Linux Audio

Check our new training course

Loading...
v5.4
  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	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
 71struct 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
 78static int bq_flush_to_queue(struct xdp_bulk_queue *bq, bool in_napi_ctx);
 79
 80static 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;
133free_percpu:
134	free_percpu(cmap->flush_list);
135free_charge:
136	bpf_map_charge_finish(&cmap->map.memory);
137free_cmap:
138	kfree(cmap);
139	return ERR_PTR(err);
140}
141
142static 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 */
148static 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
163static 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
222static 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
236static 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
249static int cpu_map_kthread_run(void *data)
250{
251	struct bpf_cpu_map_entry *rcpu = data;
252
 
253	set_current_state(TASK_INTERRUPTIBLE);
254
255	/* When kthread gives stop order, then rcpu have been disconnected
256	 * from map, thus no new packets can enter. Remaining in-flight
257	 * per CPU stored packets are flushed to this queue.  Wait honoring
258	 * kthread_stop signal until queue is empty.
259	 */
260	while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
261		unsigned int drops = 0, sched = 0;
 
 
 
262		void *frames[CPUMAP_BATCH];
263		void *skbs[CPUMAP_BATCH];
264		gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
265		int i, n, m;
266
267		/* Release CPU reschedule checks */
268		if (__ptr_ring_empty(rcpu->queue)) {
269			set_current_state(TASK_INTERRUPTIBLE);
270			/* Recheck to avoid lost wake-up */
271			if (__ptr_ring_empty(rcpu->queue)) {
272				schedule();
273				sched = 1;
274			} else {
275				__set_current_state(TASK_RUNNING);
276			}
277		} else {
278			sched = cond_resched();
279		}
280
281		/*
282		 * The bpf_cpu_map_entry is single consumer, with this
283		 * kthread CPU pinned. Lockless access to ptr_ring
284		 * consume side valid as no-resize allowed of queue.
285		 */
286		n = ptr_ring_consume_batched(rcpu->queue, frames, CPUMAP_BATCH);
287
288		for (i = 0; i < n; i++) {
289			void *f = frames[i];
290			struct page *page = virt_to_page(f);
 
 
 
 
 
 
 
 
 
 
 
291
292			/* Bring struct page memory area to curr CPU. Read by
293			 * build_skb_around via page_is_pfmemalloc(), and when
294			 * freed written by page_frag_free call.
295			 */
296			prefetchw(page);
297		}
298
299		m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, n, skbs);
300		if (unlikely(m == 0)) {
301			for (i = 0; i < n; i++)
302				skbs[i] = NULL; /* effect: xdp_return_frame */
303			drops = n;
 
 
 
 
304		}
305
306		local_bh_disable();
307		for (i = 0; i < n; i++) {
308			struct xdp_frame *xdpf = frames[i];
309			struct sk_buff *skb = skbs[i];
310			int ret;
311
312			skb = cpu_map_build_skb(rcpu, xdpf, skb);
 
313			if (!skb) {
314				xdp_return_frame(xdpf);
315				continue;
316			}
317
318			/* Inject into network stack */
319			ret = netif_receive_skb_core(skb);
320			if (ret == NET_RX_DROP)
321				drops++;
322		}
 
 
323		/* Feedback loop via tracepoint */
324		trace_xdp_cpumap_kthread(rcpu->map_id, n, drops, sched);
 
325
326		local_bh_enable(); /* resched point, may call do_softirq() */
327	}
328	__set_current_state(TASK_RUNNING);
329
330	put_cpu_map_entry(rcpu);
331	return 0;
332}
333
334static struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu,
335						       int map_id)
336{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
337	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
338	struct bpf_cpu_map_entry *rcpu;
339	struct xdp_bulk_queue *bq;
340	int numa, err, i;
341
342	/* Have map->numa_node, but choose node of redirect target CPU */
343	numa = cpu_to_node(cpu);
344
345	rcpu = kzalloc_node(sizeof(*rcpu), gfp, numa);
346	if (!rcpu)
347		return NULL;
348
349	/* Alloc percpu bulkq */
350	rcpu->bulkq = __alloc_percpu_gfp(sizeof(*rcpu->bulkq),
351					 sizeof(void *), gfp);
352	if (!rcpu->bulkq)
353		goto free_rcu;
354
355	for_each_possible_cpu(i) {
356		bq = per_cpu_ptr(rcpu->bulkq, i);
357		bq->obj = rcpu;
358	}
359
360	/* Alloc queue */
361	rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
 
362	if (!rcpu->queue)
363		goto free_bulkq;
364
365	err = ptr_ring_init(rcpu->queue, qsize, gfp);
366	if (err)
367		goto free_queue;
368
369	rcpu->cpu    = cpu;
370	rcpu->map_id = map_id;
371	rcpu->qsize  = qsize;
 
 
 
