percpu: implement new dynamic percpu allocator
authorTejun Heo <tj@kernel.org>
Fri, 20 Feb 2009 07:29:08 +0000 (16:29 +0900)
committerTejun Heo <tj@kernel.org>
Fri, 20 Feb 2009 07:29:08 +0000 (16:29 +0900)
Impact: new scalable dynamic percpu allocator which allows dynamic
        percpu areas to be accessed the same way as static ones

Implement scalable dynamic percpu allocator which can be used for both
static and dynamic percpu areas.  This will allow static and dynamic
areas to share faster direct access methods.  This feature is optional
and enabled only when CONFIG_HAVE_DYNAMIC_PER_CPU_AREA is defined by
arch.  Please read comment on top of mm/percpu.c for details.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
include/linux/percpu.h
kernel/module.c
mm/Makefile
mm/percpu.c [new file with mode: 0644]

index d99e24ae18117aa0502453e8213cc6cccfe2c7cb..18080995ff3efda8a5dac47859029fff51363e23 100644 (file)
 
 #ifdef CONFIG_SMP
 
-struct percpu_data {
-       void *ptrs[1];
-};
+#ifdef CONFIG_HAVE_DYNAMIC_PER_CPU_AREA
 
-#define __percpu_disguise(pdata) (struct percpu_data *)~(unsigned long)(pdata)
+extern void *pcpu_base_addr;
 
+typedef void (*pcpu_populate_pte_fn_t)(unsigned long addr);
+
+extern size_t __init pcpu_setup_static(pcpu_populate_pte_fn_t populate_pte_fn,
+                                      struct page **pages, size_t cpu_size);
 /*
  * Use this to get to a cpu's version of the per-cpu object
  * dynamically allocated. Non-atomic access to the current CPU's
  * version should probably be combined with get_cpu()/put_cpu().
  */
+#define per_cpu_ptr(ptr, cpu)  SHIFT_PERCPU_PTR((ptr), per_cpu_offset((cpu)))
+
+#else /* CONFIG_HAVE_DYNAMIC_PER_CPU_AREA */
+
+struct percpu_data {
+       void *ptrs[1];
+};
+
+#define __percpu_disguise(pdata) (struct percpu_data *)~(unsigned long)(pdata)
+
 #define per_cpu_ptr(ptr, cpu)                                          \
 ({                                                                     \
         struct percpu_data *__p = __percpu_disguise(ptr);              \
         (__typeof__(ptr))__p->ptrs[(cpu)];                             \
 })
 
+#endif /* CONFIG_HAVE_DYNAMIC_PER_CPU_AREA */
+
 extern void *__alloc_percpu(size_t size, size_t align);
 extern void free_percpu(void *__pdata);
 
index 52b3497b8748e9a6bdc14ecd135701521af4af07..1f0657ae555b15eadbb10658a02665bad204e828 100644 (file)
@@ -51,6 +51,7 @@
 #include <linux/tracepoint.h>
 #include <linux/ftrace.h>
 #include <linux/async.h>
+#include <linux/percpu.h>
 
 #if 0
 #define DEBUGP printk
@@ -366,6 +367,34 @@ static struct module *find_module(const char *name)
 }
 
 #ifdef CONFIG_SMP
+
+#ifdef CONFIG_HAVE_DYNAMIC_PER_CPU_AREA
+
+static void *percpu_modalloc(unsigned long size, unsigned long align,
+                            const char *name)
+{
+       void *ptr;
+
+       if (align > PAGE_SIZE) {
+               printk(KERN_WARNING "%s: per-cpu alignment %li > %li\n",
+                      name, align, PAGE_SIZE);
+               align = PAGE_SIZE;
+       }
+
+       ptr = __alloc_percpu(size, align);
+       if (!ptr)
+               printk(KERN_WARNING
+                      "Could not allocate %lu bytes percpu data\n", size);
+       return ptr;
+}
+
+static void percpu_modfree(void *freeme)
+{
+       free_percpu(freeme);
+}
+
+#else /* ... !CONFIG_HAVE_DYNAMIC_PER_CPU_AREA */
+
 /* Number of blocks used and allocated. */
 static unsigned int pcpu_num_used, pcpu_num_allocated;
 /* Size of each block.  -ve means used. */
@@ -499,6 +528,8 @@ static int percpu_modinit(void)
 }
 __initcall(percpu_modinit);
 
+#endif /* CONFIG_HAVE_DYNAMIC_PER_CPU_AREA */
+
 static unsigned int find_pcpusec(Elf_Ehdr *hdr,
                                 Elf_Shdr *sechdrs,
                                 const char *secstrings)
index 72255be57f89160cf8ff12e702c013db310d6ef4..818569b68f4652e623f5d2bd4ed91fe68fcea2b4 100644 (file)
@@ -30,6 +30,10 @@ obj-$(CONFIG_FAILSLAB) += failslab.o
 obj-$(CONFIG_MEMORY_HOTPLUG) += memory_hotplug.o
 obj-$(CONFIG_FS_XIP) += filemap_xip.o
 obj-$(CONFIG_MIGRATION) += migrate.o
+ifdef CONFIG_HAVE_DYNAMIC_PER_CPU_AREA
+obj-$(CONFIG_SMP) += percpu.o
+else
 obj-$(CONFIG_SMP) += allocpercpu.o
+endif
 obj-$(CONFIG_QUICKLIST) += quicklist.o
 obj-$(CONFIG_CGROUP_MEM_RES_CTLR) += memcontrol.o page_cgroup.o
diff --git a/mm/percpu.c b/mm/percpu.c
new file mode 100644 (file)
