Loading...
1/* bpf/cpumap.c
2 *
3 * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
4 * Released under terms in GPL version 2. See COPYING.
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
23#include <linux/sched.h>
24#include <linux/workqueue.h>
25#include <linux/kthread.h>
26#include <linux/capability.h>
27#include <trace/events/xdp.h>
28
29#include <linux/netdevice.h> /* netif_receive_skb_core */
30#include <linux/etherdevice.h> /* eth_type_trans */
31
32/* General idea: XDP packets getting XDP redirected to another CPU,
33 * will maximum be stored/queued for one driver ->poll() call. It is
34 * guaranteed that setting flush bit and flush operation happen on
35 * same CPU. Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
36 * which queue in bpf_cpu_map_entry contains packets.
37 */
38
39#define CPU_MAP_BULK_SIZE 8 /* 8 == one cacheline on 64-bit archs */
40struct xdp_bulk_queue {
41 void *q[CPU_MAP_BULK_SIZE];
42 unsigned int count;
43};
44
45/* Struct for every remote "destination" CPU in map */
46struct bpf_cpu_map_entry {
47 u32 cpu; /* kthread CPU and map index */
48 int map_id; /* Back reference to map */
49 u32 qsize; /* Queue size placeholder for map lookup */
50
51 /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
52 struct xdp_bulk_queue __percpu *bulkq;
53
54 /* Queue with potential multi-producers, and single-consumer kthread */
55 struct ptr_ring *queue;
56 struct task_struct *kthread;
57 struct work_struct kthread_stop_wq;
58
59 atomic_t refcnt; /* Control when this struct can be free'ed */
60 struct rcu_head rcu;
61};
62
63struct bpf_cpu_map {
64 struct bpf_map map;
65 /* Below members specific for map type */
66 struct bpf_cpu_map_entry **cpu_map;
67 unsigned long __percpu *flush_needed;
68};
69
70static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
71 struct xdp_bulk_queue *bq);
72
73static u64 cpu_map_bitmap_size(const union bpf_attr *attr)
74{
75 return BITS_TO_LONGS(attr->max_entries) * sizeof(unsigned long);
76}
77
78static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
79{
80 struct bpf_cpu_map *cmap;
81 int err = -ENOMEM;
82 u64 cost;
83 int ret;
84
85 if (!capable(CAP_SYS_ADMIN))
86 return ERR_PTR(-EPERM);
87
88 /* check sanity of attributes */
89 if (attr->max_entries == 0 || attr->key_size != 4 ||
90 attr->value_size != 4 || attr->map_flags & ~BPF_F_NUMA_NODE)
91 return ERR_PTR(-EINVAL);
92
93 cmap = kzalloc(sizeof(*cmap), GFP_USER);
94 if (!cmap)
95 return ERR_PTR(-ENOMEM);
96
97 bpf_map_init_from_attr(&cmap->map, attr);
98
99 /* Pre-limit array size based on NR_CPUS, not final CPU check */
100 if (cmap->map.max_entries > NR_CPUS) {
101 err = -E2BIG;
102 goto free_cmap;
103 }
104
105 /* make sure page count doesn't overflow */
106 cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *);
107 cost += cpu_map_bitmap_size(attr) * num_possible_cpus();
108 if (cost >= U32_MAX - PAGE_SIZE)
109 goto free_cmap;
110 cmap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
111
112 /* Notice returns -EPERM on if map size is larger than memlock limit */
113 ret = bpf_map_precharge_memlock(cmap->map.pages);
114 if (ret) {
115 err = ret;
116 goto free_cmap;
117 }
118
119 /* A per cpu bitfield with a bit per possible CPU in map */
120 cmap->flush_needed = __alloc_percpu(cpu_map_bitmap_size(attr),
121 __alignof__(unsigned long));
122 if (!cmap->flush_needed)
123 goto free_cmap;
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_needed);
135free_cmap:
136 kfree(cmap);
137 return ERR_PTR(err);
138}
139
140static void __cpu_map_queue_destructor(void *ptr)
141{
142 /* The tear-down procedure should have made sure that queue is
143 * empty. See __cpu_map_entry_replace() and work-queue
144 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
145 * gracefully and warn once.
