_CIL_EMPTY, &cil->xc_flags) ||1665 push_seq <= cil->xc_push_seq) {1666 spin_unlock(&cil->xc_push_lock);1667 return;1668 }1669 1670 cil->xc_push_seq = push_seq;1671 queue_work(cil->xc_push_wq, &cil->xc_ctx->push_work);1672 spin_unlock(&cil->xc_push_lock);1673}1674 1675bool1676xlog_cil_empty(1677 struct xlog *log)1678{1679 struct xfs_cil *cil = log->l_cilp;1680 bool empty = false;1681 1682 spin_lock(&cil->xc_push_lock);1683 if (test_bit(XLOG_CIL_EMPTY, &cil->xc_flags))1684 empty = true;1685 spin_unlock(&cil->xc_push_lock);1686 return empty;1687}1688 1689/*1690 * If there are intent done items in this transaction and the related intent was1691 * committed in the current (same) CIL checkpoint, we don't need to write either1692 * the intent or intent done item to the journal as the change will be1693 * journalled atomically within this checkpoint. As we cannot remove items from1694 * the CIL here, mark the related intent with a whiteout so that the CIL push1695 * can remove it rather than writing it to the journal. Then remove the intent1696 * done item from the current transaction and release it so it doesn't get put1697 * into the CIL at all.1698 */1699static uint32_t1700xlog_cil_process_intents(1701 struct xfs_cil *cil,1702 struct xfs_trans *tp)1703{1704 struct xfs_log_item *lip, *ilip, *next;1705 uint32_t len = 0;1706 1707 list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {1708 if (!(lip->li_ops->flags & XFS_ITEM_INTENT_DONE))1709 continue;1710 1711 ilip = lip->li_ops->iop_intent(lip);1712 if (!ilip || !xlog_item_in_current_chkpt(cil, ilip))1713 continue;1714 set_bit(XFS_LI_WHITEOUT, &ilip->li_flags);1715 trace_xfs_cil_whiteout_mark(ilip);1716 len += ilip->li_lv->lv_bytes;1717 kvfree(ilip->li_lv);1718 ilip->li_lv = NULL;1719 1720 xfs_trans_del_item(lip);1721 lip->li_ops->iop_release(lip);1722 }1723 return len;1724}1725 1726/*1727 * Commit a transaction with the given vector to the Committed Item List.1728 *1729 * To do this, we need to format the item, pin it in memory if required and1730 * account for the space used by the transaction. Once we have done that we1731 * need to release the unused reservation for the transaction, attach the1732 * transaction to the checkpoint context so we carry the busy extents through1733 * to checkpoint completion, and then unlock all the items in the transaction.1734 *1735 * Called with the context lock already held in read mode to lock out1736 * background commit, returns without it held once background commits are1737 * allowed again.1738 */1739void1740xlog_cil_commit(1741 struct xlog *log,1742 struct xfs_trans *tp,1743 xfs_csn_t *commit_seq,1744 bool regrant)1745{1746 struct xfs_cil *cil = log->l_cilp;1747 struct xfs_log_item *lip, *next;1748 uint32_t released_space = 0;1749 1750 /*1751 * Do all necessary memory allocation before we lock the CIL.1752 * This ensures the allocation does not deadlock with a CIL1753 * push in memory reclaim (e.g. from kswapd).1754 */1755 xlog_cil_alloc_shadow_bufs(log, tp);1756 1757 /* lock out background commit */1758 down_read(&cil->xc_ctx_lock);1759 1760 if (tp->t_flags & XFS_TRANS_HAS_INTENT_DONE)1761 released_space = xlog_cil_process_intents(cil, tp);1762 1763 xlog_cil_insert_items(log, tp, released_space);1764 1765 if (regrant && !xlog_is_shutdown(log))1766 xfs_log_ticket_regrant(log, tp->t_ticket);1767 else1768 xfs_log_ticket_ungrant(log, tp->t_ticket);1769 tp->t_ticket = NULL;1770 xfs_trans_unreserve_and_mod_sb(tp);1771 1772 /*1773 * Once all the items of the transaction have been copied to