1 | // SPDX-License-Identifier: GPL-2.0-only |
2 | /* bpf/cpumap.c |
3 | * |
4 | * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc. |
5 | */ |
6 | |
7 | /** |
8 | * DOC: cpu map |
9 | * The 'cpumap' is primarily used as a backend map for XDP BPF helper |
10 | * call bpf_redirect_map() and XDP_REDIRECT action, like 'devmap'. |
11 | * |
12 | * Unlike devmap which redirects XDP frames out to another NIC device, |
13 | * this map type redirects raw XDP frames to another CPU. The remote |
14 | * CPU will do SKB-allocation and call the normal network stack. |
15 | */ |
16 | /* |
17 | * This is a scalability and isolation mechanism, that allow |
18 | * separating the early driver network XDP layer, from the rest of the |
19 | * netstack, and assigning dedicated CPUs for this stage. This |
20 | * basically allows for 10G wirespeed pre-filtering via bpf. |
21 | */ |
22 | #include <linux/bitops.h> |
23 | #include <linux/bpf.h> |
24 | #include <linux/filter.h> |
25 | #include <linux/ptr_ring.h> |
26 | #include <net/xdp.h> |
27 | #include <net/hotdata.h> |
28 | |
29 | #include <linux/sched.h> |
30 | #include <linux/workqueue.h> |
31 | #include <linux/kthread.h> |
32 | #include <linux/completion.h> |
33 | #include <trace/events/xdp.h> |
34 | #include <linux/btf_ids.h> |
35 | |
36 | #include <linux/netdevice.h> /* netif_receive_skb_list */ |
37 | #include <linux/etherdevice.h> /* eth_type_trans */ |
38 | |
39 | /* General idea: XDP packets getting XDP redirected to another CPU, |
40 | * will maximum be stored/queued for one driver ->poll() call. It is |
41 | * guaranteed that queueing the frame and the flush operation happen on |
42 | * same CPU. Thus, cpu_map_flush operation can deduct via this_cpu_ptr() |
43 | * which queue in bpf_cpu_map_entry contains packets. |
44 | */ |
45 | |
46 | #define CPU_MAP_BULK_SIZE 8 /* 8 == one cacheline on 64-bit archs */ |
47 | struct bpf_cpu_map_entry; |
48 | struct bpf_cpu_map; |
49 | |
50 | struct xdp_bulk_queue { |
51 | void *q[CPU_MAP_BULK_SIZE]; |
52 | struct list_head flush_node; |
53 | struct bpf_cpu_map_entry *obj; |
54 | unsigned int count; |
55 | }; |
56 | |
57 | /* Struct for every remote "destination" CPU in map */ |
58 | struct bpf_cpu_map_entry { |
59 | u32 cpu; /* kthread CPU and map index */ |
60 | int map_id; /* Back reference to map */ |
61 | |
62 | /* XDP can run multiple RX-ring queues, need __percpu enqueue store */ |
63 | struct xdp_bulk_queue __percpu *bulkq; |
64 | |
65 | /* Queue with potential multi-producers, and single-consumer kthread */ |
66 | struct ptr_ring *queue; |
67 | struct task_struct *kthread; |
68 | |
69 | struct bpf_cpumap_val value; |
70 | struct bpf_prog *prog; |
71 | |
72 | struct completion kthread_running; |
73 | struct rcu_work free_work; |
74 | }; |
75 | |
76 | struct bpf_cpu_map { |
77 | struct bpf_map map; |
78 | /* Below members specific for map type */ |
79 | struct bpf_cpu_map_entry __rcu **cpu_map; |
80 | }; |
81 | |
82 | static DEFINE_PER_CPU(struct list_head, cpu_map_flush_list); |
83 | |
84 | static struct bpf_map *cpu_map_alloc(union bpf_attr *attr) |
85 | { |
86 | u32 value_size = attr->value_size; |
87 | struct bpf_cpu_map *cmap; |
88 | |
89 | /* check sanity of attributes */ |
90 | if (attr->max_entries == 0 || attr->key_size != 4 || |
91 | (value_size != offsetofend(struct bpf_cpumap_val, qsize) && |
92 | value_size != offsetofend(struct bpf_cpumap_val, bpf_prog.fd)) || |
93 | attr->map_flags & ~BPF_F_NUMA_NODE) |
94 | return ERR_PTR(error: -EINVAL); |
95 | |
96 | /* Pre-limit array size based on NR_CPUS, not final CPU check */ |
97 | if (attr->max_entries > NR_CPUS) |
98 | return ERR_PTR(error: -E2BIG); |
99 | |
100 | cmap = bpf_map_area_alloc(size: sizeof(*cmap), NUMA_NO_NODE); |
101 | if (!cmap) |
102 | return ERR_PTR(error: -ENOMEM); |
103 | |
104 | bpf_map_init_from_attr(map: &cmap->map, attr); |
105 | |
106 | /* Alloc array for possible remote "destination" CPUs */ |
107 | cmap->cpu_map = bpf_map_area_alloc(size: cmap->map.max_entries * |
108 | sizeof(struct bpf_cpu_map_entry *), |
109 | numa_node: cmap->map.numa_node); |
110 | if (!cmap->cpu_map) { |
111 | bpf_map_area_free(base: cmap); |
112 | return ERR_PTR(error: -ENOMEM); |
113 | } |
114 | |
115 | return &cmap->map; |
116 | } |
117 | |
118 | static void __cpu_map_ring_cleanup(struct ptr_ring *ring) |
119 | { |
120 | /* The tear-down procedure should have made sure that queue is |
121 | * empty. See __cpu_map_entry_replace() and work-queue |
122 | * invoked cpu_map_kthread_stop(). Catch any broken behaviour |
123 | * gracefully and warn once. |
124 | */ |
125 | void *ptr; |
126 | |
127 | while ((ptr = ptr_ring_consume(r: ring))) { |
128 | WARN_ON_ONCE(1); |
129 | if (unlikely(__ptr_test_bit(0, &ptr))) { |
130 | __ptr_clear_bit(0, &ptr); |
131 | kfree_skb(skb: ptr); |
132 | continue; |
133 | } |
134 | xdp_return_frame(xdpf: ptr); |
135 | } |
136 | } |
137 | |
138 | static void cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu, |
139 | struct list_head *listp, |
140 | struct xdp_cpumap_stats *stats) |
141 | { |
142 | struct sk_buff *skb, *tmp; |
143 | struct xdp_buff xdp; |
144 | u32 act; |
145 | int err; |
146 | |
147 | list_for_each_entry_safe(skb, tmp, listp, list) { |
148 | act = bpf_prog_run_generic_xdp(skb, xdp: &xdp, xdp_prog: rcpu->prog); |
149 | switch (act) { |
150 | case XDP_PASS: |
151 | break; |
152 | case XDP_REDIRECT: |
153 | skb_list_del_init(skb); |
154 | err = xdp_do_generic_redirect(dev: skb->dev, skb, xdp: &xdp, |
155 | prog: rcpu->prog); |
156 | if (unlikely(err)) { |
157 | kfree_skb(skb); |
158 | stats->drop++; |
159 | } else { |
160 | stats->redirect++; |
161 | } |
162 | return; |
163 | default: |
164 | bpf_warn_invalid_xdp_action(NULL, prog: rcpu->prog, act); |
165 | fallthrough; |
166 | case XDP_ABORTED: |
167 | trace_xdp_exception(dev: skb->dev, xdp: rcpu->prog, act); |
168 | fallthrough; |
169 | case XDP_DROP: |
170 | skb_list_del_init(skb); |
171 | kfree_skb(skb); |
172 | stats->drop++; |
173 | return; |
174 | } |
175 | } |
176 | } |
177 | |
178 | static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu, |
179 | void **frames, int n, |
180 | struct xdp_cpumap_stats *stats) |
181 | { |
182 | struct xdp_rxq_info rxq = {}; |
183 | struct xdp_buff xdp; |
184 | int i, nframes = 0; |
185 | |
186 | xdp_set_return_frame_no_direct(); |
187 | xdp.rxq = &rxq; |
188 | |
189 | for (i = 0; i < n; i++) { |
190 | struct xdp_frame *xdpf = frames[i]; |
191 | u32 act; |
192 | int err; |
193 | |
194 | rxq.dev = xdpf->dev_rx; |
195 | rxq.mem = xdpf->mem; |
196 | /* TODO: report queue_index to xdp_rxq_info */ |