372
373	/* Setup kthread */
 
374	rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
375					       "cpumap/%d/map:%d", cpu, map_id);
 
376	if (IS_ERR(rcpu->kthread))
377		goto free_ptr_ring;
378
379	get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
380	get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
381
382	/* Make sure kthread runs on a single CPU */
383	kthread_bind(rcpu->kthread, cpu);
384	wake_up_process(rcpu->kthread);
385
 
 
 
 
 
 
386	return rcpu;
387
 
 
 
388free_ptr_ring:
389	ptr_ring_cleanup(rcpu->queue, NULL);
390free_queue:
391	kfree(rcpu->queue);
392free_bulkq:
393	free_percpu(rcpu->bulkq);
394free_rcu:
395	kfree(rcpu);
396	return NULL;
397}
398
399static void __cpu_map_entry_free(struct rcu_head *rcu)
400{
401	struct bpf_cpu_map_entry *rcpu;
402	int cpu;
403
404	/* This cpu_map_entry have been disconnected from map and one
405	 * RCU graze-period have elapsed.  Thus, XDP cannot queue any
406	 * new packets and cannot change/set flush_needed that can
407	 * find this entry.
408	 */
409	rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
410
411	/* Flush remaining packets in percpu bulkq */
412	for_each_online_cpu(cpu) {
413		struct xdp_bulk_queue *bq = per_cpu_ptr(rcpu->bulkq, cpu);
 
 
414
415		/* No concurrent bq_enqueue can run at this point */
416		bq_flush_to_queue(bq, false);
417	}
 
 
 
418	free_percpu(rcpu->bulkq);
419	/* Cannot kthread_stop() here, last put free rcpu resources */
420	put_cpu_map_entry(rcpu);
421}
422
423/* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
424 * ensure any driver rcu critical sections have completed, but this
425 * does not guarantee a flush has happened yet. Because driver side
426 * rcu_read_lock/unlock only protects the running XDP program.  The
427 * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
428 * pending flush op doesn't fail.
429 *
430 * The bpf_cpu_map_entry is still used by the kthread, and there can
431 * still be pending packets (in queue and percpu bulkq).  A refcnt
432 * makes sure to last user (kthread_stop vs. call_rcu) free memory
433 * resources.
434 *
435 * The rcu callback __cpu_map_entry_free flush remaining packets in
436 * percpu bulkq to queue.  Due to caller map_delete_elem() disable
437 * preemption, cannot call kthread_stop() to make sure queue is empty.
438 * Instead a work_queue is started for stopping kthread,
439 * cpu_map_kthread_stop, which waits for an RCU graze period before
440 * stopping kthread, emptying the queue.
441 */
442static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
443				    u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
444{
445	struct bpf_cpu_map_entry *old_rcpu;
446
447	old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu);
448	if (old_rcpu) {
449		call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
450		INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
451		schedule_work(&old_rcpu->kthread_stop_wq);
452	}
453}
454
455static int cpu_map_delete_elem(struct bpf_map *map, void *key)
456{
457	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
458	u32 key_cpu = *(u32 *)key;
459
460	if (key_cpu >= map->max_entries)
461		return -EINVAL;
462
463	/* notice caller map_delete_elem() use preempt_disable() */
464	__cpu_map_entry_replace(cmap, key_cpu, NULL);
465	return 0;
466}
467
468static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
469			       u64 map_flags)
470{
471	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 
472	struct bpf_cpu_map_entry *rcpu;
473
474	/* Array index key correspond to CPU number */
475	u32 key_cpu = *(u32 *)key;
476	/* Value is the queue size */
477	u32 qsize = *(u32 *)value;
478
479	if (unlikely(map_flags > BPF_EXIST))
480		return -EINVAL;
481	if (unlikely(key_cpu >= cmap->map.max_entries))
482		return -E2BIG;
483	if (unlikely(map_flags == BPF_NOEXIST))
484		return -EEXIST;
485	if (unlikely(qsize > 16384)) /* sanity limit on qsize */
486		return -EOVERFLOW;
487
488	/* Make sure CPU is a valid possible cpu */
489	if (!cpu_possible(key_cpu))
490		return -ENODEV;
491
492	if (qsize == 0) {
493		rcpu = NULL; /* Same as deleting */
494	} else {
495		/* Updating qsize cause re-allocation of bpf_cpu_map_entry */
496		rcpu = __cpu_map_entry_alloc(qsize, key_cpu, map->id);
497		if (!rcpu)
498			return -ENOMEM;
499		rcpu->cmap = cmap;
500	}
501	rcu_read_lock();
502	__cpu_map_entry_replace(cmap, key_cpu, rcpu);
503	rcu_read_unlock();
504	return 0;
505}
506
507static void cpu_map_free(struct bpf_map *map)
508{
509	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
510	int cpu;
511	u32 i;
512
513	/* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
514	 * so the bpf programs (can be more than one that used this map) were
515	 * disconnected from events. Wait for outstanding critical sections in
516	 * these programs to complete. The rcu critical section only guarantees
517	 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
518	 * It does __not__ ensure pending flush operations (if any) are
519	 * complete.
520	 */
521
522	bpf_clear_redirect_map(map);
523	synchronize_rcu();
524
525	/* To ensure all pending flush operations have completed wait for flush
526	 * list be empty on _all_ cpus. Because the above synchronize_rcu()
527	 * ensures the map is disconnected from the program we can assume no new
528	 * items will be added to the list.
529	 */
530	for_each_online_cpu(cpu) {
531		struct list_head *flush_list = per_cpu_ptr(cmap->flush_list, cpu);
532
533		while (!list_empty(flush_list))
534			cond_resched();
535	}
536
537	/* For cpu_map the remote CPUs can still be using the entries
538	 * (struct bpf_cpu_map_entry).
539	 */
540	for (i = 0; i < cmap->map.max_entries; i++) {
541		struct bpf_cpu_map_entry *rcpu;
542
543		rcpu = READ_ONCE(cmap->cpu_map[i]);
544		if (!rcpu)
545			continue;
546
547		/* bq flush and cleanup happens after RCU graze-period */
548		__cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
549	}
550	free_percpu(cmap->flush_list);
551	bpf_map_area_free(cmap->cpu_map);
552	kfree(cmap);
553}
554
555struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
 
