Re: [PATCH v10 3/6] fs/proc/task_mmu: Implement IOCTL to get and/or the clear info about PTEs

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 




> On Feb 28, 2023, at 7:55 AM, Peter Xu <peterx@xxxxxxxxxx> wrote:
> 
> !! External Email
> 
> On Mon, Feb 27, 2023 at 11:09:12PM +0000, Nadav Amit wrote:
>> 
>> 
>>> On Feb 27, 2023, at 1:18 PM, Peter Xu <peterx@xxxxxxxxxx> wrote:
>>> 
>>> !! External Email
>>> 
>>> On Thu, Feb 23, 2023 at 05:11:11PM +0000, Nadav Amit wrote:
>>>> From my experience with UFFD, proper ordering of events  is crucial, although it
>>>> is not always done well. Therefore, we should aim for improvement, not
>>>> regression. I believe that utilizing the pagemap-based mechanism for WP'ing
>>>> might be a step in the wrong direction. I think that it would have been better
>>>> to emit a 'UFFD_FEATURE_WP_ASYNC' WP-log (and ordered) with UFFD #PF and
>>>> events. The 'UFFD_FEATURE_WP_ASYNC'-log may not need to wake waiters on the
>>>> file descriptor unless the log is full.
>>> 
>>> Yes this is an interesting question to think about..
>>> 
>>> Keeping the data in the pgtable has one good thing that it doesn't need any
>>> complexity on maintaining the log, and no possibility of "log full".
>> 
>> I understand your concern, but I think that eventually it might be simpler
>> to maintain, since the logic of how to process the log is moved to userspace.
>> 
>> At the same time, handling inputs from pagemap and uffd handlers and sync’ing
>> them would not be too easy for userspace.
> 
> I do not expect a common uffd-wp async user to provide a fault handler at
> all.  In my imagination it's in most cases used standalone from other uffd
> modes; it means all the faults will still be handled by the kernel.  Here
> we only leverage the accuracy of userfaultfd comparing to soft-dirty, so
> not really real "user"-faults.

If that is the only use-case, it might make sense. But I guess most users would
most likely use some library (and not syscalls directly). So slightly
complicating the API for better generality may be reasonable.

> 
>> 
>> But yes, allocation on the heap for userfaultfd_wait_queue-like entries would
>> be needed, and there are some issues of ordering the events (I think all #PF
>> and other events should be ordered regardless) and how not to traverse all
>> async-userfaultfd_wait_queue’s (except those that block if the log is full)
>> when a wakeup is needed.
> 
> Will there be an ordering requirement for an async mode?  Considering it
> should be async to whatever else, I would think it's not a problem, but
> maybe I missed something.

You may be right, but I am not sure. I am still not sure what use-cases are
targeted in this patch-set. For CRIU checkpoint use-case (when the app is
not running), I guess the current interface makes sense. But if there are
use-cases in which this you do care about UFFD-events this can become an
issue.

But even in some obvious use-cases, this might be the wrong interface for
major performance issues. If we think about some incremental copying of
modified pages (a-la pre-copy live-migration or to create point-in-time
snapshots), it seems to me much more efficient for application to have a
log than traversing all the page-tables.


>> 
>>> 
>>> If there's possible "log full" then the next question is whether we should
>>> let the worker wait the monitor if the monitor is not fast enough to
>>> collect those data.  It adds some slight dependency on the two threads, I
>>> think it can make the tracking harder or impossible in latency sensitive
>>> workloads.
>> 
>> Again, I understand your concern. But this model that I propose is not new.
>> It is used with PML (page-modification logging) and KVM, and IIRC there is
>> a similar interface between KVM and QEMU to provide this information. There
>> are endless other examples for similar producer-consumer mechanisms that
>> might lead to stall in extreme cases.
> 
> Yes, I'm not against thinking of using similar structures here.  It's just
> that it's definitely more complicated on the interface, at least we need
> yet one more interface to setup the rings and define its interfaces.
> 
> Note that although Muhammud is defining another new interface here too for
> pagemap, I don't think it's strictly needed for uffd-wp async mode.  One
> can use uffd-wp async mode with PM_UFFD_WP which is with current pagemap
> interface already.
> 
> So what Muhammud is proposing here are two things to me: (1) uffd-wp async,
> plus (2) a new pagemap interface (which will closely work with (1) only if
> we need atomicity on get-dirty and reprotect).
> 
> Defining new interface for uffd-wp async mode will be something extra, so
> IMHO besides the heap allocation on the rings, we need to also justify
> whether that is needed.  That's why I think it's fine to go with what
> Muhammud proposed, because it's a minimum changeset at least for userfault
> to support an async mode, and anything else can be done on top if necessary.
> 
> Going a bit back to the "lead to stall in extreme cases" above, just also
> want to mention that the VM use case is slightly different - dirty tracking
> is only heavily used during migration afaict, and it's a short period.  Not
> a lot of people will complain performance degrades during that period
> because that's just rare.  And, even without the ring the perf is really
> bad during migration anyway... Especially when huge pages are used to back
> the guest RAM.
> 
> Here it's slightly different to me: it's about tracking dirty pages during
> any possible workload, and it can be monitored periodically and frequently.
> So IMHO stricter than a VM use case where migration is the only period to
> use it.

I still don’t get the use-cases. "monitored periodically and frequently” is
not a use-case. And as I said before, actually, monitoring frequently is
more performant with a log than with scanning all the page-tables.

> 
>> 
>>> 
>>> The other thing is we can also make the log "never gonna full" by making it
>>> a bitmap covering any registered ranges, but I don't either know whether
>>> it'll be worth it for the effort.
>> 
>> I do not see a benefit of half-log half-scan. It tries to take the
>> data-structure of one format and combine it with another.
> 
> What I'm saying here is not half-log / half-scan, but use a single bitmap
> to store what page is dirty, just like KVM_GET_DIRTY_LOG.  I think it
> avoids any above "stall" issue.

Oh, I never went into the KVM details before - stupid me. If that’s what
eventually was proven to work for KVM/QEMU, then it really sounds like
the pagemap solution that Muhammad proposed.

But still not convoluting pagemap with userfaultfd (and especially
uffd-wp) can be beneficial. Linus already threw some comments here and
there about disliking uffd-wp, and I’m not sure adding uffd-wp specific
stuff to pagemap would be welcomed.

Anyhow, thanks for all the explanations. Eventually, I understand that
using bitmaps can be more efficient than a log if the bits are condensed.





[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux