• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 */
41 struct bpf_cpu_map_entry;
42 struct bpf_cpu_map;
43 
44 struct 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 */
52 struct bpf_cpu_map_entry {
53 	u32 cpu;    /* kthread CPU and map index */
54 	int map_id; /* Back reference to map */
55 	u32 qsize;  /* Queue size placeholder for map lookup */
56 
57 	/* XDP can run multiple RX-ring queues, need __percpu enqueue store */
58 	struct xdp_bulk_queue __percpu *bulkq;
59 
60 	struct bpf_cpu_map *cmap;
61 
62 	/* Queue with potential multi-producers, and single-consumer kthread */
63 	struct ptr_ring *queue;
64 	struct task_struct *kthread;
65 	struct work_struct kthread_stop_wq;
66 
67 	atomic_t refcnt; /* Control when this struct can be free'ed */
68 	struct rcu_head rcu;
69 };
70 
71 struct bpf_cpu_map {
72 	struct bpf_map map;
73 	/* Below members specific for map type */
74 	struct bpf_cpu_map_entry **cpu_map;
75 	struct list_head __percpu *flush_list;
76 };
77 
78 static int bq_flush_to_queue(struct xdp_bulk_queue *bq, bool in_napi_ctx);
79 
cpu_map_alloc(union bpf_attr * attr)80 static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
81 {
82 	struct bpf_cpu_map *cmap;
83 	int err = -ENOMEM;
84 	int ret, cpu;
85 	u64 cost;
86 
87 	if (!capable(CAP_SYS_ADMIN))
88 		return ERR_PTR(-EPERM);
89 
90 	/* check sanity of attributes */
91 	if (attr->max_entries == 0 || attr->key_size != 4 ||
92 	    attr->value_size != 4 || attr->map_flags & ~BPF_F_NUMA_NODE)
93 		return ERR_PTR(-EINVAL);
94 
95 	cmap = kzalloc(sizeof(*cmap), GFP_USER);
96 	if (!cmap)
97 		return ERR_PTR(-ENOMEM);
98 
99 	bpf_map_init_from_attr(&cmap->map, attr);
100 
101 	/* Pre-limit array size based on NR_CPUS, not final CPU check */
102 	if (cmap->map.max_entries > NR_CPUS) {
103 		err = -E2BIG;
104 		goto free_cmap;
105 	}
106 
107 	/* make sure page count doesn't overflow */
108 	cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *);
109 	cost += sizeof(struct list_head) * num_possible_cpus();
110 
111 	/* Notice returns -EPERM on if map size is larger than memlock limit */
112 	ret = bpf_map_charge_init(&cmap->map.memory, cost);
113 	if (ret) {
114 		err = ret;
115 		goto free_cmap;
116 	}
117 
118 	cmap->flush_list = alloc_percpu(struct list_head);
119 	if (!cmap->flush_list)
120 		goto free_charge;
121 
122 	for_each_possible_cpu(cpu)
123 		INIT_LIST_HEAD(per_cpu_ptr(cmap->flush_list, cpu));
124 
125 	/* Alloc array for possible remote "destination" CPUs */
126 	cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
127 					   sizeof(struct bpf_cpu_map_entry *),
128 					   cmap->map.numa_node);
129 	if (!cmap->cpu_map)
130 		goto free_percpu;
131 
132 	return &cmap->map;
133 free_percpu:
134 	free_percpu(cmap->flush_list);
135 free_charge:
136 	bpf_map_charge_finish(&cmap->map.memory);
137 free_cmap:
138 	kfree(cmap);
139 	return ERR_PTR(err);
140 }
141 
get_cpu_map_entry(struct bpf_cpu_map_entry * rcpu)142 static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
143 {
144 	atomic_inc(&rcpu->refcnt);
145 }
146 
147 /* called from workqueue, to workaround syscall using preempt_disable */
cpu_map_kthread_stop(struct work_struct * work)148 static void cpu_map_kthread_stop(struct work_struct *work)
149 {
150 	struct bpf_cpu_map_entry *rcpu;
151 
152 	rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
153 
154 	/* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
155 	 * as it waits until all in-flight call_rcu() callbacks complete.
