https://github.com/torvalds/linux
Revision 9230a0b65b47fe6856c4468ec0175c4987e5bede authored by Dave Chinner on 20 November 2018, 06:50:08 UTC, committed by Darrick J. Wong on 21 November 2018, 18:10:53 UTC
Long saga. There have been days spent following this through dead end
after dead end in multi-GB event traces. This morning, after writing
a trace-cmd wrapper that enabled me to be more selective about XFS
trace points, I discovered that I could get just enough essential
tracepoints enabled that there was a 50:50 chance the fsx config
would fail at ~115k ops. If it didn't fail at op 115547, I stopped
fsx at op 115548 anyway.

That gave me two traces - one where the problem manifested, and one
where it didn't. After refining the traces to have the necessary
information, I found that in the failing case there was a real
extent in the COW fork compared to an unwritten extent in the
working case.

Walking back through the two traces to the point where the CWO fork
extents actually diverged, I found that the bad case had an extra
unwritten extent in it. This is likely because the bug it led me to
had triggered multiple times in those 115k ops, leaving stray
COW extents around. What I saw was a COW delalloc conversion to an
unwritten extent (as they should always be through
xfs_iomap_write_allocate()) resulted in a /written extent/:

xfs_writepage:        dev 259:0 ino 0x83 pgoff 0x17000 size 0x79a00 offset 0 length 0
xfs_iext_remove:      dev 259:0 ino 0x83 state RC|LF|RF|COW cur 0xffff888247b899c0/2 offset 32 block 152 count 20 flag 1 caller xfs_bmap_add_extent_delay_real
xfs_bmap_pre_update:  dev 259:0 ino 0x83 state RC|LF|RF|COW cur 0xffff888247b899c0/1 offset 1 block 4503599627239429 count 31 flag 0 caller xfs_bmap_add_extent_delay_real
xfs_bmap_post_update: dev 259:0 ino 0x83 state RC|LF|RF|COW cur 0xffff888247b899c0/1 offset 1 block 121 count 51 flag 0 caller xfs_bmap_add_ex

Basically, Cow fork before:

	0 1            32          52
	+H+DDDDDDDDDDDD+UUUUUUUUUUU+
	   PREV		RIGHT

COW delalloc conversion allocates:

	  1	       32
	  +uuuuuuuuuuuu+
	  NEW

And the result according to the xfs_bmap_post_update trace was:

	0 1            32          52
	+H+wwwwwwwwwwwwwwwwwwwwwwww+
	   PREV

Which is clearly wrong - it should be a merged unwritten extent,
not an unwritten extent.

That lead me to look at the LEFT_FILLING|RIGHT_FILLING|RIGHT_CONTIG
case in xfs_bmap_add_extent_delay_real(), and sure enough, there's
the bug.

It takes the old delalloc extent (PREV) and adds the length of the
RIGHT extent to it, takes the start block from NEW, removes the
RIGHT extent and then updates PREV with the new extent.

What it fails to do is update PREV.br_state. For delalloc, this is
always XFS_EXT_NORM, while in this case we are converting the
delayed allocation to unwritten, so it needs to be updated to
XFS_EXT_UNWRITTEN. This LF|RF|RC case does not do this, and so
the resultant extent is always written.

And that's the bug I've been chasing for a week - a bmap btree bug,
not a reflink/dedupe/copy_file_range bug, but a BMBT bug introduced
with the recent in core extent tree scalability enhancements.

