linux_old1/fs/xfs/libxfs/xfs_defer.c

579 lines
18 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0+
/*
* Copyright (C) 2016 Oracle. All Rights Reserved.
* Author: Darrick J. Wong <darrick.wong@oracle.com>
*/
#include "xfs.h"
#include "xfs_fs.h"
#include "xfs_shared.h"
#include "xfs_format.h"
#include "xfs_log_format.h"
#include "xfs_trans_resv.h"
#include "xfs_bit.h"
#include "xfs_sb.h"
#include "xfs_mount.h"
#include "xfs_defer.h"
#include "xfs_trans.h"
#include "xfs_trace.h"
/*
* Deferred Operations in XFS
*
* Due to the way locking rules work in XFS, certain transactions (block
* mapping and unmapping, typically) have permanent reservations so that
* we can roll the transaction to adhere to AG locking order rules and
* to unlock buffers between metadata updates. Prior to rmap/reflink,
* the mapping code had a mechanism to perform these deferrals for
* extents that were going to be freed; this code makes that facility
* more generic.
*
* When adding the reverse mapping and reflink features, it became
* necessary to perform complex remapping multi-transactions to comply
* with AG locking order rules, and to be able to spread a single
* refcount update operation (an operation on an n-block extent can
* update as many as n records!) among multiple transactions. XFS can
* roll a transaction to facilitate this, but using this facility
* requires us to log "intent" items in case log recovery needs to
* redo the operation, and to log "done" items to indicate that redo
* is not necessary.
*
* Deferred work is tracked in xfs_defer_pending items. Each pending
* item tracks one type of deferred work. Incoming work items (which
* have not yet had an intent logged) are attached to a pending item
* on the dop_intake list, where they wait for the caller to finish
* the deferred operations.
*
* Finishing a set of deferred operations is an involved process. To
* start, we define "rolling a deferred-op transaction" as follows:
*
* > For each xfs_defer_pending item on the dop_intake list,
* - Sort the work items in AG order. XFS locking
* order rules require us to lock buffers in AG order.
* - Create a log intent item for that type.
* - Attach it to the pending item.
* - Move the pending item from the dop_intake list to the
* dop_pending list.
* > Roll the transaction.
*
* NOTE: To avoid exceeding the transaction reservation, we limit the
* number of items that we attach to a given xfs_defer_pending.
*
* The actual finishing process looks like this:
*
* > For each xfs_defer_pending in the dop_pending list,
* - Roll the deferred-op transaction as above.
* - Create a log done item for that type, and attach it to the
* log intent item.
* - For each work item attached to the log intent item,
* * Perform the described action.
* * Attach the work item to the log done item.
* * If the result of doing the work was -EAGAIN, ->finish work
* wants a new transaction. See the "Requesting a Fresh
* Transaction while Finishing Deferred Work" section below for
* details.
*
* The key here is that we must log an intent item for all pending
* work items every time we roll the transaction, and that we must log
* a done item as soon as the work is completed. With this mechanism
* we can perform complex remapping operations, chaining intent items
* as needed.
*
* Requesting a Fresh Transaction while Finishing Deferred Work
*
* If ->finish_item decides that it needs a fresh transaction to
* finish the work, it must ask its caller (xfs_defer_finish) for a
* continuation. The most likely cause of this circumstance are the
* refcount adjust functions deciding that they've logged enough items
* to be at risk of exceeding the transaction reservation.
*
* To get a fresh transaction, we want to log the existing log done
* item to prevent the log intent item from replaying, immediately log
* a new log intent item with the unfinished work items, roll the
* transaction, and re-call ->finish_item wherever it left off. The
* log done item and the new log intent item must be in the same
* transaction or atomicity cannot be guaranteed; defer_finish ensures
* that this happens.
*
* This requires some coordination between ->finish_item and
* defer_finish. Upon deciding to request a new transaction,
* ->finish_item should update the current work item to reflect the
* unfinished work. Next, it should reset the log done item's list
* count to the number of items finished, and return -EAGAIN.
