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/**
8 * DOC: cpu map
9 * The 'cpumap' is primarily used as a backend map for XDP BPF helper
10 * call bpf_redirect_map() and XDP_REDIRECT action, like 'devmap'.
11 *
12 * Unlike devmap which redirects XDP frames out to another NIC device,
13 * this map type redirects raw XDP frames to another CPU. The remote
14 * CPU will do SKB-allocation and call the normal network stack.
15 */
16/*
17 * This is a scalability and isolation mechanism, that allow
18 * separating the early driver network XDP layer, from the rest of the
19 * netstack, and assigning dedicated CPUs for this stage. This
20 * basically allows for 10G wirespeed pre-filtering via bpf.
21 */
22#include <linux/bitops.h>
23#include <linux/bpf.h>
24#include <linux/filter.h>
25#include <linux/ptr_ring.h>
26#include <net/xdp.h>
27#include <net/hotdata.h>
28
29#include <linux/sched.h>
30#include <linux/workqueue.h>
31#include <linux/kthread.h>
32#include <linux/completion.h>
33#include <trace/events/xdp.h>
34#include <linux/btf_ids.h>
35
36#include <linux/netdevice.h> /* netif_receive_skb_list */
37#include <linux/etherdevice.h> /* eth_type_trans */
38
39/* General idea: XDP packets getting XDP redirected to another CPU,
40 * will maximum be stored/queued for one driver ->poll() call. It is
41 * guaranteed that queueing the frame and the flush operation happen on
42 * same CPU. Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
43 * which queue in bpf_cpu_map_entry contains packets.
44 */
45
46#define CPU_MAP_BULK_SIZE 8 /* 8 == one cacheline on 64-bit archs */
47struct bpf_cpu_map_entry;
48struct bpf_cpu_map;
49
50struct xdp_bulk_queue {
51 void *q[CPU_MAP_BULK_SIZE];
52 struct list_head flush_node;
53 struct bpf_cpu_map_entry *obj;
54 unsigned int count;
55};
56
57/* Struct for every remote "destination" CPU in map */
58struct bpf_cpu_map_entry {
59 u32 cpu; /* kthread CPU and map index */
60 int map_id; /* Back reference to map */
61
62 /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
63 struct xdp_bulk_queue __percpu *bulkq;
64
65 /* Queue with potential multi-producers, and single-consumer kthread */
66 struct ptr_ring *queue;
67 struct task_struct *kthread;
68
69 struct bpf_cpumap_val value;
70 struct bpf_prog *prog;
71
72 struct completion kthread_running;
73 struct rcu_work free_work;
74};
75
76struct bpf_cpu_map {
77 struct bpf_map map;
78 /* Below members specific for map type */
79 struct bpf_cpu_map_entry __rcu **cpu_map;
80};
81
82static DEFINE_PER_CPU(struct list_head, cpu_map_flush_list);
83
84static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
85{
86 u32 value_size = attr->value_size;
87 struct bpf_cpu_map *cmap;
88
89 /* check sanity of attributes */
90 if (attr->max_entries == 0 || attr->key_size != 4 ||
91 (value_size != offsetofend(struct bpf_cpumap_val, qsize) &&
92 value_size != offsetofend(struct bpf_cpumap_val, bpf_prog.fd)) ||
93 attr->map_flags & ~BPF_F_NUMA_NODE)
94 return ERR_PTR(-EINVAL);
95
96 /* Pre-limit array size based on NR_CPUS, not final CPU check */
97 if (attr->max_entries > NR_CPUS)
98 return ERR_PTR(-E2BIG);
99
100 cmap = bpf_map_area_alloc(sizeof(*cmap), NUMA_NO_NODE);
101 if (!cmap)
102 return ERR_PTR(-ENOMEM);
103
104 bpf_map_init_from_attr(&cmap->map, attr);
105
106 /* Alloc array for possible remote "destination" CPUs */
107 cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
108 sizeof(struct bpf_cpu_map_entry *),
109 cmap->map.numa_node);
110 if (!cmap->cpu_map) {
111 bpf_map_area_free(cmap);
112 return ERR_PTR(-ENOMEM);
113 }
114
115 return &cmap->map;
116}
117
118static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
119{
120 /* The tear-down procedure should have made sure that queue is
121 * empty. See __cpu_map_entry_replace() and work-queue
122 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
123 * gracefully and warn once.
