Quelle  page_owner.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
#include <linux/debugfs.h>
#include <linux/mm.h>
#include <linux/slab.h>
#include <linux/uaccess.h>
#include <linux/memblock.h>
#include <linux/stacktrace.h>
#include <linux/page_owner.h>
#include <linux/jump_label.h>
#include <linux/migrate.h>
#include <linux/stackdepot.h>
#include <linux/seq_file.h>
#include <linux/memcontrol.h>
#include <linux/sched/clock.h>

#include "internal.h"

/*
 * TODO: teach PAGE_OWNER_STACK_DEPTH (__dump_page_owner and save_stack)
 * to use off stack temporal storage
 */
#define PAGE_OWNER_STACK_DEPTH (16)

struct page_owner {
 unsigned short order;
 short last_migrate_reason;
 gfp_t gfp_mask;
 depot_stack_handle_t handle;
 depot_stack_handle_t free_handle;
 u64 ts_nsec;
 u64 free_ts_nsec;
 char comm[TASK_COMM_LEN];
 pid_t pid;
 pid_t tgid;
 pid_t free_pid;
 pid_t free_tgid;
};

struct stack {
 struct stack_record *stack_record;
 struct stack *next;
};
static struct stack dummy_stack;
static struct stack failure_stack;
static struct stack *stack_list;
static DEFINE_SPINLOCK(stack_list_lock);

static bool page_owner_enabled __initdata;
DEFINE_STATIC_KEY_FALSE(page_owner_inited);

static depot_stack_handle_t dummy_handle;
static depot_stack_handle_t failure_handle;
static depot_stack_handle_t early_handle;

static void init_early_allocated_pages(void);

static inline void set_current_in_page_owner(void)
{
 /*
 * Avoid recursion.
 *
 * We might need to allocate more memory from page_owner code, so make
 * sure to signal it in order to avoid recursion.
 */
 current->in_page_owner = 1;
}

static inline void unset_current_in_page_owner(void)
{
 current->in_page_owner = 0;
}

static int __init early_page_owner_param(char *buf)
{
 int ret = kstrtobool(buf, &page_owner_enabled);

 if (page_owner_enabled)
  stack_depot_request_early_init();

 return ret;
}
early_param("page_owner", early_page_owner_param);

static __init bool need_page_owner(void)
{
 return page_owner_enabled;
}

static __always_inline depot_stack_handle_t create_dummy_stack(void)
{
 unsigned long entries[4];
 unsigned int nr_entries;

 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 0);
 return stack_depot_save(entries, nr_entries, GFP_KERNEL);
}

static noinline void register_dummy_stack(void)
{
 dummy_handle = create_dummy_stack();
}

static noinline void register_failure_stack(void)
{
 failure_handle = create_dummy_stack();
}

static noinline void register_early_stack(void)
{
 early_handle = create_dummy_stack();
}

static __init void init_page_owner(void)
{
 if (!page_owner_enabled)
  return;

 register_dummy_stack();
 register_failure_stack();
 register_early_stack();
 init_early_allocated_pages();
 /* Initialize dummy and failure stacks and link them to stack_list */
 dummy_stack.stack_record = __stack_depot_get_stack_record(dummy_handle);
 failure_stack.stack_record = __stack_depot_get_stack_record(failure_handle);
 if (dummy_stack.stack_record)
  refcount_set(&dummy_stack.stack_record->count, 1);
 if (failure_stack.stack_record)
  refcount_set(&failure_stack.stack_record->count, 1);
 dummy_stack.next = &failure_stack;
 stack_list = &dummy_stack;
 static_branch_enable(&page_owner_inited);
}

struct page_ext_operations page_owner_ops = {
 .size = sizeof(struct page_owner),
 .need = need_page_owner,
 .init = init_page_owner,
 .need_shared_flags = true,
};

static inline struct page_owner *get_page_owner(struct page_ext *page_ext)
{
 return page_ext_data(page_ext, &page_owner_ops);
}

static noinline depot_stack_handle_t save_stack(gfp_t flags)
{
 unsigned long entries[PAGE_OWNER_STACK_DEPTH];
 depot_stack_handle_t handle;
 unsigned int nr_entries;

