include/net/page_pool.h

   1 /* SPDX-License-Identifier: GPL-2.0
   2  *
   3  * page_pool.h
   4  *      Author: Jesper Dangaard Brouer <netoptimizer@brouer.com>
   5  *      Copyright (C) 2016 Red Hat, Inc.
   6  */
   7
   8 /**
   9  * DOC: page_pool allocator
  10  *
  11  * This page_pool allocator is optimized for the XDP mode that
  12  * uses one-frame-per-page, but have fallbacks that act like the
  13  * regular page allocator APIs.
  14  *
  15  * Basic use involve replacing alloc_pages() calls with the
  16  * page_pool_alloc_pages() call.  Drivers should likely use
  17  * page_pool_dev_alloc_pages() replacing dev_alloc_pages().
  18  *
  19  * If page_pool handles DMA mapping (use page->private), then API user
  20  * is responsible for invoking page_pool_put_page() once.  In-case of
  21  * elevated refcnt, the DMA state is released, assuming other users of
  22  * the page will eventually call put_page().
  23  *
  24  * If no DMA mapping is done, then it can act as shim-layer that
  25  * fall-through to alloc_page.  As no state is kept on the page, the
  26  * regular put_page() call is sufficient.
  27  */
  28 #ifndef _NET_PAGE_POOL_H
  29 #define _NET_PAGE_POOL_H
  30
  31 #include <linux/mm.h> /* Needed by ptr_ring */
  32 #include <linux/ptr_ring.h>
  33 #include <linux/dma-direction.h>
  34
  35 #define PP_FLAG_DMA_MAP 1 /* Should page_pool do the DMA map/unmap */
  36 #define PP_FLAG_ALL     PP_FLAG_DMA_MAP
  37
  38 /*
  39  * Fast allocation side cache array/stack
  40  *
  41  * The cache size and refill watermark is related to the network
  42  * use-case.  The NAPI budget is 64 packets.  After a NAPI poll the RX
  43  * ring is usually refilled and the max consumed elements will be 64,
  44  * thus a natural max size of objects needed in the cache.
  45  *
  46  * Keeping room for more objects, is due to XDP_DROP use-case.  As
  47  * XDP_DROP allows the opportunity to recycle objects directly into
  48  * this array, as it shares the same softirq/NAPI protection.  If
  49  * cache is already full (or partly full) then the XDP_DROP recycles
  50  * would have to take a slower code path.
  51  */
  52 #define PP_ALLOC_CACHE_SIZE     128
  53 #define PP_ALLOC_CACHE_REFILL   64
  54 struct pp_alloc_cache {
  55         u32 count;
  56         void *cache[PP_ALLOC_CACHE_SIZE];
  57 };
  58
  59 struct page_pool_params {
  60         unsigned int    flags;
  61         unsigned int    order;
  62         unsigned int    pool_size;
  63         int             nid;  /* Numa node id to allocate from pages from */
  64         struct device   *dev; /* device, for DMA pre-mapping purposes */
  65         enum dma_data_direction dma_dir; /* DMA mapping direction */
  66 };
  67
  68 struct page_pool {
  69         struct rcu_head rcu;
  70         struct page_pool_params p;
  71
  72         /*
  73          * Data structure for allocation side
  74          *
  75          * Drivers allocation side usually already perform some kind
  76          * of resource protection.  Piggyback on this protection, and
  77          * require driver to protect allocation side.
  78          *
  79          * For NIC drivers this means, allocate a page_pool per
  80          * RX-queue. As the RX-queue is already protected by
  81          * Softirq/BH scheduling and napi_schedule. NAPI schedule
  82          * guarantee that a single napi_struct will only be scheduled
  83          * on a single CPU (see napi_schedule).
  84          */
  85         struct pp_alloc_cache alloc ____cacheline_aligned_in_smp;
  86
  87         /* Data structure for storing recycled pages.
  88          *
  89          * Returning/freeing pages is more complicated synchronization
  90          * wise, because free's can happen on remote CPUs, with no
  91          * association with allocation resource.
  92          *
  93          * Use ptr_ring, as it separates consumer and producer
  94          * effeciently, it a way that doesn't bounce cache-lines.
  95          *
  96          * TODO: Implement bulk return pages into this structure.
  97          */
  98         struct ptr_ring ring;
  99 };
 100
 101 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp);
 102
 103 static inline struct page *page_pool_dev_alloc_pages(struct page_pool *pool)
 104 {
 105         gfp_t gfp = (GFP_ATOMIC | __GFP_NOWARN);
 106
 107         return page_pool_alloc_pages(pool, gfp);
 108 }
 109
 110 struct page_pool *page_pool_create(const struct page_pool_params *params);
 111
 112 void page_pool_destroy(struct page_pool *pool);
 113
 114 /* Never call this directly, use helpers below */
 115 void __page_pool_put_page(struct page_pool *pool,
 116                           struct page *page, bool allow_direct);
 117
 118 static inline void page_pool_put_page(struct page_pool *pool,
 119                                       struct page *page, bool allow_direct)
 120 {
 121         /* When page_pool isn't compiled-in, net/core/xdp.c doesn't
 122          * allow registering MEM_TYPE_PAGE_POOL, but shield linker.
 123          */
 124 #ifdef CONFIG_PAGE_POOL
 125         __page_pool_put_page(pool, page, allow_direct);
 126 #endif
 127 }
 128 /* Very limited use-cases allow recycle direct */
 129 static inline void page_pool_recycle_direct(struct page_pool *pool,
 130                                             struct page *page)
 131 {
 132         __page_pool_put_page(pool, page, true);
 133 }
 134
 135 static inline bool is_page_pool_compiled_in(void)
 136 {
 137 #ifdef CONFIG_PAGE_POOL
 138         return true;
 139 #else
 140         return false;
 141 #endif
 142 }
 143
 144 #endif /* _NET_PAGE_POOL_H */