Page Pool API

The page_pool allocator is optimized for recycling page or page fragment usedby skb packet and xdp frame.

Basic use involves replacing anyalloc_pages() calls withpage_pool_alloc(),which allocate memory with or without page splitting depending on therequested memory size.

If the driver knows that it always requires full pages or its allocations arealways smaller than half a page, it can use one of the more specific APIcalls:

1.page_pool_alloc_pages(): allocate memory without page splitting whendriver knows that the memory it need is always bigger than half of the pageallocated from page pool. There is no cache line dirtying for ‘structpage’when a page is recycled back to the page pool.

2.page_pool_alloc_frag(): allocate memory with page splitting when driverknows that the memory it need is always smaller than or equal to half of thepage allocated from page pool. Page splitting enables memory saving and thusavoids TLB/cache miss for data access, but there also is some cost toimplement page splitting, mainly some cache line dirtying/bouncing for‘structpage’ and atomic operation for page->pp_ref_count.

The API keeps track of in-flight pages, in order to let API users know whenit is safe to free a page_pool object, the API users must callpage_pool_put_page() orpage_pool_free_va() to free the page_pool object, orattach the page_pool object to a page_pool-aware object like skbs marked withskb_mark_for_recycle().

page_pool_put_page() may be called multiple times on the same page if a pageis split into multiple fragments. For the last fragment, it will eitherrecycle the page, or in case of page->_refcount > 1, it will release the DMAmapping and in-flight state accounting.

dma_sync_single_range_for_device() is only called for the last fragment whenpage_pool is created with PP_FLAG_DMA_SYNC_DEV flag, so it depends on thelast freed fragment to do the sync_for_device operation for all fragments inthe same page when a page is split. The API user must setup pool->p.max_lenand pool->p.offset correctly and ensure thatpage_pool_put_page() is calledwith dma_sync_size being -1 for fragment API.

Architecture overview

+------------------+|       Driver     |+------------------+        ^        |        |        |        v+--------------------------------------------+|                request memory              |+--------------------------------------------+    ^                                  ^    |                                  |    | Pool empty                       | Pool has entries    |                                  |    v                                  v+-----------------------+     +------------------------+| alloc (and map) pages |     |  get page from cache   |+-----------------------+     +------------------------+                                ^                    ^                                |                    |                                | cache available    | No entries, refill                                |                    | from ptr-ring                                |                    |                                v                    v                      +-----------------+     +------------------+                      |   Fast cache    |     |  ptr-ring cache  |                      +-----------------+     +------------------+

Monitoring

Information about page pools on the system can be accessed via the netdevgenetlink family (see Documentation/netlink/specs/netdev.yaml).

API interface

The number of pools createdmust match the number of hardware queuesunless hardware restrictions make that impossible. This would otherwise beat thepurpose of page pool, which is allocate pages fast from cache without locking.This lockless guarantee naturally comes from running under a NAPI softirq.The protection doesn’t strictly have to be NAPI, any guarantee that allocatinga page will cause no race conditions is enough.

structpage_pool*page_pool_create(conststructpage_pool_params*params)

create a page pool

Parameters

conststructpage_pool_params*params

parameters, seestructpage_pool_params

structpage_pool_params

page pool parameters

Definition:

struct page_pool_params {    struct page_pool_params_fast  fast;    unsigned int    order;    unsigned int    pool_size;    int nid;    struct device   *dev;    struct napi_struct *napi;    enum dma_data_direction dma_dir;    unsigned int    max_len;    unsigned int    offset;    struct page_pool_params_slow  slow;    STRUCT_GROUP( struct net_device *netdev;    unsigned int queue_idx;    unsigned int    flags;};

Members

fast

params accessed frequently on hotpath

order

2^order pages on allocation

pool_size

size of the ptr_ring

nid

NUMA node id to allocate from pages from

dev

device, for DMA pre-mapping purposes

napi

NAPI which is the sole consumer of pages, otherwise NULL

dma_dir

DMA mapping direction

max_len

max DMA sync memory size for PP_FLAG_DMA_SYNC_DEV

offset

DMA sync address offset for PP_FLAG_DMA_SYNC_DEV

slow

params with slowpath access only (initialization and Netlink)

netdev

netdev this pool will serve (leave as NULL if none or multiple)

queue_idx

queue idx this page_pool is being created for.

flags

PP_FLAG_DMA_MAP, PP_FLAG_DMA_SYNC_DEV, PP_FLAG_SYSTEM_POOL,PP_FLAG_ALLOW_UNREADABLE_NETMEM.

structpage*page_pool_dev_alloc_pages(structpage_pool*pool)

allocate a page.

