Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/* bpf/cpumap.c
3 *
4 * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
5 */
6
7/**
8 * DOC: cpu map
9 * The 'cpumap' is primarily used as a backend map for XDP BPF helper
10 * call bpf_redirect_map() and XDP_REDIRECT action, like 'devmap'.
11 *
12 * Unlike devmap which redirects XDP frames out to another NIC device,
13 * this map type redirects raw XDP frames to another CPU. The remote
14 * CPU will do SKB-allocation and call the normal network stack.
15 */
16/*
17 * This is a scalability and isolation mechanism, that allow
18 * separating the early driver network XDP layer, from the rest of the
19 * netstack, and assigning dedicated CPUs for this stage. This
20 * basically allows for 10G wirespeed pre-filtering via bpf.
21 */
22#include <linux/bitops.h>
23#include <linux/bpf.h>
24#include <linux/filter.h>
25#include <linux/ptr_ring.h>
26#include <net/xdp.h>
27
28#include <linux/sched.h>
29#include <linux/workqueue.h>
30#include <linux/kthread.h>
31#include <linux/capability.h>
32#include <trace/events/xdp.h>
33#include <linux/btf_ids.h>
34
35#include <linux/netdevice.h> /* netif_receive_skb_list */
36#include <linux/etherdevice.h> /* eth_type_trans */
37
38/* General idea: XDP packets getting XDP redirected to another CPU,
39 * will maximum be stored/queued for one driver ->poll() call. It is
40 * guaranteed that queueing the frame and the flush operation happen on
41 * same CPU. Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
42 * which queue in bpf_cpu_map_entry contains packets.
43 */
44
45#define CPU_MAP_BULK_SIZE 8 /* 8 == one cacheline on 64-bit archs */
46struct bpf_cpu_map_entry;
47struct bpf_cpu_map;
48
49struct xdp_bulk_queue {
50 void *q[CPU_MAP_BULK_SIZE];
51 struct list_head flush_node;
52 struct bpf_cpu_map_entry *obj;
53 unsigned int count;
54};
55
56/* Struct for every remote "destination" CPU in map */
57struct bpf_cpu_map_entry {
58 u32 cpu; /* kthread CPU and map index */
59 int map_id; /* Back reference to map */
60
61 /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
62 struct xdp_bulk_queue __percpu *bulkq;
63
64 struct bpf_cpu_map *cmap;
65
66 /* Queue with potential multi-producers, and single-consumer kthread */
67 struct ptr_ring *queue;
68 struct task_struct *kthread;
69
70 struct bpf_cpumap_val value;
71 struct bpf_prog *prog;
72
73 atomic_t refcnt; /* Control when this struct can be free'ed */
74 struct rcu_head rcu;
75
76 struct work_struct kthread_stop_wq;
77};
78
79struct bpf_cpu_map {
80 struct bpf_map map;
81 /* Below members specific for map type */
82 struct bpf_cpu_map_entry __rcu **cpu_map;
83};
84
85static DEFINE_PER_CPU(struct list_head, cpu_map_flush_list);
86
87static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
88{
89 u32 value_size = attr->value_size;
90 struct bpf_cpu_map *cmap;
91
92 if (!bpf_capable())
93 return ERR_PTR(-EPERM);
94
95 /* check sanity of attributes */
96 if (attr->max_entries == 0 || attr->key_size != 4 ||
97 (value_size != offsetofend(struct bpf_cpumap_val, qsize) &&
98 value_size != offsetofend(struct bpf_cpumap_val, bpf_prog.fd)) ||
99 attr->map_flags & ~BPF_F_NUMA_NODE)
100 return ERR_PTR(-EINVAL);
101
102 /* Pre-limit array size based on NR_CPUS, not final CPU check */
103 if (attr->max_entries > NR_CPUS)
104 return ERR_PTR(-E2BIG);
105
106 cmap = bpf_map_area_alloc(sizeof(*cmap), NUMA_NO_NODE);
107 if (!cmap)
108 return ERR_PTR(-ENOMEM);
109
110 bpf_map_init_from_attr(&cmap->map, attr);
111
112 /* Alloc array for possible remote "destination" CPUs */
113 cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
114 sizeof(struct bpf_cpu_map_entry *),
115 cmap->map.numa_node);
116 if (!cmap->cpu_map) {
117 bpf_map_area_free(cmap);
118 return ERR_PTR(-ENOMEM);
119 }
120
121 return &cmap->map;
122}
123
124static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
125{
126 atomic_inc(&rcpu->refcnt);
127}
128
129/* called from workqueue, to workaround syscall using preempt_disable */
130static void cpu_map_kthread_stop(struct work_struct *work)
131{
132 struct bpf_cpu_map_entry *rcpu;
133
134 rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
135
136 /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
137 * as it waits until all in-flight call_rcu() callbacks complete.
138 */
139 rcu_barrier();
140
141 /* kthread_stop will wake_up_process and wait for it to complete */
142 kthread_stop(rcpu->kthread);
143}
144
145static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
146{
147 /* The tear-down procedure should have made sure that queue is
148 * empty. See __cpu_map_entry_replace() and work-queue
149 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
150 * gracefully and warn once.
