Quelle  page_pool.c   Sprache: C

/* SPDX-License-Identifier: GPL-2.0
 *
 * page_pool.c
 * Author: Jesper Dangaard Brouer <netoptimizer@brouer.com>
 * Copyright (C) 2016 Red Hat, Inc.
 */

#include <linux/error-injection.h>
#include <linux/types.h>
#include <linux/kernel.h>
#include <linux/slab.h>
#include <linux/device.h>

#include <net/netdev_lock.h>
#include <net/netdev_rx_queue.h>
#include <net/page_pool/helpers.h>
#include <net/page_pool/memory_provider.h>
#include <net/xdp.h>

#include <linux/dma-direction.h>
#include <linux/dma-mapping.h>
#include <linux/page-flags.h>
#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>

#include "dev.h"
#include "mp_dmabuf_devmem.h"
#include "netmem_priv.h"
#include "page_pool_priv.h"

DEFINE_STATIC_KEY_FALSE(page_pool_mem_providers);

#define DEFER_TIME (msecs_to_jiffies(1000))
#define DEFER_WARN_INTERVAL (60 * HZ)

#define BIAS_MAX (LONG_MAX >> 1)

#ifdef CONFIG_PAGE_POOL_STATS
static DEFINE_PER_CPU(struct page_pool_recycle_stats, pp_system_recycle_stats);

/* alloc_stat_inc is intended to be used in softirq context */
#define alloc_stat_inc(pool, __stat) (pool->alloc_stats.__stat++)
/* recycle_stat_inc is safe to use when preemption is possible. */
#define recycle_stat_inc(pool, __stat)       \
 do {          \
  struct page_pool_recycle_stats __percpu *s = pool->recycle_stats; \
  this_cpu_inc(s->__stat);      \
 } while (0)

#define recycle_stat_add(pool, __stat, val)      \
 do {          \
  struct page_pool_recycle_stats __percpu *s = pool->recycle_stats; \
  this_cpu_add(s->__stat, val);      \
 } while (0)

static const char pp_stats[][ETH_GSTRING_LEN] = {
 "rx_pp_alloc_fast",
 "rx_pp_alloc_slow",
 "rx_pp_alloc_slow_ho",
 "rx_pp_alloc_empty",
 "rx_pp_alloc_refill",
 "rx_pp_alloc_waive",
 "rx_pp_recycle_cached",
 "rx_pp_recycle_cache_full",
 "rx_pp_recycle_ring",
 "rx_pp_recycle_ring_full",
 "rx_pp_recycle_released_ref",
};

/**
 * page_pool_get_stats() - fetch page pool stats
 * @pool: pool from which page was allocated
 * @stats: struct page_pool_stats to fill in
 *
 * Retrieve statistics about the page_pool. This API is only available
 * if the kernel has been configured with ``CONFIG_PAGE_POOL_STATS=y``.
 * A pointer to a caller allocated struct page_pool_stats structure
 * is passed to this API which is filled in. The caller can then report
 * those stats to the user (perhaps via ethtool, debugfs, etc.).
 */
bool page_pool_get_stats(const struct page_pool *pool,
    struct page_pool_stats *stats)
{
 int cpu = 0;

 if (!stats)
  return false;

 /* The caller is responsible to initialize stats. */
 stats->alloc_stats.fast += pool->alloc_stats.fast;
 stats->alloc_stats.slow += pool->alloc_stats.slow;
 stats->alloc_stats.slow_high_order += pool->alloc_stats.slow_high_order;
 stats->alloc_stats.empty += pool->alloc_stats.empty;
 stats->alloc_stats.refill += pool->alloc_stats.refill;
 stats->alloc_stats.waive += pool->alloc_stats.waive;

 for_each_possible_cpu(cpu) {
  const struct page_pool_recycle_stats *pcpu =
   per_cpu_ptr(pool->recycle_stats, cpu);

  stats->recycle_stats.cached += pcpu->cached;
  stats->recycle_stats.cache_full += pcpu->cache_full;
  stats->recycle_stats.ring += pcpu->ring;
  stats->recycle_stats.ring_full += pcpu->ring_full;
  stats->recycle_stats.released_refcnt += pcpu->released_refcnt;
 }

 return true;
}
EXPORT_SYMBOL(page_pool_get_stats);

u8 *page_pool_ethtool_stats_get_strings(u8 *data)
{
 int i;

 for (i = 0; i < ARRAY_SIZE(pp_stats); i++) {
  memcpy(data, pp_stats[i], ETH_GSTRING_LEN);
  data += ETH_GSTRING_LEN;
 }

 return data;
}
EXPORT_SYMBOL(page_pool_ethtool_stats_get_strings);

int page_pool_ethtool_stats_get_count(void)
{
 return ARRAY_SIZE(pp_stats);
}
EXPORT_SYMBOL(page_pool_ethtool_stats_get_count);

u64 *page_pool_ethtool_stats_get(u64 *data, const void *stats)
{
 const struct page_pool_stats *pool_stats = stats;

