Re: [PATCH] mm: fix page_mkclean_one (was: 2.6.19 file content corruption on ext3)

[Jari Sundell]

On 12/27/06, Linus Torvalds <torvalds@xxxxxxxx> wrote:
<snip>
>  - It never uses mprotect on the shared mappings, but it _does_ do:
>         "mincore()" - but the return values don't much matter (it's used
>                       as a heuristic on which parts to hash, apparently)
>
>                       I double- and triple-checked this one, because I
>                       did make changes to "mincore()", but those didn't go
>                       into the affected kernels anyway (ie they are not in
>                       plain 2.6.19, nor in 2.6.18.3 either)

Correct, mincore is only used to check if it should delay the hash checking.

>         "madvise(MADV_WILLNEED)"
>         "msync(MS_ASYNC)" (or MS_SYNC if you use a command line flag)
>         "munmap()" of course
>
>  - it never seems to mix mmap() and write() - it does _only_ mmap.
>
>  - it seems to mmap/munmap the shared files in nice 64-page chunks, all
>    64-page aligned in the file (ie it does NOT create one big mapping, it
>    has some kind of LRU of thse 64-page chunks). The only exception being
>    the last chunk, which it maps byte-accurate to the size.

The length of the chunks is only page aligned on single file torrents,
not so on multi-file torrents. I've attached a patch for rtorrent that
will extend the length to the page boundary.

>  - I haven't checked whether it only ever has the same chunk mapped once
>    at a time.

This should be the case, but two mapped chunks may share a page,
sometimes with different r/w permissions.

Jari Sundell

Attachment: extend_mapping.diff
Description: Binary data