* defer_finish sees the -EAGAIN, logs the new log intent item
* with the remaining work items, and leaves the xfs_defer_pending
* item at the head of the dop_work queue. Then it rolls the
* transaction and picks up processing where it left off. It is
* required that ->finish_item must be careful to leave enough
* transaction reservation to fit the new log intent item.
*
* This is an example of remapping the extent (E, E+B) into file X at
* offset A and dealing with the extent (C, C+B) already being mapped
* there:
* +-------------------------------------------------+
* | Unmap file X startblock C offset A length B | t0
* | Intent to reduce refcount for extent (C, B) |
* | Intent to remove rmap (X, C, A, B) |
* | Intent to free extent (D, 1) (bmbt block) |
* | Intent to map (X, A, B) at startblock E |
* +-------------------------------------------------+
* | Map file X startblock E offset A length B | t1
* | Done mapping (X, E, A, B) |
* | Intent to increase refcount for extent (E, B) |
* | Intent to add rmap (X, E, A, B) |
* +-------------------------------------------------+
* | Reduce refcount for extent (C, B) | t2
* | Done reducing refcount for extent (C, 9) |
* | Intent to reduce refcount for extent (C+9, B-9) |
* | (ran out of space after 9 refcount updates) |
* +-------------------------------------------------+
* | Reduce refcount for extent (C+9, B+9) | t3
* | Done reducing refcount for extent (C+9, B-9) |
* | Increase refcount for extent (E, B) |
* | Done increasing refcount for extent (E, B) |
* | Intent to free extent (C, B) |
* | Intent to free extent (F, 1) (refcountbt block) |
* | Intent to remove rmap (F, 1, REFC) |
* +-------------------------------------------------+
* | Remove rmap (X, C, A, B) | t4
* | Done removing rmap (X, C, A, B) |
* | Add rmap (X, E, A, B) |
* | Done adding rmap (X, E, A, B) |
* | Remove rmap (F, 1, REFC) |
* | Done removing rmap (F, 1, REFC) |
* +-------------------------------------------------+
* | Free extent (C, B) | t5
* | Done freeing extent (C, B) |
* | Free extent (D, 1) |
* | Done freeing extent (D, 1) |
* | Free extent (F, 1) |
* | Done freeing extent (F, 1) |
* +-------------------------------------------------+
*
* If we should crash before t2 commits, log recovery replays
* the following intent items:
*
* - Intent to reduce refcount for extent (C, B)
* - Intent to remove rmap (X, C, A, B)
* - Intent to free extent (D, 1) (bmbt block)
* - Intent to increase refcount for extent (E, B)
* - Intent to add rmap (X, E, A, B)
*
* In the process of recovering, it should also generate and take care
* of these intent items:
*
* - Intent to free extent (C, B)
* - Intent to free extent (F, 1) (refcountbt block)
* - Intent to remove rmap (F, 1, REFC)
*
* Note that the continuation requested between t2 and t3 is likely to
* reoccur.
*/
static const struct xfs_defer_op_type *defer_op_types[XFS_DEFER_OPS_TYPE_MAX];
/*
* For each pending item in the intake list, log its intent item and the
* associated extents, then add the entire intake list to the end of
* the pending list.
*/
STATIC void
xfs_defer_intake_work(
struct xfs_trans *tp,
struct xfs_defer_ops *dop)
{
struct list_head *li;
struct xfs_defer_pending *dfp;
list_for_each_entry(dfp, &dop->dop_intake, dfp_list) {
dfp->dfp_intent = dfp->dfp_type->create_intent(tp,
dfp->dfp_count);
trace_xfs_defer_intake_work(tp->t_mountp, dfp);
list_sort(tp->t_mountp, &dfp->dfp_work,
dfp->dfp_type->diff_items);
list_for_each(li, &dfp->dfp_work)
dfp->dfp_type->log_item(tp, dfp->dfp_intent, li);
}
list_splice_tail_init(&dop->dop_intake, &dop->dop_pending);
}
/* Abort all the intents that were committed. */
STATIC void
xfs_defer_trans_abort(
struct xfs_trans *tp,
struct xfs_defer_ops *dop,
int error)
{
struct xfs_defer_pending *dfp;
trace_xfs_defer_trans_abort(tp->t_mountp, dop, _RET_IP_);
/* Abort intent items that don't have a done item. */
list_for_each_entry(dfp, &dop->dop_pending, dfp_list) {
trace_xfs_defer_pending_abort(tp->t_mountp, dfp);
if (dfp->dfp_intent && !dfp->dfp_done) {
dfp->dfp_type->abort_intent(dfp->dfp_intent);
dfp->dfp_intent = NULL;
}
}
/* Shut down FS. */
xfs_force_shutdown(tp->t_mountp, (error == -EFSCORRUPTED) ?
