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 /* avoid overflow on round_up(map->value_size) */
74 if (attr->value_size > INT_MAX)
80 static struct bpf_map *array_map_alloc(union bpf_attr *attr)
82 bool percpu = attr->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
83 int numa_node = bpf_map_attr_numa_node(attr);
84 u32 elem_size, index_mask, max_entries;
85 bool bypass_spec_v1 = bpf_bypass_spec_v1();
86 u64 array_size, mask64;
87 struct bpf_array *array;
89 elem_size = round_up(attr->value_size, 8);
91 max_entries = attr->max_entries;
93 /* On 32 bit archs roundup_pow_of_two() with max_entries that has
94 * upper most bit set in u32 space is undefined behavior due to
95 * resulting 1U << 32, so do it manually here in u64 space.
97 mask64 = fls_long(max_entries - 1);
98 mask64 = 1ULL << mask64;
102 if (!bypass_spec_v1) {
103 /* round up array size to nearest power of 2,
104 * since cpu will speculate within index_mask limits
106 max_entries = index_mask + 1;
107 /* Check for overflows. */
108 if (max_entries < attr->max_entries)
109 return ERR_PTR(-E2BIG);
112 array_size = sizeof(*array);
114 array_size += (u64) max_entries * sizeof(void *);
116 /* rely on vmalloc() to return page-aligned memory and
117 * ensure array->value is exactly page-aligned
119 if (attr->map_flags & BPF_F_MMAPABLE) {
120 array_size = PAGE_ALIGN(array_size);
121 array_size += PAGE_ALIGN((u64) max_entries * elem_size);
123 array_size += (u64) max_entries * elem_size;
127 /* allocate all map elements and zero-initialize them */
128 if (attr->map_flags & BPF_F_MMAPABLE) {
131 /* kmalloc'ed memory can't be mmap'ed, use explicit vmalloc */
132 data = bpf_map_area_mmapable_alloc(array_size, numa_node);
134 return ERR_PTR(-ENOMEM);
135 array = data + PAGE_ALIGN(sizeof(struct bpf_array))
136 - offsetof(struct bpf_array, value);
138 array = bpf_map_area_alloc(array_size, numa_node);
141 return ERR_PTR(-ENOMEM);
142 array->index_mask = index_mask;
143 array->map.bypass_spec_v1 = bypass_spec_v1;
145 /* copy mandatory map attributes */
146 bpf_map_init_from_attr(&array->map, attr);
147 array->elem_size = elem_size;
149 if (percpu && bpf_array_alloc_percpu(array)) {
150 bpf_map_area_free(array);
151 return ERR_PTR(-ENOMEM);
157 static void *array_map_elem_ptr(struct bpf_array* array, u32 index)
159 return array->value + (u64)array->elem_size * index;
162 /* Called from syscall or from eBPF program */
163 static void *array_map_lookup_elem(struct bpf_map *map, void *key)
165 struct bpf_array *array = container_of(map, struct bpf_array, map);
166 u32 index = *(u32 *)key;
168 if (unlikely(index >= array->map.max_entries))
171 return array->value + (u64)array->elem_size * (index & array->index_mask);
174 static int array_map_direct_value_addr(const struct bpf_map *map, u64 *imm,
177 struct bpf_array *array = container_of(map, struct bpf_array, map);
179 if (map->max_entries != 1)
181 if (off >= map->value_size)
184 *imm = (unsigned long)array->value;
188 static int array_map_direct_value_meta(const struct bpf_map *map, u64 imm,
191 struct bpf_array *array = container_of(map, struct bpf_array, map);
192 u64 base = (unsigned long)array->value;
193 u64 range = array->elem_size;
195 if (map->max_entries != 1)
197 if (imm < base || imm >= base + range)
204 /* emit BPF instructions equivalent to C code of array_map_lookup_elem() */
205 static int array_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
207 struct bpf_array *array = container_of(map, struct bpf_array, map);
208 struct bpf_insn *insn = insn_buf;
209 u32 elem_size = array->elem_size;
210 const int ret = BPF_REG_0;
211 const int map_ptr = BPF_REG_1;
212 const int index = BPF_REG_2;
214 if (map->map_flags & BPF_F_INNER_MAP)
217 *insn++ = BPF_ALU64_IMM(BPF_ADD, map_ptr, offsetof(struct bpf_array, value));
218 *insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
219 if (!map->bypass_spec_v1) {
220 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 4);
221 *insn++ = BPF_ALU32_IMM(BPF_AND, ret, array->index_mask);
223 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 3);
226 if (is_power_of_2(elem_size)) {
227 *insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(elem_size));
229 *insn++ = BPF_ALU64_IMM(BPF_MUL, ret, elem_size);
231 *insn++ = BPF_ALU64_REG(BPF_ADD, ret, map_ptr);
232 *insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
233 *insn++ = BPF_MOV64_IMM(ret, 0);
234 return insn - insn_buf;
237 /* Called from eBPF program */
238 static void *percpu_array_map_lookup_elem(struct bpf_map *map, void *key)
240 struct bpf_array *array = container_of(map, struct bpf_array, map);
241 u32 index = *(u32 *)key;
243 if (unlikely(index >= array->map.max_entries))
246 return this_cpu_ptr(array->pptrs[index & array->index_mask]);
249 static void *percpu_array_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
251 struct bpf_array *array = container_of(map, struct bpf_array, map);
252 u32 index = *(u32 *)key;
254 if (cpu >= nr_cpu_ids)
