page_pool: allow caching from safely localized NAPI
Subsystems (5)
Jakub Kicinski
maintainerv2 cover page_pool: allow caching from safely localized NAPI
I went back to the explicit "are we in NAPI method", mostly because I don't like having both around :( (even tho I maintain that in_softirq() && !in_hardirq() is as safe, as softirqs do not nest).
Still returning the skbs to a CPU, tho, not to the NAPI instance. I reckon we could create a small refcounted struct per NAPI instance which would allow sockets and other users so hold a persisent and safe reference. But that's a bigger change, and I get 90+% recycling thru the cache with just these patches (for RR and streaming tests with 100% CPU use it's almost 100%).
Some numbers for streaming test with 100% CPU use (from previous version, but really they perform the same):
HW-GRO page=page before after before after
recycle:
cached: 0 138669686 0 150197505cache_full: 0 223391 0 74582
ring: 138551933 9997191 149299454 0ring_full: 0 488 3154 127590released_refcnt: 0 0 0 0
alloc:
fast: 136491361 148615710 146969587 150322859
slow: 1772 1799 144 105slow_high_order: 0 0 0 0
empty: 1772 1799 144 105
refill: 2165245 156302 2332880 2128
waive: 0 0 0 0
v2:
- minor commit message fixes (patch 1)
v1: https://lore.kernel.org/all/[email protected]/
- rename the arg in_normal_napi -> napi_safe
- also allow recycling in __kfree_skb_defer()
rfcv2: https://lore.kernel.org/all/[email protected]/
Jakub Kicinski (3):
net: skb: plumb napi state thru skb freeing paths
page_pool: allow caching from safely localized NAPI
bnxt: hook NAPIs to page pools
Documentation/networking/page_pool.rst | 1 + drivers/net/ethernet/broadcom/bnxt/bnxt.c | 1 + include/linux/netdevice.h | 3 ++ include/linux/skbuff.h | 20 +++++++---- include/net/page_pool.h | 3 +- net/core/dev.c | 3 ++ net/core/page_pool.c | 15 ++++++-- net/core/skbuff.c | 42 ++++++++++++----------- 8 files changed, 58 insertions(+), 30 deletions(-)
Signature · 1 line
2.39.2
Jakub Kicinski
maintainerPATCH v2 1/3 net: skb: plumb napi state thru skb freeing paths
We maintain a NAPI-local cache of skbs which is fed by napi_consume_skb(). Going forward we will also try to cache head and data pages. Plumb the "are we in a normal NAPI context" information thru deeper into the freeing path, up to skb_release_data() and skb_free_head()/skb_pp_recycle(). The "not normal NAPI context" comes from netpoll which passes budget of 0 to try to reap
the Tx completions but not perform any Rx.
Use "bool napi_safe" rather than bare "int budget", the further we get from NAPI the more confusing the budget argument may seem (particularly whether 0 or MAX is the correct value to pass in when not in NAPI).
---
v2:
- clarify the budget 0 and fix the name of the argument in the commit message
v1:
- feed the cache in __kfree_skb_defer(), it's in NAPI
- s/in_normal_napi/napi_safe/
--- net/core/skbuff.c | 38 ++++++++++++++++++++------------------ 1 file changed, 20 insertions(+), 18 deletions(-)
net/core/skbuff.c +20 -18
@@ -839,7 +839,7 @@ static void skb_clone_fraglist(struct sk_buff *skb)skb_get (list); } -static bool skb_pp_recycle (struct sk_buff *skb,void *data) +static bool skb_pp_recycle (struct sk_buff *skb,void *data,bool napi_safe) {if (!IS_ENABLED (CONFIG_PAGE_POOL ) || !skb->pp_recycle)return false;
@@ -856,12 +856,12 @@ static void skb_kfree_head(void *head, unsigned int end_offset)kfree (head); } -static void skb_free_head (struct sk_buff *skb) +static void skb_free_head (struct sk_buff *skb,bool napi_safe) {unsigned char *head = skb->head;if (skb->head_frag) { -if (skb_pp_recycle (skb, head)) +if (skb_pp_recycle (skb, head, napi_safe))return ;skb_free_frag (head); }else {
@@ -869,7 +869,8 @@ static void skb_free_head(struct sk_buff *skb) } } -static void skb_release_data (struct sk_buff *skb,enum skb_drop_reason reason) +static void skb_release_data (struct sk_buff *skb,enum skb_drop_reason reason, +bool napi_safe) {struct skb_shared_info *shinfo =skb_shinfo (skb);int i;
