]>
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 | ||
97 | /* mandatory map attributes */ | |
98 | cmap->map.map_type = attr->map_type; | |
99 | cmap->map.key_size = attr->key_size; | |
100 | cmap->map.value_size = attr->value_size; | |
101 | cmap->map.max_entries = attr->max_entries; | |
102 | cmap->map.map_flags = attr->map_flags; | |
103 | cmap->map.numa_node = bpf_map_attr_numa_node(attr); | |
104 | ||
105 | /* Pre-limit array size based on NR_CPUS, not final CPU check */ | |
106 | if (cmap->map.max_entries > NR_CPUS) { | |
107 | err = -E2BIG; | |
108 | goto free_cmap; | |
109 | } | |
110 | ||
111 | /* make sure page count doesn't overflow */ | |
112 | cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *); | |
113 | cost += cpu_map_bitmap_size(attr) * num_possible_cpus(); | |
114 | if (cost >= U32_MAX - PAGE_SIZE) | |
115 | goto free_cmap; | |
116 | cmap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT; | |
117 | ||
118 | /* Notice returns -EPERM on if map size is larger than memlock limit */ | |
119 | ret = bpf_map_precharge_memlock(cmap->map.pages); | |
120 | if (ret) { | |
121 | err = ret; | |
122 | goto free_cmap; | |
123 | } | |
124 | ||
125 | /* A per cpu bitfield with a bit per possible CPU in map */ | |
126 | cmap->flush_needed = __alloc_percpu(cpu_map_bitmap_size(attr), | |
127 | __alignof__(unsigned long)); | |
128 | if (!cmap->flush_needed) | |
129 | goto free_cmap; | |
130 | ||
131 | /* Alloc array for possible remote "destination" CPUs */ | |
132 | cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries * | |
133 | sizeof(struct bpf_cpu_map_entry *), | |
134 | cmap->map.numa_node); | |
135 | if (!cmap->cpu_map) | |
136 | goto free_percpu; | |
137 | ||
138 | return &cmap->map; | |
139 | free_percpu: | |
140 | free_percpu(cmap->flush_needed); | |
141 | free_cmap: | |
142 | kfree(cmap); | |
143 | return ERR_PTR(err); | |
144 | } | |
145 | ||
146 | void __cpu_map_queue_destructor(void *ptr) | |
147 | { | |
148 | /* The tear-down procedure should have made sure that queue is | |
149 | * empty. See __cpu_map_entry_replace() and work-queue | |
150 | * invoked cpu_map_kthread_stop(). Catch any broken behaviour | |
151 | * gracefully and warn once. | |
152 | */ | |
153 | if (WARN_ON_ONCE(ptr)) | |
154 | page_frag_free(ptr); | |
155 | } | |
156 | ||
157 | static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu) | |
158 | { | |
159 | if (atomic_dec_and_test(&rcpu->refcnt)) { | |
160 | /* The queue should be empty at this point */ | |
161 | ptr_ring_cleanup(rcpu->queue, __cpu_map_queue_destructor); | |
162 | kfree(rcpu->queue); | |
163 | kfree(rcpu); | |
164 | } | |
165 | } | |
166 | ||
167 | static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu) | |
168 | { | |
169 | atomic_inc(&rcpu->refcnt); | |
170 | } | |
171 | ||
172 | /* called from workqueue, to workaround syscall using preempt_disable */ | |
173 | static void cpu_map_kthread_stop(struct work_struct *work) | |
174 | { | |
175 | struct bpf_cpu_map_entry *rcpu; | |
176 | ||
177 | rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq); | |
178 | ||
179 | /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier, | |
180 | * as it waits until all in-flight call_rcu() callbacks complete. | |
