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;
91 return ERR_PTR(-EPERM);
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 cmap = kzalloc(sizeof(*cmap), GFP_USER);
102 return ERR_PTR(-ENOMEM);
104 bpf_map_init_from_attr(&cmap->map, attr);
106 /* Pre-limit array size based on NR_CPUS, not final CPU check */
107 if (cmap->map.max_entries > NR_CPUS) {
112 /* make sure page count doesn't overflow */
113 cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *);
115 /* Notice returns -EPERM on if map size is larger than memlock limit */
116 ret = bpf_map_charge_init(&cmap->map.memory, cost);
122 /* Alloc array for possible remote "destination" CPUs */
123 cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
124 sizeof(struct bpf_cpu_map_entry *),
125 cmap->map.numa_node);
131 bpf_map_charge_finish(&cmap->map.memory);
137 static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
139 atomic_inc(&rcpu->refcnt);
142 /* called from workqueue, to workaround syscall using preempt_disable */
143 static void cpu_map_kthread_stop(struct work_struct *work)
145 struct bpf_cpu_map_entry *rcpu;
147 rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
149 /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
150 * as it waits until all in-flight call_rcu() callbacks complete.
154 /* kthread_stop will wake_up_process and wait for it to complete */
155 kthread_stop(rcpu->kthread);
158 static struct sk_buff *cpu_map_build_skb(struct xdp_frame *xdpf,
161 unsigned int hard_start_headroom;
162 unsigned int frame_size;
163 void *pkt_data_start;
165 /* Part of headroom was reserved to xdpf */
166 hard_start_headroom = sizeof(struct xdp_frame) + xdpf->headroom;
168 /* Memory size backing xdp_frame data already have reserved
169 * room for build_skb to place skb_shared_info in tailroom.
171 frame_size = xdpf->frame_sz;
173 pkt_data_start = xdpf->data - hard_start_headroom;
174 skb = build_skb_around(skb, pkt_data_start, frame_size);
178 skb_reserve(skb, hard_start_headroom);
179 __skb_put(skb, xdpf->len);
181 skb_metadata_set(skb, xdpf->metasize);
183 /* Essential SKB info: protocol and skb->dev */
184 skb->protocol = eth_type_trans(skb, xdpf->dev_rx);
186 /* Optional SKB info, currently missing:
187 * - HW checksum info (skb->ip_summed)
188 * - HW RX hash (skb_set_hash)
189 * - RX ring dev queue index (skb_record_rx_queue)
192 /* Until page_pool get SKB return path, release DMA here */
193 xdp_release_frame(xdpf);
195 /* Allow SKB to reuse area used by xdp_frame */
196 xdp_scrub_frame(xdpf);
201 static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
203 /* The tear-down procedure should have made sure that queue is
204 * empty. See __cpu_map_entry_replace() and work-queue
205 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
206 * gracefully and warn once.
208 struct xdp_frame *xdpf;
210 while ((xdpf = ptr_ring_consume(ring)))
211 if (WARN_ON_ONCE(xdpf))
212 xdp_return_frame(xdpf);
215 static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
217 if (atomic_dec_and_test(&rcpu->refcnt)) {
219 bpf_prog_put(rcpu->prog);
220 /* The queue should be empty at this point */
221 __cpu_map_ring_cleanup(rcpu->queue);
222 ptr_ring_cleanup(rcpu->queue, NULL);
228 static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
229 void **frames, int n,
230 struct xdp_cpumap_stats *stats)
232 struct xdp_rxq_info rxq = {};
241 xdp_set_return_frame_no_direct();
244 for (i = 0; i < n; i++) {
245 struct xdp_frame *xdpf = frames[i];
249 rxq.dev = xdpf->dev_rx;
251 /* TODO: report queue_index to xdp_rxq_info */
253 xdp_convert_frame_to_buff(xdpf, &xdp);
255 act = bpf_prog_run_xdp(rcpu->prog, &xdp);
258 err = xdp_update_frame_from_buff(&xdp, xdpf);
260 xdp_return_frame(xdpf);
263 frames[nframes++] = xdpf;
268 err = xdp_do_redirect(xdpf->dev_rx, &xdp,
271 xdp_return_frame(xdpf);
278 bpf_warn_invalid_xdp_action(act);
281 xdp_return_frame(xdpf);
290 xdp_clear_return_frame_no_direct();
292 rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
297 #define CPUMAP_BATCH 8
299 static int cpu_map_kthread_run(void *data)
301 struct bpf_cpu_map_entry *rcpu = data;
302 unsigned long last_qs = jiffies;
304 set_current_state(TASK_INTERRUPTIBLE);
306 /* When kthread gives stop order, then rcpu have been disconnected
307 * from map, thus no new packets can enter. Remaining in-flight
308 * per CPU stored packets are flushed to this queue. Wait honoring
309 * kthread_stop signal until queue is empty.