@@ -894,7 +895,7 @@ static void skb_release_data(struct sk_buff *skb, enum skb_drop_reason reason)if (shinfo->frag_list)kfree_skb_list_reason (shinfo->frag_list, reason); -skb_free_head (skb); +skb_free_head (skb, napi_safe); exit:/* When we clone an SKB we copy the reycling bit. The pp_recycle * bit is only set on the head though, so in order to avoid races
@@ -955,11 +956,12 @@ void skb_release_head_state(struct sk_buff *skb) }/* Free everything but the sk_buff shell. */ -static void skb_release_all (struct sk_buff *skb,enum skb_drop_reason reason) +static void skb_release_all (struct sk_buff *skb,enum skb_drop_reason reason, +bool napi_safe) {skb_release_head_state (skb);if (likely (skb->head)) -skb_release_data (skb, reason); +skb_release_data (skb, reason, napi_safe); }/**
@@ -973,7 +975,7 @@ static void skb_release_all(struct sk_buff *skb, enum skb_drop_reason reason)void __kfree_skb (struct sk_buff *skb) { -skb_release_all (skb,SKB_DROP_REASON_NOT_SPECIFIED ); +skb_release_all (skb,SKB_DROP_REASON_NOT_SPECIFIED , false);kfree_skbmem (skb); }EXPORT_SYMBOL (__kfree_skb);
@@ -1027,7 +1029,7 @@ static void kfree_skb_add_bulk(struct sk_buff *skb,return ; } -skb_release_all (skb, reason); +skb_release_all (skb, reason, false); sa->skb_array[sa->skb_count++] = skb;if (unlikely (sa->skb_count ==KFREE_SKB_BULK_SIZE )) {
@@ -1201,7 +1203,7 @@ EXPORT_SYMBOL(consume_skb);void __consume_stateless_skb (struct sk_buff *skb) {trace_consume_skb (skb,__builtin_return_address (0 )); -skb_release_data (skb,SKB_CONSUMED ); +skb_release_data (skb,SKB_CONSUMED , false);kfree_skbmem (skb); }
@@ -1226,7 +1228,7 @@ static void napi_skb_cache_put(struct sk_buff *skb)void __kfree_skb_defer (struct sk_buff *skb) { -skb_release_all (skb,SKB_DROP_REASON_NOT_SPECIFIED ); +skb_release_all (skb,SKB_DROP_REASON_NOT_SPECIFIED , true);napi_skb_cache_put (skb); }
@@ -1264,7 +1266,7 @@ void napi_consume_skb(struct sk_buff *skb, int budget)return ; } -skb_release_all (skb,SKB_CONSUMED ); +skb_release_all (skb,SKB_CONSUMED , !!budget);napi_skb_cache_put (skb); }EXPORT_SYMBOL (napi_consume_skb);
@@ -1395,7 +1397,7 @@ EXPORT_SYMBOL_GPL(alloc_skb_for_msg);*/ struct sk_buff *skb_morph (struct sk_buff *dst,struct sk_buff *src) { -skb_release_all (dst,SKB_CONSUMED ); +skb_release_all (dst,SKB_CONSUMED , false);return __skb_clone (dst, src); }EXPORT_SYMBOL_GPL (skb_morph);
@@ -2018,9 +2020,9 @@ int pskb_expand_head(struct sk_buff *skb, int nhead, int ntail,if (skb_has_frag_list (skb))skb_clone_fraglist (skb); -skb_release_data (skb,SKB_CONSUMED ); +skb_release_data (skb,SKB_CONSUMED , false); }else { -skb_free_head (skb); +skb_free_head (skb, false); } off = (data + nhead) - skb->head;
@@ -6389,12 +6391,12 @@ static int pskb_carve_inside_header(struct sk_buff *skb, const u32 off,skb_frag_ref (skb, i);if (skb_has_frag_list (skb))skb_clone_fraglist (skb); -skb_release_data (skb,SKB_CONSUMED ); +skb_release_data (skb,SKB_CONSUMED , false); }else {/* we can reuse existing recount- all we did was * relocate values */ -skb_free_head (skb); +skb_free_head (skb, false); } skb->head = data;
@@ -6529,7 +6531,7 @@ static int pskb_carve_inside_nonlinear(struct sk_buff *skb, const u32 off,skb_kfree_head (data, size);return -ENOMEM ; } -skb_release_data (skb,SKB_CONSUMED ); +skb_release_data (skb,SKB_CONSUMED , false); skb->head = data; skb->head_frag =0 ;
Signature · 1 line
2.39.2
Re: [PATCH net-next v2 1/3] net: skb: plumb napi state thru skb freeing paths
Quoted 7 lines · Jakub Kicinski
We maintain a NAPI-local cache of skbs which is fed by
napi_consume_skb(). Going forward we will also try to cache head and data pages. Plumb the "are we in a normal NAPI context" information thru deeper into the freeing path, up toskb_release_data()and skb_free_head()/skb_pp_recycle(). The "not normal NAPI context" comes from netpoll which passes budget of 0 to try to reap
the Tx completions but not perform any Rx.
