1 // SPDX-License-Identifier: GPL-2.0-only
4 * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
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'.
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.
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.
19 #include <linux/bpf.h>
20 #include <linux/filter.h>
21 #include <linux/ptr_ring.h>
24 #include <linux/sched.h>
25 #include <linux/workqueue.h>
26 #include <linux/kthread.h>
27 #include <linux/capability.h>
28 #include <trace/events/xdp.h>
30 #include <linux/netdevice.h> /* netif_receive_skb_core */
31 #include <linux/etherdevice.h> /* eth_type_trans */
33 /* General idea: XDP packets getting XDP redirected to another CPU,
34 * will maximum be stored/queued for one driver ->poll() call. It is
35 * guaranteed that queueing the frame and the flush operation happen on
36 * same CPU. Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
37 * which queue in bpf_cpu_map_entry contains packets.
40 #define CPU_MAP_BULK_SIZE 8 /* 8 == one cacheline on 64-bit archs */
41 struct bpf_cpu_map_entry;
44 struct xdp_bulk_queue {
45 void *q[CPU_MAP_BULK_SIZE];
46 struct list_head flush_node;
47 struct bpf_cpu_map_entry *obj;
51 /* Struct for every remote "destination" CPU in map */
52 struct bpf_cpu_map_entry {
53 u32 cpu; /* kthread CPU and map index */
54 int map_id; /* Back reference to map */
56 /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
57 struct xdp_bulk_queue __percpu *bulkq;
59 struct bpf_cpu_map *cmap;
61 /* Queue with potential multi-producers, and single-consumer kthread */
62 struct ptr_ring *queue;
63 struct task_struct *kthread;
65 struct bpf_cpumap_val value;
66 struct bpf_prog *prog;
68 atomic_t refcnt; /* Control when this struct can be free'ed */
71 struct work_struct kthread_stop_wq;
76 /* Below members specific for map type */
77 struct bpf_cpu_map_entry **cpu_map;
80 static DEFINE_PER_CPU(struct list_head, cpu_map_flush_list);
82 static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
84 u32 value_size = attr->value_size;
85 struct bpf_cpu_map *cmap;
89 return ERR_PTR(-EPERM);
91 /* check sanity of attributes */
92 if (attr->max_entries == 0 || attr->key_size != 4 ||
93 (value_size != offsetofend(struct bpf_cpumap_val, qsize) &&
94 value_size != offsetofend(struct bpf_cpumap_val, bpf_prog.fd)) ||
95 attr->map_flags & ~BPF_F_NUMA_NODE)
96 return ERR_PTR(-EINVAL);
98 cmap = kzalloc(sizeof(*cmap), GFP_USER | __GFP_ACCOUNT);
100 return ERR_PTR(-ENOMEM);
102 bpf_map_init_from_attr(&cmap->map, attr);
104 /* Pre-limit array size based on NR_CPUS, not final CPU check */
105 if (cmap->map.max_entries > NR_CPUS) {
110 /* Alloc array for possible remote "destination" CPUs */
111 cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
112 sizeof(struct bpf_cpu_map_entry *),
113 cmap->map.numa_node);
123 static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
125 atomic_inc(&rcpu->refcnt);
128 /* called from workqueue, to workaround syscall using preempt_disable */
129 static void cpu_map_kthread_stop(struct work_struct *work)
131 struct bpf_cpu_map_entry *rcpu;
133 rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
135 /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
136 * as it waits until all in-flight call_rcu() callbacks complete.
140 /* kthread_stop will wake_up_process and wait for it to complete */
141 kthread_stop(rcpu->kthread);
144 static struct sk_buff *cpu_map_build_skb(struct xdp_frame *xdpf,
147 unsigned int hard_start_headroom;
148 unsigned int frame_size;
149 void *pkt_data_start;
151 /* Part of headroom was reserved to xdpf */
152 hard_start_headroom = sizeof(struct xdp_frame) + xdpf->headroom;
154 /* Memory size backing xdp_frame data already have reserved
155 * room for build_skb to place skb_shared_info in tailroom.
157 frame_size = xdpf->frame_sz;
159 pkt_data_start = xdpf->data - hard_start_headroom;
160 skb = build_skb_around(skb, pkt_data_start, frame_size);
164 skb_reserve(skb, hard_start_headroom);
165 __skb_put(skb, xdpf->len);
167 skb_metadata_set(skb, xdpf->metasize);
169 /* Essential SKB info: protocol and skb->dev */
170 skb->protocol = eth_type_trans(skb, xdpf->dev_rx);
172 /* Optional SKB info, currently missing:
173 * - HW checksum info (skb->ip_summed)
174 * - HW RX hash (skb_set_hash)
175 * - RX ring dev queue index (skb_record_rx_queue)
178 /* Until page_pool get SKB return path, release DMA here */
179 xdp_release_frame(xdpf);
181 /* Allow SKB to reuse area used by xdp_frame */
182 xdp_scrub_frame(xdpf);
187 static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
189 /* The tear-down procedure should have made sure that queue is
190 * empty. See __cpu_map_entry_replace() and work-queue
191 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
192 * gracefully and warn once.
194 struct xdp_frame *xdpf;
196 while ((xdpf = ptr_ring_consume(ring)))
197 if (WARN_ON_ONCE(xdpf))
198 xdp_return_frame(xdpf);
201 static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
203 if (atomic_dec_and_test(&rcpu->refcnt)) {
205 bpf_prog_put(rcpu->prog);
206 /* The queue should be empty at this point */
207 __cpu_map_ring_cleanup(rcpu->queue);
208 ptr_ring_cleanup(rcpu->queue, NULL);
214 static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
215 void **frames, int n,
216 struct xdp_cpumap_stats *stats)
218 struct xdp_rxq_info rxq;
227 xdp_set_return_frame_no_direct();
230 for (i = 0; i < n; i++) {
231 struct xdp_frame *xdpf = frames[i];
235 rxq.dev = xdpf->dev_rx;
237 /* TODO: report queue_index to xdp_rxq_info */
239 xdp_convert_frame_to_buff(xdpf, &xdp);
241 act = bpf_prog_run_xdp(rcpu->prog, &xdp);
244 err = xdp_update_frame_from_buff(&xdp, xdpf);
246 xdp_return_frame(xdpf);
249 frames[nframes++] = xdpf;
254 err = xdp_do_redirect(xdpf->dev_rx, &xdp,
257 xdp_return_frame(xdpf);
264 bpf_warn_invalid_xdp_action(act);
267 xdp_return_frame(xdpf);
276 xdp_clear_return_frame_no_direct();
278 rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
283 #define CPUMAP_BATCH 8
285 static int cpu_map_kthread_run(void *data)
287 struct bpf_cpu_map_entry *rcpu = data;
289 set_current_state(TASK_INTERRUPTIBLE);
291 /* When kthread gives stop order, then rcpu have been disconnected
292 * from map, thus no new packets can enter. Remaining in-flight
293 * per CPU stored packets are flushed to this queue. Wait honoring
294 * kthread_stop signal until queue is empty.
296 while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
297 struct xdp_cpumap_stats stats = {}; /* zero stats */
298 gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
299 unsigned int drops = 0, sched = 0;
300 void *frames[CPUMAP_BATCH];
301 void *skbs[CPUMAP_BATCH];
302 int i, n, m, nframes;
304 /* Release CPU reschedule checks */
305 if (__ptr_ring_empty(rcpu->queue)) {
306 set_current_state(TASK_INTERRUPTIBLE);
307 /* Recheck to avoid lost wake-up */
308 if (__ptr_ring_empty(rcpu->queue)) {
312 __set_current_state(TASK_RUNNING);
315 sched = cond_resched();
319 * The bpf_cpu_map_entry is single consumer, with this
320 * kthread CPU pinned. Lockless access to ptr_ring
321 * consume side valid as no-resize allowed of queue.
323 n = __ptr_ring_consume_batched(rcpu->queue, frames,
325 for (i = 0; i < n; i++) {
327 struct page *page = virt_to_page(f);
329 /* Bring struct page memory area to curr CPU. Read by
330 * build_skb_around via page_is_pfmemalloc(), and when
331 * freed written by page_frag_free call.
336 /* Support running another XDP prog on this CPU */
337 nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, n, &stats);
339 m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, nframes, skbs);
340 if (unlikely(m == 0)) {
341 for (i = 0; i < nframes; i++)
342 skbs[i] = NULL; /* effect: xdp_return_frame */
348 for (i = 0; i < nframes; i++) {
349 struct xdp_frame *xdpf = frames[i];
350 struct sk_buff *skb = skbs[i];
353 skb = cpu_map_build_skb(xdpf, skb);
355 xdp_return_frame(xdpf);
359 /* Inject into network stack */
360 ret = netif_receive_skb_core(skb);
361 if (ret == NET_RX_DROP)
364 /* Feedback loop via tracepoint */
365 trace_xdp_cpumap_kthread(rcpu->map_id, n, drops, sched, &stats);
367 local_bh_enable(); /* resched point, may call do_softirq() */
369 __set_current_state(TASK_RUNNING);
371 put_cpu_map_entry(rcpu);
375 bool cpu_map_prog_allowed(struct bpf_map *map)
377 return map->map_type == BPF_MAP_TYPE_CPUMAP &&
378 map->value_size != offsetofend(struct bpf_cpumap_val, qsize);
381 static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu, int fd)
383 struct bpf_prog *prog;
385 prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
387 return PTR_ERR(prog);
389 if (prog->expected_attach_type != BPF_XDP_CPUMAP) {
394 rcpu->value.bpf_prog.id = prog->aux->id;
400 static struct bpf_cpu_map_entry *
401 __cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value,
404 int numa, err, i, fd = value->bpf_prog.fd;
405 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
406 struct bpf_cpu_map_entry *rcpu;
407 struct xdp_bulk_queue *bq;
409 /* Have map->numa_node, but choose node of redirect target CPU */
410 numa = cpu_to_node(cpu);
412 rcpu = bpf_map_kmalloc_node(map, sizeof(*rcpu), gfp | __GFP_ZERO, numa);
416 /* Alloc percpu bulkq */
417 rcpu->bulkq = bpf_map_alloc_percpu(map, sizeof(*rcpu->bulkq),
418 sizeof(void *), gfp);
422 for_each_possible_cpu(i) {
423 bq = per_cpu_ptr(rcpu->bulkq, i);
428 rcpu->queue = bpf_map_kmalloc_node(map, sizeof(*rcpu->queue), gfp,
433 err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
438 rcpu->map_id = map->id;
439 rcpu->value.qsize = value->qsize;
441 if (fd > 0 && __cpu_map_load_bpf_program(rcpu, fd))
445 rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
446 "cpumap/%d/map:%d", cpu,
448 if (IS_ERR(rcpu->kthread))
451 get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
452 get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
454 /* Make sure kthread runs on a single CPU */
455 kthread_bind(rcpu->kthread, cpu);
456 wake_up_process(rcpu->kthread);
462 bpf_prog_put(rcpu->prog);
464 ptr_ring_cleanup(rcpu->queue, NULL);
468 free_percpu(rcpu->bulkq);
474 static void __cpu_map_entry_free(struct rcu_head *rcu)
476 struct bpf_cpu_map_entry *rcpu;
478 /* This cpu_map_entry have been disconnected from map and one
479 * RCU grace-period have elapsed. Thus, XDP cannot queue any
480 * new packets and cannot change/set flush_needed that can
483 rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
485 free_percpu(rcpu->bulkq);
486 /* Cannot kthread_stop() here, last put free rcpu resources */
487 put_cpu_map_entry(rcpu);
490 /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
491 * ensure any driver rcu critical sections have completed, but this
492 * does not guarantee a flush has happened yet. Because driver side
493 * rcu_read_lock/unlock only protects the running XDP program. The
494 * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
495 * pending flush op doesn't fail.
497 * The bpf_cpu_map_entry is still used by the kthread, and there can
498 * still be pending packets (in queue and percpu bulkq). A refcnt
499 * makes sure to last user (kthread_stop vs. call_rcu) free memory
502 * The rcu callback __cpu_map_entry_free flush remaining packets in
503 * percpu bulkq to queue. Due to caller map_delete_elem() disable
504 * preemption, cannot call kthread_stop() to make sure queue is empty.
505 * Instead a work_queue is started for stopping kthread,
506 * cpu_map_kthread_stop, which waits for an RCU grace period before
507 * stopping kthread, emptying the queue.
509 static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
510 u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
512 struct bpf_cpu_map_entry *old_rcpu;
514 old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu);
516 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
517 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
518 schedule_work(&old_rcpu->kthread_stop_wq);
522 static int cpu_map_delete_elem(struct bpf_map *map, void *key)
524 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
525 u32 key_cpu = *(u32 *)key;
527 if (key_cpu >= map->max_entries)
530 /* notice caller map_delete_elem() use preempt_disable() */
531 __cpu_map_entry_replace(cmap, key_cpu, NULL);
535 static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
538 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
539 struct bpf_cpumap_val cpumap_value = {};
540 struct bpf_cpu_map_entry *rcpu;
541 /* Array index key correspond to CPU number */
542 u32 key_cpu = *(u32 *)key;
544 memcpy(&cpumap_value, value, map->value_size);
546 if (unlikely(map_flags > BPF_EXIST))
548 if (unlikely(key_cpu >= cmap->map.max_entries))
550 if (unlikely(map_flags == BPF_NOEXIST))
552 if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
555 /* Make sure CPU is a valid possible cpu */
556 if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
559 if (cpumap_value.qsize == 0) {
560 rcpu = NULL; /* Same as deleting */
562 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
563 rcpu = __cpu_map_entry_alloc(map, &cpumap_value, key_cpu);
569 __cpu_map_entry_replace(cmap, key_cpu, rcpu);
574 static void cpu_map_free(struct bpf_map *map)
576 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
579 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
580 * so the bpf programs (can be more than one that used this map) were
581 * disconnected from events. Wait for outstanding critical sections in
582 * these programs to complete. The rcu critical section only guarantees
583 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
584 * It does __not__ ensure pending flush operations (if any) are
588 bpf_clear_redirect_map(map);
591 /* For cpu_map the remote CPUs can still be using the entries
592 * (struct bpf_cpu_map_entry).
594 for (i = 0; i < cmap->map.max_entries; i++) {
595 struct bpf_cpu_map_entry *rcpu;
597 rcpu = READ_ONCE(cmap->cpu_map[i]);
601 /* bq flush and cleanup happens after RCU grace-period */
602 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
604 bpf_map_area_free(cmap->cpu_map);
608 struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
610 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
611 struct bpf_cpu_map_entry *rcpu;
613 if (key >= map->max_entries)
616 rcpu = READ_ONCE(cmap->cpu_map[key]);
620 static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
622 struct bpf_cpu_map_entry *rcpu =
623 __cpu_map_lookup_elem(map, *(u32 *)key);
625 return rcpu ? &rcpu->value : NULL;
628 static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
630 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
631 u32 index = key ? *(u32 *)key : U32_MAX;
632 u32 *next = next_key;
634 if (index >= cmap->map.max_entries) {
639 if (index == cmap->map.max_entries - 1)
645 static int cpu_map_btf_id;
646 const struct bpf_map_ops cpu_map_ops = {
647 .map_meta_equal = bpf_map_meta_equal,
648 .map_alloc = cpu_map_alloc,
649 .map_free = cpu_map_free,
650 .map_delete_elem = cpu_map_delete_elem,
651 .map_update_elem = cpu_map_update_elem,
652 .map_lookup_elem = cpu_map_lookup_elem,
653 .map_get_next_key = cpu_map_get_next_key,
654 .map_check_btf = map_check_no_btf,
655 .map_btf_name = "bpf_cpu_map",
656 .map_btf_id = &cpu_map_btf_id,
659 static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
661 struct bpf_cpu_map_entry *rcpu = bq->obj;
662 unsigned int processed = 0, drops = 0;
663 const int to_cpu = rcpu->cpu;
667 if (unlikely(!bq->count))
671 spin_lock(&q->producer_lock);
673 for (i = 0; i < bq->count; i++) {
674 struct xdp_frame *xdpf = bq->q[i];
677 err = __ptr_ring_produce(q, xdpf);
680 xdp_return_frame_rx_napi(xdpf);
685 spin_unlock(&q->producer_lock);
687 __list_del_clearprev(&bq->flush_node);
689 /* Feedback loop via tracepoints */
690 trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
693 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
694 * Thus, safe percpu variable access.
696 static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
698 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
699 struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
701 if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
702 bq_flush_to_queue(bq);
704 /* Notice, xdp_buff/page MUST be queued here, long enough for
705 * driver to code invoking us to finished, due to driver
706 * (e.g. ixgbe) recycle tricks based on page-refcnt.
708 * Thus, incoming xdp_frame is always queued here (else we race
709 * with another CPU on page-refcnt and remaining driver code).
710 * Queue time is very short, as driver will invoke flush
711 * operation, when completing napi->poll call.
713 bq->q[bq->count++] = xdpf;
715 if (!bq->flush_node.prev)
716 list_add(&bq->flush_node, flush_list);
719 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
720 struct net_device *dev_rx)
722 struct xdp_frame *xdpf;
724 xdpf = xdp_convert_buff_to_frame(xdp);
728 /* Info needed when constructing SKB on remote CPU */
729 xdpf->dev_rx = dev_rx;
731 bq_enqueue(rcpu, xdpf);
735 void __cpu_map_flush(void)
737 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
738 struct xdp_bulk_queue *bq, *tmp;
740 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
741 bq_flush_to_queue(bq);
743 /* If already running, costs spin_lock_irqsave + smb_mb */
744 wake_up_process(bq->obj->kthread);
748 static int __init cpu_map_init(void)
752 for_each_possible_cpu(cpu)
753 INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu));
757 subsys_initcall(cpu_map_init);