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 */
55 u32 qsize; /* Queue size placeholder for map lookup */
57 /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
58 struct xdp_bulk_queue __percpu *bulkq;
60 struct bpf_cpu_map *cmap;
62 /* Queue with potential multi-producers, and single-consumer kthread */
63 struct ptr_ring *queue;
64 struct task_struct *kthread;
65 struct work_struct kthread_stop_wq;
67 atomic_t refcnt; /* Control when this struct can be free'ed */
73 /* Below members specific for map type */
74 struct bpf_cpu_map_entry **cpu_map;
75 struct list_head __percpu *flush_list;
78 static int bq_flush_to_queue(struct xdp_bulk_queue *bq, bool in_napi_ctx);
80 static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
82 struct bpf_cpu_map *cmap;
87 if (!capable(CAP_SYS_ADMIN))
88 return ERR_PTR(-EPERM);
90 /* check sanity of attributes */
91 if (attr->max_entries == 0 || attr->key_size != 4 ||
92 attr->value_size != 4 || attr->map_flags & ~BPF_F_NUMA_NODE)
93 return ERR_PTR(-EINVAL);
95 cmap = kzalloc(sizeof(*cmap), GFP_USER);
97 return ERR_PTR(-ENOMEM);
99 bpf_map_init_from_attr(&cmap->map, attr);
101 /* Pre-limit array size based on NR_CPUS, not final CPU check */
102 if (cmap->map.max_entries > NR_CPUS) {
107 /* make sure page count doesn't overflow */
108 cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *);
109 cost += sizeof(struct list_head) * num_possible_cpus();
111 /* Notice returns -EPERM on if map size is larger than memlock limit */
112 ret = bpf_map_charge_init(&cmap->map.memory, cost);
118 cmap->flush_list = alloc_percpu(struct list_head);
119 if (!cmap->flush_list)
122 for_each_possible_cpu(cpu)
123 INIT_LIST_HEAD(per_cpu_ptr(cmap->flush_list, cpu));
125 /* Alloc array for possible remote "destination" CPUs */
126 cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
127 sizeof(struct bpf_cpu_map_entry *),
128 cmap->map.numa_node);
134 free_percpu(cmap->flush_list);
136 bpf_map_charge_finish(&cmap->map.memory);
142 static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
144 atomic_inc(&rcpu->refcnt);
147 /* called from workqueue, to workaround syscall using preempt_disable */
148 static void cpu_map_kthread_stop(struct work_struct *work)
150 struct bpf_cpu_map_entry *rcpu;
152 rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
154 /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
155 * as it waits until all in-flight call_rcu() callbacks complete.
159 /* kthread_stop will wake_up_process and wait for it to complete */
160 kthread_stop(rcpu->kthread);
163 static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu,
164 struct xdp_frame *xdpf,
167 unsigned int hard_start_headroom;
168 unsigned int frame_size;
169 void *pkt_data_start;
171 /* Part of headroom was reserved to xdpf */
172 hard_start_headroom = sizeof(struct xdp_frame) + xdpf->headroom;
174 /* build_skb need to place skb_shared_info after SKB end, and
175 * also want to know the memory "truesize". Thus, need to
176 * know the memory frame size backing xdp_buff.
178 * XDP was designed to have PAGE_SIZE frames, but this
179 * assumption is not longer true with ixgbe and i40e. It
180 * would be preferred to set frame_size to 2048 or 4096
181 * depending on the driver.
183 * frame_len = frame_size - sizeof(*xdp_frame);
185 * Instead, with info avail, skb_shared_info in placed after
186 * packet len. This, unfortunately fakes the truesize.
187 * Another disadvantage of this approach, the skb_shared_info
188 * is not at a fixed memory location, with mixed length
189 * packets, which is bad for cache-line hotness.
