mm: Close races between THP migration and PMD numa clearing
[linux-3.10.git] / mm / percpu.c
index cc9c4c6..8c8e08f 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * linux/mm/percpu.c - percpu memory allocator
+ * mm/percpu.c - percpu memory allocator
  *
  * Copyright (C) 2009          SUSE Linux Products GmbH
  * Copyright (C) 2009          Tejun Heo <tj@kernel.org>
@@ -7,14 +7,13 @@
  * 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 boot-time determined number of units and the
- * first chunk is used for static percpu variables in the kernel image
+ * areas.  Percpu areas are allocated in chunks.  Each chunk is
+ * consisted of boot-time determined number of 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
+ * When a chunk is filled up, another chunk is allocated.
  *
  *  c0                           c1                         c2
  *  -------------------          -------------------        ------------
@@ -32,7 +31,7 @@
  * 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
+ * guaranteed to be equal to or larger than the maximum contiguous
  * area in the chunk.  This helps the allocator not to iterate the
  * chunk maps unnecessarily.
  *
@@ -46,8 +45,6 @@
  *
  * To use this allocator, arch code should do the followings.
  *
- * - drop CONFIG_HAVE_LEGACY_PER_CPU_AREA
- *
  * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
  *   regular address to percpu pointer and back if they need to be
  *   different from the default
 #include <linux/spinlock.h>
 #include <linux/vmalloc.h>
 #include <linux/workqueue.h>
+#include <linux/kmemleak.h>
 
 #include <asm/cacheflush.h>
 #include <asm/sections.h>
 #include <asm/tlbflush.h>
+#include <asm/io.h>
 
 #define PCPU_SLOT_BASE_SHIFT           5       /* 1-31 shares the same slot */
 #define PCPU_DFL_MAP_ALLOC             16      /* start a map with 16 ents */
 
+#ifdef CONFIG_SMP
 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
 #ifndef __addr_to_pcpu_ptr
 #define __addr_to_pcpu_ptr(addr)                                       \
-       (void *)((unsigned long)(addr) - (unsigned long)pcpu_base_addr  \
-                + (unsigned long)__per_cpu_start)
+       (void __percpu *)((unsigned long)(addr) -                       \
+                         (unsigned long)pcpu_base_addr +               \
+                         (unsigned long)__per_cpu_start)
 #endif
 #ifndef __pcpu_ptr_to_addr
 #define __pcpu_ptr_to_addr(ptr)                                                \
-       (void *)((unsigned long)(ptr) + (unsigned long)pcpu_base_addr   \
-                - (unsigned long)__per_cpu_start)
+       (void __force *)((unsigned long)(ptr) +                         \
+                        (unsigned long)pcpu_base_addr -                \
+                        (unsigned long)__per_cpu_start)
 #endif
+#else  /* CONFIG_SMP */
+/* on UP, it's always identity mapped */
+#define __addr_to_pcpu_ptr(addr)       (void __percpu *)(addr)
+#define __pcpu_ptr_to_addr(ptr)                (void __force *)(ptr)
+#endif /* CONFIG_SMP */
 
 struct pcpu_chunk {
        struct list_head        list;           /* linked to pcpu_slot lists */
@@ -98,7 +105,7 @@ struct pcpu_chunk {
        int                     map_used;       /* # of map entries used */
        int                     map_alloc;      /* # of map entries allocated */
        int                     *map;           /* allocation map */
-       struct vm_struct        **vms;          /* mapped vmalloc regions */
+       void                    *data;          /* chunk data */
        bool                    immutable;      /* no [de]population allowed */
        unsigned long           populated[];    /* populated bitmap */
 };
@@ -110,9 +117,9 @@ static int pcpu_atom_size __read_mostly;
 static int pcpu_nr_slots __read_mostly;
 static size_t pcpu_chunk_struct_size __read_mostly;
 
-/* cpus with the lowest and highest unit numbers */
-static unsigned int pcpu_first_unit_cpu __read_mostly;
-static unsigned int pcpu_last_unit_cpu __read_mostly;
+/* cpus with the lowest and highest unit addresses */
+static unsigned int pcpu_low_unit_cpu __read_mostly;
+static unsigned int pcpu_high_unit_cpu __read_mostly;
 
 /* the address of the first chunk which starts with the kernel static area */
 void *pcpu_base_addr __read_mostly;
@@ -153,7 +160,10 @@ static int pcpu_reserved_chunk_limit;
  *
  * During allocation, pcpu_alloc_mutex is kept locked all the time and
  * pcpu_lock is grabbed and released as necessary.  All actual memory
- * allocations are done using GFP_KERNEL with pcpu_lock released.
+ * allocations are done using GFP_KERNEL with pcpu_lock released.  In
+ * general, percpu memory can't be allocated with irq off but
+ * irqsave/restore are still used in alloc path so that it can be used
+ * from early init path - sched_init() specifically.
  *
  * Free path accesses and alters only the index data structures, so it
  * can be safely called from atomic context.  When memory needs to be
@@ -173,6 +183,21 @@ static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
 static void pcpu_reclaim(struct work_struct *work);
 static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
 
+static bool pcpu_addr_in_first_chunk(void *addr)
+{
+       void *first_start = pcpu_first_chunk->base_addr;
+
+       return addr >= first_start && addr < first_start + pcpu_unit_size;
+}
+
+static bool pcpu_addr_in_reserved_chunk(void *addr)
+{
+       void *first_start = pcpu_first_chunk->base_addr;
+
+       return addr >= first_start &&
+               addr < first_start + pcpu_reserved_chunk_limit;
+}
+
 static int __pcpu_size_to_slot(int size)
 {
        int highbit = fls(size);        /* size is in bytes */
@@ -194,27 +219,6 @@ static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
        return pcpu_size_to_slot(chunk->free_size);
 }
 
-static int pcpu_page_idx(unsigned int cpu, int page_idx)
-{
-       return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
-}
-
-static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
-                                    unsigned int cpu, int page_idx)
-{
-       return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
-               (page_idx << PAGE_SHIFT);
-}
-
-static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk,
-                                   unsigned int cpu, int page_idx)
-{
-       /* must not be used on pre-mapped chunk */
-       WARN_ON(chunk->immutable);
-
-       return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx));
-}
-
 /* set the pointer to a chunk in a page struct */
 static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
 {
@@ -227,13 +231,27 @@ static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
        return (struct pcpu_chunk *)page->index;
 }
 
-static void pcpu_next_unpop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
+static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
+{
+       return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
+}
+
+static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
+                                    unsigned int cpu, int page_idx)
+{
+       return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
+               (page_idx << PAGE_SHIFT);
+}
+
+static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
+                                          int *rs, int *re, int end)
 {
        *rs = find_next_zero_bit(chunk->populated, end, *rs);
        *re = find_next_bit(chunk->populated, end, *rs + 1);
 }
 
-static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
+static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
+                                        int *rs, int *re, int end)
 {
        *rs = find_next_bit(chunk->populated, end, *rs);
        *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
@@ -241,7 +259,7 @@ static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
 
 /*
  * (Un)populated page region iterators.  Iterate over (un)populated
- * page regions betwen @start and @end in @chunk.  @rs and @re should
+ * page regions between @start and @end in @chunk.  @rs and @re should
  * be integer variables and will be set to start and end page index of
  * the current region.
  */
@@ -256,11 +274,11 @@ static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
             (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
 
 /**
- * pcpu_mem_alloc - allocate memory
+ * pcpu_mem_zalloc - allocate memory
  * @size: bytes to allocate
  *
  * Allocate @size bytes.  If @size is smaller than PAGE_SIZE,
- * kzalloc() is used; otherwise, vmalloc() is used.  The returned
+ * kzalloc() is used; otherwise, vzalloc() is used.  The returned
  * memory is always zeroed.
  *
  * CONTEXT:
@@ -269,16 +287,15 @@ static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
  * RETURNS:
  * Pointer to the allocated area on success, NULL on failure.
  */
-static void *pcpu_mem_alloc(size_t size)
+static void *pcpu_mem_zalloc(size_t size)
 {
+       if (WARN_ON_ONCE(!slab_is_available()))
+               return NULL;
+
        if (size <= PAGE_SIZE)
                return kzalloc(size, GFP_KERNEL);
-       else {
-               void *ptr = vmalloc(size);
-               if (ptr)
-                       memset(ptr, 0, size);
-               return ptr;
-       }
+       else
+               return vzalloc(size);
 }
 
 /**
@@ -286,7 +303,7 @@ static void *pcpu_mem_alloc(size_t size)
  * @ptr: memory to free
  * @size: size of the area
  *
- * Free @ptr.  @ptr should have been allocated using pcpu_mem_alloc().
+ * Free @ptr.  @ptr should have been allocated using pcpu_mem_zalloc().
  */
 static void pcpu_mem_free(void *ptr, size_t size)
 {
@@ -322,92 +339,81 @@ static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
 }
 
 /**
- * pcpu_chunk_addr_search - determine chunk containing specified address
- * @addr: address for which the chunk needs to be determined.
+ * pcpu_need_to_extend - determine whether chunk area map needs to be extended
+ * @chunk: chunk of interest
+ *
+ * Determine whether area map of @chunk needs to be extended to
+ * accommodate a new allocation.
+ *
+ * CONTEXT:
+ * pcpu_lock.
  *
  * RETURNS:
- * The address of the found chunk.
+ * New target map allocation length if extension is necessary, 0
+ * otherwise.
  */
-static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
+static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
 {
-       void *first_start = pcpu_first_chunk->base_addr;
+       int new_alloc;
 
