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>
23 #include <linux/sched.h>
24 #include <linux/workqueue.h>
25 #include <linux/kthread.h>
26 #include <linux/capability.h>
28 /* General idea: XDP packets getting XDP redirected to another CPU,
29 * will maximum be stored/queued for one driver ->poll() call. It is
30 * guaranteed that setting flush bit and flush operation happen on
31 * same CPU. Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
32 * which queue in bpf_cpu_map_entry contains packets.
35 #define CPU_MAP_BULK_SIZE 8 /* 8 == one cacheline on 64-bit archs */
36 struct xdp_bulk_queue {
37 void *q[CPU_MAP_BULK_SIZE];
41 /* Struct for every remote "destination" CPU in map */
42 struct bpf_cpu_map_entry {
43 u32 qsize; /* Queue size placeholder for map lookup */
45 /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
46 struct xdp_bulk_queue __percpu *bulkq;
48 /* Queue with potential multi-producers, and single-consumer kthread */
49 struct ptr_ring *queue;
50 struct task_struct *kthread;
51 struct work_struct kthread_stop_wq;
53 atomic_t refcnt; /* Control when this struct can be free'ed */
59 /* Below members specific for map type */
60 struct bpf_cpu_map_entry **cpu_map;
61 unsigned long __percpu *flush_needed;
64 static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
65 struct xdp_bulk_queue *bq);
67 static u64 cpu_map_bitmap_size(const union bpf_attr *attr)
69 return BITS_TO_LONGS(attr->max_entries) * sizeof(unsigned long);
72 static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
74 struct bpf_cpu_map *cmap;
79 if (!capable(CAP_SYS_ADMIN))
80 return ERR_PTR(-EPERM);
82 /* check sanity of attributes */
83 if (attr->max_entries == 0 || attr->key_size != 4 ||
84 attr->value_size != 4 || attr->map_flags & ~BPF_F_NUMA_NODE)
85 return ERR_PTR(-EINVAL);
87 cmap = kzalloc(sizeof(*cmap), GFP_USER);
89 return ERR_PTR(-ENOMEM);
91 /* mandatory map attributes */
92 cmap->map.map_type = attr->map_type;
93 cmap->map.key_size = attr->key_size;
94 cmap->map.value_size = attr->value_size;
95 cmap->map.max_entries = attr->max_entries;
96 cmap->map.map_flags = attr->map_flags;
97 cmap->map.numa_node = bpf_map_attr_numa_node(attr);
99 /* Pre-limit array size based on NR_CPUS, not final CPU check */
100 if (cmap->map.max_entries > NR_CPUS) {
105 /* make sure page count doesn't overflow */
106 cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *);
107 cost += cpu_map_bitmap_size(attr) * num_possible_cpus();
108 if (cost >= U32_MAX - PAGE_SIZE)
110 cmap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
112 /* Notice returns -EPERM on if map size is larger than memlock limit */
113 ret = bpf_map_precharge_memlock(cmap->map.pages);
119 /* A per cpu bitfield with a bit per possible CPU in map */
120 cmap->flush_needed = __alloc_percpu(cpu_map_bitmap_size(attr),
121 __alignof__(unsigned long));
122 if (!cmap->flush_needed)
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_needed);
140 void __cpu_map_queue_destructor(void *ptr)
142 /* The tear-down procedure should have made sure that queue is
143 * empty. See __cpu_map_entry_replace() and work-queue
144 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
145 * gracefully and warn once.
147 if (WARN_ON_ONCE(ptr))
151 static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
153 if (atomic_dec_and_test(&rcpu->refcnt)) {
154 /* The queue should be empty at this point */
155 ptr_ring_cleanup(rcpu->queue, __cpu_map_queue_destructor);
161 static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
163 atomic_inc(&rcpu->refcnt);
166 /* called from workqueue, to workaround syscall using preempt_disable */
167 static void cpu_map_kthread_stop(struct work_struct *work)
169 struct bpf_cpu_map_entry *rcpu;
171 rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
173 /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
174 * as it waits until all in-flight call_rcu() callbacks complete.
178 /* kthread_stop will wake_up_process and wait for it to complete */
179 kthread_stop(rcpu->kthread);
182 static int cpu_map_kthread_run(void *data)
184 struct bpf_cpu_map_entry *rcpu = data;
186 set_current_state(TASK_INTERRUPTIBLE);
188 /* When kthread gives stop order, then rcpu have been disconnected
189 * from map, thus no new packets can enter. Remaining in-flight
190 * per CPU stored packets are flushed to this queue. Wait honoring
191 * kthread_stop signal until queue is empty.
193 while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
194 struct xdp_pkt *xdp_pkt;
198 while ((xdp_pkt = ptr_ring_consume(rcpu->queue))) {
199 /* For now just "refcnt-free" */
200 page_frag_free(xdp_pkt);
202 __set_current_state(TASK_INTERRUPTIBLE);
204 __set_current_state(TASK_RUNNING);
206 put_cpu_map_entry(rcpu);
210 struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu, int map_id)
212 gfp_t gfp = GFP_ATOMIC|__GFP_NOWARN;
213 struct bpf_cpu_map_entry *rcpu;
216 /* Have map->numa_node, but choose node of redirect target CPU */
217 numa = cpu_to_node(cpu);
219 rcpu = kzalloc_node(sizeof(*rcpu), gfp, numa);
223 /* Alloc percpu bulkq */
224 rcpu->bulkq = __alloc_percpu_gfp(sizeof(*rcpu->bulkq),
225 sizeof(void *), gfp);
230 rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
234 err = ptr_ring_init(rcpu->queue, qsize, gfp);
241 rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
242 "cpumap/%d/map:%d", cpu, map_id);
243 if (IS_ERR(rcpu->kthread))
246 get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
247 get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
249 /* Make sure kthread runs on a single CPU */
250 kthread_bind(rcpu->kthread, cpu);
251 wake_up_process(rcpu->kthread);
256 ptr_ring_cleanup(rcpu->queue, NULL);
260 free_percpu(rcpu->bulkq);
266 void __cpu_map_entry_free(struct rcu_head *rcu)
268 struct bpf_cpu_map_entry *rcpu;
271 /* This cpu_map_entry have been disconnected from map and one
272 * RCU graze-period have elapsed. Thus, XDP cannot queue any
273 * new packets and cannot change/set flush_needed that can
276 rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
278 /* Flush remaining packets in percpu bulkq */
279 for_each_online_cpu(cpu) {
280 struct xdp_bulk_queue *bq = per_cpu_ptr(rcpu->bulkq, cpu);
282 /* No concurrent bq_enqueue can run at this point */
283 bq_flush_to_queue(rcpu, bq);
285 free_percpu(rcpu->bulkq);
286 /* Cannot kthread_stop() here, last put free rcpu resources */
287 put_cpu_map_entry(rcpu);
290 /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
291 * ensure any driver rcu critical sections have completed, but this
292 * does not guarantee a flush has happened yet. Because driver side
293 * rcu_read_lock/unlock only protects the running XDP program. The
294 * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
295 * pending flush op doesn't fail.
297 * The bpf_cpu_map_entry is still used by the kthread, and there can
298 * still be pending packets (in queue and percpu bulkq). A refcnt
299 * makes sure to last user (kthread_stop vs. call_rcu) free memory
302 * The rcu callback __cpu_map_entry_free flush remaining packets in
303 * percpu bulkq to queue. Due to caller map_delete_elem() disable
304 * preemption, cannot call kthread_stop() to make sure queue is empty.
305 * Instead a work_queue is started for stopping kthread,
306 * cpu_map_kthread_stop, which waits for an RCU graze period before
307 * stopping kthread, emptying the queue.
309 void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
310 u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
312 struct bpf_cpu_map_entry *old_rcpu;
314 old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu);
316 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
317 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
318 schedule_work(&old_rcpu->kthread_stop_wq);
322 int cpu_map_delete_elem(struct bpf_map *map, void *key)
324 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
325 u32 key_cpu = *(u32 *)key;
327 if (key_cpu >= map->max_entries)
330 /* notice caller map_delete_elem() use preempt_disable() */
331 __cpu_map_entry_replace(cmap, key_cpu, NULL);
335 int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
338 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
339 struct bpf_cpu_map_entry *rcpu;
341 /* Array index key correspond to CPU number */
342 u32 key_cpu = *(u32 *)key;
343 /* Value is the queue size */
344 u32 qsize = *(u32 *)value;
346 if (unlikely(map_flags > BPF_EXIST))
348 if (unlikely(key_cpu >= cmap->map.max_entries))
350 if (unlikely(map_flags == BPF_NOEXIST))
352 if (unlikely(qsize > 16384)) /* sanity limit on qsize */
355 /* Make sure CPU is a valid possible cpu */
356 if (!cpu_possible(key_cpu))
360 rcpu = NULL; /* Same as deleting */
362 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
363 rcpu = __cpu_map_entry_alloc(qsize, key_cpu, map->id);
368 __cpu_map_entry_replace(cmap, key_cpu, rcpu);
373 void cpu_map_free(struct bpf_map *map)
375 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
379 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
380 * so the bpf programs (can be more than one that used this map) were
381 * disconnected from events. Wait for outstanding critical sections in
382 * these programs to complete. The rcu critical section only guarantees
383 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
384 * It does __not__ ensure pending flush operations (if any) are
389 /* To ensure all pending flush operations have completed wait for flush
390 * bitmap to indicate all flush_needed bits to be zero on _all_ cpus.
391 * Because the above synchronize_rcu() ensures the map is disconnected
392 * from the program we can assume no new bits will be set.
394 for_each_online_cpu(cpu) {
395 unsigned long *bitmap = per_cpu_ptr(cmap->flush_needed, cpu);
397 while (!bitmap_empty(bitmap, cmap->map.max_entries))
401 /* For cpu_map the remote CPUs can still be using the entries
402 * (struct bpf_cpu_map_entry).
404 for (i = 0; i < cmap->map.max_entries; i++) {
405 struct bpf_cpu_map_entry *rcpu;
407 rcpu = READ_ONCE(cmap->cpu_map[i]);
411 /* bq flush and cleanup happens after RCU graze-period */
412 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
414 free_percpu(cmap->flush_needed);
415 bpf_map_area_free(cmap->cpu_map);
419 struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
421 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
422 struct bpf_cpu_map_entry *rcpu;
424 if (key >= map->max_entries)
427 rcpu = READ_ONCE(cmap->cpu_map[key]);
431 static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
433 struct bpf_cpu_map_entry *rcpu =
434 __cpu_map_lookup_elem(map, *(u32 *)key);
436 return rcpu ? &rcpu->qsize : NULL;
439 static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
441 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
442 u32 index = key ? *(u32 *)key : U32_MAX;
443 u32 *next = next_key;
445 if (index >= cmap->map.max_entries) {
450 if (index == cmap->map.max_entries - 1)
456 const struct bpf_map_ops cpu_map_ops = {
457 .map_alloc = cpu_map_alloc,
458 .map_free = cpu_map_free,
459 .map_delete_elem = cpu_map_delete_elem,
460 .map_update_elem = cpu_map_update_elem,
461 .map_lookup_elem = cpu_map_lookup_elem,
462 .map_get_next_key = cpu_map_get_next_key,
465 static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
466 struct xdp_bulk_queue *bq)
471 if (unlikely(!bq->count))
475 spin_lock(&q->producer_lock);
477 for (i = 0; i < bq->count; i++) {
478 void *xdp_pkt = bq->q[i];
481 err = __ptr_ring_produce(q, xdp_pkt);
484 page_frag_free(xdp_pkt);
488 spin_unlock(&q->producer_lock);
493 /* Notice: Will change in later patch */
500 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
501 * Thus, safe percpu variable access.
503 static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_pkt *xdp_pkt)
505 struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
507 if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
508 bq_flush_to_queue(rcpu, bq);
510 /* Notice, xdp_buff/page MUST be queued here, long enough for
511 * driver to code invoking us to finished, due to driver
512 * (e.g. ixgbe) recycle tricks based on page-refcnt.
514 * Thus, incoming xdp_pkt is always queued here (else we race
515 * with another CPU on page-refcnt and remaining driver code).
516 * Queue time is very short, as driver will invoke flush
517 * operation, when completing napi->poll call.
519 bq->q[bq->count++] = xdp_pkt;
523 int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
524 struct net_device *dev_rx)
526 struct xdp_pkt *xdp_pkt;
529 /* For now this is just used as a void pointer to data_hard_start.
530 * Followup patch will generalize this.
532 xdp_pkt = xdp->data_hard_start;
534 /* Fake writing into xdp_pkt->data to measure overhead */
535 headroom = xdp->data - xdp->data_hard_start;
536 if (headroom < sizeof(*xdp_pkt))
537 xdp_pkt->data = xdp->data;
539 bq_enqueue(rcpu, xdp_pkt);
543 void __cpu_map_insert_ctx(struct bpf_map *map, u32 bit)
545 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
546 unsigned long *bitmap = this_cpu_ptr(cmap->flush_needed);
548 __set_bit(bit, bitmap);
551 void __cpu_map_flush(struct bpf_map *map)
553 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
554 unsigned long *bitmap = this_cpu_ptr(cmap->flush_needed);
557 /* The napi->poll softirq makes sure __cpu_map_insert_ctx()
558 * and __cpu_map_flush() happen on same CPU. Thus, the percpu
559 * bitmap indicate which percpu bulkq have packets.
561 for_each_set_bit(bit, bitmap, map->max_entries) {
562 struct bpf_cpu_map_entry *rcpu = READ_ONCE(cmap->cpu_map[bit]);
563 struct xdp_bulk_queue *bq;
565 /* This is possible if entry is removed by user space
566 * between xdp redirect and flush op.
571 __clear_bit(bit, bitmap);
573 /* Flush all frames in bulkq to real queue */
574 bq = this_cpu_ptr(rcpu->bulkq);
575 bq_flush_to_queue(rcpu, bq);
577 /* If already running, costs spin_lock_irqsave + smb_mb */
578 wake_up_process(rcpu->kthread);