Commit | Line | Data |
---|---|---|
6710e112 JDB |
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> | |
f9419f7b | 27 | #include <trace/events/xdp.h> |
6710e112 | 28 | |
1c601d82 JDB |
29 | #include <linux/netdevice.h> /* netif_receive_skb_core */ |
30 | #include <linux/etherdevice.h> /* eth_type_trans */ | |
31 | ||
6710e112 JDB |
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 */ | |
40 | struct 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 */ | |
46 | struct bpf_cpu_map_entry { | |
f9419f7b JDB |
47 | u32 cpu; /* kthread CPU and map index */ |
48 | int map_id; /* Back reference to map */ | |
6710e112 JDB |
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 | ||
63 | struct 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 | ||
70 | static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu, | |
71 | struct xdp_bulk_queue *bq); | |
72 | ||
73 | static u64 cpu_map_bitmap_size(const union bpf_attr *attr) | |
74 | { | |
75 | return BITS_TO_LONGS(attr->max_entries) * sizeof(unsigned long); | |
76 | } | |
77 | ||
78 | static 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 | ||
bd475643 | 97 | bpf_map_init_from_attr(&cmap->map, attr); |
6710e112 JDB |
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; | |
133 | free_percpu: | |
134 | free_percpu(cmap->flush_needed); | |
135 | free_cmap: | |
136 | kfree(cmap); | |
137 | return ERR_PTR(err); | |
138 | } | |
139 | ||
0fe875c5 | 140 | static void __cpu_map_queue_destructor(void *ptr) |
6710e112 JDB |
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 | ||
151 | static 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 | ||
161 | static 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 */ | |
167 | static 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 | ||
1c601d82 JDB |
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 | */ | |
186 | struct 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 */ | |
195 | static 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; | |
03c4cc38 | 205 | if (unlikely((headroom - metasize) < sizeof(*xdp_pkt))) |
1c601d82 JDB |
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 | ||
0fe875c5 WY |
219 | static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu, |
220 | struct xdp_pkt *xdp_pkt) | |
1c601d82 JDB |
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 | ||
6710e112 JDB |
268 | static 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)) { | |
f9419f7b | 280 | unsigned int processed = 0, drops = 0, sched = 0; |
6710e112 JDB |
281 | struct xdp_pkt *xdp_pkt; |
282 | ||
1c601d82 JDB |
283 | /* Release CPU reschedule checks */ |
284 | if (__ptr_ring_empty(rcpu->queue)) { | |
31749468 JDB |
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 | } | |
1c601d82 | 293 | } else { |
f9419f7b | 294 | sched = cond_resched(); |
6710e112 | 295 | } |
1c601d82 JDB |
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 | } | |
f9419f7b JDB |
323 | /* Feedback loop via tracepoint */ |
324 | trace_xdp_cpumap_kthread(rcpu->map_id, processed, drops, sched); | |
325 | ||
1c601d82 | 326 | local_bh_enable(); /* resched point, may call do_softirq() */ |
6710e112 JDB |
327 | } |
328 | __set_current_state(TASK_RUNNING); | |
329 | ||
330 | put_cpu_map_entry(rcpu); | |
331 | return 0; | |
332 | } | |
333 | ||
0fe875c5 WY |
334 | static struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu, |
335 | int map_id) | |
6710e112 | 336 | { |
7fc17e90 | 337 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
6710e112 JDB |
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 | ||
f9419f7b JDB |
363 | rcpu->cpu = cpu; |
364 | rcpu->map_id = map_id; | |
365 | rcpu->qsize = qsize; | |
6710e112 JDB |
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 | ||
382 | free_ptr_ring: | |
383 | ptr_ring_cleanup(rcpu->queue, NULL); | |
384 | free_queue: | |
385 | kfree(rcpu->queue); | |
386 | free_bulkq: | |
387 | free_percpu(rcpu->bulkq); | |
388 | free_rcu: | |
389 | kfree(rcpu); | |
390 | return NULL; | |
391 | } | |
392 | ||
0fe875c5 | 393 | static void __cpu_map_entry_free(struct rcu_head *rcu) |
6710e112 JDB |
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 | */ | |
0fe875c5 WY |
436 | static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap, |
437 | u32 key_cpu, struct bpf_cpu_map_entry *rcpu) | |
6710e112 JDB |
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 | ||
0fe875c5 | 449 | static int cpu_map_delete_elem(struct bpf_map *map, void *key) |
6710e112 JDB |
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 | ||
0fe875c5 WY |
462 | static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value, |
463 | u64 map_flags) | |
6710e112 JDB |
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 | ||
0fe875c5 | 500 | static void cpu_map_free(struct bpf_map *map) |
6710e112 JDB |
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 | ||
546 | struct 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 | ||
558 | static 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 | ||
566 | static 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 | ||
583 | const 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 | ||
592 | static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu, | |
593 | struct xdp_bulk_queue *bq) | |
594 | { | |
f9419f7b JDB |
595 | unsigned int processed = 0, drops = 0; |
596 | const int to_cpu = rcpu->cpu; | |
6710e112 JDB |
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) { | |
f9419f7b JDB |
612 | drops++; |
613 | page_frag_free(xdp_pkt); /* Free xdp_pkt */ | |
6710e112 | 614 | } |
f9419f7b | 615 | processed++; |
6710e112 JDB |
616 | } |
617 | bq->count = 0; | |
618 | spin_unlock(&q->producer_lock); | |
619 | ||
f9419f7b JDB |
620 | /* Feedback loop via tracepoints */ |
621 | trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu); | |
6710e112 JDB |
622 | return 0; |
623 | } | |
624 | ||
6710e112 JDB |
625 | /* Runs under RCU-read-side, plus in softirq under NAPI protection. |
626 | * Thus, safe percpu variable access. | |
627 | */ | |
9c270af3 | 628 | static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_pkt *xdp_pkt) |
6710e112 JDB |
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 | ||
9c270af3 JDB |
648 | int 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; | |
9c270af3 | 652 | |
1c601d82 | 653 | xdp_pkt = convert_to_xdp_pkt(xdp); |
03c4cc38 | 654 | if (unlikely(!xdp_pkt)) |
1c601d82 | 655 | return -EOVERFLOW; |
9c270af3 | 656 | |
1c601d82 JDB |
657 | /* Info needed when constructing SKB on remote CPU */ |
658 | xdp_pkt->dev_rx = dev_rx; | |
9c270af3 JDB |
659 | |
660 | bq_enqueue(rcpu, xdp_pkt); | |
661 | return 0; | |
662 | } | |
663 | ||
6710e112 JDB |
664 | void __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 | ||
672 | void __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 | } |