181 | */ | |
182 | rcu_barrier(); | |
183 | ||
184 | /* kthread_stop will wake_up_process and wait for it to complete */ | |
185 | kthread_stop(rcpu->kthread); | |
186 | } | |
187 | ||
1c601d82 JDB |
188 | /* For now, xdp_pkt is a cpumap internal data structure, with info |
189 | * carried between enqueue to dequeue. It is mapped into the top | |
190 | * headroom of the packet, to avoid allocating separate mem. | |
191 | */ | |
192 | struct xdp_pkt { | |
193 | void *data; | |
194 | u16 len; | |
195 | u16 headroom; | |
196 | u16 metasize; | |
197 | struct net_device *dev_rx; | |
198 | }; | |
199 | ||
200 | /* Convert xdp_buff to xdp_pkt */ | |
201 | static struct xdp_pkt *convert_to_xdp_pkt(struct xdp_buff *xdp) | |
202 | { | |
203 | struct xdp_pkt *xdp_pkt; | |
204 | int metasize; | |
205 | int headroom; | |
206 | ||
207 | /* Assure headroom is available for storing info */ | |
208 | headroom = xdp->data - xdp->data_hard_start; | |
209 | metasize = xdp->data - xdp->data_meta; | |
210 | metasize = metasize > 0 ? metasize : 0; | |
03c4cc38 | 211 | if (unlikely((headroom - metasize) < sizeof(*xdp_pkt))) |
1c601d82 JDB |
212 | return NULL; |
213 | ||
214 | /* Store info in top of packet */ | |
215 | xdp_pkt = xdp->data_hard_start; | |
216 | ||
217 | xdp_pkt->data = xdp->data; | |
218 | xdp_pkt->len = xdp->data_end - xdp->data; | |
219 | xdp_pkt->headroom = headroom - sizeof(*xdp_pkt); | |
220 | xdp_pkt->metasize = metasize; | |
221 | ||
222 | return xdp_pkt; | |
223 | } | |
224 | ||
225 | struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu, | |
226 | struct xdp_pkt *xdp_pkt) | |
227 | { | |
228 | unsigned int frame_size; | |
229 | void *pkt_data_start; | |
230 | struct sk_buff *skb; | |
231 | ||
232 | /* build_skb need to place skb_shared_info after SKB end, and | |
233 | * also want to know the memory "truesize". Thus, need to | |
234 | * know the memory frame size backing xdp_buff. | |
235 | * | |
236 | * XDP was designed to have PAGE_SIZE frames, but this | |
237 | * assumption is not longer true with ixgbe and i40e. It | |
238 | * would be preferred to set frame_size to 2048 or 4096 | |
239 | * depending on the driver. | |
240 | * frame_size = 2048; | |
241 | * frame_len = frame_size - sizeof(*xdp_pkt); | |
242 | * | |
243 | * Instead, with info avail, skb_shared_info in placed after | |
244 | * packet len. This, unfortunately fakes the truesize. | |
245 | * Another disadvantage of this approach, the skb_shared_info | |
246 | * is not at a fixed memory location, with mixed length | |
247 | * packets, which is bad for cache-line hotness. | |
248 | */ | |
249 | frame_size = SKB_DATA_ALIGN(xdp_pkt->len) + xdp_pkt->headroom + | |
250 | SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); | |
251 | ||
252 | pkt_data_start = xdp_pkt->data - xdp_pkt->headroom; | |
253 | skb = build_skb(pkt_data_start, frame_size); | |
254 | if (!skb) | |
255 | return NULL; | |
256 | ||
257 | skb_reserve(skb, xdp_pkt->headroom); | |
258 | __skb_put(skb, xdp_pkt->len); | |
259 | if (xdp_pkt->metasize) | |
260 | skb_metadata_set(skb, xdp_pkt->metasize); | |
261 | ||
262 | /* Essential SKB info: protocol and skb->dev */ | |
263 | skb->protocol = eth_type_trans(skb, xdp_pkt->dev_rx); | |
