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 bpf_long_memcpy(value + off, per_cpu_ptr(pptr, cpu), size);
289 /* Called from syscall */
290 static int array_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
292 struct bpf_array *array = container_of(map, struct bpf_array, map);
293 u32 index = key ? *(u32 *)key : U32_MAX;
294 u32 *next = (u32 *)next_key;
296 if (index >= array->map.max_entries) {
301 if (index == array->map.max_entries - 1)
308 static void check_and_free_fields(struct bpf_array *arr, void *val)
310 if (map_value_has_timer(&arr->map))
311 bpf_timer_cancel_and_free(val + arr->map.timer_off);
312 if (map_value_has_kptrs(&arr->map))
313 bpf_map_free_kptrs(&arr->map, val);
316 /* Called from syscall or from eBPF program */
317 static int array_map_update_elem(struct bpf_map *map, void *key, void *value,
320 struct bpf_array *array = container_of(map, struct bpf_array, map);
321 u32 index = *(u32 *)key;
324 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
328 if (unlikely(index >= array->map.max_entries))
329 /* all elements were pre-allocated, cannot insert a new one */
332 if (unlikely(map_flags & BPF_NOEXIST))
333 /* all elements already exist */
336 if (unlikely((map_flags & BPF_F_LOCK) &&
337 !map_value_has_spin_lock(map)))
340 if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
341 memcpy(this_cpu_ptr(array->pptrs[index & array->index_mask]),
342 value, map->value_size);
345 (u64)array->elem_size * (index & array->index_mask);
346 if (map_flags & BPF_F_LOCK)
347 copy_map_value_locked(map, val, value, false);
349 copy_map_value(map, val, value);
350 check_and_free_fields(array, val);
355 int bpf_percpu_array_update(struct bpf_map *map, void *key, void *value,
358 struct bpf_array *array = container_of(map, struct bpf_array, map);
359 u32 index = *(u32 *)key;
364 if (unlikely(map_flags > BPF_EXIST))
368 if (unlikely(index >= array->map.max_entries))
369 /* all elements were pre-allocated, cannot insert a new one */
372 if (unlikely(map_flags == BPF_NOEXIST))
373 /* all elements already exist */
376 /* the user space will provide round_up(value_size, 8) bytes that
377 * will be copied into per-cpu area. bpf programs can only access
378 * value_size of it. During lookup the same extra bytes will be
379 * returned or zeros which were zero-filled by percpu_alloc,
380 * so no kernel data leaks possible
382 size = array->elem_size;
384 pptr = array->pptrs[index & array->index_mask];
385 for_each_possible_cpu(cpu) {
386 bpf_long_memcpy(per_cpu_ptr(pptr, cpu), value + off, size);
393 /* Called from syscall or from eBPF program */
394 static int array_map_delete_elem(struct bpf_map *map, void *key)
399 static void *array_map_vmalloc_addr(struct bpf_array *array)
401 return (void *)round_down((unsigned long)array, PAGE_SIZE);
404 static void array_map_free_timers(struct bpf_map *map)
406 struct bpf_array *array = container_of(map, struct bpf_array, map);
409 /* We don't reset or free kptr on uref dropping to zero. */
410 if (!map_value_has_timer(map))
413 for (i = 0; i < array->map.max_entries; i++)
414 bpf_timer_cancel_and_free(array_map_elem_ptr(array, i) + map->timer_off);
417 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
418 static void array_map_free(struct bpf_map *map)
420 struct bpf_array *array = container_of(map, struct bpf_array, map);
423 if (map_value_has_kptrs(map)) {
424 for (i = 0; i < array->map.max_entries; i++)