146 */
147 if (WARN_ON_ONCE(ptr))
148 page_frag_free(ptr);
149}
150
151static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
152{
153 if (atomic_dec_and_test(&rcpu->refcnt)) {
154 /* The queue should be empty at this point */
155 ptr_ring_cleanup(rcpu->queue, __cpu_map_queue_destructor);
156 kfree(rcpu->queue);
157 kfree(rcpu);
158 }
159}
160
161static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
162{
163 atomic_inc(&rcpu->refcnt);
164}
165
166/* called from workqueue, to workaround syscall using preempt_disable */
167static void cpu_map_kthread_stop(struct work_struct *work)
168{
169 struct bpf_cpu_map_entry *rcpu;
170
171 rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
172
173 /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
174 * as it waits until all in-flight call_rcu() callbacks complete.
175 */
176 rcu_barrier();
177
178 /* kthread_stop will wake_up_process and wait for it to complete */
179 kthread_stop(rcpu->kthread);
180}
181
182/* For now, xdp_pkt is a cpumap internal data structure, with info
183 * carried between enqueue to dequeue. It is mapped into the top
184 * headroom of the packet, to avoid allocating separate mem.
185 */
186struct xdp_pkt {
187 void *data;
188 u16 len;
189 u16 headroom;
190 u16 metasize;
191 struct net_device *dev_rx;
192};
193
194/* Convert xdp_buff to xdp_pkt */
195static struct xdp_pkt *convert_to_xdp_pkt(struct xdp_buff *xdp)
196{
197 struct xdp_pkt *xdp_pkt;
198 int metasize;
199 int headroom;
200
201 /* Assure headroom is available for storing info */
202 headroom = xdp->data - xdp->data_hard_start;
203 metasize = xdp->data - xdp->data_meta;
204 metasize = metasize > 0 ? metasize : 0;
205 if (unlikely((headroom - metasize) < sizeof(*xdp_pkt)))
206 return NULL;
207
208 /* Store info in top of packet */
209 xdp_pkt = xdp->data_hard_start;
210
211 xdp_pkt->data = xdp->data;
212 xdp_pkt->len = xdp->data_end - xdp->data;
213 xdp_pkt->headroom = headroom - sizeof(*xdp_pkt);
214 xdp_pkt->metasize = metasize;
215
216 return xdp_pkt;
217}
218
219static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu,
220 struct xdp_pkt *xdp_pkt)
221{
222 unsigned int frame_size;
223 void *pkt_data_start;
224 struct sk_buff *skb;
225
226 /* build_skb need to place skb_shared_info after SKB end, and
227 * also want to know the memory "truesize". Thus, need to
228 * know the memory frame size backing xdp_buff.
229 *
230 * XDP was designed to have PAGE_SIZE frames, but this
231 * assumption is not longer true with ixgbe and i40e. It
232 * would be preferred to set frame_size to 2048 or 4096
233 * depending on the driver.
234 * frame_size = 2048;
235 * frame_len = frame_size - sizeof(*xdp_pkt);
236 *
237 * Instead, with info avail, skb_shared_info in placed after
238 * packet len. This, unfortunately fakes the truesize.
239 * Another disadvantage of this approach, the skb_shared_info
240 * is not at a fixed memory location, with mixed length
241 * packets, which is bad for cache-line hotness.
242 */
243 frame_size = SKB_DATA_ALIGN(xdp_pkt->len) + xdp_pkt->headroom +
244 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
245
246 pkt_data_start = xdp_pkt->data - xdp_pkt->headroom;
247 skb = build_skb(pkt_data_start, frame_size);
248 if (!skb)
249 return NULL;
250
251 skb_reserve(skb, xdp_pkt->headroom);
252 __skb_put(skb, xdp_pkt->len);
253 if (xdp_pkt->metasize)
254 skb_metadata_set(skb, xdp_pkt->metasize);
255
256 /* Essential SKB info: protocol and skb->dev */
257 skb->protocol = eth_type_trans(skb, xdp_pkt->dev_rx);
258
259 /* Optional SKB info, currently missing:
260 * - HW checksum info (skb->ip_summed)
261 * - HW RX hash (skb_set_hash)
262 * - RX ring dev queue index (skb_record_rx_queue)
263 */
264
265 return skb;
266}
267
268static int cpu_map_kthread_run(void *data)
269{
270 struct bpf_cpu_map_entry *rcpu = data;
271
272 set_current_state(TASK_INTERRUPTIBLE);
273
274 /* When kthread gives stop order, then rcpu have been disconnected
275 * from map, thus no new packets can enter. Remaining in-flight
276 * per CPU stored packets are flushed to this queue. Wait honoring
277 * kthread_stop signal until queue is empty.
