xref: /linux/kernel/bpf/cpumap.c (revision 6efc0ab3b05de0d7bab8ec0597214e4788251071)
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>   /* netif_receive_skb_list */
37 #include <linux/etherdevice.h> /* eth_type_trans */
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 */
47 struct bpf_cpu_map_entry;
48 struct bpf_cpu_map;
49 
50 struct 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 */
58 struct 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 
72 	struct completion kthread_running;
73 	struct rcu_work free_work;
74 };
75 
76 struct bpf_cpu_map {
77 	struct bpf_map map;
78 	/* Below members specific for map type */
79 	struct bpf_cpu_map_entry __rcu **cpu_map;
80 };
81 
82 static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
83 {
84 	u32 value_size = attr->value_size;
85 	struct bpf_cpu_map *cmap;
86 
87 	/* check sanity of attributes */
88 	if (attr->max_entries == 0 || attr->key_size != 4 ||
89 	    (value_size != offsetofend(struct bpf_cpumap_val, qsize) &&
90 	     value_size != offsetofend(struct bpf_cpumap_val, bpf_prog.fd)) ||
91 	    attr->map_flags & ~BPF_F_NUMA_NODE)
92 		return ERR_PTR(-EINVAL);
93 
94 	/* Pre-limit array size based on NR_CPUS, not final CPU check */
95 	if (attr->max_entries > NR_CPUS)
96 		return ERR_PTR(-E2BIG);
97 
98 	cmap = bpf_map_area_alloc(sizeof(*cmap), NUMA_NO_NODE);
99 	if (!cmap)
100 		return ERR_PTR(-ENOMEM);
101 
102 	bpf_map_init_from_attr(&cmap->map, attr);
103 
104 	/* Alloc array for possible remote "destination" CPUs */
105 	cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
106 					   sizeof(struct bpf_cpu_map_entry *),
107 					   cmap->map.numa_node);
108 	if (!cmap->cpu_map) {
109 		bpf_map_area_free(cmap);
110 		return ERR_PTR(-ENOMEM);
111 	}
112 
113 	return &cmap->map;
114 }
115 
116 static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
117 {
118 	/* The tear-down procedure should have made sure that queue is
119 	 * empty.  See __cpu_map_entry_replace() and work-queue
120 	 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
121 	 * gracefully and warn once.
122 	 */
123 	void *ptr;
124 
125 	while ((ptr = ptr_ring_consume(ring))) {
126 		WARN_ON_ONCE(1);
127 		if (unlikely(__ptr_test_bit(0, &ptr))) {
128 			__ptr_clear_bit(0, &ptr);
129 			kfree_skb(ptr);
130 			continue;
131 		}
132 		xdp_return_frame(ptr);
133 	}
134 }
135 
136 static void cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
137 				     struct list_head *listp,
138 				     struct xdp_cpumap_stats *stats)
139 {
140 	struct sk_buff *skb, *tmp;
141 	struct xdp_buff xdp;
142 	u32 act;
143 	int err;
144 
145 	list_for_each_entry_safe(skb, tmp, listp, list) {
146 		act = bpf_prog_run_generic_xdp(skb, &xdp, rcpu->prog);
147 		switch (act) {
148 		case XDP_PASS:
149 			break;
150 		case XDP_REDIRECT:
151 			skb_list_del_init(skb);
152 			err = xdp_do_generic_redirect(skb->dev, skb, &xdp,
153 						      rcpu->prog);
154 			if (unlikely(err)) {
155 				kfree_skb(skb);
156 				stats->drop++;
157 			} else {
158 				stats->redirect++;
159 			}
160 			return;
161 		default:
162 			bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
163 			fallthrough;
164 		case XDP_ABORTED:
165 			trace_xdp_exception(skb->dev, rcpu->prog, act);
166 			fallthrough;
167 		case XDP_DROP:
168 			skb_list_del_init(skb);
169 			kfree_skb(skb);
170 			stats->drop++;
171 			return;
172 		}
173 	}
174 }
175 
176 static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
177 				    void **frames, int n,
178 				    struct xdp_cpumap_stats *stats)
179 {