425 bpf_map_free_kptrs(map, array_map_elem_ptr(array, i));
426 bpf_map_free_kptr_off_tab(map);
429 if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY)
430 bpf_array_free_percpu(array);
432 if (array->map.map_flags & BPF_F_MMAPABLE)
433 bpf_map_area_free(array_map_vmalloc_addr(array));
435 bpf_map_area_free(array);
438 static void array_map_seq_show_elem(struct bpf_map *map, void *key,
445 value = array_map_lookup_elem(map, key);
451 if (map->btf_key_type_id)
452 seq_printf(m, "%u: ", *(u32 *)key);
453 btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
459 static void percpu_array_map_seq_show_elem(struct bpf_map *map, void *key,
462 struct bpf_array *array = container_of(map, struct bpf_array, map);
463 u32 index = *(u32 *)key;
469 seq_printf(m, "%u: {\n", *(u32 *)key);
470 pptr = array->pptrs[index & array->index_mask];
471 for_each_possible_cpu(cpu) {
472 seq_printf(m, "\tcpu%d: ", cpu);
473 btf_type_seq_show(map->btf, map->btf_value_type_id,
474 per_cpu_ptr(pptr, cpu), m);
482 static int array_map_check_btf(const struct bpf_map *map,
483 const struct btf *btf,
484 const struct btf_type *key_type,
485 const struct btf_type *value_type)
489 /* One exception for keyless BTF: .bss/.data/.rodata map */
490 if (btf_type_is_void(key_type)) {
491 if (map->map_type != BPF_MAP_TYPE_ARRAY ||
492 map->max_entries != 1)
495 if (BTF_INFO_KIND(value_type->info) != BTF_KIND_DATASEC)
501 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
504 int_data = *(u32 *)(key_type + 1);
505 /* bpf array can only take a u32 key. This check makes sure
506 * that the btf matches the attr used during map_create.
508 if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
514 static int array_map_mmap(struct bpf_map *map, struct vm_area_struct *vma)
516 struct bpf_array *array = container_of(map, struct bpf_array, map);
517 pgoff_t pgoff = PAGE_ALIGN(sizeof(*array)) >> PAGE_SHIFT;
519 if (!(map->map_flags & BPF_F_MMAPABLE))
522 if (vma->vm_pgoff * PAGE_SIZE + (vma->vm_end - vma->vm_start) >
523 PAGE_ALIGN((u64)array->map.max_entries * array->elem_size))
526 return remap_vmalloc_range(vma, array_map_vmalloc_addr(array),
527 vma->vm_pgoff + pgoff);
530 static bool array_map_meta_equal(const struct bpf_map *meta0,
531 const struct bpf_map *meta1)
533 if (!bpf_map_meta_equal(meta0, meta1))
535 return meta0->map_flags & BPF_F_INNER_MAP ? true :
536 meta0->max_entries == meta1->max_entries;
539 struct bpf_iter_seq_array_map_info {
541 void *percpu_value_buf;
545 static void *bpf_array_map_seq_start(struct seq_file *seq, loff_t *pos)
547 struct bpf_iter_seq_array_map_info *info = seq->private;
548 struct bpf_map *map = info->map;
549 struct bpf_array *array;
552 if (info->index >= map->max_entries)
557 array = container_of(map, struct bpf_array, map);
558 index = info->index & array->index_mask;
559 if (info->percpu_value_buf)
560 return array->pptrs[index];
561 return array_map_elem_ptr(array, index);
564 static void *bpf_array_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
566 struct bpf_iter_seq_array_map_info *info = seq->private;
567 struct bpf_map *map = info->map;
568 struct bpf_array *array;
573 if (info->index >= map->max_entries)
576 array = container_of(map, struct bpf_array, map);
577 index = info->index & array->index_mask;
578 if (info->percpu_value_buf)
579 return array->pptrs[index];
580 return array_map_elem_ptr(array, index);
583 static int __bpf_array_map_seq_show(struct seq_file *seq, void *v)
