2008-10-10 19:18:24

by dcg

[permalink] [raw]
Subject: [PATCH] Document file->f_pos as thread-unsafe

El Fri, 10 Oct 2008 13:25:15 +1100, Nick Piggin <[email protected]> escribió:

> Note that I don't think we'd want to explicitly guarantee that it is atomic
> on 64-bit machines either. It does happen to be, but I don't think we want
> anybody to rely on that...

Easily fixable. Performance testing of this new version didn't find
regressions.




Document that it is not safe to play with file->f_pos from multiple processes
if they aren't synchronizing themselves.

Signed-off-by: Diego Calleja <[email protected]>

Index: 2.6/include/linux/fs.h
===================================================================
--- 2.6.orig/include/linux/fs.h 2008-10-10 21:07:29.000000000 +0200
+++ 2.6/include/linux/fs.h 2008-10-10 21:08:25.000000000 +0200
@@ -821,6 +821,17 @@
atomic_long_t f_count;
unsigned int f_flags;
mode_t f_mode;
+ /*
+ * Linux does NOT guarantee atomic reading/writing to file->f_pos in
+ * multithreaded apps. There're several reasons for this behaviour:
+ * - Specifications don't say it must be implemented that way.
+ * - This behaviour is part of the Linux semantics.
+ * - Any application that does multithreaded access to file->f_pos
+ * should be doing its own locking: the processes should synchronize
+ * themselves when accessing a file descriptor. If an application
+ * doesn't do that, its file descriptor handling is buggy anyway and
+ * must be fixed to do its accesses to file->f_pos properly.
+ */
loff_t f_pos;
struct fown_struct f_owner;
unsigned int f_uid, f_gid;