Maybe I missed something obvious about netpoll here.
Does that mean the "normal NAPI context" and "not normal NAPI context" will call napi->poll() concurrently with different budget? Doesn't that mean two different contexts may do the tx completion concurrently? Does it break the single-producer single-consumer assumption of tx queue?
Quoted 118 lines
Use "bool
napi_safe" rather than bare "int budget", the further we get from NAPI the more confusing the budget argument may seem (particularly whether 0 or MAX is the correct value to pass in when not in NAPI).Reviewed-by Tariq Toukan <[email protected]>Signed-off-by Jakub Kicinski <[email protected]>---
v2:
- clarify the budget 0 and fix the name of the argument in the commit message
v1:
- feed the cache in__kfree_skb_defer(), it's in NAPI
- s/in_normal_napi/napi_safe/--- net/core/skbuff.c | 38 ++++++++++++++++++++------------------ 1 file changed, 20 insertions(+), 18 deletions(-)net/core/skbuff.c +13
-11@@ -839,7 +839,7 @@ static void skb_clone_fraglist(struct sk_buff *skb)skb_get (list); } -static bool skb_pp_recycle (struct sk_buff *skb,void *data) +static bool skb_pp_recycle (struct sk_buff *skb,void *data,bool napi_safe) {if (!IS_ENABLED (CONFIG_PAGE_POOL ) || !skb->pp_recycle)return false;@@ -856,12 +856,12 @@ static void skb_kfree_head(void *head, unsigned int end_offset)kfree (head); } -static void skb_free_head (struct sk_buff *skb) +static void skb_free_head (struct sk_buff *skb,bool napi_safe) {unsigned char *head = skb->head;if (skb->head_frag) { -if (skb_pp_recycle (skb, head)) +if (skb_pp_recycle (skb, head, napi_safe))return ;skb_free_frag (head); }else {@@ -869,7 +869,8 @@ static void skb_free_head(struct sk_buff *skb) } } -static void skb_release_data (struct sk_buff *skb,enum skb_drop_reason reason) +static void skb_release_data (struct sk_buff *skb,enum skb_drop_reason reason, +bool napi_safe) {struct skb_shared_info *shinfo =skb_shinfo (skb);int i;@@ -894,7 +895,7 @@ static void skb_release_data(struct sk_buff *skb, enum skb_drop_reason reason)if (shinfo->frag_list)kfree_skb_list_reason (shinfo->frag_list, reason); -skb_free_head (skb); +skb_free_head (skb, napi_safe); exit:/* When we clone an SKB we copy the reycling bit. The pp_recycle * bit is only set on the head though, so in order to avoid races @@ -955,11 +956,12 @@ void skb_release_head_state(struct sk_buff *skb) }/* Free everything but the sk_buff shell. */ -static void skb_release_all (struct sk_buff *skb,enum skb_drop_reason reason) +static void skb_release_all (struct sk_buff *skb,enum skb_drop_reason reason, +bool napi_safe) {skb_release_head_state (skb);if (likely (skb->head)) -skb_release_data (skb, reason); +skb_release_data (skb, reason, napi_safe); }/** @@ -973,7 +975,7 @@ static void skb_release_all(struct sk_buff *skb, enum skb_drop_reason reason)void __kfree_skb (struct sk_buff *skb) { -skb_release_all (skb,SKB_DROP_REASON_NOT_SPECIFIED ); +skb_release_all (skb,SKB_DROP_REASON_NOT_SPECIFIED , false);kfree_skbmem (skb); }EXPORT_SYMBOL (__kfree_skb);@@ -1027,7 +1029,7 @@ static void kfree_skb_add_bulk(struct sk_buff *skb,return ; } -skb_release_all (skb, reason); +skb_release_all (skb, reason, false); sa->skb_array[sa->skb_count++] = skb;if (unlikely (sa->skb_count ==KFREE_SKB_BULK_SIZE )) {@@ -1201,7 +1203,7 @@ EXPORT_SYMBOL(consume_skb);void __consume_stateless_skb (struct sk_buff *skb) {trace_consume_skb (skb,__builtin_return_address (0 )); -skb_release_data (skb,SKB_CONSUMED ); +skb_release_data (skb,SKB_CONSUMED , false);kfree_skbmem (skb); }@@ -1226,7 +1228,7 @@ static void napi_skb_cache_put(struct sk_buff *skb)void __kfree_skb_defer (struct sk_buff *skb) { -skb_release_all (skb,SKB_DROP_REASON_NOT_SPECIFIED ); +skb_release_all (skb,SKB_DROP_REASON_NOT_SPECIFIED , true);napi_skb_cache_put (skb); }
net/core/skbuff.c void napi_consume_skb(struct sk_buff *skb, int budget) · line 1269
Quoted 5 lines
net/core/skbuff.c +1
-1@@ -1264,7 +1266,7 @@ void napi_consume_skb(struct sk_buff *skb, int budget)return ; } -skb_release_all (skb,SKB_CONSUMED ); +skb_release_all (skb,SKB_CONSUMED , !!budget);
If it is not normal NAPI context, dev_consume_skb_any() is called and return at the begin, we may just call skb_release_all() with napi_safe being true here.