191 frame_size = SKB_DATA_ALIGN(xdpf->len + hard_start_headroom) +
192 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
194 pkt_data_start = xdpf->data - hard_start_headroom;
195 skb = build_skb_around(skb, pkt_data_start, frame_size);
199 skb_reserve(skb, hard_start_headroom);
200 __skb_put(skb, xdpf->len);
202 skb_metadata_set(skb, xdpf->metasize);
204 /* Essential SKB info: protocol and skb->dev */
205 skb->protocol = eth_type_trans(skb, xdpf->dev_rx);
207 /* Optional SKB info, currently missing:
208 * - HW checksum info (skb->ip_summed)
209 * - HW RX hash (skb_set_hash)
210 * - RX ring dev queue index (skb_record_rx_queue)
213 /* Until page_pool get SKB return path, release DMA here */
214 xdp_release_frame(xdpf);
216 /* Allow SKB to reuse area used by xdp_frame */
217 xdp_scrub_frame(xdpf);
222 static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
224 /* The tear-down procedure should have made sure that queue is
225 * empty. See __cpu_map_entry_replace() and work-queue
226 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
227 * gracefully and warn once.
229 struct xdp_frame *xdpf;
231 while ((xdpf = ptr_ring_consume(ring)))
232 if (WARN_ON_ONCE(xdpf))
233 xdp_return_frame(xdpf);
236 static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
238 if (atomic_dec_and_test(&rcpu->refcnt)) {
239 /* The queue should be empty at this point */
240 __cpu_map_ring_cleanup(rcpu->queue);
241 ptr_ring_cleanup(rcpu->queue, NULL);
247 #define CPUMAP_BATCH 8
249 static int cpu_map_kthread_run(void *data)
251 struct bpf_cpu_map_entry *rcpu = data;
253 set_current_state(TASK_INTERRUPTIBLE);
255 /* When kthread gives stop order, then rcpu have been disconnected
256 * from map, thus no new packets can enter. Remaining in-flight
257 * per CPU stored packets are flushed to this queue. Wait honoring
258 * kthread_stop signal until queue is empty.
260 while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
261 unsigned int drops = 0, sched = 0;
262 void *frames[CPUMAP_BATCH];
263 void *skbs[CPUMAP_BATCH];
264 gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
267 /* Release CPU reschedule checks */
268 if (__ptr_ring_empty(rcpu->queue)) {
269 set_current_state(TASK_INTERRUPTIBLE);
270 /* Recheck to avoid lost wake-up */
271 if (__ptr_ring_empty(rcpu->queue)) {
275 __set_current_state(TASK_RUNNING);
278 sched = cond_resched();
282 * The bpf_cpu_map_entry is single consumer, with this
283 * kthread CPU pinned. Lockless access to ptr_ring
284 * consume side valid as no-resize allowed of queue.
286 n = ptr_ring_consume_batched(rcpu->queue, frames, CPUMAP_BATCH);
288 for (i = 0; i < n; i++) {
290 struct page *page = virt_to_page(f);
292 /* Bring struct page memory area to curr CPU. Read by
293 * build_skb_around via page_is_pfmemalloc(), and when
294 * freed written by page_frag_free call.
299 m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, n, skbs);
300 if (unlikely(m == 0)) {
301 for (i = 0; i < n; i++)
302 skbs[i] = NULL; /* effect: xdp_return_frame */
307 for (i = 0; i < n; i++) {
308 struct xdp_frame *xdpf = frames[i];
309 struct sk_buff *skb = skbs[i];
312 skb = cpu_map_build_skb(rcpu, xdpf, skb);
314 xdp_return_frame(xdpf);
318 /* Inject into network stack */
319 ret = netif_receive_skb_core(skb);
320 if (ret == NET_RX_DROP)
323 /* Feedback loop via tracepoint */
324 trace_xdp_cpumap_kthread(rcpu->map_id, n, drops, sched);
326 local_bh_enable(); /* resched point, may call do_softirq() */
328 __set_current_state(TASK_RUNNING);
330 put_cpu_map_entry(rcpu);
334 static struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu,
337 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
338 struct bpf_cpu_map_entry *rcpu;
339 struct xdp_bulk_queue *bq;
342 /* Have map->numa_node, but choose node of redirect target CPU */
343 numa = cpu_to_node(cpu);
345 rcpu = kzalloc_node(sizeof(*rcpu), gfp, numa);
349 /* Alloc percpu bulkq */
350 rcpu->bulkq = __alloc_percpu_gfp(sizeof(*rcpu->bulkq),
351 sizeof(void *), gfp);
355 for_each_possible_cpu(i) {
356 bq = per_cpu_ptr(rcpu->bulkq, i);
361 rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
365 err = ptr_ring_init(rcpu->queue, qsize, gfp);
370 rcpu->map_id = map_id;
374 rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
375 "cpumap/%d/map:%d", cpu, map_id);
