| Message ID | 20180221022826.7020-2-chandan@linux.vnet.ibm.com (mailing list archive) |
|---|---|
| State | Deferred, archived |
| Headers | show |
On Wed, Feb 21, 2018 at 07:58:26AM +0530, Chandan Rajendra wrote: > generic/388 can cause the following "use after free" error to occur, > > ============================================================================= > BUG xfs_efi_item (Not tainted): Poison overwritten > ----------------------------------------------------------------------------- > > Disabling lock debugging due to kernel taint > INFO: 0x00000000292c4bd4-0x00000000292c4bd4. First byte 0x6a instead of 0x6b > INFO: Allocated in .kmem_zone_alloc+0xcc/0x190 age=79 cpu=0 pid=12436 > .__slab_alloc+0x54/0x80 > .kmem_cache_alloc+0x124/0x350 > .kmem_zone_alloc+0xcc/0x190 > .xfs_efi_init+0x48/0xf0 > .xfs_extent_free_create_intent+0x40/0x130 > .xfs_defer_intake_work+0x74/0x1e0 > .xfs_defer_finish+0xac/0x5c0 > .xfs_itruncate_extents+0x170/0x590 > .xfs_inactive_truncate+0xcc/0x170 > .xfs_inactive+0x1d8/0x2f0 > .xfs_fs_destroy_inode+0xe4/0x3d0 > .destroy_inode+0x68/0xb0 > .do_unlinkat+0x1e8/0x390 > system_call+0x58/0x6c > INFO: Freed in .xfs_efi_item_free+0x44/0x80 age=79 cpu=0 pid=12436 > .kmem_cache_free+0x120/0x2b0 > .xfs_efi_item_free+0x44/0x80 > .xfs_trans_free_items+0xd4/0x130 > .__xfs_trans_commit+0xd0/0x350 > .xfs_trans_roll+0x4c/0x90 > .xfs_defer_trans_roll+0xa4/0x2b0 > .xfs_defer_finish+0xb8/0x5c0 > .xfs_itruncate_extents+0x170/0x590 > .xfs_inactive_truncate+0xcc/0x170 > .xfs_inactive+0x1d8/0x2f0 > .xfs_fs_destroy_inode+0xe4/0x3d0 > .destroy_inode+0x68/0xb0 > .do_unlinkat+0x1e8/0x390 > system_call+0x58/0x6c > > This happens due to the following interaction, > 1. xfs_defer_finish() creates "extent free" intent item and adds it to the > per-transction list of log items. > 2. xfs_defer_trans_roll() invokes __xfs_trans_commit(). Here, if the > XFS_MOUNT_FS_SHUTDOWN flag is set, we invoke io_unlock() operation > for each of the log items in the per-transction list. For "extent > free" log items xfs_efi_item_unlock() gets invoked which then frees > the xfs_efi_log_item. Isn't this the problem? i.e. it's freeing the EFI item because the abort flag is set, but there are still other references to it? Isn't this wrong now that we have a cleanup function that will free the EFI is the commit fails? i.e: this .... > 3. xfs_defer_trans_roll() then invokes xfs_defer_trans_abort(). Since the > xfs_defer_pending->dfp_intent is still set to the "extent free" intent > item, we invoke xfs_extent_free_abort_intent(). This accesses the > previously freed xfs_efi_log_item to decrement the ref count. ... will free the EFI correctly on transaction commit failure and so we should be able to remove the "free immediately" hack from the iop_unlock() code that we used to need because there was no external reference that could free the EFI on commit failure.... > This commit fixes the bug by invoking xfs_defer_trans_abort() only when > the log items in the per-transaction list have been committed to the > CIL. The log item "committed" status is being tracked by > xfs_defer_ops->dop_committed. This was the behaviour prior to commit > 3ab78df2a59a485f479d26852a060acfd8c4ecd7 (xfs: rework xfs_bmap_free > callers to use xfs_defer_ops). Freeing in ->iop_unlock was always troublesome. We should get rid of it, not go back to it. Cheers, Dave.
