1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2024 Meta Platforms, Inc. and affiliates. */
3 #include <linux/bpf.h>
4 #include <linux/btf.h>
5 #include <linux/err.h>
6 #include <linux/btf_ids.h>
7 #include <linux/vmalloc.h>
8 #include <linux/pagemap.h>
9
10 /*
11 * bpf_arena is a sparsely populated shared memory region between bpf program and
12 * user space process.
13 *
14 * For example on x86-64 the values could be:
15 * user_vm_start 7f7d26200000 // picked by mmap()
16 * kern_vm_start ffffc90001e69000 // picked by get_vm_area()
17 * For user space all pointers within the arena are normal 8-byte addresses.
18 * In this example 7f7d26200000 is the address of the first page (pgoff=0).
19 * The bpf program will access it as: kern_vm_start + lower_32bit_of_user_ptr
20 * (u32)7f7d26200000 -> 26200000
21 * hence
22 * ffffc90001e69000 + 26200000 == ffffc90028069000 is "pgoff=0" within 4Gb
23 * kernel memory region.
24 *
25 * BPF JITs generate the following code to access arena:
26 * mov eax, eax // eax has lower 32-bit of user pointer
27 * mov word ptr [rax + r12 + off], bx
28 * where r12 == kern_vm_start and off is s16.
29 * Hence allocate 4Gb + GUARD_SZ/2 on each side.
30 *
31 * Initially kernel vm_area and user vma are not populated.
32 * User space can fault-in any address which will insert the page
33 * into kernel and user vma.
34 * bpf program can allocate a page via bpf_arena_alloc_pages() kfunc
35 * which will insert it into kernel vm_area.
36 * The later fault-in from user space will populate that page into user vma.
37 */
38
39 /* number of bytes addressable by LDX/STX insn with 16-bit 'off' field */
40 #define GUARD_SZ (1ull << sizeof_field(struct bpf_insn, off) * 8)
41 #define KERN_VM_SZ (SZ_4G + GUARD_SZ)
42
43 struct bpf_arena {
44 struct bpf_map map;
45 u64 user_vm_start;
46 u64 user_vm_end;
47 struct vm_struct *kern_vm;
48 struct maple_tree mt;
49 struct list_head vma_list;
50 struct mutex lock;
51 };
52
53 u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena)
54 {
55 return arena ? (u64) (long) arena->kern_vm->addr + GUARD_SZ / 2 : 0;
56 }
57
58 u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena)
59 {
60 return arena ? arena->user_vm_start : 0;
61 }
62
63 static long arena_map_peek_elem(struct bpf_map *map, void *value)
64 {
65 return -EOPNOTSUPP;
66 }
67
68 static long arena_map_push_elem(struct bpf_map *map, void *value, u64 flags)
69 {
70 return -EOPNOTSUPP;
71 }
72
73 static long arena_map_pop_elem(struct bpf_map *map, void *value)
74 {
75 return -EOPNOTSUPP;
76 }
77
78 static long arena_map_delete_elem(struct bpf_map *map, void *value)
79 {
80 return -EOPNOTSUPP;
81 }
82
83 static int arena_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
84 {
85 return -EOPNOTSUPP;
86 }
87
88 static long compute_pgoff(struct bpf_arena *arena, long uaddr)
89 {
90 return (u32)(uaddr - (u32)arena->user_vm_start) >> PAGE_SHIFT;
