Commit | Line | Data |
---|---|---|
ddc64d0a | 1 | // SPDX-License-Identifier: GPL-2.0-only |
6710e112 JDB |
2 | /* bpf/cpumap.c |
3 | * | |
4 | * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc. | |
6710e112 JDB |
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> | |
5ab073ff | 22 | #include <net/xdp.h> |
6710e112 JDB |
23 | |
24 | #include <linux/sched.h> | |
25 | #include <linux/workqueue.h> | |
26 | #include <linux/kthread.h> | |
27 | #include <linux/capability.h> | |
f9419f7b | 28 | #include <trace/events/xdp.h> |
6710e112 | 29 | |
1c601d82 JDB |
30 | #include <linux/netdevice.h> /* netif_receive_skb_core */ |
31 | #include <linux/etherdevice.h> /* eth_type_trans */ | |
32 | ||
6710e112 JDB |
33 | /* General idea: XDP packets getting XDP redirected to another CPU, |
34 | * will maximum be stored/queued for one driver ->poll() call. It is | |
d5df2830 | 35 | * guaranteed that queueing the frame and the flush operation happen on |
6710e112 JDB |
36 | * same CPU. Thus, cpu_map_flush operation can deduct via this_cpu_ptr() |
37 | * which queue in bpf_cpu_map_entry contains packets. | |
38 | */ | |
39 | ||
40 | #define CPU_MAP_BULK_SIZE 8 /* 8 == one cacheline on 64-bit archs */ | |
d5df2830 THJ |
41 | struct bpf_cpu_map_entry; |
42 | struct bpf_cpu_map; | |
43 | ||
6710e112 JDB |
44 | struct xdp_bulk_queue { |
45 | void *q[CPU_MAP_BULK_SIZE]; | |
d5df2830 THJ |
46 | struct list_head flush_node; |
47 | struct bpf_cpu_map_entry *obj; | |
6710e112 JDB |
48 | unsigned int count; |
49 | }; | |
50 | ||
51 | /* Struct for every remote "destination" CPU in map */ | |
52 | struct bpf_cpu_map_entry { | |
f9419f7b JDB |
53 | u32 cpu; /* kthread CPU and map index */ |
54 | int map_id; /* Back reference to map */ | |
6710e112 JDB |
55 | |
56 | /* XDP can run multiple RX-ring queues, need __percpu enqueue store */ | |
57 | struct xdp_bulk_queue __percpu *bulkq; | |
58 | ||
d5df2830 THJ |
59 | struct bpf_cpu_map *cmap; |
60 | ||
6710e112 JDB |
61 | /* Queue with potential multi-producers, and single-consumer kthread */ |
62 | struct ptr_ring *queue; | |
63 | struct task_struct *kthread; | |
644bfe51 LB |
64 | |
65 | struct bpf_cpumap_val value; | |
92164774 | 66 | struct bpf_prog *prog; |
6710e112 JDB |
67 | |
68 | atomic_t refcnt; /* Control when this struct can be free'ed */ | |
69 | struct rcu_head rcu; | |
644bfe51 LB |
70 | |
71 | struct work_struct kthread_stop_wq; | |
6710e112 JDB |
72 | }; |
73 | ||
74 | struct bpf_cpu_map { | |
75 | struct bpf_map map; | |
76 | /* Below members specific for map type */ | |
77 | struct bpf_cpu_map_entry **cpu_map; | |
6710e112 JDB |
78 | }; |
79 | ||
cdfafe98 BT |
80 | static DEFINE_PER_CPU(struct list_head, cpu_map_flush_list); |
81 | ||
4bc188c7 | 82 | static int bq_flush_to_queue(struct xdp_bulk_queue *bq); |
6710e112 JDB |
83 | |
84 | static struct bpf_map *cpu_map_alloc(union bpf_attr *attr) | |
85 | { | |
92164774 | 86 | u32 value_size = attr->value_size; |
6710e112 JDB |
87 | struct bpf_cpu_map *cmap; |
88 | int err = -ENOMEM; | |
89 | u64 cost; | |
cdfafe98 | 90 | int ret; |
6710e112 | 91 | |
2c78ee89 | 92 | if (!bpf_capable()) |
6710e112 JDB |
93 | return ERR_PTR(-EPERM); |
94 | ||
95 | /* check sanity of attributes */ | |
96 | if (attr->max_entries == 0 || attr->key_size != 4 || | |
92164774 LB |
97 | (value_size != offsetofend(struct bpf_cpumap_val, qsize) && |
98 | value_size != offsetofend(struct bpf_cpumap_val, bpf_prog.fd)) || | |
99 | attr->map_flags & ~BPF_F_NUMA_NODE) | |
6710e112 JDB |
100 | return ERR_PTR(-EINVAL); |
101 | ||
102 | cmap = kzalloc(sizeof(*cmap), GFP_USER); | |