264 | ||
265 | /* Optional SKB info, currently missing: | |
266 | * - HW checksum info (skb->ip_summed) | |
267 | * - HW RX hash (skb_set_hash) | |
268 | * - RX ring dev queue index (skb_record_rx_queue) | |
269 | */ | |
270 | ||
271 | return skb; | |
272 | } | |
273 | ||
6710e112 JDB |
274 | static int cpu_map_kthread_run(void *data) |
275 | { | |
276 | struct bpf_cpu_map_entry *rcpu = data; | |
277 | ||
278 | set_current_state(TASK_INTERRUPTIBLE); | |
279 | ||
280 | /* When kthread gives stop order, then rcpu have been disconnected | |
281 | * from map, thus no new packets can enter. Remaining in-flight | |
282 | * per CPU stored packets are flushed to this queue. Wait honoring | |
283 | * kthread_stop signal until queue is empty. | |
284 | */ | |
285 | while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) { | |
f9419f7b | 286 | unsigned int processed = 0, drops = 0, sched = 0; |
6710e112 JDB |
287 | struct xdp_pkt *xdp_pkt; |
288 | ||
1c601d82 JDB |
289 | /* Release CPU reschedule checks */ |
290 | if (__ptr_ring_empty(rcpu->queue)) { | |
31749468 JDB |
291 | set_current_state(TASK_INTERRUPTIBLE); |
292 | /* Recheck to avoid lost wake-up */ | |
293 | if (__ptr_ring_empty(rcpu->queue)) { | |
294 | schedule(); | |
295 | sched = 1; | |
296 | } else { | |
297 | __set_current_state(TASK_RUNNING); | |
298 | } | |
1c601d82 | 299 | } else { |
f9419f7b | 300 | sched = cond_resched(); |
6710e112 | 301 | } |
1c601d82 JDB |
302 | |
303 | /* Process packets in rcpu->queue */ | |
304 | local_bh_disable(); | |
305 | /* | |
306 | * The bpf_cpu_map_entry is single consumer, with this | |
307 | * kthread CPU pinned. Lockless access to ptr_ring | |
308 | * consume side valid as no-resize allowed of queue. | |
309 | */ | |
310 | while ((xdp_pkt = __ptr_ring_consume(rcpu->queue))) { | |
311 | struct sk_buff *skb; | |
312 | int ret; | |
313 | ||
314 | skb = cpu_map_build_skb(rcpu, xdp_pkt); | |
315 | if (!skb) { | |
316 | page_frag_free(xdp_pkt); | |
317 | continue; | |
318 | } | |
319 | ||
320 | /* Inject into network stack */ | |
321 | ret = netif_receive_skb_core(skb); | |
322 | if (ret == NET_RX_DROP) | |
323 | drops++; | |
324 | ||
325 | /* Limit BH-disable period */ | |
326 | if (++processed == 8) | |
327 | break; | |
328 | } | |
f9419f7b JDB |
329 | /* Feedback loop via tracepoint */ |
330 | trace_xdp_cpumap_kthread(rcpu->map_id, processed, drops, sched); | |
331 | ||
1c601d82 | 332 | local_bh_enable(); /* resched point, may call do_softirq() */ |
6710e112 JDB |
333 | } |
334 | __set_current_state(TASK_RUNNING); | |
335 | ||
336 | put_cpu_map_entry(rcpu); | |
337 | return 0; | |
338 | } | |
339 | ||
340 | struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu, int map_id) | |
341 | { | |
1d3c3c7f | 342 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
6710e112 JDB |
343 | struct bpf_cpu_map_entry *rcpu; |
344 | int numa, err; | |
345 | ||
346 | /* Have map->numa_node, but choose node of redirect target CPU */ | |
347 | numa = cpu_to_node(cpu); | |
348 | ||
349 | rcpu = kzalloc_node(sizeof(*rcpu), gfp, numa); | |
350 | if (!rcpu) | |
351 | return NULL; | |
352 | ||
353 | /* Alloc percpu bulkq */ | |
354 | rcpu->bulkq = __alloc_percpu_gfp(sizeof(*rcpu->bulkq), | |