-       /* is it in the first chunk? */
-       if (addr >= first_start && addr < first_start + pcpu_unit_size) {
-               /* is it in the reserved area? */
-               if (addr < first_start + pcpu_reserved_chunk_limit)
-                       return pcpu_reserved_chunk;
-               return pcpu_first_chunk;
-       }
+       if (chunk->map_alloc >= chunk->map_used + 2)
+               return 0;
 
-       /*
-        * The address is relative to unit0 which might be unused and
-        * thus unmapped.  Offset the address to the unit space of the
-        * current processor before looking it up in the vmalloc
-        * space.  Note that any possible cpu id can be used here, so
-        * there's no need to worry about preemption or cpu hotplug.
-        */
-       addr += pcpu_unit_offsets[smp_processor_id()];
-       return pcpu_get_page_chunk(vmalloc_to_page(addr));
+       new_alloc = PCPU_DFL_MAP_ALLOC;
+       while (new_alloc < chunk->map_used + 2)
+               new_alloc *= 2;
+
+       return new_alloc;
 }
 
 /**
- * pcpu_extend_area_map - extend area map for allocation
- * @chunk: target chunk
+ * pcpu_extend_area_map - extend area map of a chunk
+ * @chunk: chunk of interest
+ * @new_alloc: new target allocation length of the area map
  *
- * Extend area map of @chunk so that it can accomodate an allocation.
- * A single allocation can split an area into three areas, so this
- * function makes sure that @chunk->map has at least two extra slots.
+ * Extend area map of @chunk to have @new_alloc entries.
  *
  * CONTEXT:
- * pcpu_alloc_mutex, pcpu_lock.  pcpu_lock is released and reacquired
- * if area map is extended.
+ * Does GFP_KERNEL allocation.  Grabs and releases pcpu_lock.
  *
  * RETURNS:
- * 0 if noop, 1 if successfully extended, -errno on failure.
+ * 0 on success, -errno on failure.
  */
-static int pcpu_extend_area_map(struct pcpu_chunk *chunk)
+static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
 {
-       int new_alloc;
-       int *new;
-       size_t size;
+       int *old = NULL, *new = NULL;
+       size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
+       unsigned long flags;
 
-       /* has enough? */
-       if (chunk->map_alloc >= chunk->map_used + 2)
-               return 0;
+       new = pcpu_mem_zalloc(new_size);
+       if (!new)
+               return -ENOMEM;
 
-       spin_unlock_irq(&pcpu_lock);
+       /* acquire pcpu_lock and switch to new area map */
+       spin_lock_irqsave(&pcpu_lock, flags);
 
-       new_alloc = PCPU_DFL_MAP_ALLOC;
-       while (new_alloc < chunk->map_used + 2)
-               new_alloc *= 2;
+       if (new_alloc <= chunk->map_alloc)
+               goto out_unlock;
 
-       new = pcpu_mem_alloc(new_alloc * sizeof(new[0]));
-       if (!new) {
-               spin_lock_irq(&pcpu_lock);
-               return -ENOMEM;
-       }
+       old_size = chunk->map_alloc * sizeof(chunk->map[0]);
+       old = chunk->map;
 
-       /*
-        * Acquire pcpu_lock and switch to new area map.  Only free
-        * could have happened inbetween, so map_used couldn't have
-        * grown.
-        */
-       spin_lock_irq(&pcpu_lock);
-       BUG_ON(new_alloc < chunk->map_used + 2);
+       memcpy(new, old, old_size);
+
+       chunk->map_alloc = new_alloc;
+       chunk->map = new;
+       new = NULL;
 
-       size = chunk->map_alloc * sizeof(chunk->map[0]);
-       memcpy(new, chunk->map, size);
+out_unlock:
+       spin_unlock_irqrestore(&pcpu_lock, flags);
 
        /*
-        * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
-        * one of the first chunks and still using static map.
+        * pcpu_mem_free() might end up calling vfree() which uses
+        * IRQ-unsafe lock and thus can't be called under pcpu_lock.
         */
-       if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC)
-               pcpu_mem_free(chunk->map, size);
+       pcpu_mem_free(old, old_size);
+       pcpu_mem_free(new, new_size);
 
-       chunk->map_alloc = new_alloc;
-       chunk->map = new;
        return 0;
 }
 
@@ -426,7 +432,7 @@ static int pcpu_extend_area_map(struct pcpu_chunk *chunk)
  * depending on @head, is reduced by @tail bytes and @tail byte block
  * is inserted after the target block.
  *
- * @chunk->map must have enough free slots to accomodate the split.
+ * @chunk->map must have enough free slots to accommodate the split.
  *
  * CONTEXT:
  * pcpu_lock.
@@ -595,436 +601,93 @@ static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
        pcpu_chunk_relocate(chunk, oslot);
 }
 
-/**
- * pcpu_get_pages_and_bitmap - get temp pages array and bitmap
- * @chunk: chunk of interest
- * @bitmapp: output parameter for bitmap
- * @may_alloc: may allocate the array
- *
- * Returns pointer to array of pointers to struct page and bitmap,
- * both of which can be indexed with pcpu_page_idx().  The returned
- * array is cleared to zero and *@bitmapp is copied from
- * @chunk->populated.  Note that there is only one array and bitmap
- * and access exclusion is the caller's responsibility.
- *
- * CONTEXT:
- * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
- * Otherwise, don't care.
- *
- * RETURNS:
- * Pointer to temp pages array on success, NULL on failure.
- */
-static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk,
-                                              unsigned long **bitmapp,
-                                              bool may_alloc)
-{
-       static struct page **pages;
-       static unsigned long *bitmap;
-       size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
-       size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) *
-                            sizeof(unsigned long);
-
-       if (!pages || !bitmap) {
-               if (may_alloc && !pages)
-                       pages = pcpu_mem_alloc(pages_size);
-               if (may_alloc && !bitmap)
-                       bitmap = pcpu_mem_alloc(bitmap_size);
-               if (!pages || !bitmap)
-                       return NULL;
-       }
-
-       memset(pages, 0, pages_size);
-       bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages);
-
-       *bitmapp = bitmap;
-       return pages;
-}
-
-/**
- * pcpu_free_pages - free pages which were allocated for @chunk
- * @chunk: chunk pages were allocated for
- * @pages: array of pages to be freed, indexed by pcpu_page_idx()
- * @populated: populated bitmap
- * @page_start: page index of the first page to be freed
- * @page_end: page index of the last page to be freed + 1
- *
- * Free pages [@page_start and @page_end) in @pages for all units.
- * The pages were allocated for @chunk.
- */
-static void pcpu_free_pages(struct pcpu_chunk *chunk,
-                           struct page **pages, unsigned long *populated,
-                           int page_start, int page_end)
+static struct pcpu_chunk *pcpu_alloc_chunk(void)
 {
-       unsigned int cpu;
-       int i;
-
-       for_each_possible_cpu(cpu) {
-               for (i = page_start; i < page_end; i++) {
-                       struct page *page = pages[pcpu_page_idx(cpu, i)];
-
-                       if (page)
-                               __free_page(page);
-               }
-       }
-}
-
-/**
- * pcpu_alloc_pages - allocates pages for @chunk
- * @chunk: target chunk
- * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
- * @populated: populated bitmap
- * @page_start: page index of the first page to be allocated
- * @page_end: page index of the last page to be allocated + 1
- *
- * Allocate pages [@page_start,@page_end) into @pages for all units.
- * The allocation is for @chunk.  Percpu core doesn't care about the
- * content of @pages and will pass it verbatim to pcpu_map_pages().
- */
-static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
-                           struct page **pages, unsigned long *populated,
-                           int page_start, int page_end)
-{
-       const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
-       unsigned int cpu;
-       int i;
-
-       for_each_possible_cpu(cpu) {
-               for (i = page_start; i < page_end; i++) {
-                       struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
-
-                       *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
-                       if (!*pagep) {
-                               pcpu_free_pages(chunk, pages, populated,
-                                               page_start, page_end);
-                               return -ENOMEM;
-                       }
-               }
-       }
-       return 0;
-}
-
-/**
- * pcpu_pre_unmap_flush - flush cache prior to unmapping
- * @chunk: chunk the regions to be flushed belongs to
- * @page_start: page index of the first page to be flushed
- * @page_end: page index of the last page to be flushed + 1
- *
- * Pages in [@page_start,@page_end) of @chunk are about to be
- * unmapped.  Flush cache.  As 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.
- */
-static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
-                                int page_start, int page_end)
-{
-       flush_cache_vunmap(
-               pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
-               pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
-}
-
-static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
-{
-       unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
-}
-
-/**
- * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
- * @chunk: chunk of interest
- * @pages: pages array which can be used to pass information to free
- * @populated: populated bitmap
- * @page_start: page index of the first page to unmap
- * @page_end: page index of the last page to unmap + 1
- *
- * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
- * Corresponding elements in @pages were cleared by the caller and can
- * be used to carry information to pcpu_free_pages() which will be
- * called after all unmaps are finished.  The caller should call
- * proper pre/post flush functions.
- */
-static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
-                            struct page **pages, unsigned long *populated,
-                            int page_start, int page_end)
-{
-       unsigned int cpu;
-       int i;
+       struct pcpu_chunk *chunk;
 
-       for_each_possible_cpu(cpu) {
-               for (i = page_start; i < page_end; i++) {
-                       struct page *page;
+       chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
+       if (!chunk)
+               return NULL;
 
