mc hangs accessing files within a .tar.gz within a cpio within an rpm for first 60 seconds

[nvwarr]

In lib/vfs/gc.c in the function vfs_expire, curr_time and exp_time are declared guint64. curr_time is initialised with a timestamp and exp_time with this timestamp minus 60 seconds. Later there is if (stamping->time <= exp_time). Prior to commit a94dd7d2ded0dd3ce3f6cd2c56612327d065b5ac curr_time was initialised with a value larger than 60 seconds, so everything was fine. This commit changed the initialisation to a timer starting when mc is started. So for the first 60 seconds, the result of the subtraction is negative, but it is a guint64, so we just get a VERY large unsigned value and the if (stamping->time <= exp_time) is always true. So mc thinks the vfs hasn't been used recently and goes into an infinite loop.

If one opens a .rpm file with mc and goes into the CONTENTS.cpio and then tries to go into the .tar.gz there (this is the usual structure of a .rpm) after waiting 60 seconds, everything is fine. However, before 60 seconds, mc hangs. There may be other ways to trigger this, but this is how I discovered it.

The fix is to remove the subtraction of 60 seconds when initialising exp_time and add them to the left-hand side of the comparison.

[nvwarr]

Patch to fix the bug

[andrew_b]

Remembering discussion #3859, I think it's time to drop mc_timer and use g_get_real_time(). There is no difference what time to use: relative to mc start moment or relative to UNIX epoch.

[andrew_b]

Branch: 4147_vfs_timeout
Initial changeset:e612ea50a12c70dd1d636a3086c6889db52f9ce0

Please test.

[nvwarr]

Branch 4147_vfs_timeout works for me. So that is another way to solve the problem. I still think that you should consider avoiding the subtraction of vfs_timeout seconds from an unsigned type, though. That's always a bit dangerous. I would consider that the real bug, which was exposed by using mc_timer and could potentially reoccur if a future change goes back to a relative time since startup of mc. That said, your fix does avoid the issue.

[andrew_b]

Replying to nvwarr:

I still think that you should consider avoiding the subtraction of vfs_timeout seconds from an unsigned type, though.

guint64 is replaced with gint64 for all related variables and structure members.

That's always a bit dangerous. I would consider that the real bug, which was exposed by using mc_timer and could potentially reoccur if a future change goes back to a relative time since startup of mc.

mc_timer is removed in the 2nd commit in this branch.

[nvwarr]

Oops, I missed the change to gint64. That's perfect then.

[andrew_b]

Merged to master: [446a031350d4da12f521cc37881bdaf016b33c1f].

git log --pretty=oneline f0a3794b0..446a03135