151 */
152 struct xdp_frame *xdpf;
153
154 while ((xdpf = ptr_ring_consume(ring)))
155 if (WARN_ON_ONCE(xdpf))
156 xdp_return_frame(xdpf);
157}
158
159static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
160{
161 if (atomic_dec_and_test(&rcpu->refcnt)) {
162 if (rcpu->prog)
163 bpf_prog_put(rcpu->prog);
164 /* The queue should be empty at this point */
165 __cpu_map_ring_cleanup(rcpu->queue);
166 ptr_ring_cleanup(rcpu->queue, NULL);
167 kfree(rcpu->queue);
168 kfree(rcpu);
169 }
170}
171
172static void cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
173 struct list_head *listp,
174 struct xdp_cpumap_stats *stats)
175{
176 struct sk_buff *skb, *tmp;
177 struct xdp_buff xdp;
178 u32 act;
179 int err;
180
181 list_for_each_entry_safe(skb, tmp, listp, list) {
182 act = bpf_prog_run_generic_xdp(skb, &xdp, rcpu->prog);
183 switch (act) {
184 case XDP_PASS:
185 break;
186 case XDP_REDIRECT:
187 skb_list_del_init(skb);
188 err = xdp_do_generic_redirect(skb->dev, skb, &xdp,
189 rcpu->prog);
190 if (unlikely(err)) {
191 kfree_skb(skb);
192 stats->drop++;
193 } else {
194 stats->redirect++;
195 }
196 return;
197 default:
198 bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
199 fallthrough;
200 case XDP_ABORTED:
201 trace_xdp_exception(skb->dev, rcpu->prog, act);
202 fallthrough;
203 case XDP_DROP:
204 skb_list_del_init(skb);
205 kfree_skb(skb);
206 stats->drop++;
207 return;
208 }
209 }
210}
211
212static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
213 void **frames, int n,
214 struct xdp_cpumap_stats *stats)
215{
216 struct xdp_rxq_info rxq;
217 struct xdp_buff xdp;
218 int i, nframes = 0;
219
220 xdp_set_return_frame_no_direct();
221 xdp.rxq = &rxq;
222
223 for (i = 0; i < n; i++) {
224 struct xdp_frame *xdpf = frames[i];
225 u32 act;
226 int err;
227
228 rxq.dev = xdpf->dev_rx;
229 rxq.mem = xdpf->mem;
230 /* TODO: report queue_index to xdp_rxq_info */
231
232 xdp_convert_frame_to_buff(xdpf, &xdp);
233
234 act = bpf_prog_run_xdp(rcpu->prog, &xdp);
235 switch (act) {
236 case XDP_PASS:
237 err = xdp_update_frame_from_buff(&xdp, xdpf);
238 if (err < 0) {
239 xdp_return_frame(xdpf);
240 stats->drop++;
241 } else {
242 frames[nframes++] = xdpf;
243 stats->pass++;
244 }
245 break;
246 case XDP_REDIRECT:
247 err = xdp_do_redirect(xdpf->dev_rx, &xdp,
248 rcpu->prog);
249 if (unlikely(err)) {
250 xdp_return_frame(xdpf);
251 stats->drop++;
252 } else {
253 stats->redirect++;
254 }
255 break;
256 default:
257 bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
258 fallthrough;
259 case XDP_DROP:
260 xdp_return_frame(xdpf);
261 stats->drop++;
262 break;
263 }
264 }
265
266 xdp_clear_return_frame_no_direct();
267
268 return nframes;
269}
270
271#define CPUMAP_BATCH 8
272
273static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
274 int xdp_n, struct xdp_cpumap_stats *stats,
275 struct list_head *list)
276{
277 int nframes;
278
279 if (!rcpu->prog)
280 return xdp_n;
281
282 rcu_read_lock_bh();
283
284 nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, xdp_n, stats);
285
286 if (stats->redirect)
287 xdp_do_flush();
288
289 if (unlikely(!list_empty(list)))
290 cpu_map_bpf_prog_run_skb(rcpu, list, stats);
291
292 rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
293
294 return nframes;
295}
296
297
298static int cpu_map_kthread_run(void *data)
299{
300 struct bpf_cpu_map_entry *rcpu = data;
301
302 set_current_state(TASK_INTERRUPTIBLE);
303
304 /* When kthread gives stop order, then rcpu have been disconnected
305 * from map, thus no new packets can enter. Remaining in-flight
306 * per CPU stored packets are flushed to this queue. Wait honoring
307 * kthread_stop signal until queue is empty.
308 */
309 while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
310 struct xdp_cpumap_stats stats = {}; /* zero stats */
311 unsigned int kmem_alloc_drops = 0, sched = 0;
312 gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
313 int i, n, m, nframes, xdp_n;
314 void *frames[CPUMAP_BATCH];
315 void *skbs[CPUMAP_BATCH];
316 LIST_HEAD(list);
317
318 /* Release CPU reschedule checks */
319 if (__ptr_ring_empty(rcpu->queue)) {
320 set_current_state(TASK_INTERRUPTIBLE);
321 /* Recheck to avoid lost wake-up */
322 if (__ptr_ring_empty(rcpu->queue)) {
323 schedule();
324 sched = 1;
325 } else {
326 __set_current_state(TASK_RUNNING);
327 }
328 } else {
329 sched = cond_resched();
330 }
331
332 /*
333 * The bpf_cpu_map_entry is single consumer, with this
334 * kthread CPU pinned. Lockless access to ptr_ring
335 * consume side valid as no-resize allowed of queue.
336 */
337 n = __ptr_ring_consume_batched(rcpu->queue, frames,
338 CPUMAP_BATCH);
339 for (i = 0, xdp_n = 0; i < n; i++) {
340 void *f = frames[i];
341 struct page *page;
342
343 if (unlikely(__ptr_test_bit(0, &f))) {
344 struct sk_buff *skb = f;
345
346 __ptr_clear_bit(0, &skb);
347 list_add_tail(&skb->list, &list);
348 continue;
349 }
350
351 frames[xdp_n++] = f;
352 page = virt_to_page(f);
353
354 /* Bring struct page memory area to curr CPU. Read by
355 * build_skb_around via page_is_pfmemalloc(), and when
356 * freed written by page_frag_free call.
