1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016,2017 Facebook
8 #include <linux/slab.h>
10 #include <linux/filter.h>
11 #include <linux/perf_event.h>
12 #include <uapi/linux/btf.h>
13 #include <linux/rcupdate_trace.h>
14 #include <linux/btf_ids.h>
16 #include "map_in_map.h"
18 #define ARRAY_CREATE_FLAG_MASK \
19 (BPF_F_NUMA_NODE | BPF_F_MMAPABLE | BPF_F_ACCESS_MASK | \
20 BPF_F_PRESERVE_ELEMS | BPF_F_INNER_MAP)
22 static void bpf_array_free_percpu(struct bpf_array *array)
26 for (i = 0; i < array->map.max_entries; i++) {
27 free_percpu(array->pptrs[i]);
32 static int bpf_array_alloc_percpu(struct bpf_array *array)
37 for (i = 0; i < array->map.max_entries; i++) {
38 ptr = bpf_map_alloc_percpu(&array->map, array->elem_size, 8,
39 GFP_USER | __GFP_NOWARN);
41 bpf_array_free_percpu(array);
44 array->pptrs[i] = ptr;
51 /* Called from syscall */
52 int array_map_alloc_check(union bpf_attr *attr)
54 bool percpu = attr->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
55 int numa_node = bpf_map_attr_numa_node(attr);
57 /* check sanity of attributes */
58 if (attr->max_entries == 0 || attr->key_size != 4 ||
59 attr->value_size == 0 ||
60 attr->map_flags & ~ARRAY_CREATE_FLAG_MASK ||
61 !bpf_map_flags_access_ok(attr->map_flags) ||
62 (percpu && numa_node != NUMA_NO_NODE))
65 if (attr->map_type != BPF_MAP_TYPE_ARRAY &&
66 attr->map_flags & (BPF_F_MMAPABLE | BPF_F_INNER_MAP))
69 if (attr->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY &&
70 attr->map_flags & BPF_F_PRESERVE_ELEMS)
73 if (attr->value_size > KMALLOC_MAX_SIZE)
74 /* if value_size is bigger, the user space won't be able to
75 * access the elements.
82 static struct bpf_map *array_map_alloc(union bpf_attr *attr)
84 bool percpu = attr->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
85 int numa_node = bpf_map_attr_numa_node(attr);
86 u32 elem_size, index_mask, max_entries;
87 bool bypass_spec_v1 = bpf_bypass_spec_v1();
88 u64 array_size, mask64;
89 struct bpf_array *array;
91 elem_size = round_up(attr->value_size, 8);
93 max_entries = attr->max_entries;
95 /* On 32 bit archs roundup_pow_of_two() with max_entries that has
96 * upper most bit set in u32 space is undefined behavior due to
97 * resulting 1U << 32, so do it manually here in u64 space.
99 mask64 = fls_long(max_entries - 1);
100 mask64 = 1ULL << mask64;
104 if (!bypass_spec_v1) {
105 /* round up array size to nearest power of 2,
106 * since cpu will speculate within index_mask limits
108 max_entries = index_mask + 1;
109 /* Check for overflows. */
110 if (max_entries < attr->max_entries)
111 return ERR_PTR(-E2BIG);
114 array_size = sizeof(*array);
116 array_size += (u64) max_entries * sizeof(void *);
118 /* rely on vmalloc() to return page-aligned memory and
119 * ensure array->value is exactly page-aligned
121 if (attr->map_flags & BPF_F_MMAPABLE) {
122 array_size = PAGE_ALIGN(array_size);
123 array_size += PAGE_ALIGN((u64) max_entries * elem_size);
125 array_size += (u64) max_entries * elem_size;
129 /* allocate all map elements and zero-initialize them */
130 if (attr->map_flags & BPF_F_MMAPABLE) {
133 /* kmalloc'ed memory can't be mmap'ed, use explicit vmalloc */
134 data = bpf_map_area_mmapable_alloc(array_size, numa_node);
136 return ERR_PTR(-ENOMEM);
137 array = data + PAGE_ALIGN(sizeof(struct bpf_array))
138 - offsetof(struct bpf_array, value);
140 array = bpf_map_area_alloc(array_size, numa_node);
143 return ERR_PTR(-ENOMEM);
144 array->index_mask = index_mask;
145 array->map.bypass_spec_v1 = bypass_spec_v1;
147 /* copy mandatory map attributes */
148 bpf_map_init_from_attr(&array->map, attr);
149 array->elem_size = elem_size;
151 if (percpu && bpf_array_alloc_percpu(array)) {
152 bpf_map_area_free(array);
153 return ERR_PTR(-ENOMEM);
159 static void *array_map_elem_ptr(struct bpf_array* array, u32 index)
161 return array->value + (u64)array->elem_size * index;
164 /* Called from syscall or from eBPF program */
165 static void *array_map_lookup_elem(struct bpf_map *map, void *key)