 if (current->in_page_owner)
  return dummy_handle;

 set_current_in_page_owner();
 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 2);
 handle = stack_depot_save(entries, nr_entries, flags);
 if (!handle)
  handle = failure_handle;
 unset_current_in_page_owner();

 return handle;
}

static void add_stack_record_to_list(struct stack_record *stack_record,
         gfp_t gfp_mask)
{
 unsigned long flags;
 struct stack *stack;

 if (!gfpflags_allow_spinning(gfp_mask))
  return;

 set_current_in_page_owner();
 stack = kmalloc(sizeof(*stack), gfp_nested_mask(gfp_mask));
 if (!stack) {
  unset_current_in_page_owner();
  return;
 }
 unset_current_in_page_owner();

 stack->stack_record = stack_record;
 stack->next = NULL;

 spin_lock_irqsave(&stack_list_lock, flags);
 stack->next = stack_list;
 /*
 * This pairs with smp_load_acquire() from function
 * stack_start(). This guarantees that stack_start()
 * will see an updated stack_list before starting to
 * traverse the list.
 */
 smp_store_release(&stack_list, stack);
 spin_unlock_irqrestore(&stack_list_lock, flags);
}

static void inc_stack_record_count(depot_stack_handle_t handle, gfp_t gfp_mask,
       int nr_base_pages)
{
 struct stack_record *stack_record = __stack_depot_get_stack_record(handle);

 if (!stack_record)
  return;

 /*
 * New stack_record's that do not use STACK_DEPOT_FLAG_GET start
 * with REFCOUNT_SATURATED to catch spurious increments of their
 * refcount.
 * Since we do not use STACK_DEPOT_FLAG_GET API, let us
 * set a refcount of 1 ourselves.
 */
 if (refcount_read(&stack_record->count) == REFCOUNT_SATURATED) {
  int old = REFCOUNT_SATURATED;

  if (atomic_try_cmpxchg_relaxed(&stack_record->count.refs, &old, 1))
   /* Add the new stack_record to our list */
   add_stack_record_to_list(stack_record, gfp_mask);
 }
 refcount_add(nr_base_pages, &stack_record->count);
}

static void dec_stack_record_count(depot_stack_handle_t handle,
       int nr_base_pages)
{
 struct stack_record *stack_record = __stack_depot_get_stack_record(handle);

 if (!stack_record)
  return;

 if (refcount_sub_and_test(nr_base_pages, &stack_record->count))
  pr_warn("%s: refcount went to 0 for %u handle\n", __func__,
   handle);
}

static inline void __update_page_owner_handle(struct page *page,
           depot_stack_handle_t handle,
           unsigned short order,
           gfp_t gfp_mask,
           short last_migrate_reason, u64 ts_nsec,
           pid_t pid, pid_t tgid, char *comm)
{
 struct page_ext_iter iter;
 struct page_ext *page_ext;
 struct page_owner *page_owner;

 rcu_read_lock();
 for_each_page_ext(page, 1 << order, page_ext, iter) {
  page_owner = get_page_owner(page_ext);
  page_owner->handle = handle;
  page_owner->order = order;
  page_owner->gfp_mask = gfp_mask;
  page_owner->last_migrate_reason = last_migrate_reason;
  page_owner->pid = pid;
  page_owner->tgid = tgid;
  page_owner->ts_nsec = ts_nsec;
  strscpy(page_owner->comm, comm,
   sizeof(page_owner->comm));
  __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
  __set_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
 }
 rcu_read_unlock();
}

static inline void __update_page_owner_free_handle(struct page *page,
         depot_stack_handle_t handle,
         unsigned short order,
         pid_t pid, pid_t tgid,
         u64 free_ts_nsec)
{
 struct page_ext_iter iter;
 struct page_ext *page_ext;
 struct page_owner *page_owner;

