Merge tag 'pinctrl-v6.1-1' of git://git.kernel.org/pub/scm/linux/kernel/git/linusw...
[platform/kernel/linux-rpi.git] / kernel / bpf / ringbuf.c
1 #include <linux/bpf.h>
2 #include <linux/btf.h>
3 #include <linux/err.h>
4 #include <linux/irq_work.h>
5 #include <linux/slab.h>
6 #include <linux/filter.h>
7 #include <linux/mm.h>
8 #include <linux/vmalloc.h>
9 #include <linux/wait.h>
10 #include <linux/poll.h>
11 #include <linux/kmemleak.h>
12 #include <uapi/linux/btf.h>
13 #include <linux/btf_ids.h>
14
15 #define RINGBUF_CREATE_FLAG_MASK (BPF_F_NUMA_NODE)
16
17 /* non-mmap()'able part of bpf_ringbuf (everything up to consumer page) */
18 #define RINGBUF_PGOFF \
19         (offsetof(struct bpf_ringbuf, consumer_pos) >> PAGE_SHIFT)
20 /* consumer page and producer page */
21 #define RINGBUF_POS_PAGES 2
22
23 #define RINGBUF_MAX_RECORD_SZ (UINT_MAX/4)
24
25 /* Maximum size of ring buffer area is limited by 32-bit page offset within
26  * record header, counted in pages. Reserve 8 bits for extensibility, and take
27  * into account few extra pages for consumer/producer pages and
28  * non-mmap()'able parts. This gives 64GB limit, which seems plenty for single
29  * ring buffer.
30  */
31 #define RINGBUF_MAX_DATA_SZ \
32         (((1ULL << 24) - RINGBUF_POS_PAGES - RINGBUF_PGOFF) * PAGE_SIZE)
33
34 struct bpf_ringbuf {
35         wait_queue_head_t waitq;
36         struct irq_work work;
37         u64 mask;
38         struct page **pages;
39         int nr_pages;
40         spinlock_t spinlock ____cacheline_aligned_in_smp;
41         /* For user-space producer ring buffers, an atomic_t busy bit is used
42          * to synchronize access to the ring buffers in the kernel, rather than
43          * the spinlock that is used for kernel-producer ring buffers. This is
44          * done because the ring buffer must hold a lock across a BPF program's
45          * callback:
46          *
47          *    __bpf_user_ringbuf_peek() // lock acquired
48          * -> program callback_fn()
49          * -> __bpf_user_ringbuf_sample_release() // lock released
50          *
51          * It is unsafe and incorrect to hold an IRQ spinlock across what could
52          * be a long execution window, so we instead simply disallow concurrent
53          * access to the ring buffer by kernel consumers, and return -EBUSY from
54          * __bpf_user_ringbuf_peek() if the busy bit is held by another task.
55          */
56         atomic_t busy ____cacheline_aligned_in_smp;
57         /* Consumer and producer counters are put into separate pages to
58          * allow each position to be mapped with different permissions.
59          * This prevents a user-space application from modifying the
60          * position and ruining in-kernel tracking. The permissions of the
61          * pages depend on who is producing samples: user-space or the
62          * kernel.
63          *
64          * Kernel-producer
65          * ---------------
66          * The producer position and data pages are mapped as r/o in
67          * userspace. For this approach, bits in the header of samples are
68          * used to signal to user-space, and to other producers, whether a
69          * sample is currently being written.
70          *
71          * User-space producer
72          * -------------------
73          * Only the page containing the consumer position is mapped r/o in
74          * user-space. User-space producers also use bits of the header to
75          * communicate to the kernel, but the kernel must carefully check and
76          * validate each sample to ensure that they're correctly formatted, and
77          * fully contained within the ring buffer.