167 struct bpf_array *array = container_of(map, struct bpf_array, map);
168 u32 index = *(u32 *)key;
170 if (unlikely(index >= array->map.max_entries))
173 return array->value + (u64)array->elem_size * (index & array->index_mask);
176 static int array_map_direct_value_addr(const struct bpf_map *map, u64 *imm,
179 struct bpf_array *array = container_of(map, struct bpf_array, map);
181 if (map->max_entries != 1)
183 if (off >= map->value_size)
186 *imm = (unsigned long)array->value;
190 static int array_map_direct_value_meta(const struct bpf_map *map, u64 imm,
193 struct bpf_array *array = container_of(map, struct bpf_array, map);
194 u64 base = (unsigned long)array->value;
195 u64 range = array->elem_size;
197 if (map->max_entries != 1)
199 if (imm < base || imm >= base + range)
206 /* emit BPF instructions equivalent to C code of array_map_lookup_elem() */
207 static int array_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
209 struct bpf_array *array = container_of(map, struct bpf_array, map);
210 struct bpf_insn *insn = insn_buf;
211 u32 elem_size = array->elem_size;
212 const int ret = BPF_REG_0;
213 const int map_ptr = BPF_REG_1;
214 const int index = BPF_REG_2;
216 if (map->map_flags & BPF_F_INNER_MAP)
219 *insn++ = BPF_ALU64_IMM(BPF_ADD, map_ptr, offsetof(struct bpf_array, value));
220 *insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
221 if (!map->bypass_spec_v1) {
222 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 4);
223 *insn++ = BPF_ALU32_IMM(BPF_AND, ret, array->index_mask);
225 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 3);
228 if (is_power_of_2(elem_size)) {
229 *insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(elem_size));
231 *insn++ = BPF_ALU64_IMM(BPF_MUL, ret, elem_size);
233 *insn++ = BPF_ALU64_REG(BPF_ADD, ret, map_ptr);
234 *insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
235 *insn++ = BPF_MOV64_IMM(ret, 0);
236 return insn - insn_buf;
239 /* Called from eBPF program */
240 static void *percpu_array_map_lookup_elem(struct bpf_map *map, void *key)
242 struct bpf_array *array = container_of(map, struct bpf_array, map);
243 u32 index = *(u32 *)key;
245 if (unlikely(index >= array->map.max_entries))
248 return this_cpu_ptr(array->pptrs[index & array->index_mask]);
251 static void *percpu_array_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
253 struct bpf_array *array = container_of(map, struct bpf_array, map);
254 u32 index = *(u32 *)key;
256 if (cpu >= nr_cpu_ids)
259 if (unlikely(index >= array->map.max_entries))
262 return per_cpu_ptr(array->pptrs[index & array->index_mask], cpu);
265 int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value)
267 struct bpf_array *array = container_of(map, struct bpf_array, map);
268 u32 index = *(u32 *)key;
273 if (unlikely(index >= array->map.max_entries))
276 /* per_cpu areas are zero-filled and bpf programs can only
277 * access 'value_size' of them, so copying rounded areas
278 * will not leak any kernel data
280 size = array->elem_size;
282 pptr = array->pptrs[index & array->index_mask];
283 for_each_possible_cpu(cpu) {
284 bpf_long_memcpy(value + off, per_cpu_ptr(pptr, cpu), size);
291 /* Called from syscall */
292 static int array_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
294 struct bpf_array *array = container_of(map, struct bpf_array, map);
295 u32 index = key ? *(u32 *)key : U32_MAX;
296 u32 *next = (u32 *)next_key;
298 if (index >= array->map.max_entries) {
303 if (index == array->map.max_entries - 1)
310 static void check_and_free_fields(struct bpf_array *arr, void *val)
312 if (map_value_has_timer(&arr->map))
313 bpf_timer_cancel_and_free(val + arr->map.timer_off);
314 if (map_value_has_kptrs(&arr->map))
315 bpf_map_free_kptrs(&arr->map, val);
318 /* Called from syscall or from eBPF program */
319 static int array_map_update_elem(struct bpf_map *map, void *key, void *value,
322 struct bpf_array *array = container_of(map, struct bpf_array, map);
323 u32 index = *(u32 *)key;
326 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
330 if (unlikely(index >= array->map.max_entries))
331 /* all elements were pre-allocated, cannot insert a new one */
334 if (unlikely(map_flags & BPF_NOEXIST))
335 /* all elements already exist */
338 if (unlikely((map_flags & BPF_F_LOCK) &&
339 !map_value_has_spin_lock(map)))
342 if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
343 memcpy(this_cpu_ptr(array->pptrs[index & array->index_mask]),
344 value, map->value_size);
347 (u64)array->elem_size * (index & array->index_mask);
348 if (map_flags & BPF_F_LOCK)
349 copy_map_value_locked(map, val, value, false);
351 copy_map_value(map, val, value);
352 check_and_free_fields(array, val);
357 int bpf_percpu_array_update(struct bpf_map *map, void *key, void *value,
360 struct bpf_array *array = container_of(map, struct bpf_array, map);
361 u32 index = *(u32 *)key;
366 if (unlikely(map_flags > BPF_EXIST))
370 if (unlikely(index >= array->map.max_entries))
371 /* all elements were pre-allocated, cannot insert a new one */
374 if (unlikely(map_flags == BPF_NOEXIST))
375 /* all elements already exist */
378 /* the user space will provide round_up(value_size, 8) bytes that
379 * will be copied into per-cpu area. bpf programs can only access
380 * value_size of it. During lookup the same extra bytes will be
381 * returned or zeros which were zero-filled by percpu_alloc,
382 * so no kernel data leaks possible
384 size = array->elem_size;
386 pptr = array->pptrs[index & array->index_mask];
387 for_each_possible_cpu(cpu) {
388 bpf_long_memcpy(per_cpu_ptr(pptr, cpu), value + off, size);
395 /* Called from syscall or from eBPF program */
396 static int array_map_delete_elem(struct bpf_map *map, void *key)
401 static void *array_map_vmalloc_addr(struct bpf_array *array)
403 return (void *)round_down((unsigned long)array, PAGE_SIZE);
406 static void array_map_free_timers(struct bpf_map *map)
408 struct bpf_array *array = container_of(map, struct bpf_array, map);
411 /* We don't reset or free kptr on uref dropping to zero. */
412 if (!map_value_has_timer(map))
415 for (i = 0; i < array->map.max_entries; i++)
416 bpf_timer_cancel_and_free(array_map_elem_ptr(array, i) + map->timer_off);
419 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
420 static void array_map_free(struct bpf_map *map)
422 struct bpf_array *array = container_of(map, struct bpf_array, map);
425 if (map_value_has_kptrs(map)) {
426 for (i = 0; i < array->map.max_entries; i++)
427 bpf_map_free_kptrs(map, array_map_elem_ptr(array, i));
428 bpf_map_free_kptr_off_tab(map);
431 if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY)
432 bpf_array_free_percpu(array);
434 if (array->map.map_flags & BPF_F_MMAPABLE)
435 bpf_map_area_free(array_map_vmalloc_addr(array));
437 bpf_map_area_free(array);
440 static void array_map_seq_show_elem(struct bpf_map *map, void *key,
447 value = array_map_lookup_elem(map, key);
453 if (map->btf_key_type_id)
454 seq_printf(m, "%u: ", *(u32 *)key);
455 btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
461 static void percpu_array_map_seq_show_elem(struct bpf_map *map, void *key,
464 struct bpf_array *array = container_of(map, struct bpf_array, map);
465 u32 index = *(u32 *)key;
471 seq_printf(m, "%u: {\n", *(u32 *)key);
472 pptr = array->pptrs[index & array->index_mask];
473 for_each_possible_cpu(cpu) {
474 seq_printf(m, "\tcpu%d: ", cpu);
475 btf_type_seq_show(map->btf, map->btf_value_type_id,
476 per_cpu_ptr(pptr, cpu), m);
484 static int array_map_check_btf(const struct bpf_map *map,
485 const struct btf *btf,
486 const struct btf_type *key_type,
487 const struct btf_type *value_type)
491 /* One exception for keyless BTF: .bss/.data/.rodata map */
492 if (btf_type_is_void(key_type)) {
493 if (map->map_type != BPF_MAP_TYPE_ARRAY ||
494 map->max_entries != 1)
497 if (BTF_INFO_KIND(value_type->info) != BTF_KIND_DATASEC)
503 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
506 int_data = *(u32 *)(key_type + 1);
507 /* bpf array can only take a u32 key. This check makes sure
508 * that the btf matches the attr used during map_create.