197 | |
198 | xdp_convert_frame_to_buff(frame: xdpf, xdp: &xdp); |
199 | |
200 | act = bpf_prog_run_xdp(prog: rcpu->prog, xdp: &xdp); |
201 | switch (act) { |
202 | case XDP_PASS: |
203 | err = xdp_update_frame_from_buff(xdp: &xdp, xdp_frame: xdpf); |
204 | if (err < 0) { |
205 | xdp_return_frame(xdpf); |
206 | stats->drop++; |
207 | } else { |
208 | frames[nframes++] = xdpf; |
209 | stats->pass++; |
210 | } |
211 | break; |
212 | case XDP_REDIRECT: |
213 | err = xdp_do_redirect(dev: xdpf->dev_rx, xdp: &xdp, |
214 | prog: rcpu->prog); |
215 | if (unlikely(err)) { |
216 | xdp_return_frame(xdpf); |
217 | stats->drop++; |
218 | } else { |
219 | stats->redirect++; |
220 | } |
221 | break; |
222 | default: |
223 | bpf_warn_invalid_xdp_action(NULL, prog: rcpu->prog, act); |
224 | fallthrough; |
225 | case XDP_DROP: |
226 | xdp_return_frame(xdpf); |
227 | stats->drop++; |
228 | break; |
229 | } |
230 | } |
231 | |
232 | xdp_clear_return_frame_no_direct(); |
233 | |
234 | return nframes; |
235 | } |
236 | |
237 | #define CPUMAP_BATCH 8 |
238 | |
239 | static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames, |
240 | int xdp_n, struct xdp_cpumap_stats *stats, |
241 | struct list_head *list) |
242 | { |
243 | int nframes; |
244 | |
245 | if (!rcpu->prog) |
246 | return xdp_n; |
247 | |
248 | rcu_read_lock_bh(); |
249 | |
250 | nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, n: xdp_n, stats); |
251 | |
252 | if (stats->redirect) |
253 | xdp_do_flush(); |
254 | |
255 | if (unlikely(!list_empty(list))) |
256 | cpu_map_bpf_prog_run_skb(rcpu, listp: list, stats); |
257 | |
258 | rcu_read_unlock_bh(); /* resched point, may call do_softirq() */ |
259 | |
260 | return nframes; |
261 | } |
262 | |
263 | static int cpu_map_kthread_run(void *data) |
264 | { |
265 | struct bpf_cpu_map_entry *rcpu = data; |
266 | unsigned long last_qs = jiffies; |
267 | |
268 | complete(&rcpu->kthread_running); |
269 | set_current_state(TASK_INTERRUPTIBLE); |
270 | |
271 | /* When kthread gives stop order, then rcpu have been disconnected |
272 | * from map, thus no new packets can enter. Remaining in-flight |
273 | * per CPU stored packets are flushed to this queue. Wait honoring |
274 | * kthread_stop signal until queue is empty. |
275 | */ |
276 | while (!kthread_should_stop() || !__ptr_ring_empty(r: rcpu->queue)) { |
277 | struct xdp_cpumap_stats stats = {}; /* zero stats */ |
278 | unsigned int kmem_alloc_drops = 0, sched = 0; |
279 | gfp_t gfp = __GFP_ZERO | GFP_ATOMIC; |
280 | int i, n, m, nframes, xdp_n; |
281 | void *frames[CPUMAP_BATCH]; |
282 | void *skbs[CPUMAP_BATCH]; |
283 | LIST_HEAD(list); |
284 | |
285 | /* Release CPU reschedule checks */ |
286 | if (__ptr_ring_empty(r: rcpu->queue)) { |
287 | set_current_state(TASK_INTERRUPTIBLE); |
288 | /* Recheck to avoid lost wake-up */ |
289 | if (__ptr_ring_empty(r: rcpu->queue)) { |
290 | schedule(); |
291 | sched = 1; |
292 | last_qs = jiffies; |
293 | } else { |
294 | __set_current_state(TASK_RUNNING); |
295 | } |
296 | } else { |
297 | rcu_softirq_qs_periodic(last_qs); |
298 | sched = cond_resched(); |
299 | } |
300 | |
301 | /* |
302 | * The bpf_cpu_map_entry is single consumer, with this |
303 | * kthread CPU pinned. Lockless access to ptr_ring |
304 | * consume side valid as no-resize allowed of queue. |
