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#include <net/hotdata.h>
28
29#include <linux/sched.h>
30#include <linux/workqueue.h>
31#include <linux/kthread.h>
32#include <linux/completion.h>
33#include <trace/events/xdp.h>
34#include <linux/btf_ids.h>
35
36#include <linux/netdevice.h>
37#include <net/gro.h>
38
39/* General idea: XDP packets getting XDP redirected to another CPU,
40 * will maximum be stored/queued for one driver ->poll() call. It is
41 * guaranteed that queueing the frame and the flush operation happen on
42 * same CPU. Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
43 * which queue in bpf_cpu_map_entry contains packets.
44 */
45
46#define CPU_MAP_BULK_SIZE 8 /* 8 == one cacheline on 64-bit archs */
47struct bpf_cpu_map_entry;
48struct bpf_cpu_map;
49
50struct xdp_bulk_queue {
51 void *q[CPU_MAP_BULK_SIZE];
52 struct list_head flush_node;
53 struct bpf_cpu_map_entry *obj;
54 unsigned int count;
55};
56
57/* Struct for every remote "destination" CPU in map */
58struct bpf_cpu_map_entry {
59 u32 cpu; /* kthread CPU and map index */
60 int map_id; /* Back reference to map */
61
62 /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
63 struct xdp_bulk_queue __percpu *bulkq;
64
65 /* Queue with potential multi-producers, and single-consumer kthread */
66 struct ptr_ring *queue;
67 struct task_struct *kthread;
68
69 struct bpf_cpumap_val value;
70 struct bpf_prog *prog;
71 struct gro_node gro;
72
73 struct completion kthread_running;
74 struct rcu_work free_work;
75};
76
77struct bpf_cpu_map {
78 struct bpf_map map;
79 /* Below members specific for map type */
80 struct bpf_cpu_map_entry __rcu **cpu_map;
81};
82
83static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
84{
85 u32 value_size = attr->value_size;
86 struct bpf_cpu_map *cmap;
87
88 /* check sanity of attributes */
89 if (attr->max_entries == 0 || attr->key_size != 4 ||
90 (value_size != offsetofend(struct bpf_cpumap_val, qsize) &&
91 value_size != offsetofend(struct bpf_cpumap_val, bpf_prog.fd)) ||
92 attr->map_flags & ~BPF_F_NUMA_NODE)
93 return ERR_PTR(error: -EINVAL);
94
95 /* Pre-limit array size based on NR_CPUS, not final CPU check */
96 if (attr->max_entries > NR_CPUS)
97 return ERR_PTR(error: -E2BIG);
98
99 cmap = bpf_map_area_alloc(size: sizeof(*cmap), NUMA_NO_NODE);
100 if (!cmap)
101 return ERR_PTR(error: -ENOMEM);
102
103 bpf_map_init_from_attr(map: &cmap->map, attr);
104
105 /* Alloc array for possible remote "destination" CPUs */
106 cmap->cpu_map = bpf_map_area_alloc(size: cmap->map.max_entries *
107 sizeof(struct bpf_cpu_map_entry *),
108 numa_node: cmap->map.numa_node);
109 if (!cmap->cpu_map) {
110 bpf_map_area_free(base: cmap);
111 return ERR_PTR(error: -ENOMEM);
112 }
113
114 return &cmap->map;
115}
116
117static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
118{
119 /* The tear-down procedure should have made sure that queue is
120 * empty. See __cpu_map_entry_replace() and work-queue
121 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
122 * gracefully and warn once.