510 if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
516 static int array_map_mmap(struct bpf_map *map, struct vm_area_struct *vma)
518 struct bpf_array *array = container_of(map, struct bpf_array, map);
519 pgoff_t pgoff = PAGE_ALIGN(sizeof(*array)) >> PAGE_SHIFT;
521 if (!(map->map_flags & BPF_F_MMAPABLE))
524 if (vma->vm_pgoff * PAGE_SIZE + (vma->vm_end - vma->vm_start) >
525 PAGE_ALIGN((u64)array->map.max_entries * array->elem_size))
528 return remap_vmalloc_range(vma, array_map_vmalloc_addr(array),
529 vma->vm_pgoff + pgoff);
532 static bool array_map_meta_equal(const struct bpf_map *meta0,
533 const struct bpf_map *meta1)
535 if (!bpf_map_meta_equal(meta0, meta1))
537 return meta0->map_flags & BPF_F_INNER_MAP ? true :
538 meta0->max_entries == meta1->max_entries;
541 struct bpf_iter_seq_array_map_info {
543 void *percpu_value_buf;
547 static void *bpf_array_map_seq_start(struct seq_file *seq, loff_t *pos)
549 struct bpf_iter_seq_array_map_info *info = seq->private;
550 struct bpf_map *map = info->map;
551 struct bpf_array *array;
554 if (info->index >= map->max_entries)
559 array = container_of(map, struct bpf_array, map);
560 index = info->index & array->index_mask;
561 if (info->percpu_value_buf)
562 return array->pptrs[index];
563 return array_map_elem_ptr(array, index);
566 static void *bpf_array_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
568 struct bpf_iter_seq_array_map_info *info = seq->private;
569 struct bpf_map *map = info->map;
570 struct bpf_array *array;
575 if (info->index >= map->max_entries)
578 array = container_of(map, struct bpf_array, map);
579 index = info->index & array->index_mask;
580 if (info->percpu_value_buf)
581 return array->pptrs[index];
582 return array_map_elem_ptr(array, index);
585 static int __bpf_array_map_seq_show(struct seq_file *seq, void *v)
587 struct bpf_iter_seq_array_map_info *info = seq->private;
588 struct bpf_iter__bpf_map_elem ctx = {};
589 struct bpf_map *map = info->map;
590 struct bpf_array *array = container_of(map, struct bpf_array, map);
591 struct bpf_iter_meta meta;
592 struct bpf_prog *prog;
593 int off = 0, cpu = 0;
594 void __percpu **pptr;
598 prog = bpf_iter_get_info(&meta, v == NULL);
605 ctx.key = &info->index;
607 if (!info->percpu_value_buf) {
611 size = array->elem_size;
612 for_each_possible_cpu(cpu) {
613 bpf_long_memcpy(info->percpu_value_buf + off,
614 per_cpu_ptr(pptr, cpu),
618 ctx.value = info->percpu_value_buf;
622 return bpf_iter_run_prog(prog, &ctx);
625 static int bpf_array_map_seq_show(struct seq_file *seq, void *v)
627 return __bpf_array_map_seq_show(seq, v);
630 static void bpf_array_map_seq_stop(struct seq_file *seq, void *v)
633 (void)__bpf_array_map_seq_show(seq, NULL);
636 static int bpf_iter_init_array_map(void *priv_data,
637 struct bpf_iter_aux_info *aux)
639 struct bpf_iter_seq_array_map_info *seq_info = priv_data;
640 struct bpf_map *map = aux->map;
641 struct bpf_array *array = container_of(map, struct bpf_array, map);
645 if (map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
646 buf_size = array->elem_size * num_possible_cpus();
647 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
651 seq_info->percpu_value_buf = value_buf;
658 static void bpf_iter_fini_array_map(void *priv_data)
660 struct bpf_iter_seq_array_map_info *seq_info = priv_data;
662 kfree(seq_info->percpu_value_buf);
665 static const struct seq_operations bpf_array_map_seq_ops = {
666 .start = bpf_array_map_seq_start,
667 .next = bpf_array_map_seq_next,
668 .stop = bpf_array_map_seq_stop,
669 .show = bpf_array_map_seq_show,
672 static const struct bpf_iter_seq_info iter_seq_info = {
673 .seq_ops = &bpf_array_map_seq_ops,
674 .init_seq_private = bpf_iter_init_array_map,
675 .fini_seq_private = bpf_iter_fini_array_map,
676 .seq_priv_size = sizeof(struct bpf_iter_seq_array_map_info),
679 static int bpf_for_each_array_elem(struct bpf_map *map, bpf_callback_t callback_fn,
680 void *callback_ctx, u64 flags)
682 u32 i, key, num_elems = 0;
683 struct bpf_array *array;
691 is_percpu = map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
692 array = container_of(map, struct bpf_array, map);
695 for (i = 0; i < map->max_entries; i++) {
697 val = this_cpu_ptr(array->pptrs[i]);
699 val = array_map_elem_ptr(array, i);
702 ret = callback_fn((u64)(long)map, (u64)(long)&key,
703 (u64)(long)val, (u64)(long)callback_ctx, 0);
704 /* return value: 0 - continue, 1 - stop and return */
714 BTF_ID_LIST_SINGLE(array_map_btf_ids, struct, bpf_array)
715 const struct bpf_map_ops array_map_ops = {
716 .map_meta_equal = array_map_meta_equal,
717 .map_alloc_check = array_map_alloc_check,
718 .map_alloc = array_map_alloc,
719 .map_free = array_map_free,
720 .map_get_next_key = array_map_get_next_key,
721 .map_release_uref = array_map_free_timers,
722 .map_lookup_elem = array_map_lookup_elem,
723 .map_update_elem = array_map_update_elem,
724 .map_delete_elem = array_map_delete_elem,
725 .map_gen_lookup = array_map_gen_lookup,
726 .map_direct_value_addr = array_map_direct_value_addr,
727 .map_direct_value_meta = array_map_direct_value_meta,
728 .map_mmap = array_map_mmap,
729 .map_seq_show_elem = array_map_seq_show_elem,