257 if (unlikely(index >= array->map.max_entries))
260 return per_cpu_ptr(array->pptrs[index & array->index_mask], cpu);
263 int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value)
265 struct bpf_array *array = container_of(map, struct bpf_array, map);
266 u32 index = *(u32 *)key;
271 if (unlikely(index >= array->map.max_entries))
274 /* per_cpu areas are zero-filled and bpf programs can only
275 * access 'value_size' of them, so copying rounded areas
276 * will not leak any kernel data
278 size = array->elem_size;
280 pptr = array->pptrs[index & array->index_mask];
281 for_each_possible_cpu(cpu) {
282 copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));
283 check_and_init_map_value(map, value + off);
290 /* Called from syscall */
291 static int array_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
293 struct bpf_array *array = container_of(map, struct bpf_array, map);
294 u32 index = key ? *(u32 *)key : U32_MAX;
295 u32 *next = (u32 *)next_key;
297 if (index >= array->map.max_entries) {
302 if (index == array->map.max_entries - 1)
309 static void check_and_free_fields(struct bpf_array *arr, void *val)
311 if (map_value_has_timer(&arr->map))
312 bpf_timer_cancel_and_free(val + arr->map.timer_off);
313 if (map_value_has_kptrs(&arr->map))
314 bpf_map_free_kptrs(&arr->map, val);
317 /* Called from syscall or from eBPF program */
318 static int array_map_update_elem(struct bpf_map *map, void *key, void *value,
321 struct bpf_array *array = container_of(map, struct bpf_array, map);
322 u32 index = *(u32 *)key;
325 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
329 if (unlikely(index >= array->map.max_entries))
330 /* all elements were pre-allocated, cannot insert a new one */
333 if (unlikely(map_flags & BPF_NOEXIST))
334 /* all elements already exist */
337 if (unlikely((map_flags & BPF_F_LOCK) &&
338 !map_value_has_spin_lock(map)))
341 if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
342 val = this_cpu_ptr(array->pptrs[index & array->index_mask]);
343 copy_map_value(map, val, value);
344 check_and_free_fields(array, val);
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 copy_map_value_long(map, per_cpu_ptr(pptr, cpu), value + off);
389 check_and_free_fields(array, per_cpu_ptr(pptr, cpu));
396 /* Called from syscall or from eBPF program */
397 static int array_map_delete_elem(struct bpf_map *map, void *key)
402 static void *array_map_vmalloc_addr(struct bpf_array *array)
404 return (void *)round_down((unsigned long)array, PAGE_SIZE);
407 static void array_map_free_timers(struct bpf_map *map)
409 struct bpf_array *array = container_of(map, struct bpf_array, map);
412 /* We don't reset or free kptr on uref dropping to zero. */
413 if (!map_value_has_timer(map))
416 for (i = 0; i < array->map.max_entries; i++)
417 bpf_timer_cancel_and_free(array_map_elem_ptr(array, i) + map->timer_off);
420 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
421 static void array_map_free(struct bpf_map *map)
423 struct bpf_array *array = container_of(map, struct bpf_array, map);
426 if (map_value_has_kptrs(map)) {
427 if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
428 for (i = 0; i < array->map.max_entries; i++) {
429 void __percpu *pptr = array->pptrs[i & array->index_mask];
432 for_each_possible_cpu(cpu) {
433 bpf_map_free_kptrs(map, per_cpu_ptr(pptr, cpu));
438 for (i = 0; i < array->map.max_entries; i++)
439 bpf_map_free_kptrs(map, array_map_elem_ptr(array, i));
441 bpf_map_free_kptr_off_tab(map);
444 if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY)
445 bpf_array_free_percpu(array);
447 if (array->map.map_flags & BPF_F_MMAPABLE)
448 bpf_map_area_free(array_map_vmalloc_addr(array));
450 bpf_map_area_free(array);
453 static void array_map_seq_show_elem(struct bpf_map *map, void *key,
460 value = array_map_lookup_elem(map, key);
466 if (map->btf_key_type_id)
467 seq_printf(m, "%u: ", *(u32 *)key);
468 btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
474 static void percpu_array_map_seq_show_elem(struct bpf_map *map, void *key,
477 struct bpf_array *array = container_of(map, struct bpf_array, map);
478 u32 index = *(u32 *)key;
484 seq_printf(m, "%u: {\n", *(u32 *)key);
485 pptr = array->pptrs[index & array->index_mask];
486 for_each_possible_cpu(cpu) {
487 seq_printf(m, "\tcpu%d: ", cpu);
488 btf_type_seq_show(map->btf, map->btf_value_type_id,
489 per_cpu_ptr(pptr, cpu), m);
497 static int array_map_check_btf(const struct bpf_map *map,
498 const struct btf *btf,
499 const struct btf_type *key_type,
500 const struct btf_type *value_type)
504 /* One exception for keyless BTF: .bss/.data/.rodata map */
505 if (btf_type_is_void(key_type)) {
506 if (map->map_type != BPF_MAP_TYPE_ARRAY ||
507 map->max_entries != 1)
510 if (BTF_INFO_KIND(value_type->info) != BTF_KIND_DATASEC)
516 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
519 int_data = *(u32 *)(key_type + 1);
520 /* bpf array can only take a u32 key. This check makes sure
521 * that the btf matches the attr used during map_create.
