aboutsummaryrefslogtreecommitdiffstats
path: root/mm
diff options
context:
space:
mode:
Diffstat (limited to 'mm')
-rw-r--r--mm/Makefile4
-rw-r--r--mm/allocpercpu.c32
-rw-r--r--mm/bootmem.c35
-rw-r--r--mm/filemap.c7
-rw-r--r--mm/memory.c10
-rw-r--r--mm/page_alloc.c5
-rw-r--r--mm/percpu.c1226
-rw-r--r--mm/slab.c75
-rw-r--r--mm/slob.c39
-rw-r--r--mm/slub.c98
-rw-r--r--mm/vmalloc.c97
-rw-r--r--mm/vmscan.c2
12 files changed, 1576 insertions, 54 deletions
diff --git a/mm/Makefile b/mm/Makefile
index 72255be57f89..818569b68f46 100644
--- a/mm/Makefile
+++ b/mm/Makefile
@@ -30,6 +30,10 @@ obj-$(CONFIG_FAILSLAB) += failslab.o
30obj-$(CONFIG_MEMORY_HOTPLUG) += memory_hotplug.o 30obj-$(CONFIG_MEMORY_HOTPLUG) += memory_hotplug.o
31obj-$(CONFIG_FS_XIP) += filemap_xip.o 31obj-$(CONFIG_FS_XIP) += filemap_xip.o
32obj-$(CONFIG_MIGRATION) += migrate.o 32obj-$(CONFIG_MIGRATION) += migrate.o
33ifdef CONFIG_HAVE_DYNAMIC_PER_CPU_AREA
34obj-$(CONFIG_SMP) += percpu.o
35else
33obj-$(CONFIG_SMP) += allocpercpu.o 36obj-$(CONFIG_SMP) += allocpercpu.o
37endif
34obj-$(CONFIG_QUICKLIST) += quicklist.o 38obj-$(CONFIG_QUICKLIST) += quicklist.o
35obj-$(CONFIG_CGROUP_MEM_RES_CTLR) += memcontrol.o page_cgroup.o 39obj-$(CONFIG_CGROUP_MEM_RES_CTLR) += memcontrol.o page_cgroup.o
diff --git a/mm/allocpercpu.c b/mm/allocpercpu.c
index 4297bc41bfd2..3653c570232b 100644
--- a/mm/allocpercpu.c
+++ b/mm/allocpercpu.c
@@ -99,45 +99,51 @@ static int __percpu_populate_mask(void *__pdata, size_t size, gfp_t gfp,
99 __percpu_populate_mask((__pdata), (size), (gfp), &(mask)) 99 __percpu_populate_mask((__pdata), (size), (gfp), &(mask))
100 100
101/** 101/**
102 * percpu_alloc_mask - initial setup of per-cpu data 102 * alloc_percpu - initial setup of per-cpu data
103 * @size: size of per-cpu object 103 * @size: size of per-cpu object
104 * @gfp: may sleep or not etc. 104 * @align: alignment
105 * @mask: populate per-data for cpu's selected through mask bits
106 * 105 *
107 * Populating per-cpu data for all online cpu's would be a typical use case, 106 * Allocate dynamic percpu area. Percpu objects are populated with
108 * which is simplified by the percpu_alloc() wrapper. 107 * zeroed buffers.
109 * Per-cpu objects are populated with zeroed buffers.
110 */ 108 */
111void *__percpu_alloc_mask(size_t size, gfp_t gfp, cpumask_t *mask) 109void *__alloc_percpu(size_t size, size_t align)
112{ 110{
113 /* 111 /*
114 * We allocate whole cache lines to avoid false sharing 112 * We allocate whole cache lines to avoid false sharing
115 */ 113 */
116 size_t sz = roundup(nr_cpu_ids * sizeof(void *), cache_line_size()); 114 size_t sz = roundup(nr_cpu_ids * sizeof(void *), cache_line_size());
117 void *pdata = kzalloc(sz, gfp); 115 void *pdata = kzalloc(sz, GFP_KERNEL);
118 void *__pdata = __percpu_disguise(pdata); 116 void *__pdata = __percpu_disguise(pdata);
119 117
118 /*
119 * Can't easily make larger alignment work with kmalloc. WARN
120 * on it. Larger alignment should only be used for module
121 * percpu sections on SMP for which this path isn't used.
122 */
123 WARN_ON_ONCE(align > __alignof__(unsigned long long));
124
120 if (unlikely(!pdata)) 125 if (unlikely(!pdata))
121 return NULL; 126 return NULL;
122 if (likely(!__percpu_populate_mask(__pdata, size, gfp, mask))) 127 if (likely(!__percpu_populate_mask(__pdata, size, GFP_KERNEL,
128 &cpu_possible_map)))
123 return __pdata; 129 return __pdata;
124 kfree(pdata); 130 kfree(pdata);
125 return NULL; 131 return NULL;
126} 132}
127EXPORT_SYMBOL_GPL(__percpu_alloc_mask); 133EXPORT_SYMBOL_GPL(__alloc_percpu);
128 134
129/** 135/**
130 * percpu_free - final cleanup of per-cpu data 136 * free_percpu - final cleanup of per-cpu data
131 * @__pdata: object to clean up 137 * @__pdata: object to clean up
132 * 138 *
133 * We simply clean up any per-cpu object left. No need for the client to 139 * We simply clean up any per-cpu object left. No need for the client to
134 * track and specify through a bis mask which per-cpu objects are to free. 140 * track and specify through a bis mask which per-cpu objects are to free.
135 */ 141 */
136void percpu_free(void *__pdata) 142void free_percpu(void *__pdata)
137{ 143{
138 if (unlikely(!__pdata)) 144 if (unlikely(!__pdata))
139 return; 145 return;
140 __percpu_depopulate_mask(__pdata, &cpu_possible_map); 146 __percpu_depopulate_mask(__pdata, &cpu_possible_map);
141 kfree(__percpu_disguise(__pdata)); 147 kfree(__percpu_disguise(__pdata));
142} 148}
143EXPORT_SYMBOL_GPL(percpu_free); 149EXPORT_SYMBOL_GPL(free_percpu);
diff --git a/mm/bootmem.c b/mm/bootmem.c
index 51a0ccf61e0e..daf92713f7de 100644
--- a/mm/bootmem.c
+++ b/mm/bootmem.c
@@ -382,7 +382,6 @@ int __init reserve_bootmem_node(pg_data_t *pgdat, unsigned long physaddr,
382 return mark_bootmem_node(pgdat->bdata, start, end, 1, flags); 382 return mark_bootmem_node(pgdat->bdata, start, end, 1, flags);
383} 383}
384 384
385#ifndef CONFIG_HAVE_ARCH_BOOTMEM_NODE
386/** 385/**
387 * reserve_bootmem - mark a page range as usable 386 * reserve_bootmem - mark a page range as usable
388 * @addr: starting address of the range 387 * @addr: starting address of the range
@@ -403,7 +402,6 @@ int __init reserve_bootmem(unsigned long addr, unsigned long size,
403 402
404 return mark_bootmem(start, end, 1, flags); 403 return mark_bootmem(start, end, 1, flags);
405} 404}
406#endif /* !CONFIG_HAVE_ARCH_BOOTMEM_NODE */
407 405
408static unsigned long align_idx(struct bootmem_data *bdata, unsigned long idx, 406static unsigned long align_idx(struct bootmem_data *bdata, unsigned long idx,
409 unsigned long step) 407 unsigned long step)
@@ -429,8 +427,8 @@ static unsigned long align_off(struct bootmem_data *bdata, unsigned long off,
429} 427}
430 428
431static void * __init alloc_bootmem_core(struct bootmem_data *bdata, 429static void * __init alloc_bootmem_core(struct bootmem_data *bdata,
432 unsigned long size, unsigned long align, 430 unsigned long size, unsigned long align,
433 unsigned long goal, unsigned long limit) 431 unsigned long goal, unsigned long limit)
434{ 432{
435 unsigned long fallback = 0; 433 unsigned long fallback = 0;
436 unsigned long min, max, start, sidx, midx, step; 434 unsigned long min, max, start, sidx, midx, step;
@@ -530,17 +528,34 @@ find_block:
530 return NULL; 528 return NULL;
531} 529}
532 530
531static void * __init alloc_arch_preferred_bootmem(bootmem_data_t *bdata,
532 unsigned long size, unsigned long align,
533 unsigned long goal, unsigned long limit)
534{
535#ifdef CONFIG_HAVE_ARCH_BOOTMEM
536 bootmem_data_t *p_bdata;
537
538 p_bdata = bootmem_arch_preferred_node(bdata, size, align, goal, limit);
539 if (p_bdata)
540 return alloc_bootmem_core(p_bdata, size, align, goal, limit);
541#endif
542 return NULL;
543}
544
533static void * __init ___alloc_bootmem_nopanic(unsigned long size, 545static void * __init ___alloc_bootmem_nopanic(unsigned long size,
534 unsigned long align, 546 unsigned long align,
535 unsigned long goal, 547 unsigned long goal,
536 unsigned long limit) 548 unsigned long limit)
537{ 549{
538 bootmem_data_t *bdata; 550 bootmem_data_t *bdata;
551 void *region;
539 552
540restart: 553restart:
541 list_for_each_entry(bdata, &bdata_list, list) { 554 region = alloc_arch_preferred_bootmem(NULL, size, align, goal, limit);
542 void *region; 555 if (region)
556 return region;
543 557
558 list_for_each_entry(bdata, &bdata_list, list) {
544 if (goal && bdata->node_low_pfn <= PFN_DOWN(goal)) 559 if (goal && bdata->node_low_pfn <= PFN_DOWN(goal))
545 continue; 560 continue;
546 if (limit && bdata->node_min_pfn >= PFN_DOWN(limit)) 561 if (limit && bdata->node_min_pfn >= PFN_DOWN(limit))
@@ -618,6 +633,10 @@ static void * __init ___alloc_bootmem_node(bootmem_data_t *bdata,
618{ 633{
619 void *ptr; 634 void *ptr;
620 635
636 ptr = alloc_arch_preferred_bootmem(bdata, size, align, goal, limit);
637 if (ptr)
638 return ptr;
639
621 ptr = alloc_bootmem_core(bdata, size, align, goal, limit); 640 ptr = alloc_bootmem_core(bdata, size, align, goal, limit);
622 if (ptr) 641 if (ptr)
623 return ptr; 642 return ptr;
@@ -674,6 +693,10 @@ void * __init __alloc_bootmem_node_nopanic(pg_data_t *pgdat, unsigned long size,
674{ 693{
675 void *ptr; 694 void *ptr;
676 695
696 ptr = alloc_arch_preferred_bootmem(pgdat->bdata, size, align, goal, 0);
697 if (ptr)
698 return ptr;
699
677 ptr = alloc_bootmem_core(pgdat->bdata, size, align, goal, 0); 700 ptr = alloc_bootmem_core(pgdat->bdata, size, align, goal, 0);
678 if (ptr) 701 if (ptr)
679 return ptr; 702 return ptr;
diff --git a/mm/filemap.c b/mm/filemap.c
index 23acefe51808..126d3973b3d1 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1823,7 +1823,7 @@ static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1823 int copy = min(bytes, iov->iov_len - base); 1823 int copy = min(bytes, iov->iov_len - base);
1824 1824
1825 base = 0; 1825 base = 0;
1826 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy); 1826 left = __copy_from_user_inatomic(vaddr, buf, copy);
1827 copied += copy; 1827 copied += copy;
1828 bytes -= copy; 1828 bytes -= copy;
1829 vaddr += copy; 1829 vaddr += copy;
@@ -1851,8 +1851,7 @@ size_t iov_iter_copy_from_user_atomic(struct page *page,
1851 if (likely(i->nr_segs == 1)) { 1851 if (likely(i->nr_segs == 1)) {
1852 int left; 1852 int left;
1853 char __user *buf = i->iov->iov_base + i->iov_offset; 1853 char __user *buf = i->iov->iov_base + i->iov_offset;
1854 left = __copy_from_user_inatomic_nocache(kaddr + offset, 1854 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1855 buf, bytes);
1856 copied = bytes - left; 1855 copied = bytes - left;
1857 } else { 1856 } else {
1858 copied = __iovec_copy_from_user_inatomic(kaddr + offset, 1857 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
@@ -1880,7 +1879,7 @@ size_t iov_iter_copy_from_user(struct page *page,
1880 if (likely(i->nr_segs == 1)) { 1879 if (likely(i->nr_segs == 1)) {
1881 int left; 1880 int left;
1882 char __user *buf = i->iov->iov_base + i->iov_offset; 1881 char __user *buf = i->iov->iov_base + i->iov_offset;
1883 left = __copy_from_user_nocache(kaddr + offset, buf, bytes); 1882 left = __copy_from_user(kaddr + offset, buf, bytes);
1884 copied = bytes - left; 1883 copied = bytes - left;
1885 } else { 1884 } else {
1886 copied = __iovec_copy_from_user_inatomic(kaddr + offset, 1885 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
diff --git a/mm/memory.c b/mm/memory.c
index baa999e87cd2..05fab3bc5b4b 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -48,6 +48,8 @@
48#include <linux/rmap.h> 48#include <linux/rmap.h>
49#include <linux/module.h> 49#include <linux/module.h>
50#include <linux/delayacct.h> 50#include <linux/delayacct.h>
51#include <linux/kprobes.h>
52#include <linux/mutex.h>
51#include <linux/init.h> 53#include <linux/init.h>
52#include <linux/writeback.h> 54#include <linux/writeback.h>
53#include <linux/memcontrol.h> 55#include <linux/memcontrol.h>
@@ -99,6 +101,14 @@ int randomize_va_space __read_mostly =
99 2; 101 2;
100#endif 102#endif
101 103
104/*
105 * mutex protecting text section modification (dynamic code patching).
