blob: 04c8eb11cd90bac6b6a9d91942d3df8c54a96306 [file] [log] [blame]
Thomas Gleixnerddc64d02019-05-31 01:09:24 -07001// SPDX-License-Identifier: GPL-2.0-only
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +02002/* bpf/cpumap.c
3 *
4 * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +02005 */
6
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'.
9 *
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.
13 *
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.
18 */
19#include <linux/bpf.h>
20#include <linux/filter.h>
21#include <linux/ptr_ring.h>
Jesper Dangaard Brouer5ab073f2018-04-17 16:45:26 +020022#include <net/xdp.h>
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020023
24#include <linux/sched.h>
25#include <linux/workqueue.h>
26#include <linux/kthread.h>
27#include <linux/capability.h>
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +020028#include <trace/events/xdp.h>
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020029
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +020030#include <linux/netdevice.h> /* netif_receive_skb_core */
31#include <linux/etherdevice.h> /* eth_type_trans */
32
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020033/* General idea: XDP packets getting XDP redirected to another CPU,
34 * will maximum be stored/queued for one driver ->poll() call. It is
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +020035 * guaranteed that queueing the frame and the flush operation happen on
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020036 * same CPU. Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
37 * which queue in bpf_cpu_map_entry contains packets.
38 */
39
40#define CPU_MAP_BULK_SIZE 8 /* 8 == one cacheline on 64-bit archs */
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +020041struct bpf_cpu_map_entry;
42struct bpf_cpu_map;
43
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020044struct xdp_bulk_queue {
45 void *q[CPU_MAP_BULK_SIZE];
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +020046 struct list_head flush_node;
47 struct bpf_cpu_map_entry *obj;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020048 unsigned int count;
49};
50
51/* Struct for every remote "destination" CPU in map */
52struct bpf_cpu_map_entry {
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +020053 u32 cpu; /* kthread CPU and map index */
54 int map_id; /* Back reference to map */
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020055 u32 qsize; /* Queue size placeholder for map lookup */
56
57 /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
58 struct xdp_bulk_queue __percpu *bulkq;
59
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +020060 struct bpf_cpu_map *cmap;
61
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020062 /* 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;
66
67 atomic_t refcnt; /* Control when this struct can be free'ed */
68 struct rcu_head rcu;
69};
70
71struct bpf_cpu_map {
72 struct bpf_map map;
73 /* Below members specific for map type */
74 struct bpf_cpu_map_entry **cpu_map;
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +020075 struct list_head __percpu *flush_list;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020076};
77
Björn Töpel4bc188c2019-12-19 07:10:00 +010078static int bq_flush_to_queue(struct xdp_bulk_queue *bq);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020079
80static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
81{
82 struct bpf_cpu_map *cmap;
83 int err = -ENOMEM;
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +020084 int ret, cpu;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020085 u64 cost;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020086
87 if (!capable(CAP_SYS_ADMIN))
88 return ERR_PTR(-EPERM);
89
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);
94
95 cmap = kzalloc(sizeof(*cmap), GFP_USER);
96 if (!cmap)
97 return ERR_PTR(-ENOMEM);
98
Jakub Kicinskibd475642018-01-11 20:29:06 -080099 bpf_map_init_from_attr(&cmap->map, attr);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200100
101 /* Pre-limit array size based on NR_CPUS, not final CPU check */
102 if (cmap->map.max_entries > NR_CPUS) {
103 err = -E2BIG;
104 goto free_cmap;
105 }
106
107 /* make sure page count doesn't overflow */
108 cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *);
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200109 cost += sizeof(struct list_head) * num_possible_cpus();
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200110
111 /* Notice returns -EPERM on if map size is larger than memlock limit */
Roman Gushchinc85d6912019-05-29 18:03:59 -0700112 ret = bpf_map_charge_init(&cmap->map.memory, cost);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200113 if (ret) {
114 err = ret;
115 goto free_cmap;
116 }
117
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200118 cmap->flush_list = alloc_percpu(struct list_head);
119 if (!cmap->flush_list)
Roman Gushchinb936ca62019-05-29 18:03:58 -0700120 goto free_charge;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200121
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200122 for_each_possible_cpu(cpu)
123 INIT_LIST_HEAD(per_cpu_ptr(cmap->flush_list, cpu));
124
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200125 /* 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);
129 if (!cmap->cpu_map)
130 goto free_percpu;
131
132 return &cmap->map;
133free_percpu:
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200134 free_percpu(cmap->flush_list);
Roman Gushchinb936ca62019-05-29 18:03:58 -0700135free_charge:
136 bpf_map_charge_finish(&cmap->map.memory);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200137free_cmap:
138 kfree(cmap);
139 return ERR_PTR(err);
140}
141
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200142static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
143{
144 atomic_inc(&rcpu->refcnt);
145}
146
147/* called from workqueue, to workaround syscall using preempt_disable */
148static void cpu_map_kthread_stop(struct work_struct *work)
149{
150 struct bpf_cpu_map_entry *rcpu;
151
152 rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
153
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.