 *data++ = pool_stats->alloc_stats.fast;
 *data++ = pool_stats->alloc_stats.slow;
 *data++ = pool_stats->alloc_stats.slow_high_order;
 *data++ = pool_stats->alloc_stats.empty;
 *data++ = pool_stats->alloc_stats.refill;
 *data++ = pool_stats->alloc_stats.waive;
 *data++ = pool_stats->recycle_stats.cached;
 *data++ = pool_stats->recycle_stats.cache_full;
 *data++ = pool_stats->recycle_stats.ring;
 *data++ = pool_stats->recycle_stats.ring_full;
 *data++ = pool_stats->recycle_stats.released_refcnt;

 return data;
}
EXPORT_SYMBOL(page_pool_ethtool_stats_get);

#else
#define alloc_stat_inc(...) do { } while (0)
#define recycle_stat_inc(...) do { } while (0)
#define recycle_stat_add(...) do { } while (0)
#endif

static bool page_pool_producer_lock(struct page_pool *pool)
 __acquires(&pool->ring.producer_lock)
{
 bool in_softirq = in_softirq();

 if (in_softirq)
  spin_lock(&pool->ring.producer_lock);
 else
  spin_lock_bh(&pool->ring.producer_lock);

 return in_softirq;
}

static void page_pool_producer_unlock(struct page_pool *pool,
          bool in_softirq)
 __releases(&pool->ring.producer_lock)
{
 if (in_softirq)
  spin_unlock(&pool->ring.producer_lock);
 else
  spin_unlock_bh(&pool->ring.producer_lock);
}

static void page_pool_struct_check(void)
{
 CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_users);
 CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_page);
 CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_offset);
 CACHELINE_ASSERT_GROUP_SIZE(struct page_pool, frag,
        PAGE_POOL_FRAG_GROUP_ALIGN);
}

static int page_pool_init(struct page_pool *pool,
     const struct page_pool_params *params,
     int cpuid)
{
 unsigned int ring_qsize = 1024; /* Default */
 struct netdev_rx_queue *rxq;
 int err;

 page_pool_struct_check();

 memcpy(&pool->p, ¶ms->fast, sizeof(pool->p));
 memcpy(&pool->slow, ¶ms->slow, sizeof(pool->slow));

 pool->cpuid = cpuid;
 pool->dma_sync_for_cpu = true;

 /* Validate only known flags were used */
 if (pool->slow.flags & ~PP_FLAG_ALL)
  return -EINVAL;

 if (pool->p.pool_size)
  ring_qsize = min(pool->p.pool_size, 16384);

 /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
  * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
  * which is the XDP_TX use-case.
 */
 if (pool->slow.flags & PP_FLAG_DMA_MAP) {
  if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
      (pool->p.dma_dir != DMA_BIDIRECTIONAL))
   return -EINVAL;

  pool->dma_map = true;
 }

 if (pool->slow.flags & PP_FLAG_DMA_SYNC_DEV) {
  /* In order to request DMA-sync-for-device the page
   * needs to be mapped
 */
  if (!(pool->slow.flags & PP_FLAG_DMA_MAP))
   return -EINVAL;

  if (!pool->p.max_len)
   return -EINVAL;

  pool->dma_sync = true;

  /* pool->p.offset has to be set according to the address
   * offset used by the DMA engine to start copying rx data
 */
 }

 pool->has_init_callback = !!pool->slow.init_callback;

#ifdef CONFIG_PAGE_POOL_STATS
 if (!(pool->slow.flags & PP_FLAG_SYSTEM_POOL)) {
  pool->recycle_stats = alloc_percpu(struct page_pool_recycle_stats);
  if (!pool->recycle_stats)
   return -ENOMEM;
 } else {
  /* For system page pool instance we use a singular stats object
   * instead of allocating a separate percpu variable for each
   * (also percpu) page pool instance.
 */
  pool->recycle_stats = &pp_system_recycle_stats;
  pool->system = true;
 }
#endif

 if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0) {
#ifdef CONFIG_PAGE_POOL_STATS
  if (!pool->system)
   free_percpu(pool->recycle_stats);
#endif
  return -ENOMEM;
 }

 atomic_set(&pool->pages_state_release_cnt, 0);

 /* Driver calling page_pool_create() also call page_pool_destroy() */
 refcount_set(&pool->user_cnt, 1);

 xa_init_flags(&pool->dma_mapped, XA_FLAGS_ALLOC1);

 if (pool->slow.flags & PP_FLAG_ALLOW_UNREADABLE_NETMEM) {
  netdev_assert_locked(pool->slow.netdev);
  rxq = __netif_get_rx_queue(pool->slow.netdev,
        pool->slow.queue_idx);
  pool->mp_priv = rxq->mp_params.mp_priv;
  pool->mp_ops = rxq->mp_params.mp_ops;
 }

 if (pool->mp_ops) {
  if (!pool->dma_map || !pool->dma_sync) {
   err = -EOPNOTSUPP;
   goto free_ptr_ring;
  }

  if (WARN_ON(!is_kernel_rodata((unsigned long)pool->mp_ops))) {
   err = -EFAULT;
   goto free_ptr_ring;
  }

  err = pool->mp_ops->init(pool);
  if (err) {
   pr_warn("%s() mem-provider init failed %d\n", __func__,
    err);
   goto free_ptr_ring;
  }

  static_branch_inc(&page_pool_mem_providers);
 }

 return 0;

free_ptr_ring:
 ptr_ring_cleanup(&pool->ring, NULL);
#ifdef CONFIG_PAGE_POOL_STATS
 if (!pool->system)
  free_percpu(pool->recycle_stats);
#endif
 return err;
}

static void page_pool_uninit(struct page_pool *pool)
{
 ptr_ring_cleanup(&pool->ring, NULL);
 xa_destroy(&pool->dma_mapped);