103 | if (!cmap) | |
104 | return ERR_PTR(-ENOMEM); | |
105 | ||
bd475643 | 106 | bpf_map_init_from_attr(&cmap->map, attr); |
6710e112 JDB |
107 | |
108 | /* Pre-limit array size based on NR_CPUS, not final CPU check */ | |
109 | if (cmap->map.max_entries > NR_CPUS) { | |
110 | err = -E2BIG; | |
111 | goto free_cmap; | |
112 | } | |
113 | ||
114 | /* make sure page count doesn't overflow */ | |
115 | cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *); | |
6710e112 JDB |
116 | |
117 | /* Notice returns -EPERM on if map size is larger than memlock limit */ | |
c85d6913 | 118 | ret = bpf_map_charge_init(&cmap->map.memory, cost); |
6710e112 JDB |
119 | if (ret) { |
120 | err = ret; | |
121 | goto free_cmap; | |
122 | } | |
123 | ||
6710e112 JDB |
124 | /* Alloc array for possible remote "destination" CPUs */ |
125 | cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries * | |
126 | sizeof(struct bpf_cpu_map_entry *), | |
127 | cmap->map.numa_node); | |
128 | if (!cmap->cpu_map) | |
cdfafe98 | 129 | goto free_charge; |
6710e112 JDB |
130 | |
131 | return &cmap->map; | |
b936ca64 RG |
132 | free_charge: |
133 | bpf_map_charge_finish(&cmap->map.memory); | |
6710e112 JDB |
134 | free_cmap: |
135 | kfree(cmap); | |
136 | return ERR_PTR(err); | |
137 | } | |
138 | ||
6710e112 JDB |
139 | static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu) |
140 | { | |
141 | atomic_inc(&rcpu->refcnt); | |
142 | } | |
143 | ||
144 | /* called from workqueue, to workaround syscall using preempt_disable */ | |
145 | static void cpu_map_kthread_stop(struct work_struct *work) | |
146 | { | |
147 | struct bpf_cpu_map_entry *rcpu; | |
148 | ||
149 | rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq); | |
150 | ||
151 | /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier, | |
152 | * as it waits until all in-flight call_rcu() callbacks complete. | |
153 | */ | |
154 | rcu_barrier(); | |
155 | ||
156 | /* kthread_stop will wake_up_process and wait for it to complete */ | |
157 | kthread_stop(rcpu->kthread); | |
158 | } | |
159 | ||
0fe875c5 | 160 | static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu, |
8f0504a9 JDB |
161 | struct xdp_frame *xdpf, |
162 | struct sk_buff *skb) | |
1c601d82 | 163 | { |
676e4a6f | 164 | unsigned int hard_start_headroom; |
1c601d82 JDB |
165 | unsigned int frame_size; |
166 | void *pkt_data_start; | |
1c601d82 | 167 | |
676e4a6f JDB |
168 | /* Part of headroom was reserved to xdpf */ |
169 | hard_start_headroom = sizeof(struct xdp_frame) + xdpf->headroom; | |
170 | ||
db612f74 JDB |
171 | /* Memory size backing xdp_frame data already have reserved |
172 | * room for build_skb to place skb_shared_info in tailroom. | |
1c601d82 | 173 | */ |
db612f74 | 174 | frame_size = xdpf->frame_sz; |
1c601d82 | 175 | |
676e4a6f | 176 | pkt_data_start = xdpf->data - hard_start_headroom; |
8f0504a9 JDB |
177 | skb = build_skb_around(skb, pkt_data_start, frame_size); |
178 | if (unlikely(!skb)) | |
1c601d82 JDB |
179 | return NULL; |
180 | ||
676e4a6f | 181 | skb_reserve(skb, hard_start_headroom); |
70280ed9 JDB |
182 | __skb_put(skb, xdpf->len); |
183 | if (xdpf->metasize) | |
184 | skb_metadata_set(skb, xdpf->metasize); | |
1c601d82 JDB |
185 | |
186 | /* Essential SKB info: protocol and skb->dev */ | |
70280ed9 | 187 | skb->protocol = eth_type_trans(skb, xdpf->dev_rx); |
1c601d82 JDB |
188 | |
189 | /* Optional SKB info, currently missing: | |
190 | * - HW checksum info (skb->ip_summed) | |
191 | * - HW RX hash (skb_set_hash) | |
192 | * - RX ring dev queue index (skb_record_rx_queue) | |
193 | */ | |