180 	struct xdp_rxq_info rxq = {};
181 	struct xdp_buff xdp;
182 	int i, nframes = 0;
183 
184 	xdp_set_return_frame_no_direct();
185 	xdp.rxq = &rxq;
186 
187 	for (i = 0; i < n; i++) {
188 		struct xdp_frame *xdpf = frames[i];
189 		u32 act;
190 		int err;
191 
192 		rxq.dev = xdpf->dev_rx;
193 		rxq.mem = xdpf->mem;
194 		/* TODO: report queue_index to xdp_rxq_info */
195 
196 		xdp_convert_frame_to_buff(xdpf, &xdp);
197 
198 		act = bpf_prog_run_xdp(rcpu->prog, &xdp);
199 		switch (act) {
200 		case XDP_PASS:
201 			err = xdp_update_frame_from_buff(&xdp, xdpf);
202 			if (err < 0) {
203 				xdp_return_frame(xdpf);
204 				stats->drop++;
205 			} else {
206 				frames[nframes++] = xdpf;
207 				stats->pass++;
208 			}
209 			break;
210 		case XDP_REDIRECT:
211 			err = xdp_do_redirect(xdpf->dev_rx, &xdp,
212 					      rcpu->prog);
213 			if (unlikely(err)) {
214 				xdp_return_frame(xdpf);
215 				stats->drop++;
216 			} else {
217 				stats->redirect++;
218 			}
219 			break;
220 		default:
221 			bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
222 			fallthrough;
223 		case XDP_DROP:
224 			xdp_return_frame(xdpf);
225 			stats->drop++;
226 			break;
227 		}
228 	}
229 
230 	xdp_clear_return_frame_no_direct();
231 
232 	return nframes;
233 }
234 
235 #define CPUMAP_BATCH 8
236 
237 static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
238 				int xdp_n, struct xdp_cpumap_stats *stats,
239 				struct list_head *list)
240 {
241 	struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
242 	int nframes;
243 
244 	if (!rcpu->prog)
245 		return xdp_n;
246 
247 	rcu_read_lock_bh();
248 	bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx);
249 
250 	nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, xdp_n, stats);
251 
252 	if (stats->redirect)
253 		xdp_do_flush();
254 
255 	if (unlikely(!list_empty(list)))
256 		cpu_map_bpf_prog_run_skb(rcpu, list, stats);
257 
258 	bpf_net_ctx_clear(bpf_net_ctx);
259 	rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
260 
261 	return nframes;
262 }
263 
264 static int cpu_map_kthread_run(void *data)
265 {
266 	struct bpf_cpu_map_entry *rcpu = data;
267 	unsigned long last_qs = jiffies;
268 
269 	complete(&rcpu->kthread_running);
270 	set_current_state(TASK_INTERRUPTIBLE);
271 
272 	/* When kthread gives stop order, then rcpu have been disconnected
273 	 * from map, thus no new packets can enter. Remaining in-flight
274 	 * per CPU stored packets are flushed to this queue.  Wait honoring
275 	 * kthread_stop signal until queue is empty.
276 	 */
277 	while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
278 		struct xdp_cpumap_stats stats = {}; /* zero stats */
279 		unsigned int kmem_alloc_drops = 0, sched = 0;
280 		gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
281 		int i, n, m, nframes, xdp_n;
282 		void *frames[CPUMAP_BATCH];
283 		void *skbs[CPUMAP_BATCH];
284 		LIST_HEAD(list);
285 
286 		/* Release CPU reschedule checks */
287 		if (__ptr_ring_empty(rcpu->queue)) {
288 			set_current_state(TASK_INTERRUPTIBLE);
289 			/* Recheck to avoid lost wake-up */
290 			if (__ptr_ring_empty(rcpu->queue)) {
291 				schedule();
292 				sched = 1;
293 				last_qs = jiffies;
294 			} else {
295 				__set_current_state(TASK_RUNNING);
296 			}
297 		} else {
298 			rcu_softirq_qs_periodic(last_qs);
299 			sched = cond_resched();
300 		}
301 
302 		/*
303 		 * The bpf_cpu_map_entry is single consumer, with this
304 		 * kthread CPU pinned. Lockless access to ptr_ring
305 		 * consume side valid as no-resize allowed of queue.