278 */
279 while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
280 unsigned int processed = 0, drops = 0, sched = 0;
281 struct xdp_pkt *xdp_pkt;
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 /* Process packets in rcpu->queue */
298 local_bh_disable();
299 /*
300 * The bpf_cpu_map_entry is single consumer, with this
301 * kthread CPU pinned. Lockless access to ptr_ring
302 * consume side valid as no-resize allowed of queue.
303 */
304 while ((xdp_pkt = __ptr_ring_consume(rcpu->queue))) {
305 struct sk_buff *skb;
306 int ret;
307
308 skb = cpu_map_build_skb(rcpu, xdp_pkt);
309 if (!skb) {
310 page_frag_free(xdp_pkt);
311 continue;
312 }
313
314 /* Inject into network stack */
315 ret = netif_receive_skb_core(skb);
316 if (ret == NET_RX_DROP)
317 drops++;
318
319 /* Limit BH-disable period */
320 if (++processed == 8)
321 break;
322 }
323 /* Feedback loop via tracepoint */
324 trace_xdp_cpumap_kthread(rcpu->map_id, processed, 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 int numa, err;
340
341 /* Have map->numa_node, but choose node of redirect target CPU */
342 numa = cpu_to_node(cpu);
343
344 rcpu = kzalloc_node(sizeof(*rcpu), gfp, numa);
345 if (!rcpu)
346 return NULL;
347
348 /* Alloc percpu bulkq */
349 rcpu->bulkq = __alloc_percpu_gfp(sizeof(*rcpu->bulkq),
350 sizeof(void *), gfp);
351 if (!rcpu->bulkq)
352 goto free_rcu;
353
354 /* Alloc queue */
355 rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
356 if (!rcpu->queue)
357 goto free_bulkq;
358
359 err = ptr_ring_init(rcpu->queue, qsize, gfp);
360 if (err)
361 goto free_queue;
362
363 rcpu->cpu = cpu;
364 rcpu->map_id = map_id;
365 rcpu->qsize = qsize;
366
367 /* Setup kthread */
368 rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
369 "cpumap/%d/map:%d", cpu, map_id);
370 if (IS_ERR(rcpu->kthread))
371 goto free_ptr_ring;
372
373 get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
374 get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
375
376 /* Make sure kthread runs on a single CPU */
377 kthread_bind(rcpu->kthread, cpu);
378 wake_up_process(rcpu->kthread);
379
380 return rcpu;
381
382free_ptr_ring:
383 ptr_ring_cleanup(rcpu->queue, NULL);
384free_queue:
385 kfree(rcpu->queue);
386free_bulkq:
387 free_percpu(rcpu->bulkq);
388free_rcu:
389 kfree(rcpu);
390 return NULL;
391}
392
393static void __cpu_map_entry_free(struct rcu_head *rcu)
394{
395 struct bpf_cpu_map_entry *rcpu;
396 int cpu;
397
398 /* This cpu_map_entry have been disconnected from map and one
399 * RCU graze-period have elapsed. Thus, XDP cannot queue any
400 * new packets and cannot change/set flush_needed that can
401 * find this entry.
402 */
403 rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
404
405 /* Flush remaining packets in percpu bulkq */
406 for_each_online_cpu(cpu) {
407 struct xdp_bulk_queue *bq = per_cpu_ptr(rcpu->bulkq, cpu);
408
409 /* No concurrent bq_enqueue can run at this point */
410 bq_flush_to_queue(rcpu, bq);
411 }
412 free_percpu(rcpu->bulkq);
413 /* Cannot kthread_stop() here, last put free rcpu resources */
414 put_cpu_map_entry(rcpu);
415}
416
417/* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
418 * ensure any driver rcu critical sections have completed, but this
419 * does not guarantee a flush has happened yet. Because driver side
420 * rcu_read_lock/unlock only protects the running XDP program. The
421 * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
422 * pending flush op doesn't fail.
423 *
424 * The bpf_cpu_map_entry is still used by the kthread, and there can
425 * still be pending packets (in queue and percpu bulkq). A refcnt
426 * makes sure to last user (kthread_stop vs. call_rcu) free memory
427 * resources.