156 	 */
157 	rcu_barrier();
158 
159 	/* kthread_stop will wake_up_process and wait for it to complete */
160 	kthread_stop(rcpu->kthread);
161 }
162 
cpu_map_build_skb(struct bpf_cpu_map_entry * rcpu,struct xdp_frame * xdpf,struct sk_buff * skb)163 static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu,
164 					 struct xdp_frame *xdpf,
165 					 struct sk_buff *skb)
166 {
167 	unsigned int hard_start_headroom;
168 	unsigned int frame_size;
169 	void *pkt_data_start;
170 
171 	/* Part of headroom was reserved to xdpf */
172 	hard_start_headroom = sizeof(struct xdp_frame) +  xdpf->headroom;
173 
174 	/* build_skb need to place skb_shared_info after SKB end, and
175 	 * also want to know the memory "truesize".  Thus, need to
176 	 * know the memory frame size backing xdp_buff.
177 	 *
178 	 * XDP was designed to have PAGE_SIZE frames, but this
179 	 * assumption is not longer true with ixgbe and i40e.  It
180 	 * would be preferred to set frame_size to 2048 or 4096
181 	 * depending on the driver.
182 	 *   frame_size = 2048;
183 	 *   frame_len  = frame_size - sizeof(*xdp_frame);
184 	 *
185 	 * Instead, with info avail, skb_shared_info in placed after
186 	 * packet len.  This, unfortunately fakes the truesize.
187 	 * Another disadvantage of this approach, the skb_shared_info
188 	 * is not at a fixed memory location, with mixed length
189 	 * packets, which is bad for cache-line hotness.
190 	 */
191 	frame_size = SKB_DATA_ALIGN(xdpf->len + hard_start_headroom) +
192 		SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
193 
194 	pkt_data_start = xdpf->data - hard_start_headroom;
195 	skb = build_skb_around(skb, pkt_data_start, frame_size);
196 	if (unlikely(!skb))
197 		return NULL;
198 
199 	skb_reserve(skb, hard_start_headroom);
200 	__skb_put(skb, xdpf->len);
201 	if (xdpf->metasize)
202 		skb_metadata_set(skb, xdpf->metasize);
203 
204 	/* Essential SKB info: protocol and skb->dev */
205 	skb->protocol = eth_type_trans(skb, xdpf->dev_rx);
206 
207 	/* Optional SKB info, currently missing:
208 	 * - HW checksum info		(skb->ip_summed)
209 	 * - HW RX hash			(skb_set_hash)
210 	 * - RX ring dev queue index	(skb_record_rx_queue)
211 	 */
212 
213 	/* Until page_pool get SKB return path, release DMA here */
214 	xdp_release_frame(xdpf);
215 
216 	/* Allow SKB to reuse area used by xdp_frame */
217 	xdp_scrub_frame(xdpf);
218 
219 	return skb;
220 }
221 
__cpu_map_ring_cleanup(struct ptr_ring * ring)222 static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
223 {
224 	/* The tear-down procedure should have made sure that queue is
225 	 * empty.  See __cpu_map_entry_replace() and work-queue
226 	 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
227 	 * gracefully and warn once.
228 	 */
229 	struct xdp_frame *xdpf;
230 
231 	while ((xdpf = ptr_ring_consume(ring)))
232 		if (WARN_ON_ONCE(xdpf))
233 			xdp_return_frame(xdpf);
234 }
235 
put_cpu_map_entry(struct bpf_cpu_map_entry * rcpu)236 static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
237 {
238 	if (atomic_dec_and_test(&rcpu->refcnt)) {
239 		/* The queue should be empty at this point */
240 		__cpu_map_ring_cleanup(rcpu->queue);
241 		ptr_ring_cleanup(rcpu->queue, NULL);
242 		kfree(rcpu->queue);
243 		kfree(rcpu);
244 	}
245 }
246 
247 #define CPUMAP_BATCH 8
248 
cpu_map_kthread_run(void * data)249 static int cpu_map_kthread_run(void *data)
250 {
251 	struct bpf_cpu_map_entry *rcpu = data;
252 	unsigned long last_qs = jiffies;
253 
254 	set_current_state(TASK_INTERRUPTIBLE);
255 
256 	/* When kthread gives stop order, then rcpu have been disconnected
257 	 * from map, thus no new packets can enter. Remaining in-flight
258 	 * per CPU stored packets are flushed to this queue.  Wait honoring
259 	 * kthread_stop signal until queue is empty.
