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 * API keeps track of in-flight pages, in-order to let API user know
20 * when it is safe to dealloactor page_pool object.  Thus, API users
21 * must make sure to call page_pool_release_page() when a page is
22 * "leaving" the page_pool.  Or call page_pool_put_page() where
23 * appropiate.  For maintaining correct accounting.
24 *
25 * API user must only call page_pool_put_page() once on a page, as it
26 * will either recycle the page, or in case of elevated refcnt, it
27 * will release the DMA mapping and in-flight state accounting.  We
28 * hope to lift this requirement in the future.
29 */
30#ifndef _NET_PAGE_POOL_H
31#define _NET_PAGE_POOL_H
32
33#include <linux/mm.h> /* Needed by ptr_ring */
34#include <linux/ptr_ring.h>
35#include <linux/dma-direction.h>
36
37#define PP_FLAG_DMA_MAP		BIT(0) /* Should page_pool do the DMA
38					* map/unmap
39					*/
40#define PP_FLAG_DMA_SYNC_DEV	BIT(1) /* If set all pages that the driver gets
41					* from page_pool will be
42					* DMA-synced-for-device according to
43					* the length provided by the device
44					* driver.
45					* Please note DMA-sync-for-CPU is still
46					* device driver responsibility
47					*/
48#define PP_FLAG_ALL		(PP_FLAG_DMA_MAP | PP_FLAG_DMA_SYNC_DEV)
49
50/*
51 * Fast allocation side cache array/stack
52 *
53 * The cache size and refill watermark is related to the network
54 * use-case.  The NAPI budget is 64 packets.  After a NAPI poll the RX
55 * ring is usually refilled and the max consumed elements will be 64,
56 * thus a natural max size of objects needed in the cache.
57 *
58 * Keeping room for more objects, is due to XDP_DROP use-case.  As
59 * XDP_DROP allows the opportunity to recycle objects directly into
60 * this array, as it shares the same softirq/NAPI protection.  If
61 * cache is already full (or partly full) then the XDP_DROP recycles
62 * would have to take a slower code path.
63 */
64#define PP_ALLOC_CACHE_SIZE	128
65#define PP_ALLOC_CACHE_REFILL	64
66struct pp_alloc_cache {
67	u32 count;
68	void *cache[PP_ALLOC_CACHE_SIZE];
69};
70
71struct page_pool_params {
72	unsigned int	flags;
73	unsigned int	order;
74	unsigned int	pool_size;
75	int		nid;  /* Numa node id to allocate from pages from */
76	struct device	*dev; /* device, for DMA pre-mapping purposes */
77	enum dma_data_direction dma_dir; /* DMA mapping direction */
78	unsigned int	max_len; /* max DMA sync memory size */
79	unsigned int	offset;  /* DMA addr offset */
80};
81
82struct page_pool {
83	struct page_pool_params p;
84
85	struct delayed_work release_dw;
86	void (*disconnect)(void *);
87	unsigned long defer_start;
88	unsigned long defer_warn;
89
90	u32 pages_state_hold_cnt;
91
92	/*
93	 * Data structure for allocation side
94	 *
95	 * Drivers allocation side usually already perform some kind
96	 * of resource protection.  Piggyback on this protection, and
97	 * require driver to protect allocation side.
98	 *
99	 * For NIC drivers this means, allocate a page_pool per
100	 * RX-queue. As the RX-queue is already protected by
101	 * Softirq/BH scheduling and napi_schedule. NAPI schedule
102	 * guarantee that a single napi_struct will only be scheduled
103	 * on a single CPU (see napi_schedule).
104	 */
105	struct pp_alloc_cache alloc ____cacheline_aligned_in_smp;
106
107	/* Data structure for storing recycled pages.
108	 *
109	 * Returning/freeing pages is more complicated synchronization
110	 * wise, because free's can happen on remote CPUs, with no
111	 * association with allocation resource.
112	 *
113	 * Use ptr_ring, as it separates consumer and producer
114	 * effeciently, it a way that doesn't bounce cache-lines.