355 | sizeof(void *), gfp); | |
356 | if (!rcpu->bulkq) | |
357 | goto free_rcu; | |
358 | ||
359 | /* Alloc queue */ | |
360 | rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa); | |
361 | if (!rcpu->queue) | |
362 | goto free_bulkq; | |
363 | ||
364 | err = ptr_ring_init(rcpu->queue, qsize, gfp); | |
365 | if (err) | |
366 | goto free_queue; | |
367 | ||
f9419f7b JDB |
368 | rcpu->cpu = cpu; |
369 | rcpu->map_id = map_id; | |
370 | rcpu->qsize = qsize; | |
6710e112 JDB |
371 | |
372 | /* Setup kthread */ | |
373 | rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa, | |
374 | "cpumap/%d/map:%d", cpu, map_id); | |
375 | if (IS_ERR(rcpu->kthread)) | |
376 | goto free_ptr_ring; | |
377 | ||
378 | get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */ | |
379 | get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */ | |
380 | ||
381 | /* Make sure kthread runs on a single CPU */ | |
382 | kthread_bind(rcpu->kthread, cpu); | |
383 | wake_up_process(rcpu->kthread); | |
384 | ||
385 | return rcpu; | |
386 | ||
387 | free_ptr_ring: | |
388 | ptr_ring_cleanup(rcpu->queue, NULL); | |
389 | free_queue: | |
390 | kfree(rcpu->queue); | |
391 | free_bulkq: | |
392 | free_percpu(rcpu->bulkq); | |
393 | free_rcu: | |
394 | kfree(rcpu); | |
395 | return NULL; | |
396 | } | |
397 | ||
398 | void __cpu_map_entry_free(struct rcu_head *rcu) | |
399 | { | |
400 | struct bpf_cpu_map_entry *rcpu; | |
401 | int cpu; | |
402 | ||
403 | /* This cpu_map_entry have been disconnected from map and one | |
404 | * RCU graze-period have elapsed. Thus, XDP cannot queue any | |
405 | * new packets and cannot change/set flush_needed that can | |
406 | * find this entry. | |
407 | */ | |
408 | rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu); | |
409 | ||
410 | /* Flush remaining packets in percpu bulkq */ | |
411 | for_each_online_cpu(cpu) { | |
412 | struct xdp_bulk_queue *bq = per_cpu_ptr(rcpu->bulkq, cpu); | |
413 | ||
414 | /* No concurrent bq_enqueue can run at this point */ | |
415 | bq_flush_to_queue(rcpu, bq); | |
416 | } | |
417 | free_percpu(rcpu->bulkq); | |
418 | /* Cannot kthread_stop() here, last put free rcpu resources */ | |
419 | put_cpu_map_entry(rcpu); | |
420 | } | |
421 | ||
422 | /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to | |
423 | * ensure any driver rcu critical sections have completed, but this | |
424 | * does not guarantee a flush has happened yet. Because driver side | |
425 | * rcu_read_lock/unlock only protects the running XDP program. The | |
426 | * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a | |
427 | * pending flush op doesn't fail. | |
428 | * | |
429 | * The bpf_cpu_map_entry is still used by the kthread, and there can | |
430 | * still be pending packets (in queue and percpu bulkq). A refcnt | |
431 | * makes sure to last user (kthread_stop vs. call_rcu) free memory | |
432 | * resources. | |
433 | * | |
434 | * The rcu callback __cpu_map_entry_free flush remaining packets in | |
435 | * percpu bulkq to queue. Due to caller map_delete_elem() disable | |
436 | * preemption, cannot call kthread_stop() to make sure queue is empty. | |
437 | * Instead a work_queue is started for stopping kthread, | |