123 */
124 void *ptr;
125
126 while ((ptr = ptr_ring_consume(r: ring))) {
127 WARN_ON_ONCE(1);
128 if (unlikely(__ptr_test_bit(0, &ptr))) {
129 __ptr_clear_bit(0, &ptr);
130 kfree_skb(skb: ptr);
131 continue;
132 }
133 xdp_return_frame(xdpf: ptr);
134 }
135}
136
137static u32 cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
138 void **skbs, u32 skb_n,
139 struct xdp_cpumap_stats *stats)
140{
141 struct xdp_buff xdp;
142 u32 act, pass = 0;
143 int err;
144
145 for (u32 i = 0; i < skb_n; i++) {
146 struct sk_buff *skb = skbs[i];
147
148 act = bpf_prog_run_generic_xdp(skb, xdp: &xdp, xdp_prog: rcpu->prog);
149 switch (act) {
150 case XDP_PASS:
151 skbs[pass++] = skb;
152 break;
153 case XDP_REDIRECT:
154 err = xdp_do_generic_redirect(dev: skb->dev, skb, xdp: &xdp,
155 prog: rcpu->prog);
156 if (unlikely(err)) {
157 kfree_skb(skb);
158 stats->drop++;
159 } else {
160 stats->redirect++;
161 }
162 break;
163 default:
164 bpf_warn_invalid_xdp_action(NULL, prog: rcpu->prog, act);
165 fallthrough;
166 case XDP_ABORTED:
167 trace_xdp_exception(dev: skb->dev, xdp: rcpu->prog, act);
168 fallthrough;
169 case XDP_DROP:
170 napi_consume_skb(skb, budget: true);
171 stats->drop++;
172 break;
173 }
174 }
175
176 stats->pass += pass;
177
178 return pass;
179}
180
181static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
182 void **frames, int n,
183 struct xdp_cpumap_stats *stats)
184{
185 struct xdp_rxq_info rxq = {};
186 struct xdp_buff xdp;
187 int i, nframes = 0;
188
189 xdp.rxq = &rxq;
190
191 for (i = 0; i < n; i++) {
192 struct xdp_frame *xdpf = frames[i];
193 u32 act;
194 int err;
195
196 rxq.dev = xdpf->dev_rx;
197 rxq.mem.type = xdpf->mem_type;
198 /* TODO: report queue_index to xdp_rxq_info */
199
200 xdp_convert_frame_to_buff(frame: xdpf, xdp: &xdp);
201
202 act = bpf_prog_run_xdp(prog: rcpu->prog, xdp: &xdp);
203 switch (act) {
204 case XDP_PASS:
205 err = xdp_update_frame_from_buff(xdp: &xdp, xdp_frame: xdpf);
206 if (err < 0) {
207 xdp_return_frame(xdpf);
208 stats->drop++;
209 } else {
210 frames[nframes++] = xdpf;
211 }
212 break;
213 case XDP_REDIRECT:
214 err = xdp_do_redirect(dev: xdpf->dev_rx, xdp: &xdp,
215 prog: rcpu->prog);
216 if (unlikely(err)) {
217 xdp_return_frame(xdpf);
218 stats->drop++;
219 } else {
220 stats->redirect++;
221 }
222 break;
223 default:
224 bpf_warn_invalid_xdp_action(NULL, prog: rcpu->prog, act);
225 fallthrough;
226 case XDP_DROP:
227 xdp_return_frame(xdpf);
228 stats->drop++;
229 break;
230 }
231 }
232
233 stats->pass += nframes;
234
235 return nframes;
236}
237
238#define CPUMAP_BATCH 8
239
240struct cpu_map_ret {
241 u32 xdp_n;
242 u32 skb_n;
243};
244
245static void cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
246 void **skbs, struct cpu_map_ret *ret,
247 struct xdp_cpumap_stats *stats)
248{
249 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
250
251 if (!rcpu->prog)
252 goto out;
253
254 rcu_read_lock();
255 bpf_net_ctx = bpf_net_ctx_set(bpf_net_ctx: &__bpf_net_ctx);
256 xdp_set_return_frame_no_direct();
257
258 ret->xdp_n = cpu_map_bpf_prog_run_xdp(rcpu, frames, n: ret->xdp_n, stats);
259 if (unlikely(ret->skb_n))
260 ret->skb_n = cpu_map_bpf_prog_run_skb(rcpu, skbs, skb_n: ret->skb_n,
261 stats);
262
263 if (stats->redirect)
264 xdp_do_flush();
265
266 xdp_clear_return_frame_no_direct();
267 bpf_net_ctx_clear(bpf_net_ctx);
268 rcu_read_unlock();
269
270out:
271 if (unlikely(ret->skb_n) && ret->xdp_n)
272 memmove(&skbs[ret->xdp_n], skbs, ret->skb_n * sizeof(*skbs));
273}
274
275static void cpu_map_gro_flush(struct bpf_cpu_map_entry *rcpu, bool empty)
276{
277 /*
278 * If the ring is not empty, there'll be a new iteration soon, and we
279 * only need to do a full flush if a tick is long (> 1 ms).