78          */
79         unsigned long consumer_pos __aligned(PAGE_SIZE);
80         unsigned long producer_pos __aligned(PAGE_SIZE);
81         char data[] __aligned(PAGE_SIZE);
82 };
83
84 struct bpf_ringbuf_map {
85         struct bpf_map map;
86         struct bpf_ringbuf *rb;
87 };
88
89 /* 8-byte ring buffer record header structure */
90 struct bpf_ringbuf_hdr {
91         u32 len;
92         u32 pg_off;
93 };
94
95 static struct bpf_ringbuf *bpf_ringbuf_area_alloc(size_t data_sz, int numa_node)
96 {
97         const gfp_t flags = GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL |
98                             __GFP_NOWARN | __GFP_ZERO;
99         int nr_meta_pages = RINGBUF_PGOFF + RINGBUF_POS_PAGES;
100         int nr_data_pages = data_sz >> PAGE_SHIFT;
101         int nr_pages = nr_meta_pages + nr_data_pages;
102         struct page **pages, *page;
103         struct bpf_ringbuf *rb;
104         size_t array_size;
105         int i;
106
107         /* Each data page is mapped twice to allow "virtual"
108          * continuous read of samples wrapping around the end of ring
109          * buffer area:
110          * ------------------------------------------------------
111          * | meta pages |  real data pages  |  same data pages  |
112          * ------------------------------------------------------
113          * |            | 1 2 3 4 5 6 7 8 9 | 1 2 3 4 5 6 7 8 9 |
114          * ------------------------------------------------------
115          * |            | TA             DA | TA             DA |
116          * ------------------------------------------------------
117          *                               ^^^^^^^
118          *                                  |
119          * Here, no need to worry about special handling of wrapped-around
120          * data due to double-mapped data pages. This works both in kernel and
121          * when mmap()'ed in user-space, simplifying both kernel and
122          * user-space implementations significantly.
123          */
124         array_size = (nr_meta_pages + 2 * nr_data_pages) * sizeof(*pages);
125         pages = bpf_map_area_alloc(array_size, numa_node);
126         if (!pages)
127                 return NULL;
128
129         for (i = 0; i < nr_pages; i++) {
130                 page = alloc_pages_node(numa_node, flags, 0);
131                 if (!page) {
132                         nr_pages = i;
133                         goto err_free_pages;
134                 }
135                 pages[i] = page;
136                 if (i >= nr_meta_pages)
137                         pages[nr_data_pages + i] = page;
138         }
139
140         rb = vmap(pages, nr_meta_pages + 2 * nr_data_pages,
141                   VM_MAP | VM_USERMAP, PAGE_KERNEL);
142         if (rb) {
143                 kmemleak_not_leak(pages);
144                 rb->pages = pages;
145                 rb->nr_pages = nr_pages;
146                 return rb;
147         }
148
149 err_free_pages:
150         for (i = 0; i < nr_pages; i++)
151                 __free_page(pages[i]);
152         bpf_map_area_free(pages);
153         return NULL;
154 }
155
156 static void bpf_ringbuf_notify(struct irq_work *work)
157 {
158         struct bpf_ringbuf *rb = container_of(work, struct bpf_ringbuf, work);
159
160         wake_up_all(&rb->waitq);
161 }
162
163 static struct bpf_ringbuf *bpf_ringbuf_alloc(size_t data_sz, int numa_node)
164 {
165         struct bpf_ringbuf *rb;
166
167         rb = bpf_ringbuf_area_alloc(data_sz, numa_node);
168         if (!rb)
169                 return NULL;
170
171         spin_lock_init(&rb->spinlock);
172         atomic_set(&rb->busy, 0);
173         init_waitqueue_head(&rb->waitq);
174         init_irq_work(&rb->work, bpf_ringbuf_notify);
175
176         rb->mask = data_sz - 1;
177         rb->consumer_pos = 0;