106 * some users need to sleep (allocating memory...) while they hold this lock.
107 *
108 * NOT exported to modules - patching kernel text is a really delicate matter.
109 */
110DEFINE_MUTEX(text_mutex);
111
102static int __init disable_randmaps(char *s) 112static int __init disable_randmaps(char *s)
103{ 113{
104 randomize_va_space = 0; 114 randomize_va_space = 0;
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 5c44ed49ca93..a3803ea8c27d 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1479,6 +1479,8 @@ __alloc_pages_internal(gfp_t gfp_mask, unsigned int order,
1479 unsigned long did_some_progress; 1479 unsigned long did_some_progress;
1480 unsigned long pages_reclaimed = 0; 1480 unsigned long pages_reclaimed = 0;
1481 1481
1482 lockdep_trace_alloc(gfp_mask);
1483
1482 might_sleep_if(wait); 1484 might_sleep_if(wait);
1483 1485
1484 if (should_fail_alloc_page(gfp_mask, order)) 1486 if (should_fail_alloc_page(gfp_mask, order))
@@ -1578,12 +1580,15 @@ nofail_alloc:
1578 */ 1580 */
1579 cpuset_update_task_memory_state(); 1581 cpuset_update_task_memory_state();
1580 p->flags |= PF_MEMALLOC; 1582 p->flags |= PF_MEMALLOC;
1583
1584 lockdep_set_current_reclaim_state(gfp_mask);
1581 reclaim_state.reclaimed_slab = 0; 1585 reclaim_state.reclaimed_slab = 0;
1582 p->reclaim_state = &reclaim_state; 1586 p->reclaim_state = &reclaim_state;
1583 1587
1584 did_some_progress = try_to_free_pages(zonelist, order, gfp_mask); 1588 did_some_progress = try_to_free_pages(zonelist, order, gfp_mask);
1585 1589
1586 p->reclaim_state = NULL; 1590 p->reclaim_state = NULL;
1591 lockdep_clear_current_reclaim_state();
1587 p->flags &= ~PF_MEMALLOC; 1592 p->flags &= ~PF_MEMALLOC;
1588 1593
1589 cond_resched(); 1594 cond_resched();
diff --git a/mm/percpu.c b/mm/percpu.c
new file mode 100644
index 000000000000..bfe6a3afaf45
--- /dev/null
+++ b/mm/percpu.c
@@ -0,0 +1,1226 @@
1/*
2 * linux/mm/percpu.c - percpu memory allocator
3 *
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *
9 * This is percpu allocator which can handle both static and dynamic
10 * areas. Percpu areas are allocated in chunks in vmalloc area. Each
11 * chunk is consisted of num_possible_cpus() units and the first chunk
12 * is used for static percpu variables in the kernel image (special
13 * boot time alloc/init handling necessary as these areas need to be
14 * brought up before allocation services are running). Unit grows as
15 * necessary and all units grow or shrink in unison. When a chunk is
16 * filled up, another chunk is allocated. ie. in vmalloc area
17 *
18 * c0 c1 c2
19 * ------------------- ------------------- ------------
20 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
21 * ------------------- ...... ------------------- .... ------------
22 *
23 * Allocation is done in offset-size areas of single unit space. Ie,
24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
25 * c1:u1, c1:u2 and c1:u3. Percpu access can be done by configuring
26 * percpu base registers UNIT_SIZE apart.
27 *
28 * There are usually many small percpu allocations many of them as
29 * small as 4 bytes. The allocator organizes chunks into lists
30 * according to free size and tries to allocate from the fullest one.
31 * Each chunk keeps the maximum contiguous area size hint which is
32 * guaranteed to be eqaul to or larger than the maximum contiguous
33 * area in the chunk. This helps the allocator not to iterate the
34 * chunk maps unnecessarily.
35 *
36 * Allocation state in each chunk is kept using an array of integers
37 * on chunk->map. A positive value in the map represents a free
38 * region and negative allocated. Allocation inside a chunk is done
39 * by scanning this map sequentially and serving the first matching
40 * entry. This is mostly copied from the percpu_modalloc() allocator.
41 * Chunks are also linked into a rb tree to ease address to chunk
42 * mapping during free.
43 *
44 * To use this allocator, arch code should do the followings.
45 *
46 * - define CONFIG_HAVE_DYNAMIC_PER_CPU_AREA
47 *
48 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
49 * regular address to percpu pointer and back
50 *
51 * - use pcpu_setup_first_chunk() during percpu area initialization to
52 * setup the first chunk containing the kernel static percpu area
53 */
54
55#include <linux/bitmap.h>
56#include <linux/bootmem.h>
57#include <linux/list.h>
58#include <linux/mm.h>
59#include <linux/module.h>
60#include <linux/mutex.h>
61#include <linux/percpu.h>
62#include <linux/pfn.h>
63#include <linux/rbtree.h>
64#include <linux/slab.h>
65#include <linux/spinlock.h>
66#include <linux/vmalloc.h>
67#include <linux/workqueue.h>
68
69#include <asm/cacheflush.h>
70#include <asm/tlbflush.h>
71
72#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
73#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
74
75struct pcpu_chunk {
76 struct list_head list; /* linked to pcpu_slot lists */
77 struct rb_node rb_node; /* key is chunk->vm->addr */
78 int free_size; /* free bytes in the chunk */
79 int contig_hint; /* max contiguous size hint */
80 struct vm_struct *vm; /* mapped vmalloc region */
81 int map_used; /* # of map entries used */
82 int map_alloc; /* # of map entries allocated */
83 int *map; /* allocation map */
84 bool immutable; /* no [de]population allowed */
85 struct page **page; /* points to page array */
86 struct page *page_ar[]; /* #cpus * UNIT_PAGES */
87};
88
89static int pcpu_unit_pages __read_mostly;
90static int pcpu_unit_size __read_mostly;
91static int pcpu_chunk_size __read_mostly;
92static int pcpu_nr_slots __read_mostly;
93static size_t pcpu_chunk_struct_size __read_mostly;
94
95/* the address of the first chunk which starts with the kernel static area */
96void *pcpu_base_addr __read_mostly;
97EXPORT_SYMBOL_GPL(pcpu_base_addr);
98
99/* optional reserved chunk, only accessible for reserved allocations */
100static struct pcpu_chunk *pcpu_reserved_chunk;
101/* offset limit of the reserved chunk */
102static int pcpu_reserved_chunk_limit;
103
104/*
105 * Synchronization rules.
106 *
107 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
108 * protects allocation/reclaim paths, chunks and chunk->page arrays.
109 * The latter is a spinlock and protects the index data structures -
110 * chunk slots, rbtree, chunks and area maps in chunks.
111 *
112 * During allocation, pcpu_alloc_mutex is kept locked all the time and
113 * pcpu_lock is grabbed and released as necessary. All actual memory
114 * allocations are done using GFP_KERNEL with pcpu_lock released.
115 *
116 * Free path accesses and alters only the index data structures, so it
117 * can be safely called from atomic context. When memory needs to be
118 * returned to the system, free path schedules reclaim_work which
119 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
120 * reclaimed, release both locks and frees the chunks. Note that it's
121 * necessary to grab both locks to remove a chunk from circulation as
122 * allocation path might be referencing the chunk with only
123 * pcpu_alloc_mutex locked.