280 * If the ring is empty, to not hold GRO packets in the stack for too
281 * long, do a full flush.
282 * This is equivalent to how NAPI decides whether to perform a full
283 * flush.
284 */
285 gro_flush_normal(gro: &rcpu->gro, flush_old: !empty && HZ >= 1000);
286}
287
288static int cpu_map_kthread_run(void *data)
289{
290 struct bpf_cpu_map_entry *rcpu = data;
291 unsigned long last_qs = jiffies;
292 u32 packets = 0;
293
294 complete(&rcpu->kthread_running);
295 set_current_state(TASK_INTERRUPTIBLE);
296
297 /* When kthread gives stop order, then rcpu have been disconnected
298 * from map, thus no new packets can enter. Remaining in-flight
299 * per CPU stored packets are flushed to this queue. Wait honoring
300 * kthread_stop signal until queue is empty.
301 */
302 while (!kthread_should_stop() || !__ptr_ring_empty(r: rcpu->queue)) {
303 struct xdp_cpumap_stats stats = {}; /* zero stats */
304 unsigned int kmem_alloc_drops = 0, sched = 0;
305 struct cpu_map_ret ret = { };
306 void *frames[CPUMAP_BATCH];
307 void *skbs[CPUMAP_BATCH];
308 u32 i, n, m;
309 bool empty;
310
311 /* Release CPU reschedule checks */
312 if (__ptr_ring_empty(r: rcpu->queue)) {
313 set_current_state(TASK_INTERRUPTIBLE);
314 /* Recheck to avoid lost wake-up */
315 if (__ptr_ring_empty(r: rcpu->queue)) {
316 schedule();
317 sched = 1;
318 last_qs = jiffies;
319 } else {
320 __set_current_state(TASK_RUNNING);
321 }
322 } else {
323 rcu_softirq_qs_periodic(last_qs);
324 sched = cond_resched();
325 }
326
327 /*
328 * The bpf_cpu_map_entry is single consumer, with this
329 * kthread CPU pinned. Lockless access to ptr_ring
330 * consume side valid as no-resize allowed of queue.
331 */
332 n = __ptr_ring_consume_batched(r: rcpu->queue, array: frames,
333 CPUMAP_BATCH);
334 for (i = 0; i < n; i++) {
335 void *f = frames[i];
336 struct page *page;
337
338 if (unlikely(__ptr_test_bit(0, &f))) {
339 struct sk_buff *skb = f;
340
341 __ptr_clear_bit(0, &skb);
342 skbs[ret.skb_n++] = skb;
343 continue;
344 }
345
346 frames[ret.xdp_n++] = f;
347 page = virt_to_page(f);
348
349 /* Bring struct page memory area to curr CPU. Read by
350 * build_skb_around via page_is_pfmemalloc(), and when
351 * freed written by page_frag_free call.
352 */
353 prefetchw(x: page);
354 }
355
356 local_bh_disable();
357
358 /* Support running another XDP prog on this CPU */
359 cpu_map_bpf_prog_run(rcpu, frames, skbs, ret: &ret, stats: &stats);
360 if (!ret.xdp_n)
361 goto stats;
362
363 m = napi_skb_cache_get_bulk(skbs, n: ret.xdp_n);
364 if (unlikely(m < ret.xdp_n)) {
365 for (i = m; i < ret.xdp_n; i++)
366 xdp_return_frame(xdpf: frames[i]);
367
368 if (ret.skb_n)
369 memmove(&skbs[m], &skbs[ret.xdp_n],
370 ret.skb_n * sizeof(*skbs));
371
372 kmem_alloc_drops += ret.xdp_n - m;
373 ret.xdp_n = m;
374 }
375
376 for (i = 0; i < ret.xdp_n; i++) {
377 struct xdp_frame *xdpf = frames[i];
378
379 /* Can fail only when !skb -- already handled above */
380 __xdp_build_skb_from_frame(xdpf, skb: skbs[i], dev: xdpf->dev_rx);
381 }
382
383stats:
384 /* Feedback loop via tracepoint.