305 | */ |
306 | n = __ptr_ring_consume_batched(r: rcpu->queue, array: frames, |
307 | CPUMAP_BATCH); |
308 | for (i = 0, xdp_n = 0; i < n; i++) { |
309 | void *f = frames[i]; |
310 | struct page *page; |
311 | |
312 | if (unlikely(__ptr_test_bit(0, &f))) { |
313 | struct sk_buff *skb = f; |
314 | |
315 | __ptr_clear_bit(0, &skb); |
316 | list_add_tail(new: &skb->list, head: &list); |
317 | continue; |
318 | } |
319 | |
320 | frames[xdp_n++] = f; |
321 | page = virt_to_page(f); |
322 | |
323 | /* Bring struct page memory area to curr CPU. Read by |
324 | * build_skb_around via page_is_pfmemalloc(), and when |
325 | * freed written by page_frag_free call. |
326 | */ |
327 | prefetchw(x: page); |
328 | } |
329 | |
330 | /* Support running another XDP prog on this CPU */ |
331 | nframes = cpu_map_bpf_prog_run(rcpu, frames, xdp_n, stats: &stats, list: &list); |
332 | if (nframes) { |
333 | m = kmem_cache_alloc_bulk(s: net_hotdata.skbuff_cache, |
334 | flags: gfp, size: nframes, p: skbs); |
335 | if (unlikely(m == 0)) { |
336 | for (i = 0; i < nframes; i++) |
337 | skbs[i] = NULL; /* effect: xdp_return_frame */ |
338 | kmem_alloc_drops += nframes; |
339 | } |
340 | } |
341 | |
342 | local_bh_disable(); |
343 | for (i = 0; i < nframes; i++) { |
344 | struct xdp_frame *xdpf = frames[i]; |
345 | struct sk_buff *skb = skbs[i]; |
346 | |
347 | skb = __xdp_build_skb_from_frame(xdpf, skb, |
348 | dev: xdpf->dev_rx); |
349 | if (!skb) { |
350 | xdp_return_frame(xdpf); |
351 | continue; |
352 | } |
353 | |
354 | list_add_tail(new: &skb->list, head: &list); |
355 | } |
356 | netif_receive_skb_list(head: &list); |
357 | |
358 | /* Feedback loop via tracepoint */ |
359 | trace_xdp_cpumap_kthread(map_id: rcpu->map_id, processed: n, drops: kmem_alloc_drops, |
360 | sched, xdp_stats: &stats); |
361 | |
362 | local_bh_enable(); /* resched point, may call do_softirq() */ |
363 | } |
364 | __set_current_state(TASK_RUNNING); |
365 | |
366 | return 0; |
367 | } |
368 | |
369 | static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu, |
370 | struct bpf_map *map, int fd) |
371 | { |
372 | struct bpf_prog *prog; |
373 | |
374 | prog = bpf_prog_get_type(ufd: fd, type: BPF_PROG_TYPE_XDP); |
375 | if (IS_ERR(ptr: prog)) |
376 | return PTR_ERR(ptr: prog); |
377 | |
378 | if (prog->expected_attach_type != BPF_XDP_CPUMAP || |
379 | !bpf_prog_map_compatible(map, fp: prog)) { |
380 | bpf_prog_put(prog); |
381 | return -EINVAL; |
382 | } |
383 | |
384 | rcpu->value.bpf_prog.id = prog->aux->id; |
385 | rcpu->prog = prog; |
386 | |
387 | return 0; |
388 | } |
389 | |
390 | static struct bpf_cpu_map_entry * |
391 | __cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value, |
392 | u32 cpu) |
393 | { |
394 | int numa, err, i, fd = value->bpf_prog.fd; |
395 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
396 | struct bpf_cpu_map_entry *rcpu; |
397 | struct xdp_bulk_queue *bq; |
398 | |
399 | /* Have map->numa_node, but choose node of redirect target CPU */ |
400 | numa = cpu_to_node(cpu); |
401 | |
402 | rcpu = bpf_map_kmalloc_node(map, size: sizeof(*rcpu), flags: gfp | __GFP_ZERO, node: numa); |
403 | if (!rcpu) |
404 | return NULL; |
405 | |
406 | /* Alloc percpu bulkq */ |
407 | rcpu->bulkq = bpf_map_alloc_percpu(map, size: sizeof(*rcpu->bulkq), |