124 */
125static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
126static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
127
128static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
129static struct rb_root pcpu_addr_root = RB_ROOT; /* chunks by address */
130
131/* reclaim work to release fully free chunks, scheduled from free path */
132static void pcpu_reclaim(struct work_struct *work);
133static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
134
135static int __pcpu_size_to_slot(int size)
136{
137 int highbit = fls(size); /* size is in bytes */
138 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
139}
140
141static int pcpu_size_to_slot(int size)
142{
143 if (size == pcpu_unit_size)
144 return pcpu_nr_slots - 1;
145 return __pcpu_size_to_slot(size);
146}
147
148static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
149{
150 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
151 return 0;
152
153 return pcpu_size_to_slot(chunk->free_size);
154}
155
156static int pcpu_page_idx(unsigned int cpu, int page_idx)
157{
158 return cpu * pcpu_unit_pages + page_idx;
159}
160
161static struct page **pcpu_chunk_pagep(struct pcpu_chunk *chunk,
162 unsigned int cpu, int page_idx)
163{
164 return &chunk->page[pcpu_page_idx(cpu, page_idx)];
165}
166
167static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
168 unsigned int cpu, int page_idx)
169{
170 return (unsigned long)chunk->vm->addr +
171 (pcpu_page_idx(cpu, page_idx) << PAGE_SHIFT);
172}
173
174static bool pcpu_chunk_page_occupied(struct pcpu_chunk *chunk,
175 int page_idx)
176{
177 return *pcpu_chunk_pagep(chunk, 0, page_idx) != NULL;
178}
179
180/**
181 * pcpu_mem_alloc - allocate memory
182 * @size: bytes to allocate
183 *
184 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
185 * kzalloc() is used; otherwise, vmalloc() is used. The returned
186 * memory is always zeroed.
187 *
188 * CONTEXT:
189 * Does GFP_KERNEL allocation.
190 *
191 * RETURNS:
192 * Pointer to the allocated area on success, NULL on failure.
193 */
194static void *pcpu_mem_alloc(size_t size)
195{
196 if (size <= PAGE_SIZE)
197 return kzalloc(size, GFP_KERNEL);
198 else {
199 void *ptr = vmalloc(size);
200 if (ptr)
201 memset(ptr, 0, size);
202 return ptr;
203 }
204}
205
206/**
207 * pcpu_mem_free - free memory
208 * @ptr: memory to free
209 * @size: size of the area
210 *
211 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
212 */
213static void pcpu_mem_free(void *ptr, size_t size)
214{
215 if (size <= PAGE_SIZE)
216 kfree(ptr);
217 else
218 vfree(ptr);
219}
220
221/**
222 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
223 * @chunk: chunk of interest
224 * @oslot: the previous slot it was on
225 *
226 * This function is called after an allocation or free changed @chunk.
227 * New slot according to the changed state is determined and @chunk is
228 * moved to the slot. Note that the reserved chunk is never put on
229 * chunk slots.
230 *
231 * CONTEXT:
232 * pcpu_lock.
233 */
234static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
235{
236 int nslot = pcpu_chunk_slot(chunk);
237
238 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
239 if (oslot < nslot)
240 list_move(&chunk->list, &pcpu_slot[nslot]);
241 else
242 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
243 }
244}
245
246static struct rb_node **pcpu_chunk_rb_search(void *addr,
247 struct rb_node **parentp)
248{
249 struct rb_node **p = &pcpu_addr_root.rb_node;
250 struct rb_node *parent = NULL;
251 struct pcpu_chunk *chunk;
252
253 while (*p) {
254 parent = *p;
255 chunk = rb_entry(parent, struct pcpu_chunk, rb_node);
256
257 if (addr < chunk->vm->addr)
258 p = &(*p)->rb_left;
259 else if (addr > chunk->vm->addr)
260 p = &(*p)->rb_right;
261 else
262 break;
263 }
264
265 if (parentp)
266 *parentp = parent;
267 return p;
268}
269
270/**
271 * pcpu_chunk_addr_search - search for chunk containing specified address
272 * @addr: address to search for
273 *
274 * Look for chunk which might contain @addr. More specifically, it
275 * searchs for the chunk with the highest start address which isn't
276 * beyond @addr.
277 *
278 * CONTEXT:
279 * pcpu_lock.
280 *
281 * RETURNS:
282 * The address of the found chunk.
283 */
284static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
285{
286 struct rb_node *n, *parent;
287 struct pcpu_chunk *chunk;
288
289 /* is it in the reserved chunk? */
290 if (pcpu_reserved_chunk) {
291 void *start = pcpu_reserved_chunk->vm->addr;
292
293 if (addr >= start && addr < start + pcpu_reserved_chunk_limit)
294 return pcpu_reserved_chunk;
295 }
296
297 /* nah... search the regular ones */
298 n = *pcpu_chunk_rb_search(addr, &parent);
299 if (!n) {
300 /* no exactly matching chunk, the parent is the closest */
301 n = parent;
302 BUG_ON(!n);
303 }
304 chunk = rb_entry(n, struct pcpu_chunk, rb_node);
305
306 if (addr < chunk->vm->addr) {
307 /* the parent was the next one, look for the previous one */
308 n = rb_prev(n);
309 BUG_ON(!n);
310 chunk = rb_entry(n, struct pcpu_chunk, rb_node);
311 }
312
313 return chunk;
314}
315
316/**
317 * pcpu_chunk_addr_insert - insert chunk into address rb tree
318 * @new: chunk to insert
319 *
320 * Insert @new into address rb tree.
321 *
322 * CONTEXT:
323 * pcpu_lock.
324 */
325static void pcpu_chunk_addr_insert(struct pcpu_chunk *new)
326{
327 struct rb_node **p, *parent;
328
329 p = pcpu_chunk_rb_search(new->vm->addr, &parent);
330 BUG_ON(*p);
331 rb_link_node(&new->rb_node, parent, p);
332 rb_insert_color(&new->rb_node, &pcpu_addr_root);
333}
334
335/**
336 * pcpu_extend_area_map - extend area map for allocation
337 * @chunk: target chunk
338 *
339 * Extend area map of @chunk so that it can accomodate an allocation.
340 * A single allocation can split an area into three areas, so this
341 * function makes sure that @chunk->map has at least two extra slots.
342 *
343 * CONTEXT:
344 * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired
345 * if area map is extended.
346 *
347 * RETURNS:
348 * 0 if noop, 1 if successfully extended, -errno on failure.
349 */
350static int pcpu_extend_area_map(struct pcpu_chunk *chunk)
351{
352 int new_alloc;
353 int *new;
354 size_t size;
355
356 /* has enough? */
357 if (chunk->map_alloc >= chunk->map_used + 2)
358 return 0;
359
360 spin_unlock_irq(&pcpu_lock);
361
362 new_alloc = PCPU_DFL_MAP_ALLOC;
363 while (new_alloc < chunk->map_used + 2)
364 new_alloc *= 2;
365
366 new = pcpu_mem_alloc(new_alloc * sizeof(new[0]));
367 if (!new) {
368 spin_lock_irq(&pcpu_lock);
369 return -ENOMEM;
370 }
371
372 /*
373 * Acquire pcpu_lock and switch to new area map. Only free
374 * could have happened inbetween, so map_used couldn't have
375 * grown.
376 */
377 spin_lock_irq(&pcpu_lock);
378 BUG_ON(new_alloc < chunk->map_used + 2);
379
380 size = chunk->map_alloc * sizeof(chunk->map[0]);
381 memcpy(new, chunk->map, size);
382
383 /*
384 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
385 * one of the first chunks and still using static map.
386 */
387 if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC)
388 pcpu_mem_free(chunk->map, size);
389
390 chunk->map_alloc = new_alloc;
391 chunk->map = new;
392 return 0;
393}
394
395/**
396 * pcpu_split_block - split a map block
397 * @chunk: chunk of interest
398 * @i: index of map block to split
399 * @head: head size in bytes (can be 0)
400 * @tail: tail size in bytes (can be 0)
401 *
402 * Split the @i'th map block into two or three blocks. If @head is
403 * non-zero, @head bytes block is inserted before block @i moving it
404 * to @i+1 and reducing its size by @head bytes.
405 *
406 * If @tail is non-zero, the target block, which can be @i or @i+1
407 * depending on @head, is reduced by @tail bytes and @tail byte block
408 * is inserted after the target block.
409 *
410 * @chunk->map must have enough free slots to accomodate the split.
411 *
412 * CONTEXT:
413 * pcpu_lock.
414 */
415static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
416 int head, int tail)
417{
418 int nr_extra = !!head + !!tail;
419
420 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
421
422 /* insert new subblocks */
423 memmove(&chunk->map[i + nr_extra], &chunk->map[i],
424 sizeof(chunk->map[0]) * (chunk->map_used - i));
425 chunk->map_used += nr_extra;
426
427 if (head) {
428 chunk->map[i + 1] = chunk->map[i] - head;
429 chunk->map[i++] = head;
430 }
431 if (tail) {
432 chunk->map[i++] -= tail;
433 chunk->map[i] = tail;
434 }
435}
436
437/**
438 * pcpu_alloc_area - allocate area from a pcpu_chunk
439 * @chunk: chunk of interest
440 * @size: wanted size in bytes
441 * @align: wanted align
442 *
443 * Try to allocate @size bytes area aligned at @align from @chunk.
444 * Note that this function only allocates the offset. It doesn't
445 * populate or map the area.
446 *
447 * @chunk->map must have at least two free slots.
448 *
449 * CONTEXT:
450 * pcpu_lock.
451 *
452 * RETURNS:
453 * Allocated offset in @chunk on success, -1 if no matching area is
454 * found.
455 */
456static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
457{
458 int oslot = pcpu_chunk_slot(chunk);
459 int max_contig = 0;
460 int i, off;
461
462 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
463 bool is_last = i + 1 == chunk->map_used;
464 int head, tail;
465
466 /* extra for alignment requirement */
467 head = ALIGN(off, align) - off;
468 BUG_ON(i == 0 && head != 0);
469
470 if (chunk->map[i] < 0)
471 continue;
472 if (chunk->map[i] < head + size) {
473 max_contig = max(chunk->map[i], max_contig);
474 continue;
475 }
476
477 /*
478 * If head is small or the previous block is free,
479 * merge'em. Note that 'small' is defined as smaller
480 * than sizeof(int), which is very small but isn't too
481 * uncommon for percpu allocations.