124 */
125 void *ptr;
126
127 while ((ptr = ptr_ring_consume(ring))) {
128 WARN_ON_ONCE(1);
129 if (unlikely(__ptr_test_bit(0, &ptr))) {
130 __ptr_clear_bit(0, &ptr);
131 kfree_skb(ptr);
132 continue;
133 }
134 xdp_return_frame(ptr);
135 }
136}
137
138static void cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
139 struct list_head *listp,
140 struct xdp_cpumap_stats *stats)
141{
142 struct sk_buff *skb, *tmp;
143 struct xdp_buff xdp;
144 u32 act;
145 int err;
146
147 list_for_each_entry_safe(skb, tmp, listp, list) {
148 act = bpf_prog_run_generic_xdp(skb, &xdp, rcpu->prog);
149 switch (act) {
150 case XDP_PASS:
151 break;
152 case XDP_REDIRECT:
153 skb_list_del_init(skb);
154 err = xdp_do_generic_redirect(skb->dev, skb, &xdp,
155 rcpu->prog);
156 if (unlikely(err)) {
157 kfree_skb(skb);
158 stats->drop++;
159 } else {
160 stats->redirect++;
161 }
162 return;
163 default:
164 bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
165 fallthrough;
166 case XDP_ABORTED:
167 trace_xdp_exception(skb->dev, rcpu->prog, act);
168 fallthrough;
169 case XDP_DROP:
170 skb_list_del_init(skb);
171 kfree_skb(skb);
172 stats->drop++;
173 return;
174 }
175 }
176}
177
178static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
179 void **frames, int n,
180 struct xdp_cpumap_stats *stats)
181{
182 struct xdp_rxq_info rxq = {};
183 struct xdp_buff xdp;
184 int i, nframes = 0;
185
186 xdp_set_return_frame_no_direct();
187 xdp.rxq = &rxq;
188
189 for (i = 0; i < n; i++) {
190 struct xdp_frame *xdpf = frames[i];
191 u32 act;
192 int err;
193
194 rxq.dev = xdpf->dev_rx;
195 rxq.mem = xdpf->mem;
196 /* TODO: report queue_index to xdp_rxq_info */
197
198 xdp_convert_frame_to_buff(xdpf, &xdp);
199
200 act = bpf_prog_run_xdp(rcpu->prog, &xdp);
201 switch (act) {
202 case XDP_PASS:
203 err = xdp_update_frame_from_buff(&xdp, xdpf);
204 if (err < 0) {
205 xdp_return_frame(xdpf);
206 stats->drop++;
207 } else {
208 frames[nframes++] = xdpf;
209 stats->pass++;
210 }
211 break;
212 case XDP_REDIRECT:
213 err = xdp_do_redirect(xdpf->dev_rx, &xdp,
214 rcpu->prog);
215 if (unlikely(err)) {
216 xdp_return_frame(xdpf);
217 stats->drop++;
218 } else {
219 stats->redirect++;
220 }
221 break;
222 default:
223 bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
224 fallthrough;
225 case XDP_DROP:
226 xdp_return_frame(xdpf);
227 stats->drop++;
228 break;
229 }
230 }
231
232 xdp_clear_return_frame_no_direct();
233
234 return nframes;
235}
236
237#define CPUMAP_BATCH 8
238
239static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
240 int xdp_n, struct xdp_cpumap_stats *stats,
241 struct list_head *list)
242{
243 int nframes;
244
245 if (!rcpu->prog)
246 return xdp_n;
247
248 rcu_read_lock_bh();
249
250 nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, xdp_n, stats);
251
252 if (stats->redirect)
253 xdp_do_flush();
254
255 if (unlikely(!list_empty(list)))
256 cpu_map_bpf_prog_run_skb(rcpu, list, stats);
257
258 rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
259
260 return nframes;
261}
262
263static int cpu_map_kthread_run(void *data)
264{
265 struct bpf_cpu_map_entry *rcpu = data;
266 unsigned long last_qs = jiffies;
267
268 complete(&rcpu->kthread_running);
269 set_current_state(TASK_INTERRUPTIBLE);
270
271 /* When kthread gives stop order, then rcpu have been disconnected
272 * from map, thus no new packets can enter. Remaining in-flight
273 * per CPU stored packets are flushed to this queue. Wait honoring
274 * kthread_stop signal until queue is empty.