#ifdef CONFIG_PAGE_POOL_STATS
 if (!pool->system)
  free_percpu(pool->recycle_stats);
#endif
}

/**
 * page_pool_create_percpu() - create a page pool for a given cpu.
 * @params: parameters, see struct page_pool_params
 * @cpuid: cpu identifier
 */
struct page_pool *
page_pool_create_percpu(const struct page_pool_params *params, int cpuid)
{
 struct page_pool *pool;
 int err;

 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
 if (!pool)
  return ERR_PTR(-ENOMEM);

 err = page_pool_init(pool, params, cpuid);
 if (err < 0)
  goto err_free;

 err = page_pool_list(pool);
 if (err)
  goto err_uninit;

 return pool;

err_uninit:
 page_pool_uninit(pool);
err_free:
 pr_warn("%s() gave up with errno %d\n", __func__, err);
 kfree(pool);
 return ERR_PTR(err);
}
EXPORT_SYMBOL(page_pool_create_percpu);

/**
 * page_pool_create() - create a page pool
 * @params: parameters, see struct page_pool_params
 */
struct page_pool *page_pool_create(const struct page_pool_params *params)
{
 return page_pool_create_percpu(params, -1);
}
EXPORT_SYMBOL(page_pool_create);

static void page_pool_return_netmem(struct page_pool *pool, netmem_ref netmem);

static noinline netmem_ref page_pool_refill_alloc_cache(struct page_pool *pool)
{
 struct ptr_ring *r = &pool->ring;
 netmem_ref netmem;
 int pref_nid; /* preferred NUMA node */

 /* Quicker fallback, avoid locks when ring is empty */
 if (__ptr_ring_empty(r)) {
  alloc_stat_inc(pool, empty);
  return 0;
 }

 /* Softirq guarantee CPU and thus NUMA node is stable. This,
  * assumes CPU refilling driver RX-ring will also run RX-NAPI.
 */
#ifdef CONFIG_NUMA
 pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
#else
 /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
 pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
#endif

 /* Refill alloc array, but only if NUMA match */
 do {
  netmem = (__force netmem_ref)__ptr_ring_consume(r);
  if (unlikely(!netmem))
   break;

  if (likely(netmem_is_pref_nid(netmem, pref_nid))) {
   pool->alloc.cache[pool->alloc.count++] = netmem;
  } else {
   /* NUMA mismatch;
    * (1) release 1 page to page-allocator and
    * (2) break out to fallthrough to alloc_pages_node.
    * This limit stress on page buddy alloactor.
 */
   page_pool_return_netmem(pool, netmem);
   alloc_stat_inc(pool, waive);
   netmem = 0;
   break;
  }
 } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);

 /* Return last page */
 if (likely(pool->alloc.count > 0)) {
  netmem = pool->alloc.cache[--pool->alloc.count];
  alloc_stat_inc(pool, refill);
 }

 return netmem;
}

/* fast path */
static netmem_ref __page_pool_get_cached(struct page_pool *pool)
{
 netmem_ref netmem;

 /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
 if (likely(pool->alloc.count)) {
  /* Fast-path */
  netmem = pool->alloc.cache[--pool->alloc.count];
  alloc_stat_inc(pool, fast);
 } else {
  netmem = page_pool_refill_alloc_cache(pool);
 }

 return netmem;
}

static void __page_pool_dma_sync_for_device(const struct page_pool *pool,
         netmem_ref netmem,
         u32 dma_sync_size)
{
#if defined(CONFIG_HAS_DMA) && defined(CONFIG_DMA_NEED_SYNC)
 dma_addr_t dma_addr = page_pool_get_dma_addr_netmem(netmem);

 dma_sync_size = min(dma_sync_size, pool->p.max_len);
 __dma_sync_single_for_device(pool->p.dev, dma_addr + pool->p.offset,
         dma_sync_size, pool->p.dma_dir);
#endif
}

static __always_inline void
page_pool_dma_sync_for_device(const struct page_pool *pool,
         netmem_ref netmem,
         u32 dma_sync_size)
{
 if (pool->dma_sync && dma_dev_need_sync(pool->p.dev)) {
  rcu_read_lock();
  /* re-check under rcu_read_lock() to sync with page_pool_scrub() */
  if (pool->dma_sync)
   __page_pool_dma_sync_for_device(pool, netmem,
       dma_sync_size);
  rcu_read_unlock();
 }
}

static int page_pool_register_dma_index(struct page_pool *pool,
     netmem_ref netmem, gfp_t gfp)
{
 int err = 0;
 u32 id;

 if (unlikely(!PP_DMA_INDEX_BITS))
  goto out;

 if (in_softirq())
  err = xa_alloc(&pool->dma_mapped, &id, netmem_to_page(netmem),
          PP_DMA_INDEX_LIMIT, gfp);
 else
  err = xa_alloc_bh(&pool->dma_mapped, &id, netmem_to_page(netmem),
      PP_DMA_INDEX_LIMIT, gfp);
 if (err) {
  WARN_ONCE(err != -ENOMEM, "couldn't track DMA mapping, please report to netdev@");
  goto out;
 }