260 	 */
261 	while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
262 		unsigned int drops = 0, sched = 0;
263 		void *frames[CPUMAP_BATCH];
264 		void *skbs[CPUMAP_BATCH];
265 		gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
266 		int i, n, m;
267 
268 		/* Release CPU reschedule checks */
269 		if (__ptr_ring_empty(rcpu->queue)) {
270 			set_current_state(TASK_INTERRUPTIBLE);
271 			/* Recheck to avoid lost wake-up */
272 			if (__ptr_ring_empty(rcpu->queue)) {
273 				schedule();
274 				sched = 1;
275 				last_qs = jiffies;
276 			} else {
277 				__set_current_state(TASK_RUNNING);
278 			}
279 		} else {
280 			rcu_softirq_qs_periodic(last_qs);
281 			sched = cond_resched();
282 		}
283 
284 		/*
285 		 * The bpf_cpu_map_entry is single consumer, with this
286 		 * kthread CPU pinned. Lockless access to ptr_ring
287 		 * consume side valid as no-resize allowed of queue.
288 		 */
289 		n = ptr_ring_consume_batched(rcpu->queue, frames, CPUMAP_BATCH);
290 
291 		for (i = 0; i < n; i++) {
292 			void *f = frames[i];
293 			struct page *page = virt_to_page(f);
294 
295 			/* Bring struct page memory area to curr CPU. Read by
296 			 * build_skb_around via page_is_pfmemalloc(), and when
297 			 * freed written by page_frag_free call.
298 			 */
299 			prefetchw(page);
300 		}
301 
302 		m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, n, skbs);
303 		if (unlikely(m == 0)) {
304 			for (i = 0; i < n; i++)
305 				skbs[i] = NULL; /* effect: xdp_return_frame */
306 			drops = n;
307 		}
308 
309 		local_bh_disable();
310 		for (i = 0; i < n; i++) {
311 			struct xdp_frame *xdpf = frames[i];
312 			struct sk_buff *skb = skbs[i];
313 			int ret;
314 
315 			skb = cpu_map_build_skb(rcpu, xdpf, skb);
316 			if (!skb) {
317 				xdp_return_frame(xdpf);
318 				continue;
319 			}
320 
321 			/* Inject into network stack */
322 			ret = netif_receive_skb_core(skb);
323 			if (ret == NET_RX_DROP)
324 				drops++;
325 		}
326 		/* Feedback loop via tracepoint */
327 		trace_xdp_cpumap_kthread(rcpu->map_id, n, drops, sched);
328 
329 		local_bh_enable(); /* resched point, may call do_softirq() */
330 	}
331 	__set_current_state(TASK_RUNNING);
332 
333 	put_cpu_map_entry(rcpu);
334 	return 0;
335 }
336 
__cpu_map_entry_alloc(u32 qsize,u32 cpu,int map_id)337 static struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu,
338 						       int map_id)
339 {
340 	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
341 	struct bpf_cpu_map_entry *rcpu;
342 	struct xdp_bulk_queue *bq;
343 	int numa, err, i;
344 
345 	/* Have map->numa_node, but choose node of redirect target CPU */
346 	numa = cpu_to_node(cpu);
347 
348 	rcpu = kzalloc_node(sizeof(*rcpu), gfp, numa);
349 	if (!rcpu)
350 		return NULL;
351 
352 	/* Alloc percpu bulkq */
353 	rcpu->bulkq = __alloc_percpu_gfp(sizeof(*rcpu->bulkq),
354 					 sizeof(void *), gfp);
355 	if (!rcpu->bulkq)
356 		goto free_rcu;
357 
358 	for_each_possible_cpu(i) {