Parameters

structpage_pool*pool

pool from which to allocate

Description

Get a page from the page allocator or page_pool caches.

structpage*page_pool_dev_alloc_frag(structpage_pool*pool,unsignedint*offset,unsignedintsize)

allocate a page fragment.

Parameters

structpage_pool*pool

pool from which to allocate

unsignedint*offset

offset to the allocated page

unsignedintsize

requested size

Description

Get a page fragment from the page allocator or page_pool caches.

Return

allocated page fragment, otherwise return NULL.

structpage*page_pool_dev_alloc(structpage_pool*pool,unsignedint*offset,unsignedint*size)

allocate a page or a page fragment.

Parameters

structpage_pool*pool

pool from which to allocate

unsignedint*offset

offset to the allocated page

unsignedint*size

in as the requested size, out as the allocated size

Description

Get a page or a page fragment from the page allocator or page_pool cachesdepending on the requested size in order to allocate memory with least memoryutilization and performance penalty.

Return

allocated page or page fragment, otherwise return NULL.

void*page_pool_dev_alloc_va(structpage_pool*pool,unsignedint*size)

allocate a page or a page fragment and return its va.

Parameters

structpage_pool*pool

pool from which to allocate

unsignedint*size

in as the requested size, out as the allocated size

Description

This is just a thin wrapper around thepage_pool_alloc() API, andit returns va of the allocated page or page fragment.

Return

the va for the allocated page or page fragment, otherwise return NULL.

enumdma_data_directionpage_pool_get_dma_dir(conststructpage_pool*pool)

Retrieve the stored DMA direction.

Parameters

conststructpage_pool*pool

pool from which page was allocated

Description

Get the stored dma direction. A driver might decide to store this locallyand avoid the extra cache line from page_pool to determine the direction.

voidpage_pool_put_page(structpage_pool*pool,structpage*page,unsignedintdma_sync_size,boolallow_direct)

release a reference to a page pool page

Parameters

structpage_pool*pool

pool from which page was allocated

structpage*page

page to release a reference on

unsignedintdma_sync_size

how much of the page may have been touched by the device

boolallow_direct

released by the consumer, allow lockless caching

Description

The outcome of this depends on the page refcnt. If the driver bumpsthe refcnt > 1 this will unmap the page. If the page refcnt is 1the allocator owns the page and will try to recycle it in one of the poolcaches. If PP_FLAG_DMA_SYNC_DEV is set, the page will be synced for_deviceusingdma_sync_single_range_for_device().

voidpage_pool_put_full_page(structpage_pool*pool,structpage*page,boolallow_direct)

release a reference on a page pool page

Parameters

structpage_pool*pool

pool from which page was allocated

structpage*page

page to release a reference on

boolallow_direct

released by the consumer, allow lockless caching

Description

Similar topage_pool_put_page(), but will DMA sync the entire memory areaas configured inpage_pool_params.max_len.

voidpage_pool_recycle_direct(structpage_pool*pool,structpage*page)

release a reference on a page pool page

Parameters

structpage_pool*pool

pool from which page was allocated

structpage*page

page to release a reference on

Description

Similar topage_pool_put_full_page() but caller must guarantee safe context(e.g NAPI), since it will recycle the page directly into the pool fast cache.

voidpage_pool_free_va(structpage_pool*pool,void*va,boolallow_direct)

free a va into the page_pool

Parameters

structpage_pool*pool

pool from which va was allocated

void*va

va to be freed

boolallow_direct

freed by the consumer, allow lockless caching

Description

Free a va allocated frompage_pool_allo_va().

dma_addr_tpage_pool_get_dma_addr(conststructpage*page)

Retrieve the stored DMA address.

Parameters

conststructpage*page

page allocated from a page pool

Description

Fetch the DMA address of the page. The page pool to which the page belongsmust had been created with PP_FLAG_DMA_MAP.

boolpage_pool_get_stats(conststructpage_pool*pool,structpage_pool_stats*stats)

fetch page pool stats

Parameters

conststructpage_pool*pool

pool from which page was allocated

structpage_pool_stats*stats

structpage_pool_stats to fill in

Description

Retrieve statistics about the page_pool. This API is only availableif the kernel has been configured withCONFIG_PAGE_POOL_STATS=y.A pointer to a caller allocatedstructpage_pool_stats structureis passed to this API which is filled in. The caller can then reportthose stats to the user (perhaps via ethtool, debugfs, etc.).