194 | ||
6bf071bf JDB |
195 | /* Until page_pool get SKB return path, release DMA here */ |
196 | xdp_release_frame(xdpf); | |
197 | ||
676e4a6f JDB |
198 | /* Allow SKB to reuse area used by xdp_frame */ |
199 | xdp_scrub_frame(xdpf); | |
200 | ||
1c601d82 JDB |
201 | return skb; |
202 | } | |
203 | ||
5ab073ff JDB |
204 | static void __cpu_map_ring_cleanup(struct ptr_ring *ring) |
205 | { | |
206 | /* The tear-down procedure should have made sure that queue is | |
207 | * empty. See __cpu_map_entry_replace() and work-queue | |
208 | * invoked cpu_map_kthread_stop(). Catch any broken behaviour | |
209 | * gracefully and warn once. | |
210 | */ | |
70280ed9 | 211 | struct xdp_frame *xdpf; |
5ab073ff | 212 | |
70280ed9 JDB |
213 | while ((xdpf = ptr_ring_consume(ring))) |
214 | if (WARN_ON_ONCE(xdpf)) | |
03993094 | 215 | xdp_return_frame(xdpf); |
5ab073ff JDB |
216 | } |
217 | ||
218 | static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu) | |
219 | { | |
220 | if (atomic_dec_and_test(&rcpu->refcnt)) { | |
92164774 LB |
221 | if (rcpu->prog) |
222 | bpf_prog_put(rcpu->prog); | |
5ab073ff JDB |
223 | /* The queue should be empty at this point */ |
224 | __cpu_map_ring_cleanup(rcpu->queue); | |
225 | ptr_ring_cleanup(rcpu->queue, NULL); | |
226 | kfree(rcpu->queue); | |
227 | kfree(rcpu); | |
228 | } | |
229 | } | |
230 | ||
92164774 LB |
231 | static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu, |
232 | void **frames, int n, | |
233 | struct xdp_cpumap_stats *stats) | |
234 | { | |
235 | struct xdp_rxq_info rxq; | |
236 | struct xdp_buff xdp; | |
237 | int i, nframes = 0; | |
238 | ||
239 | if (!rcpu->prog) | |
240 | return n; | |
241 | ||
28b1520e | 242 | rcu_read_lock_bh(); |
92164774 LB |
243 | |
244 | xdp_set_return_frame_no_direct(); | |
245 | xdp.rxq = &rxq; | |
246 | ||
247 | for (i = 0; i < n; i++) { | |
248 | struct xdp_frame *xdpf = frames[i]; | |
249 | u32 act; | |
250 | int err; | |
251 | ||
252 | rxq.dev = xdpf->dev_rx; | |
253 | rxq.mem = xdpf->mem; | |
254 | /* TODO: report queue_index to xdp_rxq_info */ | |
255 | ||
256 | xdp_convert_frame_to_buff(xdpf, &xdp); | |
257 | ||
258 | act = bpf_prog_run_xdp(rcpu->prog, &xdp); | |
259 | switch (act) { | |
260 | case XDP_PASS: | |
261 | err = xdp_update_frame_from_buff(&xdp, xdpf); | |
262 | if (err < 0) { | |
263 | xdp_return_frame(xdpf); | |
264 | stats->drop++; | |
265 | } else { | |
266 | frames[nframes++] = xdpf; | |
267 | stats->pass++; | |
268 | } | |
269 | break; | |
28b1520e LB |
270 | case XDP_REDIRECT: |
271 | err = xdp_do_redirect(xdpf->dev_rx, &xdp, | |
272 | rcpu->prog); | |
273 | if (unlikely(err)) { | |
274 | xdp_return_frame(xdpf); | |
275 | stats->drop++; | |
276 | } else { | |
277 | stats->redirect++; | |
278 | } | |
279 | break; | |
92164774 LB |
280 | default: |
281 | bpf_warn_invalid_xdp_action(act); | |
282 | /* fallthrough */ | |
283 | case XDP_DROP: | |
284 | xdp_return_frame(xdpf); | |
285 | stats->drop++; | |
286 | break; | |
287 | } | |
288 | } | |
289 | ||
28b1520e LB |
290 | if (stats->redirect) |
291 | xdp_do_flush_map(); | |
292 | ||
92164774 LB |
293 | xdp_clear_return_frame_no_direct(); |
294 | ||
28b1520e | 295 | rcu_read_unlock_bh(); /* resched point, may call do_softirq() */ |
92164774 LB |
296 | |
297 | return nframes; | |
298 | } | |
299 | ||
77361825 JDB |
300 | #define CPUMAP_BATCH 8 |
301 | ||
6710e112 JDB |
302 | static int cpu_map_kthread_run(void *data) |
303 | { | |
304 | struct bpf_cpu_map_entry *rcpu = data; | |
305 | ||
306 | set_current_state(TASK_INTERRUPTIBLE); | |
307 | ||