178         rb->producer_pos = 0;
179
180         return rb;
181 }
182
183 static struct bpf_map *ringbuf_map_alloc(union bpf_attr *attr)
184 {
185         struct bpf_ringbuf_map *rb_map;
186
187         if (attr->map_flags & ~RINGBUF_CREATE_FLAG_MASK)
188                 return ERR_PTR(-EINVAL);
189
190         if (attr->key_size || attr->value_size ||
191             !is_power_of_2(attr->max_entries) ||
192             !PAGE_ALIGNED(attr->max_entries))
193                 return ERR_PTR(-EINVAL);
194
195 #ifdef CONFIG_64BIT
196         /* on 32-bit arch, it's impossible to overflow record's hdr->pgoff */
197         if (attr->max_entries > RINGBUF_MAX_DATA_SZ)
198                 return ERR_PTR(-E2BIG);
199 #endif
200
201         rb_map = bpf_map_area_alloc(sizeof(*rb_map), NUMA_NO_NODE);
202         if (!rb_map)
203                 return ERR_PTR(-ENOMEM);
204
205         bpf_map_init_from_attr(&rb_map->map, attr);
206
207         rb_map->rb = bpf_ringbuf_alloc(attr->max_entries, rb_map->map.numa_node);
208         if (!rb_map->rb) {
209                 bpf_map_area_free(rb_map);
210                 return ERR_PTR(-ENOMEM);
211         }
212
213         return &rb_map->map;
214 }
215
216 static void bpf_ringbuf_free(struct bpf_ringbuf *rb)
217 {
218         /* copy pages pointer and nr_pages to local variable, as we are going
219          * to unmap rb itself with vunmap() below
220          */
221         struct page **pages = rb->pages;
222         int i, nr_pages = rb->nr_pages;
223
224         vunmap(rb);
225         for (i = 0; i < nr_pages; i++)
226                 __free_page(pages[i]);
227         bpf_map_area_free(pages);
228 }
229
230 static void ringbuf_map_free(struct bpf_map *map)
231 {
232         struct bpf_ringbuf_map *rb_map;
233
234         rb_map = container_of(map, struct bpf_ringbuf_map, map);
235         bpf_ringbuf_free(rb_map->rb);
236         bpf_map_area_free(rb_map);
237 }
238
239 static void *ringbuf_map_lookup_elem(struct bpf_map *map, void *key)
240 {
241         return ERR_PTR(-ENOTSUPP);
242 }
243
244 static int ringbuf_map_update_elem(struct bpf_map *map, void *key, void *value,
245                                    u64 flags)
246 {
247         return -ENOTSUPP;
248 }
249
250 static int ringbuf_map_delete_elem(struct bpf_map *map, void *key)
251 {
252         return -ENOTSUPP;
253 }
254
255 static int ringbuf_map_get_next_key(struct bpf_map *map, void *key,
256                                     void *next_key)
257 {
258         return -ENOTSUPP;
259 }
260
261 static int ringbuf_map_mmap_kern(struct bpf_map *map, struct vm_area_struct *vma)
262 {
263         struct bpf_ringbuf_map *rb_map;
264
265         rb_map = container_of(map, struct bpf_ringbuf_map, map);
266
267         if (vma->vm_flags & VM_WRITE) {
268                 /* allow writable mapping for the consumer_pos only */
269                 if (vma->vm_pgoff != 0 || vma->vm_end - vma->vm_start != PAGE_SIZE)
270                         return -EPERM;
271         } else {
272                 vma->vm_flags &= ~VM_MAYWRITE;
273         }
274         /* remap_vmalloc_range() checks size and offset constraints */
275         return remap_vmalloc_range(vma, rb_map->rb,
276                                    vma->vm_pgoff + RINGBUF_PGOFF);
277 }
278
279 static int ringbuf_map_mmap_user(struct bpf_map *map, struct vm_area_struct *vma)
280 {
281         struct bpf_ringbuf_map *rb_map;
282
283         rb_map = container_of(map, struct bpf_ringbuf_map, map);
284
285         if (vma->vm_flags & VM_WRITE) {
286                 if (vma->vm_pgoff == 0)
287                         /* Disallow writable mappings to the consumer pointer,
288                          * and allow writable mappings to both the producer
289                          * position, and the ring buffer data itself.
