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/completion.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;
77 struct completion kthread_running;
82 /* Below members specific for map type */
83 struct bpf_cpu_map_entry __rcu **cpu_map;
86 static DEFINE_PER_CPU(struct list_head, cpu_map_flush_list);
88 static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
90 u32 value_size = attr->value_size;
91 struct bpf_cpu_map *cmap;
93 /* check sanity of attributes */
94 if (attr->max_entries == 0 || attr->key_size != 4 ||
95 (value_size != offsetofend(struct bpf_cpumap_val, qsize) &&
96 value_size != offsetofend(struct bpf_cpumap_val, bpf_prog.fd)) ||
97 attr->map_flags & ~BPF_F_NUMA_NODE)
98 return ERR_PTR(-EINVAL);
100 /* Pre-limit array size based on NR_CPUS, not final CPU check */
101 if (attr->max_entries > NR_CPUS)
102 return ERR_PTR(-E2BIG);
104 cmap = bpf_map_area_alloc(sizeof(*cmap), NUMA_NO_NODE);
106 return ERR_PTR(-ENOMEM);
108 bpf_map_init_from_attr(&cmap->map, attr);
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);
114 if (!cmap->cpu_map) {
115 bpf_map_area_free(cmap);
116 return ERR_PTR(-ENOMEM);
122 static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
124 atomic_inc(&rcpu->refcnt);
127 static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
129 /* The tear-down procedure should have made sure that queue is
130 * empty. See __cpu_map_entry_replace() and work-queue
131 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
132 * gracefully and warn once.
136 while ((ptr = ptr_ring_consume(ring))) {
138 if (unlikely(__ptr_test_bit(0, &ptr))) {
139 __ptr_clear_bit(0, &ptr);
143 xdp_return_frame(ptr);
147 static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
149 if (atomic_dec_and_test(&rcpu->refcnt)) {
151 bpf_prog_put(rcpu->prog);
152 /* The queue should be empty at this point */
153 __cpu_map_ring_cleanup(rcpu->queue);
154 ptr_ring_cleanup(rcpu->queue, NULL);
160 /* called from workqueue, to workaround syscall using preempt_disable */
161 static void cpu_map_kthread_stop(struct work_struct *work)
163 struct bpf_cpu_map_entry *rcpu;
165 rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
167 /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
168 * as it waits until all in-flight call_rcu() callbacks complete.
172 /* kthread_stop will wake_up_process and wait for it to complete */
173 kthread_stop(rcpu->kthread);
176 static void cpu_map_bpf_prog_run_skb(struct bpf_cpu_map_entry *rcpu,
177 struct list_head *listp,
178 struct xdp_cpumap_stats *stats)
180 struct sk_buff *skb, *tmp;
185 list_for_each_entry_safe(skb, tmp, listp, list) {
186 act = bpf_prog_run_generic_xdp(skb, &xdp, rcpu->prog);
191 skb_list_del_init(skb);
192 err = xdp_do_generic_redirect(skb->dev, skb, &xdp,
202 bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
205 trace_xdp_exception(skb->dev, rcpu->prog, act);
208 skb_list_del_init(skb);
216 static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
217 void **frames, int n,
218 struct xdp_cpumap_stats *stats)
220 struct xdp_rxq_info rxq;
224 xdp_set_return_frame_no_direct();
227 for (i = 0; i < n; i++) {
228 struct xdp_frame *xdpf = frames[i];
232 rxq.dev = xdpf->dev_rx;
234 /* TODO: report queue_index to xdp_rxq_info */
236 xdp_convert_frame_to_buff(xdpf, &xdp);
238 act = bpf_prog_run_xdp(rcpu->prog, &xdp);
241 err = xdp_update_frame_from_buff(&xdp, xdpf);
243 xdp_return_frame(xdpf);
246 frames[nframes++] = xdpf;
251 err = xdp_do_redirect(xdpf->dev_rx, &xdp,
254 xdp_return_frame(xdpf);
261 bpf_warn_invalid_xdp_action(NULL, rcpu->prog, act);
264 xdp_return_frame(xdpf);
270 xdp_clear_return_frame_no_direct();
275 #define CPUMAP_BATCH 8
277 static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
278 int xdp_n, struct xdp_cpumap_stats *stats,
279 struct list_head *list)
288 nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, xdp_n, stats);
293 if (unlikely(!list_empty(list)))
294 cpu_map_bpf_prog_run_skb(rcpu, list, stats);
296 rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
301 static int cpu_map_kthread_run(void *data)
303 struct bpf_cpu_map_entry *rcpu = data;
305 complete(&rcpu->kthread_running);
306 set_current_state(TASK_INTERRUPTIBLE);
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.