357 */
358 prefetchw(page);
359 }
360
361 /* Support running another XDP prog on this CPU */
362 nframes = cpu_map_bpf_prog_run(rcpu, frames, xdp_n, &stats, &list);
363 if (nframes) {
364 m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, nframes, skbs);
365 if (unlikely(m == 0)) {
366 for (i = 0; i < nframes; i++)
367 skbs[i] = NULL; /* effect: xdp_return_frame */
368 kmem_alloc_drops += nframes;
369 }
370 }
371
372 local_bh_disable();
373 for (i = 0; i < nframes; i++) {
374 struct xdp_frame *xdpf = frames[i];
375 struct sk_buff *skb = skbs[i];
376
377 skb = __xdp_build_skb_from_frame(xdpf, skb,
378 xdpf->dev_rx);
379 if (!skb) {
380 xdp_return_frame(xdpf);
381 continue;
382 }
383
384 list_add_tail(&skb->list, &list);
385 }
386 netif_receive_skb_list(&list);
387
388 /* Feedback loop via tracepoint */
389 trace_xdp_cpumap_kthread(rcpu->map_id, n, kmem_alloc_drops,
390 sched, &stats);
391
392 local_bh_enable(); /* resched point, may call do_softirq() */
393 }
394 __set_current_state(TASK_RUNNING);
395
396 put_cpu_map_entry(rcpu);
397 return 0;
398}
399
400static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu,
401 struct bpf_map *map, int fd)
402{
403 struct bpf_prog *prog;
404
405 prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
406 if (IS_ERR(prog))
407 return PTR_ERR(prog);
408
409 if (prog->expected_attach_type != BPF_XDP_CPUMAP ||
410 !bpf_prog_map_compatible(map, prog)) {
411 bpf_prog_put(prog);
412 return -EINVAL;
413 }
414
415 rcpu->value.bpf_prog.id = prog->aux->id;
416 rcpu->prog = prog;
417
418 return 0;
419}
420
421static struct bpf_cpu_map_entry *
422__cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value,
423 u32 cpu)
424{
425 int numa, err, i, fd = value->bpf_prog.fd;
426 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
427 struct bpf_cpu_map_entry *rcpu;
428 struct xdp_bulk_queue *bq;
429
430 /* Have map->numa_node, but choose node of redirect target CPU */
431 numa = cpu_to_node(cpu);
432
433 rcpu = bpf_map_kmalloc_node(map, sizeof(*rcpu), gfp | __GFP_ZERO, numa);
434 if (!rcpu)
435 return NULL;
436
437 /* Alloc percpu bulkq */
438 rcpu->bulkq = bpf_map_alloc_percpu(map, sizeof(*rcpu->bulkq),
439 sizeof(void *), gfp);
440 if (!rcpu->bulkq)
441 goto free_rcu;
442
443 for_each_possible_cpu(i) {
444 bq = per_cpu_ptr(rcpu->bulkq, i);
445 bq->obj = rcpu;
446 }
447
448 /* Alloc queue */
449 rcpu->queue = bpf_map_kmalloc_node(map, sizeof(*rcpu->queue), gfp,
450 numa);
451 if (!rcpu->queue)
452 goto free_bulkq;
453
454 err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
455 if (err)
456 goto free_queue;
457
458 rcpu->cpu = cpu;
459 rcpu->map_id = map->id;
460 rcpu->value.qsize = value->qsize;
461
462 if (fd > 0 && __cpu_map_load_bpf_program(rcpu, map, fd))
463 goto free_ptr_ring;
464
465 /* Setup kthread */
466 rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
467 "cpumap/%d/map:%d", cpu,
468 map->id);
469 if (IS_ERR(rcpu->kthread))
470 goto free_prog;
471
472 get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
473 get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
474
475 /* Make sure kthread runs on a single CPU */
476 kthread_bind(rcpu->kthread, cpu);
477 wake_up_process(rcpu->kthread);
478
479 return rcpu;
480
481free_prog:
482 if (rcpu->prog)
483 bpf_prog_put(rcpu->prog);
484free_ptr_ring:
485 ptr_ring_cleanup(rcpu->queue, NULL);
486free_queue:
487 kfree(rcpu->queue);
488free_bulkq:
489 free_percpu(rcpu->bulkq);
490free_rcu:
491 kfree(rcpu);
492 return NULL;
493}
494
495static void __cpu_map_entry_free(struct rcu_head *rcu)
496{
497 struct bpf_cpu_map_entry *rcpu;
498
499 /* This cpu_map_entry have been disconnected from map and one
500 * RCU grace-period have elapsed. Thus, XDP cannot queue any
501 * new packets and cannot change/set flush_needed that can
502 * find this entry.
503 */
504 rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
505
506 free_percpu(rcpu->bulkq);
507 /* Cannot kthread_stop() here, last put free rcpu resources */
508 put_cpu_map_entry(rcpu);
509}
510
511/* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
512 * ensure any driver rcu critical sections have completed, but this
513 * does not guarantee a flush has happened yet. Because driver side
514 * rcu_read_lock/unlock only protects the running XDP program. The
515 * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
516 * pending flush op doesn't fail.
517 *
518 * The bpf_cpu_map_entry is still used by the kthread, and there can
519 * still be pending packets (in queue and percpu bulkq). A refcnt
520 * makes sure to last user (kthread_stop vs. call_rcu) free memory
521 * resources.
522 *
523 * The rcu callback __cpu_map_entry_free flush remaining packets in
524 * percpu bulkq to queue. Due to caller map_delete_elem() disable
525 * preemption, cannot call kthread_stop() to make sure queue is empty.
526 * Instead a work_queue is started for stopping kthread,
527 * cpu_map_kthread_stop, which waits for an RCU grace period before
528 * stopping kthread, emptying the queue.
529 */
530static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
531 u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
532{
533 struct bpf_cpu_map_entry *old_rcpu;
534
535 old_rcpu = unrcu_pointer(xchg(&cmap->cpu_map[key_cpu], RCU_INITIALIZER(rcpu)));
536 if (old_rcpu) {
537 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
538 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
539 schedule_work(&old_rcpu->kthread_stop_wq);
540 }
541}
542
543static int cpu_map_delete_elem(struct bpf_map *map, void *key)
544{
545 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
546 u32 key_cpu = *(u32 *)key;
547
548 if (key_cpu >= map->max_entries)
549 return -EINVAL;
550
551 /* notice caller map_delete_elem() use preempt_disable() */
552 __cpu_map_entry_replace(cmap, key_cpu, NULL);
553 return 0;
554}
555
556static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
557 u64 map_flags)
558{
559 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
560 struct bpf_cpumap_val cpumap_value = {};
561 struct bpf_cpu_map_entry *rcpu;
562 /* Array index key correspond to CPU number */
563 u32 key_cpu = *(u32 *)key;
564
565 memcpy(&cpumap_value, value, map->value_size);
566
567 if (unlikely(map_flags > BPF_EXIST))
568 return -EINVAL;
569 if (unlikely(key_cpu >= cmap->map.max_entries))
570 return -E2BIG;
571 if (unlikely(map_flags == BPF_NOEXIST))
572 return -EEXIST;
573 if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
574 return -EOVERFLOW;
575
576 /* Make sure CPU is a valid possible cpu */
577 if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
578 return -ENODEV;
579
580 if (cpumap_value.qsize == 0) {
581 rcpu = NULL; /* Same as deleting */
582 } else {
583 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
584 rcpu = __cpu_map_entry_alloc(map, &cpumap_value, key_cpu);
585 if (!rcpu)
586 return -ENOMEM;
587 rcpu->cmap = cmap;
588 }
589 rcu_read_lock();
590 __cpu_map_entry_replace(cmap, key_cpu, rcpu);
591 rcu_read_unlock();
592 return 0;
593}
594
595static void cpu_map_free(struct bpf_map *map)
596{
597 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
598 u32 i;
599
600 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
601 * so the bpf programs (can be more than one that used this map) were
602 * disconnected from events. Wait for outstanding critical sections in
603 * these programs to complete. The rcu critical section only guarantees
604 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
605 * It does __not__ ensure pending flush operations (if any) are
606 * complete.