730 .map_check_btf = array_map_check_btf,
731 .map_lookup_batch = generic_map_lookup_batch,
732 .map_update_batch = generic_map_update_batch,
733 .map_set_for_each_callback_args = map_set_for_each_callback_args,
734 .map_for_each_callback = bpf_for_each_array_elem,
735 .map_btf_id = &array_map_btf_ids[0],
736 .iter_seq_info = &iter_seq_info,
739 const struct bpf_map_ops percpu_array_map_ops = {
740 .map_meta_equal = bpf_map_meta_equal,
741 .map_alloc_check = array_map_alloc_check,
742 .map_alloc = array_map_alloc,
743 .map_free = array_map_free,
744 .map_get_next_key = array_map_get_next_key,
745 .map_lookup_elem = percpu_array_map_lookup_elem,
746 .map_update_elem = array_map_update_elem,
747 .map_delete_elem = array_map_delete_elem,
748 .map_lookup_percpu_elem = percpu_array_map_lookup_percpu_elem,
749 .map_seq_show_elem = percpu_array_map_seq_show_elem,
750 .map_check_btf = array_map_check_btf,
751 .map_lookup_batch = generic_map_lookup_batch,
752 .map_update_batch = generic_map_update_batch,
753 .map_set_for_each_callback_args = map_set_for_each_callback_args,
754 .map_for_each_callback = bpf_for_each_array_elem,
755 .map_btf_id = &array_map_btf_ids[0],
756 .iter_seq_info = &iter_seq_info,
759 static int fd_array_map_alloc_check(union bpf_attr *attr)
761 /* only file descriptors can be stored in this type of map */
762 if (attr->value_size != sizeof(u32))
764 /* Program read-only/write-only not supported for special maps yet. */
765 if (attr->map_flags & (BPF_F_RDONLY_PROG | BPF_F_WRONLY_PROG))
767 return array_map_alloc_check(attr);
770 static void fd_array_map_free(struct bpf_map *map)
772 struct bpf_array *array = container_of(map, struct bpf_array, map);
775 /* make sure it's empty */
776 for (i = 0; i < array->map.max_entries; i++)
777 BUG_ON(array->ptrs[i] != NULL);
779 bpf_map_area_free(array);
782 static void *fd_array_map_lookup_elem(struct bpf_map *map, void *key)
784 return ERR_PTR(-EOPNOTSUPP);
787 /* only called from syscall */
788 int bpf_fd_array_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
793 if (!map->ops->map_fd_sys_lookup_elem)
797 elem = array_map_lookup_elem(map, key);
798 if (elem && (ptr = READ_ONCE(*elem)))
799 *value = map->ops->map_fd_sys_lookup_elem(ptr);
807 /* only called from syscall */
808 int bpf_fd_array_map_update_elem(struct bpf_map *map, struct file *map_file,
809 void *key, void *value, u64 map_flags)
811 struct bpf_array *array = container_of(map, struct bpf_array, map);
812 void *new_ptr, *old_ptr;
813 u32 index = *(u32 *)key, ufd;
815 if (map_flags != BPF_ANY)
818 if (index >= array->map.max_entries)
822 new_ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
824 return PTR_ERR(new_ptr);
826 if (map->ops->map_poke_run) {
827 mutex_lock(&array->aux->poke_mutex);
828 old_ptr = xchg(array->ptrs + index, new_ptr);
829 map->ops->map_poke_run(map, index, old_ptr, new_ptr);
830 mutex_unlock(&array->aux->poke_mutex);
832 old_ptr = xchg(array->ptrs + index, new_ptr);
836 map->ops->map_fd_put_ptr(old_ptr);
840 static int fd_array_map_delete_elem(struct bpf_map *map, void *key)
842 struct bpf_array *array = container_of(map, struct bpf_array, map);
844 u32 index = *(u32 *)key;
846 if (index >= array->map.max_entries)
849 if (map->ops->map_poke_run) {
850 mutex_lock(&array->aux->poke_mutex);
851 old_ptr = xchg(array->ptrs + index, NULL);
852 map->ops->map_poke_run(map, index, old_ptr, NULL);
853 mutex_unlock(&array->aux->poke_mutex);
855 old_ptr = xchg(array->ptrs + index, NULL);
859 map->ops->map_fd_put_ptr(old_ptr);
866 static void *prog_fd_array_get_ptr(struct bpf_map *map,
867 struct file *map_file, int fd)
869 struct bpf_prog *prog = bpf_prog_get(fd);
874 if (!bpf_prog_map_compatible(map, prog)) {
876 return ERR_PTR(-EINVAL);
882 static void prog_fd_array_put_ptr(void *ptr)
887 static u32 prog_fd_array_sys_lookup_elem(void *ptr)
889 return ((struct bpf_prog *)ptr)->aux->id;
892 /* decrement refcnt of all bpf_progs that are stored in this map */
893 static void bpf_fd_array_map_clear(struct bpf_map *map)
895 struct bpf_array *array = container_of(map, struct bpf_array, map);
898 for (i = 0; i < array->map.max_entries; i++)
899 fd_array_map_delete_elem(map, &i);
902 static void prog_array_map_seq_show_elem(struct bpf_map *map, void *key,
910 elem = array_map_lookup_elem(map, key);
912 ptr = READ_ONCE(*elem);
914 seq_printf(m, "%u: ", *(u32 *)key);
915 prog_id = prog_fd_array_sys_lookup_elem(ptr);
916 btf_type_seq_show(map->btf, map->btf_value_type_id,
925 struct prog_poke_elem {
926 struct list_head list;
927 struct bpf_prog_aux *aux;
930 static int prog_array_map_poke_track(struct bpf_map *map,
931 struct bpf_prog_aux *prog_aux)
933 struct prog_poke_elem *elem;
934 struct bpf_array_aux *aux;
937 aux = container_of(map, struct bpf_array, map)->aux;
938 mutex_lock(&aux->poke_mutex);
939 list_for_each_entry(elem, &aux->poke_progs, list) {
940 if (elem->aux == prog_aux)
944 elem = kmalloc(sizeof(*elem), GFP_KERNEL);
950 INIT_LIST_HEAD(&elem->list);
951 /* We must track the program's aux info at this point in time
952 * since the program pointer itself may not be stable yet, see
953 * also comment in prog_array_map_poke_run().