408 | align: sizeof(void *), flags: gfp); |
409 | if (!rcpu->bulkq) |
410 | goto free_rcu; |
411 | |
412 | for_each_possible_cpu(i) { |
413 | bq = per_cpu_ptr(rcpu->bulkq, i); |
414 | bq->obj = rcpu; |
415 | } |
416 | |
417 | /* Alloc queue */ |
418 | rcpu->queue = bpf_map_kmalloc_node(map, size: sizeof(*rcpu->queue), flags: gfp, |
419 | node: numa); |
420 | if (!rcpu->queue) |
421 | goto free_bulkq; |
422 | |
423 | err = ptr_ring_init(r: rcpu->queue, size: value->qsize, gfp); |
424 | if (err) |
425 | goto free_queue; |
426 | |
427 | rcpu->cpu = cpu; |
428 | rcpu->map_id = map->id; |
429 | rcpu->value.qsize = value->qsize; |
430 | |
431 | if (fd > 0 && __cpu_map_load_bpf_program(rcpu, map, fd)) |
432 | goto free_ptr_ring; |
433 | |
434 | /* Setup kthread */ |
435 | init_completion(x: &rcpu->kthread_running); |
436 | rcpu->kthread = kthread_create_on_node(threadfn: cpu_map_kthread_run, data: rcpu, node: numa, |
437 | namefmt: "cpumap/%d/map:%d" , cpu, |
438 | map->id); |
439 | if (IS_ERR(ptr: rcpu->kthread)) |
440 | goto free_prog; |
441 | |
442 | /* Make sure kthread runs on a single CPU */ |
443 | kthread_bind(k: rcpu->kthread, cpu); |
444 | wake_up_process(tsk: rcpu->kthread); |
445 | |
446 | /* Make sure kthread has been running, so kthread_stop() will not |
447 | * stop the kthread prematurely and all pending frames or skbs |
448 | * will be handled by the kthread before kthread_stop() returns. |
449 | */ |
450 | wait_for_completion(&rcpu->kthread_running); |
451 | |
452 | return rcpu; |
453 | |
454 | free_prog: |
455 | if (rcpu->prog) |
456 | bpf_prog_put(prog: rcpu->prog); |
457 | free_ptr_ring: |
458 | ptr_ring_cleanup(r: rcpu->queue, NULL); |
459 | free_queue: |
460 | kfree(objp: rcpu->queue); |
461 | free_bulkq: |
462 | free_percpu(pdata: rcpu->bulkq); |
463 | free_rcu: |
464 | kfree(objp: rcpu); |
465 | return NULL; |
466 | } |
467 | |
468 | static void __cpu_map_entry_free(struct work_struct *work) |
469 | { |
470 | struct bpf_cpu_map_entry *rcpu; |
471 | |
472 | /* This cpu_map_entry have been disconnected from map and one |
473 | * RCU grace-period have elapsed. Thus, XDP cannot queue any |
474 | * new packets and cannot change/set flush_needed that can |
475 | * find this entry. |
476 | */ |
477 | rcpu = container_of(to_rcu_work(work), struct bpf_cpu_map_entry, free_work); |
478 | |
479 | /* kthread_stop will wake_up_process and wait for it to complete. |
480 | * cpu_map_kthread_run() makes sure the pointer ring is empty |
481 | * before exiting. |
482 | */ |
483 | kthread_stop(k: rcpu->kthread); |
484 | |
485 | if (rcpu->prog) |
486 | bpf_prog_put(prog: rcpu->prog); |
487 | /* The queue should be empty at this point */ |
488 | __cpu_map_ring_cleanup(ring: rcpu->queue); |
489 | ptr_ring_cleanup(r: rcpu->queue, NULL); |
490 | kfree(objp: rcpu->queue); |
491 | free_percpu(pdata: rcpu->bulkq); |
492 | kfree(objp: rcpu); |
493 | } |
494 | |
495 | /* After the xchg of the bpf_cpu_map_entry pointer, we need to make sure the old |
496 | * entry is no longer in use before freeing. We use queue_rcu_work() to call |
497 | * __cpu_map_entry_free() in a separate workqueue after waiting for an RCU grace |
498 | * period. This means that (a) all pending enqueue and flush operations have |