607 */
608
609 synchronize_rcu();
610
611 /* For cpu_map the remote CPUs can still be using the entries
612 * (struct bpf_cpu_map_entry).
613 */
614 for (i = 0; i < cmap->map.max_entries; i++) {
615 struct bpf_cpu_map_entry *rcpu;
616
617 rcpu = rcu_dereference_raw(cmap->cpu_map[i]);
618 if (!rcpu)
619 continue;
620
621 /* bq flush and cleanup happens after RCU grace-period */
622 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
623 }
624 bpf_map_area_free(cmap->cpu_map);
625 bpf_map_area_free(cmap);
626}
627
628/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
629 * by local_bh_disable() (from XDP calls inside NAPI). The
630 * rcu_read_lock_bh_held() below makes lockdep accept both.
631 */
632static void *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
633{
634 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
635 struct bpf_cpu_map_entry *rcpu;
636
637 if (key >= map->max_entries)
638 return NULL;
639
640 rcpu = rcu_dereference_check(cmap->cpu_map[key],
641 rcu_read_lock_bh_held());
642 return rcpu;
643}
644
645static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
646{
647 struct bpf_cpu_map_entry *rcpu =
648 __cpu_map_lookup_elem(map, *(u32 *)key);
649
650 return rcpu ? &rcpu->value : NULL;
651}
652
653static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
654{
655 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
656 u32 index = key ? *(u32 *)key : U32_MAX;
657 u32 *next = next_key;
658
659 if (index >= cmap->map.max_entries) {
660 *next = 0;
661 return 0;
662 }
663
664 if (index == cmap->map.max_entries - 1)
665 return -ENOENT;
666 *next = index + 1;
667 return 0;
668}
669
670static int cpu_map_redirect(struct bpf_map *map, u64 index, u64 flags)
671{
672 return __bpf_xdp_redirect_map(map, index, flags, 0,
673 __cpu_map_lookup_elem);
674}
675
676BTF_ID_LIST_SINGLE(cpu_map_btf_ids, struct, bpf_cpu_map)
677const struct bpf_map_ops cpu_map_ops = {
678 .map_meta_equal = bpf_map_meta_equal,
679 .map_alloc = cpu_map_alloc,
680 .map_free = cpu_map_free,
681 .map_delete_elem = cpu_map_delete_elem,
682 .map_update_elem = cpu_map_update_elem,
683 .map_lookup_elem = cpu_map_lookup_elem,
684 .map_get_next_key = cpu_map_get_next_key,
685 .map_check_btf = map_check_no_btf,
686 .map_btf_id = &cpu_map_btf_ids[0],
687 .map_redirect = cpu_map_redirect,
688};
689
690static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
691{
692 struct bpf_cpu_map_entry *rcpu = bq->obj;
693 unsigned int processed = 0, drops = 0;
694 const int to_cpu = rcpu->cpu;
695 struct ptr_ring *q;
696 int i;
697
698 if (unlikely(!bq->count))
699 return;
700
701 q = rcpu->queue;
702 spin_lock(&q->producer_lock);
703
704 for (i = 0; i < bq->count; i++) {
705 struct xdp_frame *xdpf = bq->q[i];
706 int err;
707
708 err = __ptr_ring_produce(q, xdpf);
709 if (err) {
710 drops++;
711 xdp_return_frame_rx_napi(xdpf);
712 }
713 processed++;
714 }
715 bq->count = 0;
716 spin_unlock(&q->producer_lock);
717
718 __list_del_clearprev(&bq->flush_node);
719
720 /* Feedback loop via tracepoints */
721 trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
722}
723
724/* Runs under RCU-read-side, plus in softirq under NAPI protection.
725 * Thus, safe percpu variable access.
726 */
727static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
728{
729 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
730 struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
731
732 if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
733 bq_flush_to_queue(bq);
734
735 /* Notice, xdp_buff/page MUST be queued here, long enough for
736 * driver to code invoking us to finished, due to driver
737 * (e.g. ixgbe) recycle tricks based on page-refcnt.
738 *
739 * Thus, incoming xdp_frame is always queued here (else we race
740 * with another CPU on page-refcnt and remaining driver code).
741 * Queue time is very short, as driver will invoke flush
742 * operation, when completing napi->poll call.
743 */
744 bq->q[bq->count++] = xdpf;
745
746 if (!bq->flush_node.prev)
747 list_add(&bq->flush_node, flush_list);
748}
749
750int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf,
751 struct net_device *dev_rx)
752{
753 /* Info needed when constructing SKB on remote CPU */
754 xdpf->dev_rx = dev_rx;
755
756 bq_enqueue(rcpu, xdpf);
757 return 0;
758}
759
760int cpu_map_generic_redirect(struct bpf_cpu_map_entry *rcpu,
761 struct sk_buff *skb)
762{
763 int ret;
764
765 __skb_pull(skb, skb->mac_len);
766 skb_set_redirected(skb, false);
767 __ptr_set_bit(0, &skb);
768
769 ret = ptr_ring_produce(rcpu->queue, skb);
770 if (ret < 0)
771 goto trace;
772
773 wake_up_process(rcpu->kthread);
774trace:
775 trace_xdp_cpumap_enqueue(rcpu->map_id, !ret, !!ret, rcpu->cpu);
776 return ret;
777}
778
779void __cpu_map_flush(void)
780{
781 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
782 struct xdp_bulk_queue *bq, *tmp;
783
784 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
785 bq_flush_to_queue(bq);
786
787 /* If already running, costs spin_lock_irqsave + smb_mb */
788 wake_up_process(bq->obj->kthread);
789 }
790}
791
792static int __init cpu_map_init(void)
793{
794 int cpu;
795
796 for_each_possible_cpu(cpu)
797 INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu));
798 return 0;
799}
800
801subsys_initcall(cpu_map_init);
1// SPDX-License-Identifier: GPL-2.0-only
2/* bpf/cpumap.c
3 *
4 * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
5 */
6
7/* The 'cpumap' is primarily used as a backend map for XDP BPF helper
8 * call bpf_redirect_map() and XDP_REDIRECT action, like 'devmap'.