275 */
276 while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
277 struct xdp_cpumap_stats stats = {}; /* zero stats */
278 unsigned int kmem_alloc_drops = 0, sched = 0;
279 gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
280 int i, n, m, nframes, xdp_n;
281 void *frames[CPUMAP_BATCH];
282 void *skbs[CPUMAP_BATCH];
283 LIST_HEAD(list);
284
285 /* Release CPU reschedule checks */
286 if (__ptr_ring_empty(rcpu->queue)) {
287 set_current_state(TASK_INTERRUPTIBLE);
288 /* Recheck to avoid lost wake-up */
289 if (__ptr_ring_empty(rcpu->queue)) {
290 schedule();
291 sched = 1;
292 last_qs = jiffies;
293 } else {
294 __set_current_state(TASK_RUNNING);
295 }
296 } else {
297 rcu_softirq_qs_periodic(last_qs);
298 sched = cond_resched();
299 }
300
301 /*
302 * The bpf_cpu_map_entry is single consumer, with this
303 * kthread CPU pinned. Lockless access to ptr_ring
304 * consume side valid as no-resize allowed of queue.
305 */
306 n = __ptr_ring_consume_batched(rcpu->queue, frames,
307 CPUMAP_BATCH);
308 for (i = 0, xdp_n = 0; i < n; i++) {
309 void *f = frames[i];
310 struct page *page;
311
312 if (unlikely(__ptr_test_bit(0, &f))) {
313 struct sk_buff *skb = f;
314
315 __ptr_clear_bit(0, &skb);
316 list_add_tail(&skb->list, &list);
317 continue;
318 }
319
320 frames[xdp_n++] = f;
321 page = virt_to_page(f);
322
323 /* Bring struct page memory area to curr CPU. Read by
324 * build_skb_around via page_is_pfmemalloc(), and when
325 * freed written by page_frag_free call.
326 */
327 prefetchw(page);
328 }
329
330 /* Support running another XDP prog on this CPU */
331 nframes = cpu_map_bpf_prog_run(rcpu, frames, xdp_n, &stats, &list);
332 if (nframes) {
333 m = kmem_cache_alloc_bulk(net_hotdata.skbuff_cache,
334 gfp, nframes, skbs);
335 if (unlikely(m == 0)) {
336 for (i = 0; i < nframes; i++)
337 skbs[i] = NULL; /* effect: xdp_return_frame */
338 kmem_alloc_drops += nframes;
339 }
340 }
341
342 local_bh_disable();
343 for (i = 0; i < nframes; i++) {
344 struct xdp_frame *xdpf = frames[i];
345 struct sk_buff *skb = skbs[i];
346
347 skb = __xdp_build_skb_from_frame(xdpf, skb,
348 xdpf->dev_rx);
349 if (!skb) {
350 xdp_return_frame(xdpf);
351 continue;
352 }
353
354 list_add_tail(&skb->list, &list);
355 }
356 netif_receive_skb_list(&list);
357
358 /* Feedback loop via tracepoint */
359 trace_xdp_cpumap_kthread(rcpu->map_id, n, kmem_alloc_drops,
360 sched, &stats);
361
362 local_bh_enable(); /* resched point, may call do_softirq() */
363 }
364 __set_current_state(TASK_RUNNING);
365
366 return 0;
367}
368
369static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu,
370 struct bpf_map *map, int fd)
371{
372 struct bpf_prog *prog;
373
374 prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
375 if (IS_ERR(prog))
376 return PTR_ERR(prog);
377
378 if (prog->expected_attach_type != BPF_XDP_CPUMAP ||
379 !bpf_prog_map_compatible(map, prog)) {
380 bpf_prog_put(prog);
381 return -EINVAL;
382 }
383
384 rcpu->value.bpf_prog.id = prog->aux->id;
385 rcpu->prog = prog;
386
387 return 0;
388}
389
390static struct bpf_cpu_map_entry *
391__cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value,
392 u32 cpu)
393{
394 int numa, err, i, fd = value->bpf_prog.fd;
395 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