306 		 */
307 		n = __ptr_ring_consume_batched(rcpu->queue, frames,
308 					       CPUMAP_BATCH);
309 		for (i = 0, xdp_n = 0; i < n; i++) {
310 			void *f = frames[i];
311 			struct page *page;
312 
313 			if (unlikely(__ptr_test_bit(0, &f))) {
314 				struct sk_buff *skb = f;
315 
316 				__ptr_clear_bit(0, &skb);
317 				list_add_tail(&skb->list, &list);
318 				continue;
319 			}
320 
321 			frames[xdp_n++] = f;
322 			page = virt_to_page(f);
323 
324 			/* Bring struct page memory area to curr CPU. Read by
325 			 * build_skb_around via page_is_pfmemalloc(), and when
326 			 * freed written by page_frag_free call.
327 			 */
328 			prefetchw(page);
329 		}
330 
331 		/* Support running another XDP prog on this CPU */
332 		nframes = cpu_map_bpf_prog_run(rcpu, frames, xdp_n, &stats, &list);
333 		if (nframes) {
334 			m = kmem_cache_alloc_bulk(net_hotdata.skbuff_cache,
335 						  gfp, nframes, skbs);
336 			if (unlikely(m == 0)) {
337 				for (i = 0; i < nframes; i++)
338 					skbs[i] = NULL; /* effect: xdp_return_frame */
339 				kmem_alloc_drops += nframes;
340 			}
341 		}
342 
343 		local_bh_disable();
344 		for (i = 0; i < nframes; i++) {
345 			struct xdp_frame *xdpf = frames[i];
346 			struct sk_buff *skb = skbs[i];
347 
348 			skb = __xdp_build_skb_from_frame(xdpf, skb,
349 							 xdpf->dev_rx);
350 			if (!skb) {
351 				xdp_return_frame(xdpf);
352 				continue;
353 			}
354 
355 			list_add_tail(&skb->list, &list);
356 		}
357 		netif_receive_skb_list(&list);
358 
359 		/* Feedback loop via tracepoint */
360 		trace_xdp_cpumap_kthread(rcpu->map_id, n, kmem_alloc_drops,
361 					 sched, &stats);
362 
363 		local_bh_enable(); /* resched point, may call do_softirq() */
364 	}
365 	__set_current_state(TASK_RUNNING);
366 
367 	return 0;
368 }
369 
370 static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu,
371 				      struct bpf_map *map, int fd)
372 {
373 	struct bpf_prog *prog;
374 
375 	prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
376 	if (IS_ERR(prog))
377 		return PTR_ERR(prog);
378 
379 	if (prog->expected_attach_type != BPF_XDP_CPUMAP ||
380 	    !bpf_prog_map_compatible(map, prog)) {
381 		bpf_prog_put(prog);
382 		return -EINVAL;
383 	}
384 
385 	rcpu->value.bpf_prog.id = prog->aux->id;
386 	rcpu->prog = prog;
387 
388 	return 0;
389 }
390 
391 static struct bpf_cpu_map_entry *
392 __cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value,
393 		      u32 cpu)
394 {
395 	int numa, err, i, fd = value->bpf_prog.fd;
396 	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
397 	struct bpf_cpu_map_entry *rcpu;
398 	struct xdp_bulk_queue *bq;
399 
400 	/* Have map->numa_node, but choose node of redirect target CPU */
401 	numa = cpu_to_node(cpu);
402 
403 	rcpu = bpf_map_kmalloc_node(map, sizeof(*rcpu), gfp | __GFP_ZERO, numa);
404 	if (!rcpu)
405 		return NULL;
406 
407 	/* Alloc percpu bulkq */
408 	rcpu->bulkq = bpf_map_alloc_percpu(map, sizeof(*rcpu->bulkq),
409 					   sizeof(void *), gfp);
410 	if (!rcpu->bulkq)
411 		goto free_rcu;
412 
413 	for_each_possible_cpu(i) {
414 		bq = per_cpu_ptr(rcpu->bulkq, i);
415 		bq->obj = rcpu;
416 	}
417 
418 	/* Alloc queue */
419 	rcpu->queue = bpf_map_kmalloc_node(map, sizeof(*rcpu->queue), gfp,
420 					   numa);
421 	if (!rcpu->queue)
422 		goto free_bulkq;
423 
424 	err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
425 	if (err)
426 		goto free_queue;
427 
428 	rcpu->cpu    = cpu;
429 	rcpu->map_id = map->id;
430 	rcpu->value.qsize  = value->qsize;
431 
432 	if (fd > 0 && __cpu_map_load_bpf_program(rcpu, map, fd))
433 		goto free_ptr_ring;
434 
435 	/* Setup kthread */
436 	init_completion(&rcpu->kthread_running);
437 	rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
438 					       "cpumap/%d/map:%d", cpu,
439 					       map->id);
440 	if (IS_ERR(rcpu->kthread))
441 		goto free_prog;
442 
443 	/* Make sure kthread runs on a single CPU */
444 	kthread_bind(rcpu->kthread, cpu);
445 	wake_up_process(rcpu->kthread);
446 
447 	/* Make sure kthread has been running, so kthread_stop() will not
448 	 * stop the kthread prematurely and all pending frames or skbs
449 	 * will be handled by the kthread before kthread_stop() returns.