Signed-off-by: Dave Chinner <dchinner@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>
1 parent 2c30717
Raw File
Tip revision: 9230a0b65b47fe6856c4468ec0175c4987e5bede authored by Dave Chinner on 20 November 2018, 06:50:08 UTC
xfs: delalloc -> unwritten COW fork allocation can go wrong
Tip revision: 9230a0b
list_sort.c
// SPDX-License-Identifier: GPL-2.0
#include <linux/kernel.h>
#include <linux/bug.h>
#include <linux/compiler.h>
#include <linux/export.h>
#include <linux/string.h>
#include <linux/list_sort.h>
#include <linux/list.h>

#define MAX_LIST_LENGTH_BITS 20

/*
 * Returns a list organized in an intermediate format suited
 * to chaining of merge() calls: null-terminated, no reserved or
 * sentinel head node, "prev" links not maintained.
 */
static struct list_head *merge(void *priv,
				int (*cmp)(void *priv, struct list_head *a,
					struct list_head *b),
				struct list_head *a, struct list_head *b)
{
	struct list_head head, *tail = &head;

	while (a && b) {
		/* if equal, take 'a' -- important for sort stability */
		if ((*cmp)(priv, a, b) <= 0) {
			tail->next = a;
			a = a->next;
		} else {
			tail->next = b;
			b = b->next;
		}
		tail = tail->next;
	}
	tail->next = a?:b;
	return head.next;
}

/*
 * Combine final list merge with restoration of standard doubly-linked
 * list structure.  This approach duplicates code from merge(), but
 * runs faster than the tidier alternatives of either a separate final
 * prev-link restoration pass, or maintaining the prev links
 * throughout.
 */
static void merge_and_restore_back_links(void *priv,
				int (*cmp)(void *priv, struct list_head *a,
					struct list_head *b),
				struct list_head *head,
				struct list_head *a, struct list_head *b)
{
	struct list_head *tail = head;
	u8 count = 0;

	while (a && b) {
		/* if equal, take 'a' -- important for sort stability */
		if ((*cmp)(priv, a, b) <= 0) {
			tail->next = a;
			a->prev = tail;
			a = a->next;
		} else {
			tail->next = b;
			b->prev = tail;
			b = b->next;
		}
		tail = tail->next;
	}
	tail->next = a ? : b;

	do {
		/*
		 * In worst cases this loop may run many iterations.
		 * Continue callbacks to the client even though no
		 * element comparison is needed, so the client's cmp()
		 * routine can invoke cond_resched() periodically.
		 */
		if (unlikely(!(++count)))
			(*cmp)(priv, tail->next, tail->next);

		tail->next->prev = tail;
		tail = tail->next;
	} while (tail->next);

	tail->next = head;
	head->prev = tail;
}

/**
 * list_sort - sort a list
 * @priv: private data, opaque to list_sort(), passed to @cmp
 * @head: the list to sort
 * @cmp: the elements comparison function
 *
 * This function implements "merge sort", which has O(nlog(n))
 * complexity.
 *
 * The comparison function @cmp must return a negative value if @a
 * should sort before @b, and a positive value if @a should sort after
 * @b. If @a and @b are equivalent, and their original relative
 * ordering is to be preserved, @cmp must return 0.
 */
void list_sort(void *priv, struct list_head *head,
		int (*cmp)(void *priv, struct list_head *a,
			struct list_head *b))
{
	struct list_head *part[MAX_LIST_LENGTH_BITS+1]; /* sorted partial lists
						-- last slot is a sentinel */
	int lev;  /* index into part[] */
	int max_lev = 0;
	struct list_head *list;

	if (list_empty(head))
		return;

	memset(part, 0, sizeof(part));

	head->prev->next = NULL;
	list = head->next;

	while (list) {
		struct list_head *cur = list;
		list = list->next;
		cur->next = NULL;

		for (lev = 0; part[lev]; lev++) {
			cur = merge(priv, cmp, part[lev], cur);
			part[lev] = NULL;
		}
		if (lev > max_lev) {
			if (unlikely(lev >= ARRAY_SIZE(part)-1)) {
				printk_once(KERN_DEBUG "list too long for efficiency\n");
				lev--;
			}
			max_lev = lev;
		}
		part[lev] = cur;
	}

	for (lev = 0; lev < max_lev; lev++)
		if (part[lev])
			list = merge(priv, cmp, part[lev], list);

	merge_and_restore_back_links(priv, cmp, head, part[max_lev], list);
}
EXPORT_SYMBOL(list_sort);
back to top