438 | * cpu_map_kthread_stop, which waits for an RCU graze period before | |
439 | * stopping kthread, emptying the queue. | |
440 | */ | |
441 | void __cpu_map_entry_replace(struct bpf_cpu_map *cmap, | |
442 | u32 key_cpu, struct bpf_cpu_map_entry *rcpu) | |
443 | { | |
444 | struct bpf_cpu_map_entry *old_rcpu; | |
445 | ||
446 | old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu); | |
447 | if (old_rcpu) { | |
448 | call_rcu(&old_rcpu->rcu, __cpu_map_entry_free); | |
449 | INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop); | |
450 | schedule_work(&old_rcpu->kthread_stop_wq); | |
451 | } | |
452 | } | |
453 | ||
454 | int cpu_map_delete_elem(struct bpf_map *map, void *key) | |
455 | { | |
456 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); | |
457 | u32 key_cpu = *(u32 *)key; | |
458 | ||
459 | if (key_cpu >= map->max_entries) | |
460 | return -EINVAL; | |
461 | ||
462 | /* notice caller map_delete_elem() use preempt_disable() */ | |
463 | __cpu_map_entry_replace(cmap, key_cpu, NULL); | |
464 | return 0; | |
465 | } | |
466 | ||
467 | int cpu_map_update_elem(struct bpf_map *map, void *key, void *value, | |
468 | u64 map_flags) | |
469 | { | |
470 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); | |
471 | struct bpf_cpu_map_entry *rcpu; | |
472 | ||
473 | /* Array index key correspond to CPU number */ | |
474 | u32 key_cpu = *(u32 *)key; | |
475 | /* Value is the queue size */ | |
476 | u32 qsize = *(u32 *)value; | |
477 | ||
478 | if (unlikely(map_flags > BPF_EXIST)) | |
479 | return -EINVAL; | |
480 | if (unlikely(key_cpu >= cmap->map.max_entries)) | |
481 | return -E2BIG; | |
482 | if (unlikely(map_flags == BPF_NOEXIST)) | |
483 | return -EEXIST; | |
484 | if (unlikely(qsize > 16384)) /* sanity limit on qsize */ | |
485 | return -EOVERFLOW; | |
486 | ||
487 | /* Make sure CPU is a valid possible cpu */ | |
488 | if (!cpu_possible(key_cpu)) | |
489 | return -ENODEV; | |
490 | ||
491 | if (qsize == 0) { | |
492 | rcpu = NULL; /* Same as deleting */ | |
493 | } else { | |
494 | /* Updating qsize cause re-allocation of bpf_cpu_map_entry */ | |
495 | rcpu = __cpu_map_entry_alloc(qsize, key_cpu, map->id); | |
496 | if (!rcpu) | |
497 | return -ENOMEM; | |
498 | } | |
499 | rcu_read_lock(); | |
500 | __cpu_map_entry_replace(cmap, key_cpu, rcpu); | |
501 | rcu_read_unlock(); | |
502 | return 0; | |
503 | } | |
504 | ||
505 | void cpu_map_free(struct bpf_map *map) | |
506 | { | |
507 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); | |
508 | int cpu; | |
509 | u32 i; | |
510 | ||
511 | /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0, | |
512 | * so the bpf programs (can be more than one that used this map) were | |
513 | * disconnected from events. Wait for outstanding critical sections in | |
514 | * these programs to complete. The rcu critical section only guarantees | |
515 | * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map. | |
516 | * It does __not__ ensure pending flush operations (if any) are | |
517 | * complete. | |
518 | */ | |
519 | synchronize_rcu(); | |
520 | ||
521 | /* To ensure all pending flush operations have completed wait for flush | |
522 | * bitmap to indicate all flush_needed bits to be zero on _all_ cpus. | |