-                       page = pcpu_chunk_page(chunk, cpu, i);
-                       WARN_ON(!page);
-                       pages[pcpu_page_idx(cpu, i)] = page;
-               }
-               __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
-                                  page_end - page_start);
+       chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
+                                               sizeof(chunk->map[0]));
+       if (!chunk->map) {
+               kfree(chunk);
+               return NULL;
        }
 
-       for (i = page_start; i < page_end; i++)
-               __clear_bit(i, populated);
-}
+       chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
+       chunk->map[chunk->map_used++] = pcpu_unit_size;
 
-/**
- * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
- * @chunk: pcpu_chunk the regions to be flushed belong to
- * @page_start: page index of the first page to be flushed
- * @page_end: page index of the last page to be flushed + 1
- *
- * Pages [@page_start,@page_end) of @chunk have been unmapped.  Flush
- * TLB for the regions.  This can be skipped if the area is to be
- * returned to vmalloc as vmalloc will handle TLB flushing lazily.
- *
- * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
- * for the whole region.
- */
-static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
-                                     int page_start, int page_end)
-{
-       flush_tlb_kernel_range(
-               pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
-               pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
-}
+       INIT_LIST_HEAD(&chunk->list);
+       chunk->free_size = pcpu_unit_size;
+       chunk->contig_hint = pcpu_unit_size;
 
-static int __pcpu_map_pages(unsigned long addr, struct page **pages,
-                           int nr_pages)
-{
-       return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
-                                       PAGE_KERNEL, pages);
+       return chunk;
 }
 
-/**
- * pcpu_map_pages - map pages into a pcpu_chunk
- * @chunk: chunk of interest
- * @pages: pages array containing pages to be mapped
- * @populated: populated bitmap
- * @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.  The
- * caller is responsible for calling pcpu_post_map_flush() after all
- * mappings are complete.
- *
- * This function is responsible for setting corresponding bits in
- * @chunk->populated bitmap and whatever is necessary for reverse
- * lookup (addr -> chunk).
- */
-static int pcpu_map_pages(struct pcpu_chunk *chunk,
-                         struct page **pages, unsigned long *populated,
-                         int page_start, int page_end)
+static void pcpu_free_chunk(struct pcpu_chunk *chunk)
 {
-       unsigned int cpu, tcpu;
-       int i, err;
-
-       for_each_possible_cpu(cpu) {
-               err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
-                                      &pages[pcpu_page_idx(cpu, page_start)],
-                                      page_end - page_start);
-               if (err < 0)
-                       goto err;
-       }
-
-       /* mapping successful, link chunk and mark populated */
-       for (i = page_start; i < page_end; i++) {
-               for_each_possible_cpu(cpu)
-                       pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
-                                           chunk);
-               __set_bit(i, populated);
-       }
-
-       return 0;
-
-err:
-       for_each_possible_cpu(tcpu) {
-               if (tcpu == cpu)
-                       break;
-               __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
-                                  page_end - page_start);
-       }
-       return err;
+       if (!chunk)
+               return;
+       pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
+       pcpu_mem_free(chunk, pcpu_chunk_struct_size);
 }
 
-/**
- * pcpu_post_map_flush - flush cache after mapping
- * @chunk: pcpu_chunk the regions to be flushed belong to
- * @page_start: page index of the first page to be flushed
- * @page_end: page index of the last page to be flushed + 1
- *
- * Pages [@page_start,@page_end) of @chunk have been mapped.  Flush
- * cache.
- *
- * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
- * for the whole region.
+/*
+ * Chunk management implementation.
+ *
+ * To allow different implementations, chunk alloc/free and
+ * [de]population are implemented in a separate file which is pulled
+ * into this file and compiled together.  The following functions
+ * should be implemented.
+ *
+ * pcpu_populate_chunk         - populate the specified range of a chunk
+ * pcpu_depopulate_chunk       - depopulate the specified range of a chunk
+ * pcpu_create_chunk           - create a new chunk
+ * pcpu_destroy_chunk          - destroy a chunk, always preceded by full depop
+ * pcpu_addr_to_page           - translate address to physical address
+ * pcpu_verify_alloc_info      - check alloc_info is acceptable during init
  */
-static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
-                               int page_start, int page_end)
-{
-       flush_cache_vmap(
-               pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
-               pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
-}
+static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
+static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
+static struct pcpu_chunk *pcpu_create_chunk(void);
+static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
+static struct page *pcpu_addr_to_page(void *addr);
+static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
+
+#ifdef CONFIG_NEED_PER_CPU_KM
+#include "percpu-km.c"
+#else
+#include "percpu-vm.c"
+#endif
 
 /**
- * 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 in bytes
- * @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.
+ * pcpu_chunk_addr_search - determine chunk containing specified address
+ * @addr: address for which the chunk needs to be determined.
  *
- * CONTEXT:
- * pcpu_alloc_mutex.
+ * RETURNS:
+ * The address of the found chunk.
  */
-static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size)
+static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
 {
-       int page_start = PFN_DOWN(off);
-       int page_end = PFN_UP(off + size);
-       struct page **pages;
-       unsigned long *populated;
-       int rs, re;
-
-       /* quick path, check whether it's empty already */
-       pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
-               if (rs == page_start && re == page_end)
-                       return;
-               break;
+       /* is it in the first chunk? */
+       if (pcpu_addr_in_first_chunk(addr)) {
+               /* is it in the reserved area? */
+               if (pcpu_addr_in_reserved_chunk(addr))
+                       return pcpu_reserved_chunk;
+               return pcpu_first_chunk;
        }
 
-       /* immutable chunks can't be depopulated */
-       WARN_ON(chunk->immutable);
-
        /*
-        * If control reaches here, there must have been at least one
-        * successful population attempt so the temp pages array must
-        * be available now.
+        * The address is relative to unit0 which might be unused and
+        * thus unmapped.  Offset the address to the unit space of the
+        * current processor before looking it up in the vmalloc
+        * space.  Note that any possible cpu id can be used here, so
+        * there's no need to worry about preemption or cpu hotplug.
         */
-       pages = pcpu_get_pages_and_bitmap(chunk, &populated, false);
-       BUG_ON(!pages);
-
-       /* unmap and free */
-       pcpu_pre_unmap_flush(chunk, page_start, page_end);
-
-       pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
-               pcpu_unmap_pages(chunk, pages, populated, rs, re);
-
-       /* no need to flush tlb, vmalloc will handle it lazily */
-
-       pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
-               pcpu_free_pages(chunk, pages, populated, rs, re);
-
-       /* commit new bitmap */
-       bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
-}
-
-/**
- * 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 in bytes
- *
- * For each cpu, populate and map pages [@page_start,@page_end) into
- * @chunk.  The area is cleared on return.
- *
- * CONTEXT:
- * pcpu_alloc_mutex, does GFP_KERNEL allocation.
- */
-static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
-{
-       int page_start = PFN_DOWN(off);
-       int page_end = PFN_UP(off + size);
-       int free_end = page_start, unmap_end = page_start;
-       struct page **pages;
-       unsigned long *populated;
-       unsigned int cpu;
-       int rs, re, rc;
-
-       /* quick path, check whether all pages are already there */
-       pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) {
-               if (rs == page_start && re == page_end)
-                       goto clear;
-               break;
-       }
-
-       /* need to allocate and map pages, this chunk can't be immutable */
-       WARN_ON(chunk->immutable);
-
-       pages = pcpu_get_pages_and_bitmap(chunk, &populated, true);
-       if (!pages)
-               return -ENOMEM;
-
-       /* alloc and map */
-       pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
-               rc = pcpu_alloc_pages(chunk, pages, populated, rs, re);
-               if (rc)
-                       goto err_free;
-               free_end = re;
-       }
-
-       pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
-               rc = pcpu_map_pages(chunk, pages, populated, rs, re);
-               if (rc)
-                       goto err_unmap;
-               unmap_end = re;
-       }
-       pcpu_post_map_flush(chunk, page_start, page_end);
-
-       /* commit new bitmap */
-       bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
-clear:
-       for_each_possible_cpu(cpu)
-               memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
-       return 0;
-
-err_unmap:
-       pcpu_pre_unmap_flush(chunk, page_start, unmap_end);
-       pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end)
-               pcpu_unmap_pages(chunk, pages, populated, rs, re);
-       pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end);
-err_free:
-       pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end)
-               pcpu_free_pages(chunk, pages, populated, rs, re);
-       return rc;
-}
-
-static void free_pcpu_chunk(struct pcpu_chunk *chunk)
-{
-       if (!chunk)
-               return;
-       if (chunk->vms)
-               pcpu_free_vm_areas(chunk->vms, pcpu_nr_groups);
-       pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[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_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
-       chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
-       chunk->map[chunk->map_used++] = pcpu_unit_size;
-
-       chunk->vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
-                                      pcpu_nr_groups, pcpu_atom_size,
-                                      GFP_KERNEL);
-       if (!chunk->vms) {
-               free_pcpu_chunk(chunk);
-               return NULL;
-       }
-
-       INIT_LIST_HEAD(&chunk->list);
-       chunk->free_size = pcpu_unit_size;
-       chunk->contig_hint = pcpu_unit_size;
-       chunk->base_addr = chunk->vms[0]->addr - pcpu_group_offsets[0];
-
-       return chunk;
+       addr += pcpu_unit_offsets[raw_smp_processor_id()];
+       return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
 }
 
