Re: [RFC PATCH] mm/madvise: remove redundant mmap_lock operations from process_madvise()

[SeongJae Park]

On Tue, 14 Jan 2025 22:44:53 -0500 "Liam R. Howlett" <Liam.Howlett@xxxxxxxxxx> wrote:

> * Shakeel Butt <shakeel.butt@xxxxxxxxx> [250114 13:14]:
> > Ccing relevant folks.
> 
> Thanks Shakeel.
> 
> > 
> > On Fri, Jan 10, 2025 at 04:46:18PM -0800, SeongJae Park wrote:
> > > process_madvise() calls do_madvise() for each address range.  Then, each
> > > do_madvise() invocation holds and releases same mmap_lock.  Optimize the
> > > redundant lock operations by doing the locking in process_madvise(), and
> > > inform do_madvise() that the lock is already held and therefore can be
> > > skipped.
> > > 
> > > Evaluation
> > > ==========
> > > 
> > > I measured the time to apply MADV_DONTNEED advice to 256 MiB memory
> > > using multiple madvise() calls, 4 KiB per each call.  I also do the same
> > > with process_madvise(), but with varying iovec size from 1 to 1024.
> > > The source code for the measurement is available at GitHub[1].
> > > 
> > > The measurement results are as below.  'sz_batches' column shows the
> > > iovec size of process_madvise() calls.  '0' is for madvise() calls case.
> > > 'before' and 'after' columns are the measured time to apply
> > > MADV_DONTNEED to the 256 MiB memory buffer in nanoseconds, on kernels
> > > that built without and with this patch, respectively.  So lower value
> > > means better efficiency.  'after/before' column is the ratio of 'after'
> > > to 'before'.
> > > 
> > >     sz_batches  before     after      after/before
> > >     0           124062365  96670188   0.779206393494111
> > >     1           136341258  113915688  0.835518827323714
> > >     2           105314942  78898211   0.749164453796119
> > >     4           82012858   59778998   0.728897875989153
> > >     8           82562651   51003069   0.617749895167489
> > >     16          71474930   47575960   0.665631431888076
> > >     32          71391211   42902076   0.600943385033768
> > >     64          68225932   41337835   0.605896230190011
> > >     128         71053578   42467240   0.597679120395598
> > >     256         85094126   41630463   0.489228398679364
> > >     512         68531628   44049763   0.6427654542221
> > >     1024        79338892   43370866   0.546653285755491
> > > 
> > > The measurement shows this patch reduces the process_madvise() latency,
> > > proportional to the batching size, from about 25% with the batch size 2
> > > to about 55% with the batch size 1,024.  The trend is somewhat we can
> > > expect.
> > > 
> > > Interestingly, this patch has also optimize madvise() and single batch
> > > size process_madvise(), though.  I ran this test multiple times, but the
> > > results are consistent.  I'm still investigating if there are something
> > > I'm missing.  But I believe the investigation may not necessarily be a
> > > blocker of this RFC, so just posting this.  I will add updates of the
> > > madvise() and single batch size process_madvise() investigation later.
> > > 
> > > [1] https://github.com/sjp38/eval_proc_madvise
> > > 
> > > Signed-off-by: SeongJae Park <sj@xxxxxxxxxx>
> > > ---
> > >  include/linux/mm.h |  3 ++-
> > >  io_uring/advise.c  |  2 +-
> > >  mm/damon/vaddr.c   |  2 +-
> > >  mm/madvise.c       | 54 +++++++++++++++++++++++++++++++++++-----------
> > >  4 files changed, 45 insertions(+), 16 deletions(-)
> > > 
> > > diff --git a/include/linux/mm.h b/include/linux/mm.h
> > > index 612b513ebfbd..e3ca5967ebd4 100644
> > > --- a/include/linux/mm.h
> > > +++ b/include/linux/mm.h
> > > @@ -3459,7 +3459,8 @@ int do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
> > >  		    unsigned long end, struct list_head *uf, bool unlock);
> > >  extern int do_munmap(struct mm_struct *, unsigned long, size_t,
> > >  		     struct list_head *uf);
> > > -extern int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior);
> > > +extern int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in,
> > > +		int behavior, bool lock_held);
> 
> We are dropping externs when it is not needed as things are changed.

Good point.  I will drop it if I result in changing something here in next
versions.

> 
> Also, please don't use a flags for this.  It will have a single user of
> true, probably ever.

Ok, that sounds fair.

> 
> It might be better to break do_madvise up into more parts:
> 1. is_madvise_valid(), which does the checking.
> 2. madivse_do_behavior()
> 
> The locking type is already extracted to madivse_need_mmap_write().
> 
> What do you think?

Sounds good to me :)

I will make the v2 of this patch following your suggestion after waiting for
any possible more comments a bit.


Thanks,
SJ

[...]