523 | * Because the above synchronize_rcu() ensures the map is disconnected | |
524 | * from the program we can assume no new bits will be set. | |
525 | */ | |
526 | for_each_online_cpu(cpu) { | |
527 | unsigned long *bitmap = per_cpu_ptr(cmap->flush_needed, cpu); | |
528 | ||
529 | while (!bitmap_empty(bitmap, cmap->map.max_entries)) | |
530 | cond_resched(); | |
531 | } | |
532 | ||
533 | /* For cpu_map the remote CPUs can still be using the entries | |
534 | * (struct bpf_cpu_map_entry). | |
535 | */ | |
536 | for (i = 0; i < cmap->map.max_entries; i++) { | |
537 | struct bpf_cpu_map_entry *rcpu; | |
538 | ||
539 | rcpu = READ_ONCE(cmap->cpu_map[i]); | |
540 | if (!rcpu) | |
541 | continue; | |
542 | ||
543 | /* bq flush and cleanup happens after RCU graze-period */ | |
544 | __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */ | |
545 | } | |
546 | free_percpu(cmap->flush_needed); | |
547 | bpf_map_area_free(cmap->cpu_map); | |
548 | kfree(cmap); | |
549 | } | |
550 | ||
551 | struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key) | |
552 | { | |
553 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); | |
554 | struct bpf_cpu_map_entry *rcpu; | |
555 | ||
556 | if (key >= map->max_entries) | |
557 | return NULL; | |
558 | ||
559 | rcpu = READ_ONCE(cmap->cpu_map[key]); | |
560 | return rcpu; | |
561 | } | |
562 | ||
563 | static void *cpu_map_lookup_elem(struct bpf_map *map, void *key) | |
564 | { | |
565 | struct bpf_cpu_map_entry *rcpu = | |
566 | __cpu_map_lookup_elem(map, *(u32 *)key); | |
567 | ||
568 | return rcpu ? &rcpu->qsize : NULL; | |
569 | } | |
570 | ||
571 | static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key) | |
572 | { | |
573 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); | |
574 | u32 index = key ? *(u32 *)key : U32_MAX; | |
575 | u32 *next = next_key; | |
576 | ||
577 | if (index >= cmap->map.max_entries) { | |
578 | *next = 0; | |
579 | return 0; | |
580 | } | |
581 | ||
582 | if (index == cmap->map.max_entries - 1) | |
583 | return -ENOENT; | |
584 | *next = index + 1; | |
585 | return 0; | |
586 | } | |
587 | ||
588 | const struct bpf_map_ops cpu_map_ops = { | |
589 | .map_alloc = cpu_map_alloc, | |
590 | .map_free = cpu_map_free, | |
591 | .map_delete_elem = cpu_map_delete_elem, | |
592 | .map_update_elem = cpu_map_update_elem, | |
593 | .map_lookup_elem = cpu_map_lookup_elem, | |
594 | .map_get_next_key = cpu_map_get_next_key, | |
595 | }; | |
596 | ||
597 | static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu, | |
598 | struct xdp_bulk_queue *bq) | |
599 | { | |
f9419f7b JDB |
600 | unsigned int processed = 0, drops = 0; |
601 | const int to_cpu = rcpu->cpu; | |
6710e112 JDB |
602 | struct ptr_ring *q; |
603 | int i; | |
604 | ||
605 | if (unlikely(!bq->count)) | |
606 | return 0; | |
607 | ||
608 | q = rcpu->queue; | |
609 | spin_lock(&q->producer_lock); | |
610 | ||
611 | for (i = 0; i < bq->count; i++) { | |
612 | void *xdp_pkt = bq->q[i]; | |
613 | int err; | |
614 | ||
615 | err = __ptr_ring_produce(q, xdp_pkt); | |
616 | if (err) { | |
f9419f7b JDB |
617 | drops++; |
618 | page_frag_free(xdp_pkt); /* Free xdp_pkt */ | |
6710e112 | 619 | } |
f9419f7b | 620 | processed++; |
6710e112 JDB |
621 | } |
622 | bq->count = 0; | |