index 0000000..4617d97
--- /dev/null
@@ -0,0 +1,890 @@
+/*
+ * linux/mm/percpu.c - percpu memory allocator
+ *
+ * Copyright (C) 2009          SUSE Linux Products GmbH
+ * Copyright (C) 2009          Tejun Heo <tj@kernel.org>
+ *
+ * This file is released under the GPLv2.
+ *
+ * This is percpu allocator which can handle both static and dynamic
+ * areas.  Percpu areas are allocated in chunks in vmalloc area.  Each
+ * chunk is consisted of num_possible_cpus() units and the first chunk
+ * is used for static percpu variables in the kernel image (special
+ * boot time alloc/init handling necessary as these areas need to be
+ * brought up before allocation services are running).  Unit grows as
+ * necessary and all units grow or shrink in unison.  When a chunk is
+ * filled up, another chunk is allocated.  ie. in vmalloc area
+ *
+ *  c0                           c1                         c2
+ *  -------------------          -------------------        ------------
+ * | u0 | u1 | u2 | u3 |        | u0 | u1 | u2 | u3 |      | u0 | u1 | u
+ *  -------------------  ......  -------------------  ....  ------------
+ *
+ * Allocation is done in offset-size areas of single unit space.  Ie,
+ * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
+ * c1:u1, c1:u2 and c1:u3.  Percpu access can be done by configuring
+ * percpu base registers UNIT_SIZE apart.
+ *
+ * There are usually many small percpu allocations many of them as
+ * small as 4 bytes.  The allocator organizes chunks into lists
+ * according to free size and tries to allocate from the fullest one.
+ * Each chunk keeps the maximum contiguous area size hint which is
+ * guaranteed to be eqaul to or larger than the maximum contiguous
+ * area in the chunk.  This helps the allocator not to iterate the
+ * chunk maps unnecessarily.
+ *
+ * Allocation state in each chunk is kept using an array of integers
+ * on chunk->map.  A positive value in the map represents a free
+ * region and negative allocated.  Allocation inside a chunk is done
+ * by scanning this map sequentially and serving the first matching
+ * entry.  This is mostly copied from the percpu_modalloc() allocator.
+ * Chunks are also linked into a rb tree to ease address to chunk
+ * mapping during free.
+ *
+ * To use this allocator, arch code should do the followings.
+ *
+ * - define CONFIG_HAVE_DYNAMIC_PER_CPU_AREA
+ *
+ * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
+ *   regular address to percpu pointer and back
+ *
+ * - use pcpu_setup_static() during percpu area initialization to
+ *   setup kernel static percpu area
+ */
+
+#include <linux/bitmap.h>
+#include <linux/bootmem.h>
+#include <linux/list.h>
+#include <linux/mm.h>
+#include <linux/module.h>
+#include <linux/mutex.h>
+#include <linux/percpu.h>
+#include <linux/pfn.h>
+#include <linux/rbtree.h>
+#include <linux/slab.h>
+#include <linux/vmalloc.h>
+
+#include <asm/cacheflush.h>
+#include <asm/tlbflush.h>
+
+#define PCPU_MIN_UNIT_PAGES_SHIFT      4       /* also max alloc size */
+#define PCPU_SLOT_BASE_SHIFT           5       /* 1-31 shares the same slot */
+#define PCPU_DFL_MAP_ALLOC             16      /* start a map with 16 ents */
+
+struct pcpu_chunk {
+       struct list_head        list;           /* linked to pcpu_slot lists */
+       struct rb_node          rb_node;        /* key is chunk->vm->addr */
+       int                     free_size;      /* free bytes in the chunk */
+       int                     contig_hint;    /* max contiguous size hint */
+       struct vm_struct        *vm;            /* mapped vmalloc region */
+       int                     map_used;       /* # of map entries used */
+       int                     map_alloc;      /* # of map entries allocated */
+       int                     *map;           /* allocation map */
+       struct page             *page[];        /* #cpus * UNIT_PAGES */
+};
+
+static int pcpu_unit_pages_shift;
+static int pcpu_unit_pages;
+static int pcpu_unit_shift;
+static int pcpu_unit_size;
+static int pcpu_chunk_size;
+static int pcpu_nr_slots;
+static size_t pcpu_chunk_struct_size;
+
+/* the address of the first chunk which starts with the kernel static area */
+void *pcpu_base_addr;
+EXPORT_SYMBOL_GPL(pcpu_base_addr);
+
+/* the size of kernel static area */
+static int pcpu_static_size;
+
+/*
+ * One mutex to rule them all.