585 struct bpf_iter_seq_array_map_info *info = seq->private;
586 struct bpf_iter__bpf_map_elem ctx = {};
587 struct bpf_map *map = info->map;
588 struct bpf_array *array = container_of(map, struct bpf_array, map);
589 struct bpf_iter_meta meta;
590 struct bpf_prog *prog;
591 int off = 0, cpu = 0;
592 void __percpu **pptr;
596 prog = bpf_iter_get_info(&meta, v == NULL);
603 ctx.key = &info->index;
605 if (!info->percpu_value_buf) {
609 size = array->elem_size;
610 for_each_possible_cpu(cpu) {
611 bpf_long_memcpy(info->percpu_value_buf + off,
612 per_cpu_ptr(pptr, cpu),
616 ctx.value = info->percpu_value_buf;
620 return bpf_iter_run_prog(prog, &ctx);
623 static int bpf_array_map_seq_show(struct seq_file *seq, void *v)
625 return __bpf_array_map_seq_show(seq, v);
628 static void bpf_array_map_seq_stop(struct seq_file *seq, void *v)
631 (void)__bpf_array_map_seq_show(seq, NULL);
634 static int bpf_iter_init_array_map(void *priv_data,
635 struct bpf_iter_aux_info *aux)
637 struct bpf_iter_seq_array_map_info *seq_info = priv_data;
638 struct bpf_map *map = aux->map;
639 struct bpf_array *array = container_of(map, struct bpf_array, map);
643 if (map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
644 buf_size = array->elem_size * num_possible_cpus();
645 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
649 seq_info->percpu_value_buf = value_buf;
656 static void bpf_iter_fini_array_map(void *priv_data)
658 struct bpf_iter_seq_array_map_info *seq_info = priv_data;
660 kfree(seq_info->percpu_value_buf);
663 static const struct seq_operations bpf_array_map_seq_ops = {
664 .start = bpf_array_map_seq_start,
665 .next = bpf_array_map_seq_next,
666 .stop = bpf_array_map_seq_stop,
667 .show = bpf_array_map_seq_show,
670 static const struct bpf_iter_seq_info iter_seq_info = {
671 .seq_ops = &bpf_array_map_seq_ops,
672 .init_seq_private = bpf_iter_init_array_map,
673 .fini_seq_private = bpf_iter_fini_array_map,
674 .seq_priv_size = sizeof(struct bpf_iter_seq_array_map_info),
677 static int bpf_for_each_array_elem(struct bpf_map *map, bpf_callback_t callback_fn,
678 void *callback_ctx, u64 flags)
680 u32 i, key, num_elems = 0;
681 struct bpf_array *array;
689 is_percpu = map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
690 array = container_of(map, struct bpf_array, map);
693 for (i = 0; i < map->max_entries; i++) {
695 val = this_cpu_ptr(array->pptrs[i]);
697 val = array_map_elem_ptr(array, i);
700 ret = callback_fn((u64)(long)map, (u64)(long)&key,
701 (u64)(long)val, (u64)(long)callback_ctx, 0);
702 /* return value: 0 - continue, 1 - stop and return */
712 BTF_ID_LIST_SINGLE(array_map_btf_ids, struct, bpf_array)
713 const struct bpf_map_ops array_map_ops = {
714 .map_meta_equal = array_map_meta_equal,
715 .map_alloc_check = array_map_alloc_check,
716 .map_alloc = array_map_alloc,
717 .map_free = array_map_free,
718 .map_get_next_key = array_map_get_next_key,
719 .map_release_uref = array_map_free_timers,
720 .map_lookup_elem = array_map_lookup_elem,
721 .map_update_elem = array_map_update_elem,
722 .map_delete_elem = array_map_delete_elem,
723 .map_gen_lookup = array_map_gen_lookup,
724 .map_direct_value_addr = array_map_direct_value_addr,
725 .map_direct_value_meta = array_map_direct_value_meta,
726 .map_mmap = array_map_mmap,
727 .map_seq_show_elem = array_map_seq_show_elem,
728 .map_check_btf = array_map_check_btf,
729 .map_lookup_batch = generic_map_lookup_batch,