 
 
 
556{
557	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
558	struct bpf_cpu_map_entry *rcpu;
559
560	if (key >= map->max_entries)
561		return NULL;
562
563	rcpu = READ_ONCE(cmap->cpu_map[key]);
 
564	return rcpu;
565}
566
567static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
568{
569	struct bpf_cpu_map_entry *rcpu =
570		__cpu_map_lookup_elem(map, *(u32 *)key);
571
572	return rcpu ? &rcpu->qsize : NULL;
573}
574
575static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
576{
577	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
578	u32 index = key ? *(u32 *)key : U32_MAX;
579	u32 *next = next_key;
580
581	if (index >= cmap->map.max_entries) {
582		*next = 0;
583		return 0;
584	}
585
586	if (index == cmap->map.max_entries - 1)
587		return -ENOENT;
588	*next = index + 1;
589	return 0;
590}
591
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
592const struct bpf_map_ops cpu_map_ops = {
 
593	.map_alloc		= cpu_map_alloc,
594	.map_free		= cpu_map_free,
595	.map_delete_elem	= cpu_map_delete_elem,
596	.map_update_elem	= cpu_map_update_elem,
597	.map_lookup_elem	= cpu_map_lookup_elem,
598	.map_get_next_key	= cpu_map_get_next_key,
599	.map_check_btf		= map_check_no_btf,
 
 
 
600};
601
602static int bq_flush_to_queue(struct xdp_bulk_queue *bq, bool in_napi_ctx)
603{
604	struct bpf_cpu_map_entry *rcpu = bq->obj;
605	unsigned int processed = 0, drops = 0;
606	const int to_cpu = rcpu->cpu;
607	struct ptr_ring *q;
608	int i;
609
610	if (unlikely(!bq->count))
611		return 0;
612
613	q = rcpu->queue;
614	spin_lock(&q->producer_lock);
615
616	for (i = 0; i < bq->count; i++) {
617		struct xdp_frame *xdpf = bq->q[i];
618		int err;
619
620		err = __ptr_ring_produce(q, xdpf);
621		if (err) {
622			drops++;
623			if (likely(in_napi_ctx))
624				xdp_return_frame_rx_napi(xdpf);
625			else
626				xdp_return_frame(xdpf);
627		}
628		processed++;
629	}
630	bq->count = 0;
631	spin_unlock(&q->producer_lock);
632
633	__list_del_clearprev(&bq->flush_node);
634
635	/* Feedback loop via tracepoints */
636	trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
637	return 0;
638}
639
640/* Runs under RCU-read-side, plus in softirq under NAPI protection.
641 * Thus, safe percpu variable access.
642 */
643static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
644{
645	struct list_head *flush_list = this_cpu_ptr(rcpu->cmap->flush_list);
646	struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
647
648	if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
649		bq_flush_to_queue(bq, true);
650
651	/* Notice, xdp_buff/page MUST be queued here, long enough for
652	 * driver to code invoking us to finished, due to driver
653	 * (e.g. ixgbe) recycle tricks based on page-refcnt.
654	 *
655	 * Thus, incoming xdp_frame is always queued here (else we race
656	 * with another CPU on page-refcnt and remaining driver code).
657	 * Queue time is very short, as driver will invoke flush
658	 * operation, when completing napi->poll call.
659	 */
660	bq->q[bq->count++] = xdpf;
661
662	if (!bq->flush_node.prev)
663		list_add(&bq->flush_node, flush_list);
664
665	return 0;
666}
667
668int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
669		    struct net_device *dev_rx)
670{
671	struct xdp_frame *xdpf;
672
673	xdpf = convert_to_xdp_frame(xdp);
674	if (unlikely(!xdpf))
675		return -EOVERFLOW;
676
677	/* Info needed when constructing SKB on remote CPU */
678	xdpf->dev_rx = dev_rx;
679
680	bq_enqueue(rcpu, xdpf);
681	return 0;
682}
683
684void __cpu_map_flush(struct bpf_map *map)
 