+ *
+ * The following mutex is grabbed in the outermost public alloc/free
+ * interface functions and released only when the operation is
+ * complete.  As such, every function in this file other than the
+ * outermost functions are called under pcpu_mutex.
+ *
+ * It can easily be switched to use spinlock such that only the area
+ * allocation and page population commit are protected with it doing
+ * actual [de]allocation without holding any lock.  However, given
+ * what this allocator does, I think it's better to let them run
+ * sequentially.
+ */
+static DEFINE_MUTEX(pcpu_mutex);
+
+static struct list_head *pcpu_slot;            /* chunk list slots */
+static struct rb_root pcpu_addr_root = RB_ROOT;        /* chunks by address */
+
+static int pcpu_size_to_slot(int size)
+{
+       int highbit = fls(size);
+       return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
+}
+
+static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
+{
+       if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
+               return 0;
+
+       return pcpu_size_to_slot(chunk->free_size);
+}
+
+static int pcpu_page_idx(unsigned int cpu, int page_idx)
+{
+       return (cpu << pcpu_unit_pages_shift) + page_idx;
+}
+
+static struct page **pcpu_chunk_pagep(struct pcpu_chunk *chunk,
+                                     unsigned int cpu, int page_idx)
+{
+       return &chunk->page[pcpu_page_idx(cpu, page_idx)];
+}
+
+static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
+                                    unsigned int cpu, int page_idx)
+{
+       return (unsigned long)chunk->vm->addr +
+               (pcpu_page_idx(cpu, page_idx) << PAGE_SHIFT);
+}
+
+static bool pcpu_chunk_page_occupied(struct pcpu_chunk *chunk,
+                                    int page_idx)
+{
+       return *pcpu_chunk_pagep(chunk, 0, page_idx) != NULL;
+}
+
+/**
+ * pcpu_realloc - versatile realloc
+ * @p: the current pointer (can be NULL for new allocations)
+ * @size: the current size (can be 0 for new allocations)
+ * @new_size: the wanted new size (can be 0 for free)
+ *
+ * More robust realloc which can be used to allocate, resize or free a
+ * memory area of arbitrary size.  If the needed size goes over
+ * PAGE_SIZE, kernel VM is used.
+ *
+ * RETURNS:
+ * The new pointer on success, NULL on failure.
+ */
+static void *pcpu_realloc(void *p, size_t size, size_t new_size)
+{
+       void *new;
+
+       if (new_size <= PAGE_SIZE)
+               new = kmalloc(new_size, GFP_KERNEL);
+       else
+               new = vmalloc(new_size);
+       if (new_size && !new)
+               return NULL;
+
+       memcpy(new, p, min(size, new_size));
+       if (new_size > size)
+               memset(new + size, 0, new_size - size);
+
+       if (size <= PAGE_SIZE)
+               kfree(p);
+       else
+               vfree(p);
+
+       return new;
+}
+
+/**
+ * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
+ * @chunk: chunk of interest
+ * @oslot: the previous slot it was on
+ *
+ * This function is called after an allocation or free changed @chunk.
+ * New slot according to the changed state is determined and @chunk is
+ * moved to the slot.