308 | /* When kthread gives stop order, then rcpu have been disconnected | |
309 | * from map, thus no new packets can enter. Remaining in-flight | |
310 | * per CPU stored packets are flushed to this queue. Wait honoring | |
311 | * kthread_stop signal until queue is empty. | |
312 | */ | |
313 | while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) { | |
92164774 LB |
314 | struct xdp_cpumap_stats stats = {}; /* zero stats */ |
315 | gfp_t gfp = __GFP_ZERO | GFP_ATOMIC; | |
77361825 JDB |
316 | unsigned int drops = 0, sched = 0; |
317 | void *frames[CPUMAP_BATCH]; | |
8f0504a9 | 318 | void *skbs[CPUMAP_BATCH]; |
92164774 | 319 | int i, n, m, nframes; |
6710e112 | 320 | |
1c601d82 JDB |
321 | /* Release CPU reschedule checks */ |
322 | if (__ptr_ring_empty(rcpu->queue)) { | |
31749468 JDB |
323 | set_current_state(TASK_INTERRUPTIBLE); |
324 | /* Recheck to avoid lost wake-up */ | |
325 | if (__ptr_ring_empty(rcpu->queue)) { | |
326 | schedule(); | |
327 | sched = 1; | |
328 | } else { | |
329 | __set_current_state(TASK_RUNNING); | |
330 | } | |
1c601d82 | 331 | } else { |
f9419f7b | 332 | sched = cond_resched(); |
6710e112 | 333 | } |
1c601d82 | 334 | |
1c601d82 JDB |
335 | /* |
336 | * The bpf_cpu_map_entry is single consumer, with this | |
337 | * kthread CPU pinned. Lockless access to ptr_ring | |
338 | * consume side valid as no-resize allowed of queue. | |
339 | */ | |
92164774 LB |
340 | n = __ptr_ring_consume_batched(rcpu->queue, frames, |
341 | CPUMAP_BATCH); | |
86d23145 JDB |
342 | for (i = 0; i < n; i++) { |
343 | void *f = frames[i]; | |
344 | struct page *page = virt_to_page(f); | |
345 | ||
346 | /* Bring struct page memory area to curr CPU. Read by | |
347 | * build_skb_around via page_is_pfmemalloc(), and when | |
348 | * freed written by page_frag_free call. | |
349 | */ | |
350 | prefetchw(page); | |
351 | } | |
352 | ||
92164774 LB |
353 | /* Support running another XDP prog on this CPU */ |
354 | nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, n, &stats); | |
355 | if (nframes) { | |
356 | m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, nframes, skbs); | |
357 | if (unlikely(m == 0)) { | |
358 | for (i = 0; i < nframes; i++) | |
359 | skbs[i] = NULL; /* effect: xdp_return_frame */ | |
360 | drops += nframes; | |
361 | } | |
8f0504a9 | 362 | } |
77361825 JDB |
363 | |
364 | local_bh_disable(); | |
92164774 | 365 | for (i = 0; i < nframes; i++) { |
77361825 | 366 | struct xdp_frame *xdpf = frames[i]; |
8f0504a9 | 367 | struct sk_buff *skb = skbs[i]; |
1c601d82 JDB |
368 | int ret; |
369 | ||
8f0504a9 | 370 | skb = cpu_map_build_skb(rcpu, xdpf, skb); |
1c601d82 | 371 | if (!skb) { |
03993094 | 372 | xdp_return_frame(xdpf); |
1c601d82 JDB |
373 | continue; |
374 | } | |
375 | ||
376 | /* Inject into network stack */ | |
377 | ret = netif_receive_skb_core(skb); | |
378 | if (ret == NET_RX_DROP) | |
379 | drops++; | |
1c601d82 | 380 | } |
f9419f7b | 381 | /* Feedback loop via tracepoint */ |
92164774 | 382 | trace_xdp_cpumap_kthread(rcpu->map_id, n, drops, sched, &stats); |
f9419f7b | 383 | |
1c601d82 | 384 | local_bh_enable(); /* resched point, may call do_softirq() */ |
6710e112 JDB |
385 | } |
386 | __set_current_state(TASK_RUNNING); | |
387 | ||
388 | put_cpu_map_entry(rcpu); | |
389 | return 0; | |
390 | } | |
391 | ||
92164774 LB |
392 | bool cpu_map_prog_allowed(struct bpf_map *map) |
393 | { | |
394 | return map->map_type == BPF_MAP_TYPE_CPUMAP && | |
395 | map->value_size != offsetofend(struct bpf_cpumap_val, qsize); | |
396 | } | |
397 | ||
398 | static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu, int fd) | |