91 }
92
93 static struct bpf_map *arena_map_alloc(union bpf_attr *attr)
94 {
95 struct vm_struct *kern_vm;
96 int numa_node = bpf_map_attr_numa_node(attr);
97 struct bpf_arena *arena;
98 u64 vm_range;
99 int err = -ENOMEM;
100
101 if (attr->key_size || attr->value_size || attr->max_entries == 0 ||
102 /* BPF_F_MMAPABLE must be set */
103 !(attr->map_flags & BPF_F_MMAPABLE) ||
104 /* No unsupported flags present */
105 (attr->map_flags & ~(BPF_F_SEGV_ON_FAULT | BPF_F_MMAPABLE | BPF_F_NO_USER_CONV)))
106 return ERR_PTR(-EINVAL);
107
108 if (attr->map_extra & ~PAGE_MASK)
109 /* If non-zero the map_extra is an expected user VMA start address */
110 return ERR_PTR(-EINVAL);
111
112 vm_range = (u64)attr->max_entries * PAGE_SIZE;
113 if (vm_range > SZ_4G)
114 return ERR_PTR(-E2BIG);
115
116 if ((attr->map_extra >> 32) != ((attr->map_extra + vm_range - 1) >> 32))
117 /* user vma must not cross 32-bit boundary */
118 return ERR_PTR(-ERANGE);
119
120 kern_vm = get_vm_area(KERN_VM_SZ, VM_SPARSE | VM_USERMAP);
121 if (!kern_vm)
122 return ERR_PTR(-ENOMEM);
123
124 arena = bpf_map_area_alloc(sizeof(*arena), numa_node);
125 if (!arena)
126 goto err;
127
128 arena->kern_vm = kern_vm;
129 arena->user_vm_start = attr->map_extra;
130 if (arena->user_vm_start)
131 arena->user_vm_end = arena->user_vm_start + vm_range;
132
133 INIT_LIST_HEAD(&arena->vma_list);
134 bpf_map_init_from_attr(&arena->map, attr);
135 mt_init_flags(&arena->mt, MT_FLAGS_ALLOC_RANGE);
136 mutex_init(&arena->lock);
137
138 return &arena->map;
139 err:
140 free_vm_area(kern_vm);
141 return ERR_PTR(err);
142 }
143
144 static int existing_page_cb(pte_t *ptep, unsigned long addr, void *data)
145 {
146 struct page *page;
147 pte_t pte;
148
149 pte = ptep_get(ptep);
150 if (!pte_present(pte)) /* sanity check */
151 return 0;
152 page = pte_page(pte);
153 /*
154 * We do not update pte here:
155 * 1. Nobody should be accessing bpf_arena's range outside of a kernel bug
156 * 2. TLB flushing is batched or deferred. Even if we clear pte,
157 * the TLB entries can stick around and continue to permit access to
158 * the freed page. So it all relies on 1.
159 */
160 __free_page(page);
161 return 0;
162 }
163
164 static void arena_map_free(struct bpf_map *map)
165 {
166 struct bpf_arena *arena = container_of(map, struct bpf_arena, map);
167
168 /*
169 * Check that user vma-s are not around when bpf map is freed.
170 * mmap() holds vm_file which holds bpf_map refcnt.
171 * munmap() must have happened on vma followed by arena_vm_close()
172 * which would clear arena->vma_list.
173 */
174 if (WARN_ON_ONCE(!list_empty(&arena->vma_list)))
175 return;
176
177 /*
178 * free_vm_area() calls remove_vm_area() that calls free_unmap_vmap_area().
179 * It unmaps everything from vmalloc area and clears pgtables.
180 * Call apply_to_existing_page_range() first to find populated ptes and
181 * free those pages.