482 */
483 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
484 if (chunk->map[i - 1] > 0)
485 chunk->map[i - 1] += head;
486 else {
487 chunk->map[i - 1] -= head;
488 chunk->free_size -= head;
489 }
490 chunk->map[i] -= head;
491 off += head;
492 head = 0;
493 }
494
495 /* if tail is small, just keep it around */
496 tail = chunk->map[i] - head - size;
497 if (tail < sizeof(int))
498 tail = 0;
499
500 /* split if warranted */
501 if (head || tail) {
502 pcpu_split_block(chunk, i, head, tail);
503 if (head) {
504 i++;
505 off += head;
506 max_contig = max(chunk->map[i - 1], max_contig);
507 }
508 if (tail)
509 max_contig = max(chunk->map[i + 1], max_contig);
510 }
511
512 /* update hint and mark allocated */
513 if (is_last)
514 chunk->contig_hint = max_contig; /* fully scanned */
515 else
516 chunk->contig_hint = max(chunk->contig_hint,
517 max_contig);
518
519 chunk->free_size -= chunk->map[i];
520 chunk->map[i] = -chunk->map[i];
521
522 pcpu_chunk_relocate(chunk, oslot);
523 return off;
524 }
525
526 chunk->contig_hint = max_contig; /* fully scanned */
527 pcpu_chunk_relocate(chunk, oslot);
528
529 /* tell the upper layer that this chunk has no matching area */
530 return -1;
531}
532
533/**
534 * pcpu_free_area - free area to a pcpu_chunk
535 * @chunk: chunk of interest
536 * @freeme: offset of area to free
537 *
538 * Free area starting from @freeme to @chunk. Note that this function
539 * only modifies the allocation map. It doesn't depopulate or unmap
540 * the area.
541 *
542 * CONTEXT:
543 * pcpu_lock.
544 */
545static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
546{
547 int oslot = pcpu_chunk_slot(chunk);
548 int i, off;
549
550 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
551 if (off == freeme)
552 break;
553 BUG_ON(off != freeme);
554 BUG_ON(chunk->map[i] > 0);
555
556 chunk->map[i] = -chunk->map[i];
557 chunk->free_size += chunk->map[i];
558
559 /* merge with previous? */
560 if (i > 0 && chunk->map[i - 1] >= 0) {
561 chunk->map[i - 1] += chunk->map[i];
562 chunk->map_used--;
563 memmove(&chunk->map[i], &chunk->map[i + 1],
564 (chunk->map_used - i) * sizeof(chunk->map[0]));
565 i--;
566 }
567 /* merge with next? */
568 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
569 chunk->map[i] += chunk->map[i + 1];
570 chunk->map_used--;
571 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
572 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
573 }
574
575 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
576 pcpu_chunk_relocate(chunk, oslot);
577}
578
579/**
580 * pcpu_unmap - unmap pages out of a pcpu_chunk
581 * @chunk: chunk of interest
582 * @page_start: page index of the first page to unmap
583 * @page_end: page index of the last page to unmap + 1
584 * @flush: whether to flush cache and tlb or not
585 *
586 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
587 * If @flush is true, vcache is flushed before unmapping and tlb
588 * after.
589 */
590static void pcpu_unmap(struct pcpu_chunk *chunk, int page_start, int page_end,
591 bool flush)
592{
593 unsigned int last = num_possible_cpus() - 1;
594 unsigned int cpu;
595
596 /* unmap must not be done on immutable chunk */
597 WARN_ON(chunk->immutable);
598
599 /*
600 * Each flushing trial can be very expensive, issue flush on
601 * the whole region at once rather than doing it for each cpu.
602 * This could be an overkill but is more scalable.
603 */
604 if (flush)
605 flush_cache_vunmap(pcpu_chunk_addr(chunk, 0, page_start),
606 pcpu_chunk_addr(chunk, last, page_end));
607
608 for_each_possible_cpu(cpu)
609 unmap_kernel_range_noflush(
610 pcpu_chunk_addr(chunk, cpu, page_start),
611 (page_end - page_start) << PAGE_SHIFT);
612
613 /* ditto as flush_cache_vunmap() */
614 if (flush)
615 flush_tlb_kernel_range(pcpu_chunk_addr(chunk, 0, page_start),
616 pcpu_chunk_addr(chunk, last, page_end));
617}
618
619/**
620 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
621 * @chunk: chunk to depopulate
622 * @off: offset to the area to depopulate
623 * @size: size of the area to depopulate in bytes
624 * @flush: whether to flush cache and tlb or not
625 *
626 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
627 * from @chunk. If @flush is true, vcache is flushed before unmapping
628 * and tlb after.
629 *
630 * CONTEXT:
631 * pcpu_alloc_mutex.
632 */
633static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size,
634 bool flush)
635{
636 int page_start = PFN_DOWN(off);
637 int page_end = PFN_UP(off + size);
638 int unmap_start = -1;
639 int uninitialized_var(unmap_end);
640 unsigned int cpu;
641 int i;
642
643 for (i = page_start; i < page_end; i++) {
644 for_each_possible_cpu(cpu) {
645 struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i);
646
647 if (!*pagep)
648 continue;
649
650 __free_page(*pagep);
651
652 /*
653 * If it's partial depopulation, it might get
654 * populated or depopulated again. Mark the
655 * page gone.
656 */
657 *pagep = NULL;
658
659 unmap_start = unmap_start < 0 ? i : unmap_start;
660 unmap_end = i + 1;
661 }
662 }
663
664 if (unmap_start >= 0)
665 pcpu_unmap(chunk, unmap_start, unmap_end, flush);
666}
667
668/**
669 * pcpu_map - map pages into a pcpu_chunk
670 * @chunk: chunk of interest
671 * @page_start: page index of the first page to map
672 * @page_end: page index of the last page to map + 1
673 *
674 * For each cpu, map pages [@page_start,@page_end) into @chunk.
675 * vcache is flushed afterwards.
676 */
677static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end)
678{
679 unsigned int last = num_possible_cpus() - 1;
680 unsigned int cpu;
681 int err;
682
683 /* map must not be done on immutable chunk */
684 WARN_ON(chunk->immutable);
685
686 for_each_possible_cpu(cpu) {
687 err = map_kernel_range_noflush(
688 pcpu_chunk_addr(chunk, cpu, page_start),
689 (page_end - page_start) << PAGE_SHIFT,
690 PAGE_KERNEL,
691 pcpu_chunk_pagep(chunk, cpu, page_start));
692 if (err < 0)
693 return err;
694 }
695
696 /* flush at once, please read comments in pcpu_unmap() */
697 flush_cache_vmap(pcpu_chunk_addr(chunk, 0, page_start),
698 pcpu_chunk_addr(chunk, last, page_end));
699 return 0;
700}
701
702/**
703 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
704 * @chunk: chunk of interest
705 * @off: offset to the area to populate
706 * @size: size of the area to populate in bytes
707 *
708 * For each cpu, populate and map pages [@page_start,@page_end) into
709 * @chunk. The area is cleared on return.
710 *
711 * CONTEXT:
712 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
713 */
714static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
715{
716 const gfp_t alloc_mask = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
717 int page_start = PFN_DOWN(off);
718 int page_end = PFN_UP(off + size);
719 int map_start = -1;
720 int uninitialized_var(map_end);
721 unsigned int cpu;
722 int i;
723
724 for (i = page_start; i < page_end; i++) {
725 if (pcpu_chunk_page_occupied(chunk, i)) {
726 if (map_start >= 0) {
727 if (pcpu_map(chunk, map_start, map_end))
728 goto err;
729 map_start = -1;
730 }
731 continue;
732 }
733
734 map_start = map_start < 0 ? i : map_start;
735 map_end = i + 1;
736
737 for_each_possible_cpu(cpu) {
738 struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i);
739
740 *pagep = alloc_pages_node(cpu_to_node(cpu),
741 alloc_mask, 0);
742 if (!*pagep)
743 goto err;
744 }
745 }
746
747 if (map_start >= 0 && pcpu_map(chunk, map_start, map_end))
748 goto err;
749
750 for_each_possible_cpu(cpu)
751 memset(chunk->vm->addr + cpu * pcpu_unit_size + off, 0,
752 size);
753
754 return 0;
755err:
756 /* likely under heavy memory pressure, give memory back */
757 pcpu_depopulate_chunk(chunk, off, size, true);
758 return -ENOMEM;
759}
760
761static void free_pcpu_chunk(struct pcpu_chunk *chunk)
762{
763 if (!chunk)
764 return;
765 if (chunk->vm)
766 free_vm_area(chunk->vm);
767 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
768 kfree(chunk);
769}
770
771static struct pcpu_chunk *alloc_pcpu_chunk(void)
772{
773 struct pcpu_chunk *chunk;
774
775 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
776 if (!chunk)
777 return NULL;
778
779 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
780 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
781 chunk->map[chunk->map_used++] = pcpu_unit_size;
782 chunk->page = chunk->page_ar;
783
784 chunk->vm = get_vm_area(pcpu_chunk_size, GFP_KERNEL);
785 if (!chunk->vm) {
786 free_pcpu_chunk(chunk);
787 return NULL;
788 }
789
790 INIT_LIST_HEAD(&chunk->list);
791 chunk->free_size = pcpu_unit_size;
792 chunk->contig_hint = pcpu_unit_size;
793
794 return chunk;
795}
796
797/**
798 * pcpu_alloc - the percpu allocator
799 * @size: size of area to allocate in bytes
800 * @align: alignment of area (max PAGE_SIZE)
801 * @reserved: allocate from the reserved chunk if available
802 *
803 * Allocate percpu area of @size bytes aligned at @align.
804 *
805 * CONTEXT:
806 * Does GFP_KERNEL allocation.
807 *
808 * RETURNS:
809 * Percpu pointer to the allocated area on success, NULL on failure.
810 */
811static void *pcpu_alloc(size_t size, size_t align, bool reserved)
812{
813 struct pcpu_chunk *chunk;
814 int slot, off;
815
816 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
817 WARN(true, "illegal size (%zu) or align (%zu) for "
818 "percpu allocation\n", size, align);
819 return NULL;
820 }
821
822 mutex_lock(&pcpu_alloc_mutex);
823 spin_lock_irq(&pcpu_lock);
824
825 /* serve reserved allocations from the reserved chunk if available */
826 if (reserved && pcpu_reserved_chunk) {
827 chunk = pcpu_reserved_chunk;
828 if (size > chunk->contig_hint ||
829 pcpu_extend_area_map(chunk) < 0)
830 goto fail_unlock;
831 off = pcpu_alloc_area(chunk, size, align);
832 if (off >= 0)
833 goto area_found;
834 goto fail_unlock;
835 }
836
837restart:
838 /* search through normal chunks */
839 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
840 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
841 if (size > chunk->contig_hint)
842 continue;
843
844 switch (pcpu_extend_area_map(chunk)) {
845 case 0:
846 break;
847 case 1:
848 goto restart; /* pcpu_lock dropped, restart */
849 default:
850 goto fail_unlock;
851 }
852
853 off = pcpu_alloc_area(chunk, size, align);
854 if (off >= 0)
855 goto area_found;
856 }
857 }
858
859 /* hmmm... no space left, create a new chunk */
860 spin_unlock_irq(&pcpu_lock);
861
862 chunk = alloc_pcpu_chunk();
863 if (!chunk)
864 goto fail_unlock_mutex;
865
866 spin_lock_irq(&pcpu_lock);
867 pcpu_chunk_relocate(chunk, -1);
868 pcpu_chunk_addr_insert(chunk);
869 goto restart;
870
871area_found:
872 spin_unlock_irq(&pcpu_lock);
873
874 /* populate, map and clear the area */
875 if (pcpu_populate_chunk(chunk, off, size)) {
876 spin_lock_irq(&pcpu_lock);
877 pcpu_free_area(chunk, off);
878 goto fail_unlock;
879 }
880
881 mutex_unlock(&pcpu_alloc_mutex);
882
883 return __addr_to_pcpu_ptr(chunk->vm->addr + off);
884
885fail_unlock:
886 spin_unlock_irq(&pcpu_lock);
887fail_unlock_mutex:
888 mutex_unlock(&pcpu_alloc_mutex);
889 return NULL;
890}
891
892/**
893 * __alloc_percpu - allocate dynamic percpu area
894 * @size: size of area to allocate in bytes
895 * @align: alignment of area (max PAGE_SIZE)
896 *
897 * Allocate percpu area of @size bytes aligned at @align. Might
898 * sleep. Might trigger writeouts.