523 if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
529 static int array_map_mmap(struct bpf_map *map, struct vm_area_struct *vma)
531 struct bpf_array *array = container_of(map, struct bpf_array, map);
532 pgoff_t pgoff = PAGE_ALIGN(sizeof(*array)) >> PAGE_SHIFT;
534 if (!(map->map_flags & BPF_F_MMAPABLE))
537 if (vma->vm_pgoff * PAGE_SIZE + (vma->vm_end - vma->vm_start) >
538 PAGE_ALIGN((u64)array->map.max_entries * array->elem_size))
541 return remap_vmalloc_range(vma, array_map_vmalloc_addr(array),
542 vma->vm_pgoff + pgoff);
545 static bool array_map_meta_equal(const struct bpf_map *meta0,
546 const struct bpf_map *meta1)
548 if (!bpf_map_meta_equal(meta0, meta1))
550 return meta0->map_flags & BPF_F_INNER_MAP ? true :
551 meta0->max_entries == meta1->max_entries;
554 struct bpf_iter_seq_array_map_info {
556 void *percpu_value_buf;
560 static void *bpf_array_map_seq_start(struct seq_file *seq, loff_t *pos)
562 struct bpf_iter_seq_array_map_info *info = seq->private;
563 struct bpf_map *map = info->map;
564 struct bpf_array *array;
567 if (info->index >= map->max_entries)
572 array = container_of(map, struct bpf_array, map);
573 index = info->index & array->index_mask;
574 if (info->percpu_value_buf)
575 return array->pptrs[index];
576 return array_map_elem_ptr(array, index);
579 static void *bpf_array_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
581 struct bpf_iter_seq_array_map_info *info = seq->private;
582 struct bpf_map *map = info->map;
583 struct bpf_array *array;
588 if (info->index >= map->max_entries)
591 array = container_of(map, struct bpf_array, map);
592 index = info->index & array->index_mask;
593 if (info->percpu_value_buf)
594 return array->pptrs[index];
595 return array_map_elem_ptr(array, index);
598 static int __bpf_array_map_seq_show(struct seq_file *seq, void *v)
600 struct bpf_iter_seq_array_map_info *info = seq->private;
601 struct bpf_iter__bpf_map_elem ctx = {};
602 struct bpf_map *map = info->map;
603 struct bpf_array *array = container_of(map, struct bpf_array, map);
604 struct bpf_iter_meta meta;
605 struct bpf_prog *prog;
606 int off = 0, cpu = 0;
607 void __percpu **pptr;
611 prog = bpf_iter_get_info(&meta, v == NULL);
618 ctx.key = &info->index;
620 if (!info->percpu_value_buf) {
624 size = array->elem_size;
625 for_each_possible_cpu(cpu) {
626 copy_map_value_long(map, info->percpu_value_buf + off,
627 per_cpu_ptr(pptr, cpu));
628 check_and_init_map_value(map, info->percpu_value_buf + off);
631 ctx.value = info->percpu_value_buf;
635 return bpf_iter_run_prog(prog, &ctx);
638 static int bpf_array_map_seq_show(struct seq_file *seq, void *v)
640 return __bpf_array_map_seq_show(seq, v);
643 static void bpf_array_map_seq_stop(struct seq_file *seq, void *v)
646 (void)__bpf_array_map_seq_show(seq, NULL);
649 static int bpf_iter_init_array_map(void *priv_data,
650 struct bpf_iter_aux_info *aux)
652 struct bpf_iter_seq_array_map_info *seq_info = priv_data;
653 struct bpf_map *map = aux->map;
654 struct bpf_array *array = container_of(map, struct bpf_array, map);
658 if (map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
659 buf_size = array->elem_size * num_possible_cpus();
660 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
664 seq_info->percpu_value_buf = value_buf;
667 /* bpf_iter_attach_map() acquires a map uref, and the uref may be
668 * released before or in the middle of iterating map elements, so
669 * acquire an extra map uref for iterator.