290                          */
291                         return -EPERM;
292         } else {
293                 vma->vm_flags &= ~VM_MAYWRITE;
294         }
295         /* remap_vmalloc_range() checks size and offset constraints */
296         return remap_vmalloc_range(vma, rb_map->rb, vma->vm_pgoff + RINGBUF_PGOFF);
297 }
298
299 static unsigned long ringbuf_avail_data_sz(struct bpf_ringbuf *rb)
300 {
301         unsigned long cons_pos, prod_pos;
302
303         cons_pos = smp_load_acquire(&rb->consumer_pos);
304         prod_pos = smp_load_acquire(&rb->producer_pos);
305         return prod_pos - cons_pos;
306 }
307
308 static u32 ringbuf_total_data_sz(const struct bpf_ringbuf *rb)
309 {
310         return rb->mask + 1;
311 }
312
313 static __poll_t ringbuf_map_poll_kern(struct bpf_map *map, struct file *filp,
314                                       struct poll_table_struct *pts)
315 {
316         struct bpf_ringbuf_map *rb_map;
317
318         rb_map = container_of(map, struct bpf_ringbuf_map, map);
319         poll_wait(filp, &rb_map->rb->waitq, pts);
320
321         if (ringbuf_avail_data_sz(rb_map->rb))
322                 return EPOLLIN | EPOLLRDNORM;
323         return 0;
324 }
325
326 static __poll_t ringbuf_map_poll_user(struct bpf_map *map, struct file *filp,
327                                       struct poll_table_struct *pts)
328 {
329         struct bpf_ringbuf_map *rb_map;
330
331         rb_map = container_of(map, struct bpf_ringbuf_map, map);
332         poll_wait(filp, &rb_map->rb->waitq, pts);
333
334         if (ringbuf_avail_data_sz(rb_map->rb) < ringbuf_total_data_sz(rb_map->rb))
335                 return EPOLLOUT | EPOLLWRNORM;
336         return 0;
337 }
338
339 BTF_ID_LIST_SINGLE(ringbuf_map_btf_ids, struct, bpf_ringbuf_map)
340 const struct bpf_map_ops ringbuf_map_ops = {
341         .map_meta_equal = bpf_map_meta_equal,
342         .map_alloc = ringbuf_map_alloc,
343         .map_free = ringbuf_map_free,
344         .map_mmap = ringbuf_map_mmap_kern,
345         .map_poll = ringbuf_map_poll_kern,
346         .map_lookup_elem = ringbuf_map_lookup_elem,
347         .map_update_elem = ringbuf_map_update_elem,
348         .map_delete_elem = ringbuf_map_delete_elem,
349         .map_get_next_key = ringbuf_map_get_next_key,
350         .map_btf_id = &ringbuf_map_btf_ids[0],
351 };
352
353 BTF_ID_LIST_SINGLE(user_ringbuf_map_btf_ids, struct, bpf_ringbuf_map)
354 const struct bpf_map_ops user_ringbuf_map_ops = {
355         .map_meta_equal = bpf_map_meta_equal,
356         .map_alloc = ringbuf_map_alloc,
357         .map_free = ringbuf_map_free,
358         .map_mmap = ringbuf_map_mmap_user,
359         .map_poll = ringbuf_map_poll_user,
360         .map_lookup_elem = ringbuf_map_lookup_elem,
361         .map_update_elem = ringbuf_map_update_elem,
362         .map_delete_elem = ringbuf_map_delete_elem,
363         .map_get_next_key = ringbuf_map_get_next_key,
364         .map_btf_id = &user_ringbuf_map_btf_ids[0],
365 };
366
367 /* Given pointer to ring buffer record metadata and struct bpf_ringbuf itself,
368  * calculate offset from record metadata to ring buffer in pages, rounded
369  * down. This page offset is stored as part of record metadata and allows to
370  * restore struct bpf_ringbuf * from record pointer. This page offset is
371  * stored at offset 4 of record metadata header.