685{
686	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
687	struct list_head *flush_list = this_cpu_ptr(cmap->flush_list);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
688	struct xdp_bulk_queue *bq, *tmp;
689
690	list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
691		bq_flush_to_queue(bq, true);
692
693		/* If already running, costs spin_lock_irqsave + smb_mb */
694		wake_up_process(bq->obj->kthread);
695	}
696}
v6.8
  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/completion.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	/* Queue with potential multi-producers, and single-consumer kthread */
 65	struct ptr_ring *queue;
 66	struct task_struct *kthread;
 
 67
 68	struct bpf_cpumap_val value;
 69	struct bpf_prog *prog;
 70
 71	struct completion kthread_running;
 72	struct rcu_work free_work;
 73};
 74
 75struct bpf_cpu_map {
 76	struct bpf_map map;
 77	/* Below members specific for map type */
 78	struct bpf_cpu_map_entry __rcu **cpu_map;
 
 79};
 80
 81static DEFINE_PER_CPU(struct list_head, cpu_map_flush_list);
 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(-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(-E2BIG);
 98
 99	cmap = bpf_map_area_alloc(sizeof(*cmap), NUMA_NO_NODE);
100	if (!cmap)
101		return ERR_PTR(-ENOMEM);
102
103	bpf_map_init_from_attr(&cmap->map, attr);
104
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
105	/* Alloc array for possible remote "destination" CPUs */
106	cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
107					   sizeof(struct bpf_cpu_map_entry *),
108					   cmap->map.numa_node);
109	if (!cmap->cpu_map) {
110		bpf_map_area_free(cmap);
111		return ERR_PTR(-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(ring))) {
127		WARN_ON_ONCE(1);
128		if (unlikely(__ptr_test_bit(0, &ptr))) {
129			__ptr_clear_bit(0, &ptr);
130			kfree_skb(ptr);
131			continue;
132		}
133		xdp_return_frame(ptr);
134	}
135}
136
137static void cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
138				     struct list_head *listp,
139				     struct xdp_cpumap_stats *stats)
140{
141	struct sk_buff *skb, *tmp;
142	struct xdp_buff xdp;
143	u32 act;
144	int err;
145
146	list_for_each_entry_safe(skb, tmp, listp, list) {
147		act = bpf_prog_run_generic_xdp(skb, &xdp, rcpu->prog);
148		switch (act) {
149		case XDP_PASS:
150			break;
151		case XDP_REDIRECT:
152			skb_list_del_init(skb);
153			err = xdp_do_generic_redirect(skb->dev, skb, &xdp,
154						      rcpu->prog);
155			if (unlikely(err)) {
156				kfree_skb(skb);
157				stats->drop++;
158			} else {
159				stats->redirect++;
160			}
161			return;
162		default:
163			bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
164			fallthrough;
165		case XDP_ABORTED:
166			trace_xdp_exception(skb->dev, rcpu->prog, act);
167			fallthrough;
168		case XDP_DROP:
169			skb_list_del_init(skb);
170			kfree_skb(skb);
171			stats->drop++;
172			return;
173		}
174	}
175}
176
177static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
178				    void **frames, int n,
179				    struct xdp_cpumap_stats *stats)
180{
181	struct xdp_rxq_info rxq = {};
182	struct xdp_buff xdp;
183	int i, nframes = 0;
184
185	xdp_set_return_frame_no_direct();
186	xdp.rxq = &rxq;
187
188	for (i = 0; i < n; i++) {
189		struct xdp_frame *xdpf = frames[i];
190		u32 act;
191		int err;
192
193		rxq.dev = xdpf->dev_rx;
194		rxq.mem = xdpf->mem;
195		/* TODO: report queue_index to xdp_rxq_info */
196
197		xdp_convert_frame_to_buff(xdpf, &xdp);
198
199		act = bpf_prog_run_xdp(rcpu->prog, &xdp);
200		switch (act) {
201		case XDP_PASS:
202			err = xdp_update_frame_from_buff(&xdp, xdpf);
203			if (err < 0) {
204				xdp_return_frame(xdpf);
205				stats->drop++;
206			} else {
207				frames[nframes++] = xdpf;
208				stats->pass++;
209			}
210			break;
211		case XDP_REDIRECT:
212			err = xdp_do_redirect(xdpf->dev_rx, &xdp,
213					      rcpu->prog);
214			if (unlikely(err)) {
215				xdp_return_frame(xdpf);
216				stats->drop++;
217			} else {
218				stats->redirect++;
219			}
220			break;
221		default:
222			bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
223			fallthrough;
224		case XDP_DROP:
225			xdp_return_frame(xdpf);
226			stats->drop++;
227			break;
228		}
229	}
230
231	xdp_clear_return_frame_no_direct();
232
233	return nframes;
234}
235
236#define CPUMAP_BATCH 8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
237
238static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
239				int xdp_n, struct xdp_cpumap_stats *stats,
240				struct list_head *list)
241{
242	int nframes;
243
244	if (!rcpu->prog)
245		return xdp_n;
 