399 | { | |
400 | struct bpf_prog *prog; | |
401 | ||
402 | prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP); | |
403 | if (IS_ERR(prog)) | |
404 | return PTR_ERR(prog); | |
405 | ||
406 | if (prog->expected_attach_type != BPF_XDP_CPUMAP) { | |
407 | bpf_prog_put(prog); | |
408 | return -EINVAL; | |
409 | } | |
410 | ||
411 | rcpu->value.bpf_prog.id = prog->aux->id; | |
412 | rcpu->prog = prog; | |
413 | ||
414 | return 0; | |
415 | } | |
416 | ||
644bfe51 LB |
417 | static struct bpf_cpu_map_entry * |
418 | __cpu_map_entry_alloc(struct bpf_cpumap_val *value, u32 cpu, int map_id) | |
6710e112 | 419 | { |
92164774 | 420 | int numa, err, i, fd = value->bpf_prog.fd; |
7fc17e90 | 421 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
6710e112 | 422 | struct bpf_cpu_map_entry *rcpu; |
d5df2830 | 423 | struct xdp_bulk_queue *bq; |
6710e112 JDB |
424 | |
425 | /* Have map->numa_node, but choose node of redirect target CPU */ | |
426 | numa = cpu_to_node(cpu); | |
427 | ||
428 | rcpu = kzalloc_node(sizeof(*rcpu), gfp, numa); | |
429 | if (!rcpu) | |
430 | return NULL; | |
431 | ||
432 | /* Alloc percpu bulkq */ | |
433 | rcpu->bulkq = __alloc_percpu_gfp(sizeof(*rcpu->bulkq), | |
434 | sizeof(void *), gfp); | |
435 | if (!rcpu->bulkq) | |
436 | goto free_rcu; | |
437 | ||
d5df2830 THJ |
438 | for_each_possible_cpu(i) { |
439 | bq = per_cpu_ptr(rcpu->bulkq, i); | |
440 | bq->obj = rcpu; | |
441 | } | |
442 | ||
6710e112 JDB |
443 | /* Alloc queue */ |
444 | rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa); | |
445 | if (!rcpu->queue) | |
446 | goto free_bulkq; | |
447 | ||
644bfe51 | 448 | err = ptr_ring_init(rcpu->queue, value->qsize, gfp); |
6710e112 JDB |
449 | if (err) |
450 | goto free_queue; | |
451 | ||
f9419f7b JDB |
452 | rcpu->cpu = cpu; |
453 | rcpu->map_id = map_id; | |
644bfe51 | 454 | rcpu->value.qsize = value->qsize; |
6710e112 | 455 | |
c576b9c7 LB |
456 | if (fd > 0 && __cpu_map_load_bpf_program(rcpu, fd)) |
457 | goto free_ptr_ring; | |
458 | ||
6710e112 JDB |
459 | /* Setup kthread */ |
460 | rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa, | |
461 | "cpumap/%d/map:%d", cpu, map_id); | |
462 | if (IS_ERR(rcpu->kthread)) | |
c576b9c7 | 463 | goto free_prog; |
6710e112 JDB |
464 | |
465 | get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */ | |
466 | get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */ | |
467 | ||
468 | /* Make sure kthread runs on a single CPU */ | |
469 | kthread_bind(rcpu->kthread, cpu); | |
470 | wake_up_process(rcpu->kthread); | |
471 | ||
472 | return rcpu; | |
473 | ||
c576b9c7 LB |
474 | free_prog: |
475 | if (rcpu->prog) | |
476 | bpf_prog_put(rcpu->prog); | |
6710e112 JDB |
477 | free_ptr_ring: |
478 | ptr_ring_cleanup(rcpu->queue, NULL); | |
479 | free_queue: | |
480 | kfree(rcpu->queue); | |
481 | free_bulkq: | |
482 | free_percpu(rcpu->bulkq); | |
483 | free_rcu: | |
484 | kfree(rcpu); | |
485 | return NULL; | |
486 | } | |
487 | ||
0fe875c5 | 488 | static void __cpu_map_entry_free(struct rcu_head *rcu) |
6710e112 JDB |
489 | { |
490 | struct bpf_cpu_map_entry *rcpu; | |
6710e112 JDB |
491 | |
492 | /* This cpu_map_entry have been disconnected from map and one | |
fb5aacdf | 493 | * RCU grace-period have elapsed. Thus, XDP cannot queue any |
6710e112 JDB |
494 | * new packets and cannot change/set flush_needed that can |
495 | * find this entry. | |
496 | */ | |
497 | rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu); | |
498 | ||
6710e112 JDB |
499 | free_percpu(rcpu->bulkq); |
500 | /* Cannot kthread_stop() here, last put free rcpu resources */ | |
