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, 26 May 2023 12:25:05 -0400
From:   Peter Xu <peterx@redhat.com>
To:     Hugh Dickins <hughd@google.com>
Cc:     Andrew Morton <akpm@linux-foundation.org>,
        Mike Kravetz <mike.kravetz@oracle.com>,
        Mike Rapoport <rppt@kernel.org>,
        "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>,
        Matthew Wilcox <willy@infradead.org>,
        David Hildenbrand <david@redhat.com>,
        Suren Baghdasaryan <surenb@google.com>,
        Qi Zheng <zhengqi.arch@bytedance.com>,
        Yang Shi <shy828301@gmail.com>,
        Mel Gorman <mgorman@techsingularity.net>,
        Peter Zijlstra <peterz@infradead.org>,
        Will Deacon <will@kernel.org>, Yu Zhao <yuzhao@google.com>,
        Alistair Popple <apopple@nvidia.com>,
        Ralph Campbell <rcampbell@nvidia.com>,
        Ira Weiny <ira.weiny@intel.com>,
        Steven Price <steven.price@arm.com>,
        SeongJae Park <sj@kernel.org>,
        Naoya Horiguchi <naoya.horiguchi@nec.com>,
        Christophe Leroy <christophe.leroy@csgroup.eu>,
        Zack Rusin <zackr@vmware.com>, Jason Gunthorpe <jgg@ziepe.ca>,
        Axel Rasmussen <axelrasmussen@google.com>,
        Anshuman Khandual <anshuman.khandual@arm.com>,
        Pasha Tatashin <pasha.tatashin@soleen.com>,
        Miaohe Lin <linmiaohe@huawei.com>,
        Minchan Kim <minchan@kernel.org>,
        Christoph Hellwig <hch@infradead.org>,
        Song Liu <song@kernel.org>,
        Thomas Hellstrom <thomas.hellstrom@linux.intel.com>,
        linux-kernel@vger.kernel.org, linux-mm@kvack.org
Subject: Re: [PATCH 15/31] mm/userfaultfd: allow pte_offset_map_lock() to fail
Message-ID: <ZHDdYYQfqwiLSTzo@x1n>
References: <68a97fbe-5c1e-7ac6-72c-7b9c6290b370@google.com>
 <49d92b15-3442-4e84-39bd-c77c316bf844@google.com>
 <ZG6TOyH7NJIgXxGX@x1n>
 <8f2131ac-8996-e4b3-2aad-7a4d11bd538f@google.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
In-Reply-To: <8f2131ac-8996-e4b3-2aad-7a4d11bd538f@google.com>
Precedence: bulk

On Thu, May 25, 2023 at 03:06:27PM -0700, Hugh Dickins wrote:
> On Wed, 24 May 2023, Peter Xu wrote:
> > On Sun, May 21, 2023 at 10:07:35PM -0700, Hugh Dickins wrote:
> > > mfill_atomic_install_pte() and mfill_atomic_pte_zeropage() treat
> > > failed pte_offset_map_lock() as -EFAULT, with no attempt to retry.
> > 
> > Could you help explain why it should be -EFAULT, not -EAGAIN or -EEXIST?
> 
> Thanks a lot for looking, Peter.
> 
> No good justification for -EFAULT: I just grabbed the closest, fairly
> neutral, error code that I could see already being in use there: but now
> that you mention -EAGAIN, which I can see being used from mfill_atomic(),
> yes, that would be ideal - and consistent with how it's already being used.
> 
> I'll make that change, thanks for suggesting.  (And it had bugged me how
> my fs/userfaultfd.c was electing to retry, but this one electing to fail.)

Thanks.

> 
> > 
> > IIUC right now if pte existed we have -EEXIST returned as part of the
> > userfault ABI, no matter whether it's pte or thp.
> 
> It might or might not correspond to -EEXIST - it might even end up as
> -EFAULT on a retry after -EAGAIN: I see mfill_atomic() contains both
> -EEXIST and -EFAULT cases for pmd_trans_huge().  Actually, I could
> say that the -EFAULT case there corresponds to the -EFAULT in this
> 15/31 patch, but that would be by coincidence not design: I'm happier
> with your -EAGAIN suggestion.

I had a feeling that that 2nd -EFAULT there could crash some userapp
already if it got returned somewhere, because the userapp shouldn't expect
that.  IMHO it should also return -EAGAIN, or even -EEXIST because even if
user retries, we should highly possibly see that thp again, so the -EEXIST
should possibly follow anyway.

Not a big deal here I think - if an userapp can trigger that -EFAULT I'd
say it's also a user bug because it made two decisions already on resolving
page fault for single VA, and it's racy between them..

> 
> > 
> > IMHO it may boil down to my limited knowledge on how pte_offset_map_lock()
> > is used after this part 2 series, and I assume the core changes will be in
> > your 3rd series (besides this one and the arch one).
> > 
> > Please shed some light if there's quick answers (IIUC this is for speeding
> > up collapsing shmem thps, but still no much clue here), or I can also wait
> > for reading the 3rd part if it'll come soon in any form.
> 
> It wouldn't be particularly easy to deduce from the third series of
> patches, rather submerged in implementation details.  Just keep in mind
> that, like in the "old" pmd_trans_unstable() cases, there may be instants
> at which, when trying to get the lock on a page table, that page table
> might already have gone, or been replaced by something else e.g. a THP,
> and a retry necessary at the outer level (if it's important to persist).

I'm actually still curious how the 3rd series will look like; would love to
read it when it comes.

Thanks,

-- 
Peter Xu