372  */
373 static size_t bpf_ringbuf_rec_pg_off(struct bpf_ringbuf *rb,
374                                      struct bpf_ringbuf_hdr *hdr)
375 {
376         return ((void *)hdr - (void *)rb) >> PAGE_SHIFT;
377 }
378
379 /* Given pointer to ring buffer record header, restore pointer to struct
380  * bpf_ringbuf itself by using page offset stored at offset 4
381  */
382 static struct bpf_ringbuf *
383 bpf_ringbuf_restore_from_rec(struct bpf_ringbuf_hdr *hdr)
384 {
385         unsigned long addr = (unsigned long)(void *)hdr;
386         unsigned long off = (unsigned long)hdr->pg_off << PAGE_SHIFT;
387
388         return (void*)((addr & PAGE_MASK) - off);
389 }
390
391 static void *__bpf_ringbuf_reserve(struct bpf_ringbuf *rb, u64 size)
392 {
393         unsigned long cons_pos, prod_pos, new_prod_pos, flags;
394         u32 len, pg_off;
395         struct bpf_ringbuf_hdr *hdr;
396
397         if (unlikely(size > RINGBUF_MAX_RECORD_SZ))
398                 return NULL;
399
400         len = round_up(size + BPF_RINGBUF_HDR_SZ, 8);
401         if (len > ringbuf_total_data_sz(rb))
402                 return NULL;
403
404         cons_pos = smp_load_acquire(&rb->consumer_pos);
405
406         if (in_nmi()) {
407                 if (!spin_trylock_irqsave(&rb->spinlock, flags))
408                         return NULL;
409         } else {
410                 spin_lock_irqsave(&rb->spinlock, flags);
411         }
412
413         prod_pos = rb->producer_pos;
414         new_prod_pos = prod_pos + len;
415
416         /* check for out of ringbuf space by ensuring producer position
417          * doesn't advance more than (ringbuf_size - 1) ahead
418          */
419         if (new_prod_pos - cons_pos > rb->mask) {
420                 spin_unlock_irqrestore(&rb->spinlock, flags);
421                 return NULL;
422         }
423
424         hdr = (void *)rb->data + (prod_pos & rb->mask);
425         pg_off = bpf_ringbuf_rec_pg_off(rb, hdr);
426         hdr->len = size | BPF_RINGBUF_BUSY_BIT;
427         hdr->pg_off = pg_off;
428
429         /* pairs with consumer's smp_load_acquire() */
430         smp_store_release(&rb->producer_pos, new_prod_pos);
431
432         spin_unlock_irqrestore(&rb->spinlock, flags);
433
434         return (void *)hdr + BPF_RINGBUF_HDR_SZ;
435 }
436
437 BPF_CALL_3(bpf_ringbuf_reserve, struct bpf_map *, map, u64, size, u64, flags)
438 {
439         struct bpf_ringbuf_map *rb_map;
440
441         if (unlikely(flags))
442                 return 0;
443
444         rb_map = container_of(map, struct bpf_ringbuf_map, map);
445         return (unsigned long)__bpf_ringbuf_reserve(rb_map->rb, size);
446 }
447
448 const struct bpf_func_proto bpf_ringbuf_reserve_proto = {
449         .func           = bpf_ringbuf_reserve,
450         .ret_type       = RET_PTR_TO_ALLOC_MEM_OR_NULL,
451         .arg1_type      = ARG_CONST_MAP_PTR,
452         .arg2_type      = ARG_CONST_ALLOC_SIZE_OR_ZERO,
453         .arg3_type      = ARG_ANYTHING,
454 };
455
456 static void bpf_ringbuf_commit(void *sample, u64 flags, bool discard)
457 {
458         unsigned long rec_pos, cons_pos;
459         struct bpf_ringbuf_hdr *hdr;
460         struct bpf_ringbuf *rb;
461         u32 new_len;
462
463         hdr = sample - BPF_RINGBUF_HDR_SZ;
464         rb = bpf_ringbuf_restore_from_rec(hdr);
465         new_len = hdr->len ^ BPF_RINGBUF_BUSY_BIT;
466         if (discard)
467                 new_len |= BPF_RINGBUF_DISCARD_BIT;
468
469         /* update record header with correct final size prefix */
470         xchg(&hdr->len, new_len);
471
472         /* if consumer caught up and is waiting for our record, notify about
473          * new data availability
474          */
475         rec_pos = (void *)hdr - (void *)rb->data;