501 | put_cpu_map_entry(rcpu); | |
502 | } | |
503 | ||
504 | /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to | |
505 | * ensure any driver rcu critical sections have completed, but this | |
506 | * does not guarantee a flush has happened yet. Because driver side | |
507 | * rcu_read_lock/unlock only protects the running XDP program. The | |
508 | * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a | |
509 | * pending flush op doesn't fail. | |
510 | * | |
511 | * The bpf_cpu_map_entry is still used by the kthread, and there can | |
512 | * still be pending packets (in queue and percpu bulkq). A refcnt | |
513 | * makes sure to last user (kthread_stop vs. call_rcu) free memory | |
514 | * resources. | |
515 | * | |
516 | * The rcu callback __cpu_map_entry_free flush remaining packets in | |
517 | * percpu bulkq to queue. Due to caller map_delete_elem() disable | |
518 | * preemption, cannot call kthread_stop() to make sure queue is empty. | |
519 | * Instead a work_queue is started for stopping kthread, | |
fb5aacdf | 520 | * cpu_map_kthread_stop, which waits for an RCU grace period before |
6710e112 JDB |
521 | * stopping kthread, emptying the queue. |
522 | */ | |
0fe875c5 WY |
523 | static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap, |
524 | u32 key_cpu, struct bpf_cpu_map_entry *rcpu) | |
6710e112 JDB |
525 | { |
526 | struct bpf_cpu_map_entry *old_rcpu; | |
527 | ||
528 | old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu); | |
529 | if (old_rcpu) { | |
530 | call_rcu(&old_rcpu->rcu, __cpu_map_entry_free); | |
531 | INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop); | |
532 | schedule_work(&old_rcpu->kthread_stop_wq); | |
533 | } | |
534 | } | |
535 | ||
0fe875c5 | 536 | static int cpu_map_delete_elem(struct bpf_map *map, void *key) |
6710e112 JDB |
537 | { |
538 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); | |
539 | u32 key_cpu = *(u32 *)key; | |
540 | ||
541 | if (key_cpu >= map->max_entries) | |
542 | return -EINVAL; | |
543 | ||
544 | /* notice caller map_delete_elem() use preempt_disable() */ | |
545 | __cpu_map_entry_replace(cmap, key_cpu, NULL); | |
546 | return 0; | |
547 | } | |
548 | ||
0fe875c5 WY |
549 | static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value, |
550 | u64 map_flags) | |
6710e112 JDB |
551 | { |
552 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); | |
644bfe51 | 553 | struct bpf_cpumap_val cpumap_value = {}; |
6710e112 | 554 | struct bpf_cpu_map_entry *rcpu; |
6710e112 JDB |
555 | /* Array index key correspond to CPU number */ |
556 | u32 key_cpu = *(u32 *)key; | |
644bfe51 LB |
557 | |
558 | memcpy(&cpumap_value, value, map->value_size); | |
6710e112 JDB |
559 | |
560 | if (unlikely(map_flags > BPF_EXIST)) | |
561 | return -EINVAL; | |
562 | if (unlikely(key_cpu >= cmap->map.max_entries)) | |
563 | return -E2BIG; | |
564 | if (unlikely(map_flags == BPF_NOEXIST)) | |
565 | return -EEXIST; | |
644bfe51 | 566 | if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */ |
6710e112 JDB |
567 | return -EOVERFLOW; |
568 | ||
569 | /* Make sure CPU is a valid possible cpu */ | |
bc23d0e3 | 570 | if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu)) |
6710e112 JDB |
571 | return -ENODEV; |
572 | ||
644bfe51 | 573 | if (cpumap_value.qsize == 0) { |
6710e112 JDB |
574 | rcpu = NULL; /* Same as deleting */ |
575 | } else { | |
576 | /* Updating qsize cause re-allocation of bpf_cpu_map_entry */ | |
644bfe51 | 577 | rcpu = __cpu_map_entry_alloc(&cpumap_value, key_cpu, map->id); |
6710e112 JDB |
578 | if (!rcpu) |
579 | return -ENOMEM; | |