SHUTDOWN_CORRUPT_INCORE : SHUTDOWN_META_IO_ERROR);
}
/* Roll a transaction so we can do some deferred op processing. */
STATIC int
xfs_defer_trans_roll(
struct xfs_trans **tp)
{
struct xfs_defer_ops *dop = (*tp)->t_dfops;
int i;
int error;
/* Log all the joined inodes. */
for (i = 0; i < XFS_DEFER_OPS_NR_INODES && dop->dop_inodes[i]; i++)
xfs_trans_log_inode(*tp, dop->dop_inodes[i], XFS_ILOG_CORE);
/* Hold the (previously bjoin'd) buffer locked across the roll. */
for (i = 0; i < XFS_DEFER_OPS_NR_BUFS && dop->dop_bufs[i]; i++)
xfs_trans_dirty_buf(*tp, dop->dop_bufs[i]);
trace_xfs_defer_trans_roll((*tp)->t_mountp, dop, _RET_IP_);
/* Roll the transaction. */
error = xfs_trans_roll(tp);
dop = (*tp)->t_dfops;
if (error) {
trace_xfs_defer_trans_roll_error((*tp)->t_mountp, dop, error);
xfs_defer_trans_abort(*tp, dop, error);
return error;
}
/* Rejoin the joined inodes. */
for (i = 0; i < XFS_DEFER_OPS_NR_INODES && dop->dop_inodes[i]; i++)
xfs_trans_ijoin(*tp, dop->dop_inodes[i], 0);
/* Rejoin the buffers and dirty them so the log moves forward. */
for (i = 0; i < XFS_DEFER_OPS_NR_BUFS && dop->dop_bufs[i]; i++) {
xfs_trans_bjoin(*tp, dop->dop_bufs[i]);
xfs_trans_bhold(*tp, dop->dop_bufs[i]);
}
return error;
}
/* Do we have any work items to finish? */
bool
xfs_defer_has_unfinished_work(
struct xfs_defer_ops *dop)
{
return !list_empty(&dop->dop_pending) || !list_empty(&dop->dop_intake);
}
/*
* Add this inode to the deferred op. Each joined inode is relogged
* each time we roll the transaction.
*/
int
xfs_defer_ijoin(
struct xfs_defer_ops *dop,
struct xfs_inode *ip)
{
int i;
for (i = 0; i < XFS_DEFER_OPS_NR_INODES; i++) {
if (dop->dop_inodes[i] == ip)
return 0;
else if (dop->dop_inodes[i] == NULL) {
dop->dop_inodes[i] = ip;
return 0;
}
}
ASSERT(0);
return -EFSCORRUPTED;
}
/*
* Add this buffer to the deferred op. Each joined buffer is relogged
* each time we roll the transaction.
*/
int
xfs_defer_bjoin(
struct xfs_defer_ops *dop,
struct xfs_buf *bp)
{
int i;
for (i = 0; i < XFS_DEFER_OPS_NR_BUFS; i++) {
if (dop->dop_bufs[i] == bp)
return 0;
else if (dop->dop_bufs[i] == NULL) {
dop->dop_bufs[i] = bp;
return 0;
}
}
ASSERT(0);
return -EFSCORRUPTED;
}
/*
* Reset an already used dfops after finish.
