blob: 7473136981787155acecc7c485f6fd3ba0200411 [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
56 /* XDP can run multiple RX-ring queues, need __percpu enqueue store */
57 struct xdp_bulk_queue __percpu *bulkq;
58
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +020059 struct bpf_cpu_map *cmap;
60
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020061 /* Queue with potential multi-producers, and single-consumer kthread */
62 struct ptr_ring *queue;
63 struct task_struct *kthread;
Lorenzo Bianconi644bfe52020-07-14 15:56:37 +020064
65 struct bpf_cpumap_val value;
Lorenzo Bianconi92164772020-07-14 15:56:38 +020066 struct bpf_prog *prog;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020067
68 atomic_t refcnt; /* Control when this struct can be free'ed */
69 struct rcu_head rcu;
Lorenzo Bianconi644bfe52020-07-14 15:56:37 +020070
71 struct work_struct kthread_stop_wq;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020072};
73
74struct bpf_cpu_map {
75 struct bpf_map map;
76 /* Below members specific for map type */
77 struct bpf_cpu_map_entry **cpu_map;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020078};
79
Björn Töpelcdfafe92019-12-19 07:10:04 +010080static DEFINE_PER_CPU(struct list_head, cpu_map_flush_list);
81
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020082static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
83{
Lorenzo Bianconi92164772020-07-14 15:56:38 +020084 u32 value_size = attr->value_size;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020085 struct bpf_cpu_map *cmap;
86 int err = -ENOMEM;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020087
Alexei Starovoitov2c78ee82020-05-13 16:03:54 -070088 if (!bpf_capable())
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020089 return ERR_PTR(-EPERM);
90
91 /* check sanity of attributes */
92 if (attr->max_entries == 0 || attr->key_size != 4 ||
Lorenzo Bianconi92164772020-07-14 15:56:38 +020093 (value_size != offsetofend(struct bpf_cpumap_val, qsize) &&
94 value_size != offsetofend(struct bpf_cpumap_val, bpf_prog.fd)) ||
95 attr->map_flags & ~BPF_F_NUMA_NODE)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020096 return ERR_PTR(-EINVAL);
97
Roman Gushchine88cc052020-12-01 13:58:35 -080098 cmap = kzalloc(sizeof(*cmap), GFP_USER | __GFP_ACCOUNT);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +020099 if (!cmap)
100 return ERR_PTR(-ENOMEM);
101
Jakub Kicinskibd475642018-01-11 20:29:06 -0800102 bpf_map_init_from_attr(&cmap->map, attr);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200103
104 /* Pre-limit array size based on NR_CPUS, not final CPU check */
105 if (cmap->map.max_entries > NR_CPUS) {
106 err = -E2BIG;
107 goto free_cmap;
108 }
109
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200110 /* Alloc array for possible remote "destination" CPUs */
111 cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
112 sizeof(struct bpf_cpu_map_entry *),
113 cmap->map.numa_node);
114 if (!cmap->cpu_map)
Roman Gushchin711caba2020-12-01 13:58:46 -0800115 goto free_cmap;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200116
117 return &cmap->map;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200118free_cmap:
119 kfree(cmap);
120 return ERR_PTR(err);
121}
122
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200123static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
124{
125 atomic_inc(&rcpu->refcnt);
126}
127
128/* called from workqueue, to workaround syscall using preempt_disable */
129static void cpu_map_kthread_stop(struct work_struct *work)
130{
131 struct bpf_cpu_map_entry *rcpu;
132
133 rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
134
135 /* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
136 * as it waits until all in-flight call_rcu() callbacks complete.
137 */
138 rcu_barrier();
139
140 /* kthread_stop will wake_up_process and wait for it to complete */
141 kthread_stop(rcpu->kthread);
142}
143
Lorenzo Bianconiefa90b52020-09-28 13:24:57 +0200144static struct sk_buff *cpu_map_build_skb(struct xdp_frame *xdpf,
Jesper Dangaard Brouer8f0504a92019-04-12 17:07:43 +0200145 struct sk_buff *skb)
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200146{
Jesper Dangaard Brouer676e4a62019-03-29 10:18:00 +0100147 unsigned int hard_start_headroom;
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200148 unsigned int frame_size;
149 void *pkt_data_start;
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200150
Jesper Dangaard Brouer676e4a62019-03-29 10:18:00 +0100151 /* Part of headroom was reserved to xdpf */
152 hard_start_headroom = sizeof(struct xdp_frame) + xdpf->headroom;
153
Jesper Dangaard Brouerdb612f72020-05-14 12:49:38 +0200154 /* Memory size backing xdp_frame data already have reserved
155 * room for build_skb to place skb_shared_info in tailroom.