671 bpf_map_inc_with_uref(map);
676 static void bpf_iter_fini_array_map(void *priv_data)
678 struct bpf_iter_seq_array_map_info *seq_info = priv_data;
680 bpf_map_put_with_uref(seq_info->map);
681 kfree(seq_info->percpu_value_buf);
684 static const struct seq_operations bpf_array_map_seq_ops = {
685 .start = bpf_array_map_seq_start,
686 .next = bpf_array_map_seq_next,
687 .stop = bpf_array_map_seq_stop,
688 .show = bpf_array_map_seq_show,
691 static const struct bpf_iter_seq_info iter_seq_info = {
692 .seq_ops = &bpf_array_map_seq_ops,
693 .init_seq_private = bpf_iter_init_array_map,
694 .fini_seq_private = bpf_iter_fini_array_map,
695 .seq_priv_size = sizeof(struct bpf_iter_seq_array_map_info),
698 static int bpf_for_each_array_elem(struct bpf_map *map, bpf_callback_t callback_fn,
699 void *callback_ctx, u64 flags)
701 u32 i, key, num_elems = 0;
702 struct bpf_array *array;
710 is_percpu = map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
711 array = container_of(map, struct bpf_array, map);
714 for (i = 0; i < map->max_entries; i++) {
716 val = this_cpu_ptr(array->pptrs[i]);
718 val = array_map_elem_ptr(array, i);
721 ret = callback_fn((u64)(long)map, (u64)(long)&key,
722 (u64)(long)val, (u64)(long)callback_ctx, 0);
723 /* return value: 0 - continue, 1 - stop and return */
733 BTF_ID_LIST_SINGLE(array_map_btf_ids, struct, bpf_array)
734 const struct bpf_map_ops array_map_ops = {
735 .map_meta_equal = array_map_meta_equal,
736 .map_alloc_check = array_map_alloc_check,
737 .map_alloc = array_map_alloc,
738 .map_free = array_map_free,
739 .map_get_next_key = array_map_get_next_key,
740 .map_release_uref = array_map_free_timers,
741 .map_lookup_elem = array_map_lookup_elem,
742 .map_update_elem = array_map_update_elem,
743 .map_delete_elem = array_map_delete_elem,
744 .map_gen_lookup = array_map_gen_lookup,
745 .map_direct_value_addr = array_map_direct_value_addr,
746 .map_direct_value_meta = array_map_direct_value_meta,
747 .map_mmap = array_map_mmap,
748 .map_seq_show_elem = array_map_seq_show_elem,
749 .map_check_btf = array_map_check_btf,
750 .map_lookup_batch = generic_map_lookup_batch,
751 .map_update_batch = generic_map_update_batch,
752 .map_set_for_each_callback_args = map_set_for_each_callback_args,
753 .map_for_each_callback = bpf_for_each_array_elem,
754 .map_btf_id = &array_map_btf_ids[0],
755 .iter_seq_info = &iter_seq_info,
758 const struct bpf_map_ops percpu_array_map_ops = {
759 .map_meta_equal = bpf_map_meta_equal,
760 .map_alloc_check = array_map_alloc_check,
761 .map_alloc = array_map_alloc,
762 .map_free = array_map_free,
763 .map_get_next_key = array_map_get_next_key,
764 .map_lookup_elem = percpu_array_map_lookup_elem,
765 .map_update_elem = array_map_update_elem,
766 .map_delete_elem = array_map_delete_elem,
767 .map_lookup_percpu_elem = percpu_array_map_lookup_percpu_elem,
768 .map_seq_show_elem = percpu_array_map_seq_show_elem,
769 .map_check_btf = array_map_check_btf,
770 .map_lookup_batch = generic_map_lookup_batch,
771 .map_update_batch = generic_map_update_batch,
772 .map_set_for_each_callback_args = map_set_for_each_callback_args,
773 .map_for_each_callback = bpf_for_each_array_elem,
774 .map_btf_id = &array_map_btf_ids[0],
775 .iter_seq_info = &iter_seq_info,
778 static int fd_array_map_alloc_check(union bpf_attr *attr)
780 /* only file descriptors can be stored in this type of map */
781 if (attr->value_size != sizeof(u32))
783 /* Program read-only/write-only not supported for special maps yet. */
784 if (attr->map_flags & (BPF_F_RDONLY_PROG | BPF_F_WRONLY_PROG))
786 return array_map_alloc_check(attr);
