Andrew,
Somewhere in between 2.6.18-mm3 and 2.6.19-rc4-mm1 writev() got screwed.
It does not accept zero-length segments anymore.
Bad thing that it is extremely easy to trigger (even w/o explicit writev calls).
For example the following innocent program will fail with 2.6.19-rc4-mm1:
======================
#include <string.h>
#include <fstream>
int main()
{
char buf[1024];
memset(buf, 'A', sizeof(buf));
std::ofstream ofs("test");
//ofs << 1 << '\n';
ofs.write(buf, sizeof(buf));
return 0;
}
======================
Here is the corresponding part if strace:
======================
open("test", O_WRONLY|O_CREAT|O_TRUNC|O_LARGEFILE, 0666) = 3
writev(3, [{NULL, 0}, {"AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA"..., 1024}], 2) = -1 EFAULT (Bad address)
close(3) = 0
======================
With 2.6.18-mm3 it works
======================
open("test", O_WRONLY|O_CREAT|O_TRUNC|O_LARGEFILE, 0666) = 3
writev(3, [{NULL, 0}, {"AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA"..., 1024}], 2) = 1024
close(3) = 0
======================
It works with 2.6.19-rc4-mm1 _if_ zero-length segments are eliminated
(by uncommenting ofs << 1 << '\n'):
======================
open("test", O_WRONLY|O_CREAT|O_TRUNC|O_LARGEFILE, 0666) = 3
writev(3, [{"1\n", 2}, {"AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA"..., 1024}], 2) = 1026
close(3) = 0
======================
Given that _all_ applications using C++ streams are potentially affected
I think it's better to preserve the previous behavior even if it is
something from "undefined behavior world" (or a plain bug).
The bug is quite dangerous (I was really close to wipe out my mp3 collection).
Please let me know if you want me to try any patches.
Thank you,
Nick Orlov.
--
With best wishes,
Nick Orlov.
On Sun, 12 Nov 2006 17:30:24 -0500
Nick Orlov <[email protected]> wrote:
> Andrew,
>
> Somewhere in between 2.6.18-mm3 and 2.6.19-rc4-mm1 writev() got screwed.
> It does not accept zero-length segments anymore.
>
> Bad thing that it is extremely easy to trigger (even w/o explicit writev calls).
> For example the following innocent program will fail with 2.6.19-rc4-mm1:
>
> ======================
> #include <string.h>
> #include <fstream>
>
> int main()
> {
> char buf[1024];
> memset(buf, 'A', sizeof(buf));
> std::ofstream ofs("test");
> //ofs << 1 << '\n';
> ofs.write(buf, sizeof(buf));
> return 0;
> }
> ======================
>
>
> Here is the corresponding part if strace:
>
> ======================
> open("test", O_WRONLY|O_CREAT|O_TRUNC|O_LARGEFILE, 0666) = 3
> writev(3, [{NULL, 0}, {"AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA"..., 1024}], 2) = -1 EFAULT (Bad address)
> close(3) = 0
> ======================
>
>
> With 2.6.18-mm3 it works
>
> ======================
> open("test", O_WRONLY|O_CREAT|O_TRUNC|O_LARGEFILE, 0666) = 3
> writev(3, [{NULL, 0}, {"AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA"..., 1024}], 2) = 1024
> close(3) = 0
> ======================
>
>
> It works with 2.6.19-rc4-mm1 _if_ zero-length segments are eliminated
> (by uncommenting ofs << 1 << '\n'):
>
> ======================
> open("test", O_WRONLY|O_CREAT|O_TRUNC|O_LARGEFILE, 0666) = 3
> writev(3, [{"1\n", 2}, {"AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA"..., 1024}], 2) = 1026
> close(3) = 0
> ======================
>
>
> Given that _all_ applications using C++ streams are potentially affected
> I think it's better to preserve the previous behavior even if it is
> something from "undefined behavior world" (or a plain bug).
>
> The bug is quite dangerous (I was really close to wipe out my mp3 collection).
>
OK, thanks. Those patches do need more work. I'll shelve them for a while.
Andrew Morton wrote:
> On Sun, 12 Nov 2006 17:30:24 -0500
> Nick Orlov <[email protected]> wrote:
>
>
>>Andrew,
>>
>>Somewhere in between 2.6.18-mm3 and 2.6.19-rc4-mm1 writev() got screwed.
>>It does not accept zero-length segments anymore.
>>
>>Bad thing that it is extremely easy to trigger (even w/o explicit writev calls).
>>For example the following innocent program will fail with 2.6.19-rc4-mm1:
>>
>>======================
>>#include <string.h>
>>#include <fstream>
>>
>>int main()
>>{
>> char buf[1024];
>> memset(buf, 'A', sizeof(buf));
>> std::ofstream ofs("test");
>> //ofs << 1 << '\n';
>> ofs.write(buf, sizeof(buf));
>> return 0;
>>}
>>======================
>>
>>
>>Here is the corresponding part if strace:
>>
>>======================
>>open("test", O_WRONLY|O_CREAT|O_TRUNC|O_LARGEFILE, 0666) = 3
>>writev(3, [{NULL, 0}, {"AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA"..., 1024}], 2) = -1 EFAULT (Bad address)
>>close(3) = 0
>>======================
>>
>>
>>With 2.6.18-mm3 it works
>>
>>======================
>>open("test", O_WRONLY|O_CREAT|O_TRUNC|O_LARGEFILE, 0666) = 3
>>writev(3, [{NULL, 0}, {"AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA"..., 1024}], 2) = 1024
>>close(3) = 0
>>======================
>>
>>
>>It works with 2.6.19-rc4-mm1 _if_ zero-length segments are eliminated
>>(by uncommenting ofs << 1 << '\n'):
>>
>>======================
>>open("test", O_WRONLY|O_CREAT|O_TRUNC|O_LARGEFILE, 0666) = 3
>>writev(3, [{"1\n", 2}, {"AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA"..., 1024}], 2) = 1026
>>close(3) = 0
>>======================
>>
>>
>>Given that _all_ applications using C++ streams are potentially affected
>>I think it's better to preserve the previous behavior even if it is
>>something from "undefined behavior world" (or a plain bug).
>>
>>The bug is quite dangerous (I was really close to wipe out my mp3 collection).
>>
>
>
> OK, thanks. Those patches do need more work. I'll shelve them for a while.
Yeah, this just needs some kind of 0 size check in the access check
(arguably fault_in_pages_readable could check for zero size, but it could
go into the caller just as easily).
Thanks for reporting.
--
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com