the CIL,1774 * the items can be unlocked and possibly freed.1775 *1776 * This needs to be done before we drop the CIL context lock because we1777 * have to update state in the log items and unlock them before they go1778 * to disk. If we don't, then the CIL checkpoint can race with us and1779 * we can run checkpoint completion before we've updated and unlocked1780 * the log items. This affects (at least) processing of stale buffers,1781 * inodes and EFIs.1782 */1783 trace_xfs_trans_commit_items(tp, _RET_IP_);1784 list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {1785 xfs_trans_del_item(lip);1786 if (lip->li_ops->iop_committing)1787 lip->li_ops->iop_committing(lip, cil->xc_ctx->sequence);1788 }1789 if (commit_seq)1790 *commit_seq = cil->xc_ctx->sequence;1791 1792 /* xlog_cil_push_background() releases cil->xc_ctx_lock */1793 xlog_cil_push_background(log);1794}1795 1796/*1797 * Flush the CIL to stable storage but don't wait for it to complete. This1798 * requires the CIL push to ensure the commit record for the push hits the disk,1799 * but otherwise is no different to a push done from a log force.1800 */1801void1802xlog_cil_flush(1803 struct xlog *log)1804{1805 xfs_csn_t seq = log->l_cilp->xc_current_sequence;1806 1807 trace_xfs_log_force(log->l_mp, seq, _RET_IP_);1808 xlog_cil_push_now(log, seq, true);1809 1810 /*1811 * If the CIL is empty, make sure that any previous checkpoint that may1812 * still be in an active iclog is pushed to stable storage.1813 */1814 if (test_bit(XLOG_CIL_EMPTY, &log->l_cilp->xc_flags))1815 xfs_log_force(log->l_mp, 0);1816}1817 1818/*1819 * Conditionally push the CIL based on the sequence passed in.1820 *1821 * We only need to push if we haven't already pushed the sequence number given.1822 * Hence the only time we will trigger a push here is if the push sequence is1823 * the same as the current context.1824 *1825 * We return the current commit lsn to allow the callers to determine if a1826 * iclog flush is necessary following this call.1827 */1828xfs_lsn_t1829xlog_cil_force_seq(1830 struct xlog *log,1831 xfs_csn_t sequence)1832{1833 struct xfs_cil *cil = log->l_cilp;1834 struct xfs_cil_ctx *ctx;1835 xfs_lsn_t commit_lsn = NULLCOMMITLSN;1836 1837 ASSERT(sequence <= cil->xc_current_sequence);1838 1839 if (!sequence)1840 sequence = cil->xc_current_sequence;1841 trace_xfs_log_force(log->l_mp, sequence, _RET_IP_);1842 1843 /*1844 * check to see if we need to force out the current context.1845 * xlog_cil_push() handles racing pushes for the same sequence,1846 * so no need to deal with it here.1847 */1848restart:1849 xlog_cil_push_now(log, sequence, false);1850 1851 /*1852 * See if we can find a previous sequence still committing.1853 * We need to wait for all previous sequence commits to complete1854 * before allowing the force of push_seq to go ahead. Hence block1855 * on commits for those as well.1856 */1857 spin_lock(&cil->xc_push_lock);1858 list_for_each_entry(ctx, &cil->xc_committing, committing) {1859 /*1860 * Avoid getting stuck in this loop because we were woken by the1861 * shutdown, but then went back to sleep once already in the1862 * shutdown state.1863 */1864 if (xlog_is_shutdown(log))1865 goto out_shutdown;1866 if (ctx->sequence > sequence)1867 continue;1868 if (!ctx->commit_lsn) {1869 /*1870 * It is still being pushed! Wait for the push to1871 * complete, then start again from the beginning.1872 */1873 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);1874 xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);1875 goto