+ */
+static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
+{
+       int nslot = pcpu_chunk_slot(chunk);
+
+       if (oslot != nslot) {
+               if (oslot < nslot)
+                       list_move(&chunk->list, &pcpu_slot[nslot]);
+               else
+                       list_move_tail(&chunk->list, &pcpu_slot[nslot]);
+       }
+}
+
+static struct rb_node **pcpu_chunk_rb_search(void *addr,
+                                            struct rb_node **parentp)
+{
+       struct rb_node **p = &pcpu_addr_root.rb_node;
+       struct rb_node *parent = NULL;
+       struct pcpu_chunk *chunk;
+
+       while (*p) {
+               parent = *p;
+               chunk = rb_entry(parent, struct pcpu_chunk, rb_node);
+
+               if (addr < chunk->vm->addr)
+                       p = &(*p)->rb_left;
+               else if (addr > chunk->vm->addr)
+                       p = &(*p)->rb_right;
+               else
+                       break;
+       }
+
+       if (parentp)
+               *parentp = parent;
+       return p;
+}
+
+/**
+ * pcpu_chunk_addr_search - search for chunk containing specified address
+ * @addr: address to search for
+ *
+ * Look for chunk which might contain @addr.  More specifically, it
+ * searchs for the chunk with the highest start address which isn't
+ * beyond @addr.
+ *
+ * RETURNS:
+ * The address of the found chunk.
+ */
+static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
+{
+       struct rb_node *n, *parent;
+       struct pcpu_chunk *chunk;
+
+       n = *pcpu_chunk_rb_search(addr, &parent);
+       if (!n) {
+               /* no exactly matching chunk, the parent is the closest */
+               n = parent;
+               BUG_ON(!n);
+       }
+       chunk = rb_entry(n, struct pcpu_chunk, rb_node);
+
+       if (addr < chunk->vm->addr) {
+               /* the parent was the next one, look for the previous one */
+               n = rb_prev(n);
+               BUG_ON(!n);
+               chunk = rb_entry(n, struct pcpu_chunk, rb_node);
+       }
+
+       return chunk;
+}
+
+/**
+ * pcpu_chunk_addr_insert - insert chunk into address rb tree
+ * @new: chunk to insert
+ *
+ * Insert @new into address rb tree.
+ */
+static void pcpu_chunk_addr_insert(struct pcpu_chunk *new)
+{
+       struct rb_node **p, *parent;
+
+       p = pcpu_chunk_rb_search(new->vm->addr, &parent);
+       BUG_ON(*p);
+       rb_link_node(&new->rb_node, parent, p);
+       rb_insert_color(&new->rb_node, &pcpu_addr_root);
+}
+
+/**
+ * pcpu_split_block - split a map block
+ * @chunk: chunk of interest
+ * @i: index of map block to split
+ * @head: head size (can be 0)
+ * @tail: tail size (can be 0)
+ *
+ * Split the @i'th map block into two or three blocks.  If @head is
+ * non-zero, @head bytes block is inserted before block @i moving it
+ * to @i+1 and reducing its size by @head bytes.
+ *
+ * If @tail is non-zero, the target block, which can be @i or @i+1
+ * depending on @head, is reduced by @tail bytes and @tail byte block
+ * is inserted after the target block.
+ *
+ * RETURNS:
+ * 0 on success, -errno on failure.
+ */
+static int pcpu_split_block(struct pcpu_chunk *chunk, int i, int head, int tail)
+{
+       int nr_extra = !!head + !!tail;
+       int target = chunk->map_used + nr_extra;
+
+       /* reallocation required? */
+       if (chunk->map_alloc < target) {
+               int new_alloc = chunk->map_alloc;
+               int *new;
+
+               while (new_alloc < target)
+                       new_alloc *= 2;
+
+               new = pcpu_realloc(chunk->map,
+                                  chunk->map_alloc * sizeof(new[0]),
+                                  new_alloc * sizeof(new[0]));
+               if (!new)
+                       return -ENOMEM;
+
+               chunk->map_alloc = new_alloc;
+               chunk->map = new;
+       }
+
+       /* insert a new subblock */
+       memmove(&chunk->map[i + nr_extra], &chunk->map[i],
+               sizeof(chunk->map[0]) * (chunk->map_used - i));
+       chunk->map_used += nr_extra;
+
+       if (head) {
+               chunk->map[i + 1] = chunk->map[i] - head;
+               chunk->map[i++] = head;
+       }
+       if (tail) {
+               chunk->map[i++] -= tail;
+               chunk->map[i] = tail;
+       }
+       return 0;
+}
+
+/**
+ * pcpu_alloc_area - allocate area from a pcpu_chunk
+ * @chunk: chunk of interest
+ * @size: wanted size
+ * @align: wanted align
+ *
+ * Try to allocate @size bytes area aligned at @align from @chunk.
+ * Note that this function only allocates the offset.  It doesn't
+ * populate or map the area.
