3 * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
4 * Released under terms in GPL version 2. See COPYING.
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 setting flush bit and 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 xdp_bulk_queue {
42 void *q[CPU_MAP_BULK_SIZE];
46 /* Struct for every remote "destination" CPU in map */
47 struct bpf_cpu_map_entry {
48 u32 cpu; /* kthread CPU and map index */
49 int map_id; /* Back reference to map */
50 u32 qsize; /* Queue size placeholder for map lookup */
52 /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
53 struct xdp_bulk_queue __percpu *bulkq;
55 /* Queue with potential multi-producers, and single-consumer kthread */
56 struct ptr_ring *queue;
57 struct task_struct *kthread;
58 struct work_struct kthread_stop_wq;
60 atomic_t refcnt; /* Control when this struct can be free'ed */
66 /* Below members specific for map type */
67 struct bpf_cpu_map_entry **cpu_map;
68 unsigned long __percpu *flush_needed;
71 static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
72 struct xdp_bulk_queue *bq, bool in_napi_ctx);
74 static u64 cpu_map_bitmap_size(const union bpf_attr *attr)
76 return BITS_TO_LONGS(attr->max_entries) * sizeof(unsigned long);
79 static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
81 struct bpf_cpu_map *cmap;
86 if (!capable(CAP_SYS_ADMIN))
87 return ERR_PTR(-EPERM);
89 /* check sanity of attributes */
90 if (attr->max_entries == 0 || attr->key_size != 4 ||
91 attr->value_size != 4 || attr->map_flags & ~BPF_F_NUMA_NODE)
92 return ERR_PTR(-EINVAL);
94 cmap = kzalloc(sizeof(*cmap), GFP_USER);
96 return ERR_PTR(-ENOMEM);
98 bpf_map_init_from_attr(&cmap->map, attr);
100 /* Pre-limit array size based on NR_CPUS, not final CPU check */
101 if (cmap->map.max_entries > NR_CPUS) {
106 /* make sure page count doesn't overflow */
107 cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *);
108 cost += cpu_map_bitmap_size(attr) * num_possible_cpus();
109 if (cost >= U32_MAX - PAGE_SIZE)
111 cmap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
113 /* Notice returns -EPERM on if map size is larger than memlock limit */
114 ret = bpf_map_precharge_memlock(cmap->map.pages);
120 /* A per cpu bitfield with a bit per possible CPU in map */
121 cmap->flush_needed = __alloc_percpu(cpu_map_bitmap_size(attr),
122 __alignof__(unsigned long));
123 if (!cmap->flush_needed)
126 /* Alloc array for possible remote "destination" CPUs */
127 cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
128 sizeof(struct bpf_cpu_map_entry *),
129 cmap->map.numa_node);
135 free_percpu(cmap->flush_needed);
141 static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
143 atomic_inc(&rcpu->refcnt);
146 /* called from workqueue, to workaround syscall using preempt_disable */
147 static void cpu_map_kthread_stop(struct work_struct *work)
149 struct bpf_cpu_map_entry *rcpu;
151 rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
153 /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
154 * as it waits until all in-flight call_rcu() callbacks complete.
158 /* kthread_stop will wake_up_process and wait for it to complete */
159 kthread_stop(rcpu->kthread);
162 static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu,
163 struct xdp_frame *xdpf)
165 unsigned int frame_size;
166 void *pkt_data_start;
169 /* build_skb need to place skb_shared_info after SKB end, and
170 * also want to know the memory "truesize". Thus, need to
171 * know the memory frame size backing xdp_buff.
173 * XDP was designed to have PAGE_SIZE frames, but this
174 * assumption is not longer true with ixgbe and i40e. It
175 * would be preferred to set frame_size to 2048 or 4096
176 * depending on the driver.
178 * frame_len = frame_size - sizeof(*xdp_frame);
180 * Instead, with info avail, skb_shared_info in placed after
181 * packet len. This, unfortunately fakes the truesize.
182 * Another disadvantage of this approach, the skb_shared_info
183 * is not at a fixed memory location, with mixed length
184 * packets, which is bad for cache-line hotness.