396 struct bpf_cpu_map_entry *rcpu;
397 struct xdp_bulk_queue *bq;
398
399 /* Have map->numa_node, but choose node of redirect target CPU */
400 numa = cpu_to_node(cpu);
401
402 rcpu = bpf_map_kmalloc_node(map, sizeof(*rcpu), gfp | __GFP_ZERO, numa);
403 if (!rcpu)
404 return NULL;
405
406 /* Alloc percpu bulkq */
407 rcpu->bulkq = bpf_map_alloc_percpu(map, sizeof(*rcpu->bulkq),
408 sizeof(void *), gfp);
409 if (!rcpu->bulkq)
410 goto free_rcu;
411
412 for_each_possible_cpu(i) {
413 bq = per_cpu_ptr(rcpu->bulkq, i);
414 bq->obj = rcpu;
415 }
416
417 /* Alloc queue */
418 rcpu->queue = bpf_map_kmalloc_node(map, sizeof(*rcpu->queue), gfp,
419 numa);
420 if (!rcpu->queue)
421 goto free_bulkq;
422
423 err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
424 if (err)
425 goto free_queue;
426
427 rcpu->cpu = cpu;
428 rcpu->map_id = map->id;
429 rcpu->value.qsize = value->qsize;
430
431 if (fd > 0 && __cpu_map_load_bpf_program(rcpu, map, fd))
432 goto free_ptr_ring;
433
434 /* Setup kthread */
435 init_completion(&rcpu->kthread_running);
436 rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
437 "cpumap/%d/map:%d", cpu,
438 map->id);
439 if (IS_ERR(rcpu->kthread))
440 goto free_prog;
441
442 /* Make sure kthread runs on a single CPU */
443 kthread_bind(rcpu->kthread, cpu);
444 wake_up_process(rcpu->kthread);
445
446 /* Make sure kthread has been running, so kthread_stop() will not
447 * stop the kthread prematurely and all pending frames or skbs
448 * will be handled by the kthread before kthread_stop() returns.
449 */
450 wait_for_completion(&rcpu->kthread_running);
451
452 return rcpu;
453
454free_prog:
455 if (rcpu->prog)
456 bpf_prog_put(rcpu->prog);
457free_ptr_ring:
458 ptr_ring_cleanup(rcpu->queue, NULL);
459free_queue:
460 kfree(rcpu->queue);
461free_bulkq:
462 free_percpu(rcpu->bulkq);
463free_rcu:
464 kfree(rcpu);
465 return NULL;
466}
467
468static void __cpu_map_entry_free(struct work_struct *work)
469{
470 struct bpf_cpu_map_entry *rcpu;
471
472 /* This cpu_map_entry have been disconnected from map and one
473 * RCU grace-period have elapsed. Thus, XDP cannot queue any
474 * new packets and cannot change/set flush_needed that can
475 * find this entry.
476 */
477 rcpu = container_of(to_rcu_work(work), struct bpf_cpu_map_entry, free_work);
478
479 /* kthread_stop will wake_up_process and wait for it to complete.
480 * cpu_map_kthread_run() makes sure the pointer ring is empty
481 * before exiting.
482 */
483 kthread_stop(rcpu->kthread);
484
485 if (rcpu->prog)
486 bpf_prog_put(rcpu->prog);
487 /* The queue should be empty at this point */
488 __cpu_map_ring_cleanup(rcpu->queue);
489 ptr_ring_cleanup(rcpu->queue, NULL);
490 kfree(rcpu->queue);
491 free_percpu(rcpu->bulkq);
492 kfree(rcpu);
493}
494
495/* After the xchg of the bpf_cpu_map_entry pointer, we need to make sure the old
496 * entry is no longer in use before freeing. We use queue_rcu_work() to call
497 * __cpu_map_entry_free() in a separate workqueue after waiting for an RCU grace
498 * period. This means that (a) all pending enqueue and flush operations have
499 * completed (because of the RCU callback), and (b) we are in a workqueue
500 * context where we can stop the kthread and wait for it to exit before freeing
501 * everything.
