[Bug 275436] tmpfs does not honor memory limits on writes

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=275436

--- Comment #9 from Mike Karels <karels@freebsd.org> ---
(In reply to Konstantin Belousov from comment #8)

> Proposals to limit max tmpfs page uses to some percentage of the total of
> memory and swap would not work.

My proof-of-concept does not use total memory + swap.  Instead, it computes a
reserve which is (100 - memlimit%) * (free memory + swap) at time of
tmpfs_init().  This seems to work out reasonably in my test cases so far.  I
need to try with mapped files too.  I'll attach my current patch.

> There is simply no reasonable answer to question 'how to limit tmpfs physical
> memory consumption' without knowing the system load pattern.

There is no perfect solution, that is true.  But when memory + swap is low
enough, allowing tmpfs to proceed makes bad things happen, like killing
processes and hanging the system.  Backing off a little makes things better in
at least some cases, which seems worthwhile to me.  And tmpfs can report a
failure (ENOSPC) without being killed, unlike processes touching memory.  Yes,
a compile may fail, etc, but at least that is related to the shortage.  It
seems bad if writes to /tmp by an unprivileged user cause root or other
important processes to be killed.

-- 
You are receiving this mail because:
You are the assignee for the bug.