385 * NB: keep before recv to allow measuring enqueue/dequeue latency.
386 */
387 trace_xdp_cpumap_kthread(map_id: rcpu->map_id, processed: n, drops: kmem_alloc_drops,
388 sched, xdp_stats: &stats);
389
390 for (i = 0; i < ret.xdp_n + ret.skb_n; i++)
391 gro_receive_skb(gro: &rcpu->gro, skb: skbs[i]);
392
393 /* Flush either every 64 packets or in case of empty ring */
394 packets += n;
395 empty = __ptr_ring_empty(r: rcpu->queue);
396 if (packets >= NAPI_POLL_WEIGHT || empty) {
397 cpu_map_gro_flush(rcpu, empty);
398 packets = 0;
399 }
400
401 local_bh_enable(); /* resched point, may call do_softirq() */
402 }
403 __set_current_state(TASK_RUNNING);
404
405 return 0;
406}
407
408static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu,
409 struct bpf_map *map, int fd)
410{
411 struct bpf_prog *prog;
412
413 prog = bpf_prog_get_type(ufd: fd, type: BPF_PROG_TYPE_XDP);
414 if (IS_ERR(ptr: prog))
415 return PTR_ERR(ptr: prog);
416
417 if (prog->expected_attach_type != BPF_XDP_CPUMAP ||
418 !bpf_prog_map_compatible(map, fp: prog)) {
419 bpf_prog_put(prog);
420 return -EINVAL;
421 }
422
423 rcpu->value.bpf_prog.id = prog->aux->id;
424 rcpu->prog = prog;
425
426 return 0;
427}
428
429static struct bpf_cpu_map_entry *
430__cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value,
431 u32 cpu)
432{
433 int numa, err, i, fd = value->bpf_prog.fd;
434 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
435 struct bpf_cpu_map_entry *rcpu;
436 struct xdp_bulk_queue *bq;
437
438 /* Have map->numa_node, but choose node of redirect target CPU */
439 numa = cpu_to_node(cpu);
440
441 rcpu = bpf_map_kmalloc_node(map, size: sizeof(*rcpu), flags: gfp | __GFP_ZERO, node: numa);
442 if (!rcpu)
443 return NULL;
444
445 /* Alloc percpu bulkq */
446 rcpu->bulkq = bpf_map_alloc_percpu(map, size: sizeof(*rcpu->bulkq),
447 align: sizeof(void *), flags: gfp);
448 if (!rcpu->bulkq)
449 goto free_rcu;
450
451 for_each_possible_cpu(i) {
452 bq = per_cpu_ptr(rcpu->bulkq, i);
453 bq->obj = rcpu;
454 }
455
456 /* Alloc queue */
457 rcpu->queue = bpf_map_kmalloc_node(map, size: sizeof(*rcpu->queue), flags: gfp,
458 node: numa);
459 if (!rcpu->queue)
460 goto free_bulkq;
461
462 err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
463 if (err)
464 goto free_queue;
465
466 rcpu->cpu = cpu;
467 rcpu->map_id = map->id;
468 rcpu->value.qsize = value->qsize;
469 gro_init(gro: &rcpu->gro);
470
471 if (fd > 0 && __cpu_map_load_bpf_program(rcpu, map, fd))
472 goto free_ptr_ring;
473
474 /* Setup kthread */
475 init_completion(x: &rcpu->kthread_running);
476 rcpu->kthread = kthread_create_on_node(threadfn: cpu_map_kthread_run, data: rcpu, node: numa,
477 namefmt: "cpumap/%d/map:%d", cpu,
478 map->id);
479 if (IS_ERR(ptr: rcpu->kthread))
480 goto free_prog;
481
482 /* Make sure kthread runs on a single CPU */
483 kthread_bind(k: rcpu->kthread, cpu);
484 wake_up_process(tsk: rcpu->kthread);
485
486 /* Make sure kthread has been running, so kthread_stop() will not
487 * stop the kthread prematurely and all pending frames or skbs
488 * will be handled by the kthread before kthread_stop() returns.