156 */
157 rcu_barrier();
158
159 /* kthread_stop will wake_up_process and wait for it to complete */
160 kthread_stop(rcpu->kthread);
161}
162
Wei Yongjun0fe875c2018-01-16 11:27:05 +0000163static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu,
Jesper Dangaard Brouer8f0504a92019-04-12 17:07:43 +0200164 struct xdp_frame *xdpf,
165 struct sk_buff *skb)
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200166{
Jesper Dangaard Brouer676e4a62019-03-29 10:18:00 +0100167 unsigned int hard_start_headroom;
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200168 unsigned int frame_size;
169 void *pkt_data_start;
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200170
Jesper Dangaard Brouer676e4a62019-03-29 10:18:00 +0100171 /* Part of headroom was reserved to xdpf */
172 hard_start_headroom = sizeof(struct xdp_frame) + xdpf->headroom;
173
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200174 /* 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.
177 *
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.
182 * frame_size = 2048;
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200183 * frame_len = frame_size - sizeof(*xdp_frame);
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200184 *
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.
190 */
Jesper Dangaard Brouer676e4a62019-03-29 10:18:00 +0100191 frame_size = SKB_DATA_ALIGN(xdpf->len + hard_start_headroom) +
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200192 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
193
Jesper Dangaard Brouer676e4a62019-03-29 10:18:00 +0100194 pkt_data_start = xdpf->data - hard_start_headroom;
Jesper Dangaard Brouer8f0504a92019-04-12 17:07:43 +0200195 skb = build_skb_around(skb, pkt_data_start, frame_size);
196 if (unlikely(!skb))
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200197 return NULL;
198
Jesper Dangaard Brouer676e4a62019-03-29 10:18:00 +0100199 skb_reserve(skb, hard_start_headroom);
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200200 __skb_put(skb, xdpf->len);
201 if (xdpf->metasize)
202 skb_metadata_set(skb, xdpf->metasize);
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200203
204 /* Essential SKB info: protocol and skb->dev */
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200205 skb->protocol = eth_type_trans(skb, xdpf->dev_rx);
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200206
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)
211 */
212
Jesper Dangaard Brouer6bf071b2019-06-18 15:05:27 +0200213 /* Until page_pool get SKB return path, release DMA here */
214 xdp_release_frame(xdpf);
215
Jesper Dangaard Brouer676e4a62019-03-29 10:18:00 +0100216 /* Allow SKB to reuse area used by xdp_frame */
217 xdp_scrub_frame(xdpf);
218
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200219 return skb;
220}
221
Jesper Dangaard Brouer5ab073f2018-04-17 16:45:26 +0200222static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
223{
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.