955 elem->aux = prog_aux;
957 list_add_tail(&elem->list, &aux->poke_progs);
959 mutex_unlock(&aux->poke_mutex);
963 static void prog_array_map_poke_untrack(struct bpf_map *map,
964 struct bpf_prog_aux *prog_aux)
966 struct prog_poke_elem *elem, *tmp;
967 struct bpf_array_aux *aux;
969 aux = container_of(map, struct bpf_array, map)->aux;
970 mutex_lock(&aux->poke_mutex);
971 list_for_each_entry_safe(elem, tmp, &aux->poke_progs, list) {
972 if (elem->aux == prog_aux) {
973 list_del_init(&elem->list);
978 mutex_unlock(&aux->poke_mutex);
981 static void prog_array_map_poke_run(struct bpf_map *map, u32 key,
982 struct bpf_prog *old,
983 struct bpf_prog *new)
985 u8 *old_addr, *new_addr, *old_bypass_addr;
986 struct prog_poke_elem *elem;
987 struct bpf_array_aux *aux;
989 aux = container_of(map, struct bpf_array, map)->aux;
990 WARN_ON_ONCE(!mutex_is_locked(&aux->poke_mutex));
992 list_for_each_entry(elem, &aux->poke_progs, list) {
993 struct bpf_jit_poke_descriptor *poke;
996 for (i = 0; i < elem->aux->size_poke_tab; i++) {
997 poke = &elem->aux->poke_tab[i];
999 /* Few things to be aware of:
1001 * 1) We can only ever access aux in this context, but
1002 * not aux->prog since it might not be stable yet and
1003 * there could be danger of use after free otherwise.
1004 * 2) Initially when we start tracking aux, the program
1005 * is not JITed yet and also does not have a kallsyms
1006 * entry. We skip these as poke->tailcall_target_stable
1007 * is not active yet. The JIT will do the final fixup
1008 * before setting it stable. The various
1009 * poke->tailcall_target_stable are successively
1010 * activated, so tail call updates can arrive from here
1011 * while JIT is still finishing its final fixup for
1012 * non-activated poke entries.
1013 * 3) On program teardown, the program's kallsym entry gets
1014 * removed out of RCU callback, but we can only untrack
1015 * from sleepable context, therefore bpf_arch_text_poke()
1016 * might not see that this is in BPF text section and
1017 * bails out with -EINVAL. As these are unreachable since
1018 * RCU grace period already passed, we simply skip them.
1019 * 4) Also programs reaching refcount of zero while patching
1020 * is in progress is okay since we're protected under
1021 * poke_mutex and untrack the programs before the JIT
1022 * buffer is freed. When we're still in the middle of
1023 * patching and suddenly kallsyms entry of the program
1024 * gets evicted, we just skip the rest which is fine due
1026 * 5) Any other error happening below from bpf_arch_text_poke()
1027 * is a unexpected bug.