Quoted 49 lines
net/core/skbuff.c +6
-6napi_skb_cache_put (skb); }EXPORT_SYMBOL (napi_consume_skb);@@ -1395,7 +1397,7 @@ EXPORT_SYMBOL_GPL(alloc_skb_for_msg);*/ struct sk_buff *skb_morph (struct sk_buff *dst,struct sk_buff *src) { -skb_release_all (dst,SKB_CONSUMED ); +skb_release_all (dst,SKB_CONSUMED , false);return __skb_clone (dst, src); }EXPORT_SYMBOL_GPL (skb_morph);@@ -2018,9 +2020,9 @@ int pskb_expand_head(struct sk_buff *skb, int nhead, int ntail,if (skb_has_frag_list (skb))skb_clone_fraglist (skb); -skb_release_data (skb,SKB_CONSUMED ); +skb_release_data (skb,SKB_CONSUMED , false); }else { -skb_free_head (skb); +skb_free_head (skb, false); } off = (data + nhead) - skb->head;@@ -6389,12 +6391,12 @@ static int pskb_carve_inside_header(struct sk_buff *skb, const u32 off,skb_frag_ref (skb, i);if (skb_has_frag_list (skb))skb_clone_fraglist (skb); -skb_release_data (skb,SKB_CONSUMED ); +skb_release_data (skb,SKB_CONSUMED , false); }else {/* we can reuse existing recount- all we did was * relocate values */ -skb_free_head (skb); +skb_free_head (skb, false); } skb->head = data;@@ -6529,7 +6531,7 @@ static int pskb_carve_inside_nonlinear(struct sk_buff *skb, const u32 off,skb_kfree_head (data, size);return -ENOMEM ; } -skb_release_data (skb,SKB_CONSUMED ); +skb_release_data (skb,SKB_CONSUMED , false); skb->head = data; skb->head_frag =0 ;
Re: [PATCH net-next v2 1/3] net: skb: plumb napi state thru skb freeing paths
Quoted 5 lines · Yunsheng Lin
Maybe I missed something obvious about netpoll here.
Does that mean the "normal NAPI context" and "not normal NAPI context" will callnapi->poll()concurrently with different budget? Doesn't that mean two different contexts may do the tx completion concurrently?
Please take a look at netpoll code:netpoll_poll_lock, poll_napi() and poll_one_napi()
Quoted 1 line
Does it break the single-producer single-consumer assumption of tx queue?
We do not think so.
Re: [PATCH net-next v2 1/3] net: skb: plumb napi state thru skb freeing paths
Quoted 7 lines · Eric Dumazet
Quoted 5 lines · Yunsheng Lin
Maybe I missed something obvious about netpoll here.
Does that mean the "normal NAPI context" and "not normal NAPI context" will callnapi->poll()concurrently with different budget? Doesn't that mean two different contexts may do the tx completion concurrently?
Quoted 6 lines
Please take a look at netpoll code:
netpoll_poll_lock,poll_napi()andpoll_one_napi()Quoted 1 line
Does it break the single-producer single-consumer assumption of tx queue?
We do not think so.
Then I guess it is ok to do direct recycling for page pool case as it is per napi?
It is per cpu cache case we are using !!bugget to protect it from preemption while netpoll_poll_dev() is running?
Quoted 2 lines
.
Jakub Kicinski
maintainerRe: [PATCH net-next v2 1/3] net: skb: plumb napi state thru skb freeing paths
Quoted 6 lines · On Fri, 14 Apr 2023 08:57:03 +0800 Yunsheng Lin wrote:
Quoted 3 lines
Quoted 1 line
Does it break the single-producer single-consumer assumption of tx queue?
We do not think so.
Then I guess it is ok to do direct recycling for page pool case as it is per napi?
We're talking about the tx queue or the pp cache?
Those have different producers and consumers.
Quoted 2 lines
It is per cpu cache case we are using !!bugget to protect it from preemption while
netpoll_poll_dev()is running?