d5df2830 | 580 | rcpu->cmap = cmap; |
6710e112 JDB |
581 | } |
582 | rcu_read_lock(); | |
583 | __cpu_map_entry_replace(cmap, key_cpu, rcpu); | |
584 | rcu_read_unlock(); | |
585 | return 0; | |
586 | } | |
587 | ||
0fe875c5 | 588 | static void cpu_map_free(struct bpf_map *map) |
6710e112 JDB |
589 | { |
590 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); | |
6710e112 JDB |
591 | u32 i; |
592 | ||
593 | /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0, | |
594 | * so the bpf programs (can be more than one that used this map) were | |
595 | * disconnected from events. Wait for outstanding critical sections in | |
596 | * these programs to complete. The rcu critical section only guarantees | |
597 | * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map. | |
598 | * It does __not__ ensure pending flush operations (if any) are | |
599 | * complete. | |
600 | */ | |
f6069b9a DB |
601 | |
602 | bpf_clear_redirect_map(map); | |
6710e112 JDB |
603 | synchronize_rcu(); |
604 | ||
6710e112 JDB |
605 | /* For cpu_map the remote CPUs can still be using the entries |
606 | * (struct bpf_cpu_map_entry). | |
607 | */ | |
608 | for (i = 0; i < cmap->map.max_entries; i++) { | |
609 | struct bpf_cpu_map_entry *rcpu; | |
610 | ||
611 | rcpu = READ_ONCE(cmap->cpu_map[i]); | |
612 | if (!rcpu) | |
613 | continue; | |
614 | ||
fb5aacdf | 615 | /* bq flush and cleanup happens after RCU grace-period */ |
6710e112 JDB |
616 | __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */ |
617 | } | |
6710e112 JDB |
618 | bpf_map_area_free(cmap->cpu_map); |
619 | kfree(cmap); | |
620 | } | |
621 | ||
622 | struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key) | |
623 | { | |
624 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); | |
625 | struct bpf_cpu_map_entry *rcpu; | |
626 | ||
627 | if (key >= map->max_entries) | |
628 | return NULL; | |
629 | ||
630 | rcpu = READ_ONCE(cmap->cpu_map[key]); | |
631 | return rcpu; | |
632 | } | |
633 | ||
634 | static void *cpu_map_lookup_elem(struct bpf_map *map, void *key) | |
635 | { | |
636 | struct bpf_cpu_map_entry *rcpu = | |
637 | __cpu_map_lookup_elem(map, *(u32 *)key); | |
638 | ||
644bfe51 | 639 | return rcpu ? &rcpu->value : NULL; |
6710e112 JDB |
640 | } |
641 | ||
642 | static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key) | |
643 | { | |
644 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); | |
645 | u32 index = key ? *(u32 *)key : U32_MAX; | |
646 | u32 *next = next_key; | |
647 | ||
648 | if (index >= cmap->map.max_entries) { | |
649 | *next = 0; | |
650 | return 0; | |
651 | } | |
652 | ||
653 | if (index == cmap->map.max_entries - 1) | |
654 | return -ENOENT; | |
655 | *next = index + 1; | |
656 | return 0; | |
657 | } | |
658 | ||
2872e9ac | 659 | static int cpu_map_btf_id; |
6710e112 JDB |
660 | const struct bpf_map_ops cpu_map_ops = { |
661 | .map_alloc = cpu_map_alloc, | |
662 | .map_free = cpu_map_free, | |
663 | .map_delete_elem = cpu_map_delete_elem, | |
664 | .map_update_elem = cpu_map_update_elem, | |
665 | .map_lookup_elem = cpu_map_lookup_elem, | |
666 | .map_get_next_key = cpu_map_get_next_key, | |
e8d2bec0 | 667 | .map_check_btf = map_check_no_btf, |
2872e9ac AI |
668 | .map_btf_name = "bpf_cpu_map", |
669 | .map_btf_id = &cpu_map_btf_id, | |
6710e112 JDB |
670 | }; |
671 | ||
4bc188c7 | 672 | static int bq_flush_to_queue(struct xdp_bulk_queue *bq) |
6710e112 | 673 | { |
d5df2830 | 674 | struct bpf_cpu_map_entry *rcpu = bq->obj; |
f9419f7b JDB |
675 | unsigned int processed = 0, drops = 0; |
676 | const int to_cpu = rcpu->cpu; | |