502 */
503static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
504 u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
505{
506 struct bpf_cpu_map_entry *old_rcpu;
507
508 old_rcpu = unrcu_pointer(xchg(&cmap->cpu_map[key_cpu], RCU_INITIALIZER(rcpu)));
509 if (old_rcpu) {
510 INIT_RCU_WORK(&old_rcpu->free_work, __cpu_map_entry_free);
511 queue_rcu_work(system_wq, &old_rcpu->free_work);
512 }
513}
514
515static long cpu_map_delete_elem(struct bpf_map *map, void *key)
516{
517 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
518 u32 key_cpu = *(u32 *)key;
519
520 if (key_cpu >= map->max_entries)
521 return -EINVAL;
522
523 /* notice caller map_delete_elem() uses rcu_read_lock() */
524 __cpu_map_entry_replace(cmap, key_cpu, NULL);
525 return 0;
526}
527
528static long cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
529 u64 map_flags)
530{
531 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
532 struct bpf_cpumap_val cpumap_value = {};
533 struct bpf_cpu_map_entry *rcpu;
534 /* Array index key correspond to CPU number */
535 u32 key_cpu = *(u32 *)key;
536
537 memcpy(&cpumap_value, value, map->value_size);
538
539 if (unlikely(map_flags > BPF_EXIST))
540 return -EINVAL;
541 if (unlikely(key_cpu >= cmap->map.max_entries))
542 return -E2BIG;
543 if (unlikely(map_flags == BPF_NOEXIST))
544 return -EEXIST;
545 if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
546 return -EOVERFLOW;
547
548 /* Make sure CPU is a valid possible cpu */
549 if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
550 return -ENODEV;
551
552 if (cpumap_value.qsize == 0) {
553 rcpu = NULL; /* Same as deleting */
554 } else {
555 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
556 rcpu = __cpu_map_entry_alloc(map, &cpumap_value, key_cpu);
557 if (!rcpu)
558 return -ENOMEM;
559 }
560 rcu_read_lock();
561 __cpu_map_entry_replace(cmap, key_cpu, rcpu);
562 rcu_read_unlock();
563 return 0;
564}
565
566static void cpu_map_free(struct bpf_map *map)
567{
568 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
569 u32 i;
570
571 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
572 * so the bpf programs (can be more than one that used this map) were
573 * disconnected from events. Wait for outstanding critical sections in
574 * these programs to complete. synchronize_rcu() below not only
575 * guarantees no further "XDP/bpf-side" reads against
576 * bpf_cpu_map->cpu_map, but also ensure pending flush operations
577 * (if any) are completed.
578 */
579 synchronize_rcu();
580
581 /* The only possible user of bpf_cpu_map_entry is
582 * cpu_map_kthread_run().
583 */
584 for (i = 0; i < cmap->map.max_entries; i++) {
585 struct bpf_cpu_map_entry *rcpu;
586
587 rcpu = rcu_dereference_raw(cmap->cpu_map[i]);
588 if (!rcpu)
589 continue;
590
591 /* Stop kthread and cleanup entry directly */
592 __cpu_map_entry_free(&rcpu->free_work.work);
593 }
594 bpf_map_area_free(cmap->cpu_map);
595 bpf_map_area_free(cmap);
596}
597
598/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
599 * by local_bh_disable() (from XDP calls inside NAPI). The
600 * rcu_read_lock_bh_held() below makes lockdep accept both.
601 */
602static void *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
603{
604 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
605 struct bpf_cpu_map_entry *rcpu;
606
607 if (key >= map->max_entries)
608 return NULL;
609
610 rcpu = rcu_dereference_check(cmap->cpu_map[key],
611 rcu_read_lock_bh_held());
612 return rcpu;
613}
614
615static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
616{
617 struct bpf_cpu_map_entry *rcpu =
618 __cpu_map_lookup_elem(map, *(u32 *)key);
619
620 return rcpu ? &rcpu->value : NULL;
621}
622
623static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
624{
625 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
626 u32 index = key ? *(u32 *)key : U32_MAX;
627 u32 *next = next_key;
628
629 if (index >= cmap->map.max_entries) {
630 *next = 0;
631 return 0;
632 }
633
634 if (index == cmap->map.max_entries - 1)
635 return -ENOENT;
636 *next = index + 1;
637 return 0;
638}
639
640static long cpu_map_redirect(struct bpf_map *map, u64 index, u64 flags)
