Skip to contentSkip to conversation

fs: remove local variable copy of f_pos to enable thread-safe updates

PATCH Matti Linnanvuori 2 participants 2 messages May 2, 2008 VFS

lore
Share & tools
lore
#0 lore raw

----- Original Message ----

Quoted 9 lines

From: Al Viro <[email protected]>
To: Matti Linnanvuori <[email protected]>
Cc: [email protected]; [email protected]
Sent: Friday, May 2, 2008 12:33:02 PM
Subject: Re: [patch] fs: remove local variable copy of f_pos to enable thread-safe updates

Quoted 2 lines · Matti Linnanvuori

From: Matti Linnanvuori

Quoted 8 lines
Quoted 2 lines

Remove local variable copy of f_pos to enable thread-safe updates in implementation functions of system calls read, readv, write and writev.

... and we are back to square one with racy instances of ->read() that will use *pos twice, leaving a window for e.g. lseek() to invalidate e.g. wraparound tests. Not that deciding what to do at the end of e.g. read() with lseek() happening in the middle would be fun anyway - esp. since you don't remember the kind of seek it had been.

File systems can be written so that implementations of read, lseek etc. acquire a lock before using *pos, which precludes races. That was the idea behind my patch. So, my patch does not necessarily cause races. read instances can be racy in Linux now because there is no lock to serialize read etc. system calls.

      ____________________________________________________________________________________

Be a better friend, newshound, and
know-it-all with Yahoo! Mobile. Try it now. http://mobile.yahoo.com/;_ylt=Ahu06i62sR8HDtDypao8Wcj9tAcJ

maintainer
#1 lore raw
Quoted 5 lines · Matti Linnanvuori

File systems can be written so that implementations of read, lseek etc. acquire a lock before using *pos, which precludes races. That was the idea behind my patch. So, my patch does not necessarily cause races. read instances can be racy in Linux now because there is no lock to serialize read etc. system calls.

Oh, yes it does - current tree guarantees that variable passed to ->read() will not change under it, so any code that relied on that got broken.

And no, modifying every driver's ->read() is not going to be fun - it's not just filesystems.