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