 netmem_set_dma_index(netmem, id);
out:
 return err;
}

static int page_pool_release_dma_index(struct page_pool *pool,
           netmem_ref netmem)
{
 struct page *old, *page = netmem_to_page(netmem);
 unsigned long id;

 if (unlikely(!PP_DMA_INDEX_BITS))
  return 0;

 id = netmem_get_dma_index(netmem);
 if (!id)
  return -1;

 if (in_softirq())
  old = xa_cmpxchg(&pool->dma_mapped, id, page, NULL, 0);
 else
  old = xa_cmpxchg_bh(&pool->dma_mapped, id, page, NULL, 0);
 if (old != page)
  return -1;

 netmem_set_dma_index(netmem, 0);

 return 0;
}

static bool page_pool_dma_map(struct page_pool *pool, netmem_ref netmem, gfp_t gfp)
{
 dma_addr_t dma;
 int err;

 /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
  * since dma_addr_t can be either 32 or 64 bits and does not always fit
  * into page private data (i.e 32bit cpu with 64bit DMA caps)
  * This mapping is kept for lifetime of page, until leaving pool.
 */
 dma = dma_map_page_attrs(pool->p.dev, netmem_to_page(netmem), 0,
     (PAGE_SIZE << pool->p.order), pool->p.dma_dir,
     DMA_ATTR_SKIP_CPU_SYNC |
      DMA_ATTR_WEAK_ORDERING);
 if (dma_mapping_error(pool->p.dev, dma))
  return false;

 if (page_pool_set_dma_addr_netmem(netmem, dma)) {
  WARN_ONCE(1, "unexpected DMA address, please report to netdev@");
  goto unmap_failed;
 }

 err = page_pool_register_dma_index(pool, netmem, gfp);
 if (err)
  goto unset_failed;

 page_pool_dma_sync_for_device(pool, netmem, pool->p.max_len);

 return true;

unset_failed:
 page_pool_set_dma_addr_netmem(netmem, 0);
unmap_failed:
 dma_unmap_page_attrs(pool->p.dev, dma,
        PAGE_SIZE << pool->p.order, pool->p.dma_dir,
        DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
 return false;
}

static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
       gfp_t gfp)
{
 struct page *page;

 gfp |= __GFP_COMP;
 page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 if (unlikely(!page))
  return NULL;

 if (pool->dma_map && unlikely(!page_pool_dma_map(pool, page_to_netmem(page), gfp))) {
  put_page(page);
  return NULL;
 }

 alloc_stat_inc(pool, slow_high_order);
 page_pool_set_pp_info(pool, page_to_netmem(page));

 /* Track how many pages are held 'in-flight' */
 pool->pages_state_hold_cnt++;
 trace_page_pool_state_hold(pool, page_to_netmem(page),
       pool->pages_state_hold_cnt);
 return page;
}

/* slow path */
static noinline netmem_ref __page_pool_alloc_netmems_slow(struct page_pool *pool,
         gfp_t gfp)
{
 const int bulk = PP_ALLOC_CACHE_REFILL;
 unsigned int pp_order = pool->p.order;
 bool dma_map = pool->dma_map;
 netmem_ref netmem;
 int i, nr_pages;

 /* Unconditionally set NOWARN if allocating from NAPI.
  * Drivers forget to set it, and OOM reports on packet Rx are useless.
 */
 if ((gfp & GFP_ATOMIC) == GFP_ATOMIC)
  gfp |= __GFP_NOWARN;

 /* Don't support bulk alloc for high-order pages */
 if (unlikely(pp_order))
  return page_to_netmem(__page_pool_alloc_page_order(pool, gfp));

 /* Unnecessary as alloc cache is empty, but guarantees zero count */
 if (unlikely(pool->alloc.count > 0))
  return pool->alloc.cache[--pool->alloc.count];

 /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk */
 memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);

 nr_pages = alloc_pages_bulk_node(gfp, pool->p.nid, bulk,
      (struct page **)pool->alloc.cache);
 if (unlikely(!nr_pages))
  return 0;

 /* Pages have been filled into alloc.cache array, but count is zero and
  * page element have not been (possibly) DMA mapped.
 */
 for (i = 0; i < nr_pages; i++) {
  netmem = pool->alloc.cache[i];
  if (dma_map && unlikely(!page_pool_dma_map(pool, netmem, gfp))) {
   put_page(netmem_to_page(netmem));
   continue;
  }

  page_pool_set_pp_info(pool, netmem);
  pool->alloc.cache[pool->alloc.count++] = netmem;
  /* Track how many pages are held 'in-flight' */
  pool->pages_state_hold_cnt++;
  trace_page_pool_state_hold(pool, netmem,
        pool->pages_state_hold_cnt);
 }

 /* Return last page */
 if (likely(pool->alloc.count > 0)) {
  netmem = pool->alloc.cache[--pool->alloc.count];
  alloc_stat_inc(pool, slow);
 } else {
  netmem = 0;
 }

 /* When page just alloc'ed is should/must have refcnt 1. */
 return netmem;
}

/* For using page_pool replace: alloc_pages() API calls, but provide
 * synchronization guarantee for allocation side.
 */
netmem_ref page_pool_alloc_netmems(struct page_pool *pool, gfp_t gfp)
{
 netmem_ref netmem;