428 *
429 * The rcu callback __cpu_map_entry_free flush remaining packets in
430 * percpu bulkq to queue. Due to caller map_delete_elem() disable
431 * preemption, cannot call kthread_stop() to make sure queue is empty.
432 * Instead a work_queue is started for stopping kthread,
433 * cpu_map_kthread_stop, which waits for an RCU graze period before
434 * stopping kthread, emptying the queue.
435 */
436static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
437 u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
438{
439 struct bpf_cpu_map_entry *old_rcpu;
440
441 old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu);
442 if (old_rcpu) {
443 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
444 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
445 schedule_work(&old_rcpu->kthread_stop_wq);
446 }
447}
448
449static int cpu_map_delete_elem(struct bpf_map *map, void *key)
450{
451 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
452 u32 key_cpu = *(u32 *)key;
453
454 if (key_cpu >= map->max_entries)
455 return -EINVAL;
456
457 /* notice caller map_delete_elem() use preempt_disable() */
458 __cpu_map_entry_replace(cmap, key_cpu, NULL);
459 return 0;
460}
461
462static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
463 u64 map_flags)
464{
465 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
466 struct bpf_cpu_map_entry *rcpu;
467
468 /* Array index key correspond to CPU number */
469 u32 key_cpu = *(u32 *)key;
470 /* Value is the queue size */
471 u32 qsize = *(u32 *)value;
472
473 if (unlikely(map_flags > BPF_EXIST))
474 return -EINVAL;
475 if (unlikely(key_cpu >= cmap->map.max_entries))
476 return -E2BIG;
477 if (unlikely(map_flags == BPF_NOEXIST))
478 return -EEXIST;
479 if (unlikely(qsize > 16384)) /* sanity limit on qsize */
480 return -EOVERFLOW;
481
482 /* Make sure CPU is a valid possible cpu */
483 if (!cpu_possible(key_cpu))
484 return -ENODEV;
485
486 if (qsize == 0) {
487 rcpu = NULL; /* Same as deleting */
488 } else {
489 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
490 rcpu = __cpu_map_entry_alloc(qsize, key_cpu, map->id);
491 if (!rcpu)
492 return -ENOMEM;
493 }
494 rcu_read_lock();
495 __cpu_map_entry_replace(cmap, key_cpu, rcpu);
496 rcu_read_unlock();
497 return 0;
498}
499
500static void cpu_map_free(struct bpf_map *map)
501{
502 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
503 int cpu;
504 u32 i;
505
506 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
507 * so the bpf programs (can be more than one that used this map) were
508 * disconnected from events. Wait for outstanding critical sections in
509 * these programs to complete. The rcu critical section only guarantees
510 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
511 * It does __not__ ensure pending flush operations (if any) are
512 * complete.
513 */
514 synchronize_rcu();
515
516 /* To ensure all pending flush operations have completed wait for flush
517 * bitmap to indicate all flush_needed bits to be zero on _all_ cpus.
518 * Because the above synchronize_rcu() ensures the map is disconnected
519 * from the program we can assume no new bits will be set.
520 */
521 for_each_online_cpu(cpu) {
522 unsigned long *bitmap = per_cpu_ptr(cmap->flush_needed, cpu);
523
524 while (!bitmap_empty(bitmap, cmap->map.max_entries))
525 cond_resched();
526 }
527
528 /* For cpu_map the remote CPUs can still be using the entries
529 * (struct bpf_cpu_map_entry).