115	 *
116	 * TODO: Implement bulk return pages into this structure.
117	 */
118	struct ptr_ring ring;
119
120	atomic_t pages_state_release_cnt;
121
122	/* A page_pool is strictly tied to a single RX-queue being
123	 * protected by NAPI, due to above pp_alloc_cache. This
124	 * refcnt serves purpose is to simplify drivers error handling.
125	 */
126	refcount_t user_cnt;
127
128	u64 destroy_cnt;
129};
130
131struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp);
132
133static inline struct page *page_pool_dev_alloc_pages(struct page_pool *pool)
134{
135	gfp_t gfp = (GFP_ATOMIC | __GFP_NOWARN);
136
137	return page_pool_alloc_pages(pool, gfp);
138}
139
140/* get the stored dma direction. A driver might decide to treat this locally and
141 * avoid the extra cache line from page_pool to determine the direction
142 */
143static
144inline enum dma_data_direction page_pool_get_dma_dir(struct page_pool *pool)
145{
146	return pool->p.dma_dir;
147}
148
149struct page_pool *page_pool_create(const struct page_pool_params *params);
150
151#ifdef CONFIG_PAGE_POOL
152void page_pool_destroy(struct page_pool *pool);
153void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *));
154void page_pool_release_page(struct page_pool *pool, struct page *page);
155#else
156static inline void page_pool_destroy(struct page_pool *pool)
157{
158}
159
160static inline void page_pool_use_xdp_mem(struct page_pool *pool,
161					 void (*disconnect)(void *))
162{
163}
164static inline void page_pool_release_page(struct page_pool *pool,
165					  struct page *page)
166{
167}
168#endif
169
170void page_pool_put_page(struct page_pool *pool, struct page *page,
171			unsigned int dma_sync_size, bool allow_direct);
172
173/* Same as above but will try to sync the entire area pool->max_len */
174static inline void page_pool_put_full_page(struct page_pool *pool,
175					   struct page *page, bool allow_direct)
176{
177	/* When page_pool isn't compiled-in, net/core/xdp.c doesn't
178	 * allow registering MEM_TYPE_PAGE_POOL, but shield linker.
179	 */
180#ifdef CONFIG_PAGE_POOL
181	page_pool_put_page(pool, page, -1, allow_direct);
182#endif
183}
184
185/* Same as above but the caller must guarantee safe context. e.g NAPI */
186static inline void page_pool_recycle_direct(struct page_pool *pool,
187					    struct page *page)
188{
189	page_pool_put_full_page(pool, page, true);
190}
191
192static inline dma_addr_t page_pool_get_dma_addr(struct page *page)
193{
194	dma_addr_t ret = page->dma_addr[0];
195	if (sizeof(dma_addr_t) > sizeof(unsigned long))
196		ret |= (dma_addr_t)page->dma_addr[1] << 16 << 16;
197	return ret;
198}
199
200static inline void page_pool_set_dma_addr(struct page *page, dma_addr_t addr)
201{
202	page->dma_addr[0] = addr;
203	if (sizeof(dma_addr_t) > sizeof(unsigned long))
204		page->dma_addr[1] = upper_32_bits(addr);
205}
206
207static inline bool is_page_pool_compiled_in(void)
208{
209#ifdef CONFIG_PAGE_POOL
210	return true;
211#else
212	return false;
213#endif
214}
215
216static inline bool page_pool_put(struct page_pool *pool)
217{
218	return refcount_dec_and_test(&pool->user_cnt);
219}
220
221/* Caller must provide appropriate safe context, e.g. NAPI. */
222void page_pool_update_nid(struct page_pool *pool, int new_nid);
223static inline void page_pool_nid_changed(struct page_pool *pool, int new_nid)
224{
225	if (unlikely(pool->p.nid != new_nid))
226		page_pool_update_nid(pool, new_nid);
227}
228#endif /* _NET_PAGE_POOL_H */
229