 /* Fast-path: Get a page from cache */
 netmem = __page_pool_get_cached(pool);
 if (netmem)
  return netmem;

 /* Slow-path: cache empty, do real allocation */
 if (static_branch_unlikely(&page_pool_mem_providers) && pool->mp_ops)
  netmem = pool->mp_ops->alloc_netmems(pool, gfp);
 else
  netmem = __page_pool_alloc_netmems_slow(pool, gfp);
 return netmem;
}
EXPORT_SYMBOL(page_pool_alloc_netmems);
ALLOW_ERROR_INJECTION(page_pool_alloc_netmems, NULL);

struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
{
 return netmem_to_page(page_pool_alloc_netmems(pool, gfp));
}
EXPORT_SYMBOL(page_pool_alloc_pages);

/* 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))

s32 page_pool_inflight(const struct page_pool *pool, bool strict)
{
 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);

 if (strict) {
  trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
  WARN(inflight < 0, "Negative(%d) inflight packet-pages",
       inflight);
 } else {
  inflight = max(0, inflight);
 }

 return inflight;
}

void page_pool_set_pp_info(struct page_pool *pool, netmem_ref netmem)
{
 netmem_set_pp(netmem, pool);
 netmem_or_pp_magic(netmem, PP_SIGNATURE);

 /* Ensuring all pages have been split into one fragment initially:
  * page_pool_set_pp_info() is only called once for every page when it
  * is allocated from the page allocator and page_pool_fragment_page()
  * is dirtying the same cache line as the page->pp_magic above, so
  * the overhead is negligible.
 */
 page_pool_fragment_netmem(netmem, 1);
 if (pool->has_init_callback)
  pool->slow.init_callback(netmem, pool->slow.init_arg);
}

void page_pool_clear_pp_info(netmem_ref netmem)
{
 netmem_clear_pp_magic(netmem);
 netmem_set_pp(netmem, NULL);
}

static __always_inline void __page_pool_release_netmem_dma(struct page_pool *pool,
          netmem_ref netmem)
{
 dma_addr_t dma;

 if (!pool->dma_map)
  /* Always account for inflight pages, even if we didn't
   * map them
 */
  return;

 if (page_pool_release_dma_index(pool, netmem))
  return;

 dma = page_pool_get_dma_addr_netmem(netmem);

 /* When page is unmapped, it cannot be returned to our pool */
 dma_unmap_page_attrs(pool->p.dev, dma,
        PAGE_SIZE << pool->p.order, pool->p.dma_dir,
        DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
 page_pool_set_dma_addr_netmem(netmem, 0);
}

/* 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
 * page-allocator via put_page).
 */
static void page_pool_return_netmem(struct page_pool *pool, netmem_ref netmem)
{
 int count;
 bool put;

 put = true;
 if (static_branch_unlikely(&page_pool_mem_providers) && pool->mp_ops)
  put = pool->mp_ops->release_netmem(pool, netmem);
 else
  __page_pool_release_netmem_dma(pool, netmem);

 /* 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, netmem, count);

 if (put) {
  page_pool_clear_pp_info(netmem);
  put_page(netmem_to_page(netmem));
 }
 /* An optimization would be to call __free_pages(page, pool->p.order)
  * knowing page is not part of page-cache (thus avoiding a
  * __page_cache_release() call).
 */
}

static bool page_pool_recycle_in_ring(struct page_pool *pool, netmem_ref netmem)
{
 bool in_softirq, ret;

 /* BH protection not needed if current is softirq */
 in_softirq = page_pool_producer_lock(pool);
 ret = !__ptr_ring_produce(&pool->ring, (__force void *)netmem);
 if (ret)
  recycle_stat_inc(pool, ring);
 page_pool_producer_unlock(pool, in_softirq);

 return ret;
}

/* Only allow direct recycling in special circumstances, into the
 * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 *
 * Caller must provide appropriate safe context.
 */
static bool page_pool_recycle_in_cache(netmem_ref netmem,
           struct page_pool *pool)
{
 if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) {
  recycle_stat_inc(pool, cache_full);
  return false;
 }

 /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 pool->alloc.cache[pool->alloc.count++] = netmem;
 recycle_stat_inc(pool, cached);
 return true;
}

static bool __page_pool_page_can_be_recycled(netmem_ref netmem)
{
 return netmem_is_net_iov(netmem) ||
        (page_ref_count(netmem_to_page(netmem)) == 1 &&
  !page_is_pfmemalloc(netmem_to_page(netmem)));
}

/* If the page refcnt == 1, this will try to recycle the page.
 * If pool->dma_sync is set, we'll try to sync the DMA area for
 * the configured size min(dma_sync_size, pool->max_len).
 * If the page refcnt != 1, then the page will be returned to memory
 * subsystem.
 */
static __always_inline netmem_ref
__page_pool_put_page(struct page_pool *pool, netmem_ref netmem,
       unsigned int dma_sync_size, bool allow_direct)
{
 lockdep_assert_no_hardirq();