228 */
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200229 struct xdp_frame *xdpf;
Jesper Dangaard Brouer5ab073f2018-04-17 16:45:26 +0200230
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200231 while ((xdpf = ptr_ring_consume(ring)))
232 if (WARN_ON_ONCE(xdpf))
Jesper Dangaard Brouer03993092018-04-17 16:46:32 +0200233 xdp_return_frame(xdpf);
Jesper Dangaard Brouer5ab073f2018-04-17 16:45:26 +0200234}
235
236static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
237{
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);
242 kfree(rcpu->queue);
243 kfree(rcpu);
244 }
245}
246
Jesper Dangaard Brouer77361822019-04-12 17:07:32 +0200247#define CPUMAP_BATCH 8
248
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200249static int cpu_map_kthread_run(void *data)
250{
251 struct bpf_cpu_map_entry *rcpu = data;
252
253 set_current_state(TASK_INTERRUPTIBLE);
254
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.
259 */
260 while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
Jesper Dangaard Brouer77361822019-04-12 17:07:32 +0200261 unsigned int drops = 0, sched = 0;
262 void *frames[CPUMAP_BATCH];
Jesper Dangaard Brouer8f0504a92019-04-12 17:07:43 +0200263 void *skbs[CPUMAP_BATCH];
264 gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
265 int i, n, m;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200266
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200267 /* Release CPU reschedule checks */
268 if (__ptr_ring_empty(rcpu->queue)) {
Jesper Dangaard Brouer31749462017-10-23 19:39:28 +0200269 set_current_state(TASK_INTERRUPTIBLE);
270 /* Recheck to avoid lost wake-up */
271 if (__ptr_ring_empty(rcpu->queue)) {
272 schedule();
273 sched = 1;
274 } else {
275 __set_current_state(TASK_RUNNING);
276 }
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200277 } else {
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200278 sched = cond_resched();
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200279 }
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200280
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200281 /*
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.
285 */
Jesper Dangaard Brouer77361822019-04-12 17:07:32 +0200286 n = ptr_ring_consume_batched(rcpu->queue, frames, CPUMAP_BATCH);
Jesper Dangaard Brouer86d23142019-04-12 17:07:48 +0200287
288 for (i = 0; i < n; i++) {
289 void *f = frames[i];
290 struct page *page = virt_to_page(f);
291
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.
295 */
296 prefetchw(page);
297 }
298
Jesper Dangaard Brouer8f0504a92019-04-12 17:07:43 +0200299 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 */
303 drops = n;
304 }
Jesper Dangaard Brouer77361822019-04-12 17:07:32 +0200305
306 local_bh_disable();
307 for (i = 0; i < n; i++) {
308 struct xdp_frame *xdpf = frames[i];
Jesper Dangaard Brouer8f0504a92019-04-12 17:07:43 +0200309 struct sk_buff *skb = skbs[i];
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200310 int ret;
311
Jesper Dangaard Brouer8f0504a92019-04-12 17:07:43 +0200312 skb = cpu_map_build_skb(rcpu, xdpf, skb);
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200313 if (!skb) {
Jesper Dangaard Brouer03993092018-04-17 16:46:32 +0200314 xdp_return_frame(xdpf);
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200315 continue;
316 }
317
318 /* Inject into network stack */
319 ret = netif_receive_skb_core(skb);
320 if (ret == NET_RX_DROP)
321 drops++;
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200322 }
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200323 /* Feedback loop via tracepoint */
Jesper Dangaard Brouer77361822019-04-12 17:07:32 +0200324 trace_xdp_cpumap_kthread(rcpu->map_id, n, drops, sched);
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200325
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200326 local_bh_enable(); /* resched point, may call do_softirq() */
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200327 }
328 __set_current_state(TASK_RUNNING);
329
330 put_cpu_map_entry(rcpu);
331 return 0;
332}
333
Wei Yongjun0fe875c2018-01-16 11:27:05 +0000334static struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu,
335 int map_id)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200336{
Jason Wang7fc17e92018-02-14 22:17:34 +0800337 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200338 struct bpf_cpu_map_entry *rcpu;
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200339 struct xdp_bulk_queue *bq;
340 int numa, err, i;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200341
342 /* Have map->numa_node, but choose node of redirect target CPU */
343 numa = cpu_to_node(cpu);
344
345 rcpu = kzalloc_node(sizeof(*rcpu), gfp, numa);
346 if (!rcpu)
347 return NULL;
348
349 /* Alloc percpu bulkq */
350 rcpu->bulkq = __alloc_percpu_gfp(sizeof(*rcpu->bulkq),
351 sizeof(void *), gfp);
352 if (!rcpu->bulkq)
353 goto free_rcu;
354
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200355 for_each_possible_cpu(i) {
356 bq = per_cpu_ptr(rcpu->bulkq, i);
357 bq->obj = rcpu;
358 }
359
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200360 /* Alloc queue */
361 rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
362 if (!rcpu->queue)
363 goto free_bulkq;
364
365 err = ptr_ring_init(rcpu->queue, qsize, gfp);
366 if (err)
367 goto free_queue;
368
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200369 rcpu->cpu = cpu;
370 rcpu->map_id = map_id;
371 rcpu->qsize = qsize;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200372
373 /* Setup kthread */
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))
377 goto free_ptr_ring;
378
379 get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
380 get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
381
382 /* Make sure kthread runs on a single CPU */
383 kthread_bind(rcpu->kthread, cpu);
384 wake_up_process(rcpu->kthread);
385
386 return rcpu;
387
388free_ptr_ring:
389 ptr_ring_cleanup(rcpu->queue, NULL);
390free_queue:
391 kfree(rcpu->queue);
392free_bulkq:
393 free_percpu(rcpu->bulkq);
394free_rcu:
395 kfree(rcpu);
396 return NULL;
397}
398
Wei Yongjun0fe875c2018-01-16 11:27:05 +0000399static void __cpu_map_entry_free(struct rcu_head *rcu)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200400{
401 struct bpf_cpu_map_entry *rcpu;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200402
403 /* This cpu_map_entry have been disconnected from map and one
404 * RCU graze-period have elapsed. Thus, XDP cannot queue any
405 * new packets and cannot change/set flush_needed that can
406 * find this entry.
407 */
408 rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
409
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200410 free_percpu(rcpu->bulkq);
411 /* Cannot kthread_stop() here, last put free rcpu resources */
412 put_cpu_map_entry(rcpu);
413}
414
415/* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
416 * ensure any driver rcu critical sections have completed, but this
417 * does not guarantee a flush has happened yet. Because driver side
418 * rcu_read_lock/unlock only protects the running XDP program. The
419 * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
420 * pending flush op doesn't fail.
421 *
422 * The bpf_cpu_map_entry is still used by the kthread, and there can
423 * still be pending packets (in queue and percpu bulkq). A refcnt
424 * makes sure to last user (kthread_stop vs. call_rcu) free memory
425 * resources.
426 *
427 * The rcu callback __cpu_map_entry_free flush remaining packets in
428 * percpu bulkq to queue. Due to caller map_delete_elem() disable
429 * preemption, cannot call kthread_stop() to make sure queue is empty.
430 * Instead a work_queue is started for stopping kthread,
431 * cpu_map_kthread_stop, which waits for an RCU graze period before
432 * stopping kthread, emptying the queue.
433 */
Wei Yongjun0fe875c2018-01-16 11:27:05 +0000434static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
435 u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200436{
437 struct bpf_cpu_map_entry *old_rcpu;
438
439 old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu);
440 if (old_rcpu) {
441 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
442 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
443 schedule_work(&old_rcpu->kthread_stop_wq);
444 }
445}
446
Wei Yongjun0fe875c2018-01-16 11:27:05 +0000447static int cpu_map_delete_elem(struct bpf_map *map, void *key)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200448{
449 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
450 u32 key_cpu = *(u32 *)key;
451
452 if (key_cpu >= map->max_entries)
453 return -EINVAL;
454
455 /* notice caller map_delete_elem() use preempt_disable() */
456 __cpu_map_entry_replace(cmap, key_cpu, NULL);
457 return 0;
458}
459
Wei Yongjun0fe875c2018-01-16 11:27:05 +0000460static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
461 u64 map_flags)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200462{
463 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
464 struct bpf_cpu_map_entry *rcpu;
465
466 /* Array index key correspond to CPU number */
467 u32 key_cpu = *(u32 *)key;
468 /* Value is the queue size */
469 u32 qsize = *(u32 *)value;
470
471 if (unlikely(map_flags > BPF_EXIST))
472 return -EINVAL;
473 if (unlikely(key_cpu >= cmap->map.max_entries))
474 return -E2BIG;
475 if (unlikely(map_flags == BPF_NOEXIST))
476 return -EEXIST;
477 if (unlikely(qsize > 16384)) /* sanity limit on qsize */
478 return -EOVERFLOW;
479
480 /* Make sure CPU is a valid possible cpu */
481 if (!cpu_possible(key_cpu))
482 return -ENODEV;
483
484 if (qsize == 0) {
485 rcpu = NULL; /* Same as deleting */
486 } else {
487 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
488 rcpu = __cpu_map_entry_alloc(qsize, key_cpu, map->id);
489 if (!rcpu)
490 return -ENOMEM;
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200491 rcpu->cmap = cmap;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200492 }
493 rcu_read_lock();
494 __cpu_map_entry_replace(cmap, key_cpu, rcpu);
495 rcu_read_unlock();
496 return 0;
497}
498
Wei Yongjun0fe875c2018-01-16 11:27:05 +0000499static void cpu_map_free(struct bpf_map *map)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200500{
501 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200502 u32 i;
503
504 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
505 * so the bpf programs (can be more than one that used this map) were
506 * disconnected from events. Wait for outstanding critical sections in
507 * these programs to complete. The rcu critical section only guarantees
508 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
509 * It does __not__ ensure pending flush operations (if any) are
510 * complete.
511 */
Daniel Borkmannf6069b92018-08-17 23:26:14 +0200512
513 bpf_clear_redirect_map(map);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200514 synchronize_rcu();
515
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200516 /* For cpu_map the remote CPUs can still be using the entries
517 * (struct bpf_cpu_map_entry).
518 */
519 for (i = 0; i < cmap->map.max_entries; i++) {
520 struct bpf_cpu_map_entry *rcpu;
521
522 rcpu = READ_ONCE(cmap->cpu_map[i]);
523 if (!rcpu)
524 continue;
525
526 /* bq flush and cleanup happens after RCU graze-period */
527 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
528 }
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200529 free_percpu(cmap->flush_list);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200530 bpf_map_area_free(cmap->cpu_map);
531 kfree(cmap);
532}
533
534struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
535{
536 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
537 struct bpf_cpu_map_entry *rcpu;
538
539 if (key >= map->max_entries)
540 return NULL;
541
542 rcpu = READ_ONCE(cmap->cpu_map[key]);
543 return rcpu;
544}
545
546static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
547{
548 struct bpf_cpu_map_entry *rcpu =
549 __cpu_map_lookup_elem(map, *(u32 *)key);
550
551 return rcpu ? &rcpu->qsize : NULL;
552}
553
554static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
555{
556 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
557 u32 index = key ? *(u32 *)key : U32_MAX;
558 u32 *next = next_key;
559
560 if (index >= cmap->map.max_entries) {
561 *next = 0;
562 return 0;
563 }
564
565 if (index == cmap->map.max_entries - 1)
566 return -ENOENT;
567 *next = index + 1;
568 return 0;
569}
570
571const struct bpf_map_ops cpu_map_ops = {
572 .map_alloc = cpu_map_alloc,
573 .map_free = cpu_map_free,
574 .map_delete_elem = cpu_map_delete_elem,
575 .map_update_elem = cpu_map_update_elem,
576 .map_lookup_elem = cpu_map_lookup_elem,
577 .map_get_next_key = cpu_map_get_next_key,
Daniel Borkmanne8d2bec2018-08-12 01:59:17 +0200578 .map_check_btf = map_check_no_btf,
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200579};
580
Björn Töpel4bc188c2019-12-19 07:10:00 +0100581static int bq_flush_to_queue(struct xdp_bulk_queue *bq)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200582{
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200583 struct bpf_cpu_map_entry *rcpu = bq->obj;
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200584 unsigned int processed = 0, drops = 0;
585 const int to_cpu = rcpu->cpu;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200586 struct ptr_ring *q;
587 int i;
588
589 if (unlikely(!bq->count))
590 return 0;
591
592 q = rcpu->queue;
593 spin_lock(&q->producer_lock);
594
595 for (i = 0; i < bq->count; i++) {
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200596 struct xdp_frame *xdpf = bq->q[i];
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200597 int err;
598
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200599 err = __ptr_ring_produce(q, xdpf);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200600 if (err) {
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200601 drops++;
Björn Töpel4bc188c2019-12-19 07:10:00 +0100602 xdp_return_frame_rx_napi(xdpf);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200603 }
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200604 processed++;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200605 }
606 bq->count = 0;
607 spin_unlock(&q->producer_lock);
608
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200609 __list_del_clearprev(&bq->flush_node);
610
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200611 /* Feedback loop via tracepoints */
612 trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200613 return 0;
614}
615
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200616/* Runs under RCU-read-side, plus in softirq under NAPI protection.
617 * Thus, safe percpu variable access.
618 */
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200619static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200620{
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200621 struct list_head *flush_list = this_cpu_ptr(rcpu->cmap->flush_list);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200622 struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
623
624 if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
Björn Töpel4bc188c2019-12-19 07:10:00 +0100625 bq_flush_to_queue(bq);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200626
627 /* Notice, xdp_buff/page MUST be queued here, long enough for
628 * driver to code invoking us to finished, due to driver
629 * (e.g. ixgbe) recycle tricks based on page-refcnt.
630 *
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200631 * Thus, incoming xdp_frame is always queued here (else we race
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200632 * with another CPU on page-refcnt and remaining driver code).
633 * Queue time is very short, as driver will invoke flush
634 * operation, when completing napi->poll call.
635 */
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200636 bq->q[bq->count++] = xdpf;
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200637
638 if (!bq->flush_node.prev)
639 list_add(&bq->flush_node, flush_list);
640
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200641 return 0;
642}
643
Jesper Dangaard Brouer9c270af2017-10-16 12:19:34 +0200644int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
645 struct net_device *dev_rx)
646{
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200647 struct xdp_frame *xdpf;
Jesper Dangaard Brouer9c270af2017-10-16 12:19:34 +0200648
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200649 xdpf = convert_to_xdp_frame(xdp);
650 if (unlikely(!xdpf))
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200651 return -EOVERFLOW;
Jesper Dangaard Brouer9c270af2017-10-16 12:19:34 +0200652
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200653 /* Info needed when constructing SKB on remote CPU */
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200654 xdpf->dev_rx = dev_rx;
Jesper Dangaard Brouer9c270af2017-10-16 12:19:34 +0200655
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200656 bq_enqueue(rcpu, xdpf);
Jesper Dangaard Brouer9c270af2017-10-16 12:19:34 +0200657 return 0;
658}
659
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200660void __cpu_map_flush(struct bpf_map *map)
661{
662 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200663 struct list_head *flush_list = this_cpu_ptr(cmap->flush_list);
664 struct xdp_bulk_queue *bq, *tmp;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200665
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200666 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
Björn Töpel4bc188c2019-12-19 07:10:00 +0100667 bq_flush_to_queue(bq);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200668
669 /* If already running, costs spin_lock_irqsave + smb_mb */
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200670 wake_up_process(bq->obj->kthread);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200671 }
672}