On Wednesday, February 21, 2018 9:31:09 AM IST Dave Chinner wrote: > On Wed, Feb 21, 2018 at 07:58:26AM +0530, Chandan Rajendra wrote: > > generic/388 can cause the following "use after free" error to occur, > > > > ============================================================================= > > BUG xfs_efi_item (Not tainted): Poison overwritten > > ----------------------------------------------------------------------------- > > > > Disabling lock debugging due to kernel taint > > INFO: 0x00000000292c4bd4-0x00000000292c4bd4. First byte 0x6a instead of 0x6b > > INFO: Allocated in .kmem_zone_alloc+0xcc/0x190 age=79 cpu=0 pid=12436 > > .__slab_alloc+0x54/0x80 > > .kmem_cache_alloc+0x124/0x350 > > .kmem_zone_alloc+0xcc/0x190 > > .xfs_efi_init+0x48/0xf0 > > .xfs_extent_free_create_intent+0x40/0x130 > > .xfs_defer_intake_work+0x74/0x1e0 > > .xfs_defer_finish+0xac/0x5c0 > > .xfs_itruncate_extents+0x170/0x590 > > .xfs_inactive_truncate+0xcc/0x170 > > .xfs_inactive+0x1d8/0x2f0 > > .xfs_fs_destroy_inode+0xe4/0x3d0 > > .destroy_inode+0x68/0xb0 > > .do_unlinkat+0x1e8/0x390 > > system_call+0x58/0x6c > > INFO: Freed in .xfs_efi_item_free+0x44/0x80 age=79 cpu=0 pid=12436 > > .kmem_cache_free+0x120/0x2b0 > > .xfs_efi_item_free+0x44/0x80 > > .xfs_trans_free_items+0xd4/0x130 > > .__xfs_trans_commit+0xd0/0x350 > > .xfs_trans_roll+0x4c/0x90 > > .xfs_defer_trans_roll+0xa4/0x2b0 > > .xfs_defer_finish+0xb8/0x5c0 > > .xfs_itruncate_extents+0x170/0x590 > > .xfs_inactive_truncate+0xcc/0x170 > > .xfs_inactive+0x1d8/0x2f0 > > .xfs_fs_destroy_inode+0xe4/0x3d0 > > .destroy_inode+0x68/0xb0 > > .do_unlinkat+0x1e8/0x390 > > system_call+0x58/0x6c > > > > This happens due to the following interaction, > > 1. xfs_defer_finish() creates "extent free" intent item and adds it to the > > per-transction list of log items. > > 2. xfs_defer_trans_roll() invokes __xfs_trans_commit(). Here, if the > > XFS_MOUNT_FS_SHUTDOWN flag is set, we invoke io_unlock() operation > > for each of the log items in the per-transction list. For "extent > > free" log items xfs_efi_item_unlock() gets invoked which then frees > > the xfs_efi_log_item. > > Isn't this the problem? i.e. it's freeing the EFI item because the > abort flag is set, but there are still other references to it? Isn't > this wrong now that we have a cleanup function that will free the > EFI is the commit fails? i.e: this .... > > > 3. xfs_defer_trans_roll() then invokes xfs_defer_trans_abort(). Since the > > xfs_defer_pending->dfp_intent is still set to the "extent free" intent > > item, we invoke xfs_extent_free_abort_intent(). This accesses the > > previously freed xfs_efi_log_item to decrement the ref count. > > ... will free the EFI correctly on transaction commit failure and so > we should be able to remove the "free immediately" hack from the > iop_unlock() code that we used to need because there was no external > reference that could free the EFI on commit failure.... The EFI log item would have its ref count initialized to 2 (One for the CIL/AIL and one for EFD). In the "use after free" case the EFI was not committed to the CIL. At this point, we don't have any entity owning a reference to the log item since the EFI was not committed to the CIL nor was the EFD created. Hence IMHO, freeing the log item is the right approach. Also, the cleanup function (i.e. xfs_extent_free_abort_intent()) would decrement the ref count by 1 i.e. Its refcount would go from the initial value of 2 to 1. Hence the cleanup function won't free the EFI log item causing a memory leak. > > > This commit fixes the bug by invoking xfs_defer_trans_abort() only when > > the log items in the per-transaction list have been committed to the > > CIL. The log item "committed" status is being tracked by > > xfs_defer_ops->dop_committed. This was the behaviour prior to commit > > 3ab78df2a59a485f479d26852a060acfd8c4ecd7 (xfs: rework xfs_bmap_free > > callers to use xfs_defer_ops). > > Freeing in ->iop_unlock was always troublesome. We should get > rid of it, not go back to it. > > Cheers, > > Dave. >