186 frame_size = SKB_DATA_ALIGN(xdpf->len) + xdpf->headroom +
187 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
189 pkt_data_start = xdpf->data - xdpf->headroom;
190 skb = build_skb(pkt_data_start, frame_size);
194 skb_reserve(skb, xdpf->headroom);
195 __skb_put(skb, xdpf->len);
197 skb_metadata_set(skb, xdpf->metasize);
199 /* Essential SKB info: protocol and skb->dev */
200 skb->protocol = eth_type_trans(skb, xdpf->dev_rx);
202 /* Optional SKB info, currently missing:
203 * - HW checksum info (skb->ip_summed)
204 * - HW RX hash (skb_set_hash)
205 * - RX ring dev queue index (skb_record_rx_queue)
211 static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
213 /* The tear-down procedure should have made sure that queue is
214 * empty. See __cpu_map_entry_replace() and work-queue
215 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
216 * gracefully and warn once.
218 struct xdp_frame *xdpf;
220 while ((xdpf = ptr_ring_consume(ring)))
221 if (WARN_ON_ONCE(xdpf))
222 xdp_return_frame(xdpf);
225 static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
227 if (atomic_dec_and_test(&rcpu->refcnt)) {
228 /* The queue should be empty at this point */
229 __cpu_map_ring_cleanup(rcpu->queue);
230 ptr_ring_cleanup(rcpu->queue, NULL);
236 static int cpu_map_kthread_run(void *data)
238 struct bpf_cpu_map_entry *rcpu = data;
240 set_current_state(TASK_INTERRUPTIBLE);
242 /* When kthread gives stop order, then rcpu have been disconnected
243 * from map, thus no new packets can enter. Remaining in-flight
244 * per CPU stored packets are flushed to this queue. Wait honoring
245 * kthread_stop signal until queue is empty.
247 while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
248 unsigned int processed = 0, drops = 0, sched = 0;
249 struct xdp_frame *xdpf;
251 /* Release CPU reschedule checks */
252 if (__ptr_ring_empty(rcpu->queue)) {
253 set_current_state(TASK_INTERRUPTIBLE);
254 /* Recheck to avoid lost wake-up */
255 if (__ptr_ring_empty(rcpu->queue)) {
259 __set_current_state(TASK_RUNNING);
262 sched = cond_resched();
265 /* Process packets in rcpu->queue */
268 * The bpf_cpu_map_entry is single consumer, with this
269 * kthread CPU pinned. Lockless access to ptr_ring
270 * consume side valid as no-resize allowed of queue.
272 while ((xdpf = __ptr_ring_consume(rcpu->queue))) {
276 skb = cpu_map_build_skb(rcpu, xdpf);
278 xdp_return_frame(xdpf);
282 /* Inject into network stack */
283 ret = netif_receive_skb_core(skb);
284 if (ret == NET_RX_DROP)
287 /* Limit BH-disable period */
288 if (++processed == 8)
291 /* Feedback loop via tracepoint */
292 trace_xdp_cpumap_kthread(rcpu->map_id, processed, drops, sched);
294 local_bh_enable(); /* resched point, may call do_softirq() */
296 __set_current_state(TASK_RUNNING);
298 put_cpu_map_entry(rcpu);
302 static struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu,
305 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
306 struct bpf_cpu_map_entry *rcpu;
309 /* Have map->numa_node, but choose node of redirect target CPU */
310 numa = cpu_to_node(cpu);
312 rcpu = kzalloc_node(sizeof(*rcpu), gfp, numa);
316 /* Alloc percpu bulkq */
317 rcpu->bulkq = __alloc_percpu_gfp(sizeof(*rcpu->bulkq),
318 sizeof(void *), gfp);
323 rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
327 err = ptr_ring_init(rcpu->queue, qsize, gfp);
332 rcpu->map_id = map_id;
336 rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
337 "cpumap/%d/map:%d", cpu, map_id);
338 if (IS_ERR(rcpu->kthread))
341 get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
342 get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
344 /* Make sure kthread runs on a single CPU */
345 kthread_bind(rcpu->kthread, cpu);
346 wake_up_process(rcpu->kthread);
351 ptr_ring_cleanup(rcpu->queue, NULL);
355 free_percpu(rcpu->bulkq);
361 static void __cpu_map_entry_free(struct rcu_head *rcu)
363 struct bpf_cpu_map_entry *rcpu;
366 /* This cpu_map_entry have been disconnected from map and one
367 * RCU graze-period have elapsed. Thus, XDP cannot queue any
368 * new packets and cannot change/set flush_needed that can
371 rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
373 /* Flush remaining packets in percpu bulkq */
374 for_each_online_cpu(cpu) {
375 struct xdp_bulk_queue *bq = per_cpu_ptr(rcpu->bulkq, cpu);
377 /* No concurrent bq_enqueue can run at this point */
378 bq_flush_to_queue(rcpu, bq, false);
380 free_percpu(rcpu->bulkq);
381 /* Cannot kthread_stop() here, last put free rcpu resources */
382 put_cpu_map_entry(rcpu);
385 /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
386 * ensure any driver rcu critical sections have completed, but this
387 * does not guarantee a flush has happened yet. Because driver side
388 * rcu_read_lock/unlock only protects the running XDP program. The
389 * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
390 * pending flush op doesn't fail.
392 * The bpf_cpu_map_entry is still used by the kthread, and there can
393 * still be pending packets (in queue and percpu bulkq). A refcnt
394 * makes sure to last user (kthread_stop vs. call_rcu) free memory
397 * The rcu callback __cpu_map_entry_free flush remaining packets in
398 * percpu bulkq to queue. Due to caller map_delete_elem() disable
399 * preemption, cannot call kthread_stop() to make sure queue is empty.
400 * Instead a work_queue is started for stopping kthread,
401 * cpu_map_kthread_stop, which waits for an RCU graze period before
402 * stopping kthread, emptying the queue.
404 static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
405 u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
407 struct bpf_cpu_map_entry *old_rcpu;
409 old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu);
411 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
412 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
413 schedule_work(&old_rcpu->kthread_stop_wq);
417 static int cpu_map_delete_elem(struct bpf_map *map, void *key)
419 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
420 u32 key_cpu = *(u32 *)key;
422 if (key_cpu >= map->max_entries)
425 /* notice caller map_delete_elem() use preempt_disable() */
426 __cpu_map_entry_replace(cmap, key_cpu, NULL);
430 static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
433 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
434 struct bpf_cpu_map_entry *rcpu;
436 /* Array index key correspond to CPU number */
437 u32 key_cpu = *(u32 *)key;
438 /* Value is the queue size */
439 u32 qsize = *(u32 *)value;
441 if (unlikely(map_flags > BPF_EXIST))
443 if (unlikely(key_cpu >= cmap->map.max_entries))
445 if (unlikely(map_flags == BPF_NOEXIST))
447 if (unlikely(qsize > 16384)) /* sanity limit on qsize */
450 /* Make sure CPU is a valid possible cpu */
451 if (!cpu_possible(key_cpu))
455 rcpu = NULL; /* Same as deleting */
457 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
458 rcpu = __cpu_map_entry_alloc(qsize, key_cpu, map->id);
463 __cpu_map_entry_replace(cmap, key_cpu, rcpu);
468 static void cpu_map_free(struct bpf_map *map)
470 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
474 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
475 * so the bpf programs (can be more than one that used this map) were
476 * disconnected from events. Wait for outstanding critical sections in
477 * these programs to complete. The rcu critical section only guarantees
478 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
479 * It does __not__ ensure pending flush operations (if any) are
484 /* To ensure all pending flush operations have completed wait for flush
485 * bitmap to indicate all flush_needed bits to be zero on _all_ cpus.
486 * Because the above synchronize_rcu() ensures the map is disconnected
487 * from the program we can assume no new bits will be set.
489 for_each_online_cpu(cpu) {
490 unsigned long *bitmap = per_cpu_ptr(cmap->flush_needed, cpu);
492 while (!bitmap_empty(bitmap, cmap->map.max_entries))
496 /* For cpu_map the remote CPUs can still be using the entries
497 * (struct bpf_cpu_map_entry).
499 for (i = 0; i < cmap->map.max_entries; i++) {
500 struct bpf_cpu_map_entry *rcpu;
502 rcpu = READ_ONCE(cmap->cpu_map[i]);
506 /* bq flush and cleanup happens after RCU graze-period */
507 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
509 free_percpu(cmap->flush_needed);
510 bpf_map_area_free(cmap->cpu_map);
514 struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
516 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
517 struct bpf_cpu_map_entry *rcpu;
519 if (key >= map->max_entries)
522 rcpu = READ_ONCE(cmap->cpu_map[key]);
526 static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
528 struct bpf_cpu_map_entry *rcpu =
529 __cpu_map_lookup_elem(map, *(u32 *)key);
531 return rcpu ? &rcpu->qsize : NULL;
534 static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
536 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
537 u32 index = key ? *(u32 *)key : U32_MAX;
538 u32 *next = next_key;
540 if (index >= cmap->map.max_entries) {
545 if (index == cmap->map.max_entries - 1)
551 const struct bpf_map_ops cpu_map_ops = {
552 .map_alloc = cpu_map_alloc,
553 .map_free = cpu_map_free,
554 .map_delete_elem = cpu_map_delete_elem,
555 .map_update_elem = cpu_map_update_elem,
556 .map_lookup_elem = cpu_map_lookup_elem,
557 .map_get_next_key = cpu_map_get_next_key,
558 .map_check_btf = map_check_no_btf,
561 static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
562 struct xdp_bulk_queue *bq, bool in_napi_ctx)
564 unsigned int processed = 0, drops = 0;
565 const int to_cpu = rcpu->cpu;
569 if (unlikely(!bq->count))
573 spin_lock(&q->producer_lock);
575 for (i = 0; i < bq->count; i++) {
576 struct xdp_frame *xdpf = bq->q[i];
579 err = __ptr_ring_produce(q, xdpf);
582 if (likely(in_napi_ctx))
583 xdp_return_frame_rx_napi(xdpf);
585 xdp_return_frame(xdpf);
590 spin_unlock(&q->producer_lock);
592 /* Feedback loop via tracepoints */
593 trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
597 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
598 * Thus, safe percpu variable access.
600 static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
602 struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
604 if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
605 bq_flush_to_queue(rcpu, bq, true);
607 /* Notice, xdp_buff/page MUST be queued here, long enough for
608 * driver to code invoking us to finished, due to driver
609 * (e.g. ixgbe) recycle tricks based on page-refcnt.
611 * Thus, incoming xdp_frame is always queued here (else we race
612 * with another CPU on page-refcnt and remaining driver code).
613 * Queue time is very short, as driver will invoke flush
614 * operation, when completing napi->poll call.
616 bq->q[bq->count++] = xdpf;
620 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
621 struct net_device *dev_rx)
623 struct xdp_frame *xdpf;
625 xdpf = convert_to_xdp_frame(xdp);
629 /* Info needed when constructing SKB on remote CPU */
630 xdpf->dev_rx = dev_rx;
632 bq_enqueue(rcpu, xdpf);
636 void __cpu_map_insert_ctx(struct bpf_map *map, u32 bit)
638 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
639 unsigned long *bitmap = this_cpu_ptr(cmap->flush_needed);
641 __set_bit(bit, bitmap);
644 void __cpu_map_flush(struct bpf_map *map)
646 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
647 unsigned long *bitmap = this_cpu_ptr(cmap->flush_needed);
650 /* The napi->poll softirq makes sure __cpu_map_insert_ctx()
651 * and __cpu_map_flush() happen on same CPU. Thus, the percpu
652 * bitmap indicate which percpu bulkq have packets.
654 for_each_set_bit(bit, bitmap, map->max_entries) {
655 struct bpf_cpu_map_entry *rcpu = READ_ONCE(cmap->cpu_map[bit]);
656 struct xdp_bulk_queue *bq;
658 /* This is possible if entry is removed by user space
659 * between xdp redirect and flush op.
664 __clear_bit(bit, bitmap);
666 /* Flush all frames in bulkq to real queue */
667 bq = this_cpu_ptr(rcpu->bulkq);
668 bq_flush_to_queue(rcpu, bq, true);
670 /* If already running, costs spin_lock_irqsave + smb_mb */
671 wake_up_process(rcpu->kthread);