899 *
900 * CONTEXT:
901 * Does GFP_KERNEL allocation.
902 *
903 * RETURNS:
904 * Percpu pointer to the allocated area on success, NULL on failure.
905 */
906void *__alloc_percpu(size_t size, size_t align)
907{
908 return pcpu_alloc(size, align, false);
909}
910EXPORT_SYMBOL_GPL(__alloc_percpu);
911
912/**
913 * __alloc_reserved_percpu - allocate reserved percpu area
914 * @size: size of area to allocate in bytes
915 * @align: alignment of area (max PAGE_SIZE)
916 *
917 * Allocate percpu area of @size bytes aligned at @align from reserved
918 * percpu area if arch has set it up; otherwise, allocation is served
919 * from the same dynamic area. Might sleep. Might trigger writeouts.
920 *
921 * CONTEXT:
922 * Does GFP_KERNEL allocation.
923 *
924 * RETURNS:
925 * Percpu pointer to the allocated area on success, NULL on failure.
926 */
927void *__alloc_reserved_percpu(size_t size, size_t align)
928{
929 return pcpu_alloc(size, align, true);
930}
931
932/**
933 * pcpu_reclaim - reclaim fully free chunks, workqueue function
934 * @work: unused
935 *
936 * Reclaim all fully free chunks except for the first one.
937 *
938 * CONTEXT:
939 * workqueue context.
940 */
941static void pcpu_reclaim(struct work_struct *work)
942{
943 LIST_HEAD(todo);
944 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
945 struct pcpu_chunk *chunk, *next;
946
947 mutex_lock(&pcpu_alloc_mutex);
948 spin_lock_irq(&pcpu_lock);
949
950 list_for_each_entry_safe(chunk, next, head, list) {
951 WARN_ON(chunk->immutable);
952
953 /* spare the first one */
954 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
955 continue;
956
957 rb_erase(&chunk->rb_node, &pcpu_addr_root);
958 list_move(&chunk->list, &todo);
959 }
960
961 spin_unlock_irq(&pcpu_lock);
962 mutex_unlock(&pcpu_alloc_mutex);
963
964 list_for_each_entry_safe(chunk, next, &todo, list) {
965 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size, false);
966 free_pcpu_chunk(chunk);
967 }
968}
969
970/**
971 * free_percpu - free percpu area
972 * @ptr: pointer to area to free
973 *
974 * Free percpu area @ptr.
975 *
976 * CONTEXT:
977 * Can be called from atomic context.
978 */
979void free_percpu(void *ptr)
980{
981 void *addr = __pcpu_ptr_to_addr(ptr);
982 struct pcpu_chunk *chunk;
983 unsigned long flags;
984 int off;
985
986 if (!ptr)
987 return;
988
989 spin_lock_irqsave(&pcpu_lock, flags);
990
991 chunk = pcpu_chunk_addr_search(addr);
992 off = addr - chunk->vm->addr;
993
994 pcpu_free_area(chunk, off);
995
996 /* if there are more than one fully free chunks, wake up grim reaper */
997 if (chunk->free_size == pcpu_unit_size) {
998 struct pcpu_chunk *pos;
999
1000 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1001 if (pos != chunk) {
1002 schedule_work(&pcpu_reclaim_work);
1003 break;
1004 }
1005 }
1006
1007 spin_unlock_irqrestore(&pcpu_lock, flags);
1008}
1009EXPORT_SYMBOL_GPL(free_percpu);
1010
1011/**
1012 * pcpu_setup_first_chunk - initialize the first percpu chunk
1013 * @get_page_fn: callback to fetch page pointer
1014 * @static_size: the size of static percpu area in bytes
1015 * @reserved_size: the size of reserved percpu area in bytes
1016 * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE, -1 for auto
1017 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1018 * @base_addr: mapped address, NULL for auto
1019 * @populate_pte_fn: callback to allocate pagetable, NULL if unnecessary
1020 *
1021 * Initialize the first percpu chunk which contains the kernel static
1022 * perpcu area. This function is to be called from arch percpu area
1023 * setup path. The first two parameters are mandatory. The rest are
1024 * optional.
1025 *
1026 * @get_page_fn() should return pointer to percpu page given cpu
1027 * number and page number. It should at least return enough pages to
1028 * cover the static area. The returned pages for static area should
1029 * have been initialized with valid data. If @unit_size is specified,
1030 * it can also return pages after the static area. NULL return
1031 * indicates end of pages for the cpu. Note that @get_page_fn() must
1032 * return the same number of pages for all cpus.
1033 *
1034 * @reserved_size, if non-zero, specifies the amount of bytes to
1035 * reserve after the static area in the first chunk. This reserves
1036 * the first chunk such that it's available only through reserved
1037 * percpu allocation. This is primarily used to serve module percpu
1038 * static areas on architectures where the addressing model has
1039 * limited offset range for symbol relocations to guarantee module
1040 * percpu symbols fall inside the relocatable range.
1041 *
1042 * @unit_size, if non-negative, specifies unit size and must be
1043 * aligned to PAGE_SIZE and equal to or larger than @static_size +
1044 * @reserved_size + @dyn_size.
1045 *
1046 * @dyn_size, if non-negative, limits the number of bytes available
1047 * for dynamic allocation in the first chunk. Specifying non-negative
1048 * value make percpu leave alone the area beyond @static_size +
1049 * @reserved_size + @dyn_size.
1050 *
1051 * Non-null @base_addr means that the caller already allocated virtual
1052 * region for the first chunk and mapped it. percpu must not mess
1053 * with the chunk. Note that @base_addr with 0 @unit_size or non-NULL
1054 * @populate_pte_fn doesn't make any sense.
1055 *
1056 * @populate_pte_fn is used to populate the pagetable. NULL means the
1057 * caller already populated the pagetable.
1058 *
1059 * If the first chunk ends up with both reserved and dynamic areas, it
1060 * is served by two chunks - one to serve the core static and reserved
1061 * areas and the other for the dynamic area. They share the same vm
1062 * and page map but uses different area allocation map to stay away
1063 * from each other. The latter chunk is circulated in the chunk slots
1064 * and available for dynamic allocation like any other chunks.
1065 *
1066 * RETURNS:
1067 * The determined pcpu_unit_size which can be used to initialize
1068 * percpu access.
1069 */
1070size_t __init pcpu_setup_first_chunk(pcpu_get_page_fn_t get_page_fn,
1071 size_t static_size, size_t reserved_size,
1072 ssize_t unit_size, ssize_t dyn_size,
1073 void *base_addr,
1074 pcpu_populate_pte_fn_t populate_pte_fn)
1075{
1076 static struct vm_struct first_vm;
1077 static int smap[2], dmap[2];
1078 struct pcpu_chunk *schunk, *dchunk = NULL;
1079 unsigned int cpu;
1080 int nr_pages;
1081 int err, i;
1082
1083 /* santiy checks */
1084 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
1085 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
1086 BUG_ON(!static_size);
1087 if (unit_size >= 0) {
1088 BUG_ON(unit_size < static_size + reserved_size +
1089 (dyn_size >= 0 ? dyn_size : 0));
1090 BUG_ON(unit_size & ~PAGE_MASK);
1091 } else {
1092 BUG_ON(dyn_size >= 0);
1093 BUG_ON(base_addr);
1094 }
1095 BUG_ON(base_addr && populate_pte_fn);
1096
1097 if (unit_size >= 0)
1098 pcpu_unit_pages = unit_size >> PAGE_SHIFT;
1099 else
1100 pcpu_unit_pages = max_t(int, PCPU_MIN_UNIT_SIZE >> PAGE_SHIFT,
1101 PFN_UP(static_size + reserved_size));
1102
1103 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1104 pcpu_chunk_size = num_possible_cpus() * pcpu_unit_size;
1105 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk)
1106 + num_possible_cpus() * pcpu_unit_pages * sizeof(struct page *);
1107
1108 if (dyn_size < 0)
1109 dyn_size = pcpu_unit_size - static_size - reserved_size;
1110
1111 /*
1112 * Allocate chunk slots. The additional last slot is for
1113 * empty chunks.
1114 */
1115 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1116 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1117 for (i = 0; i < pcpu_nr_slots; i++)
1118 INIT_LIST_HEAD(&pcpu_slot[i]);
1119
1120 /*
1121 * Initialize static chunk. If reserved_size is zero, the
1122 * static chunk covers static area + dynamic allocation area
1123 * in the first chunk. If reserved_size is not zero, it
1124 * covers static area + reserved area (mostly used for module
1125 * static percpu allocation).
1126 */
1127 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1128 INIT_LIST_HEAD(&schunk->list);
1129 schunk->vm = &first_vm;
1130 schunk->map = smap;
1131 schunk->map_alloc = ARRAY_SIZE(smap);
1132 schunk->page = schunk->page_ar;
1133
1134 if (reserved_size) {
1135 schunk->free_size = reserved_size;
1136 pcpu_reserved_chunk = schunk; /* not for dynamic alloc */
1137 } else {
1138 schunk->free_size = dyn_size;
1139 dyn_size = 0; /* dynamic area covered */
1140 }
1141 schunk->contig_hint = schunk->free_size;
1142
1143 schunk->map[schunk->map_used++] = -static_size;
1144 if (schunk->free_size)
1145 schunk->map[schunk->map_used++] = schunk->free_size;
1146
1147 pcpu_reserved_chunk_limit = static_size + schunk->free_size;
1148
1149 /* init dynamic chunk if necessary */
1150 if (dyn_size) {
1151 dchunk = alloc_bootmem(sizeof(struct pcpu_chunk));
1152 INIT_LIST_HEAD(&dchunk->list);
1153 dchunk->vm = &first_vm;
1154 dchunk->map = dmap;
1155 dchunk->map_alloc = ARRAY_SIZE(dmap);
1156 dchunk->page = schunk->page_ar; /* share page map with schunk */
1157
1158 dchunk->contig_hint = dchunk->free_size = dyn_size;
1159 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1160 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1161 }
1162
1163 /* allocate vm address */
1164 first_vm.flags = VM_ALLOC;
1165 first_vm.size = pcpu_chunk_size;
1166
1167 if (!base_addr)
1168 vm_area_register_early(&first_vm, PAGE_SIZE);
1169 else {
1170 /*
1171 * Pages already mapped. No need to remap into
1172 * vmalloc area. In this case the first chunks can't
1173 * be mapped or unmapped by percpu and are marked
1174 * immutable.