On Wed, Feb 21, 2018 at 11:15:22AM +0530, Chandan Rajendra wrote: > On Wednesday, February 21, 2018 9:31:09 AM IST Dave Chinner wrote: > > On Wed, Feb 21, 2018 at 07:58:26AM +0530, Chandan Rajendra wrote: > > > generic/388 can cause the following "use after free" error to occur, > > > > > > ============================================================================= > > > BUG xfs_efi_item (Not tainted): Poison overwritten > > > ----------------------------------------------------------------------------- > > > > > > Disabling lock debugging due to kernel taint > > > INFO: 0x00000000292c4bd4-0x00000000292c4bd4. First byte 0x6a instead of 0x6b > > > INFO: Allocated in .kmem_zone_alloc+0xcc/0x190 age=79 cpu=0 pid=12436 > > > .__slab_alloc+0x54/0x80 > > > .kmem_cache_alloc+0x124/0x350 > > > .kmem_zone_alloc+0xcc/0x190 > > > .xfs_efi_init+0x48/0xf0 > > > .xfs_extent_free_create_intent+0x40/0x130 > > > .xfs_defer_intake_work+0x74/0x1e0 > > > .xfs_defer_finish+0xac/0x5c0 > > > .xfs_itruncate_extents+0x170/0x590 > > > .xfs_inactive_truncate+0xcc/0x170 > > > .xfs_inactive+0x1d8/0x2f0 > > > .xfs_fs_destroy_inode+0xe4/0x3d0 > > > .destroy_inode+0x68/0xb0 > > > .do_unlinkat+0x1e8/0x390 > > > system_call+0x58/0x6c > > > INFO: Freed in .xfs_efi_item_free+0x44/0x80 age=79 cpu=0 pid=12436 > > > .kmem_cache_free+0x120/0x2b0 > > > .xfs_efi_item_free+0x44/0x80 > > > .xfs_trans_free_items+0xd4/0x130 > > > .__xfs_trans_commit+0xd0/0x350 > > > .xfs_trans_roll+0x4c/0x90 > > > .xfs_defer_trans_roll+0xa4/0x2b0 > > > .xfs_defer_finish+0xb8/0x5c0 > > > .xfs_itruncate_extents+0x170/0x590 > > > .xfs_inactive_truncate+0xcc/0x170 > > > .xfs_inactive+0x1d8/0x2f0 > > > .xfs_fs_destroy_inode+0xe4/0x3d0 > > > .destroy_inode+0x68/0xb0 > > > .do_unlinkat+0x1e8/0x390 > > > system_call+0x58/0x6c > > > > > > This happens due to the following interaction, > > > 1. xfs_defer_finish() creates "extent free" intent item and adds it to the > > > per-transction list of log items. > > > 2. xfs_defer_trans_roll() invokes __xfs_trans_commit(). Here, if the > > > XFS_MOUNT_FS_SHUTDOWN flag is set, we invoke io_unlock() operation > > > for each of the log items in the per-transction list. For "extent > > > free" log items xfs_efi_item_unlock() gets invoked which then frees > > > the xfs_efi_log_item. > > > > Isn't this the problem? i.e. it's freeing the EFI item because the > > abort flag is set, but there are still other references to it? Isn't > > this wrong now that we have a cleanup function that will free the > > EFI is the commit fails? i.e: this .... > > > > > 3. xfs_defer_trans_roll() then invokes xfs_defer_trans_abort(). Since the > > > xfs_defer_pending->dfp_intent is still set to the "extent free" intent > > > item, we invoke xfs_extent_free_abort_intent(). This accesses the > > > previously freed xfs_efi_log_item to decrement the ref count. > > > > ... will free the EFI correctly on transaction commit failure and so > > we should be able to remove the "free immediately" hack from the > > iop_unlock() code that we used to need because there was no external > > reference that could free the EFI on commit failure.... > > The EFI log item would have its ref count initialized to 2 (One for the > CIL/AIL and one for EFD). And that's part of what has always been wrong with this code - it takes references for things that don't really have references yet, instead