+ *
+ * RETURNS:
+ * Allocated offset in @chunk on success, -errno on failure.
+ */
+static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
+{
+       int oslot = pcpu_chunk_slot(chunk);
+       int max_contig = 0;
+       int i, off;
+
+       /*
+        * The static chunk initially doesn't have map attached
+        * because kmalloc wasn't available during init.  Give it one.
+        */
+       if (unlikely(!chunk->map)) {
+               chunk->map = pcpu_realloc(NULL, 0,
+                               PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
+               if (!chunk->map)
+                       return -ENOMEM;
+
+               chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
+               chunk->map[chunk->map_used++] = -pcpu_static_size;
+               if (chunk->free_size)
+                       chunk->map[chunk->map_used++] = chunk->free_size;
+       }
+
+       for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
+               bool is_last = i + 1 == chunk->map_used;
+               int head, tail;
+
+               /* extra for alignment requirement */
+               head = ALIGN(off, align) - off;
+               BUG_ON(i == 0 && head != 0);
+
+               if (chunk->map[i] < 0)
+                       continue;
+               if (chunk->map[i] < head + size) {
+                       max_contig = max(chunk->map[i], max_contig);
+                       continue;
+               }
+
+               /*
+                * If head is small or the previous block is free,
+                * merge'em.  Note that 'small' is defined as smaller
+                * than sizeof(int), which is very small but isn't too
+                * uncommon for percpu allocations.
+                */
+               if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
+                       if (chunk->map[i - 1] > 0)
+                               chunk->map[i - 1] += head;
+                       else {
+                               chunk->map[i - 1] -= head;
+                               chunk->free_size -= head;
+                       }
+                       chunk->map[i] -= head;
+                       off += head;
+                       head = 0;
+               }
+
+               /* if tail is small, just keep it around */
+               tail = chunk->map[i] - head - size;
+               if (tail < sizeof(int))
+                       tail = 0;
+
+               /* split if warranted */
+               if (head || tail) {
+                       if (pcpu_split_block(chunk, i, head, tail))
+                               return -ENOMEM;
+                       if (head) {
+                               i++;
+                               off += head;
+                               max_contig = max(chunk->map[i - 1], max_contig);
+                       }
+                       if (tail)
+                               max_contig = max(chunk->map[i + 1], max_contig);
+               }
+
+               /* update hint and mark allocated */
+               if (is_last)
+                       chunk->contig_hint = max_contig; /* fully scanned */
+               else
+                       chunk->contig_hint = max(chunk->contig_hint,
+                                                max_contig);
+
+               chunk->free_size -= chunk->map[i];
+               chunk->map[i] = -chunk->map[i];
+
+               pcpu_chunk_relocate(chunk, oslot);
+               return off;
+       }
+
+       chunk->contig_hint = max_contig;        /* fully scanned */
+       pcpu_chunk_relocate(chunk, oslot);
+
+       /*
+        * Tell the upper layer that this chunk has no area left.
+        * Note that this is not an error condition but a notification
+        * to upper layer that it needs to look at other chunks.
+        * -ENOSPC is chosen as it isn't used in memory subsystem and
+        * matches the meaning in a way.
+        */
+       return -ENOSPC;
+}
+
+/**
+ * pcpu_free_area - free area to a pcpu_chunk
+ * @chunk: chunk of interest
+ * @freeme: offset of area to free
+ *
+ * Free area starting from @freeme to @chunk.  Note that this function
+ * only modifies the allocation map.  It doesn't depopulate or unmap
+ * the area.
+ */
+static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
+{
+       int oslot = pcpu_chunk_slot(chunk);
+       int i, off;
+
+       for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
+               if (off == freeme)
+                       break;
+       BUG_ON(off != freeme);
+       BUG_ON(chunk->map[i] > 0);
+
+       chunk->map[i] = -chunk->map[i];
+       chunk->free_size += chunk->map[i];
+
+       /* merge with previous? */
+       if (i > 0 && chunk->map[i - 1] >= 0) {
+               chunk->map[i - 1] += chunk->map[i];
+               chunk->map_used--;
+               memmove(&chunk->map[i], &chunk->map[i + 1],
+                       (chunk->map_used - i) * sizeof(chunk->map[0]));
+               i--;
+       }
+       /* merge with next? */
+       if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
+               chunk->map[i] += chunk->map[i + 1];
+               chunk->map_used--;
+               memmove(&chunk->map[i + 1], &chunk->map[i + 2],
+                       (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
+       }
+
+       chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
+       pcpu_chunk_relocate(chunk, oslot);
+}
+
+/**
+ * pcpu_unmap - unmap pages out of a pcpu_chunk
+ * @chunk: chunk of interest
+ * @page_start: page index of the first page to unmap
+ * @page_end: page index of the last page to unmap + 1
+ * @flush: whether to flush cache and tlb or not
+ *
+ * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
+ * If @flush is true, vcache is flushed before unmapping and tlb
+ * after.