730 .map_update_batch = generic_map_update_batch,
731 .map_set_for_each_callback_args = map_set_for_each_callback_args,
732 .map_for_each_callback = bpf_for_each_array_elem,
733 .map_btf_id = &array_map_btf_ids[0],
734 .iter_seq_info = &iter_seq_info,
737 const struct bpf_map_ops percpu_array_map_ops = {
738 .map_meta_equal = bpf_map_meta_equal,
739 .map_alloc_check = array_map_alloc_check,
740 .map_alloc = array_map_alloc,
741 .map_free = array_map_free,
742 .map_get_next_key = array_map_get_next_key,
743 .map_lookup_elem = percpu_array_map_lookup_elem,
744 .map_update_elem = array_map_update_elem,
745 .map_delete_elem = array_map_delete_elem,
746 .map_lookup_percpu_elem = percpu_array_map_lookup_percpu_elem,
747 .map_seq_show_elem = percpu_array_map_seq_show_elem,
748 .map_check_btf = array_map_check_btf,
749 .map_lookup_batch = generic_map_lookup_batch,
750 .map_update_batch = generic_map_update_batch,
751 .map_set_for_each_callback_args = map_set_for_each_callback_args,
752 .map_for_each_callback = bpf_for_each_array_elem,
753 .map_btf_id = &array_map_btf_ids[0],
754 .iter_seq_info = &iter_seq_info,
757 static int fd_array_map_alloc_check(union bpf_attr *attr)
759 /* only file descriptors can be stored in this type of map */
760 if (attr->value_size != sizeof(u32))
762 /* Program read-only/write-only not supported for special maps yet. */
763 if (attr->map_flags & (BPF_F_RDONLY_PROG | BPF_F_WRONLY_PROG))
765 return array_map_alloc_check(attr);
768 static void fd_array_map_free(struct bpf_map *map)
770 struct bpf_array *array = container_of(map, struct bpf_array, map);
773 /* make sure it's empty */
774 for (i = 0; i < array->map.max_entries; i++)
775 BUG_ON(array->ptrs[i] != NULL);
777 bpf_map_area_free(array);
780 static void *fd_array_map_lookup_elem(struct bpf_map *map, void *key)
782 return ERR_PTR(-EOPNOTSUPP);
785 /* only called from syscall */
786 int bpf_fd_array_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
791 if (!map->ops->map_fd_sys_lookup_elem)
795 elem = array_map_lookup_elem(map, key);
796 if (elem && (ptr = READ_ONCE(*elem)))
797 *value = map->ops->map_fd_sys_lookup_elem(ptr);
805 /* only called from syscall */
806 int bpf_fd_array_map_update_elem(struct bpf_map *map, struct file *map_file,
807 void *key, void *value, u64 map_flags)
809 struct bpf_array *array = container_of(map, struct bpf_array, map);
810 void *new_ptr, *old_ptr;
811 u32 index = *(u32 *)key, ufd;
813 if (map_flags != BPF_ANY)
816 if (index >= array->map.max_entries)
820 new_ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
822 return PTR_ERR(new_ptr);
824 if (map->ops->map_poke_run) {
825 mutex_lock(&array->aux->poke_mutex);
826 old_ptr = xchg(array->ptrs + index, new_ptr);
827 map->ops->map_poke_run(map, index, old_ptr, new_ptr);
828 mutex_unlock(&array->aux->poke_mutex);
830 old_ptr = xchg(array->ptrs + index, new_ptr);
834 map->ops->map_fd_put_ptr(old_ptr);
838 static int fd_array_map_delete_elem(struct bpf_map *map, void *key)
840 struct bpf_array *array = container_of(map, struct bpf_array, map);
842 u32 index = *(u32 *)key;
844 if (index >= array->map.max_entries)
847 if (map->ops->map_poke_run) {
848 mutex_lock(&array->aux->poke_mutex);
849 old_ptr = xchg(array->ptrs + index, NULL);
850 map->ops->map_poke_run(map, index, old_ptr, NULL);
851 mutex_unlock(&array->aux->poke_mutex);
853 old_ptr = xchg(array->ptrs + index, NULL);
