On 03/17/2013 11:01 AM, Jeff Liu wrote:
On 32-bit system, if the request pos is 64-bit and evaluate block_offset with (pos & PAGE_MASK) will result in overflows, therefore the assertion will failed. We have to check the write offset against (pos & ~0UL) to avoid this issue as it can evaluate the highest 20 bits on 32-bit correctly if the pos request is 64-bit and keep the expected result of 64-bit pos request on 64-bit system unchanged. Signed-off-by: Jie Liu <jeff.liu@xxxxxxxxxx> Reported-by: Michael L. Semon <mlsemon35@xxxxxxxxx> Cc: Dave Chinner <david@xxxxxxxxxxxxx> --- fs/xfs/xfs_aops.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c index 5f707e5..2fc7367 100644 --- a/fs/xfs/xfs_aops.c +++ b/fs/xfs/xfs_aops.c @@ -1501,7 +1501,12 @@ xfs_vm_write_failed( loff_t to = from + len; struct buffer_head *bh, *head; - ASSERT(block_offset + from == pos); + /* + * Evaluate block_offset via (pos & PAGE_MASK) on 32-bit system + * can cause overflow if the request pos is 64-bit. Hence we + * have to verify the write offset with (pos & ~0UL) to avoid it. + */ + ASSERT(block_offset + from == (pos & ~0UL)); head = page_buffers(page); block_start = 0;
Thanks! I can't help but admire the effort. That stated, I did read Dave's review and now understand the "..." that he left as comments to the original bug report...
My original reason for writing was to refine the test case a little bit. On this 32-bit Pentium III PC, xfstests #078 succeeds on a 560MB device-mapper linear target (1146880 sectors), but it fails with an oops on a 544MB dm-linear target (1114112 sectors). Looking at the output of the `df` command over and over during the test, the data does stop growing at a point between those two numbers, proving Dave's initial observation correct.
Michael _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