450 	 */
451 	wait_for_completion(&rcpu->kthread_running);
452 
453 	return rcpu;
454 
455 free_prog:
456 	if (rcpu->prog)
457 		bpf_prog_put(rcpu->prog);
458 free_ptr_ring:
459 	ptr_ring_cleanup(rcpu->queue, NULL);
460 free_queue:
461 	kfree(rcpu->queue);
462 free_bulkq:
463 	free_percpu(rcpu->bulkq);
464 free_rcu:
465 	kfree(rcpu);
466 	return NULL;
467 }
468 
469 static void __cpu_map_entry_free(struct work_struct *work)
470 {
471 	struct bpf_cpu_map_entry *rcpu;
472 
473 	/* This cpu_map_entry have been disconnected from map and one
474 	 * RCU grace-period have elapsed. Thus, XDP cannot queue any
475 	 * new packets and cannot change/set flush_needed that can
476 	 * find this entry.
477 	 */
478 	rcpu = container_of(to_rcu_work(work), struct bpf_cpu_map_entry, free_work);
479 
480 	/* kthread_stop will wake_up_process and wait for it to complete.
481 	 * cpu_map_kthread_run() makes sure the pointer ring is empty
482 	 * before exiting.
483 	 */
484 	kthread_stop(rcpu->kthread);
485 
486 	if (rcpu->prog)
487 		bpf_prog_put(rcpu->prog);
488 	/* The queue should be empty at this point */
489 	__cpu_map_ring_cleanup(rcpu->queue);
490 	ptr_ring_cleanup(rcpu->queue, NULL);
491 	kfree(rcpu->queue);
492 	free_percpu(rcpu->bulkq);
493 	kfree(rcpu);
494 }
495 
496 /* After the xchg of the bpf_cpu_map_entry pointer, we need to make sure the old
497  * entry is no longer in use before freeing. We use queue_rcu_work() to call
498  * __cpu_map_entry_free() in a separate workqueue after waiting for an RCU grace
499  * period. This means that (a) all pending enqueue and flush operations have
500  * completed (because of the RCU callback), and (b) we are in a workqueue
501  * context where we can stop the kthread and wait for it to exit before freeing
502  * everything.