9 *
10 * Unlike devmap which redirects XDP frames out another NIC device,
11 * this map type redirects raw XDP frames to another CPU. The remote
12 * CPU will do SKB-allocation and call the normal network stack.
13 *
14 * This is a scalability and isolation mechanism, that allow
15 * separating the early driver network XDP layer, from the rest of the
16 * netstack, and assigning dedicated CPUs for this stage. This
17 * basically allows for 10G wirespeed pre-filtering via bpf.
18 */
19#include <linux/bpf.h>
20#include <linux/filter.h>
21#include <linux/ptr_ring.h>
22#include <net/xdp.h>
23
24#include <linux/sched.h>
25#include <linux/workqueue.h>
26#include <linux/kthread.h>
27#include <linux/capability.h>
28#include <trace/events/xdp.h>
29
30#include <linux/netdevice.h> /* netif_receive_skb_core */
31#include <linux/etherdevice.h> /* eth_type_trans */
32
33/* General idea: XDP packets getting XDP redirected to another CPU,
34 * will maximum be stored/queued for one driver ->poll() call. It is
35 * guaranteed that queueing the frame and the flush operation happen on
36 * same CPU. Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
37 * which queue in bpf_cpu_map_entry contains packets.
38 */
39
40#define CPU_MAP_BULK_SIZE 8 /* 8 == one cacheline on 64-bit archs */
41struct bpf_cpu_map_entry;
42struct bpf_cpu_map;
43
44struct xdp_bulk_queue {
45 void *q[CPU_MAP_BULK_SIZE];
46 struct list_head flush_node;
47 struct bpf_cpu_map_entry *obj;
48 unsigned int count;
49};
50
51/* Struct for every remote "destination" CPU in map */
52struct bpf_cpu_map_entry {
53 u32 cpu; /* kthread CPU and map index */
54 int map_id; /* Back reference to map */
55
56 /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
57 struct xdp_bulk_queue __percpu *bulkq;
58
59 struct bpf_cpu_map *cmap;
60
61 /* Queue with potential multi-producers, and single-consumer kthread */
62 struct ptr_ring *queue;
63 struct task_struct *kthread;
64
65 struct bpf_cpumap_val value;
66 struct bpf_prog *prog;
67
68 atomic_t refcnt; /* Control when this struct can be free'ed */
69 struct rcu_head rcu;
70
71 struct work_struct kthread_stop_wq;
72};
73
74struct bpf_cpu_map {
75 struct bpf_map map;
76 /* Below members specific for map type */
77 struct bpf_cpu_map_entry **cpu_map;
78};
79
80static DEFINE_PER_CPU(struct list_head, cpu_map_flush_list);
81
82static int bq_flush_to_queue(struct xdp_bulk_queue *bq);
83
84static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
85{
86 u32 value_size = attr->value_size;
87 struct bpf_cpu_map *cmap;
88 int err = -ENOMEM;
89 u64 cost;
90 int ret;
91
92 if (!bpf_capable())
93 return ERR_PTR(-EPERM);
94
95 /* check sanity of attributes */
96 if (attr->max_entries == 0 || attr->key_size != 4 ||
97 (value_size != offsetofend(struct bpf_cpumap_val, qsize) &&
98 value_size != offsetofend(struct bpf_cpumap_val, bpf_prog.fd)) ||
99 attr->map_flags & ~BPF_F_NUMA_NODE)
100 return ERR_PTR(-EINVAL);
101
102 cmap = kzalloc(sizeof(*cmap), GFP_USER);
103 if (!cmap)
104 return ERR_PTR(-ENOMEM);
105
106 bpf_map_init_from_attr(&cmap->map, attr);
107
108 /* Pre-limit array size based on NR_CPUS, not final CPU check */
109 if (cmap->map.max_entries > NR_CPUS) {
110 err = -E2BIG;
111 goto free_cmap;
112 }
113
114 /* make sure page count doesn't overflow */
115 cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *);
116
117 /* Notice returns -EPERM on if map size is larger than memlock limit */
118 ret = bpf_map_charge_init(&cmap->map.memory, cost);
119 if (ret) {
120 err = ret;
121 goto free_cmap;
122 }
123
124 /* Alloc array for possible remote "destination" CPUs */
125 cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
126 sizeof(struct bpf_cpu_map_entry *),
127 cmap->map.numa_node);
128 if (!cmap->cpu_map)
129 goto free_charge;
130
131 return &cmap->map;
132free_charge:
133 bpf_map_charge_finish(&cmap->map.memory);
134free_cmap:
135 kfree(cmap);
136 return ERR_PTR(err);
137}
138
139static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
140{
141 atomic_inc(&rcpu->refcnt);
142}
143
144/* called from workqueue, to workaround syscall using preempt_disable */
145static void cpu_map_kthread_stop(struct work_struct *work)
146{
147 struct bpf_cpu_map_entry *rcpu;
148
149 rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
150
151 /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
152 * as it waits until all in-flight call_rcu() callbacks complete.
153 */
154 rcu_barrier();
155
156 /* kthread_stop will wake_up_process and wait for it to complete */
157 kthread_stop(rcpu->kthread);
158}
159
160static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu,
161 struct xdp_frame *xdpf,
162 struct sk_buff *skb)
163{
164 unsigned int hard_start_headroom;
165 unsigned int frame_size;
166 void *pkt_data_start;
167
168 /* Part of headroom was reserved to xdpf */
169 hard_start_headroom = sizeof(struct xdp_frame) + xdpf->headroom;
170
171 /* Memory size backing xdp_frame data already have reserved
172 * room for build_skb to place skb_shared_info in tailroom.