of taking them when the EFI is actually referenced. Hence we have to hack around that with things like ignoring the reference count when we get an abort signal, because we haven't reference counted the object properly. > In the "use after free" case the EFI was not > committed to the CIL. At this point, we don't have any entity owning a > reference to the log item since the EFI was not committed to the CIL nor was > the EFD created. Hence IMHO, freeing the log item is the right approach. No, we clearly have a reference - the deferops structure has a reference to it. If it didn't have a reference, then we wouldn't have a use after free situation.... > Also, the cleanup function (i.e. xfs_extent_free_abort_intent()) would > decrement the ref count by 1 i.e. Its refcount would go from the initial value > of 2 to 1. Hence the cleanup function won't free the EFI log item causing a > memory leak. No, then the xfs_extent_free_abort_intent() call releases it, dropping the remaining reference that was intended for the EFD which, because we aborted, will never take over control of that reference. i.e. the two references that are created at initialisation are for the EFI itself as it passes through the CIL and journal commit, and the second reference is used by the defer_ops it is attached to and then handed to the EFD for it's pass through the CIL and journal commit. If we never commit the EFI, then we need to release that reference (on abort), and then we need to clean up the extra reference that is destined for the EFD which is held by the deferops. That's done by xfs_extent_free_abort_intent(). And I'm guessing that we need to make sure the same fix goes through all the other items that are used as deferops intents that were modeled on the EFI/EFD infrastructure, too. Cheers, Dave.
On Thursday, February 22, 2018 2:34:04 AM IST Dave Chinner wrote: > On Wed, Feb 21, 2018 at 11:15:22AM +0530, Chandan Rajendra wrote: > > On Wednesday, February 21, 2018 9:31:09 AM IST Dave Chinner wrote: > > > On Wed, Feb 21, 2018 at 07:58:26AM +0530, Chandan Rajendra wrote: > > > > generic/388 can cause the following "use after free" error to occur, > > > > > > > > ============================================================================= > > > > BUG xfs_efi_item (Not tainted): Poison overwritten > > > > ----------------------------------------------------------------------------- > > > > > > > > Disabling lock debugging due to kernel taint > > > > INFO: 0x00000000292c4bd4-0x00000000292c4bd4. First byte 0x6a instead of 0x6b > > > > INFO: Allocated in .kmem_zone_alloc+0xcc/0x190 age=79 cpu=0 pid=12436 > > > > .__slab_alloc+0x54/0x80 > > > > .kmem_cache_alloc+0x124/0x350 > > > > .kmem_zone_alloc+0xcc/0x190 > > > > .xfs_efi_init+0x48/0xf0 > > > > .xfs_extent_free_create_intent+0x40/0x130 > > > > .xfs_defer_intake_work+0x74/0x1e0 > > > > .xfs_defer_finish+0xac/0x5c0 > > > > .xfs_itruncate_extents+0x170/0x590 > > > > .xfs_inactive_truncate+0xcc/0x170 > > > > .xfs_inactive+0x1d8/0x2f0 > > > > .xfs_fs_destroy_inode+0xe4/0x3d0 > > > > .destroy_inode+0x68/0xb0 > > > > .do_unlinkat+0x1e8/0x390 > > > > system_call+0x58/0x6c > > > > INFO: Freed in .xfs_efi_item_free+0x44/0x80 age=79 cpu=0 pid=12436 > > > > .kmem_cache_free+0x120/0x2b0 > > > > .xfs_efi_item_free+0x44/0x80 > > > > .xfs_trans_free_items+0xd4/0x130 > > > > .