489 */
490 wait_for_completion(&rcpu->kthread_running);
491
492 return rcpu;
493
494free_prog:
495 if (rcpu->prog)
496 bpf_prog_put(prog: rcpu->prog);
497free_ptr_ring:
498 gro_cleanup(gro: &rcpu->gro);
499 ptr_ring_cleanup(r: rcpu->queue, NULL);
500free_queue:
501 kfree(objp: rcpu->queue);
502free_bulkq:
503 free_percpu(pdata: rcpu->bulkq);
504free_rcu:
505 kfree(objp: rcpu);
506 return NULL;
507}
508
509static void __cpu_map_entry_free(struct work_struct *work)
510{
511 struct bpf_cpu_map_entry *rcpu;
512
513 /* This cpu_map_entry have been disconnected from map and one
514 * RCU grace-period have elapsed. Thus, XDP cannot queue any
515 * new packets and cannot change/set flush_needed that can
516 * find this entry.
517 */
518 rcpu = container_of(to_rcu_work(work), struct bpf_cpu_map_entry, free_work);
519
520 /* kthread_stop will wake_up_process and wait for it to complete.
521 * cpu_map_kthread_run() makes sure the pointer ring is empty
522 * before exiting.
523 */
524 kthread_stop(k: rcpu->kthread);
525
526 if (rcpu->prog)
527 bpf_prog_put(prog: rcpu->prog);
528 gro_cleanup(gro: &rcpu->gro);
529 /* The queue should be empty at this point */
530 __cpu_map_ring_cleanup(ring: rcpu->queue);
531 ptr_ring_cleanup(r: rcpu->queue, NULL);
532 kfree(objp: rcpu->queue);
533 free_percpu(pdata: rcpu->bulkq);
534 kfree(objp: rcpu);
535}
536
537/* After the xchg of the bpf_cpu_map_entry pointer, we need to make sure the old
538 * entry is no longer in use before freeing. We use queue_rcu_work() to call
539 * __cpu_map_entry_free() in a separate workqueue after waiting for an RCU grace
540 * period. This means that (a) all pending enqueue and flush operations have
541 * completed (because of the RCU callback), and (b) we are in a workqueue
542 * context where we can stop the kthread and wait for it to exit before freeing
543 * everything.