857 map->ops->map_fd_put_ptr(old_ptr);
864 static void *prog_fd_array_get_ptr(struct bpf_map *map,
865 struct file *map_file, int fd)
867 struct bpf_prog *prog = bpf_prog_get(fd);
872 if (!bpf_prog_map_compatible(map, prog)) {
874 return ERR_PTR(-EINVAL);
880 static void prog_fd_array_put_ptr(void *ptr)
885 static u32 prog_fd_array_sys_lookup_elem(void *ptr)
887 return ((struct bpf_prog *)ptr)->aux->id;
890 /* decrement refcnt of all bpf_progs that are stored in this map */
891 static void bpf_fd_array_map_clear(struct bpf_map *map)
893 struct bpf_array *array = container_of(map, struct bpf_array, map);
896 for (i = 0; i < array->map.max_entries; i++)
897 fd_array_map_delete_elem(map, &i);
900 static void prog_array_map_seq_show_elem(struct bpf_map *map, void *key,
908 elem = array_map_lookup_elem(map, key);
910 ptr = READ_ONCE(*elem);
912 seq_printf(m, "%u: ", *(u32 *)key);
913 prog_id = prog_fd_array_sys_lookup_elem(ptr);
914 btf_type_seq_show(map->btf, map->btf_value_type_id,
923 struct prog_poke_elem {
924 struct list_head list;
925 struct bpf_prog_aux *aux;
928 static int prog_array_map_poke_track(struct bpf_map *map,
929 struct bpf_prog_aux *prog_aux)
931 struct prog_poke_elem *elem;
932 struct bpf_array_aux *aux;
935 aux = container_of(map, struct bpf_array, map)->aux;
936 mutex_lock(&aux->poke_mutex);
937 list_for_each_entry(elem, &aux->poke_progs, list) {
938 if (elem->aux == prog_aux)
942 elem = kmalloc(sizeof(*elem), GFP_KERNEL);
948 INIT_LIST_HEAD(&elem->list);
949 /* We must track the program's aux info at this point in time
950 * since the program pointer itself may not be stable yet, see
951 * also comment in prog_array_map_poke_run().
953 elem->aux = prog_aux;
955 list_add_tail(&elem->list, &aux->poke_progs);
957 mutex_unlock(&aux->poke_mutex);
961 static void prog_array_map_poke_untrack(struct bpf_map *map,
962 struct bpf_prog_aux *prog_aux)
964 struct prog_poke_elem *elem, *tmp;
965 struct bpf_array_aux *aux;
967 aux = container_of(map, struct bpf_array, map)->aux;
968 mutex_lock(&aux->poke_mutex);
969 list_for_each_entry_safe(elem, tmp, &aux->poke_progs, list) {
970 if (elem->aux == prog_aux) {
971 list_del_init(&elem->list);
976 mutex_unlock(&aux->poke_mutex);
979 static void prog_array_map_poke_run(struct bpf_map *map, u32 key,
980 struct bpf_prog *old,
981 struct bpf_prog *new)
983 u8 *old_addr, *new_addr, *old_bypass_addr;
984 struct prog_poke_elem *elem;
985 struct bpf_array_aux *aux;
987 aux = container_of(map, struct bpf_array, map)->aux;
988 WARN_ON_ONCE(!mutex_is_locked(&aux->poke_mutex));
990 list_for_each_entry(elem, &aux->poke_progs, list) {
991 struct bpf_jit_poke_descriptor *poke;
994 for (i = 0; i < elem->aux->size_poke_tab; i++) {
995 poke = &elem->aux->poke_tab[i];
997 /* Few things to be aware of:
999 * 1) We can only ever access aux in this context, but
1000 * not aux->prog since it might not be stable yet and
1001 * there could be danger of use after free otherwise.
1002 * 2) Initially when we start tracking aux, the program
1003 * is not JITed yet and also does not have a kallsyms
1004 * entry. We skip these as poke->tailcall_target_stable
1005 * is not active yet. The JIT will do the final fixup
1006 * before setting it stable. The various
1007 * poke->tailcall_target_stable are successively
1008 * activated, so tail call updates can arrive from here
1009 * while JIT is still finishing its final fixup for
1010 * non-activated poke entries.