restart;1876 }1877 if (ctx->sequence != sequence)1878 continue;1879 /* found it! */1880 commit_lsn = ctx->commit_lsn;1881 }1882 1883 /*1884 * The call to xlog_cil_push_now() executes the push in the background.1885 * Hence by the time we have got here it our sequence may not have been1886 * pushed yet. This is true if the current sequence still matches the1887 * push sequence after the above wait loop and the CIL still contains1888 * dirty objects. This is guaranteed by the push code first adding the1889 * context to the committing list before emptying the CIL.1890 *1891 * Hence if we don't find the context in the committing list and the1892 * current sequence number is unchanged then the CIL contents are1893 * significant. If the CIL is empty, if means there was nothing to push1894 * and that means there is nothing to wait for. If the CIL is not empty,1895 * it means we haven't yet started the push, because if it had started1896 * we would have found the context on the committing list.1897 */1898 if (sequence == cil->xc_current_sequence &&1899 !test_bit(XLOG_CIL_EMPTY, &cil->xc_flags)) {1900 spin_unlock(&cil->xc_push_lock);1901 goto restart;1902 }1903 1904 spin_unlock(&cil->xc_push_lock);1905 return commit_lsn;1906 1907 /*1908 * We detected a shutdown in progress. We need to trigger the log force1909 * to pass through it's iclog state machine error handling, even though1910 * we are already in a shutdown state. Hence we can't return1911 * NULLCOMMITLSN here as that has special meaning to log forces (i.e.1912 * LSN is already stable), so we return a zero LSN instead.1913 */1914out_shutdown:1915 spin_unlock(&cil->xc_push_lock);1916 return 0;1917}1918 1919/*1920 * Perform initial CIL structure initialisation.1921 */1922int1923xlog_cil_init(1924 struct xlog *log)1925{1926 struct xfs_cil *cil;1927 struct xfs_cil_ctx *ctx;1928 struct xlog_cil_pcp *cilpcp;1929 int cpu;1930 1931 cil = kzalloc(sizeof(*cil), GFP_KERNEL | __GFP_RETRY_MAYFAIL);1932 if (!cil)1933 return -ENOMEM;1934 /*1935 * Limit the CIL pipeline depth to 4 concurrent works to bound the1936 * concurrency the log spinlocks will be exposed to.1937 */1938 cil->xc_push_wq = alloc_workqueue("xfs-cil/%s",1939 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM | WQ_UNBOUND),1940 4, log->l_mp->m_super->s_id);1941 if (!cil->xc_push_wq)1942 goto out_destroy_cil;1943 1944 cil->xc_log = log;1945 cil->xc_pcp = alloc_percpu(struct xlog_cil_pcp);1946 if (!cil->xc_pcp)1947 goto out_destroy_wq;1948 1949 for_each_possible_cpu(cpu) {1950 cilpcp = per_cpu_ptr(cil->xc_pcp, cpu);1951 INIT_LIST_HEAD(&cilpcp->busy_extents);1952 INIT_LIST_HEAD(&cilpcp->log_items);1953 }1954 1955 INIT_LIST_HEAD(&cil->xc_committing);1956 spin_lock_init(&cil->xc_push_lock);1957 init_waitqueue_head(&cil->xc_push_wait);1958 init_rwsem(&cil->xc_ctx_lock);1959 init_waitqueue_head(&cil->xc_start_wait);1960 init_waitqueue_head(&cil->xc_commit_wait);1961 log->l_cilp = cil;1962 1963 ctx = xlog_cil_ctx_alloc();1964 xlog_cil_ctx_switch(cil, ctx);1965 return 0;1966 1967out_destroy_wq:1968 destroy_workqueue(cil->xc_push_wq);1969out_destroy_cil:1970 kfree(cil);1971 return -ENOMEM;1972}1973 1974void1975xlog_cil_destroy(1976 struct xlog *log)1977{1978 struct xfs_cil *cil = log->l_cilp;1979 1980 if (cil->xc_ctx) {1981 if (cil->xc_ctx->ticket)1982 xfs_log_ticket_put(cil->xc_ctx->ticket);1983 kfree(cil->xc_ctx);1984 }1985 1986 ASSERT(test_bit(XLOG_CIL_EMPTY, &cil->xc_flags));1987 free_percpu(cil->xc_pcp);1988 destroy_workqueue(cil->xc_push_wq);1989 kfree(cil);1990}1991 1992