This is the scenario - We are feeding the page pool cache from the deferred pages. An IRQ comes and interrupts us half way thru. netpoll then tries to also feed the page pool cache. Unhappiness.
Note that netpoll is activated extremely rarely (only when something writes to the console), and even more rarely does it actually poll for space in the Tx queue.
Re: [PATCH net-next v2 1/3] net: skb: plumb napi state thru skb freeing paths
Quoted 11 lines · Jakub Kicinski
Quoted 6 lines · On Fri, 14 Apr 2023 08:57:03 +0800 Yunsheng Lin wrote:
Quoted 3 lines
Quoted 1 line
Does it break the single-producer single-consumer assumption of tx queue?
We do not think so.
Then I guess it is ok to do direct recycling for page pool case as it is per napi?
We're talking about the tx queue or the pp cache?
Those have different producers and consumers.
I was trying to confirm if there are more than one contexts that will call napi->poll() concurrently, if no, then it seems we only have one consumer for pp cache. And it does not really matter here too when napi->poll() from netpoll only do tx completion now.
If we are able to limit the context of producer for pp cache to be the same as consumer, then we might avoid the !!bugget checking.
I do ack that it is a bigger change considering when we might need to hold a persisent and safe reference to the NAPI instance for this too, so we might leave it for now like you mentioned in the cover letter.
Quoted 7 lines
Quoted 3 lines
It is per cpu cache case we are using !!bugget to protect it from preemption while
netpoll_poll_dev()is running?This is the scenario - We are feeding the page pool cache from the deferred pages. An IRQ comes and interrupts us half way thru. netpoll then tries to also feed the page pool cache. Unhappiness.
I assume feeding the page pool cache means producer for pp cache? If netpoll only do tx completion by using zero bugget, it does not seem to have any direct relation with page pool yet.
Quoted 4 lines
Note that netpoll is activated extremely rarely (only when something writes to the console), and even more rarely does it actually poll for space in the Tx queue.
That is my point, we turn out to need a checking in the fast path in order to handle the netpoll case, which is a extreme rare case.
Quoted 2 lines
.
Jakub Kicinski
maintainerPATCH v2 2/3 page_pool: allow caching from safely localized NAPI
In reply to Jakub KicinskiRecent patches to mlx5 mentioned a regression when moving from driver local page pool to only using the generic page pool code. Page pool has two recycling paths (1) direct one, which runs in safe NAPI context (basically consumer context, so producing can be lockless); and (2) via a ptr_ring, which takes a spin lock because the freeing can happen from any CPU; producer and consumer may run concurrently.
Since the page pool code was added, Eric introduced a revised version of deferred skb freeing. TCP skbs are now usually returned to the CPU which allocated them, and freed in softirq context. This places the freeing (producing of pages back to the pool) enticingly close to the allocation (consumer).
If we can prove that we're freeing in the same softirq context in which the consumer NAPI will run - lockless use of the cache is perfectly fine, no need for the lock.
Let drivers link the page pool to a NAPI instance. If the NAPI instance is scheduled on the same CPU on which we're freeing - place the pages in the direct cache.
With that and patched bnxt (XDP enabled to engage the page pool, sigh, bnxt really needs page pool work :() I see a 2.6% perf boost with a TCP stream test (app on a different physical core than softirq).
The CPU use of relevant functions decreases as expected:
page_pool_refill_alloc_cache 1.17% -> 0%
_raw_spin_lock 2.41% -> 0.98%
Only consider lockless path to be safe when NAPI is scheduled
- in practice this should cover majority if not all of steady state workloads. It's usually the NAPI kicking in that causes the skb flush.
The main case we'll miss out on is when application runs on the same CPU as NAPI. In that case we don't use the deferred skb free path.