503  */
504 static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
505 				    u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
506 {
507 	struct bpf_cpu_map_entry *old_rcpu;
508 
509 	old_rcpu = unrcu_pointer(xchg(&cmap->cpu_map[key_cpu], RCU_INITIALIZER(rcpu)));
510 	if (old_rcpu) {
511 		INIT_RCU_WORK(&old_rcpu->free_work, __cpu_map_entry_free);
512 		queue_rcu_work(system_wq, &old_rcpu->free_work);
513 	}
514 }
515 
516 static long cpu_map_delete_elem(struct bpf_map *map, void *key)
517 {
518 	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
519 	u32 key_cpu = *(u32 *)key;
520 
521 	if (key_cpu >= map->max_entries)
522 		return -EINVAL;
523 
524 	/* notice caller map_delete_elem() uses rcu_read_lock() */
525 	__cpu_map_entry_replace(cmap, key_cpu, NULL);
526 	return 0;
527 }
528 
529 static long cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
530 				u64 map_flags)
531 {
532 	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
533 	struct bpf_cpumap_val cpumap_value = {};
534 	struct bpf_cpu_map_entry *rcpu;
535 	/* Array index key correspond to CPU number */
536 	u32 key_cpu = *(u32 *)key;
537 
538 	memcpy(&cpumap_value, value, map->value_size);
539 
540 	if (unlikely(map_flags > BPF_EXIST))
541 		return -EINVAL;
542 	if (unlikely(key_cpu >= cmap->map.max_entries))
543 		return -E2BIG;
544 	if (unlikely(map_flags == BPF_NOEXIST))
545 		return -EEXIST;
546 	if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
547 		return -EOVERFLOW;
548 
549 	/* Make sure CPU is a valid possible cpu */
550 	if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
551 		return -ENODEV;
552 
553 	if (cpumap_value.qsize == 0) {
554 		rcpu = NULL; /* Same as deleting */
555 	} else {
556 		/* Updating qsize cause re-allocation of bpf_cpu_map_entry */
557 		rcpu = __cpu_map_entry_alloc(map, &cpumap_value, key_cpu);
558 		if (!rcpu)
559 			return -ENOMEM;
560 	}
561 	rcu_read_lock();
562 	__cpu_map_entry_replace(cmap, key_cpu, rcpu);
563 	rcu_read_unlock();
564 	return 0;
565 }
566 
567 static void cpu_map_free(struct bpf_map *map)
568 {
569 	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
570 	u32 i;
571 
572 	/* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
573 	 * so the bpf programs (can be more than one that used this map) were
574 	 * disconnected from events. Wait for outstanding critical sections in
575 	 * these programs to complete. synchronize_rcu() below not only
576 	 * guarantees no further "XDP/bpf-side" reads against
577 	 * bpf_cpu_map->cpu_map, but also ensure pending flush operations
578 	 * (if any) are completed.
579 	 */
580 	synchronize_rcu();
581 
582 	/* The only possible user of bpf_cpu_map_entry is
583 	 * cpu_map_kthread_run().
584 	 */
585 	for (i = 0; i < cmap->map.max_entries; i++) {
586 		struct bpf_cpu_map_entry *rcpu;
587 
588 		rcpu = rcu_dereference_raw(cmap->cpu_map[i]);
589 		if (!rcpu)
590 			continue;
591 
592 		/* Stop kthread and cleanup entry directly */
593 		__cpu_map_entry_free(&rcpu->free_work.work);
594 	}
595 	bpf_map_area_free(cmap->cpu_map);
596 	bpf_map_area_free(cmap);
597 }
598 
599 /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
600  * by local_bh_disable() (from XDP calls inside NAPI). The
601  * rcu_read_lock_bh_held() below makes lockdep accept both.
602  */
603 static void *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
604 {
605 	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
606 	struct bpf_cpu_map_entry *rcpu;
607 
608 	if (key >= map->max_entries)
609 		return NULL;
610 
611 	rcpu = rcu_dereference_check(cmap->cpu_map[key],
612 				     rcu_read_lock_bh_held());
613 	return rcpu;
614 }
615 
616 static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
617 {
618 	struct bpf_cpu_map_entry *rcpu =
619 		__cpu_map_lookup_elem(map, *(u32 *)key);
620 
621 	return rcpu ? &rcpu->value : NULL;
622 }
623 
624 static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
625 {
626 	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
627 	u32 index = key ? *(u32 *)key : U32_MAX;
628 	u32 *next = next_key;
629 
630 	if (index >= cmap->map.max_entries) {