182 */
183 apply_to_existing_page_range(&init_mm, bpf_arena_get_kern_vm_start(arena),
184 KERN_VM_SZ - GUARD_SZ, existing_page_cb, NULL);
185 free_vm_area(arena->kern_vm);
186 mtree_destroy(&arena->mt);
187 bpf_map_area_free(arena);
188 }
189
190 static void *arena_map_lookup_elem(struct bpf_map *map, void *key)
191 {
192 return ERR_PTR(-EINVAL);
193 }
194
195 static long arena_map_update_elem(struct bpf_map *map, void *key,
196 void *value, u64 flags)
197 {
198 return -EOPNOTSUPP;
199 }
200
201 static int arena_map_check_btf(const struct bpf_map *map, const struct btf *btf,
202 const struct btf_type *key_type, const struct btf_type *value_type)
203 {
204 return 0;
205 }
206
207 static u64 arena_map_mem_usage(const struct bpf_map *map)
208 {
209 return 0;
210 }
211
212 struct vma_list {
213 struct vm_area_struct *vma;
214 struct list_head head;
215 atomic_t mmap_count;
216 };
217
218 static int remember_vma(struct bpf_arena *arena, struct vm_area_struct *vma)
219 {
220 struct vma_list *vml;
221
222 vml = kmalloc(sizeof(*vml), GFP_KERNEL);
223 if (!vml)
224 return -ENOMEM;
225 atomic_set(&vml->mmap_count, 1);
226 vma->vm_private_data = vml;
227 vml->vma = vma;
228 list_add(&vml->head, &arena->vma_list);
229 return 0;
230 }
231
232 static void arena_vm_open(struct vm_area_struct *vma)
233 {
234 struct vma_list *vml = vma->vm_private_data;
235
236 atomic_inc(&vml->mmap_count);
237 }
238
239 static void arena_vm_close(struct vm_area_struct *vma)
240 {
241 struct bpf_map *map = vma->vm_file->private_data;
242 struct bpf_arena *arena = container_of(map, struct bpf_arena, map);
243 struct vma_list *vml = vma->vm_private_data;
244
245 if (!atomic_dec_and_test(&vml->mmap_count))
246 return;
247 guard(mutex)(&arena->lock);
248 /* update link list under lock */
249 list_del(&vml->head);
250 vma->vm_private_data = NULL;
251 kfree(vml);
252 }
253
254 #define MT_ENTRY ((void *)&arena_map_ops) /* unused. has to be valid pointer */
255
256 static vm_fault_t arena_vm_fault(struct vm_fault *vmf)
257 {
258 struct bpf_map *map = vmf->vma->vm_file->private_data;
259 struct bpf_arena *arena = container_of(map, struct bpf_arena, map);
260 struct page *page;
261 long kbase, kaddr;
262 int ret;
263
264 kbase = bpf_arena_get_kern_vm_start(arena);
265 kaddr = kbase + (u32)(vmf->address);
266
267 guard(mutex)(&arena->lock);
268 page = vmalloc_to_page((void *)kaddr);
269 if (page)
270 /* already have a page vmap-ed */
271 goto out;
272
273 if (arena->map.map_flags & BPF_F_SEGV_ON_FAULT)
274 /* User space requested to segfault when page is not allocated by bpf prog */
275 return VM_FAULT_SIGSEGV;
276
277 ret = mtree_insert(&arena->mt, vmf->pgoff, MT_ENTRY, GFP_KERNEL);
278 if (ret)
279 return VM_FAULT_SIGSEGV;
280
281 /* Account into memcg of the process that created bpf_arena */
282 ret = bpf_map_alloc_pages(map, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE, 1, &page);
283 if (ret) {
284 mtree_erase(&arena->mt, vmf->pgoff);
285 return VM_FAULT_SIGSEGV;
286 }
287
288 ret = vm_area_map_pages(arena->kern_vm, kaddr, kaddr + PAGE_SIZE, &page);
289 if (ret) {
290 mtree_erase(&arena->mt, vmf->pgoff);
291 __free_page(page);
292 return VM_FAULT_SIGSEGV;
293 }
294 out:
295 page_ref_add(page, 1);
296 vmf->page = page;
297 return 0;
298 }
299
300 static const struct vm_operations_struct arena_vm_ops = {
301 .open = arena_vm_open,
302 .close = arena_vm_close,
303 .fault = arena_vm_fault,
304 };
305
306 static unsigned long arena_get_unmapped_area(struct file *filp, unsigned long addr,
307 unsigned long len, unsigned long pgoff,