*/
static void
xfs_defer_reset(
struct xfs_defer_ops *dop)
{
ASSERT(!xfs_defer_has_unfinished_work(dop));
dop->dop_low = false;
memset(dop->dop_inodes, 0, sizeof(dop->dop_inodes));
memset(dop->dop_bufs, 0, sizeof(dop->dop_bufs));
}
/*
* Finish all the pending work. This involves logging intent items for
* any work items that wandered in since the last transaction roll (if
* one has even happened), rolling the transaction, and finishing the
* work items in the first item on the logged-and-pending list.
*
* If an inode is provided, relog it to the new transaction.
*/
int
xfs_defer_finish(
struct xfs_trans **tp)
{
struct xfs_defer_ops *dop = (*tp)->t_dfops;
struct xfs_defer_pending *dfp;
struct list_head *li;
struct list_head *n;
void *state;
int error = 0;
void (*cleanup_fn)(struct xfs_trans *, void *, int);
ASSERT((*tp)->t_flags & XFS_TRANS_PERM_LOG_RES);
trace_xfs_defer_finish((*tp)->t_mountp, dop, _RET_IP_);
/* Until we run out of pending work to finish... */
while (xfs_defer_has_unfinished_work(dop)) {
/* Log intents for work items sitting in the intake. */
xfs_defer_intake_work(*tp, dop);
/*
* Roll the transaction and update dop in case dfops was
* embedded in the transaction.
*/
error = xfs_defer_trans_roll(tp);
if (error)
goto out;
dop = (*tp)->t_dfops;
/* Log an intent-done item for the first pending item. */
dfp = list_first_entry(&dop->dop_pending,
struct xfs_defer_pending, dfp_list);
trace_xfs_defer_pending_finish((*tp)->t_mountp, dfp);
dfp->dfp_done = dfp->dfp_type->create_done(*tp, dfp->dfp_intent,
dfp->dfp_count);
cleanup_fn = dfp->dfp_type->finish_cleanup;
/* Finish the work items. */
state = NULL;
list_for_each_safe(li, n, &dfp->dfp_work) {
list_del(li);
dfp->dfp_count--;
error = dfp->dfp_type->finish_item(*tp, dop, li,
dfp->dfp_done, &state);
if (error == -EAGAIN) {
/*
* Caller wants a fresh transaction;
* put the work item back on the list
* and jump out.
*/
list_add(li, &dfp->dfp_work);
dfp->dfp_count++;
break;
} else if (error) {
/*
* Clean up after ourselves and jump out.
* xfs_defer_cancel will take care of freeing
* all these lists and stuff.
*/
if (cleanup_fn)
cleanup_fn(*tp, state, error);
xfs_defer_trans_abort(*tp, dop, error);
goto out;
}
}
if (error == -EAGAIN) {
/*
* Caller wants a fresh transaction, so log a
* new log intent item to replace the old one
* and roll the transaction. See "Requesting
* a Fresh Transaction while Finishing
* Deferred Work" above.