173 */
174 frame_size = xdpf->frame_sz;
175
176 pkt_data_start = xdpf->data - hard_start_headroom;
177 skb = build_skb_around(skb, pkt_data_start, frame_size);
178 if (unlikely(!skb))
179 return NULL;
180
181 skb_reserve(skb, hard_start_headroom);
182 __skb_put(skb, xdpf->len);
183 if (xdpf->metasize)
184 skb_metadata_set(skb, xdpf->metasize);
185
186 /* Essential SKB info: protocol and skb->dev */
187 skb->protocol = eth_type_trans(skb, xdpf->dev_rx);
188
189 /* Optional SKB info, currently missing:
190 * - HW checksum info (skb->ip_summed)
191 * - HW RX hash (skb_set_hash)
192 * - RX ring dev queue index (skb_record_rx_queue)
193 */
194
195 /* Until page_pool get SKB return path, release DMA here */
196 xdp_release_frame(xdpf);
197
198 /* Allow SKB to reuse area used by xdp_frame */
199 xdp_scrub_frame(xdpf);
200
201 return skb;
202}
203
204static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
205{
206 /* The tear-down procedure should have made sure that queue is
207 * empty. See __cpu_map_entry_replace() and work-queue
208 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
209 * gracefully and warn once.
210 */
211 struct xdp_frame *xdpf;
212
213 while ((xdpf = ptr_ring_consume(ring)))
214 if (WARN_ON_ONCE(xdpf))
215 xdp_return_frame(xdpf);
216}
217
218static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
219{
220 if (atomic_dec_and_test(&rcpu->refcnt)) {
221 if (rcpu->prog)
222 bpf_prog_put(rcpu->prog);
223 /* The queue should be empty at this point */
224 __cpu_map_ring_cleanup(rcpu->queue);
225 ptr_ring_cleanup(rcpu->queue, NULL);
226 kfree(rcpu->queue);
227 kfree(rcpu);
228 }
229}
230
231static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
232 void **frames, int n,
233 struct xdp_cpumap_stats *stats)
234{
235 struct xdp_rxq_info rxq;
236 struct xdp_buff xdp;
237 int i, nframes = 0;
238
239 if (!rcpu->prog)
240 return n;
241
242 rcu_read_lock_bh();
243
244 xdp_set_return_frame_no_direct();
245 xdp.rxq = &rxq;
246
247 for (i = 0; i < n; i++) {
248 struct xdp_frame *xdpf = frames[i];
249 u32 act;
250 int err;
251
252 rxq.dev = xdpf->dev_rx;
253 rxq.mem = xdpf->mem;
254 /* TODO: report queue_index to xdp_rxq_info */
255
256 xdp_convert_frame_to_buff(xdpf, &xdp);
257
258 act = bpf_prog_run_xdp(rcpu->prog, &xdp);
259 switch (act) {
260 case XDP_PASS:
261 err = xdp_update_frame_from_buff(&xdp, xdpf);
262 if (err < 0) {
263 xdp_return_frame(xdpf);
264 stats->drop++;
265 } else {
266 frames[nframes++] = xdpf;
267 stats->pass++;
268 }
269 break;
270 case XDP_REDIRECT:
271 err = xdp_do_redirect(xdpf->dev_rx, &xdp,
272 rcpu->prog);
273 if (unlikely(err)) {
274 xdp_return_frame(xdpf);
275 stats->drop++;
276 } else {
277 stats->redirect++;
278 }
279 break;
280 default:
281 bpf_warn_invalid_xdp_action(act);
282 fallthrough;
283 case XDP_DROP:
284 xdp_return_frame(xdpf);
285 stats->drop++;
286 break;
287 }
288 }
289
290 if (stats->redirect)
291 xdp_do_flush_map();
292
293 xdp_clear_return_frame_no_direct();
294
295 rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
296
297 return nframes;
298}
299
300#define CPUMAP_BATCH 8
301
302static int cpu_map_kthread_run(void *data)
303{
304 struct bpf_cpu_map_entry *rcpu = data;
305
306 set_current_state(TASK_INTERRUPTIBLE);
307
308 /* When kthread gives stop order, then rcpu have been disconnected
309 * from map, thus no new packets can enter. Remaining in-flight
310 * per CPU stored packets are flushed to this queue. Wait honoring
311 * kthread_stop signal until queue is empty.
312 */
313 while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
314 struct xdp_cpumap_stats stats = {}; /* zero stats */
315 gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
316 unsigned int drops = 0, sched = 0;
317 void *frames[CPUMAP_BATCH];
318 void *skbs[CPUMAP_BATCH];
319 int i, n, m, nframes;
320
321 /* Release CPU reschedule checks */
322 if (__ptr_ring_empty(rcpu->queue)) {
323 set_current_state(TASK_INTERRUPTIBLE);
324 /* Recheck to avoid lost wake-up */
325 if (__ptr_ring_empty(rcpu->queue)) {
326 schedule();
327 sched = 1;
328 } else {
329 __set_current_state(TASK_RUNNING);
330 }
331 } else {
332 sched = cond_resched();
333 }
334
335 /*
336 * The bpf_cpu_map_entry is single consumer, with this
337 * kthread CPU pinned. Lockless access to ptr_ring
338 * consume side valid as no-resize allowed of queue.
339 */
340 n = __ptr_ring_consume_batched(rcpu->queue, frames,
341 CPUMAP_BATCH);
342 for (i = 0; i < n; i++) {
343 void *f = frames[i];
344 struct page *page = virt_to_page(f);
345
346 /* Bring struct page memory area to curr CPU. Read by
347 * build_skb_around via page_is_pfmemalloc(), and when
348 * freed written by page_frag_free call.