 /**
@@ -1041,10 +704,14 @@ static struct pcpu_chunk *alloc_pcpu_chunk(void)
  * RETURNS:
  * Percpu pointer to the allocated area on success, NULL on failure.
  */
-static void *pcpu_alloc(size_t size, size_t align, bool reserved)
+static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
 {
+       static int warn_limit = 10;
        struct pcpu_chunk *chunk;
-       int slot, off;
+       const char *err;
+       int slot, off, new_alloc;
+       unsigned long flags;
+       void __percpu *ptr;
 
        if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
                WARN(true, "illegal size (%zu) or align (%zu) for "
@@ -1053,17 +720,31 @@ static void *pcpu_alloc(size_t size, size_t align, bool reserved)
        }
 
        mutex_lock(&pcpu_alloc_mutex);
-       spin_lock_irq(&pcpu_lock);
+       spin_lock_irqsave(&pcpu_lock, flags);
 
        /* serve reserved allocations from the reserved chunk if available */
        if (reserved && pcpu_reserved_chunk) {
                chunk = pcpu_reserved_chunk;
-               if (size > chunk->contig_hint ||
-                   pcpu_extend_area_map(chunk) < 0)
+
+               if (size > chunk->contig_hint) {
+                       err = "alloc from reserved chunk failed";
                        goto fail_unlock;
+               }
+
+               while ((new_alloc = pcpu_need_to_extend(chunk))) {
+                       spin_unlock_irqrestore(&pcpu_lock, flags);
+                       if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
+                               err = "failed to extend area map of reserved chunk";
+                               goto fail_unlock_mutex;
+                       }
+                       spin_lock_irqsave(&pcpu_lock, flags);
+               }
+
                off = pcpu_alloc_area(chunk, size, align);
                if (off >= 0)
                        goto area_found;
+
+               err = "alloc from reserved chunk failed";
                goto fail_unlock;
        }
 
@@ -1074,13 +755,20 @@ restart:
                        if (size > chunk->contig_hint)
                                continue;
 
-                       switch (pcpu_extend_area_map(chunk)) {
-                       case 0:
-                               break;
-                       case 1:
-                               goto restart;   /* pcpu_lock dropped, restart */
-                       default:
-                               goto fail_unlock;
+                       new_alloc = pcpu_need_to_extend(chunk);
+                       if (new_alloc) {
+                               spin_unlock_irqrestore(&pcpu_lock, flags);
+                               if (pcpu_extend_area_map(chunk,
+                                                        new_alloc) < 0) {
+                                       err = "failed to extend area map";
+                                       goto fail_unlock_mutex;
+                               }
+                               spin_lock_irqsave(&pcpu_lock, flags);
+                               /*
+                                * pcpu_lock has been dropped, need to
+                                * restart cpu_slot list walking.
+                                */
+                               goto restart;
                        }
 
                        off = pcpu_alloc_area(chunk, size, align);
@@ -1090,35 +778,47 @@ restart:
        }
 
        /* hmmm... no space left, create a new chunk */
-       spin_unlock_irq(&pcpu_lock);
+       spin_unlock_irqrestore(&pcpu_lock, flags);
 
-       chunk = alloc_pcpu_chunk();
-       if (!chunk)
+       chunk = pcpu_create_chunk();
+       if (!chunk) {
+               err = "failed to allocate new chunk";
                goto fail_unlock_mutex;
+       }
 
-       spin_lock_irq(&pcpu_lock);
+       spin_lock_irqsave(&pcpu_lock, flags);
        pcpu_chunk_relocate(chunk, -1);
        goto restart;
 
 area_found:
-       spin_unlock_irq(&pcpu_lock);
+       spin_unlock_irqrestore(&pcpu_lock, flags);
 
        /* populate, map and clear the area */
        if (pcpu_populate_chunk(chunk, off, size)) {
-               spin_lock_irq(&pcpu_lock);
+               spin_lock_irqsave(&pcpu_lock, flags);
                pcpu_free_area(chunk, off);
+               err = "failed to populate";
                goto fail_unlock;
        }
 
        mutex_unlock(&pcpu_alloc_mutex);
 
        /* return address relative to base address */
-       return __addr_to_pcpu_ptr(chunk->base_addr + off);
+       ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
+       kmemleak_alloc_percpu(ptr, size);
+       return ptr;
 
 fail_unlock:
-       spin_unlock_irq(&pcpu_lock);
+       spin_unlock_irqrestore(&pcpu_lock, flags);
 fail_unlock_mutex:
        mutex_unlock(&pcpu_alloc_mutex);
+       if (warn_limit) {
+               pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
+                          "%s\n", size, align, err);
+               dump_stack();
+               if (!--warn_limit)
+                       pr_info("PERCPU: limit reached, disable warning\n");
+       }
        return NULL;
 }
 
@@ -1127,8 +827,8 @@ fail_unlock_mutex:
  * @size: size of area to allocate in bytes
  * @align: alignment of area (max PAGE_SIZE)
  *
- * Allocate percpu area of @size bytes aligned at @align.  Might
- * sleep.  Might trigger writeouts.
+ * Allocate zero-filled percpu area of @size bytes aligned at @align.
+ * Might sleep.  Might trigger writeouts.
  *
  * CONTEXT:
  * Does GFP_KERNEL allocation.
@@ -1136,7 +836,7 @@ fail_unlock_mutex:
  * RETURNS:
  * Percpu pointer to the allocated area on success, NULL on failure.
  */
-void *__alloc_percpu(size_t size, size_t align)
+void __percpu *__alloc_percpu(size_t size, size_t align)
 {
        return pcpu_alloc(size, align, false);
 }
@@ -1147,9 +847,10 @@ EXPORT_SYMBOL_GPL(__alloc_percpu);
  * @size: size of area to allocate in bytes
  * @align: alignment of area (max PAGE_SIZE)
  *
- * Allocate percpu area of @size bytes aligned at @align from reserved
- * percpu area if arch has set it up; otherwise, allocation is served
- * from the same dynamic area.  Might sleep.  Might trigger writeouts.
+ * Allocate zero-filled percpu area of @size bytes aligned at @align
+ * from reserved percpu area if arch has set it up; otherwise,
+ * allocation is served from the same dynamic area.  Might sleep.
+ * Might trigger writeouts.
  *
  * CONTEXT:
  * Does GFP_KERNEL allocation.
@@ -1157,7 +858,7 @@ EXPORT_SYMBOL_GPL(__alloc_percpu);
  * RETURNS:
  * Percpu pointer to the allocated area on success, NULL on failure.
  */
-void *__alloc_reserved_percpu(size_t size, size_t align)
+void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
 {
        return pcpu_alloc(size, align, true);
 }
@@ -1194,7 +895,7 @@ static void pcpu_reclaim(struct work_struct *work)
 
        list_for_each_entry_safe(chunk, next, &todo, list) {
                pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
-               free_pcpu_chunk(chunk);
+               pcpu_destroy_chunk(chunk);
        }
 
        mutex_unlock(&pcpu_alloc_mutex);
@@ -1209,9 +910,9 @@ static void pcpu_reclaim(struct work_struct *work)
  * CONTEXT:
  * Can be called from atomic context.
  */
-void free_percpu(void *ptr)
+void free_percpu(void __percpu *ptr)
 {
-       void *addr = __pcpu_ptr_to_addr(ptr);
+       void *addr;
        struct pcpu_chunk *chunk;
        unsigned long flags;
        int off;
@@ -1219,6 +920,10 @@ void free_percpu(void *ptr)
        if (!ptr)
                return;
 
+       kmemleak_free_percpu(ptr);
+
+       addr = __pcpu_ptr_to_addr(ptr);
+
        spin_lock_irqsave(&pcpu_lock, flags);
 
        chunk = pcpu_chunk_addr_search(addr);
@@ -1241,18 +946,94 @@ void free_percpu(void *ptr)
 }
 EXPORT_SYMBOL_GPL(free_percpu);
 
-static inline size_t pcpu_calc_fc_sizes(size_t static_size,
-                                       size_t reserved_size,
-                                       ssize_t *dyn_sizep)
+/**
+ * is_kernel_percpu_address - test whether address is from static percpu area
+ * @addr: address to test
+ *
+ * Test whether @addr belongs to in-kernel static percpu area.  Module
+ * static percpu areas are not considered.  For those, use
+ * is_module_percpu_address().
+ *
+ * RETURNS:
+ * %true if @addr is from in-kernel static percpu area, %false otherwise.
+ */
+bool is_kernel_percpu_address(unsigned long addr)
 {
-       size_t size_sum;
+#ifdef CONFIG_SMP
+       const size_t static_size = __per_cpu_end - __per_cpu_start;
+       void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
+       unsigned int cpu;
 
-       size_sum = PFN_ALIGN(static_size + reserved_size +
-                            (*dyn_sizep >= 0 ? *dyn_sizep : 0));
-       if (*dyn_sizep != 0)
-               *dyn_sizep = size_sum - static_size - reserved_size;
+       for_each_possible_cpu(cpu) {
+               void *start = per_cpu_ptr(base, cpu);
+
+               if ((void *)addr >= start && (void *)addr < start + static_size)
+                       return true;
+        }
+#endif
+       /* on UP, can't distinguish from other static vars, always false */
+       return false;
+}
+
+/**
+ * per_cpu_ptr_to_phys - convert translated percpu address to physical address
+ * @addr: the address to be converted to physical address
+ *
+ * Given @addr which is dereferenceable address obtained via one of
+ * percpu access macros, this function translates it into its physical
+ * address.  The caller is responsible for ensuring @addr stays valid
+ * until this function finishes.
+ *
+ * percpu allocator has special setup for the first chunk, which currently
+ * supports either embedding in linear address space or vmalloc mapping,
+ * and, from the second one, the backing allocator (currently either vm or
+ * km) provides translation.
+ *
+ * The addr can be tranlated simply without checking if it falls into the
+ * first chunk. But the current code reflects better how percpu allocator
+ * actually works, and the verification can discover both bugs in percpu
+ * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
+ * code.
+ *
+ * RETURNS:
+ * The physical address for @addr.
+ */
+phys_addr_t per_cpu_ptr_to_phys(void *addr)
+{
+       void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
+       bool in_first_chunk = false;
+       unsigned long first_low, first_high;
+       unsigned int cpu;
+
+       /*
+        * The following test on unit_low/high isn't strictly
+        * necessary but will speed up lookups of addresses which
+        * aren't in the first chunk.
+        */
+       first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
+       first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
+                                    pcpu_unit_pages);
+       if ((unsigned long)addr >= first_low &&
+           (unsigned long)addr < first_high) {
+               for_each_possible_cpu(cpu) {
+                       void *start = per_cpu_ptr(base, cpu);
+
+                       if (addr >= start && addr < start + pcpu_unit_size) {
+                               in_first_chunk = true;
+                               break;
+                       }
+               }
+       }
 