789 static void fd_array_map_free(struct bpf_map *map)
791 struct bpf_array *array = container_of(map, struct bpf_array, map);
794 /* make sure it's empty */
795 for (i = 0; i < array->map.max_entries; i++)
796 BUG_ON(array->ptrs[i] != NULL);
798 bpf_map_area_free(array);
801 static void *fd_array_map_lookup_elem(struct bpf_map *map, void *key)
803 return ERR_PTR(-EOPNOTSUPP);
806 /* only called from syscall */
807 int bpf_fd_array_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
812 if (!map->ops->map_fd_sys_lookup_elem)
816 elem = array_map_lookup_elem(map, key);
817 if (elem && (ptr = READ_ONCE(*elem)))
818 *value = map->ops->map_fd_sys_lookup_elem(ptr);
826 /* only called from syscall */
827 int bpf_fd_array_map_update_elem(struct bpf_map *map, struct file *map_file,
828 void *key, void *value, u64 map_flags)
830 struct bpf_array *array = container_of(map, struct bpf_array, map);
831 void *new_ptr, *old_ptr;
832 u32 index = *(u32 *)key, ufd;
834 if (map_flags != BPF_ANY)
837 if (index >= array->map.max_entries)
841 new_ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
843 return PTR_ERR(new_ptr);
845 if (map->ops->map_poke_run) {
846 mutex_lock(&array->aux->poke_mutex);
847 old_ptr = xchg(array->ptrs + index, new_ptr);
848 map->ops->map_poke_run(map, index, old_ptr, new_ptr);
849 mutex_unlock(&array->aux->poke_mutex);
851 old_ptr = xchg(array->ptrs + index, new_ptr);
855 map->ops->map_fd_put_ptr(old_ptr);
859 static int fd_array_map_delete_elem(struct bpf_map *map, void *key)
861 struct bpf_array *array = container_of(map, struct bpf_array, map);
863 u32 index = *(u32 *)key;
865 if (index >= array->map.max_entries)
868 if (map->ops->map_poke_run) {
869 mutex_lock(&array->aux->poke_mutex);
870 old_ptr = xchg(array->ptrs + index, NULL);
871 map->ops->map_poke_run(map, index, old_ptr, NULL);
872 mutex_unlock(&array->aux->poke_mutex);
874 old_ptr = xchg(array->ptrs + index, NULL);
878 map->ops->map_fd_put_ptr(old_ptr);
885 static void *prog_fd_array_get_ptr(struct bpf_map *map,
886 struct file *map_file, int fd)
888 struct bpf_prog *prog = bpf_prog_get(fd);
893 if (!bpf_prog_map_compatible(map, prog)) {
895 return ERR_PTR(-EINVAL);
901 static void prog_fd_array_put_ptr(void *ptr)
906 static u32 prog_fd_array_sys_lookup_elem(void *ptr)
908 return ((struct bpf_prog *)ptr)->aux->id;
911 /* decrement refcnt of all bpf_progs that are stored in this map */
912 static void bpf_fd_array_map_clear(struct bpf_map *map)
914 struct bpf_array *array = container_of(map, struct bpf_array, map);
917 for (i = 0; i < array->map.max_entries; i++)
918 fd_array_map_delete_elem(map, &i);
921 static void prog_array_map_seq_show_elem(struct bpf_map *map, void *key,
929 elem = array_map_lookup_elem(map, key);
931 ptr = READ_ONCE(*elem);
933 seq_printf(m, "%u: ", *(u32 *)key);
934 prog_id = prog_fd_array_sys_lookup_elem(ptr);
935 btf_type_seq_show(map->btf, map->btf_value_type_id,
944 struct prog_poke_elem {
945 struct list_head list;
946 struct bpf_prog_aux *aux;
949 static int prog_array_map_poke_track(struct bpf_map *map,
950 struct bpf_prog_aux *prog_aux)
952 struct prog_poke_elem *elem;
953 struct bpf_array_aux *aux;
956 aux = container_of(map, struct bpf_array, map)->aux;
957 mutex_lock(&aux->poke_mutex);
958 list_for_each_entry(elem, &aux->poke_progs, list) {
959 if (elem->aux == prog_aux)
963 elem = kmalloc(sizeof(*elem), GFP_KERNEL);
969 INIT_LIST_HEAD(&elem->list);
970 /* We must track the program's aux info at this point in time
971 * since the program pointer itself may not be stable yet, see
972 * also comment in prog_array_map_poke_run().