311 while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
312 struct xdp_cpumap_stats stats = {}; /* zero stats */
313 gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
314 unsigned int drops = 0, sched = 0;
315 void *frames[CPUMAP_BATCH];
316 void *skbs[CPUMAP_BATCH];
317 int i, n, m, nframes;
319 /* Release CPU reschedule checks */
320 if (__ptr_ring_empty(rcpu->queue)) {
321 set_current_state(TASK_INTERRUPTIBLE);
322 /* Recheck to avoid lost wake-up */
323 if (__ptr_ring_empty(rcpu->queue)) {
328 __set_current_state(TASK_RUNNING);
331 rcu_softirq_qs_periodic(last_qs);
332 sched = cond_resched();
336 * The bpf_cpu_map_entry is single consumer, with this
337 * kthread CPU pinned. Lockless access to ptr_ring
338 * consume side valid as no-resize allowed of queue.
340 n = __ptr_ring_consume_batched(rcpu->queue, frames,
342 for (i = 0; i < n; i++) {
344 struct page *page = virt_to_page(f);
346 /* Bring struct page memory area to curr CPU. Read by
347 * build_skb_around via page_is_pfmemalloc(), and when
348 * freed written by page_frag_free call.
353 /* Support running another XDP prog on this CPU */
354 nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, n, &stats);
356 m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, nframes, skbs);
357 if (unlikely(m == 0)) {
358 for (i = 0; i < nframes; i++)
359 skbs[i] = NULL; /* effect: xdp_return_frame */
365 for (i = 0; i < nframes; i++) {
366 struct xdp_frame *xdpf = frames[i];
367 struct sk_buff *skb = skbs[i];
370 skb = cpu_map_build_skb(xdpf, skb);
372 xdp_return_frame(xdpf);
376 /* Inject into network stack */
377 ret = netif_receive_skb_core(skb);
378 if (ret == NET_RX_DROP)
381 /* Feedback loop via tracepoint */
382 trace_xdp_cpumap_kthread(rcpu->map_id, n, drops, sched, &stats);
384 local_bh_enable(); /* resched point, may call do_softirq() */
386 __set_current_state(TASK_RUNNING);
388 put_cpu_map_entry(rcpu);
392 bool cpu_map_prog_allowed(struct bpf_map *map)
394 return map->map_type == BPF_MAP_TYPE_CPUMAP &&
395 map->value_size != offsetofend(struct bpf_cpumap_val, qsize);
398 static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu, int fd)
400 struct bpf_prog *prog;
402 prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
404 return PTR_ERR(prog);
406 if (prog->expected_attach_type != BPF_XDP_CPUMAP) {
411 rcpu->value.bpf_prog.id = prog->aux->id;
417 static struct bpf_cpu_map_entry *
418 __cpu_map_entry_alloc(struct bpf_cpumap_val *value, u32 cpu, int map_id)
420 int numa, err, i, fd = value->bpf_prog.fd;
421 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
422 struct bpf_cpu_map_entry *rcpu;
423 struct xdp_bulk_queue *bq;
425 /* Have map->numa_node, but choose node of redirect target CPU */
426 numa = cpu_to_node(cpu);
428 rcpu = kzalloc_node(sizeof(*rcpu), gfp, numa);
432 /* Alloc percpu bulkq */
433 rcpu->bulkq = __alloc_percpu_gfp(sizeof(*rcpu->bulkq),
434 sizeof(void *), gfp);
438 for_each_possible_cpu(i) {
439 bq = per_cpu_ptr(rcpu->bulkq, i);
444 rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
448 err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
453 rcpu->map_id = map_id;
454 rcpu->value.qsize = value->qsize;
456 if (fd > 0 && __cpu_map_load_bpf_program(rcpu, fd))
460 rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
461 "cpumap/%d/map:%d", cpu, map_id);
462 if (IS_ERR(rcpu->kthread))
465 get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
466 get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
468 /* Make sure kthread runs on a single CPU */
469 kthread_bind(rcpu->kthread, cpu);
470 wake_up_process(rcpu->kthread);
476 bpf_prog_put(rcpu->prog);
478 ptr_ring_cleanup(rcpu->queue, NULL);
482 free_percpu(rcpu->bulkq);
