diff mbox series

[v15.1] mm, pmem, xfs: Introduce MF_MEM_PRE_REMOVE for unbind

Message ID 20231023072046.1626474-1-ruansy.fnst@fujitsu.com (mailing list archive)
State New
Headers show
Series [v15.1] mm, pmem, xfs: Introduce MF_MEM_PRE_REMOVE for unbind | expand

Commit Message

Shiyang Ruan Oct. 23, 2023, 7:20 a.m. UTC
Changes since v15:
 1. Rebased on v6.6-rc7

Now, if we suddenly remove a PMEM device(by calling unbind) which
contains FSDAX while programs are still accessing data in this device,
e.g.:
```
 $FSSTRESS_PROG -d $SCRATCH_MNT -n 99999 -p 4 &
 # $FSX_PROG -N 1000000 -o 8192 -l 500000 $SCRATCH_MNT/t001 &
 echo "pfn1.1" > /sys/bus/nd/drivers/nd_pmem/unbind
```
it could come into an unacceptable state:
  1. device has gone but mount point still exists, and umount will fail
       with "target is busy"
  2. programs will hang and cannot be killed
  3. may crash with NULL pointer dereference

To fix this, we introduce a MF_MEM_PRE_REMOVE flag to let it know that we
are going to remove the whole device, and make sure all related processes
could be notified so that they could end up gracefully.

This patch is inspired by Dan's "mm, dax, pmem: Introduce
dev_pagemap_failure()"[1].  With the help of dax_holder and
->notify_failure() mechanism, the pmem driver is able to ask filesystem
on it to unmap all files in use, and notify processes who are using
those files.

Call trace:
trigger unbind
 -> unbind_store()
  -> ... (skip)
   -> devres_release_all()
    -> kill_dax()
     -> dax_holder_notify_failure(dax_dev, 0, U64_MAX, MF_MEM_PRE_REMOVE)
      -> xfs_dax_notify_failure()
      `-> freeze_super()             // freeze (kernel call)
      `-> do xfs rmap
      ` -> mf_dax_kill_procs()
      `  -> collect_procs_fsdax()    // all associated processes
      `  -> unmap_and_kill()
      ` -> invalidate_inode_pages2_range() // drop file's cache
      `-> thaw_super()               // thaw (both kernel & user call)

Introduce MF_MEM_PRE_REMOVE to let filesystem know this is a remove
event.  Use the exclusive freeze/thaw[2] to lock the filesystem to prevent
new dax mapping from being created.  Do not shutdown filesystem directly
if configuration is not supported, or if failure range includes metadata
area.  Make sure all files and processes(not only the current progress)
are handled correctly.  Also drop the cache of associated files before
pmem is removed.

[1]: https://lore.kernel.org/linux-mm/161604050314.1463742.14151665140035795571.stgit@dwillia2-desk3.amr.corp.intel.com/
[2]: https://lore.kernel.org/linux-xfs/169116275623.3187159.16862410128731457358.stg-ugh@frogsfrogsfrogs/

Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
---
 drivers/dax/super.c         |   3 +-
 fs/xfs/xfs_notify_failure.c | 108 ++++++++++++++++++++++++++++++++++--
 include/linux/mm.h          |   1 +
 mm/memory-failure.c         |  21 +++++--
 4 files changed, 122 insertions(+), 11 deletions(-)
diff mbox series

Patch

diff --git a/drivers/dax/super.c b/drivers/dax/super.c
index 0da9232ea175..f4b635526345 100644
--- a/drivers/dax/super.c
+++ b/drivers/dax/super.c
@@ -326,7 +326,8 @@  void kill_dax(struct dax_device *dax_dev)
 		return;
 
 	if (dax_dev->holder_data != NULL)
-		dax_holder_notify_failure(dax_dev, 0, U64_MAX, 0);
+		dax_holder_notify_failure(dax_dev, 0, U64_MAX,
+				MF_MEM_PRE_REMOVE);
 
 	clear_bit(DAXDEV_ALIVE, &dax_dev->flags);
 	synchronize_srcu(&dax_srcu);
diff --git a/fs/xfs/xfs_notify_failure.c b/fs/xfs/xfs_notify_failure.c
index a7daa522e00f..fa50e5308292 100644
--- a/fs/xfs/xfs_notify_failure.c
+++ b/fs/xfs/xfs_notify_failure.c
@@ -22,6 +22,7 @@ 
 