__xfs_trans_commit+0xd0/0x350 > > > > .xfs_trans_roll+0x4c/0x90 > > > > .xfs_defer_trans_roll+0xa4/0x2b0 > > > > .xfs_defer_finish+0xb8/0x5c0 > > > > .xfs_itruncate_extents+0x170/0x590 > > > > .xfs_inactive_truncate+0xcc/0x170 > > > > .xfs_inactive+0x1d8/0x2f0 > > > > .xfs_fs_destroy_inode+0xe4/0x3d0 > > > > .destroy_inode+0x68/0xb0 > > > > .do_unlinkat+0x1e8/0x390 > > > > system_call+0x58/0x6c > > > > > > > > This happens due to the following interaction, > > > > 1. xfs_defer_finish() creates "extent free" intent item and adds it to the > > > > per-transction list of log items. > > > > 2. xfs_defer_trans_roll() invokes __xfs_trans_commit(). Here, if the > > > > XFS_MOUNT_FS_SHUTDOWN flag is set, we invoke io_unlock() operation > > > > for each of the log items in the per-transction list. For "extent > > > > free" log items xfs_efi_item_unlock() gets invoked which then frees > > > > the xfs_efi_log_item. > > > > > > Isn't this the problem? i.e. it's freeing the EFI item because the > > > abort flag is set, but there are still other references to it? Isn't > > > this wrong now that we have a cleanup function that will free the > > > EFI is the commit fails? i.e: this .... > > > > > > > 3. xfs_defer_trans_roll() then invokes xfs_defer_trans_abort(). Since the > > > > xfs_defer_pending->dfp_intent is still set to the "extent free" intent > > > > item, we invoke xfs_extent_free_abort_intent(). This accesses the > > > > previously freed xfs_efi_log_item to decrement the ref count. > > > > > > ... will free the EFI correctly on transaction commit failure and so > > > we should be able to remove the "free immediately" hack from the > > > iop_unlock() code that we used to need because there was no external > > > reference that could free the EFI on commit failure.... > > > > The EFI log item would have its ref count initialized to 2 (One for the > > CIL/AIL and one for EFD). > > And that's part of what has always been wrong with this code - it > takes references for things that don't really have references yet, > instead of taking them when the EFI is actually referenced. Hence we > have to hack around that with things like ignoring the reference > count when we get an abort signal, because we haven't reference > counted the object properly. > > > In the "use after free" case the EFI was not > > committed to the CIL. At this point, we don't have any entity owning a > > reference to the log item since the EFI was not committed to the CIL nor was > > the EFD created. Hence IMHO, freeing the log item is the right approach. > > No, we clearly have a reference - the deferops structure has a > reference to it. If it didn't have a reference, then we wouldn't > have a use after free situation.... > > > Also, the cleanup function (i.e. xfs_extent_free_abort_intent()) would > > decrement the ref count by 1 i.e. Its refcount would go from the initial value > > of 2 to 1. Hence the cleanup function won't free the EFI log item causing a > > memory leak. > > No, then the xfs_extent_free_abort_intent() call releases it, > dropping the remaining reference that was intended for the EFD > which, because we aborted, will never take over control of that > reference. > > i.e. the two references that are created at initialisation are for > the EFI itself as it passes through the CIL and journal commit, and > the second reference is used by the defer_ops it is attached to and > then handed to the EFD for it's pass through the CIL and journal > commit. > > If we never commit the EFI, then we need to release that reference > (on abort), and then we need to clean up the extra reference that is > destined for the EFD which is held by the deferops. That's done by > xfs_extent_free_abort_intent(). > > And I'm guessing that we need to make sure the same fix goes through > all the other items that are used as deferops intents that were > modeled on the EFI/EFD infrastructure, too. > I agree. Thanks for the explaination. I will work on writing up a patch to fix this issue for all the log items which have corresponding intent and done items.
