Re: sendfile data corruption

Pawel Worach wrote:
> Andre Oppermann wrote:
>> andre       2006-11-02 16:53:26 UTC
>>
>>   FreeBSD src repository
>>
> ...
>>   Rewrite kern_sendfile() to work in two loops, the inner which turns 
>> as many
> ...
>>     Revision  Changes    Path
>>   1.240     +280 -242  src/sys/kern/uipc_syscalls.c
>>   1.55      +2 -0      src/sys/sys/libkern.h
>>   1.91      +2 -1      src/sys/sys/socket.h
> 
> Hi Andre,
> 
> I'm seeing some strange data corruption with this change.
> Using apache 2.0.59 from ports, hardware is SMP i386.
> 
> 0>root_at_cookie /usr/local/www/data# md5 sh
> MD5 (sh) = e090ae9fc697b6ec84165af920034dc4
> 0>root_at_cookie /usr/local/www/data# unsetenv http_proxy
> 0>root_at_cookie /usr/local/www/data# fetch -o /tmp/sh http://127.0.0.1/sh
> /tmp/sh                                       100% of  109 kB 6516 kBps
> 0>root_at_cookie /usr/local/www/data# md5 /tmp/sh
> MD5 (/tmp/sh) = 1b6b9786ce7aa74b7ecbc7ee82c230dd
> 
> It seems to be consistent...
> 0>root_at_zero /usr/local/www/data# fetch -o /tmp/sh2 http://127.0.0.1/sh
> /tmp/sh2                                      100% of  109 kB   41 MBps
> 0>root_at_zero /usr/local/www/data# md5 /tmp/sh2
> MD5 (/tmp/sh2) = 1b6b9786ce7aa74b7ecbc7ee82c230dd
> 
> Checking with hd(1) the changed data always seems to start at offset 
> 0x0000e000.
> 
> cvs up -D '2006/11/02 17:00:00' OK
>    Repository revision: 1.239 
> /export/ctm/cvs/src/sys/kern/uipc_syscalls.c,v
> 
> cvs up -D '2006/11/02 18:00:00' BROKEN
>    Repository revision: 1.240 
> /export/ctm/cvs/src/sys/kern/uipc_syscalls.c,v
> 
> Files changed with update:
> P geom/journal/g_journal.c
> P kern/uipc_syscalls.c
> P sys/libkern.h
> P sys/socket.h
> 
> Any other information I can provide?

I'm looking into the problem.  Please try a binary FTP transfer as well
and check if the checksums match. ftpd uses sendfile(2) as well but w/o
headers or trailers and does the send in one swoop.

-- 
Andre