974 elem->aux = prog_aux;
976 list_add_tail(&elem->list, &aux->poke_progs);
978 mutex_unlock(&aux->poke_mutex);
982 static void prog_array_map_poke_untrack(struct bpf_map *map,
983 struct bpf_prog_aux *prog_aux)
985 struct prog_poke_elem *elem, *tmp;
986 struct bpf_array_aux *aux;
988 aux = container_of(map, struct bpf_array, map)->aux;
989 mutex_lock(&aux->poke_mutex);
990 list_for_each_entry_safe(elem, tmp, &aux->poke_progs, list) {
991 if (elem->aux == prog_aux) {
992 list_del_init(&elem->list);
997 mutex_unlock(&aux->poke_mutex);
1000 static void prog_array_map_poke_run(struct bpf_map *map, u32 key,
1001 struct bpf_prog *old,
1002 struct bpf_prog *new)
1004 u8 *old_addr, *new_addr, *old_bypass_addr;
1005 struct prog_poke_elem *elem;
1006 struct bpf_array_aux *aux;
1008 aux = container_of(map, struct bpf_array, map)->aux;
1009 WARN_ON_ONCE(!mutex_is_locked(&aux->poke_mutex));
1011 list_for_each_entry(elem, &aux->poke_progs, list) {
1012 struct bpf_jit_poke_descriptor *poke;
1015 for (i = 0; i < elem->aux->size_poke_tab; i++) {
1016 poke = &elem->aux->poke_tab[i];
1018 /* Few things to be aware of:
1020 * 1) We can only ever access aux in this context, but
1021 * not aux->prog since it might not be stable yet and
1022 * there could be danger of use after free otherwise.
1023 * 2) Initially when we start tracking aux, the program
1024 * is not JITed yet and also does not have a kallsyms
1025 * entry. We skip these as poke->tailcall_target_stable
1026 * is not active yet. The JIT will do the final fixup
1027 * before setting it stable. The various
1028 * poke->tailcall_target_stable are successively
1029 * activated, so tail call updates can arrive from here
1030 * while JIT is still finishing its final fixup for
1031 * non-activated poke entries.
1032 * 3) On program teardown, the program's kallsym entry gets
1033 * removed out of RCU callback, but we can only untrack
1034 * from sleepable context, therefore bpf_arch_text_poke()
1035 * might not see that this is in BPF text section and
1036 * bails out with -EINVAL. As these are unreachable since
1037 * RCU grace period already passed, we simply skip them.
1038 * 4) Also programs reaching refcount of zero while patching
1039 * is in progress is okay since we're protected under
1040 * poke_mutex and untrack the programs before the JIT
1041 * buffer is freed. When we're still in the middle of
1042 * patching and suddenly kallsyms entry of the program
1043 * gets evicted, we just skip the rest which is fine due
1045 * 5) Any other error happening below from bpf_arch_text_poke()
1046 * is a unexpected bug.
1048 if (!READ_ONCE(poke->tailcall_target_stable))
1050 if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
1052 if (poke->tail_call.map != map ||
1053 poke->tail_call.key != key)
1056 old_bypass_addr = old ? NULL : poke->bypass_addr;
1057 old_addr = old ? (u8 *)old->bpf_func + poke->adj_off : NULL;
1058 new_addr = new ? (u8 *)new->bpf_func + poke->adj_off : NULL;
1061 ret = bpf_arch_text_poke(poke->tailcall_target,
1063 old_addr, new_addr);
1064 BUG_ON(ret < 0 && ret != -EINVAL);
1066 ret = bpf_arch_text_poke(poke->tailcall_bypass,
1070 BUG_ON(ret < 0 && ret != -EINVAL);
1073 ret = bpf_arch_text_poke(poke->tailcall_bypass,
1077 BUG_ON(ret < 0 && ret != -EINVAL);
1078 /* let other CPUs finish the execution of program
1079 * so that it will not possible to expose them
1080 * to invalid nop, stack unwind, nop state
1084 ret = bpf_arch_text_poke(poke->tailcall_target,
1087 BUG_ON(ret < 0 && ret != -EINVAL);
1093 static void prog_array_map_clear_deferred(struct work_struct *work)
1095 struct bpf_map *map = container_of(work, struct bpf_array_aux,
1097 bpf_fd_array_map_clear(map);
1101 static void prog_array_map_clear(struct bpf_map *map)
1103 struct bpf_array_aux *aux = container_of(map, struct bpf_array,
1106 schedule_work(&aux->work);
1109 static struct bpf_map *prog_array_map_alloc(union bpf_attr *attr)
1111 struct bpf_array_aux *aux;
1112 struct bpf_map *map;
1114 aux = kzalloc(sizeof(*aux), GFP_KERNEL_ACCOUNT);
1116 return ERR_PTR(-ENOMEM);
1118 INIT_WORK(&aux->work, prog_array_map_clear_deferred);
1119 INIT_LIST_HEAD(&aux->poke_progs);
1120 mutex_init(&aux->poke_mutex);
1122 map = array_map_alloc(attr);
1128 container_of(map, struct bpf_array, map)->aux = aux;
1134 static void prog_array_map_free(struct bpf_map *map)
1136 struct prog_poke_elem *elem, *tmp;
1137 struct bpf_array_aux *aux;
1139 aux = container_of(map, struct bpf_array, map)->aux;
1140 list_for_each_entry_safe(elem, tmp, &aux->poke_progs, list) {
1141 list_del_init(&elem->list);
1145 fd_array_map_free(map);
1148 /* prog_array->aux->{type,jited} is a runtime binding.