 #include <linux/mm.h>
 #include <linux/dax.h>
+#include <linux/fs.h>
 
 struct xfs_failure_info {
 	xfs_agblock_t		startblock;
@@ -73,10 +74,16 @@  xfs_dax_failure_fn(
 	struct xfs_mount		*mp = cur->bc_mp;
 	struct xfs_inode		*ip;
 	struct xfs_failure_info		*notify = data;
+	struct address_space		*mapping;
+	pgoff_t				pgoff;
+	unsigned long			pgcnt;
 	int				error = 0;
 
 	if (XFS_RMAP_NON_INODE_OWNER(rec->rm_owner) ||
 	    (rec->rm_flags & (XFS_RMAP_ATTR_FORK | XFS_RMAP_BMBT_BLOCK))) {
+		/* Continue the query because this isn't a failure. */
+		if (notify->mf_flags & MF_MEM_PRE_REMOVE)
+			return 0;
 		notify->want_shutdown = true;
 		return 0;
 	}
@@ -92,14 +99,60 @@  xfs_dax_failure_fn(
 		return 0;
 	}
 
-	error = mf_dax_kill_procs(VFS_I(ip)->i_mapping,
-				  xfs_failure_pgoff(mp, rec, notify),
-				  xfs_failure_pgcnt(mp, rec, notify),
-				  notify->mf_flags);
+	mapping = VFS_I(ip)->i_mapping;
+	pgoff = xfs_failure_pgoff(mp, rec, notify);
+	pgcnt = xfs_failure_pgcnt(mp, rec, notify);
+
+	/* Continue the rmap query if the inode isn't a dax file. */
+	if (dax_mapping(mapping))
+		error = mf_dax_kill_procs(mapping, pgoff, pgcnt,
+					  notify->mf_flags);
+
+	/* Invalidate the cache in dax pages. */
+	if (notify->mf_flags & MF_MEM_PRE_REMOVE)
+		invalidate_inode_pages2_range(mapping, pgoff,
+					      pgoff + pgcnt - 1);
+
 	xfs_irele(ip);
 	return error;
 }
 
+static int
+xfs_dax_notify_failure_freeze(
+	struct xfs_mount	*mp)
+{
+	struct super_block	*sb = mp->m_super;
+	int			error;
+
+	error = freeze_super(sb, FREEZE_HOLDER_KERNEL);
+	if (error)
+		xfs_emerg(mp, "already frozen by kernel, err=%d", error);
+
+	return error;
+}
+
+static void
+xfs_dax_notify_failure_thaw(
+	struct xfs_mount	*mp,
+	bool			kernel_frozen)
+{
+	struct super_block	*sb = mp->m_super;
+	int			error;
+
+	if (kernel_frozen) {
+		error = thaw_super(sb, FREEZE_HOLDER_KERNEL);
+		if (error)
+			xfs_emerg(mp, "still frozen after notify failure, err=%d",
+				error);
+	}
+
+	/*
+	 * Also thaw userspace call anyway because the device is about to be
+	 * removed immediately.
+	 */
+	thaw_super(sb, FREEZE_HOLDER_USERSPACE);
+}
+
 static int
 xfs_dax_notify_ddev_failure(
 	struct xfs_mount	*mp,
@@ -112,15 +165,29 @@  xfs_dax_notify_ddev_failure(
 	struct xfs_btree_cur	*cur = NULL;
 	struct xfs_buf		*agf_bp = NULL;
 	int			error = 0;
+	bool			kernel_frozen = false;
 	xfs_fsblock_t		fsbno = XFS_DADDR_TO_FSB(mp, daddr);
 	xfs_agnumber_t		agno = XFS_FSB_TO_AGNO(mp, fsbno);
 	xfs_fsblock_t		end_fsbno = XFS_DADDR_TO_FSB(mp,
 							     daddr + bblen - 1);
 	xfs_agnumber_t		end_agno = XFS_FSB_TO_AGNO(mp, end_fsbno);
 
+	if (mf_flags & MF_MEM_PRE_REMOVE) {
+		xfs_info(mp, "Device is about to be removed!");
+		/*
+		 * Freeze fs to prevent new mappings from being created.
+		 * - Keep going on if others already hold the kernel forzen.
+		 * - Keep going on if other errors too because this device is
+		 *   starting to fail.
+		 * - If kernel frozen state is hold successfully here, thaw it
+		 *   here as well at the end.
+		 */
+		kernel_frozen = xfs_dax_notify_failure_freeze(mp) == 0;
+	}
+
 	error = xfs_trans_alloc_empty(mp, &tp);
 	if (error)
-		return error;
+		goto out;
 
 	for (; agno <= end_agno; agno++) {
 		struct xfs_rmap_irec	ri_low = { };
@@ -165,11 +232,26 @@  xfs_dax_notify_ddev_failure(
 	}
 
 	xfs_trans_cancel(tp);
-	if (error || notify.want_shutdown) {
+
+	/*
+	 * Shutdown fs from a force umount in pre-remove case which won't fail,
+	 * so errors can be ignored.  Otherwise, shutdown the filesystem with
+	 * CORRUPT flag if error occured or notify.want_shutdown was set during
+	 * RMAP querying.
+	 */
+	if (mf_flags & MF_MEM_PRE_REMOVE)
+		xfs_force_shutdown(mp, SHUTDOWN_FORCE_UMOUNT);
+	else if (error || notify.want_shutdown) {
 		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_ONDISK);
 		if (!error)
 			error = -EFSCORRUPTED;
 	}
+
+out:
+	/* Thaw the fs if it has been frozen before. */
+	if (mf_flags & MF_MEM_PRE_REMOVE)
+		xfs_dax_notify_failure_thaw(mp, kernel_frozen);
+
 	return error;
 }
 