544 */
545static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
546 u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
547{
548 struct bpf_cpu_map_entry *old_rcpu;
549
550 old_rcpu = unrcu_pointer(xchg(&cmap->cpu_map[key_cpu], RCU_INITIALIZER(rcpu)));
551 if (old_rcpu) {
552 INIT_RCU_WORK(&old_rcpu->free_work, __cpu_map_entry_free);
553 queue_rcu_work(wq: system_percpu_wq, rwork: &old_rcpu->free_work);
554 }
555}
556
557static long cpu_map_delete_elem(struct bpf_map *map, void *key)
558{
559 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
560 u32 key_cpu = *(u32 *)key;
561
562 if (key_cpu >= map->max_entries)
563 return -EINVAL;
564
565 /* notice caller map_delete_elem() uses rcu_read_lock() */
566 __cpu_map_entry_replace(cmap, key_cpu, NULL);
567 return 0;
568}
569
570static long cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
571 u64 map_flags)
572{
573 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
574 struct bpf_cpumap_val cpumap_value = {};
575 struct bpf_cpu_map_entry *rcpu;
576 /* Array index key correspond to CPU number */
577 u32 key_cpu = *(u32 *)key;
578
579 memcpy(&cpumap_value, value, map->value_size);
580
581 if (unlikely(map_flags > BPF_EXIST))
582 return -EINVAL;
583 if (unlikely(key_cpu >= cmap->map.max_entries))
584 return -E2BIG;
585 if (unlikely(map_flags == BPF_NOEXIST))
586 return -EEXIST;
587 if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
588 return -EOVERFLOW;
589
590 /* Make sure CPU is a valid possible cpu */
591 if (key_cpu >= nr_cpumask_bits || !cpu_possible(cpu: key_cpu))
592 return -ENODEV;
593
594 if (cpumap_value.qsize == 0) {
595 rcpu = NULL; /* Same as deleting */
596 } else {
597 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
598 rcpu = __cpu_map_entry_alloc(map, value: &cpumap_value, cpu: key_cpu);
599 if (!rcpu)
600 return -ENOMEM;
601 }
602 rcu_read_lock();
603 __cpu_map_entry_replace(cmap, key_cpu, rcpu);
604 rcu_read_unlock();
605 return 0;
606}
607
608static void cpu_map_free(struct bpf_map *map)
609{
610 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
611 u32 i;
612
613 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
614 * so the bpf programs (can be more than one that used this map) were
615 * disconnected from events. Wait for outstanding critical sections in
616 * these programs to complete. synchronize_rcu() below not only
617 * guarantees no further "XDP/bpf-side" reads against
618 * bpf_cpu_map->cpu_map, but also ensure pending flush operations
619 * (if any) are completed.
620 */
621 synchronize_rcu();
622
623 /* The only possible user of bpf_cpu_map_entry is
624 * cpu_map_kthread_run().
625 */
626 for (i = 0; i < cmap->map.max_entries; i++) {
627 struct bpf_cpu_map_entry *rcpu;
628
629 rcpu = rcu_dereference_raw(cmap->cpu_map[i]);
630 if (!rcpu)
631 continue;
632
633 /* Stop kthread and cleanup entry directly */
634 __cpu_map_entry_free(work: &rcpu->free_work.work);
635 }
636 bpf_map_area_free(base: cmap->cpu_map);
637 bpf_map_area_free(base: cmap);
638}
639
640/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
641 * by local_bh_disable() (from XDP calls inside NAPI). The
642 * rcu_read_lock_bh_held() below makes lockdep accept both.
643 */
644static void *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
645{
646 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
647 struct bpf_cpu_map_entry *rcpu;
648
649 if (key >= map->max_entries)
650 return NULL;
651
652 rcpu = rcu_dereference_check(cmap->cpu_map[key],
653 rcu_read_lock_bh_held());
654 return rcpu;
655}
656
657static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
658{
659 struct bpf_cpu_map_entry *rcpu =
660 __cpu_map_lookup_elem(map, key: *(u32 *)key);
661
662 return rcpu ? &rcpu->value : NULL;
663}
664
665static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
666{
667 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
668 u32 index = key ? *(u32 *)key : U32_MAX;
669 u32 *next = next_key;
670
671 if (index >= cmap->map.max_entries) {
672 *next = 0;
673 return 0;
674 }
675