308 unsigned long flags)
309 {
310 struct bpf_map *map = filp->private_data;
311 struct bpf_arena *arena = container_of(map, struct bpf_arena, map);
312 long ret;
313
314 if (pgoff)
315 return -EINVAL;
316 if (len > SZ_4G)
317 return -E2BIG;
318
319 /* if user_vm_start was specified at arena creation time */
320 if (arena->user_vm_start) {
321 if (len > arena->user_vm_end - arena->user_vm_start)
322 return -E2BIG;
323 if (len != arena->user_vm_end - arena->user_vm_start)
324 return -EINVAL;
325 if (addr != arena->user_vm_start)
326 return -EINVAL;
327 }
328
329 ret = mm_get_unmapped_area(current->mm, filp, addr, len * 2, 0, flags);
330 if (IS_ERR_VALUE(ret))
331 return ret;
332 if ((ret >> 32) == ((ret + len - 1) >> 32))
333 return ret;
334 if (WARN_ON_ONCE(arena->user_vm_start))
335 /* checks at map creation time should prevent this */
336 return -EFAULT;
337 return round_up(ret, SZ_4G);
338 }
339
340 static int arena_map_mmap(struct bpf_map *map, struct vm_area_struct *vma)
341 {
342 struct bpf_arena *arena = container_of(map, struct bpf_arena, map);
343
344 guard(mutex)(&arena->lock);
345 if (arena->user_vm_start && arena->user_vm_start != vma->vm_start)
346 /*
347 * If map_extra was not specified at arena creation time then
348 * 1st user process can do mmap(NULL, ...) to pick user_vm_start
349 * 2nd user process must pass the same addr to mmap(addr, MAP_FIXED..);
350 * or
351 * specify addr in map_extra and
352 * use the same addr later with mmap(addr, MAP_FIXED..);
353 */
354 return -EBUSY;
355
356 if (arena->user_vm_end && arena->user_vm_end != vma->vm_end)
357 /* all user processes must have the same size of mmap-ed region */
358 return -EBUSY;
359
360 /* Earlier checks should prevent this */
361 if (WARN_ON_ONCE(vma->vm_end - vma->vm_start > SZ_4G || vma->vm_pgoff))
362 return -EFAULT;
363
364 if (remember_vma(arena, vma))
365 return -ENOMEM;
366
367 arena->user_vm_start = vma->vm_start;
368 arena->user_vm_end = vma->vm_end;
369 /*
370 * bpf_map_mmap() checks that it's being mmaped as VM_SHARED and
371 * clears VM_MAYEXEC. Set VM_DONTEXPAND as well to avoid
372 * potential change of user_vm_start.
373 */
374 vm_flags_set(vma, VM_DONTEXPAND);
375 vma->vm_ops = &arena_vm_ops;
376 return 0;
377 }
378
379 static int arena_map_direct_value_addr(const struct bpf_map *map, u64 *imm, u32 off)
380 {
381 struct bpf_arena *arena = container_of(map, struct bpf_arena, map);
382
383 if ((u64)off > arena->user_vm_end - arena->user_vm_start)
384 return -ERANGE;
385 *imm = (unsigned long)arena->user_vm_start;
386 return 0;
387 }
388
389 BTF_ID_LIST_SINGLE(bpf_arena_map_btf_ids, struct, bpf_arena)
390 const struct bpf_map_ops arena_map_ops = {
391 .map_meta_equal = bpf_map_meta_equal,
392 .map_alloc = arena_map_alloc,
393 .map_free = arena_map_free,
394 .map_direct_value_addr = arena_map_direct_value_addr,
395 .map_mmap = arena_map_mmap,
396 .map_get_unmapped_area = arena_get_unmapped_area,
397 .map_get_next_key = arena_map_get_next_key,
398 .map_push_elem = arena_map_push_elem,
399 .map_peek_elem = arena_map_peek_elem,
400 .map_pop_elem = arena_map_pop_elem,
401 .map_lookup_elem = arena_map_lookup_elem,
402 .map_update_elem = arena_map_update_elem,
403 .map_delete_elem = arena_map_delete_elem,
404 .map_check_btf = arena_map_check_btf,
405 .map_mem_usage = arena_map_mem_usage,
406 .map_btf_id = &bpf_arena_map_btf_ids[0],
407 };
408
409 static u64 clear_lo32(u64 val)
410 {
411 return val & ~(u64)~0U;
412 }
413
414 /*
415 * Allocate pages and vmap them into kernel vmalloc area.