 
 
 
 
 
 
 
 
 
 
246
247	rcu_read_lock_bh();
 
248
249	nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, xdp_n, stats);
 
250
251	if (stats->redirect)
252		xdp_do_flush();
253
254	if (unlikely(!list_empty(list)))
255		cpu_map_bpf_prog_run_skb(rcpu, list, stats);
 
 
 
 
 
 
256
257	rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
 
 
 
258
259	return nframes;
 
 
 
 
 
 
 
 
260}
261
 
 
262static int cpu_map_kthread_run(void *data)
263{
264	struct bpf_cpu_map_entry *rcpu = data;
265
266	complete(&rcpu->kthread_running);
267	set_current_state(TASK_INTERRUPTIBLE);
268
269	/* When kthread gives stop order, then rcpu have been disconnected
270	 * from map, thus no new packets can enter. Remaining in-flight
271	 * per CPU stored packets are flushed to this queue.  Wait honoring
272	 * kthread_stop signal until queue is empty.
273	 */
274	while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
275		struct xdp_cpumap_stats stats = {}; /* zero stats */
276		unsigned int kmem_alloc_drops = 0, sched = 0;
277		gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
278		int i, n, m, nframes, xdp_n;
279		void *frames[CPUMAP_BATCH];
280		void *skbs[CPUMAP_BATCH];
281		LIST_HEAD(list);
 
282
283		/* Release CPU reschedule checks */
284		if (__ptr_ring_empty(rcpu->queue)) {
285			set_current_state(TASK_INTERRUPTIBLE);
286			/* Recheck to avoid lost wake-up */
287			if (__ptr_ring_empty(rcpu->queue)) {
288				schedule();
289				sched = 1;
290			} else {
291				__set_current_state(TASK_RUNNING);
292			}
293		} else {
294			sched = cond_resched();
295		}
296
297		/*
298		 * The bpf_cpu_map_entry is single consumer, with this
299		 * kthread CPU pinned. Lockless access to ptr_ring
300		 * consume side valid as no-resize allowed of queue.
301		 */
302		n = __ptr_ring_consume_batched(rcpu->queue, frames,
303					       CPUMAP_BATCH);
304		for (i = 0, xdp_n = 0; i < n; i++) {
305			void *f = frames[i];
306			struct page *page;
307
308			if (unlikely(__ptr_test_bit(0, &f))) {
309				struct sk_buff *skb = f;
310
311				__ptr_clear_bit(0, &skb);
312				list_add_tail(&skb->list, &list);
313				continue;
314			}
315
316			frames[xdp_n++] = f;
317			page = virt_to_page(f);
318
319			/* Bring struct page memory area to curr CPU. Read by
320			 * build_skb_around via page_is_pfmemalloc(), and when
321			 * freed written by page_frag_free call.
322			 */
323			prefetchw(page);
324		}
325
326		/* Support running another XDP prog on this CPU */
327		nframes = cpu_map_bpf_prog_run(rcpu, frames, xdp_n, &stats, &list);
328		if (nframes) {
329			m = kmem_cache_alloc_bulk(skbuff_cache, gfp, nframes, skbs);
330			if (unlikely(m == 0)) {
331				for (i = 0; i < nframes; i++)
332					skbs[i] = NULL; /* effect: xdp_return_frame */
333				kmem_alloc_drops += nframes;
334			}
335		}
336
337		local_bh_disable();
338		for (i = 0; i < nframes; i++) {
339			struct xdp_frame *xdpf = frames[i];
340			struct sk_buff *skb = skbs[i];
 
341
342			skb = __xdp_build_skb_from_frame(xdpf, skb,
343							 xdpf->dev_rx);
344			if (!skb) {
345				xdp_return_frame(xdpf);
346				continue;
347			}
348
349			list_add_tail(&skb->list, &list);
 
 
 
350		}
351		netif_receive_skb_list(&list);
352
353		/* Feedback loop via tracepoint */
354		trace_xdp_cpumap_kthread(rcpu->map_id, n, kmem_alloc_drops,
355					 sched, &stats);
356
357		local_bh_enable(); /* resched point, may call do_softirq() */
358	}
359	__set_current_state(TASK_RUNNING);
360
 
