Message-ID: <871qk582tn.fsf@toke.dk>
Date:   Thu, 27 Apr 2023 22:53:08 +0200
From:   Toke Høiland-Jørgensen <toke@...hat.com>
To:     Jesper Dangaard Brouer <brouer@...hat.com>,
        Ilias Apalodimas <ilias.apalodimas@...aro.org>,
        netdev@...r.kernel.org, Eric Dumazet <eric.dumazet@...il.com>,
        linux-mm@...ck.org, Mel Gorman <mgorman@...hsingularity.net>
Cc:     Jesper Dangaard Brouer <brouer@...hat.com>, lorenzo@...nel.org,
        linyunsheng@...wei.com, bpf@...r.kernel.org,
        "David S. Miller" <davem@...emloft.net>,
        Jakub Kicinski <kuba@...nel.org>,
        Paolo Abeni <pabeni@...hat.com>,
        Andrew Morton <akpm@...ux-foundation.org>, willy@...radead.org
Subject: Re: [PATCH RFC net-next/mm V2 1/2] page_pool: Remove workqueue in
 new shutdown scheme

Jesper Dangaard Brouer <brouer@...hat.com> writes:

> This removes the workqueue scheme that periodically tests when
> inflight reach zero such that page_pool memory can be freed.
>
> This change adds code to fast-path free checking for a shutdown flags
> bit after returning PP pages.

I think the general approach is workable, but spotted a few issues with
the details, see below.

> Performance is very important for PP, as the fast path is used for
> XDP_DROP use-cases where NIC drivers recycle PP pages directly into PP
> alloc cache.
>
> The goal were that this code change should have zero impact on this
> fast-path. The slight code reorg of likely() are deliberate. Micro
> benchmarking done via kernel module[1] on x86_64, shows this code
> change only cost a single instruction extra (approx 0.3 nanosec on CPU
> E5-1650 @3.60GHz).
>
> It is possible to make this code zero impact via static_key, but that
> change is not considered worth the complexity.
>
> [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/lib/bench_page_pool_simple.c
>
> Signed-off-by: Jesper Dangaard Brouer <brouer@...hat.com>
> ---
>  include/net/page_pool.h |    9 ++--
>  net/core/page_pool.c    |  100 +++++++++++++++++++++++++++++------------------
>  2 files changed, 66 insertions(+), 43 deletions(-)
>
> diff --git a/include/net/page_pool.h b/include/net/page_pool.h
> index c8ec2f34722b..a71c0f2695b0 100644
> --- a/include/net/page_pool.h
> +++ b/include/net/page_pool.h
> @@ -50,6 +50,9 @@
>  				 PP_FLAG_DMA_SYNC_DEV |\
>  				 PP_FLAG_PAGE_FRAG)
>  
> +/* Internal flag: PP in shutdown phase, waiting for inflight pages */
> +#define PP_FLAG_SHUTDOWN	BIT(8)
> +
>  /*
>   * Fast allocation side cache array/stack
>   *
> @@ -151,11 +154,6 @@ static inline u64 *page_pool_ethtool_stats_get(u64 *data, void *stats)
>  struct page_pool {
>  	struct page_pool_params p;
>  
> -	struct delayed_work release_dw;
> -	void (*disconnect)(void *);
> -	unsigned long defer_start;
> -	unsigned long defer_warn;
> -
>  	u32 pages_state_hold_cnt;
>  	unsigned int frag_offset;
>  	struct page *frag_page;
> @@ -165,6 +163,7 @@ struct page_pool {
>  	/* these stats are incremented while in softirq context */
>  	struct page_pool_alloc_stats alloc_stats;
>  #endif
> +	void (*disconnect)(void *);
>  	u32 xdp_mem_id;
>  
>  	/*
> diff --git a/net/core/page_pool.c b/net/core/page_pool.c
> index e212e9d7edcb..b8359d84e30f 100644
> --- a/net/core/page_pool.c
> +++ b/net/core/page_pool.c
> @@ -23,9 +23,6 @@
>  
>  #include <trace/events/page_pool.h>
>  
> -#define DEFER_TIME (msecs_to_jiffies(1000))
> -#define DEFER_WARN_INTERVAL (60 * HZ)
> -
>  #define BIAS_MAX	LONG_MAX
>  
>  #ifdef CONFIG_PAGE_POOL_STATS
> @@ -380,6 +377,10 @@ static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
>  	struct page *page;
>  	int i, nr_pages;
>  
> +	/* API usage BUG: PP in shutdown phase, cannot alloc new pages */
> +	if (WARN_ON(pool->p.flags & PP_FLAG_SHUTDOWN))
> +		return NULL;
> +
>  	/* Don't support bulk alloc for high-order pages */
>  	if (unlikely(pp_order))
>  		return __page_pool_alloc_page_order(pool, gfp);
> @@ -445,15 +446,20 @@ struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
>  }
>  EXPORT_SYMBOL(page_pool_alloc_pages);
>  
> +/* Avoid inlining code to avoid speculative fetching cacheline */
> +noinline u32 pp_read_hold_cnt(struct page_pool *pool)
> +{
> +	return READ_ONCE(pool->pages_state_hold_cnt);
> +}
> +
>  /* Calculate distance between two u32 values, valid if distance is below 2^(31)
>   *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
>   */
>  #define _distance(a, b)	(s32)((a) - (b))
>  
> -static s32 page_pool_inflight(struct page_pool *pool)
> +static s32 __page_pool_inflight(struct page_pool *pool,
> +				u32 hold_cnt, u32 release_cnt)
>  {
> -	u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
> -	u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
>  	s32 inflight;
>  
>  	inflight = _distance(hold_cnt, release_cnt);
> @@ -464,6 +470,16 @@ static s32 page_pool_inflight(struct page_pool *pool)
>  	return inflight;
>  }
>  
> +static s32 page_pool_inflight(struct page_pool *pool)
> +{
> +	u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
> +	u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
> +	return __page_pool_inflight(pool, hold_cnt, release_cnt);
> +}
> +
> +static int page_pool_free_attempt(struct page_pool *pool,
> +				  u32 hold_cnt, u32 release_cnt);
> +
>  /* Disconnects a page (from a page_pool).  API users can have a need
>   * to disconnect a page (from a page_pool), to allow it to be used as
>   * a regular page (that will eventually be returned to the normal
> @@ -471,8 +487,10 @@ static s32 page_pool_inflight(struct page_pool *pool)
>   */
>  void page_pool_release_page(struct page_pool *pool, struct page *page)
>  {
> +	unsigned int flags = READ_ONCE(pool->p.flags);
>  	dma_addr_t dma;
> -	int count;
> +	u32 release_cnt;
> +	u32 hold_cnt;
>  
>  	if (!(pool->p.flags & PP_FLAG_DMA_MAP))
>  		/* Always account for inflight pages, even if we didn't
> @@ -490,11 +508,15 @@ void page_pool_release_page(struct page_pool *pool, struct page *page)
>  skip_dma_unmap:
>  	page_pool_clear_pp_info(page);
>  
> -	/* This may be the last page returned, releasing the pool, so
> -	 * it is not safe to reference pool afterwards.
> -	 */
> -	count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
> -	trace_page_pool_state_release(pool, page, count);
> +	if (flags & PP_FLAG_SHUTDOWN)
> +		hold_cnt = pp_read_hold_cnt(pool);
> +
> +	release_cnt = atomic_inc_return(&pool->pages_state_release_cnt);
> +	trace_page_pool_state_release(pool, page, release_cnt);
> +
> +	/* In shutdown phase, last page will free pool instance */
> +	if (flags & PP_FLAG_SHUTDOWN)
> +		page_pool_free_attempt(pool, hold_cnt, release_cnt);