+ */
+static void pcpu_unmap(struct pcpu_chunk *chunk, int page_start, int page_end,
+                      bool flush)
+{
+       unsigned int last = num_possible_cpus() - 1;
+       unsigned int cpu;
+
+       /*
+        * Each flushing trial can be very expensive, issue flush on
+        * the whole region at once rather than doing it for each cpu.
+        * This could be an overkill but is more scalable.
+        */
+       if (flush)
+               flush_cache_vunmap(pcpu_chunk_addr(chunk, 0, page_start),
+                                  pcpu_chunk_addr(chunk, last, page_end));
+
+       for_each_possible_cpu(cpu)
+               unmap_kernel_range_noflush(
+                               pcpu_chunk_addr(chunk, cpu, page_start),
+                               (page_end - page_start) << PAGE_SHIFT);
+
+       /* ditto as flush_cache_vunmap() */
+       if (flush)
+               flush_tlb_kernel_range(pcpu_chunk_addr(chunk, 0, page_start),
+                                      pcpu_chunk_addr(chunk, last, page_end));
+}
+
+/**
+ * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
+ * @chunk: chunk to depopulate
+ * @off: offset to the area to depopulate
+ * @size: size of the area to depopulate
+ * @flush: whether to flush cache and tlb or not
+ *
+ * For each cpu, depopulate and unmap pages [@page_start,@page_end)
+ * from @chunk.  If @flush is true, vcache is flushed before unmapping
+ * and tlb after.
+ */
+static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, size_t off,
+                                 size_t size, bool flush)
+{
+       int page_start = PFN_DOWN(off);
+       int page_end = PFN_UP(off + size);
+       int unmap_start = -1;
+       int uninitialized_var(unmap_end);
+       unsigned int cpu;
+       int i;
+
+       for (i = page_start; i < page_end; i++) {
+               for_each_possible_cpu(cpu) {
+                       struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i);
+
+                       if (!*pagep)
+                               continue;
+
+                       __free_page(*pagep);
+
+                       /*
+                        * If it's partial depopulation, it might get
+                        * populated or depopulated again.  Mark the
+                        * page gone.
+                        */
+                       *pagep = NULL;
+
+                       unmap_start = unmap_start < 0 ? i : unmap_start;
+                       unmap_end = i + 1;
+               }
+       }
+
+       if (unmap_start >= 0)
+               pcpu_unmap(chunk, unmap_start, unmap_end, flush);
+}
+
+/**
+ * pcpu_map - map pages into a pcpu_chunk
+ * @chunk: chunk of interest
+ * @page_start: page index of the first page to map
+ * @page_end: page index of the last page to map + 1
+ *
+ * For each cpu, map pages [@page_start,@page_end) into @chunk.
+ * vcache is flushed afterwards.
+ */
+static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end)
+{
+       unsigned int last = num_possible_cpus() - 1;
+       unsigned int cpu;
+       int err;
+
+       for_each_possible_cpu(cpu) {
+               err = map_kernel_range_noflush(
+                               pcpu_chunk_addr(chunk, cpu, page_start),
+                               (page_end - page_start) << PAGE_SHIFT,
+                               PAGE_KERNEL,
+                               pcpu_chunk_pagep(chunk, cpu, page_start));
+               if (err < 0)
+                       return err;
+       }
+
+       /* flush at once, please read comments in pcpu_unmap() */
+       flush_cache_vmap(pcpu_chunk_addr(chunk, 0, page_start),
+                        pcpu_chunk_addr(chunk, last, page_end));
+       return 0;
+}
+
+/**
+ * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
+ * @chunk: chunk of interest
+ * @off: offset to the area to populate
+ * @size: size of the area to populate
+ *
+ * For each cpu, populate and map pages [@page_start,@page_end) into
+ * @chunk.  The area is cleared on return.