---
v1:
- s/in_napi/napi_safe/
rfc v2: https://lore.kernel.org/all/[email protected]/
- plumb thru "are we in NAPI" bool rather than guessing based on softirq && !hardirq
--- Documentation/networking/page_pool.rst | 1 + include/linux/netdevice.h | 3 +++ include/linux/skbuff.h | 20 +++++++++++++------- include/net/page_pool.h | 3 ++- net/core/dev.c | 3 +++ net/core/page_pool.c | 15 +++++++++++++-- net/core/skbuff.c | 4 ++-- 7 files changed, 37 insertions(+), 12 deletions(-)
Documentation/networking/page_pool.rst +1 -0
@@ -165,6 +165,7 @@ Registration pp_params.pool_size = DESC_NUM; pp_params.nid = NUMA_NO_NODE; pp_params.dev = priv->dev; + pp_params.napi = napi; /* only if locking is tied to NAPI */ pp_params.dma_dir = xdp_prog ? DMA_BIDIRECTIONAL : DMA_FROM_DEVICE; page_pool = page_pool_create(&pp_params);
include/linux/netdevice.h +3 -0
@@ -360,8 +360,11 @@ struct napi_struct {unsigned long gro_bitmask;int (*poll)(struct napi_struct *,int );#ifdef CONFIG_NETPOLL +/* CPU actively polling if netpoll is configured */ int poll_owner;#endif +/* CPU on which NAPI has been scheduled for processing */ +int list_owner;struct net_device *dev;struct gro_list gro_hash[GRO_HASH_BUCKETS ];struct sk_buff *skb;
include/linux/skbuff.h +13 -7
@@ -3386,6 +3386,18 @@ static inline void skb_frag_ref(struct sk_buff *skb, int f)__skb_frag_ref (&skb_shinfo (skb)->frags[f]); } +static inline void +napi_frag_unref (skb_frag_t *frag,bool recycle,bool napi_safe) +{ +struct page *page =skb_frag_page (frag); + +#ifdef CONFIG_PAGE_POOL +if (recycle &&page_pool_return_skb_page (page, napi_safe)) +return ; +#endif +put_page (page); +} +/** * __skb_frag_unref - release a reference on a paged fragment. * @frag: the paged fragment
@@ -3396,13 +3408,7 @@ static inline void skb_frag_ref(struct sk_buff *skb, int f)*/ static inline void __skb_frag_unref (skb_frag_t *frag,bool recycle) { -struct page *page =skb_frag_page (frag); - -#ifdef CONFIG_PAGE_POOL -if (recycle &&page_pool_return_skb_page (page)) -return ; -#endif -put_page (page); +napi_frag_unref (frag, recycle, false); }/**
include/net/page_pool.h +2 -1
@@ -77,6 +77,7 @@ struct page_pool_params {unsigned int pool_size;int nid;/* Numa node id to allocate from pages from */ struct device *dev;/* device, for DMA pre-mapping purposes */ +struct napi_struct *napi;/* Sole consumer of pages, otherwise NULL */ enum dma_data_direction dma_dir;/* DMA mapping direction */ unsigned int max_len;/* max DMA sync memory size */ unsigned int offset;/* DMA addr offset */
@@ -239,7 +240,7 @@ inline enum dma_data_direction page_pool_get_dma_dir(struct page_pool *pool)return pool->p.dma_dir; } -bool page_pool_return_skb_page (struct page *page); +bool page_pool_return_skb_page (struct page *page,bool napi_safe);struct page_pool *page_pool_create (const struct page_pool_params *params);
net/core/dev.c +3 -0
@@ -4358,6 +4358,7 @@ static inline void ____napi_schedule(struct softnet_data *sd, }list_add_tail (&napi->poll_list, &sd->poll_list); +WRITE_ONCE (napi->list_owner,smp_processor_id ());/* If not called from net_rx_action() * we have to raise NET_RX_SOFTIRQ. */
@@ -6068,6 +6069,7 @@ bool napi_complete_done(struct napi_struct *n, int work_done)list_del_init (&n->poll_list);local_irq_restore (flags); } +WRITE_ONCE (n->list_owner,-1 ); val =READ_ONCE (n->state); do {
@@ -6383,6 +6385,7 @@ void netif_napi_add_weight(struct net_device *dev, struct napi_struct *napi,#ifdef CONFIG_NETPOLL napi->poll_owner =-1 ;#endif + napi->list_owner =-1 ;set_bit (NAPI_STATE_SCHED , &napi->state);set_bit (NAPI_STATE_NPSVC , &napi->state);list_add_rcu (&napi->dev_list, &dev->napi_list);
net/core/page_pool.c +13 -2
@@ -19,6 +19,7 @@#include <linux/mm.h> /* for put_page() */ #include <linux/poison.h> #include <linux/ethtool.h> +#include <linux/netdevice.h> #include <trace/events/page_pool.h>
@@ -874,9 +875,11 @@ void page_pool_update_nid(struct page_pool *pool, int new_nid) }EXPORT_SYMBOL (page_pool_update_nid); -bool page_pool_return_skb_page (struct page *page) +bool page_pool_return_skb_page (struct page *page,bool napi_safe) { +struct napi_struct *napi;struct page_pool *pp; +bool allow_direct; page =compound_head (page);