 rcu_read_lock();
 for_each_page_ext(page, 1 << order, page_ext, iter) {
  page_owner = get_page_owner(page_ext);
  /* Only __reset_page_owner() wants to clear the bit */
  if (handle) {
   __clear_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
   page_owner->free_handle = handle;
  }
  page_owner->free_ts_nsec = free_ts_nsec;
  page_owner->free_pid = current->pid;
  page_owner->free_tgid = current->tgid;
 }
 rcu_read_unlock();
}

void __reset_page_owner(struct page *page, unsigned short order)
{
 struct page_ext *page_ext;
 depot_stack_handle_t handle;
 depot_stack_handle_t alloc_handle;
 struct page_owner *page_owner;
 u64 free_ts_nsec = local_clock();

 page_ext = page_ext_get(page);
 if (unlikely(!page_ext))
  return;

 page_owner = get_page_owner(page_ext);
 alloc_handle = page_owner->handle;
 page_ext_put(page_ext);

 /*
 * Do not specify GFP_NOWAIT to make gfpflags_allow_spinning() == false
 * to prevent issues in stack_depot_save().
 * This is similar to alloc_pages_nolock() gfp flags, but only used
 * to signal stack_depot to avoid spin_locks.
 */
 handle = save_stack(__GFP_NOWARN);
 __update_page_owner_free_handle(page, handle, order, current->pid,
     current->tgid, free_ts_nsec);

 if (alloc_handle != early_handle)
  /*
 * early_handle is being set as a handle for all those
 * early allocated pages. See init_pages_in_zone().
 * Since their refcount is not being incremented because
 * the machinery is not ready yet, we cannot decrement
 * their refcount either.
 */
  dec_stack_record_count(alloc_handle, 1 << order);
}

noinline void __set_page_owner(struct page *page, unsigned short order,
     gfp_t gfp_mask)
{
 u64 ts_nsec = local_clock();
 depot_stack_handle_t handle;

 handle = save_stack(gfp_mask);
 __update_page_owner_handle(page, handle, order, gfp_mask, -1,
       ts_nsec, current->pid, current->tgid,
       current->comm);
 inc_stack_record_count(handle, gfp_mask, 1 << order);
}

void __folio_set_owner_migrate_reason(struct folio *folio, int reason)
{
 struct page_ext *page_ext = page_ext_get(&folio->page);
 struct page_owner *page_owner;

 if (unlikely(!page_ext))
  return;

 page_owner = get_page_owner(page_ext);
 page_owner->last_migrate_reason = reason;
 page_ext_put(page_ext);
}

void __split_page_owner(struct page *page, int old_order, int new_order)
{
 struct page_ext_iter iter;
 struct page_ext *page_ext;
 struct page_owner *page_owner;

 rcu_read_lock();
 for_each_page_ext(page, 1 << old_order, page_ext, iter) {
  page_owner = get_page_owner(page_ext);
  page_owner->order = new_order;
 }
 rcu_read_unlock();
}

void __folio_copy_owner(struct folio *newfolio, struct folio *old)
{
 struct page_ext *page_ext;
 struct page_ext_iter iter;
 struct page_owner *old_page_owner;
 struct page_owner *new_page_owner;
 depot_stack_handle_t migrate_handle;

 page_ext = page_ext_get(&old->page);
 if (unlikely(!page_ext))
  return;

 old_page_owner = get_page_owner(page_ext);
 page_ext_put(page_ext);

 page_ext = page_ext_get(&newfolio->page);
 if (unlikely(!page_ext))
  return;

 new_page_owner = get_page_owner(page_ext);
 page_ext_put(page_ext);

 migrate_handle = new_page_owner->handle;
 __update_page_owner_handle(&newfolio->page, old_page_owner->handle,
       old_page_owner->order, old_page_owner->gfp_mask,
       old_page_owner->last_migrate_reason,
       old_page_owner->ts_nsec, old_page_owner->pid,
       old_page_owner->tgid, old_page_owner->comm);
 /*
 * Do not proactively clear PAGE_EXT_OWNER{_ALLOCATED} bits as the folio
 * will be freed after migration. Keep them until then as they may be
 * useful.
 */
 __update_page_owner_free_handle(&newfolio->page, 0, old_page_owner->order,
     old_page_owner->free_pid,
     old_page_owner->free_tgid,
     old_page_owner->free_ts_nsec);
 /*
 * We linked the original stack to the new folio, we need to do the same
 * for the new one and the old folio otherwise there will be an imbalance
 * when subtracting those pages from the stack.
 */
 rcu_read_lock();
 for_each_page_ext(&old->page, 1 << new_page_owner->order, page_ext, iter) {
  old_page_owner = get_page_owner(page_ext);
  old_page_owner->handle = migrate_handle;
 }
 rcu_read_unlock();
}