1175 */
1176 first_vm.addr = base_addr;
1177 schunk->immutable = true;
1178 if (dchunk)
1179 dchunk->immutable = true;
1180 }
1181
1182 /* assign pages */
1183 nr_pages = -1;
1184 for_each_possible_cpu(cpu) {
1185 for (i = 0; i < pcpu_unit_pages; i++) {
1186 struct page *page = get_page_fn(cpu, i);
1187
1188 if (!page)
1189 break;
1190 *pcpu_chunk_pagep(schunk, cpu, i) = page;
1191 }
1192
1193 BUG_ON(i < PFN_UP(static_size));
1194
1195 if (nr_pages < 0)
1196 nr_pages = i;
1197 else
1198 BUG_ON(nr_pages != i);
1199 }
1200
1201 /* map them */
1202 if (populate_pte_fn) {
1203 for_each_possible_cpu(cpu)
1204 for (i = 0; i < nr_pages; i++)
1205 populate_pte_fn(pcpu_chunk_addr(schunk,
1206 cpu, i));
1207
1208 err = pcpu_map(schunk, 0, nr_pages);
1209 if (err)
1210 panic("failed to setup static percpu area, err=%d\n",
1211 err);
1212 }
1213
1214 /* link the first chunk in */
1215 if (!dchunk) {
1216 pcpu_chunk_relocate(schunk, -1);
1217 pcpu_chunk_addr_insert(schunk);
1218 } else {
1219 pcpu_chunk_relocate(dchunk, -1);
1220 pcpu_chunk_addr_insert(dchunk);
1221 }
1222
1223 /* we're done */
1224 pcpu_base_addr = (void *)pcpu_chunk_addr(schunk, 0, 0);
1225 return pcpu_unit_size;
1226}
diff --git a/mm/slab.c b/mm/slab.c
index 4d00855629c4..9ec66c3e6ee0 100644
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -102,6 +102,7 @@
102#include <linux/cpu.h> 102#include <linux/cpu.h>
103#include <linux/sysctl.h> 103#include <linux/sysctl.h>
104#include <linux/module.h> 104#include <linux/module.h>
105#include <trace/kmemtrace.h>
105#include <linux/rcupdate.h> 106#include <linux/rcupdate.h>
106#include <linux/string.h> 107#include <linux/string.h>
107#include <linux/uaccess.h> 108#include <linux/uaccess.h>
@@ -568,6 +569,14 @@ static void **dbg_userword(struct kmem_cache *cachep, void *objp)
568 569
569#endif 570#endif
570 571
572#ifdef CONFIG_KMEMTRACE
573size_t slab_buffer_size(struct kmem_cache *cachep)
574{
575 return cachep->buffer_size;
576}
577EXPORT_SYMBOL(slab_buffer_size);
578#endif
579
571/* 580/*
572 * Do not go above this order unless 0 objects fit into the slab. 581 * Do not go above this order unless 0 objects fit into the slab.
573 */ 582 */
@@ -3318,6 +3327,8 @@ __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3318 unsigned long save_flags; 3327 unsigned long save_flags;
3319 void *ptr; 3328 void *ptr;
3320 3329
3330 lockdep_trace_alloc(flags);
3331
3321 if (slab_should_failslab(cachep, flags)) 3332 if (slab_should_failslab(cachep, flags))
3322 return NULL; 3333 return NULL;
3323 3334
@@ -3394,6 +3405,8 @@ __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3394 unsigned long save_flags; 3405 unsigned long save_flags;
3395 void *objp; 3406 void *objp;
3396 3407
3408 lockdep_trace_alloc(flags);
3409
3397 if (slab_should_failslab(cachep, flags)) 3410 if (slab_should_failslab(cachep, flags))
3398 return NULL; 3411 return NULL;
3399 3412
@@ -3550,10 +3563,23 @@ static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3550 */ 3563 */
3551void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3564void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3552{ 3565{
3553 return __cache_alloc(cachep, flags, __builtin_return_address(0)); 3566 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3567
3568 kmemtrace_mark_alloc(KMEMTRACE_TYPE_CACHE, _RET_IP_, ret,
3569 obj_size(cachep), cachep->buffer_size, flags);
3570
3571 return ret;
3554} 3572}
3555EXPORT_SYMBOL(kmem_cache_alloc); 3573EXPORT_SYMBOL(kmem_cache_alloc);
3556 3574
3575#ifdef CONFIG_KMEMTRACE
3576void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3577{
3578 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3579}
3580EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3581#endif
3582
3557/** 3583/**
3558 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry. 3584 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
3559 * @cachep: the cache we're checking against 3585 * @cachep: the cache we're checking against
@@ -3598,23 +3624,47 @@ out:
3598#ifdef CONFIG_NUMA 3624#ifdef CONFIG_NUMA
3599void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) 3625void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3600{ 3626{
3601 return __cache_alloc_node(cachep, flags, nodeid, 3627 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3602 __builtin_return_address(0)); 3628 __builtin_return_address(0));
3629
3630 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_CACHE, _RET_IP_, ret,
3631 obj_size(cachep), cachep->buffer_size,
3632 flags, nodeid);
3633
3634 return ret;
3603} 3635}
3604EXPORT_SYMBOL(kmem_cache_alloc_node); 3636EXPORT_SYMBOL(kmem_cache_alloc_node);
3605 3637
3638#ifdef CONFIG_KMEMTRACE
3639void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3640 gfp_t flags,
3641 int nodeid)
3642{
3643 return __cache_alloc_node(cachep, flags, nodeid,
3644 __builtin_return_address(0));
3645}
3646EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3647#endif
3648
3606static __always_inline void * 3649static __always_inline void *
3607__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) 3650__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3608{ 3651{
3609 struct kmem_cache *cachep; 3652 struct kmem_cache *cachep;
3653 void *ret;
3610 3654
3611 cachep = kmem_find_general_cachep(size, flags); 3655 cachep = kmem_find_general_cachep(size, flags);
3612 if (unlikely(ZERO_OR_NULL_PTR(cachep))) 3656 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3613 return cachep; 3657 return cachep;
3614 return kmem_cache_alloc_node(cachep, flags, node); 3658 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3659
3660 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_KMALLOC,
3661 (unsigned long) caller, ret,
3662 size, cachep->buffer_size, flags, node);
3663
3664 return ret;
3615} 3665}
3616 3666
3617#ifdef CONFIG_DEBUG_SLAB 3667#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
3618void *__kmalloc_node(size_t size, gfp_t flags, int node) 3668void *__kmalloc_node(size_t size, gfp_t flags, int node)
3619{ 3669{
3620 return __do_kmalloc_node(size, flags, node, 3670 return __do_kmalloc_node(size, flags, node,
@@ -3647,6 +3697,7 @@ static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3647 void *caller) 3697 void *caller)
3648{ 3698{
3649 struct kmem_cache *cachep; 3699 struct kmem_cache *cachep;
3700 void *ret;
3650 3701
3651 /* If you want to save a few bytes .text space: replace 3702 /* If you want to save a few bytes .text space: replace
3652 * __ with kmem_. 3703 * __ with kmem_.