499 | * completed (because of the RCU callback), and (b) we are in a workqueue |
500 | * context where we can stop the kthread and wait for it to exit before freeing |
501 | * everything. |
502 | */ |
503 | static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap, |
504 | u32 key_cpu, struct bpf_cpu_map_entry *rcpu) |
505 | { |
506 | struct bpf_cpu_map_entry *old_rcpu; |
507 | |
508 | old_rcpu = unrcu_pointer(xchg(&cmap->cpu_map[key_cpu], RCU_INITIALIZER(rcpu))); |
509 | if (old_rcpu) { |
510 | INIT_RCU_WORK(&old_rcpu->free_work, __cpu_map_entry_free); |
511 | queue_rcu_work(wq: system_wq, rwork: &old_rcpu->free_work); |
512 | } |
513 | } |
514 | |
515 | static long cpu_map_delete_elem(struct bpf_map *map, void *key) |
516 | { |
517 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); |
518 | u32 key_cpu = *(u32 *)key; |
519 | |
520 | if (key_cpu >= map->max_entries) |
521 | return -EINVAL; |
522 | |
523 | /* notice caller map_delete_elem() uses rcu_read_lock() */ |
524 | __cpu_map_entry_replace(cmap, key_cpu, NULL); |
525 | return 0; |
526 | } |
527 | |
528 | static long cpu_map_update_elem(struct bpf_map *map, void *key, void *value, |
529 | u64 map_flags) |
530 | { |
531 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); |
532 | struct bpf_cpumap_val cpumap_value = {}; |
533 | struct bpf_cpu_map_entry *rcpu; |
534 | /* Array index key correspond to CPU number */ |
535 | u32 key_cpu = *(u32 *)key; |
536 | |
537 | memcpy(&cpumap_value, value, map->value_size); |
538 | |
539 | if (unlikely(map_flags > BPF_EXIST)) |
540 | return -EINVAL; |
541 | if (unlikely(key_cpu >= cmap->map.max_entries)) |
542 | return -E2BIG; |
543 | if (unlikely(map_flags == BPF_NOEXIST)) |
544 | return -EEXIST; |
545 | if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */ |
546 | return -EOVERFLOW; |
547 | |
548 | /* Make sure CPU is a valid possible cpu */ |
549 | if (key_cpu >= nr_cpumask_bits || !cpu_possible(cpu: key_cpu)) |
550 | return -ENODEV; |
551 | |
552 | if (cpumap_value.qsize == 0) { |
553 | rcpu = NULL; /* Same as deleting */ |
554 | } else { |
555 | /* Updating qsize cause re-allocation of bpf_cpu_map_entry */ |
556 | rcpu = __cpu_map_entry_alloc(map, value: &cpumap_value, cpu: key_cpu); |
557 | if (!rcpu) |
558 | return -ENOMEM; |
559 | } |
560 | rcu_read_lock(); |
561 | __cpu_map_entry_replace(cmap, key_cpu, rcpu); |
562 | rcu_read_unlock(); |
563 | return 0; |
564 | } |
565 | |
566 | static void cpu_map_free(struct bpf_map *map) |
567 | { |
568 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); |
569 | u32 i; |
570 | |
571 | /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0, |
572 | * so the bpf programs (can be more than one that used this map) were |
573 | * disconnected from events. Wait for outstanding critical sections in |
574 | * these programs to complete. synchronize_rcu() below not only |
575 | * guarantees no further "XDP/bpf-side" reads against |
576 | * bpf_cpu_map->cpu_map, but also ensure pending flush operations |
577 | * (if any) are completed. |
578 | */ |
579 | synchronize_rcu(); |
580 | |
581 | /* The only possible user of bpf_cpu_map_entry is |
582 | * cpu_map_kthread_run(). |
583 | */ |
584 | for (i = 0; i < cmap->map.max_entries; i++) { |
585 | struct bpf_cpu_map_entry *rcpu; |
586 | |
587 | rcpu = rcu_dereference_raw(cmap->cpu_map[i]); |
