Re: [PATCH 1/2] pipe: fix race with fcntl

From: Dmitry Monakhov
Date: Sun Apr 12 2015 - 13:54:52 EST

Al Viro <viro@xxxxxxxxxxxxxxxxxx> writes:

> On Sun, Apr 12, 2015 at 09:08:21PM +0400, Dmitry Monakhov wrote:
>> Fix other long standing issues caused by fcntl(,F_SETFL,):
>> - User can disable O_DIRECT for pipe[1] (paketized IO), but can not enable it again.
>> - Currently we do not set O_APPEND on pipe[1] (IMHO it is wrong, but let it be)
>> so it is reasonable to completely prohibit change O_APPEND flag on both
>> end's of pipe. Add ->check_flags method in order to diallow O_APPEND toggling.
>> Signed-off-by: Dmitry Monakhov <dmonakhov@xxxxxxxxxx>
>> ---
> TBH, all those ->direct_IO != NULL checks seem to be a wrong approach.
> If nothing else, it forces several filesystem into inventing a fake
> ->direct_IO just to fool those tests. How about we
> * introduce FMODE_MAY_DIRECT and allow ->open() explicitly set it
> * make open_check_o_direct() and fcntl.c check that instead of poking
> in ->f_mapping->a_ops, etc.
> * provide a variant of generic_file_open() that would set that
> bit and use it on the filesystems that handle dio
100% agree. FMODE is perfect place for that.

BTW: I always wondering: why we do not mark pipe[1]->f_flags with O_APPEND?
Probably the answer is that nobody care about ->f_flags since no_llseek
returns -ESPIPE, but f_flags are visiable via fcntl so IMHO it is
reasonable to fix that too.
> --
> To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at

Attachment: signature.asc
Description: PGP signature