1149 * Doing static check alone in the verifier is not enough.
1150 * Thus, prog_array_map cannot be used as an inner_map
1151 * and map_meta_equal is not implemented.
1153 const struct bpf_map_ops prog_array_map_ops = {
1154 .map_alloc_check = fd_array_map_alloc_check,
1155 .map_alloc = prog_array_map_alloc,
1156 .map_free = prog_array_map_free,
1157 .map_poke_track = prog_array_map_poke_track,
1158 .map_poke_untrack = prog_array_map_poke_untrack,
1159 .map_poke_run = prog_array_map_poke_run,
1160 .map_get_next_key = array_map_get_next_key,
1161 .map_lookup_elem = fd_array_map_lookup_elem,
1162 .map_delete_elem = fd_array_map_delete_elem,
1163 .map_fd_get_ptr = prog_fd_array_get_ptr,
1164 .map_fd_put_ptr = prog_fd_array_put_ptr,
1165 .map_fd_sys_lookup_elem = prog_fd_array_sys_lookup_elem,
1166 .map_release_uref = prog_array_map_clear,
1167 .map_seq_show_elem = prog_array_map_seq_show_elem,
1168 .map_btf_id = &array_map_btf_ids[0],
1171 static struct bpf_event_entry *bpf_event_entry_gen(struct file *perf_file,
1172 struct file *map_file)
1174 struct bpf_event_entry *ee;
1176 ee = kzalloc(sizeof(*ee), GFP_ATOMIC);
1178 ee->event = perf_file->private_data;
1179 ee->perf_file = perf_file;
1180 ee->map_file = map_file;
1186 static void __bpf_event_entry_free(struct rcu_head *rcu)
1188 struct bpf_event_entry *ee;
1190 ee = container_of(rcu, struct bpf_event_entry, rcu);
1191 fput(ee->perf_file);
1195 static void bpf_event_entry_free_rcu(struct bpf_event_entry *ee)
1197 call_rcu(&ee->rcu, __bpf_event_entry_free);
1200 static void *perf_event_fd_array_get_ptr(struct bpf_map *map,
1201 struct file *map_file, int fd)
1203 struct bpf_event_entry *ee;
1204 struct perf_event *event;
1205 struct file *perf_file;
1208 perf_file = perf_event_get(fd);
1209 if (IS_ERR(perf_file))
1212 ee = ERR_PTR(-EOPNOTSUPP);
1213 event = perf_file->private_data;
1214 if (perf_event_read_local(event, &value, NULL, NULL) == -EOPNOTSUPP)
1217 ee = bpf_event_entry_gen(perf_file, map_file);
1220 ee = ERR_PTR(-ENOMEM);
1226 static void perf_event_fd_array_put_ptr(void *ptr)
1228 bpf_event_entry_free_rcu(ptr);
1231 static void perf_event_fd_array_release(struct bpf_map *map,
1232 struct file *map_file)
1234 struct bpf_array *array = container_of(map, struct bpf_array, map);
1235 struct bpf_event_entry *ee;
1238 if (map->map_flags & BPF_F_PRESERVE_ELEMS)
1242 for (i = 0; i < array->map.max_entries; i++) {
1243 ee = READ_ONCE(array->ptrs[i]);
1244 if (ee && ee->map_file == map_file)
1245 fd_array_map_delete_elem(map, &i);
1250 static void perf_event_fd_array_map_free(struct bpf_map *map)
1252 if (map->map_flags & BPF_F_PRESERVE_ELEMS)
1253 bpf_fd_array_map_clear(map);
1254 fd_array_map_free(map);
1257 const struct bpf_map_ops perf_event_array_map_ops = {
1258 .map_meta_equal = bpf_map_meta_equal,
1259 .map_alloc_check = fd_array_map_alloc_check,
1260 .map_alloc = array_map_alloc,
1261 .map_free = perf_event_fd_array_map_free,
1262 .map_get_next_key = array_map_get_next_key,
1263 .map_lookup_elem = fd_array_map_lookup_elem,
1264 .map_delete_elem = fd_array_map_delete_elem,
1265 .map_fd_get_ptr = perf_event_fd_array_get_ptr,
1266 .map_fd_put_ptr = perf_event_fd_array_put_ptr,
1267 .map_release = perf_event_fd_array_release,
1268 .map_check_btf = map_check_no_btf,
1269 .map_btf_id = &array_map_btf_ids[0],
1272 #ifdef CONFIG_CGROUPS
1273 static void *cgroup_fd_array_get_ptr(struct bpf_map *map,
1274 struct file *map_file /* not used */,