@@ -892,12 +895,20 @@ bool page_pool_return_skb_page(struct page *page) pp = page->pp; +/* Allow direct recycle if we have reasons to believe that we are +* in the same context as the consumer would run, so there's +* no possible race. +*/ + napi = pp->p.napi; + allow_direct = napi_safe && napi && +READ_ONCE (napi->list_owner) ==smp_processor_id (); +/* Driver set this to memory recycling info. Reset it on recycle. * This will *not* work for NIC using a split-page memory model. * The page will be returned to the pool here regardless of the * 'flipped' fragment being in use or not. */ -page_pool_put_full_page (pp, page, false); +page_pool_put_full_page (pp, page, allow_direct);return true; }
net/core/skbuff.c +2 -2
@@ -843,7 +843,7 @@ static bool skb_pp_recycle(struct sk_buff *skb, void *data, bool napi_safe) {if (!IS_ENABLED (CONFIG_PAGE_POOL ) || !skb->pp_recycle)return false; -return page_pool_return_skb_page (virt_to_page (data)); +return page_pool_return_skb_page (virt_to_page (data), napi_safe); }static void skb_kfree_head (void *head,unsigned int end_offset)
@@ -889,7 +889,7 @@ static void skb_release_data(struct sk_buff *skb, enum skb_drop_reason reason, }for (i =0 ; i < shinfo->nr_frags; i++) -__skb_frag_unref (&shinfo->frags[i], skb->pp_recycle); +napi_frag_unref (&shinfo->frags[i], skb->pp_recycle, napi_safe); free_head:if (shinfo->frag_list)
Signature · 1 line
2.39.2
Quoted 49 lines · Jakub Kicinski
Recent patches to mlx5 mentioned a regression when moving from driver local page pool to only using the generic page pool code. Page pool has two recycling paths (1) direct one, which runs in safe NAPI context (basically consumer context, so producing can be lockless); and (2) via a
ptr_ring, which takes a spin lock because the freeing can happen from any CPU; producer and consumer may run concurrently.Since the page pool code was added, Eric introduced a revised version of deferred skb freeing. TCP skbs are now usually returned to the CPU which allocated them, and freed in softirq context. This places the freeing (producing of pages back to the pool) enticingly close to the allocation (consumer).
If we can prove that we're freeing in the same softirq context in which the consumer NAPI will run - lockless use of the cache is perfectly fine, no need for the lock.
Let drivers link the page pool to a NAPI instance. If the NAPI instance is scheduled on the same CPU on which we're freeing - place the pages in the direct cache.
With that and patched bnxt (XDP enabled to engage the page pool, sigh, bnxt really needs page pool work :() I see a 2.6% perf boost with a TCP stream test (app on a different physical core than softirq).
The CPU use of relevant functions decreases as expected:
page_pool_refill_alloc_cache 1.17% -> 0% _raw_spin_lock 2.41% -> 0.98%Only consider lockless path to be safe when NAPI is scheduled
- in practice this should cover majority if not all of steady state workloads. It's usually the NAPI kicking in that causes the skb flush.The main case we'll miss out on is when application runs on the same CPU as NAPI. In that case we don't use the deferred skb free path.
Reviewed-by Tariq Toukan<[email protected]>Signed-off-by Jakub Kicinski<[email protected]>---
v1:
- s/in_napi/napi_safe/
rfc v2:https://lore.kernel.org/all/[email protected]/
- plumb thru "are we in NAPI" bool rather than guessing basedon softirq && !hardirq
Quoted 9 lines
--- Documentation/networking/page_pool.rst | 1 + include/linux/netdevice.h | 3 +++ include/linux/skbuff.h | 20 +++++++++++++------- include/net/page_pool.h | 3 ++- net/core/dev.c | 3 +++ net/core/page_pool.c | 15 +++++++++++++-- net/core/skbuff.c | 4 ++-- 7 files changed, 37 insertions(+), 12 deletions(-)
BTW, does it matter if I ack with <[email protected]> or <[email protected]> ?
e.g. does the patchwork automation scripts, need to maintainer email to match?
--Jesper
Jakub Kicinski
maintainerQuoted 1 line · On Fri, 14 Apr 2023 10:45:44 +0200 Jesper Dangaard Brouer wrote:
Thanks!
Quoted 4 lines
BTW, does it matter if I ack with <[email protected]> or <[email protected]> ?
e.g. does the patchwork automation scripts, need to maintainer email to match?
For better or worse we don't have any substantial automation in the area of maintainer ack tracking. But if we did we'd run it thru various mail maps for sure, so it should be fine even then. I hope.
Jakub Kicinski
maintainerPATCH v2 3/3 bnxt: hook NAPIs to page pools
In reply to Jakub Kicinskibnxt has 1:1 mapping of page pools and NAPIs, so it's safe to hoook them up together.