 /* This allocator is optimized for the XDP mode that uses
  * one-frame-per-page, but have fallbacks that act like the
  * regular page allocator APIs.
  *
  * refcnt == 1 means page_pool owns page, and can recycle it.
  *
  * page is NOT reusable when allocated when system is under
  * some pressure. (page_is_pfmemalloc)
 */
 if (likely(__page_pool_page_can_be_recycled(netmem))) {
  /* Read barrier done in page_ref_count / READ_ONCE */

  page_pool_dma_sync_for_device(pool, netmem, dma_sync_size);

  if (allow_direct && page_pool_recycle_in_cache(netmem, pool))
   return 0;

  /* Page found as candidate for recycling */
  return netmem;
 }

 /* Fallback/non-XDP mode: API user have elevated refcnt.
  *
  * Many drivers split up the page into fragments, and some
  * want to keep doing this to save memory and do refcnt based
  * recycling. Support this use case too, to ease drivers
  * switching between XDP/non-XDP.
  *
  * In-case page_pool maintains the DMA mapping, API user must
  * call page_pool_put_page once.  In this elevated refcnt
  * case, the DMA is unmapped/released, as driver is likely
  * doing refcnt based recycle tricks, meaning another process
  * will be invoking put_page.
 */
 recycle_stat_inc(pool, released_refcnt);
 page_pool_return_netmem(pool, netmem);

 return 0;
}

static bool page_pool_napi_local(const struct page_pool *pool)
{
 const struct napi_struct *napi;
 u32 cpuid;

 /* On PREEMPT_RT the softirq can be preempted by the consumer */
 if (IS_ENABLED(CONFIG_PREEMPT_RT))
  return false;

 if (unlikely(!in_softirq()))
  return false;

 /* 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.
  * __page_pool_put_page() makes sure we're not in hardirq context
  * and interrupts are enabled prior to accessing the cache.
 */
 cpuid = smp_processor_id();
 if (READ_ONCE(pool->cpuid) == cpuid)
  return true;

 napi = READ_ONCE(pool->p.napi);

 return napi && READ_ONCE(napi->list_owner) == cpuid;
}

void page_pool_put_unrefed_netmem(struct page_pool *pool, netmem_ref netmem,
      unsigned int dma_sync_size, bool allow_direct)
{
 if (!allow_direct)
  allow_direct = page_pool_napi_local(pool);

 netmem = __page_pool_put_page(pool, netmem, dma_sync_size,
          allow_direct);
 if (netmem && !page_pool_recycle_in_ring(pool, netmem)) {
  /* Cache full, fallback to free pages */
  recycle_stat_inc(pool, ring_full);
  page_pool_return_netmem(pool, netmem);
 }
}
EXPORT_SYMBOL(page_pool_put_unrefed_netmem);

void page_pool_put_unrefed_page(struct page_pool *pool, struct page *page,
    unsigned int dma_sync_size, bool allow_direct)
{
 page_pool_put_unrefed_netmem(pool, page_to_netmem(page), dma_sync_size,
         allow_direct);
}
EXPORT_SYMBOL(page_pool_put_unrefed_page);

static void page_pool_recycle_ring_bulk(struct page_pool *pool,
     netmem_ref *bulk,
     u32 bulk_len)
{
 bool in_softirq;
 u32 i;

 /* Bulk produce into ptr_ring page_pool cache */
 in_softirq = page_pool_producer_lock(pool);

 for (i = 0; i < bulk_len; i++) {
  if (__ptr_ring_produce(&pool->ring, (__force void *)bulk[i])) {
   /* ring full */
   recycle_stat_inc(pool, ring_full);
   break;
  }
 }

 page_pool_producer_unlock(pool, in_softirq);
 recycle_stat_add(pool, ring, i);

 /* Hopefully all pages were returned into ptr_ring */
 if (likely(i == bulk_len))
  return;

 /*
  * ptr_ring cache is 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_netmem(pool, bulk[i]);
}

/**
 * page_pool_put_netmem_bulk() - release references on multiple netmems
 * @data: array holding netmem references
 * @count: number of entries in @data
 *
 * Tries to refill a number of netmems into the ptr_ring cache holding ptr_ring
 * producer lock. If the ptr_ring is full, page_pool_put_netmem_bulk()
 * will release leftover netmems to the memory provider.
 * page_pool_put_netmem_bulk() is suitable to be run inside the driver NAPI tx
 * completion loop for the XDP_REDIRECT use case.
 *
 * Please note the caller must not use data area after running
 * page_pool_put_netmem_bulk(), as this function overwrites it.
 */
void page_pool_put_netmem_bulk(netmem_ref *data, u32 count)
{
 u32 bulk_len = 0;

 for (u32 i = 0; i < count; i++) {
  netmem_ref netmem = netmem_compound_head(data[i]);

  if (page_pool_unref_and_test(netmem))
   data[bulk_len++] = netmem;
 }

 count = bulk_len;
 while (count) {
  netmem_ref bulk[XDP_BULK_QUEUE_SIZE];
  struct page_pool *pool = NULL;
  bool allow_direct;
  u32 foreign = 0;

  bulk_len = 0;

  for (u32 i = 0; i < count; i++) {
   struct page_pool *netmem_pp;
   netmem_ref netmem = data[i];