*/
dfp->dfp_intent = dfp->dfp_type->create_intent(*tp,
dfp->dfp_count);
dfp->dfp_done = NULL;
list_for_each(li, &dfp->dfp_work)
dfp->dfp_type->log_item(*tp, dfp->dfp_intent,
li);
} else {
/* Done with the dfp, free it. */
list_del(&dfp->dfp_list);
kmem_free(dfp);
}
if (cleanup_fn)
cleanup_fn(*tp, state, error);
}
xfs: return from _defer_finish with a clean transaction The following assertion was seen on generic/051: XFS: Assertion failed: tp->t_firstblock == NULLFSBLOCK, file: fs/xfs/libxfs5 ------------[ cut here ]------------ kernel BUG at fs/xfs/xfs_message.c:102! invalid opcode: 0000 [#1] SMP PTI CPU: 2 PID: 20757 Comm: fsstress Not tainted 4.18.0-rc4+ #3969 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.11.1-1 04/01/4 RIP: 0010:assfail+0x23/0x30 Code: c3 66 0f 1f 44 00 00 48 89 f1 41 89 d0 48 c7 c6 88 e0 8c 82 48 89 fa RSP: 0018:ffff88012dc43c08 EFLAGS: 00010202 RAX: 0000000000000000 RBX: ffff88012dc43ca0 RCX: 0000000000000000 RDX: 00000000ffffffc0 RSI: 000000000000000a RDI: ffffffff828480eb RBP: ffff88012aa92758 R08: 0000000000000000 R09: 0000000000000000 R10: 0000000000000000 R11: f000000000000000 R12: 0000000000000000 R13: ffff88012dc43d48 R14: ffff88013092e7e8 R15: 0000000000000014 FS: 00007f8d689b8e80(0000) GS:ffff88013fd00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007f8d689c7000 CR3: 000000012ba6a000 CR4: 00000000000006e0 Call Trace: xfs_defer_init+0xff/0x160 xfs_reflink_remap_extent+0x31b/0xa00 xfs_reflink_remap_blocks+0xec/0x4a0 xfs_reflink_remap_range+0x3a1/0x650 xfs_file_dedupe_range+0x39/0x50 vfs_dedupe_file_range+0x218/0x260 do_vfs_ioctl+0x262/0x6a0 ? __se_sys_newfstat+0x3c/0x60 ksys_ioctl+0x35/0x60 __x64_sys_ioctl+0x11/0x20 do_syscall_64+0x4b/0x190 entry_SYSCALL_64_after_hwframe+0x49/0xbe The root cause of the assertion failure is that xfs_defer_finish doesn't roll the transaction after processing all the deferred items. Therefore it returns a dirty transaction to the caller, which leaves the caller at risk of exceeding the transaction reservation if it logs more items. Brian Foster's patchset to move the defer_ops firstblock into the transaction requires t_firstblock == NULLFSBLOCK upon defer_ops initialization, which is how this was noticed at all. Reported-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> Reviewed-by: Brian Foster <bfoster@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2018-07-20 03:25:47 +08:00
/*
* Roll the transaction once more to avoid returning to the caller
* with a dirty transaction.
*/
if ((*tp)->t_flags & XFS_TRANS_DIRTY) {
error = xfs_defer_trans_roll(tp);
dop = (*tp)->t_dfops;
}
out:
if (error) {
trace_xfs_defer_finish_error((*tp)->t_mountp, dop, error);
} else {
trace_xfs_defer_finish_done((*tp)->t_mountp, dop, _RET_IP_);
xfs_defer_reset(dop);
}
return error;
}
/*
* Free up any items left in the list.
*/
void
__xfs_defer_cancel(
struct xfs_defer_ops *dop)
{
struct xfs_defer_pending *dfp;
struct xfs_defer_pending *pli;
struct list_head *pwi;
struct list_head *n;
trace_xfs_defer_cancel(NULL, dop, _RET_IP_);
/*
* Free the pending items. Caller should already have arranged
* for the intent items to be released.
*/
list_for_each_entry_safe(dfp, pli, &dop->dop_intake, dfp_list) {
trace_xfs_defer_intake_cancel(NULL, dfp);
list_del(&dfp->dfp_list);
list_for_each_safe(pwi, n, &dfp->dfp_work) {
list_del(pwi);
dfp->dfp_count--;
dfp->dfp_type->cancel_item(pwi);
}
ASSERT(dfp->dfp_count == 0);
kmem_free(dfp);
}
list_for_each_entry_safe(dfp, pli, &dop->dop_pending, dfp_list) {
trace_xfs_defer_pending_cancel(NULL, dfp);
list_del(&dfp->dfp_list);
list_for_each_safe(pwi, n, &dfp->dfp_work) {
list_del(pwi);
dfp->dfp_count--;
dfp->dfp_type->cancel_item(pwi);
}
ASSERT(dfp->dfp_count == 0);
kmem_free(dfp);
}
}
/* Add an item for later deferred processing. */
void
xfs_defer_add(
struct xfs_defer_ops *dop,
enum xfs_defer_ops_type type,
struct list_head *li)
{
struct xfs_defer_pending *dfp = NULL;
/*
* Add the item to a pending item at the end of the intake list.