void pagetypeinfo_showmixedcount_print(struct seq_file *m,
           pg_data_t *pgdat, struct zone *zone)
{
 struct page *page;
 struct page_ext *page_ext;
 struct page_owner *page_owner;
 unsigned long pfn, block_end_pfn;
 unsigned long end_pfn = zone_end_pfn(zone);
 unsigned long count[MIGRATE_TYPES] = { 0, };
 int pageblock_mt, page_mt;
 int i;

 /* Scan block by block. First and last block may be incomplete */
 pfn = zone->zone_start_pfn;

 /*
 * Walk the zone in pageblock_nr_pages steps. If a page block spans
 * a zone boundary, it will be double counted between zones. This does
 * not matter as the mixed block count will still be correct
 */
 for (; pfn < end_pfn; ) {
  page = pfn_to_online_page(pfn);
  if (!page) {
   pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
   continue;
  }

  block_end_pfn = pageblock_end_pfn(pfn);
  block_end_pfn = min(block_end_pfn, end_pfn);

  pageblock_mt = get_pageblock_migratetype(page);

  for (; pfn < block_end_pfn; pfn++) {
   /* The pageblock is online, no need to recheck. */
   page = pfn_to_page(pfn);

   if (page_zone(page) != zone)
    continue;

   if (PageBuddy(page)) {
    unsigned long freepage_order;

    freepage_order = buddy_order_unsafe(page);
    if (freepage_order <= MAX_PAGE_ORDER)
     pfn += (1UL << freepage_order) - 1;
    continue;
   }

   if (PageReserved(page))
    continue;

   page_ext = page_ext_get(page);
   if (unlikely(!page_ext))
    continue;

   if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
    goto ext_put_continue;

   page_owner = get_page_owner(page_ext);
   page_mt = gfp_migratetype(page_owner->gfp_mask);
   if (pageblock_mt != page_mt) {
    if (is_migrate_cma(pageblock_mt))
     count[MIGRATE_MOVABLE]++;
    else
     count[pageblock_mt]++;

    pfn = block_end_pfn;
    page_ext_put(page_ext);
    break;
   }
   pfn += (1UL << page_owner->order) - 1;
ext_put_continue:
   page_ext_put(page_ext);
  }
 }

 /* Print counts */
 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
 for (i = 0; i < MIGRATE_TYPES; i++)
  seq_printf(m, "%12lu ", count[i]);
 seq_putc(m, '\n');
}

/*
 * Looking for memcg information and print it out
 */
static inline int print_page_owner_memcg(char *kbuf, size_t count, int ret,
      struct page *page)
{
#ifdef CONFIG_MEMCG
 unsigned long memcg_data;
 struct mem_cgroup *memcg;
 bool online;
 char name[80];

 rcu_read_lock();
 memcg_data = READ_ONCE(page->memcg_data);
 if (!memcg_data || PageTail(page))
  goto out_unlock;

 if (memcg_data & MEMCG_DATA_OBJEXTS)
  ret += scnprintf(kbuf + ret, count - ret,
    "Slab cache page\n");

 memcg = page_memcg_check(page);
 if (!memcg)
  goto out_unlock;

 online = (memcg->css.flags & CSS_ONLINE);
 cgroup_name(memcg->css.cgroup, name, sizeof(name));
 ret += scnprintf(kbuf + ret, count - ret,
   "Charged %sto %smemcg %s\n",
   PageMemcgKmem(page) ? "(via objcg) " : "",
   online ? "" : "offline ",
   name);
out_unlock:
 rcu_read_unlock();
#endif /* CONFIG_MEMCG */