   netmem_pp = netmem_get_pp(netmem);
   if (unlikely(!pool)) {
    pool = netmem_pp;
    allow_direct = page_pool_napi_local(pool);
   } else if (netmem_pp != pool) {
    /*
     * If the netmem belongs to a different
     * page_pool, save it for another round.
 */
    data[foreign++] = netmem;
    continue;
   }

   netmem = __page_pool_put_page(pool, netmem, -1,
            allow_direct);
   /* Approved for bulk recycling in ptr_ring cache */
   if (netmem)
    bulk[bulk_len++] = netmem;
  }

  if (bulk_len)
   page_pool_recycle_ring_bulk(pool, bulk, bulk_len);

  count = foreign;
 }
}
EXPORT_SYMBOL(page_pool_put_netmem_bulk);

static netmem_ref page_pool_drain_frag(struct page_pool *pool,
           netmem_ref netmem)
{
 long drain_count = BIAS_MAX - pool->frag_users;

 /* Some user is still using the page frag */
 if (likely(page_pool_unref_netmem(netmem, drain_count)))
  return 0;

 if (__page_pool_page_can_be_recycled(netmem)) {
  page_pool_dma_sync_for_device(pool, netmem, -1);
  return netmem;
 }

 page_pool_return_netmem(pool, netmem);
 return 0;
}

static void page_pool_free_frag(struct page_pool *pool)
{
 long drain_count = BIAS_MAX - pool->frag_users;
 netmem_ref netmem = pool->frag_page;

 pool->frag_page = 0;

 if (!netmem || page_pool_unref_netmem(netmem, drain_count))
  return;

 page_pool_return_netmem(pool, netmem);
}

netmem_ref page_pool_alloc_frag_netmem(struct page_pool *pool,
           unsigned int *offset, unsigned int size,
           gfp_t gfp)
{
 unsigned int max_size = PAGE_SIZE << pool->p.order;
 netmem_ref netmem = pool->frag_page;

 if (WARN_ON(size > max_size))
  return 0;

 size = ALIGN(size, dma_get_cache_alignment());
 *offset = pool->frag_offset;

 if (netmem && *offset + size > max_size) {
  netmem = page_pool_drain_frag(pool, netmem);
  if (netmem) {
   recycle_stat_inc(pool, cached);
   alloc_stat_inc(pool, fast);
   goto frag_reset;
  }
 }

 if (!netmem) {
  netmem = page_pool_alloc_netmems(pool, gfp);
  if (unlikely(!netmem)) {
   pool->frag_page = 0;
   return 0;
  }

  pool->frag_page = netmem;

frag_reset:
  pool->frag_users = 1;
  *offset = 0;
  pool->frag_offset = size;
  page_pool_fragment_netmem(netmem, BIAS_MAX);
  return netmem;
 }

 pool->frag_users++;
 pool->frag_offset = *offset + size;
 return netmem;
}
EXPORT_SYMBOL(page_pool_alloc_frag_netmem);

struct page *page_pool_alloc_frag(struct page_pool *pool, unsigned int *offset,
      unsigned int size, gfp_t gfp)
{
 return netmem_to_page(page_pool_alloc_frag_netmem(pool, offset, size,
         gfp));
}
EXPORT_SYMBOL(page_pool_alloc_frag);

static void page_pool_empty_ring(struct page_pool *pool)
{
 netmem_ref netmem;

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

  page_pool_return_netmem(pool, netmem);
 }
}

static void __page_pool_destroy(struct page_pool *pool)
{
 if (pool->disconnect)
  pool->disconnect(pool);

 page_pool_unlist(pool);
 page_pool_uninit(pool);

 if (pool->mp_ops) {
  pool->mp_ops->destroy(pool);
  static_branch_dec(&page_pool_mem_providers);
 }

 kfree(pool);
}

static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
{
 netmem_ref netmem;

 if (pool->destroy_cnt)
  return;

 /* Empty alloc cache, assume caller made sure this is
  * no-longer in use, and page_pool_alloc_pages() cannot be
  * call concurrently.
 */
 while (pool->alloc.count) {
  netmem = pool->alloc.cache[--pool->alloc.count];
  page_pool_return_netmem(pool, netmem);
 }
}

static void page_pool_scrub(struct page_pool *pool)
{
 unsigned long id;
 void *ptr;

 page_pool_empty_alloc_cache_once(pool);
 if (!pool->destroy_cnt++ && pool->dma_map) {
  if (pool->dma_sync) {
   /* Disable page_pool_dma_sync_for_device() */
   pool->dma_sync = false;

   /* Make sure all concurrent returns that may see the old
    * value of dma_sync (and thus perform a sync) have
    * finished before doing the unmapping below. Skip the
    * wait if the device doesn't actually need syncing, or
    * if there are no outstanding mapped pages.
 */
   if (dma_dev_need_sync(pool->p.dev) &&
       !xa_empty(&pool->dma_mapped))
    synchronize_net();
  }

  xa_for_each(&pool->dma_mapped, id, ptr)
   __page_pool_release_netmem_dma(pool, page_to_netmem((struct page *)ptr));
 }