416 * Later the pages will be mmaped into user space vma.
417 */
418 static long arena_alloc_pages(struct bpf_arena *arena, long uaddr, long page_cnt, int node_id)
419 {
420 /* user_vm_end/start are fixed before bpf prog runs */
421 long page_cnt_max = (arena->user_vm_end - arena->user_vm_start) >> PAGE_SHIFT;
422 u64 kern_vm_start = bpf_arena_get_kern_vm_start(arena);
423 struct page **pages;
424 long pgoff = 0;
425 u32 uaddr32;
426 int ret, i;
427
428 if (page_cnt > page_cnt_max)
429 return 0;
430
431 if (uaddr) {
432 if (uaddr & ~PAGE_MASK)
433 return 0;
434 pgoff = compute_pgoff(arena, uaddr);
435 if (pgoff > page_cnt_max - page_cnt)
436 /* requested address will be outside of user VMA */
437 return 0;
438 }
439
440 /* zeroing is needed, since alloc_pages_bulk_array() only fills in non-zero entries */
441 pages = kvcalloc(page_cnt, sizeof(struct page *), GFP_KERNEL);
442 if (!pages)
443 return 0;
444
445 guard(mutex)(&arena->lock);
446
447 if (uaddr)
448 ret = mtree_insert_range(&arena->mt, pgoff, pgoff + page_cnt - 1,
449 MT_ENTRY, GFP_KERNEL);
450 else
451 ret = mtree_alloc_range(&arena->mt, &pgoff, MT_ENTRY,
452 page_cnt, 0, page_cnt_max - 1, GFP_KERNEL);
453 if (ret)
454 goto out_free_pages;
455
456 ret = bpf_map_alloc_pages(&arena->map, GFP_KERNEL | __GFP_ZERO,
457 node_id, page_cnt, pages);
458 if (ret)
459 goto out;
460
461 uaddr32 = (u32)(arena->user_vm_start + pgoff * PAGE_SIZE);
462 /* Earlier checks made sure that uaddr32 + page_cnt * PAGE_SIZE - 1
463 * will not overflow 32-bit. Lower 32-bit need to represent
464 * contiguous user address range.
465 * Map these pages at kern_vm_start base.
466 * kern_vm_start + uaddr32 + page_cnt * PAGE_SIZE - 1 can overflow
467 * lower 32-bit and it's ok.
468 */
469 ret = vm_area_map_pages(arena->kern_vm, kern_vm_start + uaddr32,
470 kern_vm_start + uaddr32 + page_cnt * PAGE_SIZE, pages);
471 if (ret) {
472 for (i = 0; i < page_cnt; i++)
473 __free_page(pages[i]);
474 goto out;
475 }
476 kvfree(pages);
477 return clear_lo32(arena->user_vm_start) + uaddr32;
478 out:
479 mtree_erase(&arena->mt, pgoff);
480 out_free_pages:
481 kvfree(pages);
482 return 0;
483 }
484
485 /*
486 * If page is present in vmalloc area, unmap it from vmalloc area,
487 * unmap it from all user space vma-s,
488 * and free it.