@@ -3656,11 +3707,17 @@ static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3656 cachep = __find_general_cachep(size, flags); 3707 cachep = __find_general_cachep(size, flags);
3657 if (unlikely(ZERO_OR_NULL_PTR(cachep))) 3708 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3658 return cachep; 3709 return cachep;
3659 return __cache_alloc(cachep, flags, caller); 3710 ret = __cache_alloc(cachep, flags, caller);
3711
3712 kmemtrace_mark_alloc(KMEMTRACE_TYPE_KMALLOC,
3713 (unsigned long) caller, ret,
3714 size, cachep->buffer_size, flags);
3715
3716 return ret;
3660} 3717}
3661 3718
3662 3719
3663#ifdef CONFIG_DEBUG_SLAB 3720#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
3664void *__kmalloc(size_t size, gfp_t flags) 3721void *__kmalloc(size_t size, gfp_t flags)
3665{ 3722{
3666 return __do_kmalloc(size, flags, __builtin_return_address(0)); 3723 return __do_kmalloc(size, flags, __builtin_return_address(0));
@@ -3699,6 +3756,8 @@ void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3699 debug_check_no_obj_freed(objp, obj_size(cachep)); 3756 debug_check_no_obj_freed(objp, obj_size(cachep));
3700 __cache_free(cachep, objp); 3757 __cache_free(cachep, objp);
3701 local_irq_restore(flags); 3758 local_irq_restore(flags);
3759
3760 kmemtrace_mark_free(KMEMTRACE_TYPE_CACHE, _RET_IP_, objp);
3702} 3761}
3703EXPORT_SYMBOL(kmem_cache_free); 3762EXPORT_SYMBOL(kmem_cache_free);
3704 3763
@@ -3725,6 +3784,8 @@ void kfree(const void *objp)
3725 debug_check_no_obj_freed(objp, obj_size(c)); 3784 debug_check_no_obj_freed(objp, obj_size(c));
3726 __cache_free(c, (void *)objp); 3785 __cache_free(c, (void *)objp);
3727 local_irq_restore(flags); 3786 local_irq_restore(flags);
3787
3788 kmemtrace_mark_free(KMEMTRACE_TYPE_KMALLOC, _RET_IP_, objp);
3728} 3789}
3729EXPORT_SYMBOL(kfree); 3790EXPORT_SYMBOL(kfree);
3730 3791
diff --git a/mm/slob.c b/mm/slob.c
index 52bc8a2bd9ef..596152926a8d 100644
--- a/mm/slob.c
+++ b/mm/slob.c
@@ -65,6 +65,7 @@
65#include <linux/module.h> 65#include <linux/module.h>
66#include <linux/rcupdate.h> 66#include <linux/rcupdate.h>
67#include <linux/list.h> 67#include <linux/list.h>
68#include <trace/kmemtrace.h>
68#include <asm/atomic.h> 69#include <asm/atomic.h>
69 70
70/* 71/*
@@ -463,27 +464,40 @@ void *__kmalloc_node(size_t size, gfp_t gfp, int node)
463{ 464{
464 unsigned int *m; 465 unsigned int *m;
465 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN); 466 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
467 void *ret;
468
469 lockdep_trace_alloc(flags);
466 470
467 if (size < PAGE_SIZE - align) { 471 if (size < PAGE_SIZE - align) {
468 if (!size) 472 if (!size)
469 return ZERO_SIZE_PTR; 473 return ZERO_SIZE_PTR;
470 474
471 m = slob_alloc(size + align, gfp, align, node); 475 m = slob_alloc(size + align, gfp, align, node);
476
472 if (!m) 477 if (!m)
473 return NULL; 478 return NULL;
474 *m = size; 479 *m = size;
475 return (void *)m + align; 480 ret = (void *)m + align;
481
482 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_KMALLOC,
483 _RET_IP_, ret,
484 size, size + align, gfp, node);
476 } else { 485 } else {
477 void *ret; 486 unsigned int order = get_order(size);
478 487
479 ret = slob_new_page(gfp | __GFP_COMP, get_order(size), node); 488 ret = slob_new_page(gfp | __GFP_COMP, order, node);
480 if (ret) { 489 if (ret) {
481 struct page *page; 490 struct page *page;
482 page = virt_to_page(ret); 491 page = virt_to_page(ret);
483 page->private = size; 492 page->private = size;
484 } 493 }
485 return ret; 494
495 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_KMALLOC,
496 _RET_IP_, ret,
497 size, PAGE_SIZE << order, gfp, node);
486 } 498 }
499
500 return ret;
487} 501}
488EXPORT_SYMBOL(__kmalloc_node); 502EXPORT_SYMBOL(__kmalloc_node);
489 503
@@ -501,6 +515,8 @@ void kfree(const void *block)
501 slob_free(m, *m + align); 515 slob_free(m, *m + align);
502 } else 516 } else
503 put_page(&sp->page); 517 put_page(&sp->page);
518
519 kmemtrace_mark_free(KMEMTRACE_TYPE_KMALLOC, _RET_IP_, block);
504} 520}
505EXPORT_SYMBOL(kfree); 521EXPORT_SYMBOL(kfree);
506 522
@@ -570,10 +586,19 @@ void *kmem_cache_alloc_node(struct kmem_cache *c, gfp_t flags, int node)
570{ 586{
571 void *b; 587 void *b;
572 588
573 if (c->size < PAGE_SIZE) 589 if (c->size < PAGE_SIZE) {
574 b = slob_alloc(c->size, flags, c->align, node); 590 b = slob_alloc(c->size, flags, c->align, node);
575 else 591 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_CACHE,
592 _RET_IP_, b, c->size,
593 SLOB_UNITS(c->size) * SLOB_UNIT,
594 flags, node);
595 } else {
576 b = slob_new_page(flags, get_order(c->size), node); 596 b = slob_new_page(flags, get_order(c->size), node);
597 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_CACHE,
598 _RET_IP_, b, c->size,
599 PAGE_SIZE << get_order(c->size),
600 flags, node);
601 }
577 602
578 if (c->ctor) 603 if (c->ctor)
579 c->ctor(b); 604 c->ctor(b);
@@ -609,6 +634,8 @@ void kmem_cache_free(struct kmem_cache *c, void *b)
609 } else { 634 } else {
610 __kmem_cache_free(b, c->size); 635 __kmem_cache_free(b, c->size);
611 } 636 }
637
638 kmemtrace_mark_free(KMEMTRACE_TYPE_CACHE, _RET_IP_, b);
612} 639}
613EXPORT_SYMBOL(kmem_cache_free); 640EXPORT_SYMBOL(kmem_cache_free);
614 641
diff --git a/mm/slub.c b/mm/slub.c
index 0280eee6cf37..816734ed8aa3 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -16,6 +16,7 @@
16#include <linux/slab.h> 16#include <linux/slab.h>
17#include <linux/proc_fs.h> 17#include <linux/proc_fs.h>
18#include <linux/seq_file.h> 18#include <linux/seq_file.h>
19#include <trace/kmemtrace.h>
19#include <linux/cpu.h> 20#include <linux/cpu.h>
20#include <linux/cpuset.h> 21#include <linux/cpuset.h>
21#include <linux/mempolicy.h> 22#include <linux/mempolicy.h>
@@ -1596,6 +1597,7 @@ static __always_inline void *slab_alloc(struct kmem_cache *s,
1596 unsigned long flags; 1597 unsigned long flags;
1597 unsigned int objsize; 1598 unsigned int objsize;
1598 1599
1600 lockdep_trace_alloc(gfpflags);
1599 might_sleep_if(gfpflags & __GFP_WAIT); 1601 might_sleep_if(gfpflags & __GFP_WAIT);
1600 1602
1601 if (should_failslab(s->objsize, gfpflags)) 1603 if (should_failslab(s->objsize, gfpflags))
@@ -1623,18 +1625,46 @@ static __always_inline void *slab_alloc(struct kmem_cache *s,
1623 1625
1624void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 1626void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1625{ 1627{
1626 return slab_alloc(s, gfpflags, -1, _RET_IP_); 1628 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_);
1629
1630 kmemtrace_mark_alloc(KMEMTRACE_TYPE_CACHE, _RET_IP_, ret,
1631 s->objsize, s->size, gfpflags);
1632
1633 return ret;
1627} 1634}
1628EXPORT_SYMBOL(kmem_cache_alloc); 1635EXPORT_SYMBOL(kmem_cache_alloc);
1629 1636
1637#ifdef CONFIG_KMEMTRACE
1638void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1639{
1640 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1641}
1642EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1643#endif
1644
1630#ifdef CONFIG_NUMA 1645#ifdef CONFIG_NUMA
1631void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 1646void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1632{ 1647{
1633 return slab_alloc(s, gfpflags, node, _RET_IP_); 1648 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1649
1650 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_CACHE, _RET_IP_, ret,
1651 s->objsize, s->size, gfpflags, node);
1652
1653 return ret;
1634} 1654}
1635EXPORT_SYMBOL(kmem_cache_alloc_node); 1655EXPORT_SYMBOL(kmem_cache_alloc_node);
1636#endif 1656#endif
1637 1657
1658#ifdef CONFIG_KMEMTRACE
1659void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1660 gfp_t gfpflags,
1661 int node)
1662{
1663 return slab_alloc(s, gfpflags, node, _RET_IP_);
1664}
1665EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1666#endif
1667
1638/* 1668/*
1639 * Slow patch handling. This may still be called frequently since objects 1669 * Slow patch handling. This may still be called frequently since objects
1640 * have a longer lifetime than the cpu slabs in most processing loads. 1670 * have a longer lifetime than the cpu slabs in most processing loads.
@@ -1742,6 +1772,8 @@ void kmem_cache_free(struct kmem_cache *s, void *x)
1742 page = virt_to_head_page(x); 1772 page = virt_to_head_page(x);
1743 1773
1744 slab_free(s, page, x, _RET_IP_); 1774 slab_free(s, page, x, _RET_IP_);
1775
1776 kmemtrace_mark_free(KMEMTRACE_TYPE_CACHE, _RET_IP_, x);
1745} 1777}
1746EXPORT_SYMBOL(kmem_cache_free); 1778EXPORT_SYMBOL(kmem_cache_free);
1747 1779
@@ -2475,7 +2507,7 @@ EXPORT_SYMBOL(kmem_cache_destroy);
2475 * Kmalloc subsystem 2507 * Kmalloc subsystem
2476 *******************************************************************/ 2508 *******************************************************************/
2477 2509
2478struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned; 2510struct kmem_cache kmalloc_caches[SLUB_PAGE_SHIFT] __cacheline_aligned;
2479EXPORT_SYMBOL(kmalloc_caches); 2511EXPORT_SYMBOL(kmalloc_caches);
2480 2512
2481static int __init setup_slub_min_order(char *str) 2513static int __init setup_slub_min_order(char *str)
@@ -2537,7 +2569,7 @@ panic:
2537} 2569}
2538 2570
2539#ifdef CONFIG_ZONE_DMA 2571#ifdef CONFIG_ZONE_DMA
2540static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1]; 2572static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
2541 2573
2542static void sysfs_add_func(struct work_struct *w) 2574static void sysfs_add_func(struct work_struct *w)
2543{ 2575{
@@ -2657,8 +2689,9 @@ static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2657void *__kmalloc(size_t size, gfp_t flags) 2689void *__kmalloc(size_t size, gfp_t flags)
2658{ 2690{
2659 struct kmem_cache *s; 2691 struct kmem_cache *s;
2692 void *ret;
2660 2693
2661 if (unlikely(size > PAGE_SIZE)) 2694 if (unlikely(size > SLUB_MAX_SIZE))
2662 return kmalloc_large(size, flags); 2695 return kmalloc_large(size, flags);
2663 2696
2664 s = get_slab(size, flags); 2697 s = get_slab(size, flags);
@@ -2666,7 +2699,12 @@ void *__kmalloc(size_t size, gfp_t flags)
2666 if (unlikely(ZERO_OR_NULL_PTR(s))) 2699 if (unlikely(ZERO_OR_NULL_PTR(s)))
2667 return s; 2700 return s;
2668 2701
2669 return slab_alloc(s, flags, -1, _RET_IP_); 2702 ret = slab_alloc(s, flags, -1, _RET_IP_);
2703
2704 kmemtrace_mark_alloc(KMEMTRACE_TYPE_KMALLOC, _RET_IP_, ret,
2705 size, s->size, flags);
2706
2707 return ret;
2670} 2708}
2671EXPORT_SYMBOL(__kmalloc); 2709EXPORT_SYMBOL(__kmalloc);
2672 2710
@@ -2685,16 +2723,30 @@ static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2685void *__kmalloc_node(size_t size, gfp_t flags, int node) 2723void *__kmalloc_node(size_t size, gfp_t flags, int node)
2686{ 2724{
2687 struct kmem_cache *s; 2725 struct kmem_cache *s;
2726 void *ret;