376 if (IS_ERR(rcpu->kthread))
379 get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
380 get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
382 /* Make sure kthread runs on a single CPU */
383 kthread_bind(rcpu->kthread, cpu);
384 wake_up_process(rcpu->kthread);
389 ptr_ring_cleanup(rcpu->queue, NULL);
393 free_percpu(rcpu->bulkq);
399 static void __cpu_map_entry_free(struct rcu_head *rcu)
401 struct bpf_cpu_map_entry *rcpu;
404 /* This cpu_map_entry have been disconnected from map and one
405 * RCU graze-period have elapsed. Thus, XDP cannot queue any
406 * new packets and cannot change/set flush_needed that can
409 rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
411 /* Flush remaining packets in percpu bulkq */
412 for_each_online_cpu(cpu) {
413 struct xdp_bulk_queue *bq = per_cpu_ptr(rcpu->bulkq, cpu);
415 /* No concurrent bq_enqueue can run at this point */
416 bq_flush_to_queue(bq, false);
418 free_percpu(rcpu->bulkq);
419 /* Cannot kthread_stop() here, last put free rcpu resources */
420 put_cpu_map_entry(rcpu);
423 /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
424 * ensure any driver rcu critical sections have completed, but this
425 * does not guarantee a flush has happened yet. Because driver side
426 * rcu_read_lock/unlock only protects the running XDP program. The
427 * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
428 * pending flush op doesn't fail.
430 * The bpf_cpu_map_entry is still used by the kthread, and there can
431 * still be pending packets (in queue and percpu bulkq). A refcnt
432 * makes sure to last user (kthread_stop vs. call_rcu) free memory
435 * The rcu callback __cpu_map_entry_free flush remaining packets in
436 * percpu bulkq to queue. Due to caller map_delete_elem() disable
437 * preemption, cannot call kthread_stop() to make sure queue is empty.
438 * Instead a work_queue is started for stopping kthread,
439 * cpu_map_kthread_stop, which waits for an RCU graze period before
440 * stopping kthread, emptying the queue.
442 static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
443 u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
445 struct bpf_cpu_map_entry *old_rcpu;
447 old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu);
449 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
450 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
451 schedule_work(&old_rcpu->kthread_stop_wq);
455 static int cpu_map_delete_elem(struct bpf_map *map, void *key)
457 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
458 u32 key_cpu = *(u32 *)key;
460 if (key_cpu >= map->max_entries)
463 /* notice caller map_delete_elem() use preempt_disable() */
464 __cpu_map_entry_replace(cmap, key_cpu, NULL);
468 static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
471 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
472 struct bpf_cpu_map_entry *rcpu;
474 /* Array index key correspond to CPU number */
475 u32 key_cpu = *(u32 *)key;
476 /* Value is the queue size */
477 u32 qsize = *(u32 *)value;
479 if (unlikely(map_flags > BPF_EXIST))
481 if (unlikely(key_cpu >= cmap->map.max_entries))
483 if (unlikely(map_flags == BPF_NOEXIST))
485 if (unlikely(qsize > 16384)) /* sanity limit on qsize */
488 /* Make sure CPU is a valid possible cpu */
489 if (!cpu_possible(key_cpu))
493 rcpu = NULL; /* Same as deleting */
495 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
496 rcpu = __cpu_map_entry_alloc(qsize, key_cpu, map->id);
502 __cpu_map_entry_replace(cmap, key_cpu, rcpu);
507 static void cpu_map_free(struct bpf_map *map)
509 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
513 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
514 * so the bpf programs (can be more than one that used this map) were
515 * disconnected from events. Wait for outstanding critical sections in
516 * these programs to complete. The rcu critical section only guarantees
517 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
518 * It does __not__ ensure pending flush operations (if any) are
522 bpf_clear_redirect_map(map);
525 /* To ensure all pending flush operations have completed wait for flush
526 * list be empty on _all_ cpus. Because the above synchronize_rcu()
527 * ensures the map is disconnected from the program we can assume no new
528 * items will be added to the list.