359 		bq = per_cpu_ptr(rcpu->bulkq, i);
360 		bq->obj = rcpu;
361 	}
362 
363 	/* Alloc queue */
364 	rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
365 	if (!rcpu->queue)
366 		goto free_bulkq;
367 
368 	err = ptr_ring_init(rcpu->queue, qsize, gfp);
369 	if (err)
370 		goto free_queue;
371 
372 	rcpu->cpu    = cpu;
373 	rcpu->map_id = map_id;
374 	rcpu->qsize  = qsize;
375 
376 	/* Setup kthread */
377 	rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
378 					       "cpumap/%d/map:%d", cpu, map_id);
379 	if (IS_ERR(rcpu->kthread))
380 		goto free_ptr_ring;
381 
382 	get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
383 	get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
384 
385 	/* Make sure kthread runs on a single CPU */
386 	kthread_bind(rcpu->kthread, cpu);
387 	wake_up_process(rcpu->kthread);
388 
389 	return rcpu;
390 
391 free_ptr_ring:
392 	ptr_ring_cleanup(rcpu->queue, NULL);
393 free_queue:
394 	kfree(rcpu->queue);
395 free_bulkq:
396 	free_percpu(rcpu->bulkq);
397 free_rcu:
398 	kfree(rcpu);
399 	return NULL;
400 }
401 
__cpu_map_entry_free(struct rcu_head * rcu)402 static void __cpu_map_entry_free(struct rcu_head *rcu)
403 {
404 	struct bpf_cpu_map_entry *rcpu;
405 	int cpu;
406 
407 	/* This cpu_map_entry have been disconnected from map and one
408 	 * RCU graze-period have elapsed.  Thus, XDP cannot queue any
409 	 * new packets and cannot change/set flush_needed that can
410 	 * find this entry.
411 	 */
412 	rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
413 
414 	/* Flush remaining packets in percpu bulkq */
415 	for_each_online_cpu(cpu) {
416 		struct xdp_bulk_queue *bq = per_cpu_ptr(rcpu->bulkq, cpu);
417 
418 		/* No concurrent bq_enqueue can run at this point */
419 		bq_flush_to_queue(bq, false);
420 	}
421 	free_percpu(rcpu->bulkq);
422 	/* Cannot kthread_stop() here, last put free rcpu resources */
423 	put_cpu_map_entry(rcpu);
424 }
425 
426 /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
427  * ensure any driver rcu critical sections have completed, but this
428  * does not guarantee a flush has happened yet. Because driver side
429  * rcu_read_lock/unlock only protects the running XDP program.  The
430  * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
431  * pending flush op doesn't fail.
432  *
433  * The bpf_cpu_map_entry is still used by the kthread, and there can
434  * still be pending packets (in queue and percpu bulkq).  A refcnt
435  * makes sure to last user (kthread_stop vs. call_rcu) free memory
436  * resources.
437  *
438  * The rcu callback __cpu_map_entry_free flush remaining packets in
439  * percpu bulkq to queue.  Due to caller map_delete_elem() disable
440  * preemption, cannot call kthread_stop() to make sure queue is empty.
441  * Instead a work_queue is started for stopping kthread,
442  * cpu_map_kthread_stop, which waits for an RCU graze period before
443  * stopping kthread, emptying the queue.