Since the assumption is that no new pages will be allocated once the
PP_FLAG_SHUTDOWN is set (i.e., hold_count can not increase in the case),
I don't think it matters what order you read the hold and release counts
in? So you could simplify the above to just:

> +	if (flags & PP_FLAG_SHUTDOWN)
> +		page_pool_free_attempt(pool, pp_read_hold_cnt(pool), release_cnt);

and drop the second check of the flag further up?

You could probably even lose the hold_cnt argument entirely from
page_pool_free_attempt() and just have it call pp_read_hold_cnt() directly?

>  }
>  EXPORT_SYMBOL(page_pool_release_page);
>  
> @@ -535,7 +557,7 @@ static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
>  static bool page_pool_recycle_in_cache(struct page *page,
>  				       struct page_pool *pool)
>  {
> -	if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) {
> +	if (pool->alloc.count == PP_ALLOC_CACHE_SIZE) {
>  		recycle_stat_inc(pool, cache_full);
>  		return false;
>  	}
> @@ -546,6 +568,8 @@ static bool page_pool_recycle_in_cache(struct page *page,
>  	return true;
>  }
>  
> +static void page_pool_empty_ring(struct page_pool *pool);
> +
>  /* If the page refcnt == 1, this will try to recycle the page.
>   * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
>   * the configured size min(dma_sync_size, pool->max_len).
> @@ -572,7 +596,8 @@ __page_pool_put_page(struct page_pool *pool, struct page *page,
>  			page_pool_dma_sync_for_device(pool, page,
>  						      dma_sync_size);
>  
> -		if (allow_direct && in_softirq() &&
> +		/* During PP shutdown, no direct recycle must occur */
> +		if (likely(allow_direct && in_softirq()) &&
>  		    page_pool_recycle_in_cache(page, pool))
>  			return NULL;
>  
> @@ -609,6 +634,8 @@ void page_pool_put_defragged_page(struct page_pool *pool, struct page *page,
>  		recycle_stat_inc(pool, ring_full);
>  		page_pool_return_page(pool, page);
>  	}
> +	if (pool->p.flags & PP_FLAG_SHUTDOWN)
> +		page_pool_empty_ring(pool);
>  }
>  EXPORT_SYMBOL(page_pool_put_defragged_page);
>  
> @@ -648,13 +675,17 @@ void page_pool_put_page_bulk(struct page_pool *pool, void **data,
>  
>  	/* Hopefully all pages was return into ptr_ring */
>  	if (likely(i == bulk_len))
> -		return;
> +		goto out;
>  
>  	/* ptr_ring cache full, free remaining pages outside producer lock
>  	 * since put_page() with refcnt == 1 can be an expensive operation
>  	 */
>  	for (; i < bulk_len; i++)
>  		page_pool_return_page(pool, data[i]);
> +
> +out:
> +	if (pool->p.flags & PP_FLAG_SHUTDOWN)
> +		page_pool_empty_ring(pool);
>  }
>  EXPORT_SYMBOL(page_pool_put_page_bulk);
>  
> @@ -737,6 +768,7 @@ struct page *page_pool_alloc_frag(struct page_pool *pool,
>  }
>  EXPORT_SYMBOL(page_pool_alloc_frag);
>  
> +noinline
>  static void page_pool_empty_ring(struct page_pool *pool)
>  {
>  	struct page *page;
> @@ -796,39 +828,29 @@ static void page_pool_scrub(struct page_pool *pool)
>  	page_pool_empty_ring(pool);
>  }