---
CC: [email protected]
--- drivers/net/ethernet/broadcom/bnxt/bnxt.c | 1 + 1 file changed, 1 insertion(+)
drivers/net/ethernet/broadcom/bnxt/bnxt.c +1 -0
@@ -3211,6 +3211,7 @@ static int bnxt_alloc_rx_page_pool(struct bnxt *bp, pp.pool_size = bp->rx_ring_size; pp.nid =dev_to_node (&bp->pdev->dev); + pp.napi = &rxr->bnapi->napi; pp.dev = &bp->pdev->dev; pp.dma_dir =DMA_BIDIRECTIONAL ;
Signature · 1 line
2.39.2
Quoted 30 lines · Jakub Kicinski
I went back to the explicit "are we in NAPI method", mostly because I don't like having both around :( (even tho I maintain that
in_softirq()&& !in_hardirq() is as safe, as softirqs do not nest).Still returning the skbs to a CPU, tho, not to the NAPI instance. I reckon we could create a small refcounted struct per NAPI instance which would allow sockets and other users so hold a persisent and safe reference. But that's a bigger change, and I get 90+% recycling thru the cache with just these patches (for RR and streaming tests with 100% CPU use it's almost 100%).
Some numbers for streaming test with 100% CPU use (from previous version,
but really they perform the same):HW-GRO page=page before after before after recycle:
cached: 0 138669686 0 15019 7505
cache_full: 0 223391 0 74582
ring: 138551933
9997191 149299454 0
ring_full: 0 488 3154 127590
released_refcnt: 0 0 0 0
Quoted 14 lines
alloc:
fast: 136491361 148615710 146969587 15032 2859
slow: 1772 1799 144 105
slow_high_order: 0 0 0 0
empty: 1772 1799 144 105
refill: 2165245 156302 2332880 2128
waive: 0 0 0 0
Enabled this on the mlx5 driver and seeing the following page_pool cache usage improvements for single stream iperf test case:
- For 1500 MTU and legacy rq, seeing a 20% improvement of cache usage.
- For 9K MTU, seeing 33-40 % page_pool cache usage improvements for both striding and legacy rq (depending if the app is running on the same core as the rq or not).
One thing to note is that the page_pool cache seems to get filled more often for striding rq now which is something that we could potentially improve on the mlx5 side.
Regarding my earlier comment:
Quoted 3 lines
After enabling this in the mlx5 driver, there is already improved
page_poolcache usage for our test with the application running on the same CPU with the receive queue NAPI (0 -> 98 % cache usage).
I was testing without the deferred release optimizations that we did in the mlx5 driver.
Quoted 23 lines
v2:
- minor commit message fixes (patch 1)
v1:
https://lore.kernel.org/all/[email protected]/
- rename the argin_normal_napi->napi_safe
- also allow recycling in__kfree_skb_defer()
rfcv2:
https://lore.kernel.org/all/[email protected]/Jakub Kicinski (3):
net: skb: plumb napi state thru skb freeing paths
page_pool: allow caching from safely localized NAPI
bnxt: hook NAPIs to page poolsDocumentation/networking/page_pool.rst | 1 +
drivers/net/ethernet/broadcom/bnxt/bnxt.c| 1 + include/linux/netdevice.h | 3 ++ include/linux/skbuff.h | 20 +++++++---- include/net/page_pool.h | 3 +- net/core/dev.c | 3 ++ net/core/page_pool.c | 15 ++++++-- net/core/skbuff.c | 42 ++++++++++++--------- --
Quoted 2 lines
8 files changed, 58 insertions(+), 30 deletions(-)
This series was applied to netdev/net-next.git (main) by Jakub Kicinski <[email protected]>:
Hello:
This series was applied to netdev/net-next.git (main) by Jakub Kicinski <[email protected]>:
Quoted 5 lines · On Wed, 12 Apr 2023 21:26:02 -0700 you wrote:
I went back to the explicit "are we in NAPI method", mostly because I don't like having both around :( (even tho I maintain that
in_softirq()&& !in_hardirq() is as safe, as softirqs do not nest).
Quoted 8 lines
Still returning the skbs to a CPU, tho, not to the NAPI instance. I reckon we could create a small refcounted struct per NAPI instance which would allow sockets and other users so hold a persisent and safe reference. But that's a bigger change, and I get 90+% recycling thru the cache with just these patches (for RR and streaming tests with 100% CPU use it's almost 100%).
[...]
Here is the summary with links:
- [net-next,v2,1/3] net: skb: plumb napi state thru skb freeing paths
https://git.kernel.org/netdev/net-next/c/b07a2d97ba5e - [net-next,v2,2/3] page_pool: allow caching from safely localized NAPI https://git.kernel.org/netdev/net-next/c/8c48eea3adf3 - [net-next,v2,3/3] bnxt: hook NAPIs to page pools https://git.kernel.org/netdev/net-next/c/294e39e0d034
You are awesome, thank you!
Signature · 2 lines
Deet-doot-dot, I am a bot. https://korg.docs.kernel.org/patchwork/pwbot.html