Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Fri, 28 Oct 2022 15:42:56 -0700
From:   Beau Belgrave <beaub@linux.microsoft.com>
To:     Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc:     rostedt@goodmis.org, mhiramat@kernel.org,
        dcook@linux.microsoft.com, alanau@linux.microsoft.com,
        linux-trace-devel@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [RFC PATCH 2/2] tracing/user_events: Fixup enable faults asyncly
Message-ID: <20221028224256.GA202@W11-BEAU-MD.localdomain>
References: <20221027224011.2075-1-beaub@linux.microsoft.com>
 <20221027224011.2075-3-beaub@linux.microsoft.com>
 <a18e940d-8423-0294-23b4-f2702313f3eb@efficios.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <a18e940d-8423-0294-23b4-f2702313f3eb@efficios.com>
Precedence: bulk

On Fri, Oct 28, 2022 at 06:19:10PM -0400, Mathieu Desnoyers wrote:
> On 2022-10-27 18:40, Beau Belgrave wrote:
> > When events are enabled within the various tracing facilities, such as
> > ftrace/perf, the event_mutex is held. As events are enabled pages are
> > accessed. We do not want page faults to occur under this lock. Instead
> > queue the fault to a workqueue to be handled in a process context safe
> > way without the lock.
> > 
> > The enable address is disabled while the async fault-in occurs. This
> > ensures that we don't attempt fault-in more than is necessary. Once the
> > page has been faulted in, the address write is attempted again. If the
> > page couldn't fault-in, then we wait until the next time the event is
> > enabled to prevent any potential infinite loops.
> 
> I'm also unclear about how the system call initiating the enabled state
> change is delayed (or not) when a page fault is queued.
> 

It's not, if needed we could call schedule_delayed_work(). However, I
don't think we need it. When pin_user_pages_remote is invoked, it's with
FOLL_NOFAULT. This will tell us if we need to fault pages in, we then
call fixup_user_fault with unlocked value passed. This will cause the
fixup to retry and get the page in.

It's called out in the comments for this exact purpose (lucked out
here):
mm/gup.c
 * This is meant to be called in the specific scenario where for locking reasons
 * we try to access user memory in atomic context (within a pagefault_disable()
 * section), this returns -EFAULT, and we want to resolve the user fault before
 * trying again.

The fault in happens in a workqueue, this is the same way KVM does it's
async page fault logic, so it's not a new pattern. As soon as the
fault-in is done, we have a page we should be able to use, so we
re-attempt the write immediately. If the write fails, another queue
happens and we could loop, but not like the unmap() case I replied with
earlier.

> I would expect that when a page fault is needed, after enqueuing work to the
> worker thread, the system call initiating the state change would somehow
> wait for a completion (after releasing the user events mutex). That
> completion would be signaled by the worker thread either if the page fault
> fails, or if the state change is done.
> 

I didn't see a generic fault-in + notify anywhere, do you know of one I
could use? Otherwise, it seems the pattern used is check fault, fault-in
via workqueue, re-attempt.

> Thoughts ?
> 
> Thanks,
> 
> Mathieu
> 

Thanks,
-Beau