530 for_each_online_cpu(cpu) {
531 struct list_head *flush_list = per_cpu_ptr(cmap->flush_list, cpu);
533 while (!list_empty(flush_list))
537 /* For cpu_map the remote CPUs can still be using the entries
538 * (struct bpf_cpu_map_entry).
540 for (i = 0; i < cmap->map.max_entries; i++) {
541 struct bpf_cpu_map_entry *rcpu;
543 rcpu = READ_ONCE(cmap->cpu_map[i]);
547 /* bq flush and cleanup happens after RCU graze-period */
548 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
550 free_percpu(cmap->flush_list);
551 bpf_map_area_free(cmap->cpu_map);
555 struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
557 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
558 struct bpf_cpu_map_entry *rcpu;
560 if (key >= map->max_entries)
563 rcpu = READ_ONCE(cmap->cpu_map[key]);
567 static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
569 struct bpf_cpu_map_entry *rcpu =
570 __cpu_map_lookup_elem(map, *(u32 *)key);
572 return rcpu ? &rcpu->qsize : NULL;
575 static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
577 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
578 u32 index = key ? *(u32 *)key : U32_MAX;
579 u32 *next = next_key;
581 if (index >= cmap->map.max_entries) {
586 if (index == cmap->map.max_entries - 1)
592 const struct bpf_map_ops cpu_map_ops = {
593 .map_alloc = cpu_map_alloc,
594 .map_free = cpu_map_free,
595 .map_delete_elem = cpu_map_delete_elem,
596 .map_update_elem = cpu_map_update_elem,
597 .map_lookup_elem = cpu_map_lookup_elem,
598 .map_get_next_key = cpu_map_get_next_key,
599 .map_check_btf = map_check_no_btf,
602 static int bq_flush_to_queue(struct xdp_bulk_queue *bq, bool in_napi_ctx)
604 struct bpf_cpu_map_entry *rcpu = bq->obj;
605 unsigned int processed = 0, drops = 0;
606 const int to_cpu = rcpu->cpu;
610 if (unlikely(!bq->count))
614 spin_lock(&q->producer_lock);
616 for (i = 0; i < bq->count; i++) {
617 struct xdp_frame *xdpf = bq->q[i];
620 err = __ptr_ring_produce(q, xdpf);
623 if (likely(in_napi_ctx))
624 xdp_return_frame_rx_napi(xdpf);
626 xdp_return_frame(xdpf);
631 spin_unlock(&q->producer_lock);
633 __list_del_clearprev(&bq->flush_node);
635 /* Feedback loop via tracepoints */
636 trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
640 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
641 * Thus, safe percpu variable access.
643 static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
645 struct list_head *flush_list = this_cpu_ptr(rcpu->cmap->flush_list);
646 struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
648 if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
649 bq_flush_to_queue(bq, true);
651 /* Notice, xdp_buff/page MUST be queued here, long enough for
652 * driver to code invoking us to finished, due to driver
653 * (e.g. ixgbe) recycle tricks based on page-refcnt.
655 * Thus, incoming xdp_frame is always queued here (else we race
656 * with another CPU on page-refcnt and remaining driver code).
657 * Queue time is very short, as driver will invoke flush
658 * operation, when completing napi->poll call.
660 bq->q[bq->count++] = xdpf;
662 if (!bq->flush_node.prev)
663 list_add(&bq->flush_node, flush_list);
668 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
669 struct net_device *dev_rx)
671 struct xdp_frame *xdpf;
673 xdpf = convert_to_xdp_frame(xdp);
677 /* Info needed when constructing SKB on remote CPU */
678 xdpf->dev_rx = dev_rx;
680 bq_enqueue(rcpu, xdpf);
684 void __cpu_map_flush(struct bpf_map *map)
686 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
687 struct list_head *flush_list = this_cpu_ptr(cmap->flush_list);
688 struct xdp_bulk_queue *bq, *tmp;
690 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
691 bq_flush_to_queue(bq, true);
693 /* If already running, costs spin_lock_irqsave + smb_mb */
694 wake_up_process(bq->obj->kthread);