489 */
490 static void zap_pages(struct bpf_arena *arena, long uaddr, long page_cnt)
491 {
492 struct vma_list *vml;
493
494 list_for_each_entry(vml, &arena->vma_list, head)
495 zap_page_range_single(vml->vma, uaddr,
496 PAGE_SIZE * page_cnt, NULL);
497 }
498
499 static void arena_free_pages(struct bpf_arena *arena, long uaddr, long page_cnt)
500 {
501 u64 full_uaddr, uaddr_end;
502 long kaddr, pgoff, i;
503 struct page *page;
504
505 /* only aligned lower 32-bit are relevant */
506 uaddr = (u32)uaddr;
507 uaddr &= PAGE_MASK;
508 full_uaddr = clear_lo32(arena->user_vm_start) + uaddr;
509 uaddr_end = min(arena->user_vm_end, full_uaddr + (page_cnt << PAGE_SHIFT));
510 if (full_uaddr >= uaddr_end)
511 return;
512
513 page_cnt = (uaddr_end - full_uaddr) >> PAGE_SHIFT;
514
515 guard(mutex)(&arena->lock);
516
517 pgoff = compute_pgoff(arena, uaddr);
518 /* clear range */
519 mtree_store_range(&arena->mt, pgoff, pgoff + page_cnt - 1, NULL, GFP_KERNEL);
520
521 if (page_cnt > 1)
522 /* bulk zap if multiple pages being freed */
523 zap_pages(arena, full_uaddr, page_cnt);
524
525 kaddr = bpf_arena_get_kern_vm_start(arena) + uaddr;
526 for (i = 0; i < page_cnt; i++, kaddr += PAGE_SIZE, full_uaddr += PAGE_SIZE) {
527 page = vmalloc_to_page((void *)kaddr);
528 if (!page)
529 continue;
530 if (page_cnt == 1 && page_mapped(page)) /* mapped by some user process */
531 /* Optimization for the common case of page_cnt==1:
532 * If page wasn't mapped into some user vma there
533 * is no need to call zap_pages which is slow. When
534 * page_cnt is big it's faster to do the batched zap.
535 */
536 zap_pages(arena, full_uaddr, 1);
537 vm_area_unmap_pages(arena->kern_vm, kaddr, kaddr + PAGE_SIZE);
538 __free_page(page);
539 }
540 }
541
542 __bpf_kfunc_start_defs();
543
544 __bpf_kfunc void *bpf_arena_alloc_pages(void *p__map, void *addr__ign, u32 page_cnt,
545 int node_id, u64 flags)
546 {
547 struct bpf_map *map = p__map;
548 struct bpf_arena *arena = container_of(map, struct bpf_arena, map);
549
550 if (map->map_type != BPF_MAP_TYPE_ARENA || flags || !page_cnt)
551 return NULL;
552
553 return (void *)arena_alloc_pages(arena, (long)addr__ign, page_cnt, node_id);
554 }
555
556 __bpf_kfunc void bpf_arena_free_pages(void *p__map, void *ptr__ign, u32 page_cnt)
557 {
558 struct bpf_map *map = p__map;
559 struct bpf_arena *arena = container_of(map, struct bpf_arena, map);
560
561 if (map->map_type != BPF_MAP_TYPE_ARENA || !page_cnt || !ptr__ign)
562 return;
563 arena_free_pages(arena, (long)ptr__ign, page_cnt);
564 }
565 __bpf_kfunc_end_defs();
566
567 BTF_KFUNCS_START(arena_kfuncs)
568 BTF_ID_FLAGS(func, bpf_arena_alloc_pages, KF_TRUSTED_ARGS | KF_SLEEPABLE)
569 BTF_ID_FLAGS(func, bpf_arena_free_pages, KF_TRUSTED_ARGS | KF_SLEEPABLE)
570 BTF_KFUNCS_END(arena_kfuncs)
571
572 static const struct btf_kfunc_id_set common_kfunc_set = {
573 .owner = THIS_MODULE,
574 .set = &arena_kfuncs,
575 };
576
577 static int __init kfunc_init(void)
578 {
579 return register_btf_kfunc_id_set(BPF_PROG_TYPE_UNSPEC, &common_kfunc_set);
580 }
581 late_initcall(kfunc_init);
582
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.