oops while doing block IO
With 2.16.17 I get infrequent but very consistent null pointer dereferences especially but not only when lots of block IO is going on (I backup with disk-to-disk copies).
Some more context: I use 'loop-AES' (may or may not be relevant), and the error happens most often when lots of _concurrent_ block IO takes place; usually mere single threaded copying from one disk to another does not trigger the error.
The relevant details are:
------------------------------------------------------------------------
base kernel: BUG: unable to handle kernel NULL pointer dereference at virtual address 00000040
------------------------------------------------------------------------
EIP is at generic_make_request+0x31/0x330
eax: 00000000 ebx: 00000200 ecx: a1a23480 edx: 95aa34c0
esi: 95aa34c0 edi: 00000000 ebp: b7f14cec esp: b7f14c64
ds: 007b es: 007b ss: 0068
Process pdflush (pid: 166, threadinfo=b7f14000 task=b7f15ab0)
------------------------------------------------------------------------
Call Trace: <7810402d> show_stack_log_lvl+0x9d/0xd0 <78104267> show_registers+0x1b7/0x240 <78104422> die+0x132/0x330 <78116086> do_page_fault+0x296/0x6bc <78103a9f> error_code+0x4f/0x54 <78235a18> submit_bio+0x58/0x100 <78169be3> submit_bh+0xd3/0x130 <7816b744> __block_write_full_page+0x1c4/0x3b0 <7816bd92> block_write_full_page+0x102/0x110 <78170560> blkdev_writepage+0x20/0x30 <7818f688> mpage_writepages+0x1b8/0x3d0 <78170510> generic_writepages+0x20/0x30 <7814e13d> do_writepages+0x2d/0x50 <7818da54> __writeback_single_inode+0x94/0x3c0 <7818e04b> sync_sb_inodes+0x1bb/0x2a0 <7818e6b3> writeback_inodes+0xc3/0xf9 <7814e4b0> background_writeout+0x80/0xa0 <7814ecde> pdflush+0xee/0x1b0 <78133385> kthread+0xc5/0xf0 <78100dd5> kernel_thread_helper+0x5/0x10 ------------------------------------------------------------------------
The relevant bit of code is:
------------------------------------------------------------------------
0x7823338e <generic_make_request+30>: call 0x78118410 <__might_sleep> 0x78233393 <generic_make_request+35>: call 0x783ea290 <cond_resched> 0x78233398 <generic_make_request+40>: mov 0x8(%ebp),%edx 0x7823339b <generic_make_request+43>: mov 0x8(%edx),%ecx 0x7823339e <generic_make_request+46>: mov 0x4(%ecx),%eax 0x782333a1 <generic_make_request+49>: mov 0x40(%eax),%edx 0x782333a4 <generic_make_request+52>: mov 0x3c(%eax),%eax 0x782333a7 <generic_make_request+55>: shrd $0x9,%edx,%eax
------------------------------------------------------------------------
and the null pointer is clearly in '%eax', which corresponds to '->bd_inode' at the beginning of 'generic_make_request' in 'block/ll_rw_blk.c':
------------------------------------------------------------------------
might_sleep();
/* Test device or partition size, when known. */
maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
if (maxsector) {------------------------------------------------------------------------
As a crude fix I have done this patch:
------------------------------------------------------------------------
ll_rw_blk.c renamed from ll_rw_blk.c-dist +12 -2
@@ -80,6 +80,16 @@#define BLK_BATCH_REQ 32 + +/* +* Return the maximum number of sectors for the block device, +* or 0 if unknown. +*/ +static inline sector_t bio_max_sector (const struct bio *const bio) +{ +return (bio->bi_bdev->bd_inode ==0 ) ?0 + : bio->bi_bdev->bd_inode->i_size >>9 ; +}/* * Return the threshold (number of used requests) at which the queue is * considered to be congested. It include a little hysteresis to keep the * context switch rate down.
@@ -2983,7 +2993,7 @@bdevname (bio->bi_bdev, b), bio->bi_rw, (unsigned long long )bio->bi_sector +bio_sectors (bio), - (long long)(bio->bi_bdev->bd_inode->i_size >> 9)); + (long long) bio_max_sector(bio));set_bit (BIO_EOF , &bio->bi_flags); }
@@ -3021,7 +3031,7 @@might_sleep ();/* Test device or partition size, when known. */ - maxsector = bio->bi_bdev->bd_inode->i_size >>9 ; + maxsector =bio_max_sector (bio);if (maxsector) {sector_t sector = bio->bi_sector;
------------------------------------------------------------------------
Which just prevents the null pointer dereference. The wider question is why 'bd_inode' is null, but only in some cases...