530 */
531 for (i = 0; i < cmap->map.max_entries; i++) {
532 struct bpf_cpu_map_entry *rcpu;
533
534 rcpu = READ_ONCE(cmap->cpu_map[i]);
535 if (!rcpu)
536 continue;
537
538 /* bq flush and cleanup happens after RCU graze-period */
539 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
540 }
541 free_percpu(cmap->flush_needed);
542 bpf_map_area_free(cmap->cpu_map);
543 kfree(cmap);
544}
545
546struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
547{
548 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
549 struct bpf_cpu_map_entry *rcpu;
550
551 if (key >= map->max_entries)
552 return NULL;
553
554 rcpu = READ_ONCE(cmap->cpu_map[key]);
555 return rcpu;
556}
557
558static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
559{
560 struct bpf_cpu_map_entry *rcpu =
561 __cpu_map_lookup_elem(map, *(u32 *)key);
562
563 return rcpu ? &rcpu->qsize : NULL;
564}
565
566static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
567{
568 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
569 u32 index = key ? *(u32 *)key : U32_MAX;
570 u32 *next = next_key;
571
572 if (index >= cmap->map.max_entries) {
573 *next = 0;
574 return 0;
575 }
576
577 if (index == cmap->map.max_entries - 1)
578 return -ENOENT;
579 *next = index + 1;
580 return 0;
581}
582
583const struct bpf_map_ops cpu_map_ops = {
584 .map_alloc = cpu_map_alloc,
585 .map_free = cpu_map_free,
586 .map_delete_elem = cpu_map_delete_elem,
587 .map_update_elem = cpu_map_update_elem,
588 .map_lookup_elem = cpu_map_lookup_elem,
589 .map_get_next_key = cpu_map_get_next_key,
590};
591
592static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
593 struct xdp_bulk_queue *bq)
594{
595 unsigned int processed = 0, drops = 0;
596 const int to_cpu = rcpu->cpu;
597 struct ptr_ring *q;
598 int i;
599
600 if (unlikely(!bq->count))
601 return 0;
602
603 q = rcpu->queue;
604 spin_lock(&q->producer_lock);
605
606 for (i = 0; i < bq->count; i++) {
607 void *xdp_pkt = bq->q[i];
608 int err;
609
610 err = __ptr_ring_produce(q, xdp_pkt);
611 if (err) {
612 drops++;
613 page_frag_free(xdp_pkt); /* Free xdp_pkt */
614 }
615 processed++;
616 }
617 bq->count = 0;
618 spin_unlock(&q->producer_lock);
619
620 /* Feedback loop via tracepoints */
621 trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
622 return 0;
623}
624
625/* Runs under RCU-read-side, plus in softirq under NAPI protection.
626 * Thus, safe percpu variable access.
627 */
628static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_pkt *xdp_pkt)
629{
630 struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
631
632 if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
633 bq_flush_to_queue(rcpu, bq);
634
635 /* Notice, xdp_buff/page MUST be queued here, long enough for
636 * driver to code invoking us to finished, due to driver
637 * (e.g. ixgbe) recycle tricks based on page-refcnt.
638 *
639 * Thus, incoming xdp_pkt is always queued here (else we race
640 * with another CPU on page-refcnt and remaining driver code).
641 * Queue time is very short, as driver will invoke flush
642 * operation, when completing napi->poll call.
643 */
644 bq->q[bq->count++] = xdp_pkt;
645 return 0;
646}
647
648int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
649 struct net_device *dev_rx)
650{
651 struct xdp_pkt *xdp_pkt;
652
653 xdp_pkt = convert_to_xdp_pkt(xdp);
654 if (unlikely(!xdp_pkt))
655 return -EOVERFLOW;
656
657 /* Info needed when constructing SKB on remote CPU */
658 xdp_pkt->dev_rx = dev_rx;
659
660 bq_enqueue(rcpu, xdp_pkt);
661 return 0;
662}
663
664void __cpu_map_insert_ctx(struct bpf_map *map, u32 bit)
665{
666 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
667 unsigned long *bitmap = this_cpu_ptr(cmap->flush_needed);
668
669 __set_bit(bit, bitmap);
670}
671
672void __cpu_map_flush(struct bpf_map *map)
673{
674 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
675 unsigned long *bitmap = this_cpu_ptr(cmap->flush_needed);
676 u32 bit;
677
678 /* The napi->poll softirq makes sure __cpu_map_insert_ctx()
679 * and __cpu_map_flush() happen on same CPU. Thus, the percpu
680 * bitmap indicate which percpu bulkq have packets.
681 */
682 for_each_set_bit(bit, bitmap, map->max_entries) {
683 struct bpf_cpu_map_entry *rcpu = READ_ONCE(cmap->cpu_map[bit]);
684 struct xdp_bulk_queue *bq;
685
686 /* This is possible if entry is removed by user space
687 * between xdp redirect and flush op.
688 */
689 if (unlikely(!rcpu))
690 continue;
691
692 __clear_bit(bit, bitmap);
693
694 /* Flush all frames in bulkq to real queue */
695 bq = this_cpu_ptr(rcpu->bulkq);
696 bq_flush_to_queue(rcpu, bq);
697
698 /* If already running, costs spin_lock_irqsave + smb_mb */
699 wake_up_process(rcpu->kthread);
700 }
701}
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}