+ */
+static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
+{
+       const gfp_t alloc_mask = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
+       int page_start = PFN_DOWN(off);
+       int page_end = PFN_UP(off + size);
+       int map_start = -1;
+       int map_end;
+       unsigned int cpu;
+       int i;
+
+       for (i = page_start; i < page_end; i++) {
+               if (pcpu_chunk_page_occupied(chunk, i)) {
+                       if (map_start >= 0) {
+                               if (pcpu_map(chunk, map_start, map_end))
+                                       goto err;
+                               map_start = -1;
+                       }
+                       continue;
+               }
+
+               map_start = map_start < 0 ? i : map_start;
+               map_end = i + 1;
+
+               for_each_possible_cpu(cpu) {
+                       struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i);
+
+                       *pagep = alloc_pages_node(cpu_to_node(cpu),
+                                                 alloc_mask, 0);
+                       if (!*pagep)
+                               goto err;
+               }
+       }
+
+       if (map_start >= 0 && pcpu_map(chunk, map_start, map_end))
+               goto err;
+
+       for_each_possible_cpu(cpu)
+               memset(chunk->vm->addr + (cpu << pcpu_unit_shift) + off, 0,
+                      size);
+
+       return 0;
+err:
+       /* likely under heavy memory pressure, give memory back */
+       pcpu_depopulate_chunk(chunk, off, size, true);
+       return -ENOMEM;
+}
+
+static void free_pcpu_chunk(struct pcpu_chunk *chunk)
+{
+       if (!chunk)
+               return;
+       if (chunk->vm)
+               free_vm_area(chunk->vm);
+       pcpu_realloc(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]), 0);
+       kfree(chunk);
+}
+
+static struct pcpu_chunk *alloc_pcpu_chunk(void)
+{
+       struct pcpu_chunk *chunk;
+
+       chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
+       if (!chunk)
+               return NULL;
+
+       chunk->map = pcpu_realloc(NULL, 0,
+                                 PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
+       chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
+       chunk->map[chunk->map_used++] = pcpu_unit_size;
+
+       chunk->vm = get_vm_area(pcpu_chunk_size, GFP_KERNEL);
+       if (!chunk->vm) {
+               free_pcpu_chunk(chunk);
+               return NULL;
+       }
+
+       INIT_LIST_HEAD(&chunk->list);
+       chunk->free_size = pcpu_unit_size;
+       chunk->contig_hint = pcpu_unit_size;
+
+       return chunk;
+}
+
+/**
+ * __alloc_percpu - allocate percpu area
+ * @size: size of area to allocate
+ * @align: alignment of area (max PAGE_SIZE)
+ *
+ * Allocate percpu area of @size bytes aligned at @align.  Might
+ * sleep.  Might trigger writeouts.
+ *
+ * RETURNS:
+ * Percpu pointer to the allocated area on success, NULL on failure.
+ */
+void *__alloc_percpu(size_t size, size_t align)
+{
+       void *ptr = NULL;
+       struct pcpu_chunk *chunk;
+       int slot, off;
+
+       if (unlikely(!size || size > PAGE_SIZE << PCPU_MIN_UNIT_PAGES_SHIFT ||
+                    align > PAGE_SIZE)) {
+               WARN(true, "illegal size (%zu) or align (%zu) for "
+                    "percpu allocation\n", size, align);
+               return NULL;
+       }
+
+       mutex_lock(&pcpu_mutex);
+
+       /* allocate area */
+       for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
+               list_for_each_entry(chunk, &pcpu_slot[slot], list) {
+                       if (size > chunk->contig_hint)
+                               continue;
+                       off = pcpu_alloc_area(chunk, size, align);
+                       if (off >= 0)
+                               goto area_found;
+                       if (off != -ENOSPC)
+                               goto out_unlock;
+               }
+       }
+
+       /* hmmm... no space left, create a new chunk */
+       chunk = alloc_pcpu_chunk();
+       if (!chunk)
+               goto out_unlock;
+       pcpu_chunk_relocate(chunk, -1);
+       pcpu_chunk_addr_insert(chunk);
+
+       off = pcpu_alloc_area(chunk, size, align);
+       if (off < 0)
+               goto out_unlock;
+
+area_found:
+       /* populate, map and clear the area */
+       if (pcpu_populate_chunk(chunk, off, size)) {
+               pcpu_free_area(chunk, off);
+               goto out_unlock;
+       }
+
+       ptr = __addr_to_pcpu_ptr(chunk->vm->addr + off);
+out_unlock:
+       mutex_unlock(&pcpu_mutex);
+       return ptr;
+}
+EXPORT_SYMBOL_GPL(__alloc_percpu);
+
+static void pcpu_kill_chunk(struct pcpu_chunk *chunk)
+{
+       pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size, false);
+       list_del(&chunk->list);
+       rb_erase(&chunk->rb_node, &pcpu_addr_root);
+       free_pcpu_chunk(chunk);
+}
+
+/**
+ * free_percpu - free percpu area
+ * @ptr: pointer to area to free
+ *
+ * Free percpu area @ptr.  Might sleep.