313 while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
314 struct xdp_cpumap_stats stats = {}; /* zero stats */
315 unsigned int kmem_alloc_drops = 0, sched = 0;
316 gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
317 int i, n, m, nframes, xdp_n;
318 void *frames[CPUMAP_BATCH];
319 void *skbs[CPUMAP_BATCH];
322 /* Release CPU reschedule checks */
323 if (__ptr_ring_empty(rcpu->queue)) {
324 set_current_state(TASK_INTERRUPTIBLE);
325 /* Recheck to avoid lost wake-up */
326 if (__ptr_ring_empty(rcpu->queue)) {
330 __set_current_state(TASK_RUNNING);
333 sched = cond_resched();
337 * The bpf_cpu_map_entry is single consumer, with this
338 * kthread CPU pinned. Lockless access to ptr_ring
339 * consume side valid as no-resize allowed of queue.
341 n = __ptr_ring_consume_batched(rcpu->queue, frames,
343 for (i = 0, xdp_n = 0; i < n; i++) {
347 if (unlikely(__ptr_test_bit(0, &f))) {
348 struct sk_buff *skb = f;
350 __ptr_clear_bit(0, &skb);
351 list_add_tail(&skb->list, &list);
356 page = virt_to_page(f);
358 /* Bring struct page memory area to curr CPU. Read by
359 * build_skb_around via page_is_pfmemalloc(), and when
360 * freed written by page_frag_free call.
365 /* Support running another XDP prog on this CPU */
366 nframes = cpu_map_bpf_prog_run(rcpu, frames, xdp_n, &stats, &list);
368 m = kmem_cache_alloc_bulk(skbuff_cache, gfp, nframes, skbs);
369 if (unlikely(m == 0)) {
370 for (i = 0; i < nframes; i++)
371 skbs[i] = NULL; /* effect: xdp_return_frame */
372 kmem_alloc_drops += nframes;
377 for (i = 0; i < nframes; i++) {
378 struct xdp_frame *xdpf = frames[i];
379 struct sk_buff *skb = skbs[i];
381 skb = __xdp_build_skb_from_frame(xdpf, skb,
384 xdp_return_frame(xdpf);
388 list_add_tail(&skb->list, &list);
390 netif_receive_skb_list(&list);
392 /* Feedback loop via tracepoint */
393 trace_xdp_cpumap_kthread(rcpu->map_id, n, kmem_alloc_drops,
396 local_bh_enable(); /* resched point, may call do_softirq() */
398 __set_current_state(TASK_RUNNING);
400 put_cpu_map_entry(rcpu);
404 static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu,
405 struct bpf_map *map, int fd)
407 struct bpf_prog *prog;
409 prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
411 return PTR_ERR(prog);
413 if (prog->expected_attach_type != BPF_XDP_CPUMAP ||
414 !bpf_prog_map_compatible(map, prog)) {
419 rcpu->value.bpf_prog.id = prog->aux->id;
425 static struct bpf_cpu_map_entry *
426 __cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value,
429 int numa, err, i, fd = value->bpf_prog.fd;
430 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
431 struct bpf_cpu_map_entry *rcpu;
432 struct xdp_bulk_queue *bq;
434 /* Have map->numa_node, but choose node of redirect target CPU */
435 numa = cpu_to_node(cpu);
437 rcpu = bpf_map_kmalloc_node(map, sizeof(*rcpu), gfp | __GFP_ZERO, numa);
441 /* Alloc percpu bulkq */
442 rcpu->bulkq = bpf_map_alloc_percpu(map, sizeof(*rcpu->bulkq),
443 sizeof(void *), gfp);
447 for_each_possible_cpu(i) {
448 bq = per_cpu_ptr(rcpu->bulkq, i);
453 rcpu->queue = bpf_map_kmalloc_node(map, sizeof(*rcpu->queue), gfp,
458 err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
463 rcpu->map_id = map->id;
464 rcpu->value.qsize = value->qsize;
466 if (fd > 0 && __cpu_map_load_bpf_program(rcpu, map, fd))
470 init_completion(&rcpu->kthread_running);
471 rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
472 "cpumap/%d/map:%d", cpu,
474 if (IS_ERR(rcpu->kthread))
477 get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
478 get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
480 /* Make sure kthread runs on a single CPU */
481 kthread_bind(rcpu->kthread, cpu);
482 wake_up_process(rcpu->kthread);
484 /* Make sure kthread has been running, so kthread_stop() will not
485 * stop the kthread prematurely and all pending frames or skbs
486 * will be handled by the kthread before kthread_stop() returns.
