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, 24 Mar 2023 13:31:11 +0000
From:   Matthew Wilcox <willy@infradead.org>
To:     Song Liu <songliubraving@meta.com>
Cc:     Hugh Dickins <hughd@google.com>, Song Liu <song@kernel.org>,
        Jiri Olsa <jolsa@kernel.org>,
        David Stevens <stevensd@chromium.org>,
        Linux-MM <linux-mm@kvack.org>,
        Andrew Morton <akpm@linux-foundation.org>,
        Peter Xu <peterx@redhat.com>,
        "Kirill A . Shutemov" <kirill@shutemov.name>,
        Yang Shi <shy828301@gmail.com>,
        David Hildenbrand <david@redhat.com>,
        Jiaqi Yan <jiaqiyan@google.com>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH v5 3/3] mm/khugepaged: maintain page cache uptodate flag
Message-ID: <ZB2mH/XMj3vh1RrN@casper.infradead.org>
References: <20230307052036.1520708-1-stevensd@google.com>
 <20230307052036.1520708-4-stevensd@google.com>
 <866d1a75-d462-563-dfd7-1aa2971a285b@google.com>
 <ZBzK99dSmbPB7vjA@casper.infradead.org>
 <CAPhsuW5pTh2+S7VjRbRy=WUAb7-GLQpi6HKCW0ehxs5HAup_dQ@mail.gmail.com>
 <b49b118d-893c-c749-61f6-9441539418a5@google.com>
 <ZB0ZSqw1TTp3/ec3@casper.infradead.org>
 <6F9B0A6A-B7F1-4E25-AB15-8F3782D2EF83@fb.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <6F9B0A6A-B7F1-4E25-AB15-8F3782D2EF83@fb.com>
Precedence: bulk

On Fri, Mar 24, 2023 at 06:03:37AM +0000, Song Liu wrote:
> 
> 
> > On Mar 23, 2023, at 8:30 PM, Matthew Wilcox <willy@infradead.org> wrote:
> 
> [...]
> 
> > 
> > The Uptodate flag check needs to be done by the caller; the
> > find_get_page() family return !uptodate pages.
> > 
> > But find_get_page() does not advertise itself as NMI-safe.  And I
> > think it's wrong to try to make it NMI-safe.  Most of the kernel is
> > not NMI-safe.  I think it's incumbent on the BPF people to get the
> > information they need ahead of taking the NMI.  NMI handlers are not
> > supposed to be doing a huge amount of work!  I don't really understand
> > why it needs to do work in NMI context; surely it can note the location of
> > the fault and queue work to be done later (eg on irq-enable, task-switch
> > or return-to-user)
> 
> The use case here is a profiler (similar to perf-record). Parsing the 
> build id in side the NMI makes the profiler a lot simpler. Otherwise, 
> we will need some post processing for each sample. 

Simpler for you, maybe.  But this is an NMI!  It's not supposed to
be doing printf-formatting or whatever, much less poking around
in the file cache.  Like perf, it should record a sample and then
convert that later.  Maybe it can defer to a tasklet, but i think
scheduling work is a better option.

> OTOH, it is totally fine if build_id_parse() fails some time, say < 5%. 
> The profiler output is still useful in such cases. 
> 
> I guess the next step is to replace find_get_page() with a NMI-safe
> version?

No, absolutely not.  Stop doing so much work in an NMI.