diff --git a/fs/xfs/libxfs/xfs_defer.c b/fs/xfs/libxfs/xfs_defer.c index 087fea0..eb879a0 100644 --- a/fs/xfs/libxfs/xfs_defer.c +++ b/fs/xfs/libxfs/xfs_defer.c @@ -256,13 +256,13 @@ xfs_defer_trans_roll( trace_xfs_defer_trans_roll((*tp)->t_mountp, dop); /* Roll the transaction. */ - error = xfs_trans_roll(tp); + error = xfs_trans_roll(tp, &dop->dop_committed); if (error) { trace_xfs_defer_trans_roll_error((*tp)->t_mountp, dop, error); - xfs_defer_trans_abort(*tp, dop, error); + if (dop->dop_committed == true) + xfs_defer_trans_abort(*tp, dop, error); return error; } - dop->dop_committed = true; /* Rejoin the joined inodes. */ for (i = 0; i < XFS_DEFER_OPS_NR_INODES && dop->dop_inodes[i]; i++) diff --git a/fs/xfs/xfs_bmap_util.c b/fs/xfs/xfs_bmap_util.c index c83f549..9d84d36 100644 --- a/fs/xfs/xfs_bmap_util.c +++ b/fs/xfs/xfs_bmap_util.c @@ -1830,7 +1830,7 @@ xfs_swap_change_owner( if (error != -EAGAIN) break; - error = xfs_trans_roll(tpp); + error = xfs_trans_roll(tpp, NULL); if (error) break; tp = *tpp; diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c index 37d9426..9b74539 100644 --- a/fs/xfs/xfs_inode.c +++ b/fs/xfs/xfs_inode.c @@ -1047,7 +1047,7 @@ xfs_dir_ialloc( tp->t_flags &= ~(XFS_TRANS_DQ_DIRTY); } - code = xfs_trans_roll(&tp); + code = xfs_trans_roll(&tp, NULL); /* * Re-attach the quota info that we detached from prev trx. diff --git a/fs/xfs/xfs_trans.c b/fs/xfs/xfs_trans.c index 86f92df..1bf2505 100644 --- a/fs/xfs/xfs_trans.c +++ b/fs/xfs/xfs_trans.c @@ -1057,12 +1057,15 @@ xfs_trans_cancel( */ int xfs_trans_roll( - struct xfs_trans **tpp) + struct xfs_trans **tpp, + bool *committed) { struct xfs_trans *trans = *tpp; struct xfs_trans_res tres; int error; + if (committed) + *committed = false; /* * Copy the critical parameters from one trans to the next. */ @@ -1082,6 +1085,9 @@ xfs_trans_roll( if (error) return error; + if (committed) + *committed = true; + /* * Reserve space in the log for the next transaction. * This also pushes items in the "AIL", the list of logged items, diff --git a/fs/xfs/xfs_trans.h b/fs/xfs/xfs_trans.h index 9d542df..d4deb49 100644 --- a/fs/xfs/xfs_trans.h +++ b/fs/xfs/xfs_trans.h @@ -230,7 +230,7 @@ int xfs_trans_free_extent(struct xfs_trans *, struct xfs_efd_log_item *, xfs_fsblock_t, xfs_extlen_t, struct xfs_owner_info *); int xfs_trans_commit(struct xfs_trans *); -int xfs_trans_roll(struct xfs_trans **); +int xfs_trans_roll(struct xfs_trans **, bool *); int xfs_trans_roll_inode(struct xfs_trans **, struct xfs_inode *); void xfs_trans_cancel(xfs_trans_t *); int xfs_trans_ail_init(struct xfs_mount *); diff --git a/fs/xfs/xfs_trans_inode.c b/fs/xfs/xfs_trans_inode.c index 4a89da4..bedd5fd 100644 --- a/fs/xfs/xfs_trans_inode.c +++ b/fs/xfs/xfs_trans_inode.c @@ -147,7 +147,7 @@ xfs_trans_roll_inode( int error; xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE); - error = xfs_trans_roll(tpp); + error = xfs_trans_roll(tpp, NULL); if (!error) xfs_trans_ijoin(*tpp, ip, 0); return error;