 return ret;
}

static ssize_t
print_page_owner(char __user *buf, size_t count, unsigned long pfn,
  struct page *page, struct page_owner *page_owner,
  depot_stack_handle_t handle)
{
 int ret, pageblock_mt, page_mt;
 char *kbuf;

 count = min_t(size_t, count, PAGE_SIZE);
 kbuf = kmalloc(count, GFP_KERNEL);
 if (!kbuf)
  return -ENOMEM;

 ret = scnprintf(kbuf, count,
   "Page allocated via order %u, mask %#x(%pGg), pid %d, tgid %d (%s), ts %llu ns\n",
   page_owner->order, page_owner->gfp_mask,
   &page_owner->gfp_mask, page_owner->pid,
   page_owner->tgid, page_owner->comm,
   page_owner->ts_nsec);

 /* Print information relevant to grouping pages by mobility */
 pageblock_mt = get_pageblock_migratetype(page);
 page_mt  = gfp_migratetype(page_owner->gfp_mask);
 ret += scnprintf(kbuf + ret, count - ret,
   "PFN 0x%lx type %s Block %lu type %s Flags %pGp\n",
   pfn,
   migratetype_names[page_mt],
   pfn >> pageblock_order,
   migratetype_names[pageblock_mt],
   &page->flags);

 ret += stack_depot_snprint(handle, kbuf + ret, count - ret, 0);
 if (ret >= count)
  goto err;

 if (page_owner->last_migrate_reason != -1) {
  ret += scnprintf(kbuf + ret, count - ret,
   "Page has been migrated, last migrate reason: %s\n",
   migrate_reason_names[page_owner->last_migrate_reason]);
 }

 ret = print_page_owner_memcg(kbuf, count, ret, page);

 ret += snprintf(kbuf + ret, count - ret, "\n");
 if (ret >= count)
  goto err;

 if (copy_to_user(buf, kbuf, ret))
  ret = -EFAULT;

 kfree(kbuf);
 return ret;

err:
 kfree(kbuf);
 return -ENOMEM;
}

void __dump_page_owner(const struct page *page)
{
 struct page_ext *page_ext = page_ext_get((void *)page);
 struct page_owner *page_owner;
 depot_stack_handle_t handle;
 gfp_t gfp_mask;
 int mt;

 if (unlikely(!page_ext)) {
  pr_alert("There is not page extension available.\n");
  return;
 }

 page_owner = get_page_owner(page_ext);
 gfp_mask = page_owner->gfp_mask;
 mt = gfp_migratetype(gfp_mask);

 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
  pr_alert("page_owner info is not present (never set?)\n");
  page_ext_put(page_ext);
  return;
 }

 if (test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
  pr_alert("page_owner tracks the page as allocated\n");
 else
  pr_alert("page_owner tracks the page as freed\n");

 pr_alert("page last allocated via order %u, migratetype %s, gfp_mask %#x(%pGg), pid %d, tgid %d (%s), ts %llu, free_ts %llu\n",
   page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask,
   page_owner->pid, page_owner->tgid, page_owner->comm,
   page_owner->ts_nsec, page_owner->free_ts_nsec);

 handle = READ_ONCE(page_owner->handle);
 if (!handle)
  pr_alert("page_owner allocation stack trace missing\n");
 else
  stack_depot_print(handle);

 handle = READ_ONCE(page_owner->free_handle);
 if (!handle) {
  pr_alert("page_owner free stack trace missing\n");
 } else {
  pr_alert("page last free pid %d tgid %d stack trace:\n",
     page_owner->free_pid, page_owner->free_tgid);
  stack_depot_print(handle);
 }

 if (page_owner->last_migrate_reason != -1)
  pr_alert("page has been migrated, last migrate reason: %s\n",
   migrate_reason_names[page_owner->last_migrate_reason]);
 page_ext_put(page_ext);
}

static ssize_t
read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
{
 unsigned long pfn;
 struct page *page;
 struct page_ext *page_ext;
 struct page_owner *page_owner;
 depot_stack_handle_t handle;

 if (!static_branch_unlikely(&page_owner_inited))
  return -EINVAL;

 page = NULL;
 if (*ppos == 0)
  pfn = min_low_pfn;
 else
  pfn = *ppos;
 /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
 while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
  pfn++;

