Re: [PATCH 2/4] mm: mempool: introduce page bulk allocator

From: Brian Foster
Date: Fri Oct 14 2022 - 08:04:00 EST


On Thu, Oct 06, 2022 at 11:43:21AM -0700, Yang Shi wrote:
> On Thu, Oct 6, 2022 at 7:47 AM Brian Foster <bfoster@xxxxxxxxxx> wrote:
> >
> > On Wed, Oct 05, 2022 at 11:03:39AM -0700, Yang Shi wrote:
> > > Since v5.13 the page bulk allocator was introduced to allocate order-0
> > > pages in bulk. There are a few mempool allocator callers which does
> > > order-0 page allocation in a loop, for example, dm-crypt, f2fs compress,
> > > etc. A mempool page bulk allocator seems useful. So introduce the
> > > mempool page bulk allocator.
> > >
> > > It introduces the below APIs:
> > > - mempool_init_pages_bulk()
> > > - mempool_create_pages_bulk()
> > > They initialize the mempool for page bulk allocator. The pool is filled
> > > by alloc_page() in a loop.
> > >
> > > - mempool_alloc_pages_bulk_list()
> > > - mempool_alloc_pages_bulk_array()
> > > They do bulk allocation from mempool.
> > > They do the below conceptually:
> > > 1. Call bulk page allocator
> > > 2. If the allocation is fulfilled then return otherwise try to
> > > allocate the remaining pages from the mempool
> > > 3. If it is fulfilled then return otherwise retry from #1 with sleepable
> > > gfp
> > > 4. If it is still failed, sleep for a while to wait for the mempool is
> > > refilled, then retry from #1
> > > The populated pages will stay on the list or array until the callers
> > > consume them or free them.
> > > Since mempool allocator is guaranteed to success in the sleepable context,
> > > so the two APIs return true for success or false for fail. It is the
> > > caller's responsibility to handle failure case (partial allocation), just
> > > like the page bulk allocator.
> > >
> > > The mempool typically is an object agnostic allocator, but bulk allocation
> > > is only supported by pages, so the mempool bulk allocator is for page
> > > allocation only as well.
> > >
> > > Signed-off-by: Yang Shi <shy828301@xxxxxxxxx>
> > > ---
> >
> > Hi Yang,
> >
> > I'm not terribly familiar with either component so I'm probably missing
> > context/details, but just a couple high level thoughts when reading your
> > patches...
> >
> > > include/linux/mempool.h | 19 ++++
> > > mm/mempool.c | 188 +++++++++++++++++++++++++++++++++++++---
> > > 2 files changed, 197 insertions(+), 10 deletions(-)
> > >
> > ...
> > > diff --git a/mm/mempool.c b/mm/mempool.c
> > > index ba32151f3843..7711ca2e6d66 100644
> > > --- a/mm/mempool.c
> > > +++ b/mm/mempool.c
> > > @@ -177,6 +177,7 @@ void mempool_destroy(mempool_t *pool)
> > > EXPORT_SYMBOL(mempool_destroy);
> > >
> > > static inline int __mempool_init(mempool_t *pool, int min_nr,
> > > + mempool_alloc_pages_bulk_t *alloc_pages_bulk_fn,
> > > mempool_alloc_t *alloc_fn,
> > > mempool_free_t *free_fn, void *pool_data,
> > > gfp_t gfp_mask, int node_id)
> > > @@ -186,8 +187,11 @@ static inline int __mempool_init(mempool_t *pool, int min_nr,
> > > pool->pool_data = pool_data;
> > > pool->alloc = alloc_fn;
> > > pool->free = free_fn;
> > > + pool->alloc_pages_bulk = alloc_pages_bulk_fn;
> > > init_waitqueue_head(&pool->wait);
> > >
> > > + WARN_ON_ONCE(alloc_pages_bulk_fn && alloc_fn);
> > > +
> > > pool->elements = kmalloc_array_node(min_nr, sizeof(void *),
> > > gfp_mask, node_id);
> > > if (!pool->elements)
> > > @@ -199,7 +203,10 @@ static inline int __mempool_init(mempool_t *pool, int min_nr,
> > > while (pool->curr_nr < pool->min_nr) {
> > > void *element;
> > >
> > > - element = pool->alloc(gfp_mask, pool->pool_data);
> > > + if (pool->alloc_pages_bulk)
> > > + element = alloc_page(gfp_mask);
> >
> > Any reason to not use the callback from the caller for the bulk variant
> > here? It looks like some users might expect consistency between the
> > alloc / free callbacks for the pool. I.e., I'm not familiar with
> > dm-crypt, but the code modified in the subsequent patches looks like it
> > keeps an allocated page count. Will that still work with this, assuming
> > these pages are freed through free_fn?
>
> No special reason, this implementation just end up with fewer code
> otherwise we should need to define a list, and manipulate the list,
> seems like a little bit overkilling for initialization code.
>
> Yes, that allocated page count works, just the pages in the pool are
> not counted in the count anymore, 256 pages should be not a big deal
> IMHO.
>