generic/388 can cause the following "use after free" error to occur, ============================================================================= BUG xfs_efi_item (Not tainted): Poison overwritten ----------------------------------------------------------------------------- Disabling lock debugging due to kernel taint INFO: 0x00000000292c4bd4-0x00000000292c4bd4. First byte 0x6a instead of 0x6b INFO: Allocated in .kmem_zone_alloc+0xcc/0x190 age=79 cpu=0 pid=12436 .__slab_alloc+0x54/0x80 .kmem_cache_alloc+0x124/0x350 .kmem_zone_alloc+0xcc/0x190 .xfs_efi_init+0x48/0xf0 .xfs_extent_free_create_intent+0x40/0x130 .xfs_defer_intake_work+0x74/0x1e0 .xfs_defer_finish+0xac/0x5c0 .xfs_itruncate_extents+0x170/0x590 .xfs_inactive_truncate+0xcc/0x170 .xfs_inactive+0x1d8/0x2f0 .xfs_fs_destroy_inode+0xe4/0x3d0 .destroy_inode+0x68/0xb0 .do_unlinkat+0x1e8/0x390 system_call+0x58/0x6c INFO: Freed in .xfs_efi_item_free+0x44/0x80 age=79 cpu=0 pid=12436 .kmem_cache_free+0x120/0x2b0 .xfs_efi_item_free+0x44/0x80 .xfs_trans_free_items+0xd4/0x130 .__xfs_trans_commit+0xd0/0x350 .xfs_trans_roll+0x4c/0x90 .xfs_defer_trans_roll+0xa4/0x2b0 .xfs_defer_finish+0xb8/0x5c0 .xfs_itruncate_extents+0x170/0x590 .xfs_inactive_truncate+0xcc/0x170 .xfs_inactive+0x1d8/0x2f0 .xfs_fs_destroy_inode+0xe4/0x3d0 .destroy_inode+0x68/0xb0 .do_unlinkat+0x1e8/0x390 system_call+0x58/0x6c This happens due to the following interaction, 1. xfs_defer_finish() creates "extent free" intent item and adds it to the per-transction list of log items. 2. xfs_defer_trans_roll() invokes __xfs_trans_commit(). Here, if the XFS_MOUNT_FS_SHUTDOWN flag is set, we invoke io_unlock() operation for each of the log items in the per-transction list. For "extent free" log items xfs_efi_item_unlock() gets invoked which then frees the xfs_efi_log_item. 3. xfs_defer_trans_roll() then invokes xfs_defer_trans_abort(). Since the xfs_defer_pending->dfp_intent is still set to the "extent free" intent item, we invoke xfs_extent_free_abort_intent(). This accesses the previously freed xfs_efi_log_item to decrement the ref count. This commit fixes the bug by invoking xfs_defer_trans_abort() only when the log items in the per-transaction list have been committed to the CIL. The log item "committed" status is being tracked by xfs_defer_ops->dop_committed. This was the behaviour prior to commit 3ab78df2a59a485f479d26852a060acfd8c4ecd7 (xfs: rework xfs_bmap_free callers to use xfs_defer_ops). Reported-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Chandan Rajendra <chandan@linux.vnet.ibm.com> --- fs/xfs/libxfs/xfs_defer.c | 6 +++--- fs/xfs/xfs_bmap_util.c | 2 +- fs/xfs/xfs_inode.c | 2 +- fs/xfs/xfs_trans.c | 8 +++++++- fs/xfs/xfs_trans.h | 2 +- fs/xfs/xfs_trans_inode.c | 2 +- 6 files changed, 14 insertions(+), 8 deletions(-)