 /* Find an allocated page */
 for (; pfn < max_pfn; pfn++) {
  /*
 * This temporary page_owner is required so
 * that we can avoid the context switches while holding
 * the rcu lock and copying the page owner information to
 * user through copy_to_user() or GFP_KERNEL allocations.
 */
  struct page_owner page_owner_tmp;

  /*
 * If the new page is in a new MAX_ORDER_NR_PAGES area,
 * validate the area as existing, skip it if not
 */
  if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
   pfn += MAX_ORDER_NR_PAGES - 1;
   continue;
  }

  page = pfn_to_page(pfn);
  if (PageBuddy(page)) {
   unsigned long freepage_order = buddy_order_unsafe(page);

   if (freepage_order <= MAX_PAGE_ORDER)
    pfn += (1UL << freepage_order) - 1;
   continue;
  }

  page_ext = page_ext_get(page);
  if (unlikely(!page_ext))
   continue;

  /*
 * Some pages could be missed by concurrent allocation or free,
 * because we don't hold the zone lock.
 */
  if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
   goto ext_put_continue;

  /*
 * Although we do have the info about past allocation of free
 * pages, it's not relevant for current memory usage.
 */
  if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
   goto ext_put_continue;

  page_owner = get_page_owner(page_ext);

  /*
 * Don't print "tail" pages of high-order allocations as that
 * would inflate the stats.
 */
  if (!IS_ALIGNED(pfn, 1 << page_owner->order))
   goto ext_put_continue;

  /*
 * Access to page_ext->handle isn't synchronous so we should
 * be careful to access it.
 */
  handle = READ_ONCE(page_owner->handle);
  if (!handle)
   goto ext_put_continue;

  /* Record the next PFN to read in the file offset */
  *ppos = pfn + 1;

  page_owner_tmp = *page_owner;
  page_ext_put(page_ext);
  return print_page_owner(buf, count, pfn, page,
    &page_owner_tmp, handle);
ext_put_continue:
  page_ext_put(page_ext);
 }

 return 0;
}

static loff_t lseek_page_owner(struct file *file, loff_t offset, int orig)
{
 switch (orig) {
 case SEEK_SET:
  file->f_pos = offset;
  break;
 case SEEK_CUR:
  file->f_pos += offset;
  break;
 default:
  return -EINVAL;
 }
 return file->f_pos;
}

static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
{
 unsigned long pfn = zone->zone_start_pfn;
 unsigned long end_pfn = zone_end_pfn(zone);
 unsigned long count = 0;

 /*
 * Walk the zone in pageblock_nr_pages steps. If a page block spans
 * a zone boundary, it will be double counted between zones. This does
 * not matter as the mixed block count will still be correct
 */
 for (; pfn < end_pfn; ) {
  unsigned long block_end_pfn;

  if (!pfn_valid(pfn)) {
   pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
   continue;
  }

  block_end_pfn = pageblock_end_pfn(pfn);
  block_end_pfn = min(block_end_pfn, end_pfn);

  for (; pfn < block_end_pfn; pfn++) {
   struct page *page = pfn_to_page(pfn);
   struct page_ext *page_ext;

   if (page_zone(page) != zone)
    continue;

   /*
 * To avoid having to grab zone->lock, be a little
 * careful when reading buddy page order. The only
 * danger is that we skip too much and potentially miss
 * some early allocated pages, which is better than
 * heavy lock contention.
 */
   if (PageBuddy(page)) {
    unsigned long order = buddy_order_unsafe(page);

    if (order > 0 && order <= MAX_PAGE_ORDER)
     pfn += (1UL << order) - 1;
    continue;
   }

   if (PageReserved(page))
    continue;

   page_ext = page_ext_get(page);
   if (unlikely(!page_ext))
    continue;

   /* Maybe overlapping zone */
   if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
    goto ext_put_continue;

