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;
Message-ID: <1a0a859b-1f25-5136-bb86-9efe68aabbb8@suse.cz>
Date:   Tue, 24 May 2022 22:37:15 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.8.1
Subject: Re: Memory allocation on speculative fastpaths
Content-Language: en-US
To:     Johannes Weiner <hannes@cmpxchg.org>,
        Suren Baghdasaryan <surenb@google.com>
Cc:     Matthew Wilcox <willy@infradead.org>,
        "Paul E. McKenney" <paulmck@kernel.org>,
        Michal Hocko <mhocko@suse.com>,
        "Liam R. Howlett" <liam.howlett@oracle.com>,
        Michel Lespinasse <walken.cr@gmail.com>,
        linux-mm <linux-mm@kvack.org>,
        LKML <linux-kernel@vger.kernel.org>,
        David Hildenbrand <david@redhat.com>,
        Davidlohr Bueso <dave@stgolabs.net>
References: <20220503155913.GA1187610@paulmck-ThinkPad-P17-Gen-1>
 <YnFSfc8BR8CadOtw@dhcp22.suse.cz>
 <20220503163905.GM1790663@paulmck-ThinkPad-P17-Gen-1>
 <YnF0RyBaBSC1mdKo@casper.infradead.org>
 <CAJuCfpG8mBCV8O=FWwTJj8zfoh68fH9kBraMEjxKUvCyEw2MFw@mail.gmail.com>
 <YnKocgiWrupyFki3@cmpxchg.org>
From:   Vlastimil Babka <vbabka@suse.cz>
In-Reply-To: <YnKocgiWrupyFki3@cmpxchg.org>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 5/4/22 18:23, Johannes Weiner wrote:
> On Tue, May 03, 2022 at 04:15:46PM -0700, Suren Baghdasaryan wrote:
>> On Tue, May 3, 2022 at 11:28 AM Matthew Wilcox <willy@infradead.org> wrote:
>>>
>>> On Tue, May 03, 2022 at 09:39:05AM -0700, Paul E. McKenney wrote:
>>>> On Tue, May 03, 2022 at 06:04:13PM +0200, Michal Hocko wrote:
>>>>> On Tue 03-05-22 08:59:13, Paul E. McKenney wrote:
>>>>>> Hello!
>>>>>>
>>>>>> Just following up from off-list discussions yesterday.
>>>>>>
>>>>>> The requirements to allocate on an RCU-protected speculative fastpath
>>>>>> seem to be as follows:
>>>>>>
>>>>>> 1.        Never sleep.
>>>>>> 2.        Never reclaim.
>>>>>> 3.        Leave emergency pools alone.
>>>>>>
>>>>>> Any others?
>>>>>>
>>>>>> If those rules suffice, and if my understanding of the GFP flags is
>>>>>> correct (ha!!!), then the following GFP flags should cover this:
>>>>>>
>>>>>>   __GFP_NOMEMALLOC | __GFP_NOWARN
>>>>>
>>>>> GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN
>>>>
>>>> Ah, good point on GFP_NOWAIT, thank you!
>>>
>>> Johannes (I think it was?) made the point to me that if we have another
>>> task very slowly freeing memory, a task in this path can take advantage
>>> of that other task's hard work and never go into reclaim.  So the
>>> approach we should take is:
> 
> Right, GFP_NOWAIT can starve out other allocations. It can clear out
> the freelists without the burden of having to do reclaim like
> everybody else wanting memory during a shortage. Including GFP_KERNEL.

FTR, I wonder if this is really true, given the suggested fallback. With
GFP_NOWAIT, you can either see memory (in all applicable zones) as

a) above low_watermark, just go ahead and allocate, as GFP_KERNEL would
b) between min and low watermark, wake up kswapd and allocate, as
GFP_KERNEL would
c) below min watermark, the most interesting. GFP_KERNEL fallbacks to
reclaim. If the GFP_NOWAIT path's fallback also includes reclaim, as
suggested in this thread, how is it really different from GFP_KERNEL?

So am I missing something or is GFP_NOWAIT fastpath with an immediate
fallback that includes reclaim (and not just a retry loop) fundamentally
not different from GFP_KERNEL, regardless of how often we attempt it?

> In smaller doses and/or for privileged purposes (e.g. single-argument
> kfree_rcu ;)), those allocations are fine. But because the context is
> page tables specifically, it would mean that userspace could trigger a
> large number of those and DOS other applications and the kernel.
> 
>>> p4d_alloc(GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN);
>>> pud_alloc(GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN);
>>> pmd_alloc(GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN);
>>>
>>> if (failure) {
>>>   rcu_read_unlock();
>>>   do_reclaim();
>>>   return FAULT_FLAG_RETRY;
>>> }
>>>
>>> ... but all this is now moot since the approach we agreed to yesterday
>>> is:
>>
>> I think the discussion was about the above approach and Johannes
>> suggested to fallback to the normal pagefault handling with mmap_lock
>> locked if PMD does not exist. Please correct me if I misunderstood
>> here.
> 
> Yeah. Either way works, as long as the task is held accountable.
>