1277 return cgroup_get_from_fd(fd);
1280 static void cgroup_fd_array_put_ptr(void *ptr)
1282 /* cgroup_put free cgrp after a rcu grace period */
1286 static void cgroup_fd_array_free(struct bpf_map *map)
1288 bpf_fd_array_map_clear(map);
1289 fd_array_map_free(map);
1292 const struct bpf_map_ops cgroup_array_map_ops = {
1293 .map_meta_equal = bpf_map_meta_equal,
1294 .map_alloc_check = fd_array_map_alloc_check,
1295 .map_alloc = array_map_alloc,
1296 .map_free = cgroup_fd_array_free,
1297 .map_get_next_key = array_map_get_next_key,
1298 .map_lookup_elem = fd_array_map_lookup_elem,
1299 .map_delete_elem = fd_array_map_delete_elem,
1300 .map_fd_get_ptr = cgroup_fd_array_get_ptr,
1301 .map_fd_put_ptr = cgroup_fd_array_put_ptr,
1302 .map_check_btf = map_check_no_btf,
1303 .map_btf_id = &array_map_btf_ids[0],
1307 static struct bpf_map *array_of_map_alloc(union bpf_attr *attr)
1309 struct bpf_map *map, *inner_map_meta;
1311 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1312 if (IS_ERR(inner_map_meta))
1313 return inner_map_meta;
1315 map = array_map_alloc(attr);
1317 bpf_map_meta_free(inner_map_meta);
1321 map->inner_map_meta = inner_map_meta;
1326 static void array_of_map_free(struct bpf_map *map)
1328 /* map->inner_map_meta is only accessed by syscall which
1329 * is protected by fdget/fdput.
1331 bpf_map_meta_free(map->inner_map_meta);
1332 bpf_fd_array_map_clear(map);
1333 fd_array_map_free(map);
1336 static void *array_of_map_lookup_elem(struct bpf_map *map, void *key)
1338 struct bpf_map **inner_map = array_map_lookup_elem(map, key);
1343 return READ_ONCE(*inner_map);
1346 static int array_of_map_gen_lookup(struct bpf_map *map,
1347 struct bpf_insn *insn_buf)
1349 struct bpf_array *array = container_of(map, struct bpf_array, map);
1350 u32 elem_size = array->elem_size;
1351 struct bpf_insn *insn = insn_buf;
1352 const int ret = BPF_REG_0;
1353 const int map_ptr = BPF_REG_1;
1354 const int index = BPF_REG_2;
1356 *insn++ = BPF_ALU64_IMM(BPF_ADD, map_ptr, offsetof(struct bpf_array, value));
1357 *insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
1358 if (!map->bypass_spec_v1) {
1359 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 6);
1360 *insn++ = BPF_ALU32_IMM(BPF_AND, ret, array->index_mask);
1362 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 5);
1364 if (is_power_of_2(elem_size))
1365 *insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(elem_size));
1367 *insn++ = BPF_ALU64_IMM(BPF_MUL, ret, elem_size);
1368 *insn++ = BPF_ALU64_REG(BPF_ADD, ret, map_ptr);
1369 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1370 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
1371 *insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
1372 *insn++ = BPF_MOV64_IMM(ret, 0);
1374 return insn - insn_buf;
1377 const struct bpf_map_ops array_of_maps_map_ops = {
1378 .map_alloc_check = fd_array_map_alloc_check,
1379 .map_alloc = array_of_map_alloc,
1380 .map_free = array_of_map_free,
1381 .map_get_next_key = array_map_get_next_key,
1382 .map_lookup_elem = array_of_map_lookup_elem,
1383 .map_delete_elem = fd_array_map_delete_elem,
1384 .map_fd_get_ptr = bpf_map_fd_get_ptr,
1385 .map_fd_put_ptr = bpf_map_fd_put_ptr,
1386 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1387 .map_gen_lookup = array_of_map_gen_lookup,
1388 .map_lookup_batch = generic_map_lookup_batch,
1389 .map_update_batch = generic_map_update_batch,
1390 .map_check_btf = map_check_no_btf,
1391 .map_btf_id = &array_map_btf_ids[0],