Ok. I defer to dm-crypt folks on whether/how much it might care about
pages being hidden from the accounting. My concern was partly that, but
also partly whether it was possible to break consistency between the
number of alloc and free callbacks to be expected. For example, wouldn't
these counters underflow if the mempool is torn down or shrunk (via
mempool_resize()), and thus the caller gets ->free() callbacks for pages
it never accounted for in the first place?

Brian

> >
> > > + else
> > > + element = pool->alloc(gfp_mask, pool->pool_data);
> > > if (unlikely(!element)) {
> > > mempool_exit(pool);
> > > return -ENOMEM;
> > ...
> > > @@ -457,6 +499,132 @@ void *mempool_alloc(mempool_t *pool, gfp_t gfp_mask)
> > > }
> > > EXPORT_SYMBOL(mempool_alloc);
> > >
> > > +/**
> > > + * mempool_alloc_pages_bulk - allocate a bulk of pagesfrom a specific
> > > + * memory pool
> > > + * @pool: pointer to the memory pool which was allocated via
> > > + * mempool_create().
> > > + * @gfp_mask: the usual allocation bitmask.
> > > + * @nr: the number of requested pages.
> > > + * @page_list: the list the pages will be added to.
> > > + * @page_array: the array the pages will be added to.
> > > + *
> > > + * this function only sleeps if the alloc_pages_bulk_fn() function sleeps
> > > + * or the allocation can not be satisfied even though the mempool is depleted.
> > > + * Note that due to preallocation, this function *never* fails when called
> > > + * from process contexts. (it might fail if called from an IRQ context.)
> > > + * Note: using __GFP_ZERO is not supported. And the caller should not pass
> > > + * in both valid page_list and page_array.
> > > + *
> > > + * Return: true when nr pages are allocated or false if not. It is the
> > > + * caller's responsibility to free the partial allocated pages.
> > > + */
> > > +static bool mempool_alloc_pages_bulk(mempool_t *pool, gfp_t gfp_mask,
> > > + unsigned int nr,
> > > + struct list_head *page_list,
> > > + struct page **page_array)
> > > +{
> > > + unsigned long flags;
> > > + wait_queue_entry_t wait;
> > > + gfp_t gfp_temp;
> > > + int i;
> > > + unsigned int ret, nr_remaining;
> > > + struct page *page;
> > > +
> >
> > This looks like a lot of duplicate boilerplate from mempool_alloc().
> > Could this instead do something like: rename the former to
> > __mempool_alloc() and add a count parameter, implement bulk alloc
> > support in there for count > 1, then let traditional (i.e., non-bulk)
> > mempool_alloc() callers pass a count of 1?
>
> Thanks for the suggestion. Yeah, the duplicate code is not perfect. I
> thought about this way too, but it may need to have a lot of "if
> (count > 0)" of "if (is_bulk_alloc) " if a flag is used in the code to
> handle the bulk allocation, for example, calculate remaining nr, loop
> to remove element from the pool, manipulate list or array, etc. Seems
> not that readable IMHO.
>
> We may be able to extract some common code into shared helpers, for
> example, the gfp sanitization and wait logic.
>
> >
> > Along the same lines, I also wonder if there's any value in generic bulk
> > alloc support for mempool. For example, I suppose technically this could
> > be implemented via one change to support a pool->alloc_bulk() callback
> > that any user could implement via a loop if they wanted
> > mempool_alloc_bulk() support backed by a preallocated pool. The page
> > based user could then just use that to call alloc_pages_bulk_*() as an
> > optimization without the mempool layer needing to know or care about
> > whether the underlying elements are pages or not. Hm?
>