623 | spin_unlock(&q->producer_lock); | |
624 | ||
f9419f7b JDB |
625 | /* Feedback loop via tracepoints */ |
626 | trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu); | |
6710e112 JDB |
627 | return 0; |
628 | } | |
629 | ||
6710e112 JDB |
630 | /* Runs under RCU-read-side, plus in softirq under NAPI protection. |
631 | * Thus, safe percpu variable access. | |
632 | */ | |
9c270af3 | 633 | static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_pkt *xdp_pkt) |
6710e112 JDB |
634 | { |
635 | struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq); | |
636 | ||
637 | if (unlikely(bq->count == CPU_MAP_BULK_SIZE)) | |
638 | bq_flush_to_queue(rcpu, bq); | |
639 | ||
640 | /* Notice, xdp_buff/page MUST be queued here, long enough for | |
641 | * driver to code invoking us to finished, due to driver | |
642 | * (e.g. ixgbe) recycle tricks based on page-refcnt. | |
643 | * | |
644 | * Thus, incoming xdp_pkt is always queued here (else we race | |
645 | * with another CPU on page-refcnt and remaining driver code). | |
646 | * Queue time is very short, as driver will invoke flush | |
647 | * operation, when completing napi->poll call. | |
648 | */ | |
649 | bq->q[bq->count++] = xdp_pkt; | |
650 | return 0; | |
651 | } | |
652 | ||
9c270af3 JDB |
653 | int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp, |
654 | struct net_device *dev_rx) | |
655 | { | |
656 | struct xdp_pkt *xdp_pkt; | |
9c270af3 | 657 | |
1c601d82 | 658 | xdp_pkt = convert_to_xdp_pkt(xdp); |
03c4cc38 | 659 | if (unlikely(!xdp_pkt)) |
1c601d82 | 660 | return -EOVERFLOW; |
9c270af3 | 661 | |
1c601d82 JDB |
662 | /* Info needed when constructing SKB on remote CPU */ |
663 | xdp_pkt->dev_rx = dev_rx; | |
9c270af3 JDB |
664 | |
665 | bq_enqueue(rcpu, xdp_pkt); | |
666 | return 0; | |
667 | } | |
668 | ||
6710e112 JDB |
669 | void __cpu_map_insert_ctx(struct bpf_map *map, u32 bit) |
670 | { | |
671 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); | |
672 | unsigned long *bitmap = this_cpu_ptr(cmap->flush_needed); | |
673 | ||
674 | __set_bit(bit, bitmap); | |
675 | } | |
676 | ||
677 | void __cpu_map_flush(struct bpf_map *map) | |
678 | { | |
679 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); | |
680 | unsigned long *bitmap = this_cpu_ptr(cmap->flush_needed); | |
681 | u32 bit; | |
682 | ||
683 | /* The napi->poll softirq makes sure __cpu_map_insert_ctx() | |
684 | * and __cpu_map_flush() happen on same CPU. Thus, the percpu | |
685 | * bitmap indicate which percpu bulkq have packets. | |
686 | */ | |
687 | for_each_set_bit(bit, bitmap, map->max_entries) { | |
688 | struct bpf_cpu_map_entry *rcpu = READ_ONCE(cmap->cpu_map[bit]); | |
689 | struct xdp_bulk_queue *bq; | |
690 | ||
691 | /* This is possible if entry is removed by user space | |
692 | * between xdp redirect and flush op. | |
693 | */ | |
694 | if (unlikely(!rcpu)) | |
695 | continue; | |
696 | ||
697 | __clear_bit(bit, bitmap); | |
698 | ||
699 | /* Flush all frames in bulkq to real queue */ | |
700 | bq = this_cpu_ptr(rcpu->bulkq); | |
701 | bq_flush_to_queue(rcpu, bq); | |
702 | ||
703 | /* If already running, costs spin_lock_irqsave + smb_mb */ | |
704 | wake_up_process(rcpu->kthread); | |
705 | } | |
706 | } |