361	return 0;
362}
363
364static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu,
365				      struct bpf_map *map, int fd)
366{
367	struct bpf_prog *prog;
368
369	prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
370	if (IS_ERR(prog))
371		return PTR_ERR(prog);
372
373	if (prog->expected_attach_type != BPF_XDP_CPUMAP ||
374	    !bpf_prog_map_compatible(map, prog)) {
375		bpf_prog_put(prog);
376		return -EINVAL;
377	}
378
379	rcpu->value.bpf_prog.id = prog->aux->id;
380	rcpu->prog = prog;
381
382	return 0;
383}
384
385static struct bpf_cpu_map_entry *
386__cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value,
387		      u32 cpu)
388{
389	int numa, err, i, fd = value->bpf_prog.fd;
390	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
391	struct bpf_cpu_map_entry *rcpu;
392	struct xdp_bulk_queue *bq;
 
393
394	/* Have map->numa_node, but choose node of redirect target CPU */
395	numa = cpu_to_node(cpu);
396
397	rcpu = bpf_map_kmalloc_node(map, sizeof(*rcpu), gfp | __GFP_ZERO, numa);
398	if (!rcpu)
399		return NULL;
400
401	/* Alloc percpu bulkq */
402	rcpu->bulkq = bpf_map_alloc_percpu(map, sizeof(*rcpu->bulkq),
403					   sizeof(void *), gfp);
404	if (!rcpu->bulkq)
405		goto free_rcu;
406
407	for_each_possible_cpu(i) {
408		bq = per_cpu_ptr(rcpu->bulkq, i);
409		bq->obj = rcpu;
410	}
411
412	/* Alloc queue */
413	rcpu->queue = bpf_map_kmalloc_node(map, sizeof(*rcpu->queue), gfp,
414					   numa);
415	if (!rcpu->queue)
416		goto free_bulkq;
417
418	err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
419	if (err)
420		goto free_queue;
421
422	rcpu->cpu    = cpu;
423	rcpu->map_id = map->id;
424	rcpu->value.qsize  = value->qsize;
425
426	if (fd > 0 && __cpu_map_load_bpf_program(rcpu, map, fd))
427		goto free_ptr_ring;
428
429	/* Setup kthread */
430	init_completion(&rcpu->kthread_running);
431	rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
432					       "cpumap/%d/map:%d", cpu,
433					       map->id);
434	if (IS_ERR(rcpu->kthread))
435		goto free_prog;
 
 
 
436
437	/* Make sure kthread runs on a single CPU */
438	kthread_bind(rcpu->kthread, cpu);
439	wake_up_process(rcpu->kthread);
440
441	/* Make sure kthread has been running, so kthread_stop() will not
442	 * stop the kthread prematurely and all pending frames or skbs
443	 * will be handled by the kthread before kthread_stop() returns.
444	 */
445	wait_for_completion(&rcpu->kthread_running);
446
447	return rcpu;
448
449free_prog:
450	if (rcpu->prog)
451		bpf_prog_put(rcpu->prog);
452free_ptr_ring:
453	ptr_ring_cleanup(rcpu->queue, NULL);
454free_queue:
455	kfree(rcpu->queue);
456free_bulkq:
457	free_percpu(rcpu->bulkq);
458free_rcu:
459	kfree(rcpu);
460	return NULL;
461}
462
463static void __cpu_map_entry_free(struct work_struct *work)
464{
465	struct bpf_cpu_map_entry *rcpu;
 
466
467	/* This cpu_map_entry have been disconnected from map and one
468	 * RCU grace-period have elapsed. Thus, XDP cannot queue any
469	 * new packets and cannot change/set flush_needed that can
470	 * find this entry.
471	 */
472	rcpu = container_of(to_rcu_work(work), struct bpf_cpu_map_entry, free_work);
473
474	/* kthread_stop will wake_up_process and wait for it to complete.
475	 * cpu_map_kthread_run() makes sure the pointer ring is empty
476	 * before exiting.
477	 */
478	kthread_stop(rcpu->kthread);
479
480	if (rcpu->prog)
481		bpf_prog_put(rcpu->prog);
482	/* The queue should be empty at this point */
483	__cpu_map_ring_cleanup(rcpu->queue);
484	ptr_ring_cleanup(rcpu->queue, NULL);
485	kfree(rcpu->queue);
486	free_percpu(rcpu->bulkq);
487	kfree(rcpu);
 
488}
489
490/* After the xchg of the bpf_cpu_map_entry pointer, we need to make sure the old
491 * entry is no longer in use before freeing. We use queue_rcu_work() to call
492 * __cpu_map_entry_free() in a separate workqueue after waiting for an RCU grace
493 * period. This means that (a) all pending enqueue and flush operations have
494 * completed (because of the RCU callback), and (b) we are in a workqueue
495 * context where we can stop the kthread and wait for it to exit before freeing
496 * everything.
 