488 wait_for_completion(&rcpu->kthread_running);
494 bpf_prog_put(rcpu->prog);
496 ptr_ring_cleanup(rcpu->queue, NULL);
500 free_percpu(rcpu->bulkq);
506 static void __cpu_map_entry_free(struct rcu_head *rcu)
508 struct bpf_cpu_map_entry *rcpu;
510 /* This cpu_map_entry have been disconnected from map and one
511 * RCU grace-period have elapsed. Thus, XDP cannot queue any
512 * new packets and cannot change/set flush_needed that can
515 rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
517 free_percpu(rcpu->bulkq);
518 /* Cannot kthread_stop() here, last put free rcpu resources */
519 put_cpu_map_entry(rcpu);
522 /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
523 * ensure any driver rcu critical sections have completed, but this
524 * does not guarantee a flush has happened yet. Because driver side
525 * rcu_read_lock/unlock only protects the running XDP program. The
526 * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
527 * pending flush op doesn't fail.
529 * The bpf_cpu_map_entry is still used by the kthread, and there can
530 * still be pending packets (in queue and percpu bulkq). A refcnt
531 * makes sure to last user (kthread_stop vs. call_rcu) free memory
534 * The rcu callback __cpu_map_entry_free flush remaining packets in
535 * percpu bulkq to queue. Due to caller map_delete_elem() disable
536 * preemption, cannot call kthread_stop() to make sure queue is empty.
537 * Instead a work_queue is started for stopping kthread,
538 * cpu_map_kthread_stop, which waits for an RCU grace period before
539 * stopping kthread, emptying the queue.
541 static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
542 u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
544 struct bpf_cpu_map_entry *old_rcpu;
546 old_rcpu = unrcu_pointer(xchg(&cmap->cpu_map[key_cpu], RCU_INITIALIZER(rcpu)));
548 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
549 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
550 schedule_work(&old_rcpu->kthread_stop_wq);
554 static long cpu_map_delete_elem(struct bpf_map *map, void *key)
556 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
557 u32 key_cpu = *(u32 *)key;
559 if (key_cpu >= map->max_entries)
562 /* notice caller map_delete_elem() use preempt_disable() */
563 __cpu_map_entry_replace(cmap, key_cpu, NULL);
567 static long cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
570 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
571 struct bpf_cpumap_val cpumap_value = {};
572 struct bpf_cpu_map_entry *rcpu;
573 /* Array index key correspond to CPU number */
574 u32 key_cpu = *(u32 *)key;
576 memcpy(&cpumap_value, value, map->value_size);
578 if (unlikely(map_flags > BPF_EXIST))
580 if (unlikely(key_cpu >= cmap->map.max_entries))
582 if (unlikely(map_flags == BPF_NOEXIST))
584 if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
587 /* Make sure CPU is a valid possible cpu */
588 if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
591 if (cpumap_value.qsize == 0) {
592 rcpu = NULL; /* Same as deleting */
594 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
595 rcpu = __cpu_map_entry_alloc(map, &cpumap_value, key_cpu);
601 __cpu_map_entry_replace(cmap, key_cpu, rcpu);
606 static void cpu_map_free(struct bpf_map *map)
608 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
611 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
612 * so the bpf programs (can be more than one that used this map) were
613 * disconnected from events. Wait for outstanding critical sections in
614 * these programs to complete. The rcu critical section only guarantees
615 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
616 * It does __not__ ensure pending flush operations (if any) are
622 /* For cpu_map the remote CPUs can still be using the entries
623 * (struct bpf_cpu_map_entry).
625 for (i = 0; i < cmap->map.max_entries; i++) {
626 struct bpf_cpu_map_entry *rcpu;
628 rcpu = rcu_dereference_raw(cmap->cpu_map[i]);
632 /* bq flush and cleanup happens after RCU grace-period */
633 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
635 bpf_map_area_free(cmap->cpu_map);
636 bpf_map_area_free(cmap);
639 /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
640 * by local_bh_disable() (from XDP calls inside NAPI). The
641 * rcu_read_lock_bh_held() below makes lockdep accept both.
