On 3/23/19 11:56 AM, Eric W. Biederman wrote:
> Jeff Mahoney <jeffm@xxxxxxxx> writes:
>
>> On 4/24/18 10:14 AM, Eric W. Biederman wrote:
>>> jeffm@xxxxxxxx writes:
>>>
>>>> From: Jeff Mahoney <jeffm@xxxxxxxx>
>>>>
>>>> Hi all -
>>>>
>>>> I recently encountered a customer issue where, on a machine with many TiB
>>>> of memory and a few hundred cores, after a task with a few thousand threads
>>>> and hundreds of files open exited, the system would softlockup. That
>>>> issue was (is still) being addressed by Nik Borisov's patch to add a
>>>> cond_resched call to shrink_dentry_list. The underlying issue is still
>>>> there, though. We just don't complain as loudly. When a huge task
>>>> exits, now the system is more or less unresponsive for about eight
>>>> minutes. All CPUs are pinned and every one of them is going through
>>>> dentry and inode eviction for the procfs files associated with each
>>>> thread. It's made worse by every CPU contending on the super's
>>>> inode list lock.
>>>>
>>>> The numbers get big. My test case was 4096 threads with 16384 files
>>>> open. It's a contrived example, but not that far off from the actual
>>>> customer case. In this case, a simple "find /proc" would create around
>>>> 300 million dentry/inode pairs. More practically, lsof(1) does it too,
>>>> it just takes longer. On smaller systems, memory pressure starts pushing
>>>> them out. Memory pressure isn't really an issue on this machine, so we
>>>> end up using well over 100GB for proc files. It's the combination of
>>>> the wasted CPU cycles in teardown and the wasted memory at runtime that
>>>> pushed me to take this approach.
>>>>
>>>> The biggest culprit is the "fd" and "fdinfo" directories, but those are
>>>> made worse by there being multiple copies of them even for the same
>>>> task without threads getting involved:
>>>>
>>>> - /proc/pid/fd and /proc/pid/task/pid/fd are identical but share no
>>>> resources.
>>>>
>>>> - Every /proc/pid/task/*/fd directory in a thread group has identical
>>>> contents (unless unshare(CLONE_FILES) was called), but share no
>>>> resources.
>>>>
>>>> - If we do a lookup like /proc/pid/fd on a member of a thread group,
>>>> we'll get a valid directory. Inside, there will be a complete
>>>> copy of /proc/pid/task/* just like in /proc/tgid/task. Again,
>>>> nothing is shared.
>>>>
>>>> This patch set reduces some (most) of the duplication by conditionally
>>>> replacing some of the directories with symbolic links to copies that are
>>>> identical.
>>>>
>>>> 1) Eliminate the duplication of the task directories between threads.
>>>> The task directory belongs to the thread leader and the threads
>>>> link to it: e.g. /proc/915/task -> ../910/task This mainly
>>>> reduces duplication when individual threads are looked up directly
>>>> at the tgid level. The impact varies based on the number of threads.
>>>> The user has to go out of their way in order to mess up their system
>>>> in this way. But if they were so inclined, they could create ~550
>>>> billion inodes and dentries using the test case.
>>>>
>>>> 2) Eliminate the duplication of directories that are created identically
>>>> between the tgid-level pid directory and its task directory: fd,
>>>> fdinfo, ns, net, attr. There is obviously more duplication between
>>>> the two directories, but replacing a file with a symbolic link
>>>> doesn't get us anything. This reduces the number of files associated
>>>> with fd and fdinfo by half if threads aren't involved.
>>>>
>>>> 3) Eliminate the duplication of fd and fdinfo directories among threads
>>>> that share a files_struct. We check at directory creation time if
>>>> the task is a group leader and if not, whether it shares ->files with
>>>> the group leader. If so, we create a symbolic link to ../tgid/fd*.
>>>> We use a d_revalidate callback to check whether the thread has called
>>>> unshare(CLONE_FILES) and, if so, fail the revalidation for the symlink.
>>>> Upon re-lookup, a directory will be created in its place. This is
>>>> pretty simple, so if the thread group leader calls unshare, all threads
>>>> get directories.
>>>>
>>>> With these patches applied, running the same testcase, the proc_inode
>>>> cache only gets to about 600k objects, which is about 99.7% fewer. I
>>>> get that procfs isn't supposed to be scalable, but this is kind of
>>>> extreme. :)
>>>>
>>>> Finally, I'm not a procfs expert. I'm posting this as an RFC for folks
>>>> with more knowledge of the details to pick it apart. The biggest is that
>>>> I'm not sure if any tools depend on any of these things being directories
>>>> instead of symlinks. I'd hope not, but I don't have the answer. I'm
>>>> sure there are corner cases I'm missing. Hopefully, it's not just flat
>>>> out broken since this is a problem that does need solving.
>>>>
>>>> Now I'll go put on the fireproof suit.
>>
>> Thanks for your comments. This ended up having to get back-burnered but
>> I've finally found some time to get back to it. I have new patches that
>> don't treat each entry as a special case and makes more sense, IMO.
>> They're not worth posting yet since some of the issues below remain.
>>
>>> This needs to be tested against at least apparmor to see if this breaks
>>> common policies. Changing files to symlinks in proc has a bad habit of
>>> either breaking apparmor policies or userspace assumptions. Symbolic
>>> links are unfortunately visible to userspace.
>>
>> AppArmor uses the @{pids} var in profiles that translates to a numeric
>> regex. That means that /proc/pid/task -> /proc/tgid/task won't break
>> profiles but /proc/pid/fdinfo -> /proc/pid/task/tgid/fdinfo will break.
>> Apparmor doesn't have a follow_link hook at all, so all that matters is
>> the final path. SELinux does have a follow_link hook, but I'm not
>> familiar enough with it to know whether introducing a symlink in proc
>> will make a difference.
>>
>> I've dropped the /proc/pid/{dirs} -> /proc/pid/task/pid/{dirs} part
>> since that clearly won't work.
>>
>>> Further the proc structure is tgid/task/tid where the leaf directories
>>> are per thread.
>>
>> Yes, but threads are still in /proc for lookup at the tgid level even if
>> they don't show up in readdir.
>>
>>> We more likely could get away with some magic symlinks (that would not
>>> be user visible) rather than actual symlinks.
>>
>> I think I'm missing something here. Aren't magic symlinks still
>> represented to the user as symlinks?
>>
>>> So I think you are probably on the right track to reduce the memory
>>> usage but I think some more work will be needed to make it transparently
>>> backwards compatible.
>>
>> Yeah, that's going to be the big hiccup. I think I've resolved the
>> biggest issue with AppArmor, but I don't think the problem is solvable
>> without introducing symlinks.
>
> Has anyone looked at making the fd and fdinfo files hard links.
That could work to a certain degree. It would certainly reduce the
inode count. It would still create all the dentries, though. That's
still a n^2 problem where n is the number of threads in the group.
> Alternatively it may make sense to see if there is something that we can
> do with the locking to reduce the thundering hurd problem that is being
> seen.
Yeah, that could still use some attention. The thundering herd problem
is more of a tap when you reduce the contention by 99% though.
-Jeff
--
Jeff Mahoney
SUSE Labs
Attachment:
signature.asc
Description: OpenPGP digital signature