* If the last pending item has the same type, reuse it. Else,
* create a new pending item at the end of the intake list.
*/
if (!list_empty(&dop->dop_intake)) {
dfp = list_last_entry(&dop->dop_intake,
struct xfs_defer_pending, dfp_list);
if (dfp->dfp_type->type != type ||
(dfp->dfp_type->max_items &&
dfp->dfp_count >= dfp->dfp_type->max_items))
dfp = NULL;
}
if (!dfp) {
dfp = kmem_alloc(sizeof(struct xfs_defer_pending),
KM_SLEEP | KM_NOFS);
dfp->dfp_type = defer_op_types[type];
dfp->dfp_intent = NULL;
dfp->dfp_done = NULL;
dfp->dfp_count = 0;
INIT_LIST_HEAD(&dfp->dfp_work);
list_add_tail(&dfp->dfp_list, &dop->dop_intake);
}
list_add_tail(li, &dfp->dfp_work);
dfp->dfp_count++;
}
/* Initialize a deferred operation list. */
void
xfs_defer_init_op_type(
const struct xfs_defer_op_type *type)
{
defer_op_types[type->type] = type;
}
/* Initialize a deferred operation. */
void
xfs_defer_init(
struct xfs_trans *tp,
struct xfs_defer_ops *dop)
{
struct xfs_mount *mp = NULL;
memset(dop, 0, sizeof(struct xfs_defer_ops));
INIT_LIST_HEAD(&dop->dop_intake);
INIT_LIST_HEAD(&dop->dop_pending);
if (tp) {
ASSERT(tp->t_firstblock == NULLFSBLOCK);
tp->t_dfops = dop;
mp = tp->t_mountp;
}
trace_xfs_defer_init(mp, dop, _RET_IP_);
}
xfs: support embedded dfops in transaction The dfops structure used by multi-transaction operations is typically stored on the stack and carried around by the associated transaction. The lifecycle of dfops does not quite match that of the transaction, but they are tightly related in that the former depends on the latter. The relationship of these objects is tight enough that we can avoid the cumbersome boilerplate code required in most cases to manage them separately by just embedding an xfs_defer_ops in the transaction itself. This means that a transaction allocation returns with an initialized dfops, a transaction commit finishes pending deferred items before the tx commit, a transaction cancel cancels the dfops before the transaction and a transaction dup operation transfers the current dfops state to the new transaction. The dup operation is slightly complicated by the fact that we can no longer just copy a dfops pointer from the old transaction to the new transaction. This is solved through a dfops move helper that transfers the pending items and other dfops state across the transactions. This also requires that transaction rolling code always refer to the transaction for the current dfops reference. Finally, to facilitate incremental conversion to the internal dfops and continue to support the current external dfops mode of operation, create the new ->t_dfops_internal field with a layer of indirection. On allocation, ->t_dfops points to the internal dfops. This state is overridden by callers who re-init a local dfops on the transaction. Once ->t_dfops is overridden, the external dfops reference is maintained as the transaction rolls. This patch adds the fundamental ability to support an internal dfops. All codepaths that perform deferred processing continue to override the internal dfops until they are converted over in subsequent patches. Signed-off-by: Brian Foster <bfoster@redhat.com> Reviewed-by: Bill O'Donnell <billodo@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
2018-07-25 04:43:11 +08:00
/*
* Move state from one xfs_defer_ops to another and reset the source to initial
* state. This is primarily used to carry state forward across transaction rolls
* with internal dfops.
*/
void
xfs_defer_move(
struct xfs_defer_ops *dst,
struct xfs_defer_ops *src)
{
ASSERT(dst != src);
list_splice_init(&src->dop_intake, &dst->dop_intake);
list_splice_init(&src->dop_pending, &dst->dop_pending);
memcpy(dst->dop_inodes, src->dop_inodes, sizeof(dst->dop_inodes));
memcpy(dst->dop_bufs, src->dop_bufs, sizeof(dst->dop_bufs));
dst->dop_low = src->dop_low;
xfs_defer_reset(src);
}