349 */
350 prefetchw(page);
351 }
352
353 /* Support running another XDP prog on this CPU */
354 nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, n, &stats);
355 if (nframes) {
356 m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, nframes, skbs);
357 if (unlikely(m == 0)) {
358 for (i = 0; i < nframes; i++)
359 skbs[i] = NULL; /* effect: xdp_return_frame */
360 drops += nframes;
361 }
362 }
363
364 local_bh_disable();
365 for (i = 0; i < nframes; i++) {
366 struct xdp_frame *xdpf = frames[i];
367 struct sk_buff *skb = skbs[i];
368 int ret;
369
370 skb = cpu_map_build_skb(rcpu, xdpf, skb);
371 if (!skb) {
372 xdp_return_frame(xdpf);
373 continue;
374 }
375
376 /* Inject into network stack */
377 ret = netif_receive_skb_core(skb);
378 if (ret == NET_RX_DROP)
379 drops++;
380 }
381 /* Feedback loop via tracepoint */
382 trace_xdp_cpumap_kthread(rcpu->map_id, n, drops, sched, &stats);
383
384 local_bh_enable(); /* resched point, may call do_softirq() */
385 }
386 __set_current_state(TASK_RUNNING);
387
388 put_cpu_map_entry(rcpu);
389 return 0;
390}
391
392bool cpu_map_prog_allowed(struct bpf_map *map)
393{
394 return map->map_type == BPF_MAP_TYPE_CPUMAP &&
395 map->value_size != offsetofend(struct bpf_cpumap_val, qsize);
396}
397
398static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu, int fd)
399{
400 struct bpf_prog *prog;
401
402 prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
403 if (IS_ERR(prog))
404 return PTR_ERR(prog);
405
406 if (prog->expected_attach_type != BPF_XDP_CPUMAP) {
407 bpf_prog_put(prog);
408 return -EINVAL;
409 }
410
411 rcpu->value.bpf_prog.id = prog->aux->id;
412 rcpu->prog = prog;
413
414 return 0;
415}
416
417static struct bpf_cpu_map_entry *
418__cpu_map_entry_alloc(struct bpf_cpumap_val *value, u32 cpu, int map_id)
419{
420 int numa, err, i, fd = value->bpf_prog.fd;
421 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
422 struct bpf_cpu_map_entry *rcpu;
423 struct xdp_bulk_queue *bq;
424
425 /* Have map->numa_node, but choose node of redirect target CPU */
426 numa = cpu_to_node(cpu);
427
428 rcpu = kzalloc_node(sizeof(*rcpu), gfp, numa);
429 if (!rcpu)
430 return NULL;
431
432 /* Alloc percpu bulkq */
433 rcpu->bulkq = __alloc_percpu_gfp(sizeof(*rcpu->bulkq),
434 sizeof(void *), gfp);
435 if (!rcpu->bulkq)
436 goto free_rcu;
437
438 for_each_possible_cpu(i) {
439 bq = per_cpu_ptr(rcpu->bulkq, i);
440 bq->obj = rcpu;
441 }
442
443 /* Alloc queue */
444 rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
445 if (!rcpu->queue)
446 goto free_bulkq;
447
448 err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
449 if (err)
450 goto free_queue;
451
452 rcpu->cpu = cpu;
453 rcpu->map_id = map_id;
454 rcpu->value.qsize = value->qsize;
455
456 if (fd > 0 && __cpu_map_load_bpf_program(rcpu, fd))
457 goto free_ptr_ring;
458
459 /* Setup kthread */
460 rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
461 "cpumap/%d/map:%d", cpu, map_id);
462 if (IS_ERR(rcpu->kthread))
463 goto free_prog;
464
465 get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
466 get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
467
468 /* Make sure kthread runs on a single CPU */
469 kthread_bind(rcpu->kthread, cpu);
470 wake_up_process(rcpu->kthread);
471
472 return rcpu;
473
474free_prog:
475 if (rcpu->prog)
476 bpf_prog_put(rcpu->prog);
477free_ptr_ring:
478 ptr_ring_cleanup(rcpu->queue, NULL);
479free_queue:
480 kfree(rcpu->queue);
481free_bulkq:
482 free_percpu(rcpu->bulkq);
483free_rcu:
484 kfree(rcpu);
485 return NULL;
486}
487
488static void __cpu_map_entry_free(struct rcu_head *rcu)
489{
490 struct bpf_cpu_map_entry *rcpu;
491
492 /* This cpu_map_entry have been disconnected from map and one
493 * RCU grace-period have elapsed. Thus, XDP cannot queue any
494 * new packets and cannot change/set flush_needed that can
495 * find this entry.
496 */
497 rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
498
499 free_percpu(rcpu->bulkq);
500 /* Cannot kthread_stop() here, last put free rcpu resources */
501 put_cpu_map_entry(rcpu);
502}
503
504/* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
505 * ensure any driver rcu critical sections have completed, but this
506 * does not guarantee a flush has happened yet. Because driver side
507 * rcu_read_lock/unlock only protects the running XDP program. The
508 * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
509 * pending flush op doesn't fail.
510 *
511 * The bpf_cpu_map_entry is still used by the kthread, and there can
512 * still be pending packets (in queue and percpu bulkq). A refcnt
513 * makes sure to last user (kthread_stop vs. call_rcu) free memory
514 * resources.
515 *
516 * The rcu callback __cpu_map_entry_free flush remaining packets in
517 * percpu bulkq to queue. Due to caller map_delete_elem() disable
518 * preemption, cannot call kthread_stop() to make sure queue is empty.
519 * Instead a work_queue is started for stopping kthread,
520 * cpu_map_kthread_stop, which waits for an RCU grace period before
521 * stopping kthread, emptying the queue.