1011 * 3) On program teardown, the program's kallsym entry gets
1012 * removed out of RCU callback, but we can only untrack
1013 * from sleepable context, therefore bpf_arch_text_poke()
1014 * might not see that this is in BPF text section and
1015 * bails out with -EINVAL. As these are unreachable since
1016 * RCU grace period already passed, we simply skip them.
1017 * 4) Also programs reaching refcount of zero while patching
1018 * is in progress is okay since we're protected under
1019 * poke_mutex and untrack the programs before the JIT
1020 * buffer is freed. When we're still in the middle of
1021 * patching and suddenly kallsyms entry of the program
1022 * gets evicted, we just skip the rest which is fine due
1024 * 5) Any other error happening below from bpf_arch_text_poke()
1025 * is a unexpected bug.
1027 if (!READ_ONCE(poke->tailcall_target_stable))
1029 if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
1031 if (poke->tail_call.map != map ||
1032 poke->tail_call.key != key)
1035 old_bypass_addr = old ? NULL : poke->bypass_addr;
1036 old_addr = old ? (u8 *)old->bpf_func + poke->adj_off : NULL;
1037 new_addr = new ? (u8 *)new->bpf_func + poke->adj_off : NULL;
1040 ret = bpf_arch_text_poke(poke->tailcall_target,
1042 old_addr, new_addr);
1043 BUG_ON(ret < 0 && ret != -EINVAL);
1045 ret = bpf_arch_text_poke(poke->tailcall_bypass,
1049 BUG_ON(ret < 0 && ret != -EINVAL);
1052 ret = bpf_arch_text_poke(poke->tailcall_bypass,
1056 BUG_ON(ret < 0 && ret != -EINVAL);
1057 /* let other CPUs finish the execution of program
1058 * so that it will not possible to expose them
1059 * to invalid nop, stack unwind, nop state
1063 ret = bpf_arch_text_poke(poke->tailcall_target,
1066 BUG_ON(ret < 0 && ret != -EINVAL);
1072 static void prog_array_map_clear_deferred(struct work_struct *work)
1074 struct bpf_map *map = container_of(work, struct bpf_array_aux,
1076 bpf_fd_array_map_clear(map);
1080 static void prog_array_map_clear(struct bpf_map *map)
1082 struct bpf_array_aux *aux = container_of(map, struct bpf_array,
1085 schedule_work(&aux->work);
1088 static struct bpf_map *prog_array_map_alloc(union bpf_attr *attr)
1090 struct bpf_array_aux *aux;
1091 struct bpf_map *map;
1093 aux = kzalloc(sizeof(*aux), GFP_KERNEL_ACCOUNT);
1095 return ERR_PTR(-ENOMEM);
1097 INIT_WORK(&aux->work, prog_array_map_clear_deferred);
1098 INIT_LIST_HEAD(&aux->poke_progs);
1099 mutex_init(&aux->poke_mutex);
1101 map = array_map_alloc(attr);
1107 container_of(map, struct bpf_array, map)->aux = aux;
1113 static void prog_array_map_free(struct bpf_map *map)
1115 struct prog_poke_elem *elem, *tmp;
1116 struct bpf_array_aux *aux;
1118 aux = container_of(map, struct bpf_array, map)->aux;
1119 list_for_each_entry_safe(elem, tmp, &aux->poke_progs, list) {
1120 list_del_init(&elem->list);
1124 fd_array_map_free(map);
1127 /* prog_array->aux->{type,jited} is a runtime binding.
1128 * Doing static check alone in the verifier is not enough.
1129 * Thus, prog_array_map cannot be used as an inner_map
1130 * and map_meta_equal is not implemented.