488 static void __cpu_map_entry_free(struct rcu_head *rcu)
490 struct bpf_cpu_map_entry *rcpu;
492 /* This cpu_map_entry have been disconnected from map and one
493 * RCU grace-period have elapsed. Thus, XDP cannot queue any
494 * new packets and cannot change/set flush_needed that can
497 rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
499 free_percpu(rcpu->bulkq);
500 /* Cannot kthread_stop() here, last put free rcpu resources */
501 put_cpu_map_entry(rcpu);
504 /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
505 * ensure any driver rcu critical sections have completed, but this
506 * does not guarantee a flush has happened yet. Because driver side
507 * rcu_read_lock/unlock only protects the running XDP program. The
508 * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
509 * pending flush op doesn't fail.
511 * The bpf_cpu_map_entry is still used by the kthread, and there can
512 * still be pending packets (in queue and percpu bulkq). A refcnt
513 * makes sure to last user (kthread_stop vs. call_rcu) free memory
516 * The rcu callback __cpu_map_entry_free flush remaining packets in
517 * percpu bulkq to queue. Due to caller map_delete_elem() disable
518 * preemption, cannot call kthread_stop() to make sure queue is empty.
519 * Instead a work_queue is started for stopping kthread,
520 * cpu_map_kthread_stop, which waits for an RCU grace period before
521 * stopping kthread, emptying the queue.
523 static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
524 u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
526 struct bpf_cpu_map_entry *old_rcpu;
528 old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu);
530 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
531 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
532 schedule_work(&old_rcpu->kthread_stop_wq);
536 static int cpu_map_delete_elem(struct bpf_map *map, void *key)
538 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
539 u32 key_cpu = *(u32 *)key;
541 if (key_cpu >= map->max_entries)
544 /* notice caller map_delete_elem() use preempt_disable() */
545 __cpu_map_entry_replace(cmap, key_cpu, NULL);
549 static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
552 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
553 struct bpf_cpumap_val cpumap_value = {};
554 struct bpf_cpu_map_entry *rcpu;
555 /* Array index key correspond to CPU number */
556 u32 key_cpu = *(u32 *)key;
558 memcpy(&cpumap_value, value, map->value_size);
560 if (unlikely(map_flags > BPF_EXIST))
562 if (unlikely(key_cpu >= cmap->map.max_entries))
564 if (unlikely(map_flags == BPF_NOEXIST))
566 if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
569 /* Make sure CPU is a valid possible cpu */
570 if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
573 if (cpumap_value.qsize == 0) {
574 rcpu = NULL; /* Same as deleting */
576 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
577 rcpu = __cpu_map_entry_alloc(&cpumap_value, key_cpu, map->id);
583 __cpu_map_entry_replace(cmap, key_cpu, rcpu);
588 static void cpu_map_free(struct bpf_map *map)
590 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
593 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
594 * so the bpf programs (can be more than one that used this map) were
595 * disconnected from events. Wait for outstanding critical sections in
596 * these programs to complete. The rcu critical section only guarantees
597 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
598 * It does __not__ ensure pending flush operations (if any) are
602 bpf_clear_redirect_map(map);
605 /* For cpu_map the remote CPUs can still be using the entries
606 * (struct bpf_cpu_map_entry).