631 		*next = 0;
632 		return 0;
633 	}
634 
635 	if (index == cmap->map.max_entries - 1)
636 		return -ENOENT;
637 	*next = index + 1;
638 	return 0;
639 }
640 
641 static long cpu_map_redirect(struct bpf_map *map, u64 index, u64 flags)
642 {
643 	return __bpf_xdp_redirect_map(map, index, flags, 0,
644 				      __cpu_map_lookup_elem);
645 }
646 
647 static u64 cpu_map_mem_usage(const struct bpf_map *map)
648 {
649 	u64 usage = sizeof(struct bpf_cpu_map);
650 
651 	/* Currently the dynamically allocated elements are not counted */
652 	usage += (u64)map->max_entries * sizeof(struct bpf_cpu_map_entry *);
653 	return usage;
654 }
655 
656 BTF_ID_LIST_SINGLE(cpu_map_btf_ids, struct, bpf_cpu_map)
657 const struct bpf_map_ops cpu_map_ops = {
658 	.map_meta_equal		= bpf_map_meta_equal,
659 	.map_alloc		= cpu_map_alloc,
660 	.map_free		= cpu_map_free,
661 	.map_delete_elem	= cpu_map_delete_elem,
662 	.map_update_elem	= cpu_map_update_elem,
663 	.map_lookup_elem	= cpu_map_lookup_elem,
664 	.map_get_next_key	= cpu_map_get_next_key,
665 	.map_check_btf		= map_check_no_btf,
666 	.map_mem_usage		= cpu_map_mem_usage,
667 	.map_btf_id		= &cpu_map_btf_ids[0],
668 	.map_redirect		= cpu_map_redirect,
669 };
670 
671 static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
672 {
673 	struct bpf_cpu_map_entry *rcpu = bq->obj;
674 	unsigned int processed = 0, drops = 0;
675 	const int to_cpu = rcpu->cpu;
676 	struct ptr_ring *q;
677 	int i;
678 
679 	if (unlikely(!bq->count))
680 		return;
681 
682 	q = rcpu->queue;
683 	spin_lock(&q->producer_lock);
684 
685 	for (i = 0; i < bq->count; i++) {
686 		struct xdp_frame *xdpf = bq->q[i];
687 		int err;
688 
689 		err = __ptr_ring_produce(q, xdpf);
690 		if (err) {
691 			drops++;
692 			xdp_return_frame_rx_napi(xdpf);
693 		}
694 		processed++;
695 	}
696 	bq->count = 0;
697 	spin_unlock(&q->producer_lock);
698 
699 	__list_del_clearprev(&bq->flush_node);
700 
701 	/* Feedback loop via tracepoints */
702 	trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
703 }
704 
705 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
706  * Thus, safe percpu variable access.
707  */
708 static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
709 {
710 	struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
711 
712 	if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
713 		bq_flush_to_queue(bq);
714 
715 	/* Notice, xdp_buff/page MUST be queued here, long enough for
716 	 * driver to code invoking us to finished, due to driver
717 	 * (e.g. ixgbe) recycle tricks based on page-refcnt.
718 	 *
719 	 * Thus, incoming xdp_frame is always queued here (else we race
720 	 * with another CPU on page-refcnt and remaining driver code).
721 	 * Queue time is very short, as driver will invoke flush
722 	 * operation, when completing napi->poll call.
723 	 */
724 	bq->q[bq->count++] = xdpf;
725 
726 	if (!bq->flush_node.prev) {
727 		struct list_head *flush_list = bpf_net_ctx_get_cpu_map_flush_list();
728 
729 		list_add(&bq->flush_node, flush_list);
730 	}
731 }
732 
733 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf,
734 		    struct net_device *dev_rx)
735 {
736 	/* Info needed when constructing SKB on remote CPU */
737 	xdpf->dev_rx = dev_rx;
738 
739 	bq_enqueue(rcpu, xdpf);
740 	return 0;
741 }
742 
743 int cpu_map_generic_redirect(struct bpf_cpu_map_entry *rcpu,
744 			     struct sk_buff *skb)
745 {
746 	int ret;
747 
748 	__skb_pull(skb, skb->mac_len);
749 	skb_set_redirected(skb, false);
750 	__ptr_set_bit(0, &skb);
751 
752 	ret = ptr_ring_produce(rcpu->queue, skb);
753 	if (ret < 0)
754 		goto trace;
755 
756 	wake_up_process(rcpu->kthread);
757 trace:
758 	trace_xdp_cpumap_enqueue(rcpu->map_id, !ret, !!ret, rcpu->cpu);
759 	return ret;
760 }
761 
762 void __cpu_map_flush(struct list_head *flush_list)
763 {
764 	struct xdp_bulk_queue *bq, *tmp;
765 
766 	list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
767 		bq_flush_to_queue(bq);
768 
769 		/* If already running, costs spin_lock_irqsave + smb_mb */
770 		wake_up_process(bq->obj->kthread);
771 	}
772 }
773