643 static void *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
645 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
646 struct bpf_cpu_map_entry *rcpu;
648 if (key >= map->max_entries)
651 rcpu = rcu_dereference_check(cmap->cpu_map[key],
652 rcu_read_lock_bh_held());
656 static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
658 struct bpf_cpu_map_entry *rcpu =
659 __cpu_map_lookup_elem(map, *(u32 *)key);
661 return rcpu ? &rcpu->value : NULL;
664 static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
666 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
667 u32 index = key ? *(u32 *)key : U32_MAX;
668 u32 *next = next_key;
670 if (index >= cmap->map.max_entries) {
675 if (index == cmap->map.max_entries - 1)
681 static long cpu_map_redirect(struct bpf_map *map, u64 index, u64 flags)
683 return __bpf_xdp_redirect_map(map, index, flags, 0,
684 __cpu_map_lookup_elem);
687 static u64 cpu_map_mem_usage(const struct bpf_map *map)
689 u64 usage = sizeof(struct bpf_cpu_map);
691 /* Currently the dynamically allocated elements are not counted */
692 usage += (u64)map->max_entries * sizeof(struct bpf_cpu_map_entry *);
696 BTF_ID_LIST_SINGLE(cpu_map_btf_ids, struct, bpf_cpu_map)
697 const struct bpf_map_ops cpu_map_ops = {
698 .map_meta_equal = bpf_map_meta_equal,
699 .map_alloc = cpu_map_alloc,
700 .map_free = cpu_map_free,
701 .map_delete_elem = cpu_map_delete_elem,
702 .map_update_elem = cpu_map_update_elem,
703 .map_lookup_elem = cpu_map_lookup_elem,
704 .map_get_next_key = cpu_map_get_next_key,
705 .map_check_btf = map_check_no_btf,
706 .map_mem_usage = cpu_map_mem_usage,
707 .map_btf_id = &cpu_map_btf_ids[0],
708 .map_redirect = cpu_map_redirect,
711 static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
713 struct bpf_cpu_map_entry *rcpu = bq->obj;
714 unsigned int processed = 0, drops = 0;
715 const int to_cpu = rcpu->cpu;
719 if (unlikely(!bq->count))
723 spin_lock(&q->producer_lock);
725 for (i = 0; i < bq->count; i++) {
726 struct xdp_frame *xdpf = bq->q[i];
729 err = __ptr_ring_produce(q, xdpf);
732 xdp_return_frame_rx_napi(xdpf);
737 spin_unlock(&q->producer_lock);
739 __list_del_clearprev(&bq->flush_node);
741 /* Feedback loop via tracepoints */
742 trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
745 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
746 * Thus, safe percpu variable access.
748 static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
750 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
751 struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
753 if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
754 bq_flush_to_queue(bq);
756 /* Notice, xdp_buff/page MUST be queued here, long enough for
757 * driver to code invoking us to finished, due to driver
758 * (e.g. ixgbe) recycle tricks based on page-refcnt.
760 * Thus, incoming xdp_frame is always queued here (else we race
761 * with another CPU on page-refcnt and remaining driver code).
762 * Queue time is very short, as driver will invoke flush
763 * operation, when completing napi->poll call.
765 bq->q[bq->count++] = xdpf;
767 if (!bq->flush_node.prev)
768 list_add(&bq->flush_node, flush_list);
771 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf,
772 struct net_device *dev_rx)
774 /* Info needed when constructing SKB on remote CPU */
775 xdpf->dev_rx = dev_rx;
777 bq_enqueue(rcpu, xdpf);
781 int cpu_map_generic_redirect(struct bpf_cpu_map_entry *rcpu,
786 __skb_pull(skb, skb->mac_len);
787 skb_set_redirected(skb, false);
788 __ptr_set_bit(0, &skb);
790 ret = ptr_ring_produce(rcpu->queue, skb);
794 wake_up_process(rcpu->kthread);
796 trace_xdp_cpumap_enqueue(rcpu->map_id, !ret, !!ret, rcpu->cpu);
800 void __cpu_map_flush(void)
802 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
803 struct xdp_bulk_queue *bq, *tmp;
805 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
806 bq_flush_to_queue(bq);
808 /* If already running, costs spin_lock_irqsave + smb_mb */
809 wake_up_process(bq->obj->kthread);
813 static int __init cpu_map_init(void)
817 for_each_possible_cpu(cpu)
818 INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu));
822 subsys_initcall(cpu_map_init);