2727
2728 if (unlikely(size > SLUB_MAX_SIZE)) {
2729 ret = kmalloc_large_node(size, flags, node);
2688 2730
2689 if (unlikely(size > PAGE_SIZE)) 2731 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_KMALLOC,
2690 return kmalloc_large_node(size, flags, node); 2732 _RET_IP_, ret,
2733 size, PAGE_SIZE << get_order(size),
2734 flags, node);
2735
2736 return ret;
2737 }
2691 2738
2692 s = get_slab(size, flags); 2739 s = get_slab(size, flags);
2693 2740
2694 if (unlikely(ZERO_OR_NULL_PTR(s))) 2741 if (unlikely(ZERO_OR_NULL_PTR(s)))
2695 return s; 2742 return s;
2696 2743
2697 return slab_alloc(s, flags, node, _RET_IP_); 2744 ret = slab_alloc(s, flags, node, _RET_IP_);
2745
2746 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_KMALLOC, _RET_IP_, ret,
2747 size, s->size, flags, node);
2748
2749 return ret;
2698} 2750}
2699EXPORT_SYMBOL(__kmalloc_node); 2751EXPORT_SYMBOL(__kmalloc_node);
2700#endif 2752#endif
@@ -2753,6 +2805,8 @@ void kfree(const void *x)
2753 return; 2805 return;
2754 } 2806 }
2755 slab_free(page->slab, page, object, _RET_IP_); 2807 slab_free(page->slab, page, object, _RET_IP_);
2808
2809 kmemtrace_mark_free(KMEMTRACE_TYPE_KMALLOC, _RET_IP_, x);
2756} 2810}
2757EXPORT_SYMBOL(kfree); 2811EXPORT_SYMBOL(kfree);
2758 2812
@@ -2986,7 +3040,7 @@ void __init kmem_cache_init(void)
2986 caches++; 3040 caches++;
2987 } 3041 }
2988 3042
2989 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) { 3043 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
2990 create_kmalloc_cache(&kmalloc_caches[i], 3044 create_kmalloc_cache(&kmalloc_caches[i],
2991 "kmalloc", 1 << i, GFP_KERNEL); 3045 "kmalloc", 1 << i, GFP_KERNEL);
2992 caches++; 3046 caches++;
@@ -3023,7 +3077,7 @@ void __init kmem_cache_init(void)
3023 slab_state = UP; 3077 slab_state = UP;
3024 3078
3025 /* Provide the correct kmalloc names now that the caches are up */ 3079 /* Provide the correct kmalloc names now that the caches are up */
3026 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) 3080 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
3027 kmalloc_caches[i]. name = 3081 kmalloc_caches[i]. name =
3028 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i); 3082 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3029 3083
@@ -3222,8 +3276,9 @@ static struct notifier_block __cpuinitdata slab_notifier = {
3222void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) 3276void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3223{ 3277{
3224 struct kmem_cache *s; 3278 struct kmem_cache *s;
3279 void *ret;
3225 3280
3226 if (unlikely(size > PAGE_SIZE)) 3281 if (unlikely(size > SLUB_MAX_SIZE))
3227 return kmalloc_large(size, gfpflags); 3282 return kmalloc_large(size, gfpflags);
3228 3283
3229 s = get_slab(size, gfpflags); 3284 s = get_slab(size, gfpflags);
@@ -3231,15 +3286,22 @@ void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3231 if (unlikely(ZERO_OR_NULL_PTR(s))) 3286 if (unlikely(ZERO_OR_NULL_PTR(s)))
3232 return s; 3287 return s;
3233 3288
3234 return slab_alloc(s, gfpflags, -1, caller); 3289 ret = slab_alloc(s, gfpflags, -1, caller);
3290
3291 /* Honor the call site pointer we recieved. */
3292 kmemtrace_mark_alloc(KMEMTRACE_TYPE_KMALLOC, caller, ret, size,
3293 s->size, gfpflags);
3294
3295 return ret;
3235} 3296}
3236 3297
3237void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 3298void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3238 int node, unsigned long caller) 3299 int node, unsigned long caller)
3239{ 3300{
3240 struct kmem_cache *s; 3301 struct kmem_cache *s;
3302 void *ret;
3241 3303
3242 if (unlikely(size > PAGE_SIZE)) 3304 if (unlikely(size > SLUB_MAX_SIZE))
3243 return kmalloc_large_node(size, gfpflags, node); 3305 return kmalloc_large_node(size, gfpflags, node);
3244 3306
3245 s = get_slab(size, gfpflags); 3307 s = get_slab(size, gfpflags);
@@ -3247,7 +3309,13 @@ void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3247 if (unlikely(ZERO_OR_NULL_PTR(s))) 3309 if (unlikely(ZERO_OR_NULL_PTR(s)))
3248 return s; 3310 return s;
3249 3311
3250 return slab_alloc(s, gfpflags, node, caller); 3312 ret = slab_alloc(s, gfpflags, node, caller);
3313
3314 /* Honor the call site pointer we recieved. */
3315 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_KMALLOC, caller, ret,
3316 size, s->size, gfpflags, node);
3317
3318 return ret;
3251} 3319}
3252 3320
3253#ifdef CONFIG_SLUB_DEBUG 3321#ifdef CONFIG_SLUB_DEBUG
diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 520a75980269..af58324c361a 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -24,6 +24,7 @@
24#include <linux/radix-tree.h> 24#include <linux/radix-tree.h>
25#include <linux/rcupdate.h> 25#include <linux/rcupdate.h>
26#include <linux/bootmem.h> 26#include <linux/bootmem.h>
27#include <linux/pfn.h>
27 28
28#include <asm/atomic.h> 29#include <asm/atomic.h>
29#include <asm/uaccess.h> 30#include <asm/uaccess.h>
@@ -152,8 +153,8 @@ static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
152 * 153 *
153 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N] 154 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
154 */ 155 */
155static int vmap_page_range(unsigned long start, unsigned long end, 156static int vmap_page_range_noflush(unsigned long start, unsigned long end,
156 pgprot_t prot, struct page **pages) 157 pgprot_t prot, struct page **pages)
157{ 158{
158 pgd_t *pgd; 159 pgd_t *pgd;
159 unsigned long next; 160 unsigned long next;
@@ -169,13 +170,22 @@ static int vmap_page_range(unsigned long start, unsigned long end,
169 if (err) 170 if (err)
170 break; 171 break;
171 } while (pgd++, addr = next, addr != end); 172 } while (pgd++, addr = next, addr != end);
172 flush_cache_vmap(start, end);
173 173
174 if (unlikely(err)) 174 if (unlikely(err))
175 return err; 175 return err;
176 return nr; 176 return nr;
177} 177}
178 178
179static int vmap_page_range(unsigned long start, unsigned long end,
180 pgprot_t prot, struct page **pages)
181{
182 int ret;
183
184 ret = vmap_page_range_noflush(start, end, prot, pages);
185 flush_cache_vmap(start, end);
186 return ret;
187}
188
179static inline int is_vmalloc_or_module_addr(const void *x) 189static inline int is_vmalloc_or_module_addr(const void *x)
180{ 190{
181 /* 191 /*
@@ -990,6 +1000,32 @@ void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t pro
990} 1000}
991EXPORT_SYMBOL(vm_map_ram); 1001EXPORT_SYMBOL(vm_map_ram);
992 1002
1003/**
1004 * vm_area_register_early - register vmap area early during boot
1005 * @vm: vm_struct to register
1006 * @align: requested alignment
1007 *
1008 * This function is used to register kernel vm area before
1009 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1010 * proper values on entry and other fields should be zero. On return,
1011 * vm->addr contains the allocated address.
1012 *
1013 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1014 */
1015void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1016{
1017 static size_t vm_init_off __initdata;
1018 unsigned long addr;
1019
1020 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1021 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1022
1023 vm->addr = (void *)addr;
1024
1025 vm->next = vmlist;
1026 vmlist = vm;
1027}
1028
993void __init vmalloc_init(void) 1029void __init vmalloc_init(void)
994{ 1030{
995 struct vmap_area *va; 1031 struct vmap_area *va;
@@ -1017,6 +1053,58 @@ void __init vmalloc_init(void)
1017 vmap_initialized = true; 1053 vmap_initialized = true;
1018} 1054}
1019 1055
1056/**
1057 * map_kernel_range_noflush - map kernel VM area with the specified pages
1058 * @addr: start of the VM area to map
1059 * @size: size of the VM area to map
1060 * @prot: page protection flags to use
1061 * @pages: pages to map
1062 *
1063 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1064 * specify should have been allocated using get_vm_area() and its
1065 * friends.
1066 *
1067 * NOTE:
1068 * This function does NOT do any cache flushing. The caller is
1069 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1070 * before calling this function.
1071 *
1072 * RETURNS:
1073 * The number of pages mapped on success, -errno on failure.
1074 */
1075int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1076 pgprot_t prot, struct page **pages)
1077{
1078 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1079}
1080
1081/**
1082 * unmap_kernel_range_noflush - unmap kernel VM area
1083 * @addr: start of the VM area to unmap
1084 * @size: size of the VM area to unmap
1085 *
1086 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1087 * specify should have been allocated using get_vm_area() and its
1088 * friends.
1089 *
1090 * NOTE:
1091 * This function does NOT do any cache flushing. The caller is
1092 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1093 * before calling this function and flush_tlb_kernel_range() after.
1094 */
1095void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1096{
1097 vunmap_page_range(addr, addr + size);
1098}
1099
1100/**
1101 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1102 * @addr: start of the VM area to unmap
1103 * @size: size of the VM area to unmap
1104 *
1105 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1106 * the unmapping and tlb after.
1107 */
1020void unmap_kernel_range(unsigned long addr, unsigned long size) 1108void unmap_kernel_range(unsigned long addr, unsigned long size)
1021{ 1109{
1022 unsigned long end = addr + size; 1110 unsigned long end = addr + size;
@@ -1267,6 +1355,7 @@ EXPORT_SYMBOL(vfree);
1267void vunmap(const void *addr) 1355void vunmap(const void *addr)
1268{ 1356{
1269 BUG_ON(in_interrupt()); 1357 BUG_ON(in_interrupt());
1358 might_sleep();
1270 __vunmap(addr, 0); 1359 __vunmap(addr, 0);
1271} 1360}
1272EXPORT_SYMBOL(vunmap); 1361EXPORT_SYMBOL(vunmap);
@@ -1286,6 +1375,8 @@ void *vmap(struct page **pages, unsigned int count,
1286{ 1375{
1287 struct vm_struct *area; 1376 struct vm_struct *area;
1288 1377
1378 might_sleep();
1379
1289 if (count > num_physpages) 1380 if (count > num_physpages)
1290 return NULL; 1381 return NULL;
1291 1382
diff --git a/mm/vmscan.c b/mm/vmscan.c
index 56ddf41149eb..479e46719394 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -1965,6 +1965,8 @@ static int kswapd(void *p)
1965 }; 1965 };
1966 node_to_cpumask_ptr(cpumask, pgdat->node_id); 1966 node_to_cpumask_ptr(cpumask, pgdat->node_id);
1967 1967
1968 lockdep_set_current_reclaim_state(GFP_KERNEL);
1969
1968 if (!cpumask_empty(cpumask)) 1970 if (!cpumask_empty(cpumask))
1969 set_cpus_allowed_ptr(tsk, cpumask); 1971 set_cpus_allowed_ptr(tsk, cpumask);
1970 current->reclaim_state = &reclaim_state; 1972 current->reclaim_state = &reclaim_state;
generated by cgit v1.2.2 (git 2.25.0) at 2024-11-08 00:53:35 -0500