608 for (i = 0; i < cmap->map.max_entries; i++) {
609 struct bpf_cpu_map_entry *rcpu;
611 rcpu = READ_ONCE(cmap->cpu_map[i]);
615 /* bq flush and cleanup happens after RCU grace-period */
616 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
618 bpf_map_area_free(cmap->cpu_map);
622 struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
624 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
625 struct bpf_cpu_map_entry *rcpu;
627 if (key >= map->max_entries)
630 rcpu = READ_ONCE(cmap->cpu_map[key]);
634 static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
636 struct bpf_cpu_map_entry *rcpu =
637 __cpu_map_lookup_elem(map, *(u32 *)key);
639 return rcpu ? &rcpu->value : NULL;
642 static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
644 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
645 u32 index = key ? *(u32 *)key : U32_MAX;
646 u32 *next = next_key;
648 if (index >= cmap->map.max_entries) {
653 if (index == cmap->map.max_entries - 1)
659 static int cpu_map_btf_id;
660 const struct bpf_map_ops cpu_map_ops = {
661 .map_meta_equal = bpf_map_meta_equal,
662 .map_alloc = cpu_map_alloc,
663 .map_free = cpu_map_free,
664 .map_delete_elem = cpu_map_delete_elem,
665 .map_update_elem = cpu_map_update_elem,
666 .map_lookup_elem = cpu_map_lookup_elem,
667 .map_get_next_key = cpu_map_get_next_key,
668 .map_check_btf = map_check_no_btf,
669 .map_btf_name = "bpf_cpu_map",
670 .map_btf_id = &cpu_map_btf_id,
673 static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
675 struct bpf_cpu_map_entry *rcpu = bq->obj;
676 unsigned int processed = 0, drops = 0;
677 const int to_cpu = rcpu->cpu;
681 if (unlikely(!bq->count))
685 spin_lock(&q->producer_lock);
687 for (i = 0; i < bq->count; i++) {
688 struct xdp_frame *xdpf = bq->q[i];
691 err = __ptr_ring_produce(q, xdpf);
694 xdp_return_frame_rx_napi(xdpf);
699 spin_unlock(&q->producer_lock);
701 __list_del_clearprev(&bq->flush_node);
703 /* Feedback loop via tracepoints */
704 trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
707 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
708 * Thus, safe percpu variable access.
710 static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
712 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
713 struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
715 if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
716 bq_flush_to_queue(bq);
718 /* Notice, xdp_buff/page MUST be queued here, long enough for
719 * driver to code invoking us to finished, due to driver
720 * (e.g. ixgbe) recycle tricks based on page-refcnt.
722 * Thus, incoming xdp_frame is always queued here (else we race
723 * with another CPU on page-refcnt and remaining driver code).
724 * Queue time is very short, as driver will invoke flush
725 * operation, when completing napi->poll call.
727 bq->q[bq->count++] = xdpf;
729 if (!bq->flush_node.prev)
730 list_add(&bq->flush_node, flush_list);
733 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
734 struct net_device *dev_rx)
736 struct xdp_frame *xdpf;
738 xdpf = xdp_convert_buff_to_frame(xdp);
742 /* Info needed when constructing SKB on remote CPU */
743 xdpf->dev_rx = dev_rx;
745 bq_enqueue(rcpu, xdpf);
749 void __cpu_map_flush(void)
751 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
752 struct xdp_bulk_queue *bq, *tmp;
754 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
755 bq_flush_to_queue(bq);
757 /* If already running, costs spin_lock_irqsave + smb_mb */
758 wake_up_process(bq->obj->kthread);
762 static int __init cpu_map_init(void)
766 for_each_possible_cpu(cpu)
767 INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu));
771 subsys_initcall(cpu_map_init);