Google Summer of Code Project Idea

File Locking.



Existing Work

GNU Savannah patch #332.

IRC, freenode, #hurd, 2010-12-31

<pinotree> youpi: i found the issue with python-apt
<pinotree> s/with/of/
<youpi> good!
<pinotree> lock file issue, though :/
<youpi> :/
<pinotree> this is the sample test case, derived from apt's code:
<pinotree> basically, it seems asking for a file lock in the same process
  where there's already such lock on the file, fails
<pinotree> youpi: ↑
<youpi> uh, posix doesn't even define some nesting
<pinotree> it seems it just talks about concurrency with other processes
<youpi> posix tells more about it later
<youpi> saying that if a lock already exists, then it is replaced by the
<youpi> (when inside the same process)
<pinotree> yay, found a bug in hurd :p
<youpi> well, actually it's known
<youpi> i.e. setlk is completely bogus, based on flock
<youpi> and flock doesn't have the same semantic in that regard
<youpi> so we can't fix it without really implementing setlk
<pinotree> the XXX comment in glibc/sysdeps/mach/hurd/fcntl.c, by chance?
<youpi> of course :)
<pinotree> youpi: hm, flock's man page says:
<pinotree> "A process may only hold one type of lock (shared or exclusive)
  on a file. Subsequent flock() calls on an already locked file will
  convert an existing lock to the new lock mode."
<pinotree> so a new lock in the same process over the original lock should
  replace the old one?
<youpi> uh, that's not what I had seen
<youpi> An attempt to lock the file using one of these file  descrip-
<youpi>        tors  may  be  denied  by  a  lock that the calling process
  has already
<youpi>        placed via another descriptor.
<youpi> so it's really not that easy
<pinotree> that's in case of trying to create a lock on a file with a
  different fd than the existing lock
<youpi> that's what your testcase does
<pinotree> which, hm, is python-apt's case
<youpi> that being said, the sentence I pasted does not seem to appear in
<pinotree> flock() does not seem posix
<youpi> it may have been the behavior of Linux at some point in the past
<youpi> it's not , but F_SETLK is
<youpi> and in linux world, flock <=> F_SETLK, iirc
<youpi> in glibc world, even
<youpi> (just checked it, see sysdeps/posix/flock.c
<youpi> pinotree: I guess your testcase works on Linux?
<pinotree> which means we should get a proper F_SETLK working, and then
  just use this flock version (instead of the custom one), no?
<pinotree> yes, it works on linux (and on kfreebsd, see that python-apt
<youpi> no, I mean our flock() should probably be happy with locking part
  of a file several times
<youpi> (that is, hurd's file_lock() RPC)
<youpi> ah, no, on Linux flock is its own system call
<youpi> (which is independant from lockf from the locking point of view,