   /* Found early allocated page */
   __update_page_owner_handle(page, early_handle, 0, 0,
         -1, local_clock(), current->pid,
         current->tgid, current->comm);
   count++;
ext_put_continue:
   page_ext_put(page_ext);
  }
  cond_resched();
 }

 pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
  pgdat->node_id, zone->name, count);
}

static void init_zones_in_node(pg_data_t *pgdat)
{
 struct zone *zone;
 struct zone *node_zones = pgdat->node_zones;

 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
  if (!populated_zone(zone))
   continue;

  init_pages_in_zone(pgdat, zone);
 }
}

static void init_early_allocated_pages(void)
{
 pg_data_t *pgdat;

 for_each_online_pgdat(pgdat)
  init_zones_in_node(pgdat);
}

static const struct file_operations proc_page_owner_operations = {
 .read  = read_page_owner,
 .llseek  = lseek_page_owner,
};

static void *stack_start(struct seq_file *m, loff_t *ppos)
{
 struct stack *stack;

 if (*ppos == -1UL)
  return NULL;

 if (!*ppos) {
  /*
 * This pairs with smp_store_release() from function
 * add_stack_record_to_list(), so we get a consistent
 * value of stack_list.
 */
  stack = smp_load_acquire(&stack_list);
  m->private = stack;
 } else {
  stack = m->private;
 }

 return stack;
}

static void *stack_next(struct seq_file *m, void *v, loff_t *ppos)
{
 struct stack *stack = v;

 stack = stack->next;
 *ppos = stack ? *ppos + 1 : -1UL;
 m->private = stack;

 return stack;
}

static unsigned long page_owner_pages_threshold;

static int stack_print(struct seq_file *m, void *v)
{
 int i, nr_base_pages;
 struct stack *stack = v;
 unsigned long *entries;
 unsigned long nr_entries;
 struct stack_record *stack_record = stack->stack_record;

 if (!stack->stack_record)
  return 0;

 nr_entries = stack_record->size;
 entries = stack_record->entries;
 nr_base_pages = refcount_read(&stack_record->count) - 1;

 if (nr_base_pages < 1 || nr_base_pages < page_owner_pages_threshold)
  return 0;

 for (i = 0; i < nr_entries; i++)
  seq_printf(m, " %pS\n", (void *)entries[i]);
 seq_printf(m, "nr_base_pages: %d\n\n", nr_base_pages);

 return 0;
}

static void stack_stop(struct seq_file *m, void *v)
{
}

static const struct seq_operations page_owner_stack_op = {
 .start = stack_start,
 .next = stack_next,
 .stop = stack_stop,
 .show = stack_print
};

static int page_owner_stack_open(struct inode *inode, struct file *file)
{
 return seq_open_private(file, &page_owner_stack_op, 0);
}

static const struct file_operations page_owner_stack_operations = {
 .open  = page_owner_stack_open,
 .read  = seq_read,
 .llseek  = seq_lseek,
 .release = seq_release,
};

static int page_owner_threshold_get(void *data, u64 *val)
{
 *val = READ_ONCE(page_owner_pages_threshold);
 return 0;
}

static int page_owner_threshold_set(void *data, u64 val)
{
 WRITE_ONCE(page_owner_pages_threshold, val);
 return 0;
}

DEFINE_SIMPLE_ATTRIBUTE(proc_page_owner_threshold, &page_owner_threshold_get,
   &page_owner_threshold_set, "%llu");


static int __init pageowner_init(void)
{
 struct dentry *dir;

 if (!static_branch_unlikely(&page_owner_inited)) {
  pr_info("page_owner is disabled\n");
  return 0;
 }

 debugfs_create_file("page_owner", 0400, NULL, NULL,
       &proc_page_owner_operations);
 dir = debugfs_create_dir("page_owner_stacks", NULL);
 debugfs_create_file("show_stacks", 0400, dir, NULL,
       &page_owner_stack_operations);
 debugfs_create_file("count_threshold", 0600, dir, NULL,
       &proc_page_owner_threshold);

 return 0;
}
late_initcall(pageowner_init)