 
 
 
 
 
 
 
 
 
 
497 */
498static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
499				    u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
500{
501	struct bpf_cpu_map_entry *old_rcpu;
502
503	old_rcpu = unrcu_pointer(xchg(&cmap->cpu_map[key_cpu], RCU_INITIALIZER(rcpu)));
504	if (old_rcpu) {
505		INIT_RCU_WORK(&old_rcpu->free_work, __cpu_map_entry_free);
506		queue_rcu_work(system_wq, &old_rcpu->free_work);
 
507	}
508}
509
510static long cpu_map_delete_elem(struct bpf_map *map, void *key)
511{
512	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
513	u32 key_cpu = *(u32 *)key;
514
515	if (key_cpu >= map->max_entries)
516		return -EINVAL;
517
518	/* notice caller map_delete_elem() uses rcu_read_lock() */
519	__cpu_map_entry_replace(cmap, key_cpu, NULL);
520	return 0;
521}
522
523static long cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
524				u64 map_flags)
525{
526	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
527	struct bpf_cpumap_val cpumap_value = {};
528	struct bpf_cpu_map_entry *rcpu;
 
529	/* Array index key correspond to CPU number */
530	u32 key_cpu = *(u32 *)key;
531
532	memcpy(&cpumap_value, value, map->value_size);
533
534	if (unlikely(map_flags > BPF_EXIST))
535		return -EINVAL;
536	if (unlikely(key_cpu >= cmap->map.max_entries))
537		return -E2BIG;
538	if (unlikely(map_flags == BPF_NOEXIST))
539		return -EEXIST;
540	if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
541		return -EOVERFLOW;
542
543	/* Make sure CPU is a valid possible cpu */
544	if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
545		return -ENODEV;
546
547	if (cpumap_value.qsize == 0) {
548		rcpu = NULL; /* Same as deleting */
549	} else {
550		/* Updating qsize cause re-allocation of bpf_cpu_map_entry */
551		rcpu = __cpu_map_entry_alloc(map, &cpumap_value, key_cpu);
552		if (!rcpu)
553			return -ENOMEM;
 
554	}
555	rcu_read_lock();
556	__cpu_map_entry_replace(cmap, key_cpu, rcpu);
557	rcu_read_unlock();
558	return 0;
559}
560
561static void cpu_map_free(struct bpf_map *map)
562{
563	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
 
564	u32 i;
565
566	/* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
567	 * so the bpf programs (can be more than one that used this map) were
568	 * disconnected from events. Wait for outstanding critical sections in
569	 * these programs to complete. synchronize_rcu() below not only
570	 * guarantees no further "XDP/bpf-side" reads against
571	 * bpf_cpu_map->cpu_map, but also ensure pending flush operations
572	 * (if any) are completed.
573	 */
 
 
574	synchronize_rcu();
575
576	/* The only possible user of bpf_cpu_map_entry is
577	 * cpu_map_kthread_run().
 
 
 
 
 
 
 
 
 
 
 
 
578	 */
579	for (i = 0; i < cmap->map.max_entries; i++) {
580		struct bpf_cpu_map_entry *rcpu;
581
582		rcpu = rcu_dereference_raw(cmap->cpu_map[i]);
583		if (!rcpu)
584			continue;
585
586		/* Stop kthread and cleanup entry directly */
587		__cpu_map_entry_free(&rcpu->free_work.work);
588	}
 
