Message-ID: <f7ed3125-e7ce-45b0-a077-a2979a6238b3@suse.cz>
Date: Tue, 20 Jan 2026 22:28:15 +0100
From: Vlastimil Babka <vbabka@...e.cz>
To: Lorenzo Stoakes <lorenzo.stoakes@...cle.com>
Cc: Andrew Morton <akpm@...ux-foundation.org>,
 David Hildenbrand <david@...nel.org>,
 "Liam R . Howlett" <Liam.Howlett@...cle.com>, Mike Rapoport
 <rppt@...nel.org>, Suren Baghdasaryan <surenb@...gle.com>,
 Michal Hocko <mhocko@...e.com>, Shakeel Butt <shakeel.butt@...ux.dev>,
 Jann Horn <jannh@...gle.com>, linux-mm@...ck.org,
 linux-kernel@...r.kernel.org, linux-rt-devel@...ts.linux.dev,
 Peter Zijlstra <peterz@...radead.org>, Ingo Molnar <mingo@...hat.com>,
 Will Deacon <will@...nel.org>, Boqun Feng <boqun.feng@...il.com>,
 Waiman Long <longman@...hat.com>,
 Sebastian Andrzej Siewior <bigeasy@...utronix.de>,
 Clark Williams <clrkwllms@...nel.org>, Steven Rostedt <rostedt@...dmis.org>
Subject: Re: [PATCH v2 1/2] mm/vma: use lockdep where we can, reduce
 duplication

On 1/20/26 18:49, Lorenzo Stoakes wrote:
> On Tue, Jan 20, 2026 at 02:53:30PM +0100, Vlastimil Babka wrote:
>> On 1/19/26 21:59, Lorenzo Stoakes wrote:
>> > We introduce vma_is_read_locked(), which must deal with the case in which
>> > VMA write lock sets refcnt to VMA_LOCK_OFFSET or VMA_LOCK_OFFSET +
>> > 1. Luckily is_vma_writer_only() already exists which we can use to check
>> > this.
>>
>> So I think there's a bit of a caveat in that
>>
>> - is_vma_writer_only() may be a false positive if there is a temporary
>> reader of a detached vma (per comments in vma_mark_detached() and
>> vma_mark_detached())
> 
> vma_mark_detached() and vma_mark_attached() I sasume you mean.

Right.

> OK so this is all very confusing indeed.
> 
> This function is _only_ referring to the situation between
> __vma_enter_locked() and __vma_exit_locked().
> 
> Despite their names, suggesting maybe they happen on lock and unlock
> respectively, that's not the case, they're both invoked on lock and enter
> on start of TAKING lock and exit on completion of TAKING the
> lock.

IIUC yes, for the vma "write lock".

> It seems __vma_enter_locked() is more about getting into this state with
> refcnt equal to VMA_LOCK_OFFSET (detaching - note elsewhere we say detached
> of course) or VMA_LOCK_OFFSET + 1 if attached and waiting on readers who
> are spuriously increasing the reference count.

Yes.

> So fundamentally is_vma_writer_only() is actually asking 'are we in the
> midst of a VMA write lock acquisiton having finally set the VMA's refcnt to
> VMA_LOCK_OFFSET or VMA_LOCK_OFFSET+1 but haven't yet completed acquiring
> the lock' - e.g. having not yet called __vma_exit_locked().

IIUC in the current code it's not used in that "are *we* in the midst..."
sense but "is there a writer in that phase that we are supposed to wake up
because we are the last reader", where a rare false positive answer only
results in an unnecessary wakeup of said wanna-be writer, but nothing worse.

And AFAIU this patch tries to reuse the function to ask "is the vma read
locked?" (and we presume it's by us).

> With __vma_enter_locked() called from:
> 
> vma_start_write() / vma_start_write_killable()
> -> __vma_start_write()
> -> __vma_enter_locked()
> 
> vma_mark_detached()
> -> __vma_enter_locked()
> 
> OK so in __vma_enter_locked() we add VMA_LOCK_OFFSET but then wait until we
> get to either VMA_LOCK_OFFSET + 1 (attached) or VMA_LOCK_OFFSET (detached),
> since presumably refcnt == 0 is detached, refcnt == 1 means write lock
> finally acquired (but you have to check the sequence number).
> 
> And _there_ we could have spurious readers.