+ */
+void free_percpu(void *ptr)
+{
+       void *addr = __pcpu_ptr_to_addr(ptr);
+       struct pcpu_chunk *chunk;
+       int off;
+
+       if (!ptr)
+               return;
+
+       mutex_lock(&pcpu_mutex);
+
+       chunk = pcpu_chunk_addr_search(addr);
+       off = addr - chunk->vm->addr;
+
+       pcpu_free_area(chunk, off);
+
+       /* the chunk became fully free, kill one if there are other free ones */
+       if (chunk->free_size == pcpu_unit_size) {
+               struct pcpu_chunk *pos;
+
+               list_for_each_entry(pos,
+                                   &pcpu_slot[pcpu_chunk_slot(chunk)], list)
+                       if (pos != chunk) {
+                               pcpu_kill_chunk(pos);
+                               break;
+                       }
+       }
+
+       mutex_unlock(&pcpu_mutex);
+}
+EXPORT_SYMBOL_GPL(free_percpu);
+
+/**
+ * pcpu_setup_static - initialize kernel static percpu area
+ * @populate_pte_fn: callback to allocate pagetable
+ * @pages: num_possible_cpus() * PFN_UP(cpu_size) pages
+ *
+ * Initialize kernel static percpu area.  The caller should allocate
+ * all the necessary pages and pass them in @pages.
+ * @populate_pte_fn() is called on each page to be used for percpu
+ * mapping and is responsible for making sure all the necessary page
+ * tables for the page is allocated.
+ *
+ * RETURNS:
+ * The determined pcpu_unit_size which can be used to initialize
+ * percpu access.
+ */
+size_t __init pcpu_setup_static(pcpu_populate_pte_fn_t populate_pte_fn,
+                               struct page **pages, size_t cpu_size)
+{
+       static struct vm_struct static_vm;
+       struct pcpu_chunk *static_chunk;
+       int nr_cpu_pages = DIV_ROUND_UP(cpu_size, PAGE_SIZE);
+       unsigned int cpu;
+       int err, i;
+
+       pcpu_unit_pages_shift = max_t(int, PCPU_MIN_UNIT_PAGES_SHIFT,
+                                     order_base_2(cpu_size) - PAGE_SHIFT);
+
+       pcpu_static_size = cpu_size;
+       pcpu_unit_pages = 1 << pcpu_unit_pages_shift;
+       pcpu_unit_shift = PAGE_SHIFT + pcpu_unit_pages_shift;
+       pcpu_unit_size = 1 << pcpu_unit_shift;
+       pcpu_chunk_size = num_possible_cpus() * pcpu_unit_size;
+       pcpu_nr_slots = pcpu_size_to_slot(pcpu_unit_size) + 1;
+       pcpu_chunk_struct_size = sizeof(struct pcpu_chunk)
+               + (1 << pcpu_unit_pages_shift) * sizeof(struct page *);
+
+       /* allocate chunk slots */
+       pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
+       for (i = 0; i < pcpu_nr_slots; i++)
+               INIT_LIST_HEAD(&pcpu_slot[i]);
+
+       /* init and register vm area */
+       static_vm.flags = VM_ALLOC;
+       static_vm.size = pcpu_chunk_size;
+       vm_area_register_early(&static_vm);
+
+       /* init static_chunk */
+       static_chunk = alloc_bootmem(pcpu_chunk_struct_size);
+       INIT_LIST_HEAD(&static_chunk->list);
+       static_chunk->vm = &static_vm;
+       static_chunk->free_size = pcpu_unit_size - pcpu_static_size;
+       static_chunk->contig_hint = static_chunk->free_size;
+
+       /* assign pages and map them */
+       for_each_possible_cpu(cpu) {
+               for (i = 0; i < nr_cpu_pages; i++) {
+                       *pcpu_chunk_pagep(static_chunk, cpu, i) = *pages++;
+                       populate_pte_fn(pcpu_chunk_addr(static_chunk, cpu, i));
+               }
+       }
+
+       err = pcpu_map(static_chunk, 0, nr_cpu_pages);
+       if (err)
+               panic("failed to setup static percpu area, err=%d\n", err);
+
+       /* link static_chunk in */
+       pcpu_chunk_relocate(static_chunk, -1);
+       pcpu_chunk_addr_insert(static_chunk);
+
+       /* we're done */
+       pcpu_base_addr = (void *)pcpu_chunk_addr(static_chunk, 0, 0);
+       return pcpu_unit_size;
+}