522 */
523static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
524 u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
525{
526 struct bpf_cpu_map_entry *old_rcpu;
527
528 old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu);
529 if (old_rcpu) {
530 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
531 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
532 schedule_work(&old_rcpu->kthread_stop_wq);
533 }
534}
535
536static int cpu_map_delete_elem(struct bpf_map *map, void *key)
537{
538 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
539 u32 key_cpu = *(u32 *)key;
540
541 if (key_cpu >= map->max_entries)
542 return -EINVAL;
543
544 /* notice caller map_delete_elem() use preempt_disable() */
545 __cpu_map_entry_replace(cmap, key_cpu, NULL);
546 return 0;
547}
548
549static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
550 u64 map_flags)
551{
552 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
553 struct bpf_cpumap_val cpumap_value = {};
554 struct bpf_cpu_map_entry *rcpu;
555 /* Array index key correspond to CPU number */
556 u32 key_cpu = *(u32 *)key;
557
558 memcpy(&cpumap_value, value, map->value_size);
559
560 if (unlikely(map_flags > BPF_EXIST))
561 return -EINVAL;
562 if (unlikely(key_cpu >= cmap->map.max_entries))
563 return -E2BIG;
564 if (unlikely(map_flags == BPF_NOEXIST))
565 return -EEXIST;
566 if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
567 return -EOVERFLOW;
568
569 /* Make sure CPU is a valid possible cpu */
570 if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
571 return -ENODEV;
572
573 if (cpumap_value.qsize == 0) {
574 rcpu = NULL; /* Same as deleting */
575 } else {
576 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
577 rcpu = __cpu_map_entry_alloc(&cpumap_value, key_cpu, map->id);
578 if (!rcpu)
579 return -ENOMEM;
580 rcpu->cmap = cmap;
581 }
582 rcu_read_lock();
583 __cpu_map_entry_replace(cmap, key_cpu, rcpu);
584 rcu_read_unlock();
585 return 0;
586}
587
588static void cpu_map_free(struct bpf_map *map)
589{
590 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
591 u32 i;
592
593 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
594 * so the bpf programs (can be more than one that used this map) were
595 * disconnected from events. Wait for outstanding critical sections in
596 * these programs to complete. The rcu critical section only guarantees
597 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
598 * It does __not__ ensure pending flush operations (if any) are
599 * complete.
600 */
601
602 bpf_clear_redirect_map(map);
603 synchronize_rcu();
604
605 /* For cpu_map the remote CPUs can still be using the entries
606 * (struct bpf_cpu_map_entry).
607 */
608 for (i = 0; i < cmap->map.max_entries; i++) {
609 struct bpf_cpu_map_entry *rcpu;
610
611 rcpu = READ_ONCE(cmap->cpu_map[i]);
612 if (!rcpu)
613 continue;
614
615 /* bq flush and cleanup happens after RCU grace-period */
616 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
617 }
618 bpf_map_area_free(cmap->cpu_map);
619 kfree(cmap);
620}
621
622struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
623{
624 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
625 struct bpf_cpu_map_entry *rcpu;
626
627 if (key >= map->max_entries)
628 return NULL;
629
630 rcpu = READ_ONCE(cmap->cpu_map[key]);
631 return rcpu;
632}
633
634static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
635{
636 struct bpf_cpu_map_entry *rcpu =
637 __cpu_map_lookup_elem(map, *(u32 *)key);
638
639 return rcpu ? &rcpu->value : NULL;
640}
641
642static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
643{
644 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
645 u32 index = key ? *(u32 *)key : U32_MAX;
646 u32 *next = next_key;
647
648 if (index >= cmap->map.max_entries) {
649 *next = 0;
650 return 0;
651 }
652
653 if (index == cmap->map.max_entries - 1)
654 return -ENOENT;
655 *next = index + 1;
656 return 0;
657}
658
659static int cpu_map_btf_id;
660const struct bpf_map_ops cpu_map_ops = {
661 .map_alloc = cpu_map_alloc,
662 .map_free = cpu_map_free,
663 .map_delete_elem = cpu_map_delete_elem,
664 .map_update_elem = cpu_map_update_elem,
665 .map_lookup_elem = cpu_map_lookup_elem,
666 .map_get_next_key = cpu_map_get_next_key,
667 .map_check_btf = map_check_no_btf,
668 .map_btf_name = "bpf_cpu_map",
669 .map_btf_id = &cpu_map_btf_id,
670};
671
672static int bq_flush_to_queue(struct xdp_bulk_queue *bq)
673{
674 struct bpf_cpu_map_entry *rcpu = bq->obj;
675 unsigned int processed = 0, drops = 0;
676 const int to_cpu = rcpu->cpu;
677 struct ptr_ring *q;
678 int i;
679
680 if (unlikely(!bq->count))
681 return 0;
682
683 q = rcpu->queue;
684 spin_lock(&q->producer_lock);
685
686 for (i = 0; i < bq->count; i++) {
687 struct xdp_frame *xdpf = bq->q[i];
688 int err;
689
690 err = __ptr_ring_produce(q, xdpf);
691 if (err) {
692 drops++;
693 xdp_return_frame_rx_napi(xdpf);
694 }
695 processed++;
696 }
697 bq->count = 0;
698 spin_unlock(&q->producer_lock);
699
700 __list_del_clearprev(&bq->flush_node);
701
702 /* Feedback loop via tracepoints */
703 trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
704 return 0;
705}
706
707/* Runs under RCU-read-side, plus in softirq under NAPI protection.
708 * Thus, safe percpu variable access.
709 */
710static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
711{
712 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
713 struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
714
715 if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
716 bq_flush_to_queue(bq);
717
718 /* Notice, xdp_buff/page MUST be queued here, long enough for
719 * driver to code invoking us to finished, due to driver
720 * (e.g. ixgbe) recycle tricks based on page-refcnt.
721 *
722 * Thus, incoming xdp_frame is always queued here (else we race
723 * with another CPU on page-refcnt and remaining driver code).
724 * Queue time is very short, as driver will invoke flush
725 * operation, when completing napi->poll call.
726 */
727 bq->q[bq->count++] = xdpf;
728
729 if (!bq->flush_node.prev)
730 list_add(&bq->flush_node, flush_list);
731
732 return 0;
733}
734
735int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
736 struct net_device *dev_rx)
737{
738 struct xdp_frame *xdpf;
739
740 xdpf = xdp_convert_buff_to_frame(xdp);
741 if (unlikely(!xdpf))
742 return -EOVERFLOW;
743
744 /* Info needed when constructing SKB on remote CPU */
745 xdpf->dev_rx = dev_rx;
746
747 bq_enqueue(rcpu, xdpf);
748 return 0;
749}
750
751void __cpu_map_flush(void)
752{
753 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
754 struct xdp_bulk_queue *bq, *tmp;
755
756 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
757 bq_flush_to_queue(bq);
758
759 /* If already running, costs spin_lock_irqsave + smb_mb */
760 wake_up_process(bq->obj->kthread);
761 }
762}
763
764static int __init cpu_map_init(void)
765{
766 int cpu;
767
768 for_each_possible_cpu(cpu)
769 INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu));
770 return 0;
771}
772
773subsys_initcall(cpu_map_init);