1029 if (!READ_ONCE(poke->tailcall_target_stable))
1031 if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
1033 if (poke->tail_call.map != map ||
1034 poke->tail_call.key != key)
1037 old_bypass_addr = old ? NULL : poke->bypass_addr;
1038 old_addr = old ? (u8 *)old->bpf_func + poke->adj_off : NULL;
1039 new_addr = new ? (u8 *)new->bpf_func + poke->adj_off : NULL;
1042 ret = bpf_arch_text_poke(poke->tailcall_target,
1044 old_addr, new_addr);
1045 BUG_ON(ret < 0 && ret != -EINVAL);
1047 ret = bpf_arch_text_poke(poke->tailcall_bypass,
1051 BUG_ON(ret < 0 && ret != -EINVAL);
1054 ret = bpf_arch_text_poke(poke->tailcall_bypass,
1058 BUG_ON(ret < 0 && ret != -EINVAL);
1059 /* let other CPUs finish the execution of program
1060 * so that it will not possible to expose them
1061 * to invalid nop, stack unwind, nop state
1065 ret = bpf_arch_text_poke(poke->tailcall_target,
1068 BUG_ON(ret < 0 && ret != -EINVAL);
1074 static void prog_array_map_clear_deferred(struct work_struct *work)
1076 struct bpf_map *map = container_of(work, struct bpf_array_aux,
1078 bpf_fd_array_map_clear(map);
1082 static void prog_array_map_clear(struct bpf_map *map)
1084 struct bpf_array_aux *aux = container_of(map, struct bpf_array,
1087 schedule_work(&aux->work);
1090 static struct bpf_map *prog_array_map_alloc(union bpf_attr *attr)
1092 struct bpf_array_aux *aux;
1093 struct bpf_map *map;
1095 aux = kzalloc(sizeof(*aux), GFP_KERNEL_ACCOUNT);
1097 return ERR_PTR(-ENOMEM);
1099 INIT_WORK(&aux->work, prog_array_map_clear_deferred);
1100 INIT_LIST_HEAD(&aux->poke_progs);
1101 mutex_init(&aux->poke_mutex);
1103 map = array_map_alloc(attr);
1109 container_of(map, struct bpf_array, map)->aux = aux;
1115 static void prog_array_map_free(struct bpf_map *map)
1117 struct prog_poke_elem *elem, *tmp;
1118 struct bpf_array_aux *aux;
1120 aux = container_of(map, struct bpf_array, map)->aux;
1121 list_for_each_entry_safe(elem, tmp, &aux->poke_progs, list) {
1122 list_del_init(&elem->list);
1126 fd_array_map_free(map);
1129 /* prog_array->aux->{type,jited} is a runtime binding.
1130 * Doing static check alone in the verifier is not enough.
1131 * Thus, prog_array_map cannot be used as an inner_map
1132 * and map_meta_equal is not implemented.
1134 const struct bpf_map_ops prog_array_map_ops = {
1135 .map_alloc_check = fd_array_map_alloc_check,
1136 .map_alloc = prog_array_map_alloc,
1137 .map_free = prog_array_map_free,
1138 .map_poke_track = prog_array_map_poke_track,
1139 .map_poke_untrack = prog_array_map_poke_untrack,
1140 .map_poke_run = prog_array_map_poke_run,
1141 .map_get_next_key = array_map_get_next_key,
1142 .map_lookup_elem = fd_array_map_lookup_elem,
1143 .map_delete_elem = fd_array_map_delete_elem,
1144 .map_fd_get_ptr = prog_fd_array_get_ptr,
1145 .map_fd_put_ptr = prog_fd_array_put_ptr,
1146 .map_fd_sys_lookup_elem = prog_fd_array_sys_lookup_elem,
1147 .map_release_uref = prog_array_map_clear,
1148 .map_seq_show_elem = prog_array_map_seq_show_elem,
1149 .map_btf_id = &array_map_btf_ids[0],
1152 static struct bpf_event_entry *bpf_event_entry_gen(struct file *perf_file,
1153 struct file *map_file)
1155 struct bpf_event_entry *ee;
1157 ee = kzalloc(sizeof(*ee), GFP_ATOMIC);
1159 ee->event = perf_file->private_data;
1160 ee->perf_file = perf_file;
1161 ee->map_file = map_file;
1167 static void __bpf_event_entry_free(struct rcu_head *rcu)
1169 struct bpf_event_entry *ee;
1171 ee = container_of(rcu, struct bpf_event_entry, rcu);
1172 fput(ee->perf_file);
1176 static void bpf_event_entry_free_rcu(struct bpf_event_entry *ee)
1178 call_rcu(&ee->rcu, __bpf_event_entry_free);
1181 static void *perf_event_fd_array_get_ptr(struct bpf_map *map,
1182 struct file *map_file, int fd)
1184 struct bpf_event_entry *ee;
1185 struct perf_event *event;
1186 struct file *perf_file;
1189 perf_file = perf_event_get(fd);
1190 if (IS_ERR(perf_file))
1193 ee = ERR_PTR(-EOPNOTSUPP);
1194 event = perf_file->private_data;
1195 if (perf_event_read_local(event, &value, NULL, NULL) == -EOPNOTSUPP)
1198 ee = bpf_event_entry_gen(perf_file, map_file);
1201 ee = ERR_PTR(-ENOMEM);
1207 static void perf_event_fd_array_put_ptr(void *ptr)
1209 bpf_event_entry_free_rcu(ptr);
1212 static void perf_event_fd_array_release(struct bpf_map *map,
1213 struct file *map_file)
1215 struct bpf_array *array = container_of(map, struct bpf_array, map);
1216 struct bpf_event_entry *ee;
1219 if (map->map_flags & BPF_F_PRESERVE_ELEMS)
1223 for (i = 0; i < array->map.max_entries; i++) {
1224 ee = READ_ONCE(array->ptrs[i]);