1132 const struct bpf_map_ops prog_array_map_ops = {
1133 .map_alloc_check = fd_array_map_alloc_check,
1134 .map_alloc = prog_array_map_alloc,
1135 .map_free = prog_array_map_free,
1136 .map_poke_track = prog_array_map_poke_track,
1137 .map_poke_untrack = prog_array_map_poke_untrack,
1138 .map_poke_run = prog_array_map_poke_run,
1139 .map_get_next_key = array_map_get_next_key,
1140 .map_lookup_elem = fd_array_map_lookup_elem,
1141 .map_delete_elem = fd_array_map_delete_elem,
1142 .map_fd_get_ptr = prog_fd_array_get_ptr,
1143 .map_fd_put_ptr = prog_fd_array_put_ptr,
1144 .map_fd_sys_lookup_elem = prog_fd_array_sys_lookup_elem,
1145 .map_release_uref = prog_array_map_clear,
1146 .map_seq_show_elem = prog_array_map_seq_show_elem,
1147 .map_btf_id = &array_map_btf_ids[0],
1150 static struct bpf_event_entry *bpf_event_entry_gen(struct file *perf_file,
1151 struct file *map_file)
1153 struct bpf_event_entry *ee;
1155 ee = kzalloc(sizeof(*ee), GFP_ATOMIC);
1157 ee->event = perf_file->private_data;
1158 ee->perf_file = perf_file;
1159 ee->map_file = map_file;
1165 static void __bpf_event_entry_free(struct rcu_head *rcu)
1167 struct bpf_event_entry *ee;
1169 ee = container_of(rcu, struct bpf_event_entry, rcu);
1170 fput(ee->perf_file);
1174 static void bpf_event_entry_free_rcu(struct bpf_event_entry *ee)
1176 call_rcu(&ee->rcu, __bpf_event_entry_free);
1179 static void *perf_event_fd_array_get_ptr(struct bpf_map *map,
1180 struct file *map_file, int fd)
1182 struct bpf_event_entry *ee;
1183 struct perf_event *event;
1184 struct file *perf_file;
1187 perf_file = perf_event_get(fd);
1188 if (IS_ERR(perf_file))
1191 ee = ERR_PTR(-EOPNOTSUPP);
1192 event = perf_file->private_data;
1193 if (perf_event_read_local(event, &value, NULL, NULL) == -EOPNOTSUPP)
1196 ee = bpf_event_entry_gen(perf_file, map_file);
1199 ee = ERR_PTR(-ENOMEM);
1205 static void perf_event_fd_array_put_ptr(void *ptr)
1207 bpf_event_entry_free_rcu(ptr);
1210 static void perf_event_fd_array_release(struct bpf_map *map,
1211 struct file *map_file)
1213 struct bpf_array *array = container_of(map, struct bpf_array, map);
1214 struct bpf_event_entry *ee;
1217 if (map->map_flags & BPF_F_PRESERVE_ELEMS)
1221 for (i = 0; i < array->map.max_entries; i++) {
1222 ee = READ_ONCE(array->ptrs[i]);
1223 if (ee && ee->map_file == map_file)
1224 fd_array_map_delete_elem(map, &i);
1229 static void perf_event_fd_array_map_free(struct bpf_map *map)
1231 if (map->map_flags & BPF_F_PRESERVE_ELEMS)
1232 bpf_fd_array_map_clear(map);
1233 fd_array_map_free(map);
1236 const struct bpf_map_ops perf_event_array_map_ops = {
1237 .map_meta_equal = bpf_map_meta_equal,
1238 .map_alloc_check = fd_array_map_alloc_check,
1239 .map_alloc = array_map_alloc,
1240 .map_free = perf_event_fd_array_map_free,
1241 .map_get_next_key = array_map_get_next_key,
1242 .map_lookup_elem = fd_array_map_lookup_elem,
1243 .map_delete_elem = fd_array_map_delete_elem,
1244 .map_fd_get_ptr = perf_event_fd_array_get_ptr,
1245 .map_fd_put_ptr = perf_event_fd_array_put_ptr,
1246 .map_release = perf_event_fd_array_release,
1247 .map_check_btf = map_check_no_btf,
1248 .map_btf_id = &array_map_btf_ids[0],
1251 #ifdef CONFIG_CGROUPS
1252 static void *cgroup_fd_array_get_ptr(struct bpf_map *map,
1253 struct file *map_file /* not used */,
1256 return cgroup_get_from_fd(fd);
1259 static void cgroup_fd_array_put_ptr(void *ptr)