444  */
__cpu_map_entry_replace(struct bpf_cpu_map * cmap,u32 key_cpu,struct bpf_cpu_map_entry * rcpu)445 static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
446 				    u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
447 {
448 	struct bpf_cpu_map_entry *old_rcpu;
449 
450 	old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu);
451 	if (old_rcpu) {
452 		call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
453 		INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
454 		schedule_work(&old_rcpu->kthread_stop_wq);
455 	}
456 }
457 
cpu_map_delete_elem(struct bpf_map * map,void * key)458 static int cpu_map_delete_elem(struct bpf_map *map, void *key)
459 {
460 	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
461 	u32 key_cpu = *(u32 *)key;
462 
463 	if (key_cpu >= map->max_entries)
464 		return -EINVAL;
465 
466 	/* notice caller map_delete_elem() use preempt_disable() */
467 	__cpu_map_entry_replace(cmap, key_cpu, NULL);
468 	return 0;
469 }
470 
cpu_map_update_elem(struct bpf_map * map,void * key,void * value,u64 map_flags)471 static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
472 			       u64 map_flags)
473 {
474 	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
475 	struct bpf_cpu_map_entry *rcpu;
476 
477 	/* Array index key correspond to CPU number */
478 	u32 key_cpu = *(u32 *)key;
479 	/* Value is the queue size */
480 	u32 qsize = *(u32 *)value;
481 
482 	if (unlikely(map_flags > BPF_EXIST))
483 		return -EINVAL;
484 	if (unlikely(key_cpu >= cmap->map.max_entries))
485 		return -E2BIG;
486 	if (unlikely(map_flags == BPF_NOEXIST))
487 		return -EEXIST;
488 	if (unlikely(qsize > 16384)) /* sanity limit on qsize */
489 		return -EOVERFLOW;
490 
491 	/* Make sure CPU is a valid possible cpu */
492 	if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
493 		return -ENODEV;
494 
495 	if (qsize == 0) {
496 		rcpu = NULL; /* Same as deleting */
497 	} else {
498 		/* Updating qsize cause re-allocation of bpf_cpu_map_entry */
499 		rcpu = __cpu_map_entry_alloc(qsize, key_cpu, map->id);
500 		if (!rcpu)
501 			return -ENOMEM;
502 		rcpu->cmap = cmap;
503 	}
504 	rcu_read_lock();
505 	__cpu_map_entry_replace(cmap, key_cpu, rcpu);
506 	rcu_read_unlock();
507 	return 0;
508 }
509 
cpu_map_free(struct bpf_map * map)510 static void cpu_map_free(struct bpf_map *map)
511 {
512 	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
513 	int cpu;
514 	u32 i;
515 
516 	/* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
517 	 * so the bpf programs (can be more than one that used this map) were
518 	 * disconnected from events. Wait for outstanding critical sections in
519 	 * these programs to complete. The rcu critical section only guarantees
520 	 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
521 	 * It does __not__ ensure pending flush operations (if any) are
522 	 * complete.
523 	 */
524 
525 	bpf_clear_redirect_map(map);
526 	synchronize_rcu();
527 
528 	/* To ensure all pending flush operations have completed wait for flush
529 	 * list be empty on _all_ cpus. Because the above synchronize_rcu()
530 	 * ensures the map is disconnected from the program we can assume no new
531 	 * items will be added to the list.
532 	 */
533 	for_each_online_cpu(cpu) {
534 		struct list_head *flush_list = per_cpu_ptr(cmap->flush_list, cpu);
535 
536 		while (!list_empty(flush_list))
537 			cond_resched();
538 	}
539 
540 	/* For cpu_map the remote CPUs can still be using the entries
541 	 * (struct bpf_cpu_map_entry).
542 	 */
543 	for (i = 0; i < cmap->map.max_entries; i++) {
544 		struct bpf_cpu_map_entry *rcpu;
545 
546 		rcpu = READ_ONCE(cmap->cpu_map[i]);
547 		if (!rcpu)
548 			continue;
549 
550 		/* bq flush and cleanup happens after RCU graze-period */
551 		__cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
552 	}
553 	free_percpu(cmap->flush_list);
554 	bpf_map_area_free(cmap->cpu_map);
555 	kfree(cmap);
556 }
557 
__cpu_map_lookup_elem(struct bpf_map * map,u32 key)558 struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
559 {
560 	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
561 	struct bpf_cpu_map_entry *rcpu;
562 
563 	if (key >= map->max_entries)
564 		return NULL;
