Date:   Mon, 29 Mar 2021 12:40:51 -0700
From:   Roman Gushchin <guro@...com>
To:     Dennis Zhou <dennis@...nel.org>
CC:     Tejun Heo <tj@...nel.org>, Christoph Lameter <cl@...ux.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        <linux-mm@...ck.org>, <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH rfc 1/4] percpu: implement partial chunk depopulation

On Mon, Mar 29, 2021 at 07:28:49PM +0000, Dennis Zhou wrote:
> On Mon, Mar 29, 2021 at 11:29:22AM -0700, Roman Gushchin wrote:
> > On Mon, Mar 29, 2021 at 05:20:55PM +0000, Dennis Zhou wrote:
> > > On Wed, Mar 24, 2021 at 12:06:23PM -0700, Roman Gushchin wrote:
> > > > This patch implements partial depopulation of percpu chunks.
> > > > 
> > > > As now, a chunk can be depopulated only as a part of the final
> > > > destruction, when there are no more outstanding allocations. However
> > > > to minimize a memory waste, it might be useful to depopulate a
> > > > partially filed chunk, if a small number of outstanding allocations
> > > > prevents the chunk from being reclaimed.
> > > > 
> > > > This patch implements the following depopulation process: it scans
> > > > over the chunk pages, looks for a range of empty and populated pages
> > > > and performs the depopulation. To avoid races with new allocations,
> > > > the chunk is previously isolated. After the depopulation the chunk is
> > > > returned to the original slot (but is appended to the tail of the list
> > > > to minimize the chances of population).
> > > > 
> > > > Because the pcpu_lock is dropped while calling pcpu_depopulate_chunk(),
> > > > the chunk can be concurrently moved to a different slot. So we need
> > > > to isolate it again on each step. pcpu_alloc_mutex is held, so the
> > > > chunk can't be populated/depopulated asynchronously.
> > > > 
> > > > Signed-off-by: Roman Gushchin <guro@...com>
> > > > ---
> > > >  mm/percpu.c | 90 +++++++++++++++++++++++++++++++++++++++++++++++++++++
> > > >  1 file changed, 90 insertions(+)
> > > > 
> > > > diff --git a/mm/percpu.c b/mm/percpu.c
> > > > index 6596a0a4286e..78c55c73fa28 100644
> > > > --- a/mm/percpu.c
> > > > +++ b/mm/percpu.c
> > > > @@ -2055,6 +2055,96 @@ static void __pcpu_balance_workfn(enum pcpu_chunk_type type)
> > > >  	mutex_unlock(&pcpu_alloc_mutex);
> > > >  }
> > > >  
> > > > +/**
> > > > + * pcpu_shrink_populated - scan chunks and release unused pages to the system
> > > > + * @type: chunk type
> > > > + *
> > > > + * Scan over all chunks, find those marked with the depopulate flag and
> > > > + * try to release unused pages to the system. On every attempt clear the
> > > > + * chunk's depopulate flag to avoid wasting CPU by scanning the same
> > > > + * chunk again and again.
> > > > + */
> > > > +static void pcpu_shrink_populated(enum pcpu_chunk_type type)
> > > > +{
> > > > +	struct list_head *pcpu_slot = pcpu_chunk_list(type);
> > > > +	struct pcpu_chunk *chunk;
> > > > +	int slot, i, off, start;
> > > > +
> > > > +	spin_lock_irq(&pcpu_lock);
> > > > +	for (slot = pcpu_nr_slots - 1; slot >= 0; slot--) {
> > > > +restart:
> > > > +		list_for_each_entry(chunk, &pcpu_slot[slot], list) {
> > > > +			bool isolated = false;
> > > > +
> > > > +			if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_HIGH)
> > > > +				break;
> > > > +
> > > 
> > > Deallocation makes me a little worried for the atomic case as now we
> > > could in theory pathologically scan deallocated chunks before finding a
> > > populated one.
> > > 
> > > I wonder if we should do something like once a chunk gets depopulated,
> > > it gets deprioritized and then only once we exhaust looking through
> > > allocated chunks we then find a depopulated chunk and add it back into
> > > the rotation. Possibly just add another set of slots? I guess it adds a
> > > few dimensions to pcpu_slots after the memcg change.
> > 
> > Please, take a look at patch 3 in the series ("percpu: on demand chunk depopulation").
> > Chunks considered to be a good target for the depopulation are in advance
> > marked with a special flag, so we'll actually try to depopulate only
> > few chunks at once. While the total number of chunks is fairly low,
> > I think it should work.
> > 
> > Another option is to link all such chunks into a list and scan over it,
> > instead of iterating over all slots.
> > 
> > Adding new dimensions to pcpu_slots is an option too, but I hope we can avoid
> > this, as it would complicate the code.
> > 
> 
> Yeah, depopulation has been on the todo list for a while. It adds the
> dimension/opportunity of bin packing by sidelining chunks and I'm
> wondering if that is the right thing to do.
> 
> Do you have a rough idea of the distribution of # of chunks you're
> seeing?

