bryson/pacman
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Use flocks to lock the database instead of purely relying on file existence.

Browse source 
With PackageKit or pacman itself, it sometimes occurs that the lock
file is left over erroneously. A manual administrative
    $ rm /var/lib/{pacman,PackageKit/alpm}/db.lck
is required.
This is also the preferred way to fix this issue as per the Wiki [1][2].

This is not really user-friendly when it happens, and has bothered me
personally about five or six times now to a point where I hacked this patch.

The lock file can now preexist when using libalpm, and libalpm now flock()s
the file exclusively. Linux ensures internally that no second exclusive lock
can be held. However, we still can't get rid of the unlink when unlocking
altogether, because existing bash scripts (e.g. makepkg.sh) depend on it.
Refactoring those to use flocks would be quite a mess, too. But because the
lock file should normally not exist without any lock on it, the scripts would
simply say that someone holds the lock. Refreshing the pacman database now fixes
the issue on it's own. This is now also user triggerable when using
PackageKit :)

[1] https://wiki.archlinux.org/title/Pacman
    3.3 "Failed to init transaction (unable to lock database)" error
[2] https://wiki.archlinux.org/title/KDE
    6.15 Discover stops showing updates from Arch repositories

Signed-off-by: Stephan Brunner <s.brunner@stephan-brunner.net>
This commit is contained in:
Stephan Brunner 2024-04-20 17:30:39 +02:00
parent 4dc21b965b
commit 5882e21806
No known key found for this signature in database
GPG key ID: D3CD7237AFB7EF6B
1 changed files with 56 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

59
lib/libalpm/handle.c
View file

@ -27,6 +27,7 @@
#include <sys/types.h>
#include <syslog.h>
#include <sys/stat.h>
#include <sys/file.h>
#include <fcntl.h>
/* libalpm */
@ -118,6 +119,7 @@ void _alpm_handle_free(alpm_handle_t *handle)
int _alpm_handle_lock(alpm_handle_t *handle)
{
char *dir, *ptr;
int ret;
ASSERT(handle->lockfile != NULL, return -1);
ASSERT(handle->lockfd < 0, return 0);
@ -135,22 +137,73 @@ int _alpm_handle_lock(alpm_handle_t *handle)
FREE(dir);
do {
handle->lockfd = open(handle->lockfile, O_WRONLY | O_CREAT | O_EXCL | O_CLOEXEC, 0000);
handle->lockfd = open(handle->lockfile, O_RDWR | O_CREAT | O_CLOEXEC, S_IRUSR | S_IWUSR);
} while(handle->lockfd == -1 && errno == EINTR);
return (handle->lockfd >= 0 ? 0 : -1);
if (handle->lockfd == -1) {
return -1;
}
do {
ret = flock(handle->lockfd, LOCK_EX | LOCK_NB);
} while(ret != 0 && errno == EINTR);
if(ret != 0) {
/* Do not unlink here as someone else seems to hold the lock!
* We do not check for the exact error, as many errors may occur
* and we don't want to unlock the file by deleting it without
* being absolutely sure no one else has locked it.
*/
close(handle->lockfd);
handle->lockfd = -1;
}
return ret == 0 ? 0 : -1;
}
int SYMEXPORT alpm_unlock(alpm_handle_t *handle)
{
int ret_flock, ret_unlink;
ASSERT(handle != NULL, return -1);
ASSERT(handle->lockfile != NULL, return 0);
ASSERT(handle->lockfd >= 0, return 0);
/* We can't get rid of the unlink because of backwards compatibility.
* And it would make the bash script very convoluted when implementing flocks.
* But first unlink, then unlock, so that no race condition can occur.
*
* Unlock-then-Unlink:
* 1. TX1: flock(fd1, LOCK_UN)
* 2. TX2: fd1 = open(fn)
* 3. TX1: unlink(fd1)
* 4. TX2: flock(fd1, LOCK_EX)
* ^ Succeeds, TX2 should be the exclusive owner now
* 5. TX3: fd2 = open(fn)/flock(fd2, LOCK_EX)
* ^ Would succeed too, as the file was first unlinked and is now
* opened as a new file. We have now two exclusive locks.
*
* Unlink-then-Unlock:
* 1. TX1: unlink(fd1)
* 2. TX2: fd2 = open(fn)
* ^ Creates a new file
* 3. TX1: flock(fd1, LOCK_UN)
* 4. TX2: flock(fd2, LOCK_EX)
* 5. TX3: open(fn2)/flock(fd2, LOCK_EX)
* ^ Would fail, as the file is now shared between the two TX.
*
* Basically, the unlink() is the actual unlock here, as it was without flock()
* But now, we can handle the case when the lock file already existed, too.
* Check for both unlink and flock errors, because the flock-call unlocks
* the file even without the unlink call.
*/
ret_unlink = unlink(handle->lockfile);
ret_flock = flock(handle->lockfd, LOCK_UN);
close(handle->lockfd);
handle->lockfd = -1;
if(unlink(handle->lockfile) != 0) {
if(ret_unlink != 0 || ret_flock != 0) {
RET_ERR_ASYNC_SAFE(handle, ALPM_ERR_SYSTEM, -1);
} else {
return 0;