676 if (index == cmap->map.max_entries - 1)
677 return -ENOENT;
678 *next = index + 1;
679 return 0;
680}
681
682static long cpu_map_redirect(struct bpf_map *map, u64 index, u64 flags)
683{
684 return __bpf_xdp_redirect_map(map, index, flags, flag_mask: 0,
685 lookup_elem: __cpu_map_lookup_elem);
686}
687
688static u64 cpu_map_mem_usage(const struct bpf_map *map)
689{
690 u64 usage = sizeof(struct bpf_cpu_map);
691
692 /* Currently the dynamically allocated elements are not counted */
693 usage += (u64)map->max_entries * sizeof(struct bpf_cpu_map_entry *);
694 return usage;
695}
696
697BTF_ID_LIST_SINGLE(cpu_map_btf_ids, struct, bpf_cpu_map)
698const struct bpf_map_ops cpu_map_ops = {
699 .map_meta_equal = bpf_map_meta_equal,
700 .map_alloc = cpu_map_alloc,
701 .map_free = cpu_map_free,
702 .map_delete_elem = cpu_map_delete_elem,
703 .map_update_elem = cpu_map_update_elem,
704 .map_lookup_elem = cpu_map_lookup_elem,
705 .map_get_next_key = cpu_map_get_next_key,
706 .map_check_btf = map_check_no_btf,
707 .map_mem_usage = cpu_map_mem_usage,
708 .map_btf_id = &cpu_map_btf_ids[0],
709 .map_redirect = cpu_map_redirect,
710};
711
712static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
713{
714 struct bpf_cpu_map_entry *rcpu = bq->obj;
715 unsigned int processed = 0, drops = 0;
716 const int to_cpu = rcpu->cpu;
717 struct ptr_ring *q;
718 int i;
719
720 if (unlikely(!bq->count))
721 return;
722
723 q = rcpu->queue;
724 spin_lock(lock: &q->producer_lock);
725
726 for (i = 0; i < bq->count; i++) {
727 struct xdp_frame *xdpf = bq->q[i];
728 int err;
729
730 err = __ptr_ring_produce(r: q, ptr: xdpf);
731 if (err) {
732 drops++;
733 xdp_return_frame_rx_napi(xdpf);
734 }
735 processed++;
736 }
737 bq->count = 0;
738 spin_unlock(lock: &q->producer_lock);
739
740 __list_del_clearprev(entry: &bq->flush_node);
741
742 /* Feedback loop via tracepoints */
743 trace_xdp_cpumap_enqueue(map_id: rcpu->map_id, processed, drops, to_cpu);
744}
745
746/* Runs under RCU-read-side, plus in softirq under NAPI protection.
747 * Thus, safe percpu variable access.
748 */
749static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
750{
751 struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
752
753 if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
754 bq_flush_to_queue(bq);
755
756 /* Notice, xdp_buff/page MUST be queued here, long enough for
757 * driver to code invoking us to finished, due to driver
758 * (e.g. ixgbe) recycle tricks based on page-refcnt.
759 *
760 * Thus, incoming xdp_frame is always queued here (else we race
761 * with another CPU on page-refcnt and remaining driver code).
762 * Queue time is very short, as driver will invoke flush
763 * operation, when completing napi->poll call.
764 */
765 bq->q[bq->count++] = xdpf;
766
767 if (!bq->flush_node.prev) {
768 struct list_head *flush_list = bpf_net_ctx_get_cpu_map_flush_list();
769
770 list_add(new: &bq->flush_node, head: flush_list);
771 }
772}
773
774int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf,
775 struct net_device *dev_rx)
776{
777 /* Info needed when constructing SKB on remote CPU */
778 xdpf->dev_rx = dev_rx;
779
780 bq_enqueue(rcpu, xdpf);
781 return 0;
782}
783
784int cpu_map_generic_redirect(struct bpf_cpu_map_entry *rcpu,
785 struct sk_buff *skb)
786{
787 int ret;
788
789 __skb_pull(skb, len: skb->mac_len);
790 skb_set_redirected(skb, from_ingress: false);
791 __ptr_set_bit(0, &skb);
792
793 ret = ptr_ring_produce(r: rcpu->queue, ptr: skb);
794 if (ret < 0)
795 goto trace;
796
797 wake_up_process(tsk: rcpu->kthread);
798trace:
799 trace_xdp_cpumap_enqueue(map_id: rcpu->map_id, processed: !ret, drops: !!ret, to_cpu: rcpu->cpu);
800 return ret;
801}
802
803void __cpu_map_flush(struct list_head *flush_list)
804{
805 struct xdp_bulk_queue *bq, *tmp;
806
807 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
808 bq_flush_to_queue(bq);
809
810 /* If already running, costs spin_lock_irqsave + smb_mb */
811 wake_up_process(tsk: bq->obj->kthread);
812 }
813}
814

source code of linux/kernel/bpf/cpumap.c