588 | if (!rcpu) |
589 | continue; |
590 | |
591 | /* Stop kthread and cleanup entry directly */ |
592 | __cpu_map_entry_free(work: &rcpu->free_work.work); |
593 | } |
594 | bpf_map_area_free(base: cmap->cpu_map); |
595 | bpf_map_area_free(base: cmap); |
596 | } |
597 | |
598 | /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or |
599 | * by local_bh_disable() (from XDP calls inside NAPI). The |
600 | * rcu_read_lock_bh_held() below makes lockdep accept both. |
601 | */ |
602 | static void *__cpu_map_lookup_elem(struct bpf_map *map, u32 key) |
603 | { |
604 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); |
605 | struct bpf_cpu_map_entry *rcpu; |
606 | |
607 | if (key >= map->max_entries) |
608 | return NULL; |
609 | |
610 | rcpu = rcu_dereference_check(cmap->cpu_map[key], |
611 | rcu_read_lock_bh_held()); |
612 | return rcpu; |
613 | } |
614 | |
615 | static void *cpu_map_lookup_elem(struct bpf_map *map, void *key) |
616 | { |
617 | struct bpf_cpu_map_entry *rcpu = |
618 | __cpu_map_lookup_elem(map, key: *(u32 *)key); |
619 | |
620 | return rcpu ? &rcpu->value : NULL; |
621 | } |
622 | |
623 | static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key) |
624 | { |
625 | struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map); |
626 | u32 index = key ? *(u32 *)key : U32_MAX; |
627 | u32 *next = next_key; |
628 | |
629 | if (index >= cmap->map.max_entries) { |
630 | *next = 0; |
631 | return 0; |
632 | } |
633 | |
634 | if (index == cmap->map.max_entries - 1) |
635 | return -ENOENT; |
636 | *next = index + 1; |
637 | return 0; |
638 | } |
639 | |
640 | static long cpu_map_redirect(struct bpf_map *map, u64 index, u64 flags) |
641 | { |
642 | return __bpf_xdp_redirect_map(map, index, flags, flag_mask: 0, |
643 | lookup_elem: __cpu_map_lookup_elem); |
644 | } |
645 | |
646 | static u64 cpu_map_mem_usage(const struct bpf_map *map) |
647 | { |
648 | u64 usage = sizeof(struct bpf_cpu_map); |
649 | |
650 | /* Currently the dynamically allocated elements are not counted */ |
651 | usage += (u64)map->max_entries * sizeof(struct bpf_cpu_map_entry *); |
652 | return usage; |
653 | } |
654 | |
655 | BTF_ID_LIST_SINGLE(cpu_map_btf_ids, struct, bpf_cpu_map) |
656 | const struct bpf_map_ops cpu_map_ops = { |
657 | .map_meta_equal = bpf_map_meta_equal, |
658 | .map_alloc = cpu_map_alloc, |
659 | .map_free = cpu_map_free, |
660 | .map_delete_elem = cpu_map_delete_elem, |
661 | .map_update_elem = cpu_map_update_elem, |
662 | .map_lookup_elem = cpu_map_lookup_elem, |
663 | .map_get_next_key = cpu_map_get_next_key, |
664 | .map_check_btf = map_check_no_btf, |
665 | .map_mem_usage = cpu_map_mem_usage, |
666 | .map_btf_id = &cpu_map_btf_ids[0], |
667 | .map_redirect = cpu_map_redirect, |
668 | }; |
669 | |
670 | static void bq_flush_to_queue(struct xdp_bulk_queue *bq) |
671 | { |
672 | struct bpf_cpu_map_entry *rcpu = bq->obj; |
673 | unsigned int processed = 0, drops = 0; |
674 | const int to_cpu = rcpu->cpu; |
675 | struct ptr_ring *q; |
676 | int i; |
677 | |
678 | if (unlikely(!bq->count)) |
679 | return; |
680 | |
681 | q = rcpu->queue; |
682 | spin_lock(lock: &q->producer_lock); |
683 | |
684 | for (i = 0; i < bq->count; i++) { |
685 | struct xdp_frame *xdpf = bq->q[i]; |
686 | int err; |
687 | |