476         cons_pos = smp_load_acquire(&rb->consumer_pos) & rb->mask;
477
478         if (flags & BPF_RB_FORCE_WAKEUP)
479                 irq_work_queue(&rb->work);
480         else if (cons_pos == rec_pos && !(flags & BPF_RB_NO_WAKEUP))
481                 irq_work_queue(&rb->work);
482 }
483
484 BPF_CALL_2(bpf_ringbuf_submit, void *, sample, u64, flags)
485 {
486         bpf_ringbuf_commit(sample, flags, false /* discard */);
487         return 0;
488 }
489
490 const struct bpf_func_proto bpf_ringbuf_submit_proto = {
491         .func           = bpf_ringbuf_submit,
492         .ret_type       = RET_VOID,
493         .arg1_type      = ARG_PTR_TO_ALLOC_MEM | OBJ_RELEASE,
494         .arg2_type      = ARG_ANYTHING,
495 };
496
497 BPF_CALL_2(bpf_ringbuf_discard, void *, sample, u64, flags)
498 {
499         bpf_ringbuf_commit(sample, flags, true /* discard */);
500         return 0;
501 }
502
503 const struct bpf_func_proto bpf_ringbuf_discard_proto = {
504         .func           = bpf_ringbuf_discard,
505         .ret_type       = RET_VOID,
506         .arg1_type      = ARG_PTR_TO_ALLOC_MEM | OBJ_RELEASE,
507         .arg2_type      = ARG_ANYTHING,
508 };
509
510 BPF_CALL_4(bpf_ringbuf_output, struct bpf_map *, map, void *, data, u64, size,
511            u64, flags)
512 {
513         struct bpf_ringbuf_map *rb_map;
514         void *rec;
515
516         if (unlikely(flags & ~(BPF_RB_NO_WAKEUP | BPF_RB_FORCE_WAKEUP)))
517                 return -EINVAL;
518
519         rb_map = container_of(map, struct bpf_ringbuf_map, map);
520         rec = __bpf_ringbuf_reserve(rb_map->rb, size);
521         if (!rec)
522                 return -EAGAIN;
523
524         memcpy(rec, data, size);
525         bpf_ringbuf_commit(rec, flags, false /* discard */);
526         return 0;
527 }
528
529 const struct bpf_func_proto bpf_ringbuf_output_proto = {
530         .func           = bpf_ringbuf_output,
531         .ret_type       = RET_INTEGER,
532         .arg1_type      = ARG_CONST_MAP_PTR,
533         .arg2_type      = ARG_PTR_TO_MEM | MEM_RDONLY,
534         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
535         .arg4_type      = ARG_ANYTHING,
536 };
537
538 BPF_CALL_2(bpf_ringbuf_query, struct bpf_map *, map, u64, flags)
539 {
540         struct bpf_ringbuf *rb;
541
542         rb = container_of(map, struct bpf_ringbuf_map, map)->rb;
543
544         switch (flags) {
545         case BPF_RB_AVAIL_DATA:
546                 return ringbuf_avail_data_sz(rb);
547         case BPF_RB_RING_SIZE:
548                 return ringbuf_total_data_sz(rb);
549         case BPF_RB_CONS_POS:
550                 return smp_load_acquire(&rb->consumer_pos);
551         case BPF_RB_PROD_POS:
552                 return smp_load_acquire(&rb->producer_pos);
553         default:
554                 return 0;
555         }
556 }
557
558 const struct bpf_func_proto bpf_ringbuf_query_proto = {
559         .func           = bpf_ringbuf_query,
560         .ret_type       = RET_INTEGER,
561         .arg1_type      = ARG_CONST_MAP_PTR,
562         .arg2_type      = ARG_ANYTHING,
563 };
564
565 BPF_CALL_4(bpf_ringbuf_reserve_dynptr, struct bpf_map *, map, u32, size, u64, flags,
566            struct bpf_dynptr_kern *, ptr)
567 {
568         struct bpf_ringbuf_map *rb_map;
569         void *sample;
570         int err;
571
572         if (unlikely(flags)) {
573                 bpf_dynptr_set_null(ptr);
574                 return -EINVAL;
575         }
576
577         err = bpf_dynptr_check_size(size);
578         if (err) {
579                 bpf_dynptr_set_null(ptr);
580                 return err;
581         }
582
583         rb_map = container_of(map, struct bpf_ringbuf_map, map);
584
585         sample = __bpf_ringbuf_reserve(rb_map->rb, size);
586         if (!sample) {