-       return size_sum;
+       if (in_first_chunk) {
+               if (!is_vmalloc_addr(addr))
+                       return __pa(addr);
+               else
+                       return page_to_phys(vmalloc_to_page(addr)) +
+                              offset_in_page(addr);
+       } else
+               return page_to_phys(pcpu_addr_to_page(addr)) +
+                      offset_in_page(addr);
 }
 
 /**
@@ -1311,154 +1092,6 @@ void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
 }
 
 /**
- * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
- * @reserved_size: the size of reserved percpu area in bytes
- * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
- * @atom_size: allocation atom size
- * @cpu_distance_fn: callback to determine distance between cpus, optional
- *
- * This function determines grouping of units, their mappings to cpus
- * and other parameters considering needed percpu size, allocation
- * atom size and distances between CPUs.
- *
- * Groups are always mutliples of atom size and CPUs which are of
- * LOCAL_DISTANCE both ways are grouped together and share space for
- * units in the same group.  The returned configuration is guaranteed
- * to have CPUs on different nodes on different groups and >=75% usage
- * of allocated virtual address space.
- *
- * RETURNS:
- * On success, pointer to the new allocation_info is returned.  On
- * failure, ERR_PTR value is returned.
- */
-struct pcpu_alloc_info * __init pcpu_build_alloc_info(
-                               size_t reserved_size, ssize_t dyn_size,
-                               size_t atom_size,
-                               pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
-{
-       static int group_map[NR_CPUS] __initdata;
-       static int group_cnt[NR_CPUS] __initdata;
-       const size_t static_size = __per_cpu_end - __per_cpu_start;
-       int group_cnt_max = 0, nr_groups = 1, nr_units = 0;
-       size_t size_sum, min_unit_size, alloc_size;
-       int upa, max_upa, uninitialized_var(best_upa);  /* units_per_alloc */
-       int last_allocs, group, unit;
-       unsigned int cpu, tcpu;
-       struct pcpu_alloc_info *ai;
-       unsigned int *cpu_map;
-
-       /*
-        * Determine min_unit_size, alloc_size and max_upa such that
-        * alloc_size is multiple of atom_size and is the smallest
-        * which can accomodate 4k aligned segments which are equal to
-        * or larger than min_unit_size.
-        */
-       size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size);
-       min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
-
-       alloc_size = roundup(min_unit_size, atom_size);
-       upa = alloc_size / min_unit_size;
-       while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
-               upa--;
-       max_upa = upa;
-
-       /* group cpus according to their proximity */
-       for_each_possible_cpu(cpu) {
-               group = 0;
-       next_group:
-               for_each_possible_cpu(tcpu) {
-                       if (cpu == tcpu)
-                               break;
-                       if (group_map[tcpu] == group && cpu_distance_fn &&
-                           (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
-                            cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
-                               group++;
-                               nr_groups = max(nr_groups, group + 1);
-                               goto next_group;
-                       }
-               }
-               group_map[cpu] = group;
-               group_cnt[group]++;
-               group_cnt_max = max(group_cnt_max, group_cnt[group]);
-       }
-
-       /*
-        * Expand unit size until address space usage goes over 75%
-        * and then as much as possible without using more address
-        * space.
-        */
-       last_allocs = INT_MAX;
-       for (upa = max_upa; upa; upa--) {
-               int allocs = 0, wasted = 0;
-
-               if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
-                       continue;
-
-               for (group = 0; group < nr_groups; group++) {
-                       int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
-                       allocs += this_allocs;
-                       wasted += this_allocs * upa - group_cnt[group];
-               }
-
-               /*
-                * Don't accept if wastage is over 25%.  The
-                * greater-than comparison ensures upa==1 always
-                * passes the following check.
-                */
-               if (wasted > num_possible_cpus() / 3)
-                       continue;
-
-               /* and then don't consume more memory */
-               if (allocs > last_allocs)
-                       break;
-               last_allocs = allocs;
-               best_upa = upa;
-       }
-       upa = best_upa;
-
-       /* allocate and fill alloc_info */
-       for (group = 0; group < nr_groups; group++)
-               nr_units += roundup(group_cnt[group], upa);
-
-       ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
-       if (!ai)
-               return ERR_PTR(-ENOMEM);
-       cpu_map = ai->groups[0].cpu_map;
-
-       for (group = 0; group < nr_groups; group++) {
-               ai->groups[group].cpu_map = cpu_map;
-               cpu_map += roundup(group_cnt[group], upa);
-       }
-
-       ai->static_size = static_size;
-       ai->reserved_size = reserved_size;
-       ai->dyn_size = dyn_size;
-       ai->unit_size = alloc_size / upa;
-       ai->atom_size = atom_size;
-       ai->alloc_size = alloc_size;
-
-       for (group = 0, unit = 0; group_cnt[group]; group++) {
-               struct pcpu_group_info *gi = &ai->groups[group];
-
-               /*
-                * Initialize base_offset as if all groups are located
-                * back-to-back.  The caller should update this to
-                * reflect actual allocation.
-                */
-               gi->base_offset = unit * ai->unit_size;
-
-               for_each_possible_cpu(cpu)
-                       if (group_map[cpu] == group)
-                               gi->cpu_map[gi->nr_units++] = cpu;
-               gi->nr_units = roundup(gi->nr_units, upa);
-               unit += gi->nr_units;
-       }
-       BUG_ON(unit != nr_units);
-
-       return ai;
-}
-
-/**
  * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
  * @lvl: loglevel
  * @ai: allocation info to dump
@@ -1499,20 +1132,20 @@ static void pcpu_dump_alloc_info(const char *lvl,
                for (alloc_end += gi->nr_units / upa;
                     alloc < alloc_end; alloc++) {
                        if (!(alloc % apl)) {
-                               printk("\n");
+                               printk(KERN_CONT "\n");
                                printk("%spcpu-alloc: ", lvl);
                        }
-                       printk("[%0*d] ", group_width, group);
+                       printk(KERN_CONT "[%0*d] ", group_width, group);
 
                        for (unit_end += upa; unit < unit_end; unit++)
                                if (gi->cpu_map[unit] != NR_CPUS)
-                                       printk("%0*d ", cpu_width,
+                                       printk(KERN_CONT "%0*d ", cpu_width,
                                               gi->cpu_map[unit]);
                                else
-                                       printk("%s ", empty_str);
+                                       printk(KERN_CONT "%s ", empty_str);
                }
        }
-       printk("\n");
+       printk(KERN_CONT "\n");
 }
 
 /**
@@ -1574,7 +1207,9 @@ static void pcpu_dump_alloc_info(const char *lvl,
 int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
                                  void *base_addr)
 {
-       static int smap[2], dmap[2];
+       static char cpus_buf[4096] __initdata;
+       static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
+       static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
        size_t dyn_size = ai->dyn_size;
        size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
        struct pcpu_chunk *schunk, *dchunk = NULL;
@@ -1585,17 +1220,30 @@ int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
        int *unit_map;
        int group, unit, i;
 
-       /* sanity checks */
-       BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
-                    ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
-       BUG_ON(ai->nr_groups <= 0);
-       BUG_ON(!ai->static_size);
-       BUG_ON(!base_addr);
-       BUG_ON(ai->unit_size < size_sum);
-       BUG_ON(ai->unit_size & ~PAGE_MASK);
-       BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
+       cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
 
-       pcpu_dump_alloc_info(KERN_DEBUG, ai);
+#define PCPU_SETUP_BUG_ON(cond)        do {                                    \
+       if (unlikely(cond)) {                                           \
+               pr_emerg("PERCPU: failed to initialize, %s", #cond);    \
+               pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf);   \
+               pcpu_dump_alloc_info(KERN_EMERG, ai);                   \
+               BUG();                                                  \
+       }                                                               \
+} while (0)
+
+       /* sanity checks */
+       PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
+#ifdef CONFIG_SMP
+       PCPU_SETUP_BUG_ON(!ai->static_size);
+       PCPU_SETUP_BUG_ON((unsigned long)__per_cpu_start & ~PAGE_MASK);
+#endif
+       PCPU_SETUP_BUG_ON(!base_addr);
+       PCPU_SETUP_BUG_ON((unsigned long)base_addr & ~PAGE_MASK);
+       PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
+       PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
+       PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
+       PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
+       PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
 