Yes.

> 
>>
>> - hence vma_is_read_locked() may be a false negative
> 
> Yup.
> 
>>
>> - hence vma_assert_locked() might assume wrongly that we should not assert
>> being a reader, so we vma_assert_write_locked() instead, and fail
> 
> Aside ->
> 
> 	Every time I come to this code it's like this - having to refresh
> 	my memory as to how any of it works, getting confused, etc.
> 
> 	This speaks to this being a broken abstraction similar to anon_vma.
> 
> 	What I mean by leaked abstraction is that you seem to need to
> 	maintain the _implementation_ context in your head to be able to
> 	correctly implement anything. We simply are not abstracting details
> 	here really well at all.

I think this would be true if it was applied to users of the high-level API
of the code - actual locking and unlocking for read/write. Do they have to
care about the implementation details? Hopefully not.

Here we have to think about the implementation because we are trying to
improve the API (to add assertions) so that's not surprising? If your
complaing is about an "intermediate" abstractions level like the
"is_vma_writer_only()" function then yeah it's far from perfect.

> 	The fact I got this wrong despite staring at this code for ages is
> 	indicative of that.
> 
> 	Also the fact an ostensibly simple series has turned into a
> 	'restore the context' discussion this long and taken this many
> 	hours is further suggestive.
> 
> 	I think we can do better. I'd rather not do more 'cleanup' series,
> 	but I think this badly needs it.
> 
> 	So maybe I'll convert this series into something that addresses
> 	some of this stuff.
> 
> <- Aside
> 
> 
>>
>> Howevever the above should mean it could be only us who is the temporary
>> reader. And we are not going to use vma_assert_locked() during the temporary
>> reader part (in vma_start_read()).
> 
> I don't think so? Spurious readers can arise at any time incrementing the
> refcnt via vma_start_read(), so the temporary readers could be anybody.
> 
> But they'd not get the read lock, and so shouldn't call anything asserting
> read lock.
> 
> Anyway I think my use of is_vma_writer_only() is just broken then.

AFAIU the whole thing (vma_assert_locked() after this patch) would be broken
in a case where we are really a reader and vma_is_read_locked() returns
false wrongly, and thus makes us perform vma_assert_write_locked() and fail.
So the scenarios with spurious readers can't cause that I think.

What I think could cause that is there being a writer (not us), causing
is_vma_writer_only() return true even when there's also a reader (us). And I
concluded that could happen only in case where we would be the spurious
reader racing with a detaching writer. But when we are in the temporary
spurious reader situation, we don't perform vma_is_read_locked() there.

> We _do_ need to account for the VMA write lock scenario here, but I think
> instead we should be good with refcnt > 1 && refcnt < VMA_LOCK_OFFSET no?

That check would tell us there is no writer. But there might be a writer
(not us) while we're a reader, and thus that check won't work as a signal
for "we must have the read lock"?

>> So it's probably fine, but maybe worth some comments to prevent people
>> getting suspicious and reconstructing this?
> 
> But yeah we shouldn't be asserting this anywhere during which this should
> be the case.
> 
> So hopefully the above resolves the issue?

I don't follow, but perhaps I misunderstood you above and it's late here now.

> It's still racey (write lock might have been acquired since we checked but
> that's just the nature of it.
> 
> But if we use lockdep as you mention we can actually do the same 'precise
> if lockdep, otherwise not so precise' approach in the stabilised check.
> 
> Let me respin.
>
>>
>> But I think perhaps also vma_assert_locked() could, with lockdep enabled
>> (similarly to vma_assert_stabilised() in patch 2), use the
>> "lock_is_held(&vma->vmlock_dep_map)" condition (without immediately
>> asserting it) for the primary reader vs writer decision, and not rely on
>> vma_is_read_locked()? Because lockdep has the precise information.
>>
>> It would likely make things more ugly, or require more refactoring, but
>> hopefully worthwhile?
> 
> Yeah good idea. Will do.

Ack, thanks.

> Maybe I need to make this into a broader refactoring series. Because this
> so badly needs it.

Yep :/

> Cheers, Lorenzo

Powered by blists - more mailing lists