We had some systems crash with this stack: [] ia64_leave_kernel+0x0/0x280 [] xfs_bmbt_get_startoff+0x0/0x20 [xfs] [] xfs_bmap_last_offset+0x210/0x280 [xfs] [] xfs_file_last_byte+0x70/0x1a0 [xfs] [] xfs_itruncate_start+0xc0/0x1a0 [xfs] [] xfs_inactive_free_eofblocks+0x290/0x460 [xfs] [] xfs_release+0x1b0/0x240 [xfs] [] xfs_file_release+0x70/0xa0 [xfs] [] __fput+0x1a0/0x420 [] fput+0x40/0x60 I suspect what has happened here is that xfs_bmap_last_offset() has accessed the extent list without acquiring the ilock and has raced with another thread that has changed the extent list. The other thread has reduced the extent count so we've tried to access an extent beyond the end of the list. I added some ASSERTs in xfs_iext_get_ext() and xfs_iext_idx_to_irec() to catch this and sure enough they triggered very quickly. Looking closely at the algorithm in xfs_iext_idx_to_irec(), if we try to lookup an extent beyond the end of the list the search terminates with 'low > high' and page_idx never gets updated to be an index into a specific extent buffer. Back in xfs_iext_get_ext(), if page_idx is > 255 it is off the end of the extent buffer and potentially in unmapped memory which is how the system panicked. Some debugging in xfs_iext_get_ext() on a heavily fragmented file revealed that page_idx was 167532 while the actual count of extents in the buffer was only 254 so we accessed way outside the bounds of the extent buffer. I put together a test program (attached) and ran it as: # ./extent -t -l 163840000 and it found many places where we try to index an extent beyond the end of the extent list. There's code in xfs_bmapi() and xfs_bunmapi() that uses the value of if_lastex in the data fork to index an extent so it's important to make sure that field holds a sane value. The primary culprit there is xfs_bmap_del_extent() when deleting an entire extent - it leaves if_lastex as the index of the extent that was removed. If there is another extent beyond that one then if_lastex now points to that. If the removed extent was the last extent then if_lastex is now beyond the end of the list. Index: xfs-fix/fs/xfs/xfs_bmap.c =================================================================== --- xfs-fix.orig/fs/xfs/xfs_bmap.c +++ xfs-fix/fs/xfs/xfs_bmap.c @@ -1887,7 +1887,7 @@ xfs_bmap_add_extent_hole_delay( #define SWITCH_STATE (state & MASK2(LEFT_CONTIG, RIGHT_CONTIG)) ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK); - ep = xfs_iext_get_ext(ifp, idx); + ep = NULL; state = 0; ASSERT(isnullstartblock(new->br_startblock)); /* @@ -1904,6 +1904,7 @@ xfs_bmap_add_extent_hole_delay( if (STATE_SET_TEST(RIGHT_VALID, idx < ip->i_df.if_bytes / (uint)sizeof(xfs_bmbt_rec_t))) { + ep = xfs_iext_get_ext(ifp, idx); xfs_bmbt_get_all(ep, &right); STATE_SET(RIGHT_DELAY, isnullstartblock(right.br_startblock)); } @@ -2078,7 +2079,7 @@ xfs_bmap_add_extent_hole_real( ifp = XFS_IFORK_PTR(ip, whichfork); ASSERT(idx <= ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t)); - ep = xfs_iext_get_ext(ifp, idx); + ep = NULL; state = 0; /* * Check and set flags