        /* process group information and build config tables accordingly */
        group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0]));
@@ -1604,8 +1252,10 @@ int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
        unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0]));
 
        for (cpu = 0; cpu < nr_cpu_ids; cpu++)
-               unit_map[cpu] = NR_CPUS;
-       pcpu_first_unit_cpu = NR_CPUS;
+               unit_map[cpu] = UINT_MAX;
+
+       pcpu_low_unit_cpu = NR_CPUS;
+       pcpu_high_unit_cpu = NR_CPUS;
 
        for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
                const struct pcpu_group_info *gi = &ai->groups[group];
@@ -1618,21 +1268,30 @@ int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
                        if (cpu == NR_CPUS)
                                continue;
 
-                       BUG_ON(cpu > nr_cpu_ids || !cpu_possible(cpu));
-                       BUG_ON(unit_map[cpu] != NR_CPUS);
+                       PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
+                       PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
+                       PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
 
                        unit_map[cpu] = unit + i;
                        unit_off[cpu] = gi->base_offset + i * ai->unit_size;
 
-                       if (pcpu_first_unit_cpu == NR_CPUS)
-                               pcpu_first_unit_cpu = cpu;
+                       /* determine low/high unit_cpu */
+                       if (pcpu_low_unit_cpu == NR_CPUS ||
+                           unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
+                               pcpu_low_unit_cpu = cpu;
+                       if (pcpu_high_unit_cpu == NR_CPUS ||
+                           unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
+                               pcpu_high_unit_cpu = cpu;
                }
        }
-       pcpu_last_unit_cpu = cpu;
        pcpu_nr_units = unit;
 
        for_each_possible_cpu(cpu)
-               BUG_ON(unit_map[cpu] == NR_CPUS);
+               PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
+
+       /* we're done parsing the input, undefine BUG macro and dump config */
+#undef PCPU_SETUP_BUG_ON
+       pcpu_dump_alloc_info(KERN_DEBUG, ai);
 
        pcpu_nr_groups = ai->nr_groups;
        pcpu_group_offsets = group_offsets;
@@ -1709,17 +1368,21 @@ int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
        return 0;
 }
 
-const char *pcpu_fc_names[PCPU_FC_NR] __initdata = {
+#ifdef CONFIG_SMP
+
+const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
        [PCPU_FC_AUTO]  = "auto",
        [PCPU_FC_EMBED] = "embed",
        [PCPU_FC_PAGE]  = "page",
-       [PCPU_FC_LPAGE] = "lpage",
 };
 
 enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
 
 static int __init percpu_alloc_setup(char *str)
 {
+       if (!str)
+               return -EINVAL;
+
        if (0)
                /* nada */;
 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
@@ -1730,10 +1393,6 @@ static int __init percpu_alloc_setup(char *str)
        else if (!strcmp(str, "page"))
                pcpu_chosen_fc = PCPU_FC_PAGE;
 #endif
-#ifdef CONFIG_NEED_PER_CPU_LPAGE_FIRST_CHUNK
-       else if (!strcmp(str, "lpage"))
-               pcpu_chosen_fc = PCPU_FC_LPAGE;
-#endif
        else
                pr_warning("PERCPU: unknown allocator %s specified\n", str);
 
@@ -1741,86 +1400,328 @@ static int __init percpu_alloc_setup(char *str)
 }
 early_param("percpu_alloc", percpu_alloc_setup);
 
+/*
+ * pcpu_embed_first_chunk() is used by the generic percpu setup.
+ * Build it if needed by the arch config or the generic setup is going
+ * to be used.
+ */
 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
        !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
+#define BUILD_EMBED_FIRST_CHUNK
+#endif
+
+/* build pcpu_page_first_chunk() iff needed by the arch config */
+#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
+#define BUILD_PAGE_FIRST_CHUNK
+#endif
+
+/* pcpu_build_alloc_info() is used by both embed and page first chunk */
+#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
+/**
+ * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
+ * @reserved_size: the size of reserved percpu area in bytes
+ * @dyn_size: minimum free size for dynamic allocation in bytes
+ * @atom_size: allocation atom size
+ * @cpu_distance_fn: callback to determine distance between cpus, optional
+ *
+ * This function determines grouping of units, their mappings to cpus
+ * and other parameters considering needed percpu size, allocation
+ * atom size and distances between CPUs.
+ *
+ * Groups are always mutliples of atom size and CPUs which are of
+ * LOCAL_DISTANCE both ways are grouped together and share space for
+ * units in the same group.  The returned configuration is guaranteed
+ * to have CPUs on different nodes on different groups and >=75% usage
+ * of allocated virtual address space.
+ *
+ * RETURNS:
+ * On success, pointer to the new allocation_info is returned.  On
+ * failure, ERR_PTR value is returned.
+ */
+static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
+                               size_t reserved_size, size_t dyn_size,
+                               size_t atom_size,
+                               pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
+{
+       static int group_map[NR_CPUS] __initdata;
+       static int group_cnt[NR_CPUS] __initdata;
+       const size_t static_size = __per_cpu_end - __per_cpu_start;
+       int nr_groups = 1, nr_units = 0;
+       size_t size_sum, min_unit_size, alloc_size;
+       int upa, max_upa, uninitialized_var(best_upa);  /* units_per_alloc */
+       int last_allocs, group, unit;
+       unsigned int cpu, tcpu;
+       struct pcpu_alloc_info *ai;
+       unsigned int *cpu_map;
+
+       /* this function may be called multiple times */
+       memset(group_map, 0, sizeof(group_map));
+       memset(group_cnt, 0, sizeof(group_cnt));
+
+       /* calculate size_sum and ensure dyn_size is enough for early alloc */
+       size_sum = PFN_ALIGN(static_size + reserved_size +
+                           max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
+       dyn_size = size_sum - static_size - reserved_size;
+
+       /*
+        * Determine min_unit_size, alloc_size and max_upa such that
+        * alloc_size is multiple of atom_size and is the smallest
+        * which can accommodate 4k aligned segments which are equal to
+        * or larger than min_unit_size.
+        */
+       min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
+
+       alloc_size = roundup(min_unit_size, atom_size);
+       upa = alloc_size / min_unit_size;
+       while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
+               upa--;
+       max_upa = upa;
+
+       /* group cpus according to their proximity */
+       for_each_possible_cpu(cpu) {
+               group = 0;
+       next_group:
+               for_each_possible_cpu(tcpu) {
+                       if (cpu == tcpu)
+                               break;
+                       if (group_map[tcpu] == group && cpu_distance_fn &&
+                           (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
+                            cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
+                               group++;
+                               nr_groups = max(nr_groups, group + 1);
+                               goto next_group;
+                       }
+               }
+               group_map[cpu] = group;
+               group_cnt[group]++;
+       }
+
+       /*
+        * Expand unit size until address space usage goes over 75%
+        * and then as much as possible without using more address
+        * space.
+        */
+       last_allocs = INT_MAX;
+       for (upa = max_upa; upa; upa--) {
+               int allocs = 0, wasted = 0;
+
+               if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
+                       continue;
+
+               for (group = 0; group < nr_groups; group++) {
+                       int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
+                       allocs += this_allocs;
+                       wasted += this_allocs * upa - group_cnt[group];
+               }
+
+               /*
+                * Don't accept if wastage is over 1/3.  The
+                * greater-than comparison ensures upa==1 always
+                * passes the following check.
+                */
+               if (wasted > num_possible_cpus() / 3)
+                       continue;
+
+               /* and then don't consume more memory */
+               if (allocs > last_allocs)
+                       break;
+               last_allocs = allocs;
+               best_upa = upa;
+       }
+       upa = best_upa;
+
+       /* allocate and fill alloc_info */
+       for (group = 0; group < nr_groups; group++)
+               nr_units += roundup(group_cnt[group], upa);
+
+       ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
+       if (!ai)
+               return ERR_PTR(-ENOMEM);
+       cpu_map = ai->groups[0].cpu_map;
+
+       for (group = 0; group < nr_groups; group++) {
+               ai->groups[group].cpu_map = cpu_map;
+               cpu_map += roundup(group_cnt[group], upa);
+       }
+
+       ai->static_size = static_size;
+       ai->reserved_size = reserved_size;
+       ai->dyn_size = dyn_size;
+       ai->unit_size = alloc_size / upa;
+       ai->atom_size = atom_size;
+       ai->alloc_size = alloc_size;
+
+       for (group = 0, unit = 0; group_cnt[group]; group++) {
+               struct pcpu_group_info *gi = &ai->groups[group];
+
+               /*
+                * Initialize base_offset as if all groups are located
+                * back-to-back.  The caller should update this to
+                * reflect actual allocation.
+                */
+               gi->base_offset = unit * ai->unit_size;
+
+               for_each_possible_cpu(cpu)
+                       if (group_map[cpu] == group)
+                               gi->cpu_map[gi->nr_units++] = cpu;
+               gi->nr_units = roundup(gi->nr_units, upa);
+               unit += gi->nr_units;
+       }
+       BUG_ON(unit != nr_units);
+
+       return ai;
+}
+#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
+
+#if defined(BUILD_EMBED_FIRST_CHUNK)
 /**
  * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
  * @reserved_size: the size of reserved percpu area in bytes
- * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
+ * @dyn_size: minimum free size for dynamic allocation in bytes
+ * @atom_size: allocation atom size
+ * @cpu_distance_fn: callback to determine distance between cpus, optional
+ * @alloc_fn: function to allocate percpu page
+ * @free_fn: function to free percpu page
  *
  * This is a helper to ease setting up embedded first percpu chunk and
  * can be called where pcpu_setup_first_chunk() is expected.
  *
  * If this function is used to setup the first chunk, it is allocated
- * as a contiguous area using bootmem allocator and used as-is without
- * being mapped into vmalloc area.  This enables the first chunk to
- * piggy back on the linear physical mapping which often uses larger
- * page size.
+ * by calling @alloc_fn and used as-is without being mapped into
+ * vmalloc area.  Allocations are always whole multiples of @atom_size
+ * aligned to @atom_size.
+ *
+ * This enables the first chunk to piggy back on the linear physical
+ * mapping which often uses larger page size.  Please note that this
+ * can result in very sparse cpu->unit mapping on NUMA machines thus
+ * requiring large vmalloc address space.  Don't use this allocator if
+ * vmalloc space is not orders of magnitude larger than distances
+ * between node memory addresses (ie. 32bit NUMA machines).
  *
- * When @dyn_size is positive, dynamic area might be larger than
- * specified to fill page alignment.  When @dyn_size is auto,
- * @dyn_size is just big enough to fill page alignment after static
- * and reserved areas.
+ * @dyn_size specifies the minimum dynamic area size.
  *
  * If the needed size is smaller than the minimum or specified unit
- * size, the leftover is returned to the bootmem allocator.
+ * size, the leftover is returned using @free_fn.
  *
  * RETURNS:
  * 0 on success, -errno on failure.
  */
-int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size)
+int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
+                                 size_t atom_size,
+                                 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
+                                 pcpu_fc_alloc_fn_t alloc_fn,
+                                 pcpu_fc_free_fn_t free_fn)
 {
+       void *base = (void *)ULONG_MAX;
+       void **areas = NULL;
        struct pcpu_alloc_info *ai;
-       size_t size_sum, chunk_size;
-       void *base;
-       int unit;
-       int rc;
+       size_t size_sum, areas_size, max_distance;
+       int group, i, rc;
 
