lkmm.lists.linux.dev archive mirror
 help / color / mirror / Atom feed
From: Thomas Haas <t.haas@tu-bs.de>
To: Will Deacon <will@kernel.org>,
	Hernan Ponce de Leon <hernan.poncedeleon@huaweicloud.com>
Cc: Peter Zijlstra <peterz@infradead.org>,
	Alan Stern <stern@rowland.harvard.edu>,
	Andrea Parri <parri.andrea@gmail.com>,
	Boqun Feng <boqun.feng@gmail.com>,
	Nicholas Piggin <npiggin@gmail.com>,
	David Howells <dhowells@redhat.com>,
	Jade Alglave <j.alglave@ucl.ac.uk>,
	Luc Maranget <luc.maranget@inria.fr>,
	"Paul E. McKenney" <paulmck@kernel.org>,
	Akira Yokosawa <akiyks@gmail.com>,
	Daniel Lustig <dlustig@nvidia.com>,
	Joel Fernandes <joelagnelf@nvidia.com>,
	<linux-kernel@vger.kernel.org>, <linux-arch@vger.kernel.org>,
	<lkmm@lists.linux.dev>, <jonas.oberhauser@huaweicloud.com>,
	"r.maseli@tu-bs.de" <r.maseli@tu-bs.de>
Subject: Re: [RFC] Potential problem in qspinlock due to mixed-size accesses
Date: Thu, 19 Jun 2025 16:11:47 +0200	[thread overview]
Message-ID: <5e84e9fa-38a9-45fd-a67c-2db399860480@tu-bs.de> (raw)
In-Reply-To: <20250619123002.GC21372@willie-the-truck>



On 19.06.25 14:30, Will Deacon wrote:
> On Tue, Jun 17, 2025 at 09:00:30PM +0200, Hernan Ponce de Leon wrote:
>> On 6/17/2025 4:23 PM, Thomas Haas wrote:
>>>
>>>
>>> On 17.06.25 16:17, Will Deacon wrote:
>>>> On Tue, Jun 17, 2025 at 10:42:04AM +0200, Hernan Ponce de Leon wrote:
>>>>> On 6/17/2025 8:19 AM, Thomas Haas wrote:
>>>>>> On 16.06.25 16:23, Will Deacon wrote:
>>>>>>> I'm half inclined to think that the Arm memory model
>>>>>>> should be tightened
>>>>>>> here; I can raise that with Arm and see what they say.
>>>>>>>
>>>>>>> Although the cited paper does give examples of store-forwarding from a
>>>>>>> narrow store to a wider load, the case in qspinlock is further
>>>>>>> constrained by having the store come from an atomic rmw and the load
>>>>>>> having acquire semantics. Setting aside the MSA part,
>>>>>>> that specific case
>>>>>>> _is_ ordered in the Arm memory model (and C++ release
>>>>>>> sequences rely on
>>>>>>> it iirc), so it's fair to say that Arm CPUs don't permit
>>>>>>> forwarding from
>>>>>>> an atomic rmw to an acquire load.
>>>>>>>
>>>>>>> Given that, I don't see how this is going to occur in practice.
>>>>>>
>>>>>> You are probably right. The ARM model's atomic-ordered-before relation
>>>>>>
>>>>>>         let aob = rmw | [range(rmw)]; lrs; [A | Q]
>>>>>>
>>>>>> clearly orders the rmw-store with subsequent acquire loads
>>>>>> (lrs = local-
>>>>>> read-successor, A = acquire).
>>>>>> If we treat this relation (at least the second part) as a "global
>>>>>> ordering" and extend it by "si" (same-instruction), then the
>>>>>> problematic
>>>>>> reordering under MSA should be gone.
>>>>>> I quickly ran Dartagnan on the MSA litmus tests with this change to the
>>>>>> ARM model and all the tests still pass.
>>>>>
>>>>> Even with this change I still get violations (both safety and
>>>>> termination)
>>>>> for qspinlock with dartagnan.
>>>>
>>>> Please can you be more specific about the problems you see?
>>>
>>> I talked to Hernán personally and it turned out that he used the generic
>>> implementation of smp_cond_acquire (not sure if the name is correct)
>>> which uses a relaxed load followed by a barrier. In that case, replacing
>>> aob by aob;si does not change anything.
>>> Indeed, even in the reported problem we used the generic implementation
>>> (I was unaware of this), though it is easy to check that changing the
>>> relaxed load to acquire does not give sufficient orderings.
>>
>> Yes, my bad. I was using the generic header rather than the aarch64 specific
>> one and then the changes to the model were having not effect (as they
>> should).
>>
>> Now I am using the aarch64 specific ones and I can confirm dartagnan still
>> reports the violations with the current model and making the change proposed
>> by Thomas (adding ;si just to the second part seems to be enough) indeed
>> removes all violations.
> 
> That's great! Thanks for working together to get to the bottom of it. I
> was worried that this was the tip of the iceberg.
> 
> I'll try to follow-up with Arm to see if that ';si' addition is an
> acceptable to atomic-ordered before. If not, we'll absorb the smp_wmb()
> into xchg_release() with a fat comment about RCsc and mixed-size
> accesses.
> 
> Will