1225 if (ee && ee->map_file == map_file)
1226 fd_array_map_delete_elem(map, &i);
1231 static void perf_event_fd_array_map_free(struct bpf_map *map)
1233 if (map->map_flags & BPF_F_PRESERVE_ELEMS)
1234 bpf_fd_array_map_clear(map);
1235 fd_array_map_free(map);
1238 const struct bpf_map_ops perf_event_array_map_ops = {
1239 .map_meta_equal = bpf_map_meta_equal,
1240 .map_alloc_check = fd_array_map_alloc_check,
1241 .map_alloc = array_map_alloc,
1242 .map_free = perf_event_fd_array_map_free,
1243 .map_get_next_key = array_map_get_next_key,
1244 .map_lookup_elem = fd_array_map_lookup_elem,
1245 .map_delete_elem = fd_array_map_delete_elem,
1246 .map_fd_get_ptr = perf_event_fd_array_get_ptr,
1247 .map_fd_put_ptr = perf_event_fd_array_put_ptr,
1248 .map_release = perf_event_fd_array_release,
1249 .map_check_btf = map_check_no_btf,
1250 .map_btf_id = &array_map_btf_ids[0],
1253 #ifdef CONFIG_CGROUPS
1254 static void *cgroup_fd_array_get_ptr(struct bpf_map *map,
1255 struct file *map_file /* not used */,
1258 return cgroup_get_from_fd(fd);
1261 static void cgroup_fd_array_put_ptr(void *ptr)
1263 /* cgroup_put free cgrp after a rcu grace period */
1267 static void cgroup_fd_array_free(struct bpf_map *map)
1269 bpf_fd_array_map_clear(map);
1270 fd_array_map_free(map);
1273 const struct bpf_map_ops cgroup_array_map_ops = {
1274 .map_meta_equal = bpf_map_meta_equal,
1275 .map_alloc_check = fd_array_map_alloc_check,
1276 .map_alloc = array_map_alloc,
1277 .map_free = cgroup_fd_array_free,
1278 .map_get_next_key = array_map_get_next_key,
1279 .map_lookup_elem = fd_array_map_lookup_elem,
1280 .map_delete_elem = fd_array_map_delete_elem,
1281 .map_fd_get_ptr = cgroup_fd_array_get_ptr,
1282 .map_fd_put_ptr = cgroup_fd_array_put_ptr,
1283 .map_check_btf = map_check_no_btf,
1284 .map_btf_id = &array_map_btf_ids[0],
1288 static struct bpf_map *array_of_map_alloc(union bpf_attr *attr)
1290 struct bpf_map *map, *inner_map_meta;
1292 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1293 if (IS_ERR(inner_map_meta))
1294 return inner_map_meta;
1296 map = array_map_alloc(attr);
1298 bpf_map_meta_free(inner_map_meta);
1302 map->inner_map_meta = inner_map_meta;
1307 static void array_of_map_free(struct bpf_map *map)
1309 /* map->inner_map_meta is only accessed by syscall which
1310 * is protected by fdget/fdput.
1312 bpf_map_meta_free(map->inner_map_meta);
1313 bpf_fd_array_map_clear(map);
1314 fd_array_map_free(map);
1317 static void *array_of_map_lookup_elem(struct bpf_map *map, void *key)
1319 struct bpf_map **inner_map = array_map_lookup_elem(map, key);
1324 return READ_ONCE(*inner_map);
1327 static int array_of_map_gen_lookup(struct bpf_map *map,
1328 struct bpf_insn *insn_buf)
1330 struct bpf_array *array = container_of(map, struct bpf_array, map);
1331 u32 elem_size = array->elem_size;
1332 struct bpf_insn *insn = insn_buf;
1333 const int ret = BPF_REG_0;
1334 const int map_ptr = BPF_REG_1;
1335 const int index = BPF_REG_2;
1337 *insn++ = BPF_ALU64_IMM(BPF_ADD, map_ptr, offsetof(struct bpf_array, value));
1338 *insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
1339 if (!map->bypass_spec_v1) {
1340 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 6);
1341 *insn++ = BPF_ALU32_IMM(BPF_AND, ret, array->index_mask);
1343 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 5);
1345 if (is_power_of_2(elem_size))
1346 *insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(elem_size));
1348 *insn++ = BPF_ALU64_IMM(BPF_MUL, ret, elem_size);
1349 *insn++ = BPF_ALU64_REG(BPF_ADD, ret, map_ptr);
1350 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1351 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
1352 *insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
1353 *insn++ = BPF_MOV64_IMM(ret, 0);
1355 return insn - insn_buf;
1358 const struct bpf_map_ops array_of_maps_map_ops = {
1359 .map_alloc_check = fd_array_map_alloc_check,
1360 .map_alloc = array_of_map_alloc,
1361 .map_free = array_of_map_free,
1362 .map_get_next_key = array_map_get_next_key,
1363 .map_lookup_elem = array_of_map_lookup_elem,
1364 .map_delete_elem = fd_array_map_delete_elem,
1365 .map_fd_get_ptr = bpf_map_fd_get_ptr,
1366 .map_fd_put_ptr = bpf_map_fd_put_ptr,
1367 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1368 .map_gen_lookup = array_of_map_gen_lookup,
1369 .map_lookup_batch = generic_map_lookup_batch,
1370 .map_update_batch = generic_map_update_batch,
1371 .map_check_btf = map_check_no_btf,
1372 .map_btf_id = &array_map_btf_ids[0],