6710e112 JDB |
677 | struct ptr_ring *q; |
678 | int i; | |
679 | ||
680 | if (unlikely(!bq->count)) | |
681 | return 0; | |
682 | ||
683 | q = rcpu->queue; | |
684 | spin_lock(&q->producer_lock); | |
685 | ||
686 | for (i = 0; i < bq->count; i++) { | |
70280ed9 | 687 | struct xdp_frame *xdpf = bq->q[i]; |
6710e112 JDB |
688 | int err; |
689 | ||
70280ed9 | 690 | err = __ptr_ring_produce(q, xdpf); |
6710e112 | 691 | if (err) { |
f9419f7b | 692 | drops++; |
4bc188c7 | 693 | xdp_return_frame_rx_napi(xdpf); |
6710e112 | 694 | } |
f9419f7b | 695 | processed++; |
6710e112 JDB |
696 | } |
697 | bq->count = 0; | |
698 | spin_unlock(&q->producer_lock); | |
699 | ||
d5df2830 THJ |
700 | __list_del_clearprev(&bq->flush_node); |
701 | ||
f9419f7b JDB |
702 | /* Feedback loop via tracepoints */ |
703 | trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu); | |
6710e112 JDB |
704 | return 0; |
705 | } | |
706 | ||
6710e112 JDB |
707 | /* Runs under RCU-read-side, plus in softirq under NAPI protection. |
708 | * Thus, safe percpu variable access. | |
709 | */ | |
70280ed9 | 710 | static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf) |
6710e112 | 711 | { |
cdfafe98 | 712 | struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list); |
6710e112 JDB |
713 | struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq); |
714 | ||
715 | if (unlikely(bq->count == CPU_MAP_BULK_SIZE)) | |
4bc188c7 | 716 | bq_flush_to_queue(bq); |
6710e112 JDB |
717 | |
718 | /* Notice, xdp_buff/page MUST be queued here, long enough for | |
719 | * driver to code invoking us to finished, due to driver | |
720 | * (e.g. ixgbe) recycle tricks based on page-refcnt. | |
721 | * | |
70280ed9 | 722 | * Thus, incoming xdp_frame is always queued here (else we race |
6710e112 JDB |
723 | * with another CPU on page-refcnt and remaining driver code). |
724 | * Queue time is very short, as driver will invoke flush | |
725 | * operation, when completing napi->poll call. | |
726 | */ | |
70280ed9 | 727 | bq->q[bq->count++] = xdpf; |
d5df2830 THJ |
728 | |
729 | if (!bq->flush_node.prev) | |
730 | list_add(&bq->flush_node, flush_list); | |
731 | ||
6710e112 JDB |
732 | return 0; |
733 | } | |
734 | ||
9c270af3 JDB |
735 | int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp, |
736 | struct net_device *dev_rx) | |
737 | { | |
70280ed9 | 738 | struct xdp_frame *xdpf; |
9c270af3 | 739 | |
1b698fa5 | 740 | xdpf = xdp_convert_buff_to_frame(xdp); |
70280ed9 | 741 | if (unlikely(!xdpf)) |
1c601d82 | 742 | return -EOVERFLOW; |
9c270af3 | 743 | |
1c601d82 | 744 | /* Info needed when constructing SKB on remote CPU */ |
70280ed9 | 745 | xdpf->dev_rx = dev_rx; |
9c270af3 | 746 | |
70280ed9 | 747 | bq_enqueue(rcpu, xdpf); |
9c270af3 JDB |
748 | return 0; |
749 | } | |
750 | ||
cdfafe98 | 751 | void __cpu_map_flush(void) |
6710e112 | 752 | { |
cdfafe98 | 753 | struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list); |
d5df2830 | 754 | struct xdp_bulk_queue *bq, *tmp; |
6710e112 | 755 | |
d5df2830 | 756 | list_for_each_entry_safe(bq, tmp, flush_list, flush_node) { |
4bc188c7 | 757 | bq_flush_to_queue(bq); |
6710e112 JDB |
758 | |
759 | /* If already running, costs spin_lock_irqsave + smb_mb */ | |
d5df2830 | 760 | wake_up_process(bq->obj->kthread); |
6710e112 JDB |
761 | } |
762 | } | |
cdfafe98 BT |
763 | |
764 | static int __init cpu_map_init(void) | |
765 | { | |
766 | int cpu; | |
767 | ||
768 | for_each_possible_cpu(cpu) | |
769 | INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu)); | |
770 | return 0; | |
771 | } | |
772 | ||
773 | subsys_initcall(cpu_map_init); |