587                 bpf_dynptr_set_null(ptr);
588                 return -EINVAL;
589         }
590
591         bpf_dynptr_init(ptr, sample, BPF_DYNPTR_TYPE_RINGBUF, 0, size);
592
593         return 0;
594 }
595
596 const struct bpf_func_proto bpf_ringbuf_reserve_dynptr_proto = {
597         .func           = bpf_ringbuf_reserve_dynptr,
598         .ret_type       = RET_INTEGER,
599         .arg1_type      = ARG_CONST_MAP_PTR,
600         .arg2_type      = ARG_ANYTHING,
601         .arg3_type      = ARG_ANYTHING,
602         .arg4_type      = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | MEM_UNINIT,
603 };
604
605 BPF_CALL_2(bpf_ringbuf_submit_dynptr, struct bpf_dynptr_kern *, ptr, u64, flags)
606 {
607         if (!ptr->data)
608                 return 0;
609
610         bpf_ringbuf_commit(ptr->data, flags, false /* discard */);
611
612         bpf_dynptr_set_null(ptr);
613
614         return 0;
615 }
616
617 const struct bpf_func_proto bpf_ringbuf_submit_dynptr_proto = {
618         .func           = bpf_ringbuf_submit_dynptr,
619         .ret_type       = RET_VOID,
620         .arg1_type      = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | OBJ_RELEASE,
621         .arg2_type      = ARG_ANYTHING,
622 };
623
624 BPF_CALL_2(bpf_ringbuf_discard_dynptr, struct bpf_dynptr_kern *, ptr, u64, flags)
625 {
626         if (!ptr->data)
627                 return 0;
628
629         bpf_ringbuf_commit(ptr->data, flags, true /* discard */);
630
631         bpf_dynptr_set_null(ptr);
632
633         return 0;
634 }
635
636 const struct bpf_func_proto bpf_ringbuf_discard_dynptr_proto = {
637         .func           = bpf_ringbuf_discard_dynptr,
638         .ret_type       = RET_VOID,
639         .arg1_type      = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | OBJ_RELEASE,
640         .arg2_type      = ARG_ANYTHING,
641 };
642
643 static int __bpf_user_ringbuf_peek(struct bpf_ringbuf *rb, void **sample, u32 *size)
644 {
645         int err;
646         u32 hdr_len, sample_len, total_len, flags, *hdr;
647         u64 cons_pos, prod_pos;
648
649         /* Synchronizes with smp_store_release() in user-space producer. */
650         prod_pos = smp_load_acquire(&rb->producer_pos);
651         if (prod_pos % 8)
652                 return -EINVAL;
653
654         /* Synchronizes with smp_store_release() in __bpf_user_ringbuf_sample_release() */
655         cons_pos = smp_load_acquire(&rb->consumer_pos);
656         if (cons_pos >= prod_pos)
657                 return -ENODATA;
658
659         hdr = (u32 *)((uintptr_t)rb->data + (uintptr_t)(cons_pos & rb->mask));
660         /* Synchronizes with smp_store_release() in user-space producer. */
661         hdr_len = smp_load_acquire(hdr);
662         flags = hdr_len & (BPF_RINGBUF_BUSY_BIT | BPF_RINGBUF_DISCARD_BIT);
663         sample_len = hdr_len & ~flags;
664         total_len = round_up(sample_len + BPF_RINGBUF_HDR_SZ, 8);
665
666         /* The sample must fit within the region advertised by the producer position. */
667         if (total_len > prod_pos - cons_pos)
668                 return -EINVAL;
669
670         /* The sample must fit within the data region of the ring buffer. */
671         if (total_len > ringbuf_total_data_sz(rb))
672                 return -E2BIG;
673
674         /* The sample must fit into a struct bpf_dynptr. */
675         err = bpf_dynptr_check_size(sample_len);
676         if (err)
677                 return -E2BIG;
678
679         if (flags & BPF_RINGBUF_DISCARD_BIT) {
680                 /* If the discard bit is set, the sample should be skipped.
681                  *
682                  * Update the consumer pos, and return -EAGAIN so the caller
683                  * knows to skip this sample and try to read the next one.