DMA sync

Driver is always responsible for syncing the pages for the CPU.Drivers may choose to take care of syncing for the device as wellor set thePP_FLAG_DMA_SYNC_DEV flag to request that pagesallocated from the page pool are already synced for the device.

IfPP_FLAG_DMA_SYNC_DEV is set, the driver must inform the core what portionof the buffer has to be synced. This allows the core to avoid syncing the entirepage when the drivers knows that the device only accessed a portion of the page.

Most drivers will reserve headroom in front of the frame. This partof the buffer is not touched by the device, so to avoid syncingit drivers can set theoffset field instructpage_pool_paramsappropriately.

For pages recycled on the XDP xmit and skb paths the page pool willuse themax_len member ofstructpage_pool_params to decide howmuch of the page needs to be synced (starting atoffset).When directly freeing pages in the driver (page_pool_put_page())thedma_sync_size argument specifies how much of the buffer needsto be synced.

If in doubt setoffset to 0,max_len toPAGE_SIZE andpass -1 asdma_sync_size. That combination of arguments is alwayscorrect.

Note that the syncing parameters are for the entire page.This is important to remember when using fragments (PP_FLAG_PAGE_FRAG),where allocated buffers may be smaller than a full page.Unless the driver author really understands page pool internalsit’s recommended to always useoffset=0,max_len=PAGE_SIZEwith fragmented page pools.

Stats API and structures

If the kernel is configured withCONFIG_PAGE_POOL_STATS=y, the APIpage_pool_get_stats() and structures described below are available.It takes a pointer to astructpage_pool and a pointer to astructpage_pool_stats allocated by the caller.

Older drivers expose page pool statistics via ethtool or debugfs.The same statistics are accessible via the netlink netdev familyin a driver-independent fashion.

structpage_pool_alloc_stats

allocation statistics

Definition:

struct page_pool_alloc_stats {    u64 fast;    u64 slow;    u64 slow_high_order;    u64 empty;    u64 refill;    u64 waive;};

Members

fast

successful fast path allocations

slow

slow path order-0 allocations

slow_high_order

slow path high order allocations

empty

ptr ring is empty, so a slow path allocation was forced

refill

an allocation which triggered a refill of the cache

waive

pages obtained from the ptr ring that cannot be added tothe cache due to a NUMA mismatch

structpage_pool_recycle_stats

recycling (freeing) statistics

Definition:

struct page_pool_recycle_stats {    u64 cached;    u64 cache_full;    u64 ring;    u64 ring_full;    u64 released_refcnt;};

Members

cached

recycling placed page in the page pool cache

cache_full

page pool cache was full

ring

page placed into the ptr ring

ring_full

page released from page pool because the ptr ring was full

released_refcnt

page released (and not recycled) because refcnt > 1

structpage_pool_stats

combined page pool use statistics

Definition:

struct page_pool_stats {    struct page_pool_alloc_stats alloc_stats;    struct page_pool_recycle_stats recycle_stats;};

Members

alloc_stats

seestructpage_pool_alloc_stats

recycle_stats

seestructpage_pool_recycle_stats

Description

Wrapper struct for combining page pool stats with different storagerequirements.

Coding examples

Registration

/* Page pool registration */structpage_pool_paramspp_params={0};structxdp_rxq_infoxdp_rxq;interr;pp_params.order=0;/* internal DMA mapping in page_pool */pp_params.flags=PP_FLAG_DMA_MAP;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);err=xdp_rxq_info_reg(&xdp_rxq,ndev,0);if(err)gotoerr_out;err=xdp_rxq_info_reg_mem_model(&xdp_rxq,MEM_TYPE_PAGE_POOL,page_pool);if(err)gotoerr_out;

NAPI poller

/* NAPI Rx poller */enumdma_data_directiondma_dir;dma_dir=page_pool_get_dma_dir(dring->page_pool);while(done<budget){if(someerror)page_pool_recycle_direct(page_pool,page);if(packet_is_xdp){ifXDP_DROP:page_pool_recycle_direct(page_pool,page);}else(packet_is_skb){skb_mark_for_recycle(skb);new_page=page_pool_dev_alloc_pages(page_pool);}}

Stats

#ifdef CONFIG_PAGE_POOL_STATS/* retrieve stats */structpage_pool_statsstats={0};if(page_pool_get_stats(page_pool,&stats)){/* perhaps the driver reports statistics with ethool */ethtool_print_allocation_stats(&stats.alloc_stats);ethtool_print_recycle_stats(&stats.recycle_stats);}#endif

Driver unload

/* Driver unload */page_pool_put_full_page(page_pool,page,false);xdp_rxq_info_unreg(&xdp_rxq);