688 | err = __ptr_ring_produce(r: q, ptr: xdpf); |
689 | if (err) { |
690 | drops++; |
691 | xdp_return_frame_rx_napi(xdpf); |
692 | } |
693 | processed++; |
694 | } |
695 | bq->count = 0; |
696 | spin_unlock(lock: &q->producer_lock); |
697 | |
698 | __list_del_clearprev(entry: &bq->flush_node); |
699 | |
700 | /* Feedback loop via tracepoints */ |
701 | trace_xdp_cpumap_enqueue(map_id: rcpu->map_id, processed, drops, to_cpu); |
702 | } |
703 | |
704 | /* Runs under RCU-read-side, plus in softirq under NAPI protection. |
705 | * Thus, safe percpu variable access. |
706 | */ |
707 | static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf) |
708 | { |
709 | struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list); |
710 | struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq); |
711 | |
712 | if (unlikely(bq->count == CPU_MAP_BULK_SIZE)) |
713 | bq_flush_to_queue(bq); |
714 | |
715 | /* Notice, xdp_buff/page MUST be queued here, long enough for |
716 | * driver to code invoking us to finished, due to driver |
717 | * (e.g. ixgbe) recycle tricks based on page-refcnt. |
718 | * |
719 | * Thus, incoming xdp_frame is always queued here (else we race |
720 | * with another CPU on page-refcnt and remaining driver code). |
721 | * Queue time is very short, as driver will invoke flush |
722 | * operation, when completing napi->poll call. |
723 | */ |
724 | bq->q[bq->count++] = xdpf; |
725 | |
726 | if (!bq->flush_node.prev) |
727 | list_add(new: &bq->flush_node, head: flush_list); |
728 | } |
729 | |
730 | int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf, |
731 | struct net_device *dev_rx) |
732 | { |
733 | /* Info needed when constructing SKB on remote CPU */ |
734 | xdpf->dev_rx = dev_rx; |
735 | |
736 | bq_enqueue(rcpu, xdpf); |
737 | return 0; |
738 | } |
739 | |
740 | int cpu_map_generic_redirect(struct bpf_cpu_map_entry *rcpu, |
741 | struct sk_buff *skb) |
742 | { |
743 | int ret; |
744 | |
745 | __skb_pull(skb, len: skb->mac_len); |
746 | skb_set_redirected(skb, from_ingress: false); |
747 | __ptr_set_bit(0, &skb); |
748 | |
749 | ret = ptr_ring_produce(r: rcpu->queue, ptr: skb); |
750 | if (ret < 0) |
751 | goto trace; |
752 | |
753 | wake_up_process(tsk: rcpu->kthread); |
754 | trace: |
755 | trace_xdp_cpumap_enqueue(map_id: rcpu->map_id, processed: !ret, drops: !!ret, to_cpu: rcpu->cpu); |
756 | return ret; |
757 | } |
758 | |
759 | void __cpu_map_flush(void) |
760 | { |
761 | struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list); |
762 | struct xdp_bulk_queue *bq, *tmp; |
763 | |
764 | list_for_each_entry_safe(bq, tmp, flush_list, flush_node) { |
765 | bq_flush_to_queue(bq); |
766 | |
767 | /* If already running, costs spin_lock_irqsave + smb_mb */ |
768 | wake_up_process(tsk: bq->obj->kthread); |
769 | } |
770 | } |
771 | |
772 | #ifdef CONFIG_DEBUG_NET |
773 | bool cpu_map_check_flush(void) |
774 | { |
775 | if (list_empty(this_cpu_ptr(&cpu_map_flush_list))) |
776 | return false; |
777 | __cpu_map_flush(); |
778 | return true; |
779 | } |
780 | #endif |
781 | |
782 | static int __init cpu_map_init(void) |
783 | { |
784 | int cpu; |
785 | |
786 | for_each_possible_cpu(cpu) |
787 | INIT_LIST_HEAD(list: &per_cpu(cpu_map_flush_list, cpu)); |
788 | return 0; |
789 | } |
790 | |
791 | subsys_initcall(cpu_map_init); |
792 | |