Re: [XFS] Kernel (2.6.11) deadlock (kernel hang) in user mode whenwriting data through mmap on large files (64-bit systems)

[Xavier Roche]

Hi,

Christoph Hellwig wrote:
> This is a known problem, or rather two of them:
>  1) when dirtying lots of memory via mmap writeout happens pretty randmomly
>     inside the file and the filesystem has problems creating nice clusters.
>     In the case of extent based filesystems that's really bad as the extent
>     map is fragmented now
>  2) XFS keeps each inode's extent map in one single memory block.

It seems that the file was really *badly* fragmented. The reason, as far 
as we understand the problem, was:

- a file "truncated" to _expand_ its size (using ftruncate() with a size 
MUCH larger that the current size, which is == 0), leading to create a 
"big sparse file" area
- sequential write in this file (_NOT_ random) using the corresponding 
mmapp'ed data segment
- random (!) flush from kswapd leading to allocate fragmented pages 
(sparse file)

The file appears to have over one million fragments:

$ xfs_bmap ./data | wc -l
1051397

$ xfs_bmap ./data | head -n 20
./data:
        0: [0..7]: 15103824..15103831
        1: [8..15]: 15103840..15103847
        2: [16..23]: 15103856..15103863
        3: [24..31]: 15103872..15103879
        4: [32..39]: 15103888..15103895
        5: [40..47]: 15103904..15103911
        6: [48..55]: 15103920..15103927
        7: [56..63]: 15103936..15103943
        8: [64..71]: 15103952..15103959
        9: [72..79]: 15103968..15103975
        10: [80..87]: 15104104..15104111
        11: [88..95]: 15104120..15104127
        12: [96..103]: 15104136..15104143
        13: [104..111]: 15104152..15104159
        14: [112..119]: 15104168..15104175
        15: [120..127]: 15104184..15104191
        16: [128..135]: 15104200..15104207
        17: [136..143]: 15104216..15104223
        18: [144..151]: 15104232..15104239
..

> You're seeing allocation errors where we are trying to realloc that memory
> block.
> Could you try the patches that Nikita posted to -mm that should improve
> this behaviour?

Well, the reasons seems to clearly be this anormal number fo fragments - 
is there any potential solution (in the kernel/mm), or the olny solution 
is a patch to ensure that ftruncate() is replaced by regulars 
fwrite()-zero calls ?



-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/