-       ai = pcpu_build_alloc_info(reserved_size, dyn_size, PAGE_SIZE, NULL);
+       ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
+                                  cpu_distance_fn);
        if (IS_ERR(ai))
                return PTR_ERR(ai);
-       BUG_ON(ai->nr_groups != 1);
-       BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
 
        size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
-       chunk_size = ai->unit_size * num_possible_cpus();
+       areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
 
-       base = __alloc_bootmem_nopanic(chunk_size, PAGE_SIZE,
-                                      __pa(MAX_DMA_ADDRESS));
-       if (!base) {
-               pr_warning("PERCPU: failed to allocate %zu bytes for "
-                          "embedding\n", chunk_size);
+       areas = alloc_bootmem_nopanic(areas_size);
+       if (!areas) {
                rc = -ENOMEM;
-               goto out_free_ai;
+               goto out_free;
        }
 
-       /* return the leftover and copy */
-       for (unit = 0; unit < num_possible_cpus(); unit++) {
-               void *ptr = base + unit * ai->unit_size;
+       /* allocate, copy and determine base address */
+       for (group = 0; group < ai->nr_groups; group++) {
+               struct pcpu_group_info *gi = &ai->groups[group];
+               unsigned int cpu = NR_CPUS;
+               void *ptr;
+
+               for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
+                       cpu = gi->cpu_map[i];
+               BUG_ON(cpu == NR_CPUS);
+
+               /* allocate space for the whole group */
+               ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
+               if (!ptr) {
+                       rc = -ENOMEM;
+                       goto out_free_areas;
+               }
+               /* kmemleak tracks the percpu allocations separately */
+               kmemleak_free(ptr);
+               areas[group] = ptr;
 
-               free_bootmem(__pa(ptr + size_sum), ai->unit_size - size_sum);
-               memcpy(ptr, __per_cpu_load, ai->static_size);
+               base = min(ptr, base);
+       }
+
+       /*
+        * Copy data and free unused parts.  This should happen after all
+        * allocations are complete; otherwise, we may end up with
+        * overlapping groups.
+        */
+       for (group = 0; group < ai->nr_groups; group++) {
+               struct pcpu_group_info *gi = &ai->groups[group];
+               void *ptr = areas[group];
+
+               for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
+                       if (gi->cpu_map[i] == NR_CPUS) {
+                               /* unused unit, free whole */
+                               free_fn(ptr, ai->unit_size);
+                               continue;
+                       }
+                       /* copy and return the unused part */
+                       memcpy(ptr, __per_cpu_load, ai->static_size);
+                       free_fn(ptr + size_sum, ai->unit_size - size_sum);
+               }
+       }
+
+       /* base address is now known, determine group base offsets */
+       max_distance = 0;
+       for (group = 0; group < ai->nr_groups; group++) {
+               ai->groups[group].base_offset = areas[group] - base;
+               max_distance = max_t(size_t, max_distance,
+                                    ai->groups[group].base_offset);
+       }
+       max_distance += ai->unit_size;
+
+       /* warn if maximum distance is further than 75% of vmalloc space */
+       if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) {
+               pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
+                          "space 0x%lx\n", max_distance,
+                          (unsigned long)(VMALLOC_END - VMALLOC_START));
+#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
+               /* and fail if we have fallback */
+               rc = -EINVAL;
+               goto out_free;
+#endif
        }
 
-       /* we're ready, commit */
        pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
                PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
                ai->dyn_size, ai->unit_size);
 
        rc = pcpu_setup_first_chunk(ai, base);
-out_free_ai:
+       goto out_free;
+
+out_free_areas:
+       for (group = 0; group < ai->nr_groups; group++)
+               free_fn(areas[group],
+                       ai->groups[group].nr_units * ai->unit_size);
+out_free:
        pcpu_free_alloc_info(ai);
+       if (areas)
+               free_bootmem(__pa(areas), areas_size);
        return rc;
 }
-#endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK ||
-         !CONFIG_HAVE_SETUP_PER_CPU_AREA */
+#endif /* BUILD_EMBED_FIRST_CHUNK */
 
-#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
+#ifdef BUILD_PAGE_FIRST_CHUNK
 /**
  * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
  * @reserved_size: the size of reserved percpu area in bytes
  * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
- * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
+ * @free_fn: function to free percpu page, always called with PAGE_SIZE
  * @populate_pte_fn: function to populate pte
  *
  * This is a helper to ease setting up page-remapped first percpu
@@ -1847,7 +1748,7 @@ int __init pcpu_page_first_chunk(size_t reserved_size,
 
        snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
 
-       ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL);
+       ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
        if (IS_ERR(ai))
                return PTR_ERR(ai);
        BUG_ON(ai->nr_groups != 1);
@@ -1873,6 +1774,8 @@ int __init pcpu_page_first_chunk(size_t reserved_size,
                                           "for cpu%u\n", psize_str, cpu);
                                goto enomem;
                        }
+                       /* kmemleak tracks the percpu allocations separately */
+                       kmemleak_free(ptr);
                        pages[j++] = virt_to_page(ptr);
                }
 