1261 /* cgroup_put free cgrp after a rcu grace period */
1265 static void cgroup_fd_array_free(struct bpf_map *map)
1267 bpf_fd_array_map_clear(map);
1268 fd_array_map_free(map);
1271 const struct bpf_map_ops cgroup_array_map_ops = {
1272 .map_meta_equal = bpf_map_meta_equal,
1273 .map_alloc_check = fd_array_map_alloc_check,
1274 .map_alloc = array_map_alloc,
1275 .map_free = cgroup_fd_array_free,
1276 .map_get_next_key = array_map_get_next_key,
1277 .map_lookup_elem = fd_array_map_lookup_elem,
1278 .map_delete_elem = fd_array_map_delete_elem,
1279 .map_fd_get_ptr = cgroup_fd_array_get_ptr,
1280 .map_fd_put_ptr = cgroup_fd_array_put_ptr,
1281 .map_check_btf = map_check_no_btf,
1282 .map_btf_id = &array_map_btf_ids[0],
1286 static struct bpf_map *array_of_map_alloc(union bpf_attr *attr)
1288 struct bpf_map *map, *inner_map_meta;
1290 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1291 if (IS_ERR(inner_map_meta))
1292 return inner_map_meta;
1294 map = array_map_alloc(attr);
1296 bpf_map_meta_free(inner_map_meta);
1300 map->inner_map_meta = inner_map_meta;
1305 static void array_of_map_free(struct bpf_map *map)
1307 /* map->inner_map_meta is only accessed by syscall which
1308 * is protected by fdget/fdput.
1310 bpf_map_meta_free(map->inner_map_meta);
1311 bpf_fd_array_map_clear(map);
1312 fd_array_map_free(map);
1315 static void *array_of_map_lookup_elem(struct bpf_map *map, void *key)
1317 struct bpf_map **inner_map = array_map_lookup_elem(map, key);
1322 return READ_ONCE(*inner_map);
1325 static int array_of_map_gen_lookup(struct bpf_map *map,
1326 struct bpf_insn *insn_buf)
1328 struct bpf_array *array = container_of(map, struct bpf_array, map);
1329 u32 elem_size = array->elem_size;
1330 struct bpf_insn *insn = insn_buf;
1331 const int ret = BPF_REG_0;
1332 const int map_ptr = BPF_REG_1;
1333 const int index = BPF_REG_2;
1335 *insn++ = BPF_ALU64_IMM(BPF_ADD, map_ptr, offsetof(struct bpf_array, value));
1336 *insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
1337 if (!map->bypass_spec_v1) {
1338 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 6);
1339 *insn++ = BPF_ALU32_IMM(BPF_AND, ret, array->index_mask);
1341 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 5);
1343 if (is_power_of_2(elem_size))
1344 *insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(elem_size));
1346 *insn++ = BPF_ALU64_IMM(BPF_MUL, ret, elem_size);
1347 *insn++ = BPF_ALU64_REG(BPF_ADD, ret, map_ptr);
1348 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1349 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
1350 *insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
1351 *insn++ = BPF_MOV64_IMM(ret, 0);
1353 return insn - insn_buf;
1356 const struct bpf_map_ops array_of_maps_map_ops = {
1357 .map_alloc_check = fd_array_map_alloc_check,
1358 .map_alloc = array_of_map_alloc,
1359 .map_free = array_of_map_free,
1360 .map_get_next_key = array_map_get_next_key,
1361 .map_lookup_elem = array_of_map_lookup_elem,
1362 .map_delete_elem = fd_array_map_delete_elem,
1363 .map_fd_get_ptr = bpf_map_fd_get_ptr,
1364 .map_fd_put_ptr = bpf_map_fd_put_ptr,
1365 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1366 .map_gen_lookup = array_of_map_gen_lookup,
1367 .map_lookup_batch = generic_map_lookup_batch,
1368 .map_update_batch = generic_map_update_batch,
1369 .map_check_btf = map_check_no_btf,
1370 .map_btf_id = &array_map_btf_ids[0],