The majority of chunks are almost completely empty and reside on a one
of two last slots (before the slot for empty chunks). These are chunks
I'm mostly targeting.

> 
> > > 
> > > > +			for (i = 0, start = -1; i < chunk->nr_pages; i++) {
> > > > +				if (!chunk->nr_empty_pop_pages)
> > > > +					break;
> > > > +
> > > > +				/*
> > > > +				 * If the page is empty and populated, start or
> > > > +				 * extend the [start, i) range.
> > > > +				 */
> > > > +				if (test_bit(i, chunk->populated)) {
> > > > +					off = find_first_bit(
> > > > +						pcpu_index_alloc_map(chunk, i),
> > > > +						PCPU_BITMAP_BLOCK_BITS);
> > > > +					if (off >= PCPU_BITMAP_BLOCK_BITS) {
> > > > +						if (start == -1)
> > > > +							start = i;
> > > > +						continue;
> > > > +					}
> > > 
> > > Here instead of looking at the alloc_map, you can look at the
> > > pcpu_block_md and look for a fully free contig_hint.
> > 
> > Good idea, will try in v2.
> > 
> > > 
> > > > +				}
> > > > +
> > > > +				/*
> > > > +				 * Otherwise check if there is an active range,
> > > > +				 * and if yes, depopulate it.
> > > > +				 */
> > > > +				if (start == -1)
> > > > +					continue;
> > > > +
> > > > +				/*
> > > > +				 * Isolate the chunk, so new allocations
> > > > +				 * wouldn't be served using this chunk.
> > > > +				 * Async releases can still happen.
> > > > +				 */
> > > > +				if (!list_empty(&chunk->list)) {
> > > > +					list_del_init(&chunk->list);
> > > > +					isolated = true;
> > > 
> > > Maybe when freeing a chunk, we should consider just isolating it period
> > > and preventing pcpu_free_area() from being able to add the chunk back
> > > to a pcpu_slot.
> > 
> > You mean to add a check in pcpu_free_area() if the chunks is isolated?
> > Yeah, sounds good to me, will do in v2.
> > 
> 
> Could also be done in pcpu_chunk_relocate() so it's clear an isolated
> chunk shouldn't be moved. But I think pcpu_free_area() should be the
> only way the chunk can be moved on the list.

Yeah, I thought about putting it into pcpu_chunk_relocate(), but it's
used on new chunks, so it would require more changes, so I'm not sure.

> 
> > Thank you!
> > 
> > > 
> > > > +				}
> > > > +
> > > > +				spin_unlock_irq(&pcpu_lock);
> > > > +				pcpu_depopulate_chunk(chunk, start, i);
> > > > +				cond_resched();
> > > > +				spin_lock_irq(&pcpu_lock);
> > > > +
> > > > +				pcpu_chunk_depopulated(chunk, start, i);
> > > > +
> > > > +				/*
> > > > +				 * Reset the range and continue.
> > > > +				 */
> > > > +				start = -1;
> > > > +			}
> > > > +
> > > > +			if (isolated) {
> > > > +				/*
> > > > +				 * The chunk could have been moved while
> > > > +				 * pcpu_lock wasn't held. Make sure we put
> > > > +				 * the chunk back into the slot and restart
> > > > +				 * the scanning.
> > > > +				 */
> > > > +				if (list_empty(&chunk->list))
> > > > +					list_add_tail(&chunk->list,
> > > > +						      &pcpu_slot[slot]);
> > > > +				goto restart;
> > > > +			}
> > > > +		}
> > > > +	}
> > > > +	spin_unlock_irq(&pcpu_lock);
> > > > +}
> > > > +
> > > >  /**
> > > >   * pcpu_balance_workfn - manage the amount of free chunks and populated pages
> > > >   * @work: unused
> > > > -- 
> > > > 2.30.2
> > > > 

Powered by blists - more mailing lists