565 
566 	rcpu = READ_ONCE(cmap->cpu_map[key]);
567 	return rcpu;
568 }
569 
cpu_map_lookup_elem(struct bpf_map * map,void * key)570 static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
571 {
572 	struct bpf_cpu_map_entry *rcpu =
573 		__cpu_map_lookup_elem(map, *(u32 *)key);
574 
575 	return rcpu ? &rcpu->qsize : NULL;
576 }
577 
cpu_map_get_next_key(struct bpf_map * map,void * key,void * next_key)578 static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
579 {
580 	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
581 	u32 index = key ? *(u32 *)key : U32_MAX;
582 	u32 *next = next_key;
583 
584 	if (index >= cmap->map.max_entries) {
585 		*next = 0;
586 		return 0;
587 	}
588 
589 	if (index == cmap->map.max_entries - 1)
590 		return -ENOENT;
591 	*next = index + 1;
592 	return 0;
593 }
594 
595 const struct bpf_map_ops cpu_map_ops = {
596 	.map_alloc		= cpu_map_alloc,
597 	.map_free		= cpu_map_free,
598 	.map_delete_elem	= cpu_map_delete_elem,
599 	.map_update_elem	= cpu_map_update_elem,
600 	.map_lookup_elem	= cpu_map_lookup_elem,
601 	.map_get_next_key	= cpu_map_get_next_key,
602 	.map_check_btf		= map_check_no_btf,
603 };
604 
bq_flush_to_queue(struct xdp_bulk_queue * bq,bool in_napi_ctx)605 static int bq_flush_to_queue(struct xdp_bulk_queue *bq, bool in_napi_ctx)
606 {
607 	struct bpf_cpu_map_entry *rcpu = bq->obj;
608 	unsigned int processed = 0, drops = 0;
609 	const int to_cpu = rcpu->cpu;
610 	struct ptr_ring *q;
611 	int i;
612 
613 	if (unlikely(!bq->count))
614 		return 0;
615 
616 	q = rcpu->queue;
617 	spin_lock(&q->producer_lock);
618 
619 	for (i = 0; i < bq->count; i++) {
620 		struct xdp_frame *xdpf = bq->q[i];
621 		int err;
622 
623 		err = __ptr_ring_produce(q, xdpf);
624 		if (err) {
625 			drops++;
626 			if (likely(in_napi_ctx))
627 				xdp_return_frame_rx_napi(xdpf);
628 			else
629 				xdp_return_frame(xdpf);
630 		}
631 		processed++;
632 	}
633 	bq->count = 0;
634 	spin_unlock(&q->producer_lock);
635 
636 	__list_del_clearprev(&bq->flush_node);
637 
638 	/* Feedback loop via tracepoints */
639 	trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
640 	return 0;
641 }
642 
643 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
644  * Thus, safe percpu variable access.
645  */
bq_enqueue(struct bpf_cpu_map_entry * rcpu,struct xdp_frame * xdpf)646 static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
647 {
648 	struct list_head *flush_list = this_cpu_ptr(rcpu->cmap->flush_list);
649 	struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
650 
651 	if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
652 		bq_flush_to_queue(bq, true);
653 
654 	/* Notice, xdp_buff/page MUST be queued here, long enough for
655 	 * driver to code invoking us to finished, due to driver
656 	 * (e.g. ixgbe) recycle tricks based on page-refcnt.
657 	 *
658 	 * Thus, incoming xdp_frame is always queued here (else we race
659 	 * with another CPU on page-refcnt and remaining driver code).
660 	 * Queue time is very short, as driver will invoke flush
661 	 * operation, when completing napi->poll call.
662 	 */
663 	bq->q[bq->count++] = xdpf;
664 
665 	if (!bq->flush_node.prev)
666 		list_add(&bq->flush_node, flush_list);
667 
668 	return 0;
669 }
670 
cpu_map_enqueue(struct bpf_cpu_map_entry * rcpu,struct xdp_buff * xdp,struct net_device * dev_rx)671 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
672 		    struct net_device *dev_rx)
673 {
674 	struct xdp_frame *xdpf;
675 
676 	xdpf = convert_to_xdp_frame(xdp);
677 	if (unlikely(!xdpf))
678 		return -EOVERFLOW;
679 
680 	/* Info needed when constructing SKB on remote CPU */
681 	xdpf->dev_rx = dev_rx;
682 
683 	bq_enqueue(rcpu, xdpf);
684 	return 0;
685 }
686 
__cpu_map_flush(struct bpf_map * map)687 void __cpu_map_flush(struct bpf_map *map)
688 {
689 	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
690 	struct list_head *flush_list = this_cpu_ptr(cmap->flush_list);
691 	struct xdp_bulk_queue *bq, *tmp;
692 
693 	list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
694 		bq_flush_to_queue(bq, true);
695 
696 		/* If already running, costs spin_lock_irqsave + smb_mb */
697 		wake_up_process(bq->obj->kthread);
698 	}
699 }
700