if this segment has a left neighbor. @@ -2094,6 +2095,7 @@ xfs_bmap_add_extent_hole_real( if (STATE_SET_TEST(RIGHT_VALID, idx < ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t))) { + ep = xfs_iext_get_ext(ifp, idx); xfs_bmbt_get_all(ep, &right); STATE_SET(RIGHT_DELAY, isnullstartblock(right.br_startblock)); } @@ -3206,6 +3208,8 @@ xfs_bmap_del_extent( */ XFS_BMAP_TRACE_DELETE("3", ip, idx, 1, whichfork); xfs_iext_remove(ifp, idx, 1); + if (idx >= (ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t))) + idx--; ifp->if_lastex = idx; if (delay) break; @@ -5290,12 +5294,13 @@ xfs_bmapi( /* * Else go on to the next record. */ - ep = xfs_iext_get_ext(ifp, ++lastx); prev = got; - if (lastx >= nextents) + if (++lastx >= nextents) eof = 1; - else + else { + ep = xfs_iext_get_ext(ifp, lastx); xfs_bmbt_get_all(ep, &got); + } } ifp->if_lastex = lastx; *nmap = n; @@ -5723,15 +5728,15 @@ nodelete: * If not done go on to the next (previous) record. * Reset ep in case the extents array was re-alloced. */ - ep = xfs_iext_get_ext(ifp, lastx); if (bno != (xfs_fileoff_t)-1 && bno >= start) { - if (lastx >= XFS_IFORK_NEXTENTS(ip, whichfork) || - xfs_bmbt_get_startoff(ep) > bno) { - if (--lastx >= 0) - ep = xfs_iext_get_ext(ifp, lastx); - } - if (lastx >= 0) + if (lastx >= 0) { + ep = xfs_iext_get_ext(ifp, lastx); + if (xfs_bmbt_get_startoff(ep) > bno) { + if (--lastx >= 0) + ep = xfs_iext_get_ext(ifp, lastx); + } xfs_bmbt_get_all(ep, &got); + } extno++; } } Index: xfs-fix/fs/xfs/xfs_inode.c =================================================================== --- xfs-fix.orig/fs/xfs/xfs_inode.c +++ xfs-fix/fs/xfs/xfs_inode.c @@ -1258,8 +1258,10 @@ xfs_file_last_byte( * necessary. */ if (ip->i_df.if_flags & XFS_IFEXTENTS) { + xfs_ilock(ip, XFS_ILOCK_SHARED); error = xfs_bmap_last_offset(NULL, ip, &last_block, XFS_DATA_FORK); + xfs_iunlock(ip, XFS_ILOCK_SHARED); if (error) { last_block = 0; } @@ -2645,10 +2647,8 @@ xfs_iflush_fork( case XFS_DINODE_FMT_EXTENTS: ASSERT((ifp->if_flags & XFS_IFEXTENTS) || !(iip->ili_format.ilf_fields & extflag[whichfork])); - ASSERT((xfs_iext_get_ext(ifp, 0) != NULL) || - (ifp->if_bytes == 0)); - ASSERT((xfs_iext_get_ext(ifp, 0) == NULL) || - (ifp->if_bytes > 0)); + ASSERT((ifp->if_bytes == 0) || + (xfs_iext_get_ext(ifp, 0) != NULL)); if ((iip->ili_format.ilf_fields & extflag[whichfork]) && (ifp->if_bytes > 0)) { ASSERT(XFS_IFORK_NEXTENTS(ip, whichfork) > 0); @@ -3265,6 +3265,7 @@ xfs_iext_get_ext( xfs_extnum_t idx) /* index of target extent */ { ASSERT(idx >= 0); + ASSERT(idx < (ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t))); if ((ifp->if_flags & XFS_IFEXTIREC) && (idx == 0)) { return ifp->if_u1.if_ext_irec->er_extbuf; } else if (ifp->if_flags & XFS_IFEXTIREC) { @@ -4035,6 +4036,8 @@ xfs_iext_idx_to_irec( ASSERT(ifp->if_flags & XFS_IFEXTIREC); ASSERT(page_idx >= 0 && page_idx <= ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t)); + ASSERT(page_idx < ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t) + || realloc); nlists = ifp->if_real_bytes / XFS_IEXT_BUFSZ; erp_idx = 0; low = 0;