Re: [RFC 0/2] rwsem: introduce upgrade_read interface

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Oct 16, 2024 at 04:53:45PM +0800, lizhe.67@bytedance.com wrote:
> On Wed, 16 Oct 2024 10:09:55 +0200, peterz@infradead.org wrote:
> 
> > On Wed, Oct 16, 2024 at 12:35:58PM +0800, lizhe.67@bytedance.com wrote:
> > > From: Li Zhe <lizhe.67@bytedance.com>
> > > 
> > > In the current kernel rwsem implementation, there is an interface to
> > > downgrade write lock to read lock, but there is no interface to upgrade
> > > a read lock to write lock. This means that in order to acquire write
> > > lock while holding read lock, we have to release the read lock first and
> > > then acquire the write lock, which will introduce some troubles in
> > > concurrent programming. This patch set provides the 'upgrade_read' interface
> > > to solve this problem. This interface can change a read lock to a write
> > > lock.
> > 
> > upgrade-read is fundamentally prone to deadlocks. Imagine two concurrent
> > invocations, each waiting for all readers to go away before proceeding
> > to upgrade to a writer.
> >
> > Any solution to fixing that will end up being semantically similar to
> > dropping the read lock and acquiring a write lock -- there will not be a
> > single continuous critical section.
> 
> According to the implementation of this patch, one of the invocation will

Since the premise as described here is utter nonsense, I didn't get to
actually reading the implementation -- why continue to waste time etc.

> get '-EBUSY' in this case. If -EBUSY is obtained and the invocation thread
> continues to retry instead of dropping the read lock and acquiring a write lock,
> it may cause problems.

Failure should drop the read lock, otherwise it is too easy to mess
things up.

> Of course, this patchset only try it's best to achieve a
> single continuous critical section as much as possible, and there is no guarantee.

As already stated, nothing like that was mentioned.

> > As such, this interface makes no sense.
> 
> This interface is just trying to reduce the overhead caused by the
> additional checks, which is caused by non-continuous critical
> sections, as much as possible.  Rather than eliminating it in all
> scenarios. So would it be better to change the error code to something
> else? So that the caller will not retry this interface?

You fail to quantify the gains. How am I supposed to know if the
(significant?) increase in complexity is worth it?

Why should I accept this increase in complexity for the sake of
khugepaged, something which I care very little about?