> Thanks for the suggestion. Actually I thought about this too. But the
> memory space overhead, particularly stack space seems like a
> showstopper to me. We just can put the pointers into an array, but
> this may consume a significant amount of stack memory. One pointer is
> 8 bytes, 256 objects imply 2K stack space. Of course the users could
> move the array into a dynamic allocated data structure, but this may
> need the users modify their driver. Bulk kmalloc via kmalloc_array()
> may be fine, this is the only usercase other than pages I could think
> of.
>
> >
> > Brian
> >
> > > + VM_WARN_ON_ONCE(gfp_mask & __GFP_ZERO);
> > > + might_alloc(gfp_mask);
> > > +
> > > + gfp_mask |= __GFP_NOMEMALLOC; /* don't allocate emergency reserves */
> > > + gfp_mask |= __GFP_NORETRY; /* don't loop in __alloc_pages */
> > > + gfp_mask |= __GFP_NOWARN; /* failures are OK */
> > > +
> > > + gfp_temp = gfp_mask & ~(__GFP_DIRECT_RECLAIM|__GFP_IO);
> > > +
> > > +repeat_alloc:
> > > + i = 0;
> > > + ret = pool->alloc_pages_bulk(gfp_temp, nr, pool->pool_data, page_list,
> > > + page_array);
> > > +
> > > + if (ret == nr)
> > > + return true;
> > > +
> > > + nr_remaining = nr - ret;
> > > +
> > > + spin_lock_irqsave(&pool->lock, flags);
> > > + /* Allocate page from the pool and add to the list or array */
> > > + while (pool->curr_nr && (nr_remaining > 0)) {
> > > + page = remove_element(pool);
> > > + spin_unlock_irqrestore(&pool->lock, flags);
> > > + smp_wmb();
> > > +
> > > + kmemleak_update_trace((void *)page);
> > > +
> > > + if (page_list)
> > > + list_add(&page->lru, page_list);
> > > + else
> > > + page_array[ret + i] = page;
> > > +
> > > + i++;
> > > + nr_remaining--;
> > > +
> > > + spin_lock_irqsave(&pool->lock, flags);
> > > + }
> > > +
> > > + spin_unlock_irqrestore(&pool->lock, flags);
> > > +
> > > + if (!nr_remaining)
> > > + return true;
> > > +
> > > + /*
> > > + * The bulk allocator counts in the populated pages for array,
> > > + * but don't do it for list.
> > > + */
> > > + if (page_list)
> > > + nr = nr_remaining;
> > > +
> > > + /*
> > > + * We use gfp mask w/o direct reclaim or IO for the first round. If
> > > + * alloc failed with that and @pool was empty, retry immediately.
> > > + */
> > > + if (gfp_temp != gfp_mask) {
> > > + gfp_temp = gfp_mask;
> > > + goto repeat_alloc;
> > > + }
> > > +
> > > + /* We must not sleep if !__GFP_DIRECT_RECLAIM */
> > > + if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
> > > + return false;
> > > +
> > > + /* Let's wait for someone else to return an element to @pool */
> > > + init_wait(&wait);
> > > + prepare_to_wait(&pool->wait, &wait, TASK_UNINTERRUPTIBLE);
> > > +
> > > + /*
> > > + * FIXME: this should be io_schedule(). The timeout is there as a
> > > + * workaround for some DM problems in 2.6.18.
> > > + */
> > > + io_schedule_timeout(5*HZ);
> > > +
> > > + finish_wait(&pool->wait, &wait);
> > > + goto repeat_alloc;
> > > +}
> > > +
> > > +bool mempool_alloc_pages_bulk_list(mempool_t *pool, gfp_t gfp_mask,
> > > + unsigned int nr,
> > > + struct list_head *page_list)
> > > +{
> > > + return mempool_alloc_pages_bulk(pool, gfp_mask, nr, page_list, NULL);
> > > +}
> > > +EXPORT_SYMBOL(mempool_alloc_pages_bulk_list);
> > > +
> > > +bool mempool_alloc_pages_bulk_array(mempool_t *pool, gfp_t gfp_mask,
> > > + unsigned int nr,
> > > + struct page **page_array)
> > > +{
> > > + return mempool_alloc_pages_bulk(pool, gfp_mask, nr, NULL, page_array);
> > > +}
> > > +EXPORT_SYMBOL(mempool_alloc_pages_bulk_array);
> > > +
> > > /**
> > > * mempool_free - return an element to the pool.
> > > * @element: pool element pointer.
> > > --
> > > 2.26.3
> > >
> >
>