684                  */
685                 smp_store_release(&rb->consumer_pos, cons_pos + total_len);
686                 return -EAGAIN;
687         }
688
689         if (flags & BPF_RINGBUF_BUSY_BIT)
690                 return -ENODATA;
691
692         *sample = (void *)((uintptr_t)rb->data +
693                            (uintptr_t)((cons_pos + BPF_RINGBUF_HDR_SZ) & rb->mask));
694         *size = sample_len;
695         return 0;
696 }
697
698 static void __bpf_user_ringbuf_sample_release(struct bpf_ringbuf *rb, size_t size, u64 flags)
699 {
700         u64 consumer_pos;
701         u32 rounded_size = round_up(size + BPF_RINGBUF_HDR_SZ, 8);
702
703         /* Using smp_load_acquire() is unnecessary here, as the busy-bit
704          * prevents another task from writing to consumer_pos after it was read
705          * by this task with smp_load_acquire() in __bpf_user_ringbuf_peek().
706          */
707         consumer_pos = rb->consumer_pos;
708          /* Synchronizes with smp_load_acquire() in user-space producer. */
709         smp_store_release(&rb->consumer_pos, consumer_pos + rounded_size);
710 }
711
712 BPF_CALL_4(bpf_user_ringbuf_drain, struct bpf_map *, map,
713            void *, callback_fn, void *, callback_ctx, u64, flags)
714 {
715         struct bpf_ringbuf *rb;
716         long samples, discarded_samples = 0, ret = 0;
717         bpf_callback_t callback = (bpf_callback_t)callback_fn;
718         u64 wakeup_flags = BPF_RB_NO_WAKEUP | BPF_RB_FORCE_WAKEUP;
719         int busy = 0;
720
721         if (unlikely(flags & ~wakeup_flags))
722                 return -EINVAL;
723
724         rb = container_of(map, struct bpf_ringbuf_map, map)->rb;
725
726         /* If another consumer is already consuming a sample, wait for them to finish. */
727         if (!atomic_try_cmpxchg(&rb->busy, &busy, 1))
728                 return -EBUSY;
729
730         for (samples = 0; samples < BPF_MAX_USER_RINGBUF_SAMPLES && ret == 0; samples++) {
731                 int err;
732                 u32 size;
733                 void *sample;
734                 struct bpf_dynptr_kern dynptr;
735
736                 err = __bpf_user_ringbuf_peek(rb, &sample, &size);
737                 if (err) {
738                         if (err == -ENODATA) {
739                                 break;
740                         } else if (err == -EAGAIN) {
741                                 discarded_samples++;
742                                 continue;
743                         } else {
744                                 ret = err;
745                                 goto schedule_work_return;
746                         }
747                 }
748
749                 bpf_dynptr_init(&dynptr, sample, BPF_DYNPTR_TYPE_LOCAL, 0, size);
750                 ret = callback((uintptr_t)&dynptr, (uintptr_t)callback_ctx, 0, 0, 0);
751                 __bpf_user_ringbuf_sample_release(rb, size, flags);
752         }
753         ret = samples - discarded_samples;
754
755 schedule_work_return:
756         /* Prevent the clearing of the busy-bit from being reordered before the
757          * storing of any rb consumer or producer positions.
758          */
759         smp_mb__before_atomic();
760         atomic_set(&rb->busy, 0);
761
762         if (flags & BPF_RB_FORCE_WAKEUP)
763                 irq_work_queue(&rb->work);
764         else if (!(flags & BPF_RB_NO_WAKEUP) && samples > 0)
765                 irq_work_queue(&rb->work);
766         return ret;
767 }
768
769 const struct bpf_func_proto bpf_user_ringbuf_drain_proto = {
770         .func           = bpf_user_ringbuf_drain,
771         .ret_type       = RET_INTEGER,
772         .arg1_type      = ARG_CONST_MAP_PTR,
773         .arg2_type      = ARG_PTR_TO_FUNC,
774         .arg3_type      = ARG_PTR_TO_STACK_OR_NULL,
775         .arg4_type      = ARG_ANYTHING,
776 };