I have already contacted ARM, but I think it will take some time to get 
an answer.
Interestingly, the definition of aob has changed in a newer version of 
the ARM cat model:

      let aob = [Exp & M]; rmw; [Exp & M]
          | [Exp & M]; rmw; lrs; [A | Q]
          | [Imp & TTD & R]; rmw; [HU]

Ignoring all the extra Exp/Imp/TTD stuff, you can see that the second 
part of the definition only orders the load-part of the rmw with the 
following acquire load. This suggests that a form of store forwarding 
might indeed be expected, seeing that they deliberately removed the 
rmw_store-acq_load ordering they had previously.
Nevertheless, for qspinlock to work we just need the rmw_load-acq_load 
ordering that is still provided as long as we extend it with ";si".

-- 
=====================================

Thomas Haas

Technische Universität Braunschweig
Institut für Theoretische Informatik
Mühlenpfordtstr. 23, Raum IZ 343
38106 Braunschweig | Germany

t.haas@tu-braunschweig.de
https://www.tu-braunschweig.de/tcs/team/thomas-haas


  reply	other threads:[~2025-06-19 14:11 UTC|newest]

Thread overview: 23+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2025-06-12 14:55 [RFC] Potential problem in qspinlock due to mixed-size accesses Thomas Haas
2025-06-13  7:55 ` Peter Zijlstra
2025-06-13 11:17   ` Andrea Parri
     [not found]     ` <9264df13-36db-4b25-b2c4-7a9701df2f4d@tu-bs.de>
2025-06-16  6:21       ` Andrea Parri
2025-06-16 14:11         ` Alan Stern
2025-06-17 14:09           ` Andrea Parri
2025-06-19 14:27             ` Thomas Haas
2025-06-19 14:32               ` Alan Stern
2025-06-19 14:59                 ` Thomas Haas
2025-06-19 17:56                   ` Alan Stern
2025-06-19 18:21                     ` Thomas Haas
2025-06-16 14:23   ` Will Deacon
2025-06-17  6:19     ` Thomas Haas
2025-06-17  8:42       ` Hernan Ponce de Leon
2025-06-17 14:17         ` Will Deacon
2025-06-17 14:23           ` Thomas Haas
2025-06-17 19:00             ` Hernan Ponce de Leon
2025-06-19 12:30               ` Will Deacon
2025-06-19 14:11                 ` Thomas Haas [this message]
2025-06-18  6:51   ` Paul E. McKenney
2025-06-18 12:11     ` Paul E. McKenney
2025-12-17 19:05 ` Thomas Haas
2025-12-18 22:02   ` Andrea Parri

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=5e84e9fa-38a9-45fd-a67c-2db399860480@tu-bs.de \
    --to=t.haas@tu-bs.de \
    --cc=akiyks@gmail.com \
    --cc=boqun.feng@gmail.com \
    --cc=dhowells@redhat.com \
    --cc=dlustig@nvidia.com \
    --cc=hernan.poncedeleon@huaweicloud.com \
    --cc=j.alglave@ucl.ac.uk \
    --cc=joelagnelf@nvidia.com \
    --cc=jonas.oberhauser@huaweicloud.com \
    --cc=linux-arch@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=lkmm@lists.linux.dev \
    --cc=luc.maranget@inria.fr \
    --cc=npiggin@gmail.com \
    --cc=parri.andrea@gmail.com \
    --cc=paulmck@kernel.org \
    --cc=peterz@infradead.org \
    --cc=r.maseli@tu-bs.de \
    --cc=stern@rowland.harvard.edu \
    --cc=will@kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).