So this is not in the diff context, but page_pool_empty_ring() does
this:

static void page_pool_empty_ring(struct page_pool *pool)
{
	struct page *page;

	/* Empty recycle ring */
	while ((page = ptr_ring_consume_bh(&pool->ring))) {
		/* Verify the refcnt invariant of cached pages */
		if (!(page_ref_count(page) == 1))
			pr_crit("%s() page_pool refcnt %d violation\n",
				__func__, page_ref_count(page));

		page_pool_return_page(pool, page);
	}
}

...and with this patch, that page_pool_return_page() call will now free
the pool memory entirely when the last page is returned. When it does
this, the condition in the while loop will still execute afterwards; it
would return false, but if the pool was freed, it's now referencing
freed memory when trying to read from pool->ring.

So I think page_pool_empty_ring needs to either pull out all the pages
in the ring to an on-stack buffer before calling page_pool_return_page()
on them, or there needs to be some other way to break the loop early.

There are a couple of other places where page_pool_return_page() is
called in a loop where the loop variable lives inside struct page_pool,
so we need to be absolutely sure they will never be called in the
shutdown stage, or they'll have to be fixed as well.

>  
> -static int page_pool_release(struct page_pool *pool)
> +noinline
> +static int page_pool_free_attempt(struct page_pool *pool,
> +				  u32 hold_cnt, u32 release_cnt)
>  {
>  	int inflight;
>  
> -	page_pool_scrub(pool);
> -	inflight = page_pool_inflight(pool);
> +	inflight = __page_pool_inflight(pool, hold_cnt, release_cnt);
>  	if (!inflight)
>  		page_pool_free(pool);
>  
>  	return inflight;
>  }
>  
> -static void page_pool_release_retry(struct work_struct *wq)
> +static int page_pool_release(struct page_pool *pool)
>  {
> -	struct delayed_work *dwq = to_delayed_work(wq);
> -	struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
>  	int inflight;
>  
> -	inflight = page_pool_release(pool);
> +	page_pool_scrub(pool);
> +	inflight = page_pool_inflight(pool);
>  	if (!inflight)
> -		return;
> -
> -	/* Periodic warning */
> -	if (time_after_eq(jiffies, pool->defer_warn)) {
> -		int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
> -
> -		pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
> -			__func__, inflight, sec);
> -		pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
> -	}
> +		page_pool_free(pool);
>  
> -	/* Still not ready to be disconnected, retry later */
> -	schedule_delayed_work(&pool->release_dw, DEFER_TIME);
> +	return inflight;
>  }
>  
>  void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
> @@ -868,11 +890,13 @@ void page_pool_destroy(struct page_pool *pool)
>  	if (!page_pool_release(pool))
>  		return;
>  
> -	pool->defer_start = jiffies;
> -	pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
> +	/* PP have pages inflight, thus cannot immediately release memory.
> +	 * Enter into shutdown phase.
> +	 */
> +	pool->p.flags |= PP_FLAG_SHUTDOWN;

I think there's another race here: once the flag is set in this line
(does this need a memory barrier, BTW?), another CPU can return the last
outstanding page, read the flag and call page_pool_empty_ring(). If this
happens before the call to page_pool_empty_ring() below, you'll get a
use-after-free.

To avoid this, we could artificially bump the pool->hold_cnt *before*
setting the flag above; that way we know that the page_pool_empty_ring()
won't trigger a release, because inflight pages will never go below 1.
And then, below the page_pool_empty_ring() call below, we can add an
artificial bump of the release_cnt as well, which means we'll get proper
atomic semantics on the counters and only ever release once. I.e.,:

> -	INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
> -	schedule_delayed_work(&pool->release_dw, DEFER_TIME);
> +	/* Concurrent CPUs could have returned last pages into ptr_ring */
> +	page_pool_empty_ring(pool);

        release_cnt = atomic_inc_return(&pool->pages_state_release_cnt);
        page_pool_free_attempt(pool, release_cnt);


-Toke

Powered by blists - more mailing lists