@@ -197,6 +279,14 @@  xfs_dax_notify_failure(
 
 	if (mp->m_logdev_targp && mp->m_logdev_targp->bt_daxdev == dax_dev &&
 	    mp->m_logdev_targp != mp->m_ddev_targp) {
+		/*
+		 * In the pre-remove case the failure notification is attempting
+		 * to trigger a force unmount.  The expectation is that the
+		 * device is still present, but its removal is in progress and
+		 * can not be cancelled, proceed with accessing the log device.
+		 */
+		if (mf_flags & MF_MEM_PRE_REMOVE)
+			return 0;
 		xfs_err(mp, "ondisk log corrupt, shutting down fs!");
 		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_ONDISK);
 		return -EFSCORRUPTED;
@@ -210,6 +300,12 @@  xfs_dax_notify_failure(
 	ddev_start = mp->m_ddev_targp->bt_dax_part_off;
 	ddev_end = ddev_start + bdev_nr_bytes(mp->m_ddev_targp->bt_bdev) - 1;
 
+	/* Notify failure on the whole device. */
+	if (offset == 0 && len == U64_MAX) {
+		offset = ddev_start;
+		len = bdev_nr_bytes(mp->m_ddev_targp->bt_bdev);
+	}
+
 	/* Ignore the range out of filesystem area */
 	if (offset + len - 1 < ddev_start)
 		return -ENXIO;
diff --git a/include/linux/mm.h b/include/linux/mm.h
index bf5d0b1b16f4..385eee0d05a2 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -3831,6 +3831,7 @@  enum mf_flags {
 	MF_UNPOISON = 1 << 4,
 	MF_SW_SIMULATED = 1 << 5,
 	MF_NO_RETRY = 1 << 6,
+	MF_MEM_PRE_REMOVE = 1 << 7,
 };
 int mf_dax_kill_procs(struct address_space *mapping, pgoff_t index,
 		      unsigned long count, int mf_flags);
diff --git a/mm/memory-failure.c b/mm/memory-failure.c
index 4d6e43c88489..6e43ae369fef 100644
--- a/mm/memory-failure.c
+++ b/mm/memory-failure.c
@@ -679,7 +679,7 @@  static void add_to_kill_fsdax(struct task_struct *tsk, struct page *p,
  */
 static void collect_procs_fsdax(struct page *page,
 		struct address_space *mapping, pgoff_t pgoff,
-		struct list_head *to_kill)
+		struct list_head *to_kill, bool pre_remove)
 {
 	struct vm_area_struct *vma;
 	struct task_struct *tsk;
@@ -687,8 +687,15 @@  static void collect_procs_fsdax(struct page *page,
 	i_mmap_lock_read(mapping);
 	rcu_read_lock();
 	for_each_process(tsk) {
-		struct task_struct *t = task_early_kill(tsk, true);
+		struct task_struct *t = tsk;
 
+		/*
+		 * Search for all tasks while MF_MEM_PRE_REMOVE is set, because
+		 * the current may not be the one accessing the fsdax page.
+		 * Otherwise, search for the current task.
+		 */
+		if (!pre_remove)
+			t = task_early_kill(tsk, true);
 		if (!t)
 			continue;
 		vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
@@ -1792,6 +1799,7 @@  int mf_dax_kill_procs(struct address_space *mapping, pgoff_t index,
 	dax_entry_t cookie;
 	struct page *page;
 	size_t end = index + count;
+	bool pre_remove = mf_flags & MF_MEM_PRE_REMOVE;
 
 	mf_flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
 
@@ -1803,9 +1811,14 @@  int mf_dax_kill_procs(struct address_space *mapping, pgoff_t index,
 		if (!page)
 			goto unlock;
 
-		SetPageHWPoison(page);
+		if (!pre_remove)
+			SetPageHWPoison(page);
 
-		collect_procs_fsdax(page, mapping, index, &to_kill);
+		/*
+		 * The pre_remove case is revoking access, the memory is still
+		 * good and could theoretically be put back into service.
+		 */
+		collect_procs_fsdax(page, mapping, index, &to_kill, pre_remove);
 		unmap_and_kill(&to_kill, page_to_pfn(page), mapping,
 				index, mf_flags);
 unlock: