1 // SPDX-License-Identifier: GPL-2.0-only
4 * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
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'.
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.
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.
22 #include <linux/bitops.h>
23 #include <linux/bpf.h>
24 #include <linux/filter.h>
25 #include <linux/ptr_ring.h>
28 #include <linux/sched.h>
29 #include <linux/workqueue.h>
30 #include <linux/kthread.h>
31 #include <linux/capability.h>
32 #include <trace/events/xdp.h>
33 #include <linux/btf_ids.h>
35 #include <linux/netdevice.h> /* netif_receive_skb_list */
36 #include <linux/etherdevice.h> /* eth_type_trans */
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.
45 #define CPU_MAP_BULK_SIZE 8 /* 8 == one cacheline on 64-bit archs */
46 struct bpf_cpu_map_entry;
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;
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 */
61 /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
62 struct xdp_bulk_queue __percpu *bulkq;
64 struct bpf_cpu_map *cmap;
66 /* Queue with potential multi-producers, and single-consumer kthread */
67 struct ptr_ring *queue;
68 struct task_struct *kthread;
70 struct bpf_cpumap_val value;
71 struct bpf_prog *prog;
73 atomic_t refcnt; /* Control when this struct can be free'ed */
76 struct work_struct kthread_stop_wq;
81 /* Below members specific for map type */
82 struct bpf_cpu_map_entry __rcu **cpu_map;
85 static DEFINE_PER_CPU(struct list_head, cpu_map_flush_list);
87 static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
89 u32 value_size = attr->value_size;
90 struct bpf_cpu_map *cmap;
93 return ERR_PTR(-EPERM);
95 /* check sanity of attributes */
96 if (attr->max_entries == 0 || attr->key_size != 4 ||
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)
100 return ERR_PTR(-EINVAL);
102 /* Pre-limit array size based on NR_CPUS, not final CPU check */
103 if (attr->max_entries > NR_CPUS)
104 return ERR_PTR(-E2BIG);
106 cmap = bpf_map_area_alloc(sizeof(*cmap), NUMA_NO_NODE);
108 return ERR_PTR(-ENOMEM);
110 bpf_map_init_from_attr(&cmap->map, attr);
112 /* Alloc array for possible remote "destination" CPUs */
113 cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
114 sizeof(struct bpf_cpu_map_entry *),
115 cmap->map.numa_node);
116 if (!cmap->cpu_map) {
117 bpf_map_area_free(cmap);
118 return ERR_PTR(-ENOMEM);
124 static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
126 atomic_inc(&rcpu->refcnt);
129 /* called from workqueue, to workaround syscall using preempt_disable */
130 static void cpu_map_kthread_stop(struct work_struct *work)
132 struct bpf_cpu_map_entry *rcpu;
134 rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
136 /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
137 * as it waits until all in-flight call_rcu() callbacks complete.
141 /* kthread_stop will wake_up_process and wait for it to complete */
142 kthread_stop(rcpu->kthread);
145 static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
147 /* The tear-down procedure should have made sure that queue is
148 * empty. See __cpu_map_entry_replace() and work-queue
149 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
150 * gracefully and warn once.
152 struct xdp_frame *xdpf;
154 while ((xdpf = ptr_ring_consume(ring)))
155 if (WARN_ON_ONCE(xdpf))
156 xdp_return_frame(xdpf);
159 static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
161 if (atomic_dec_and_test(&rcpu->refcnt)) {
163 bpf_prog_put(rcpu->prog);
164 /* The queue should be empty at this point */
165 __cpu_map_ring_cleanup(rcpu->queue);
166 ptr_ring_cleanup(rcpu->queue, NULL);
172 static void cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
173 struct list_head *listp,
174 struct xdp_cpumap_stats *stats)
176 struct sk_buff *skb, *tmp;
181 list_for_each_entry_safe(skb, tmp, listp, list) {
182 act = bpf_prog_run_generic_xdp(skb, &xdp, rcpu->prog);
187 skb_list_del_init(skb);
188 err = xdp_do_generic_redirect(skb->dev, skb, &xdp,
198 bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
201 trace_xdp_exception(skb->dev, rcpu->prog, act);
204 skb_list_del_init(skb);
212 static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
213 void **frames, int n,
214 struct xdp_cpumap_stats *stats)
216 struct xdp_rxq_info rxq;
220 xdp_set_return_frame_no_direct();
223 for (i = 0; i < n; i++) {
224 struct xdp_frame *xdpf = frames[i];
228 rxq.dev = xdpf->dev_rx;
230 /* TODO: report queue_index to xdp_rxq_info */
232 xdp_convert_frame_to_buff(xdpf, &xdp);
234 act = bpf_prog_run_xdp(rcpu->prog, &xdp);
237 err = xdp_update_frame_from_buff(&xdp, xdpf);
239 xdp_return_frame(xdpf);
242 frames[nframes++] = xdpf;
247 err = xdp_do_redirect(xdpf->dev_rx, &xdp,
250 xdp_return_frame(xdpf);
257 bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
260 xdp_return_frame(xdpf);
266 xdp_clear_return_frame_no_direct();
271 #define CPUMAP_BATCH 8
273 static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
274 int xdp_n, struct xdp_cpumap_stats *stats,
275 struct list_head *list)
284 nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, xdp_n, stats);
289 if (unlikely(!list_empty(list)))
290 cpu_map_bpf_prog_run_skb(rcpu, list, stats);
292 rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
298 static int cpu_map_kthread_run(void *data)
300 struct bpf_cpu_map_entry *rcpu = data;
302 set_current_state(TASK_INTERRUPTIBLE);
304 /* When kthread gives stop order, then rcpu have been disconnected
305 * from map, thus no new packets can enter. Remaining in-flight
306 * per CPU stored packets are flushed to this queue. Wait honoring
307 * kthread_stop signal until queue is empty.
309 while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
310 struct xdp_cpumap_stats stats = {}; /* zero stats */
311 unsigned int kmem_alloc_drops = 0, sched = 0;
312 gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
313 int i, n, m, nframes, xdp_n;
314 void *frames[CPUMAP_BATCH];
315 void *skbs[CPUMAP_BATCH];
318 /* Release CPU reschedule checks */
319 if (__ptr_ring_empty(rcpu->queue)) {
320 set_current_state(TASK_INTERRUPTIBLE);
321 /* Recheck to avoid lost wake-up */
322 if (__ptr_ring_empty(rcpu->queue)) {
326 __set_current_state(TASK_RUNNING);
329 sched = cond_resched();
333 * The bpf_cpu_map_entry is single consumer, with this
334 * kthread CPU pinned. Lockless access to ptr_ring
335 * consume side valid as no-resize allowed of queue.
337 n = __ptr_ring_consume_batched(rcpu->queue, frames,
339 for (i = 0, xdp_n = 0; i < n; i++) {
343 if (unlikely(__ptr_test_bit(0, &f))) {
344 struct sk_buff *skb = f;
346 __ptr_clear_bit(0, &skb);
347 list_add_tail(&skb->list, &list);
352 page = virt_to_page(f);
354 /* Bring struct page memory area to curr CPU. Read by
355 * build_skb_around via page_is_pfmemalloc(), and when
356 * freed written by page_frag_free call.
361 /* Support running another XDP prog on this CPU */
362 nframes = cpu_map_bpf_prog_run(rcpu, frames, xdp_n, &stats, &list);
364 m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, nframes, skbs);
365 if (unlikely(m == 0)) {
366 for (i = 0; i < nframes; i++)
367 skbs[i] = NULL; /* effect: xdp_return_frame */
368 kmem_alloc_drops += nframes;
373 for (i = 0; i < nframes; i++) {
374 struct xdp_frame *xdpf = frames[i];
375 struct sk_buff *skb = skbs[i];
377 skb = __xdp_build_skb_from_frame(xdpf, skb,
380 xdp_return_frame(xdpf);
384 list_add_tail(&skb->list, &list);
386 netif_receive_skb_list(&list);
388 /* Feedback loop via tracepoint */
389 trace_xdp_cpumap_kthread(rcpu->map_id, n, kmem_alloc_drops,
392 local_bh_enable(); /* resched point, may call do_softirq() */
394 __set_current_state(TASK_RUNNING);
396 put_cpu_map_entry(rcpu);
400 static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu,
401 struct bpf_map *map, int fd)
403 struct bpf_prog *prog;
405 prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
407 return PTR_ERR(prog);
409 if (prog->expected_attach_type != BPF_XDP_CPUMAP ||
410 !bpf_prog_map_compatible(map, prog)) {
415 rcpu->value.bpf_prog.id = prog->aux->id;
421 static struct bpf_cpu_map_entry *
422 __cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value,
425 int numa, err, i, fd = value->bpf_prog.fd;
426 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
427 struct bpf_cpu_map_entry *rcpu;
428 struct xdp_bulk_queue *bq;
430 /* Have map->numa_node, but choose node of redirect target CPU */
431 numa = cpu_to_node(cpu);
433 rcpu = bpf_map_kmalloc_node(map, sizeof(*rcpu), gfp | __GFP_ZERO, numa);
437 /* Alloc percpu bulkq */
438 rcpu->bulkq = bpf_map_alloc_percpu(map, sizeof(*rcpu->bulkq),
439 sizeof(void *), gfp);
443 for_each_possible_cpu(i) {
444 bq = per_cpu_ptr(rcpu->bulkq, i);
449 rcpu->queue = bpf_map_kmalloc_node(map, sizeof(*rcpu->queue), gfp,
454 err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
459 rcpu->map_id = map->id;
460 rcpu->value.qsize = value->qsize;
462 if (fd > 0 && __cpu_map_load_bpf_program(rcpu, map, fd))
466 rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
467 "cpumap/%d/map:%d", cpu,
469 if (IS_ERR(rcpu->kthread))
472 get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
473 get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
475 /* Make sure kthread runs on a single CPU */
476 kthread_bind(rcpu->kthread, cpu);
477 wake_up_process(rcpu->kthread);
483 bpf_prog_put(rcpu->prog);
485 ptr_ring_cleanup(rcpu->queue, NULL);
489 free_percpu(rcpu->bulkq);
495 static void __cpu_map_entry_free(struct rcu_head *rcu)
497 struct bpf_cpu_map_entry *rcpu;
499 /* This cpu_map_entry have been disconnected from map and one
500 * RCU grace-period have elapsed. Thus, XDP cannot queue any
501 * new packets and cannot change/set flush_needed that can
504 rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
506 free_percpu(rcpu->bulkq);
507 /* Cannot kthread_stop() here, last put free rcpu resources */
508 put_cpu_map_entry(rcpu);
511 /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
512 * ensure any driver rcu critical sections have completed, but this
513 * does not guarantee a flush has happened yet. Because driver side
514 * rcu_read_lock/unlock only protects the running XDP program. The
515 * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
516 * pending flush op doesn't fail.
518 * The bpf_cpu_map_entry is still used by the kthread, and there can
519 * still be pending packets (in queue and percpu bulkq). A refcnt
520 * makes sure to last user (kthread_stop vs. call_rcu) free memory
523 * The rcu callback __cpu_map_entry_free flush remaining packets in
524 * percpu bulkq to queue. Due to caller map_delete_elem() disable
525 * preemption, cannot call kthread_stop() to make sure queue is empty.
526 * Instead a work_queue is started for stopping kthread,
527 * cpu_map_kthread_stop, which waits for an RCU grace period before
528 * stopping kthread, emptying the queue.
530 static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
531 u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
533 struct bpf_cpu_map_entry *old_rcpu;
535 old_rcpu = unrcu_pointer(xchg(&cmap->cpu_map[key_cpu], RCU_INITIALIZER(rcpu)));
537 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
538 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
539 schedule_work(&old_rcpu->kthread_stop_wq);
543 static int cpu_map_delete_elem(struct bpf_map *map, void *key)
545 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
546 u32 key_cpu = *(u32 *)key;
548 if (key_cpu >= map->max_entries)
551 /* notice caller map_delete_elem() use preempt_disable() */
552 __cpu_map_entry_replace(cmap, key_cpu, NULL);
556 static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
559 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
560 struct bpf_cpumap_val cpumap_value = {};
561 struct bpf_cpu_map_entry *rcpu;
562 /* Array index key correspond to CPU number */
563 u32 key_cpu = *(u32 *)key;
565 memcpy(&cpumap_value, value, map->value_size);
567 if (unlikely(map_flags > BPF_EXIST))
569 if (unlikely(key_cpu >= cmap->map.max_entries))
571 if (unlikely(map_flags == BPF_NOEXIST))
573 if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
576 /* Make sure CPU is a valid possible cpu */
577 if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
580 if (cpumap_value.qsize == 0) {
581 rcpu = NULL; /* Same as deleting */
583 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
584 rcpu = __cpu_map_entry_alloc(map, &cpumap_value, key_cpu);
590 __cpu_map_entry_replace(cmap, key_cpu, rcpu);
595 static void cpu_map_free(struct bpf_map *map)
597 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
600 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
601 * so the bpf programs (can be more than one that used this map) were
602 * disconnected from events. Wait for outstanding critical sections in
603 * these programs to complete. The rcu critical section only guarantees
604 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
605 * It does __not__ ensure pending flush operations (if any) are
611 /* For cpu_map the remote CPUs can still be using the entries
612 * (struct bpf_cpu_map_entry).
614 for (i = 0; i < cmap->map.max_entries; i++) {
615 struct bpf_cpu_map_entry *rcpu;
617 rcpu = rcu_dereference_raw(cmap->cpu_map[i]);
621 /* bq flush and cleanup happens after RCU grace-period */
622 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
624 bpf_map_area_free(cmap->cpu_map);
625 bpf_map_area_free(cmap);
628 /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
629 * by local_bh_disable() (from XDP calls inside NAPI). The
630 * rcu_read_lock_bh_held() below makes lockdep accept both.
632 static void *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
634 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
635 struct bpf_cpu_map_entry *rcpu;
637 if (key >= map->max_entries)
640 rcpu = rcu_dereference_check(cmap->cpu_map[key],
641 rcu_read_lock_bh_held());
645 static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
647 struct bpf_cpu_map_entry *rcpu =
648 __cpu_map_lookup_elem(map, *(u32 *)key);
650 return rcpu ? &rcpu->value : NULL;
653 static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
655 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
656 u32 index = key ? *(u32 *)key : U32_MAX;
657 u32 *next = next_key;
659 if (index >= cmap->map.max_entries) {
664 if (index == cmap->map.max_entries - 1)
670 static int cpu_map_redirect(struct bpf_map *map, u64 index, u64 flags)
672 return __bpf_xdp_redirect_map(map, index, flags, 0,
673 __cpu_map_lookup_elem);
676 BTF_ID_LIST_SINGLE(cpu_map_btf_ids, struct, bpf_cpu_map)
677 const struct bpf_map_ops cpu_map_ops = {
678 .map_meta_equal = bpf_map_meta_equal,
679 .map_alloc = cpu_map_alloc,
680 .map_free = cpu_map_free,
681 .map_delete_elem = cpu_map_delete_elem,
682 .map_update_elem = cpu_map_update_elem,
683 .map_lookup_elem = cpu_map_lookup_elem,
684 .map_get_next_key = cpu_map_get_next_key,
685 .map_check_btf = map_check_no_btf,
686 .map_btf_id = &cpu_map_btf_ids[0],
687 .map_redirect = cpu_map_redirect,
690 static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
692 struct bpf_cpu_map_entry *rcpu = bq->obj;
693 unsigned int processed = 0, drops = 0;
694 const int to_cpu = rcpu->cpu;
698 if (unlikely(!bq->count))
702 spin_lock(&q->producer_lock);
704 for (i = 0; i < bq->count; i++) {
705 struct xdp_frame *xdpf = bq->q[i];
708 err = __ptr_ring_produce(q, xdpf);
711 xdp_return_frame_rx_napi(xdpf);
716 spin_unlock(&q->producer_lock);
718 __list_del_clearprev(&bq->flush_node);
720 /* Feedback loop via tracepoints */
721 trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
724 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
725 * Thus, safe percpu variable access.
727 static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
729 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
730 struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
732 if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
733 bq_flush_to_queue(bq);
735 /* Notice, xdp_buff/page MUST be queued here, long enough for
736 * driver to code invoking us to finished, due to driver
737 * (e.g. ixgbe) recycle tricks based on page-refcnt.
739 * Thus, incoming xdp_frame is always queued here (else we race
740 * with another CPU on page-refcnt and remaining driver code).
741 * Queue time is very short, as driver will invoke flush
742 * operation, when completing napi->poll call.
744 bq->q[bq->count++] = xdpf;
746 if (!bq->flush_node.prev)
747 list_add(&bq->flush_node, flush_list);
750 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf,
751 struct net_device *dev_rx)
753 /* Info needed when constructing SKB on remote CPU */
754 xdpf->dev_rx = dev_rx;
756 bq_enqueue(rcpu, xdpf);
760 int cpu_map_generic_redirect(struct bpf_cpu_map_entry *rcpu,
765 __skb_pull(skb, skb->mac_len);
766 skb_set_redirected(skb, false);
767 __ptr_set_bit(0, &skb);
769 ret = ptr_ring_produce(rcpu->queue, skb);
773 wake_up_process(rcpu->kthread);
775 trace_xdp_cpumap_enqueue(rcpu->map_id, !ret, !!ret, rcpu->cpu);
779 void __cpu_map_flush(void)
781 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
782 struct xdp_bulk_queue *bq, *tmp;
784 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
785 bq_flush_to_queue(bq);
787 /* If already running, costs spin_lock_irqsave + smb_mb */
788 wake_up_process(bq->obj->kthread);
792 static int __init cpu_map_init(void)
796 for_each_possible_cpu(cpu)
797 INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu));
801 subsys_initcall(cpu_map_init);