 /* No more consumers should exist, but producers could still
  * be in-flight.
 */
 page_pool_empty_ring(pool);
}

static int page_pool_release(struct page_pool *pool)
{
 bool in_softirq;
 int inflight;

 page_pool_scrub(pool);
 inflight = page_pool_inflight(pool, true);
 /* Acquire producer lock to make sure producers have exited. */
 in_softirq = page_pool_producer_lock(pool);
 page_pool_producer_unlock(pool, in_softirq);
 if (!inflight)
  __page_pool_destroy(pool);

 return inflight;
}

static void page_pool_release_retry(struct work_struct *wq)
{
 struct delayed_work *dwq = to_delayed_work(wq);
 struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 void *netdev;
 int inflight;

 inflight = page_pool_release(pool);
 /* In rare cases, a driver bug may cause inflight to go negative.
  * Don't reschedule release if inflight is 0 or negative.
  * - If 0, the page_pool has been destroyed
  * - if negative, we will never recover
  * in both cases no reschedule is necessary.
 */
 if (inflight <= 0)
  return;

 /* Periodic warning for page pools the user can't see */
 netdev = READ_ONCE(pool->slow.netdev);
 if (time_after_eq(jiffies, pool->defer_warn) &&
     (!netdev || netdev == NET_PTR_POISON)) {
  int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;

  pr_warn("%s() stalled pool shutdown: id %u, %d inflight %d sec\n",
   __func__, pool->user.id, inflight, sec);
  pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 }

 /* Still not ready to be disconnected, retry later */
 schedule_delayed_work(&pool->release_dw, DEFER_TIME);
}

void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
      const struct xdp_mem_info *mem)
{
 refcount_inc(&pool->user_cnt);
 pool->disconnect = disconnect;
 pool->xdp_mem_id = mem->id;
}

/**
 * page_pool_enable_direct_recycling() - mark page pool as owned by NAPI
 * @pool: page pool to modify
 * @napi: NAPI instance to associate the page pool with
 *
 * Associate a page pool with a NAPI instance for lockless page recycling.
 * This is useful when a new page pool has to be added to a NAPI instance
 * without disabling that NAPI instance, to mark the point at which control
 * path "hands over" the page pool to the NAPI instance. In most cases driver
 * can simply set the @napi field in struct page_pool_params, and does not
 * have to call this helper.
 *
 * The function is idempotent, but does not implement any refcounting.
 * Single page_pool_disable_direct_recycling() will disable recycling,
 * no matter how many times enable was called.
 */
void page_pool_enable_direct_recycling(struct page_pool *pool,
           struct napi_struct *napi)
{
 if (READ_ONCE(pool->p.napi) == napi)
  return;
 WARN_ON(!napi || pool->p.napi);

 mutex_lock(&page_pools_lock);
 WRITE_ONCE(pool->p.napi, napi);
 mutex_unlock(&page_pools_lock);
}
EXPORT_SYMBOL(page_pool_enable_direct_recycling);

void page_pool_disable_direct_recycling(struct page_pool *pool)
{
 /* Disable direct recycling based on pool->cpuid.
  * Paired with READ_ONCE() in page_pool_napi_local().
 */
 WRITE_ONCE(pool->cpuid, -1);

 if (!pool->p.napi)
  return;

 napi_assert_will_not_race(pool->p.napi);

 mutex_lock(&page_pools_lock);
 WRITE_ONCE(pool->p.napi, NULL);
 mutex_unlock(&page_pools_lock);
}
EXPORT_SYMBOL(page_pool_disable_direct_recycling);

void page_pool_destroy(struct page_pool *pool)
{
 if (!pool)
  return;

 if (!page_pool_put(pool))
  return;

 page_pool_disable_direct_recycling(pool);
 page_pool_free_frag(pool);

 if (!page_pool_release(pool))
  return;

 page_pool_detached(pool);
 pool->defer_start = jiffies;
 pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;

 INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 schedule_delayed_work(&pool->release_dw, DEFER_TIME);
}
EXPORT_SYMBOL(page_pool_destroy);

/* Caller must provide appropriate safe context, e.g. NAPI. */
void page_pool_update_nid(struct page_pool *pool, int new_nid)
{
 netmem_ref netmem;

 trace_page_pool_update_nid(pool, new_nid);
 pool->p.nid = new_nid;

 /* Flush pool alloc cache, as refill will check NUMA node */
 while (pool->alloc.count) {
  netmem = pool->alloc.cache[--pool->alloc.count];
  page_pool_return_netmem(pool, netmem);
 }
}
EXPORT_SYMBOL(page_pool_update_nid);

bool net_mp_niov_set_dma_addr(struct net_iov *niov, dma_addr_t addr)
{
 return page_pool_set_dma_addr_netmem(net_iov_to_netmem(niov), addr);
}

/* Associate a niov with a page pool. Should follow with a matching
 * net_mp_niov_clear_page_pool()
 */
void net_mp_niov_set_page_pool(struct page_pool *pool, struct net_iov *niov)
{
 netmem_ref netmem = net_iov_to_netmem(niov);

 page_pool_set_pp_info(pool, netmem);

 pool->pages_state_hold_cnt++;
 trace_page_pool_state_hold(pool, netmem, pool->pages_state_hold_cnt);
}

/* Disassociate a niov from a page pool. Should only be used in the
 * ->release_netmem() path.
 */
void net_mp_niov_clear_page_pool(struct net_iov *niov)
{
 netmem_ref netmem = net_iov_to_netmem(niov);

 page_pool_clear_pp_info(netmem);
}