589	bpf_map_area_free(cmap->cpu_map);
590	bpf_map_area_free(cmap);
591}
592
593/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
594 * by local_bh_disable() (from XDP calls inside NAPI). The
595 * rcu_read_lock_bh_held() below makes lockdep accept both.
596 */
597static void *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
598{
599	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
600	struct bpf_cpu_map_entry *rcpu;
601
602	if (key >= map->max_entries)
603		return NULL;
604
605	rcpu = rcu_dereference_check(cmap->cpu_map[key],
606				     rcu_read_lock_bh_held());
607	return rcpu;
608}
609
610static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
611{
612	struct bpf_cpu_map_entry *rcpu =
613		__cpu_map_lookup_elem(map, *(u32 *)key);
614
615	return rcpu ? &rcpu->value : NULL;
616}
617
618static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
619{
620	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
621	u32 index = key ? *(u32 *)key : U32_MAX;
622	u32 *next = next_key;
623
624	if (index >= cmap->map.max_entries) {
625		*next = 0;
626		return 0;
627	}
628
629	if (index == cmap->map.max_entries - 1)
630		return -ENOENT;
631	*next = index + 1;
632	return 0;
633}
634
635static long cpu_map_redirect(struct bpf_map *map, u64 index, u64 flags)
636{
637	return __bpf_xdp_redirect_map(map, index, flags, 0,
638				      __cpu_map_lookup_elem);
639}
640
641static u64 cpu_map_mem_usage(const struct bpf_map *map)
642{
643	u64 usage = sizeof(struct bpf_cpu_map);
644
645	/* Currently the dynamically allocated elements are not counted */
646	usage += (u64)map->max_entries * sizeof(struct bpf_cpu_map_entry *);
647	return usage;
648}
649
650BTF_ID_LIST_SINGLE(cpu_map_btf_ids, struct, bpf_cpu_map)
651const struct bpf_map_ops cpu_map_ops = {
652	.map_meta_equal		= bpf_map_meta_equal,
653	.map_alloc		= cpu_map_alloc,
654	.map_free		= cpu_map_free,
655	.map_delete_elem	= cpu_map_delete_elem,
656	.map_update_elem	= cpu_map_update_elem,
657	.map_lookup_elem	= cpu_map_lookup_elem,
658	.map_get_next_key	= cpu_map_get_next_key,
659	.map_check_btf		= map_check_no_btf,
660	.map_mem_usage		= cpu_map_mem_usage,
661	.map_btf_id		= &cpu_map_btf_ids[0],
662	.map_redirect		= cpu_map_redirect,
663};
664
665static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
666{
667	struct bpf_cpu_map_entry *rcpu = bq->obj;
668	unsigned int processed = 0, drops = 0;
669	const int to_cpu = rcpu->cpu;
670	struct ptr_ring *q;
671	int i;
672
673	if (unlikely(!bq->count))
674		return;
675
676	q = rcpu->queue;
677	spin_lock(&q->producer_lock);
678
679	for (i = 0; i < bq->count; i++) {
680		struct xdp_frame *xdpf = bq->q[i];
681		int err;
682
683		err = __ptr_ring_produce(q, xdpf);
684		if (err) {
685			drops++;
686			xdp_return_frame_rx_napi(xdpf);
 
 
 
687		}
688		processed++;
689	}
690	bq->count = 0;
691	spin_unlock(&q->producer_lock);
692
693	__list_del_clearprev(&bq->flush_node);
694
695	/* Feedback loop via tracepoints */
696	trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
 
697}
698
699/* Runs under RCU-read-side, plus in softirq under NAPI protection.
700 * Thus, safe percpu variable access.
701 */
702static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
703{
704	struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
705	struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
706
707	if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
708		bq_flush_to_queue(bq);
709
710	/* Notice, xdp_buff/page MUST be queued here, long enough for
711	 * driver to code invoking us to finished, due to driver
712	 * (e.g. ixgbe) recycle tricks based on page-refcnt.
713	 *
714	 * Thus, incoming xdp_frame is always queued here (else we race
715	 * with another CPU on page-refcnt and remaining driver code).
716	 * Queue time is very short, as driver will invoke flush
717	 * operation, when completing napi->poll call.
718	 */
719	bq->q[bq->count++] = xdpf;
720
721	if (!bq->flush_node.prev)
722		list_add(&bq->flush_node, flush_list);
 
 
723}
724
725int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf,
726		    struct net_device *dev_rx)
727{
 
 
 
 
 
 
728	/* Info needed when constructing SKB on remote CPU */
729	xdpf->dev_rx = dev_rx;
730
731	bq_enqueue(rcpu, xdpf);
732	return 0;
733}
734
735int cpu_map_generic_redirect(struct bpf_cpu_map_entry *rcpu,
736			     struct sk_buff *skb)
737{
738	int ret;
739
740	__skb_pull(skb, skb->mac_len);
741	skb_set_redirected(skb, false);
742	__ptr_set_bit(0, &skb);
743
744	ret = ptr_ring_produce(rcpu->queue, skb);
745	if (ret < 0)
746		goto trace;
747
748	wake_up_process(rcpu->kthread);
749trace:
750	trace_xdp_cpumap_enqueue(rcpu->map_id, !ret, !!ret, rcpu->cpu);
751	return ret;
752}
753
754void __cpu_map_flush(void)
755{
756	struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
757	struct xdp_bulk_queue *bq, *tmp;
758
759	list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
760		bq_flush_to_queue(bq);
761
762		/* If already running, costs spin_lock_irqsave + smb_mb */
763		wake_up_process(bq->obj->kthread);
764	}
765}
766
767#ifdef CONFIG_DEBUG_NET
768bool cpu_map_check_flush(void)
769{
770	if (list_empty(this_cpu_ptr(&cpu_map_flush_list)))
771		return false;
772	__cpu_map_flush();
773	return true;
774}
775#endif
776
777static int __init cpu_map_init(void)
778{
779	int cpu;
780
781	for_each_possible_cpu(cpu)
782		INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu));
783	return 0;
784}
785
786subsys_initcall(cpu_map_init);