@@ -1923,246 +1826,11 @@ out_free_ar:
        pcpu_free_alloc_info(ai);
        return rc;
 }
-#endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */
-
-#ifdef CONFIG_NEED_PER_CPU_LPAGE_FIRST_CHUNK
-struct pcpul_ent {
-       void            *ptr;
-       void            *map_addr;
-};
-
-static size_t pcpul_size;
-static size_t pcpul_lpage_size;
-static int pcpul_nr_lpages;
-static struct pcpul_ent *pcpul_map;
-
-static bool __init pcpul_unit_to_cpu(int unit, const struct pcpu_alloc_info *ai,
-                                    unsigned int *cpup)
-{
-       int group, cunit;
-
-       for (group = 0, cunit = 0; group < ai->nr_groups; group++) {
-               const struct pcpu_group_info *gi = &ai->groups[group];
-
-               if (unit < cunit + gi->nr_units) {
-                       if (cpup)
-                               *cpup = gi->cpu_map[unit - cunit];
-                       return true;
-               }
-               cunit += gi->nr_units;
-       }
-
-       return false;
-}
-
-static int __init pcpul_cpu_to_unit(int cpu, const struct pcpu_alloc_info *ai)
-{
-       int group, unit, i;
-
-       for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
-               const struct pcpu_group_info *gi = &ai->groups[group];
-
-               for (i = 0; i < gi->nr_units; i++)
-                       if (gi->cpu_map[i] == cpu)
-                               return unit + i;
-       }
-       BUG();
-}
-
-/**
- * pcpu_lpage_first_chunk - remap the first percpu chunk using large page
- * @ai: pcpu_alloc_info
- * @alloc_fn: function to allocate percpu lpage, always called with lpage_size
- * @free_fn: function to free percpu memory, @size <= lpage_size
- * @map_fn: function to map percpu lpage, always called with lpage_size
- *
- * This allocator uses large page to build and map the first chunk.
- * Unlike other helpers, the caller should provide fully initialized
- * @ai.  This can be done using pcpu_build_alloc_info().  This two
- * stage initialization is to allow arch code to evaluate the
- * parameters before committing to it.
- *
- * Large pages are allocated as directed by @unit_map and other
- * parameters and mapped to vmalloc space.  Unused holes are returned
- * to the page allocator.  Note that these holes end up being actively
- * mapped twice - once to the physical mapping and to the vmalloc area
- * for the first percpu chunk.  Depending on architecture, this might
- * cause problem when changing page attributes of the returned area.
- * These double mapped areas can be detected using
- * pcpu_lpage_remapped().
- *
- * RETURNS:
- * 0 on success, -errno on failure.
- */
-int __init pcpu_lpage_first_chunk(const struct pcpu_alloc_info *ai,
-                                 pcpu_fc_alloc_fn_t alloc_fn,
-                                 pcpu_fc_free_fn_t free_fn,
-                                 pcpu_fc_map_fn_t map_fn)
-{
-       static struct vm_struct vm;
-       const size_t lpage_size = ai->atom_size;
-       size_t chunk_size, map_size;
-       unsigned int cpu;
-       int i, j, unit, nr_units, rc;
-
-       nr_units = 0;
-       for (i = 0; i < ai->nr_groups; i++)
-               nr_units += ai->groups[i].nr_units;
-
-       chunk_size = ai->unit_size * nr_units;
-       BUG_ON(chunk_size % lpage_size);
-
-       pcpul_size = ai->static_size + ai->reserved_size + ai->dyn_size;
-       pcpul_lpage_size = lpage_size;
-       pcpul_nr_lpages = chunk_size / lpage_size;
-
-       /* allocate pointer array and alloc large pages */
-       map_size = pcpul_nr_lpages * sizeof(pcpul_map[0]);
-       pcpul_map = alloc_bootmem(map_size);
-
-       /* allocate all pages */
-       for (i = 0; i < pcpul_nr_lpages; i++) {
-               size_t offset = i * lpage_size;
-               int first_unit = offset / ai->unit_size;
-               int last_unit = (offset + lpage_size - 1) / ai->unit_size;
-               void *ptr;
-
-               /* find out which cpu is mapped to this unit */
-               for (unit = first_unit; unit <= last_unit; unit++)
-                       if (pcpul_unit_to_cpu(unit, ai, &cpu))
-                               goto found;
-               continue;
-       found:
-               ptr = alloc_fn(cpu, lpage_size, lpage_size);
-               if (!ptr) {
-                       pr_warning("PERCPU: failed to allocate large page "
-                                  "for cpu%u\n", cpu);
-                       goto enomem;
-               }
-
-               pcpul_map[i].ptr = ptr;
-       }
-
-       /* return unused holes */
-       for (unit = 0; unit < nr_units; unit++) {
-               size_t start = unit * ai->unit_size;
-               size_t end = start + ai->unit_size;
-               size_t off, next;
-
-               /* don't free used part of occupied unit */
-               if (pcpul_unit_to_cpu(unit, ai, NULL))
-                       start += pcpul_size;
-
-               /* unit can span more than one page, punch the holes */
-               for (off = start; off < end; off = next) {
-                       void *ptr = pcpul_map[off / lpage_size].ptr;
-                       next = min(roundup(off + 1, lpage_size), end);
-                       if (ptr)
-                               free_fn(ptr + off % lpage_size, next - off);
-               }
-       }
-
-       /* allocate address, map and copy */
-       vm.flags = VM_ALLOC;
-       vm.size = chunk_size;
-       vm_area_register_early(&vm, ai->unit_size);
-
-       for (i = 0; i < pcpul_nr_lpages; i++) {
-               if (!pcpul_map[i].ptr)
-                       continue;
-               pcpul_map[i].map_addr = vm.addr + i * lpage_size;
-               map_fn(pcpul_map[i].ptr, lpage_size, pcpul_map[i].map_addr);
-       }
-
-       for_each_possible_cpu(cpu)
-               memcpy(vm.addr + pcpul_cpu_to_unit(cpu, ai) * ai->unit_size,
-                      __per_cpu_load, ai->static_size);
-
-       /* we're ready, commit */
-       pr_info("PERCPU: large pages @%p s%zu r%zu d%zu u%zu\n",
-               vm.addr, ai->static_size, ai->reserved_size, ai->dyn_size,
-               ai->unit_size);
-
-       rc = pcpu_setup_first_chunk(ai, vm.addr);
-
-       /*
-        * Sort pcpul_map array for pcpu_lpage_remapped().  Unmapped
-        * lpages are pushed to the end and trimmed.
-        */
-       for (i = 0; i < pcpul_nr_lpages - 1; i++)
-               for (j = i + 1; j < pcpul_nr_lpages; j++) {
-                       struct pcpul_ent tmp;
-
-                       if (!pcpul_map[j].ptr)
-                               continue;
-                       if (pcpul_map[i].ptr &&
-                           pcpul_map[i].ptr < pcpul_map[j].ptr)
-                               continue;
-
-                       tmp = pcpul_map[i];
-                       pcpul_map[i] = pcpul_map[j];
-                       pcpul_map[j] = tmp;
-               }
-
-       while (pcpul_nr_lpages && !pcpul_map[pcpul_nr_lpages - 1].ptr)
-               pcpul_nr_lpages--;
-
-       return rc;
-
-enomem:
-       for (i = 0; i < pcpul_nr_lpages; i++)
-               if (pcpul_map[i].ptr)
-                       free_fn(pcpul_map[i].ptr, lpage_size);
-       free_bootmem(__pa(pcpul_map), map_size);
-       return -ENOMEM;
-}
-
-/**
- * pcpu_lpage_remapped - determine whether a kaddr is in pcpul recycled area
- * @kaddr: the kernel address in question
- *
- * Determine whether @kaddr falls in the pcpul recycled area.  This is
- * used by pageattr to detect VM aliases and break up the pcpu large
- * page mapping such that the same physical page is not mapped under
- * different attributes.
- *
- * The recycled area is always at the tail of a partially used large
- * page.
- *
- * RETURNS:
- * Address of corresponding remapped pcpu address if match is found;
- * otherwise, NULL.
- */
-void *pcpu_lpage_remapped(void *kaddr)
-{
-       unsigned long lpage_mask = pcpul_lpage_size - 1;
-       void *lpage_addr = (void *)((unsigned long)kaddr & ~lpage_mask);
-       unsigned long offset = (unsigned long)kaddr & lpage_mask;
-       int left = 0, right = pcpul_nr_lpages - 1;
-       int pos;
-
-       /* pcpul in use at all? */
-       if (!pcpul_map)
-               return NULL;
-
-       /* okay, perform binary search */
-       while (left <= right) {
-               pos = (left + right) / 2;
-
-               if (pcpul_map[pos].ptr < lpage_addr)
-                       left = pos + 1;
-               else if (pcpul_map[pos].ptr > lpage_addr)
-                       right = pos - 1;
-               else
-                       return pcpul_map[pos].map_addr + offset;
-       }
-
-       return NULL;
-}
-#endif /* CONFIG_NEED_PER_CPU_LPAGE_FIRST_CHUNK */
+#endif /* BUILD_PAGE_FIRST_CHUNK */
 
+#ifndef        CONFIG_HAVE_SETUP_PER_CPU_AREA
 /*
- * Generic percpu area setup.
+ * Generic SMP percpu area setup.
  *
  * The embedding helper is used because its behavior closely resembles
  * the original non-dynamic generic percpu area setup.  This is
@@ -2173,10 +1841,20 @@ void *pcpu_lpage_remapped(void *kaddr)
  * on the physical linear memory mapping which uses large page
  * mappings on applicable archs.
  */
-#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
 EXPORT_SYMBOL(__per_cpu_offset);
 
+static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
+                                      size_t align)
+{
+       return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
+}
+
+static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
+{
+       free_bootmem(__pa(ptr), size);
+}
+
 void __init setup_per_cpu_areas(void)
 {
        unsigned long delta;
@@ -2188,12 +1866,80 @@ void __init setup_per_cpu_areas(void)
         * what the legacy allocator did.
         */
        rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
-                                   PERCPU_DYNAMIC_RESERVE);
+                                   PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
+                                   pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
        if (rc < 0)
-               panic("Failed to initialized percpu areas.");
+               panic("Failed to initialize percpu areas.");
 
        delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
        for_each_possible_cpu(cpu)
                __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
 }
-#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
+#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
+
+#else  /* CONFIG_SMP */
+
+/*
+ * UP percpu area setup.
+ *
+ * UP always uses km-based percpu allocator with identity mapping.
+ * Static percpu variables are indistinguishable from the usual static
+ * variables and don't require any special preparation.
+ */
+void __init setup_per_cpu_areas(void)
+{
+       const size_t unit_size =
+               roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
+                                        PERCPU_DYNAMIC_RESERVE));
+       struct pcpu_alloc_info *ai;
+       void *fc;
+
+       ai = pcpu_alloc_alloc_info(1, 1);
+       fc = __alloc_bootmem(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
+       if (!ai || !fc)
+               panic("Failed to allocate memory for percpu areas.");
+       /* kmemleak tracks the percpu allocations separately */
+       kmemleak_free(fc);
+
+       ai->dyn_size = unit_size;
+       ai->unit_size = unit_size;
+       ai->atom_size = unit_size;
+       ai->alloc_size = unit_size;
+       ai->groups[0].nr_units = 1;
+       ai->groups[0].cpu_map[0] = 0;
+
+       if (pcpu_setup_first_chunk(ai, fc) < 0)
+               panic("Failed to initialize percpu areas.");
+}
+
+#endif /* CONFIG_SMP */
+
+/*
+ * First and reserved chunks are initialized with temporary allocation
+ * map in initdata so that they can be used before slab is online.
+ * This function is called after slab is brought up and replaces those
+ * with properly allocated maps.
+ */
+void __init percpu_init_late(void)
+{
+       struct pcpu_chunk *target_chunks[] =
+               { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
+       struct pcpu_chunk *chunk;
+       unsigned long flags;
+       int i;
+
+       for (i = 0; (chunk = target_chunks[i]); i++) {
+               int *map;
+               const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
+
+               BUILD_BUG_ON(size > PAGE_SIZE);
+
+               map = pcpu_mem_zalloc(size);
+               BUG_ON(!map);
+
+               spin_lock_irqsave(&pcpu_lock, flags);
+               memcpy(map, chunk->map, size);
+               chunk->map = map;
+               spin_unlock_irqrestore(&pcpu_lock, flags);
+       }
+}