641{
642 return __bpf_xdp_redirect_map(map, index, flags, 0,
643 __cpu_map_lookup_elem);
644}
645
646static u64 cpu_map_mem_usage(const struct bpf_map *map)
647{
648 u64 usage = sizeof(struct bpf_cpu_map);
649
650 /* Currently the dynamically allocated elements are not counted */
651 usage += (u64)map->max_entries * sizeof(struct bpf_cpu_map_entry *);
652 return usage;
653}
654
655BTF_ID_LIST_SINGLE(cpu_map_btf_ids, struct, bpf_cpu_map)
656const struct bpf_map_ops cpu_map_ops = {
657 .map_meta_equal = bpf_map_meta_equal,
658 .map_alloc = cpu_map_alloc,
659 .map_free = cpu_map_free,
660 .map_delete_elem = cpu_map_delete_elem,
661 .map_update_elem = cpu_map_update_elem,
662 .map_lookup_elem = cpu_map_lookup_elem,
663 .map_get_next_key = cpu_map_get_next_key,
664 .map_check_btf = map_check_no_btf,
665 .map_mem_usage = cpu_map_mem_usage,
666 .map_btf_id = &cpu_map_btf_ids[0],
667 .map_redirect = cpu_map_redirect,
668};
669
670static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
671{
672 struct bpf_cpu_map_entry *rcpu = bq->obj;
673 unsigned int processed = 0, drops = 0;
674 const int to_cpu = rcpu->cpu;
675 struct ptr_ring *q;
676 int i;
677
678 if (unlikely(!bq->count))
679 return;
680
681 q = rcpu->queue;
682 spin_lock(&q->producer_lock);
683
684 for (i = 0; i < bq->count; i++) {
685 struct xdp_frame *xdpf = bq->q[i];
686 int err;
687
688 err = __ptr_ring_produce(q, xdpf);
689 if (err) {
690 drops++;
691 xdp_return_frame_rx_napi(xdpf);
692 }
693 processed++;
694 }
695 bq->count = 0;
696 spin_unlock(&q->producer_lock);
697
698 __list_del_clearprev(&bq->flush_node);
699
700 /* Feedback loop via tracepoints */
701 trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
702}
703
704/* Runs under RCU-read-side, plus in softirq under NAPI protection.
705 * Thus, safe percpu variable access.
706 */
707static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
708{
709 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
710 struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
711
712 if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
713 bq_flush_to_queue(bq);
714
715 /* Notice, xdp_buff/page MUST be queued here, long enough for
716 * driver to code invoking us to finished, due to driver
717 * (e.g. ixgbe) recycle tricks based on page-refcnt.
718 *
719 * Thus, incoming xdp_frame is always queued here (else we race
720 * with another CPU on page-refcnt and remaining driver code).
721 * Queue time is very short, as driver will invoke flush
722 * operation, when completing napi->poll call.
723 */
724 bq->q[bq->count++] = xdpf;
725
726 if (!bq->flush_node.prev)
727 list_add(&bq->flush_node, flush_list);
728}
729
730int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf,
731 struct net_device *dev_rx)
732{
733 /* Info needed when constructing SKB on remote CPU */
734 xdpf->dev_rx = dev_rx;
735
736 bq_enqueue(rcpu, xdpf);
737 return 0;
738}
739
740int cpu_map_generic_redirect(struct bpf_cpu_map_entry *rcpu,
741 struct sk_buff *skb)
742{
743 int ret;
744
745 __skb_pull(skb, skb->mac_len);
746 skb_set_redirected(skb, false);
747 __ptr_set_bit(0, &skb);
748
749 ret = ptr_ring_produce(rcpu->queue, skb);
750 if (ret < 0)
751 goto trace;
752
753 wake_up_process(rcpu->kthread);
754trace:
755 trace_xdp_cpumap_enqueue(rcpu->map_id, !ret, !!ret, rcpu->cpu);
756 return ret;
757}
758
759void __cpu_map_flush(void)
760{
761 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
762 struct xdp_bulk_queue *bq, *tmp;
763
764 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
765 bq_flush_to_queue(bq);
766
767 /* If already running, costs spin_lock_irqsave + smb_mb */
768 wake_up_process(bq->obj->kthread);
769 }
770}
771
772#ifdef CONFIG_DEBUG_NET
773bool cpu_map_check_flush(void)
774{
775 if (list_empty(this_cpu_ptr(&cpu_map_flush_list)))
776 return false;
777 __cpu_map_flush();
778 return true;
779}
780#endif
781
782static int __init cpu_map_init(void)
783{
784 int cpu;
785
786 for_each_possible_cpu(cpu)
787 INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu));
788 return 0;
789}
790
791subsys_initcall(cpu_map_init);