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200156 */
Jesper Dangaard Brouerdb612f72020-05-14 12:49:38 +0200157 frame_size = xdpf->frame_sz;
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200158
Jesper Dangaard Brouer676e4a62019-03-29 10:18:00 +0100159 pkt_data_start = xdpf->data - hard_start_headroom;
Jesper Dangaard Brouer8f0504a92019-04-12 17:07:43 +0200160 skb = build_skb_around(skb, pkt_data_start, frame_size);
161 if (unlikely(!skb))
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200162 return NULL;
163
Jesper Dangaard Brouer676e4a62019-03-29 10:18:00 +0100164 skb_reserve(skb, hard_start_headroom);
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200165 __skb_put(skb, xdpf->len);
166 if (xdpf->metasize)
167 skb_metadata_set(skb, xdpf->metasize);
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200168
169 /* Essential SKB info: protocol and skb->dev */
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200170 skb->protocol = eth_type_trans(skb, xdpf->dev_rx);
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200171
172 /* Optional SKB info, currently missing:
173 * - HW checksum info (skb->ip_summed)
174 * - HW RX hash (skb_set_hash)
175 * - RX ring dev queue index (skb_record_rx_queue)
176 */
177
Jesper Dangaard Brouer6bf071b2019-06-18 15:05:27 +0200178 /* Until page_pool get SKB return path, release DMA here */
179 xdp_release_frame(xdpf);
180
Jesper Dangaard Brouer676e4a62019-03-29 10:18:00 +0100181 /* Allow SKB to reuse area used by xdp_frame */
182 xdp_scrub_frame(xdpf);
183
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200184 return skb;
185}
186
Jesper Dangaard Brouer5ab073f2018-04-17 16:45:26 +0200187static void __cpu_map_ring_cleanup(struct ptr_ring *ring)
188{
189 /* The tear-down procedure should have made sure that queue is
190 * empty. See __cpu_map_entry_replace() and work-queue
191 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
192 * gracefully and warn once.
193 */
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200194 struct xdp_frame *xdpf;
Jesper Dangaard Brouer5ab073f2018-04-17 16:45:26 +0200195
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200196 while ((xdpf = ptr_ring_consume(ring)))
197 if (WARN_ON_ONCE(xdpf))
Jesper Dangaard Brouer03993092018-04-17 16:46:32 +0200198 xdp_return_frame(xdpf);
Jesper Dangaard Brouer5ab073f2018-04-17 16:45:26 +0200199}
200
201static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
202{
203 if (atomic_dec_and_test(&rcpu->refcnt)) {
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200204 if (rcpu->prog)
205 bpf_prog_put(rcpu->prog);
Jesper Dangaard Brouer5ab073f2018-04-17 16:45:26 +0200206 /* The queue should be empty at this point */
207 __cpu_map_ring_cleanup(rcpu->queue);
208 ptr_ring_cleanup(rcpu->queue, NULL);
209 kfree(rcpu->queue);
210 kfree(rcpu);
211 }
212}
213
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200214static int cpu_map_bpf_prog_run_xdp(struct bpf_cpu_map_entry *rcpu,
215 void **frames, int n,
216 struct xdp_cpumap_stats *stats)
217{
218 struct xdp_rxq_info rxq;
219 struct xdp_buff xdp;
220 int i, nframes = 0;
221
222 if (!rcpu->prog)
223 return n;
224
Lorenzo Bianconi28b15202020-07-14 15:56:39 +0200225 rcu_read_lock_bh();
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200226
227 xdp_set_return_frame_no_direct();
228 xdp.rxq = &rxq;
229
230 for (i = 0; i < n; i++) {
231 struct xdp_frame *xdpf = frames[i];
232 u32 act;
233 int err;
234
235 rxq.dev = xdpf->dev_rx;
236 rxq.mem = xdpf->mem;
237 /* TODO: report queue_index to xdp_rxq_info */
238
239 xdp_convert_frame_to_buff(xdpf, &xdp);
240
241 act = bpf_prog_run_xdp(rcpu->prog, &xdp);
242 switch (act) {
243 case XDP_PASS:
244 err = xdp_update_frame_from_buff(&xdp, xdpf);
245 if (err < 0) {
246 xdp_return_frame(xdpf);
247 stats->drop++;
248 } else {
249 frames[nframes++] = xdpf;
250 stats->pass++;
251 }
252 break;
Lorenzo Bianconi28b15202020-07-14 15:56:39 +0200253 case XDP_REDIRECT:
254 err = xdp_do_redirect(xdpf->dev_rx, &xdp,
255 rcpu->prog);
256 if (unlikely(err)) {
257 xdp_return_frame(xdpf);
258 stats->drop++;
259 } else {
260 stats->redirect++;
261 }
262 break;
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200263 default:
264 bpf_warn_invalid_xdp_action(act);
Gustavo A. R. Silvadf561f662020-08-23 17:36:59 -0500265 fallthrough;
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200266 case XDP_DROP:
267 xdp_return_frame(xdpf);
268 stats->drop++;
269 break;
270 }
271 }
272
Lorenzo Bianconi28b15202020-07-14 15:56:39 +0200273 if (stats->redirect)
274 xdp_do_flush_map();
275
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200276 xdp_clear_return_frame_no_direct();
277
Lorenzo Bianconi28b15202020-07-14 15:56:39 +0200278 rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200279
280 return nframes;
281}
282
Jesper Dangaard Brouer77361822019-04-12 17:07:32 +0200283#define CPUMAP_BATCH 8
284
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200285static int cpu_map_kthread_run(void *data)
286{
287 struct bpf_cpu_map_entry *rcpu = data;
288
289 set_current_state(TASK_INTERRUPTIBLE);
290
291 /* When kthread gives stop order, then rcpu have been disconnected
292 * from map, thus no new packets can enter. Remaining in-flight
293 * per CPU stored packets are flushed to this queue. Wait honoring
294 * kthread_stop signal until queue is empty.
295 */
296 while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200297 struct xdp_cpumap_stats stats = {}; /* zero stats */
298 gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
Jesper Dangaard Brouer77361822019-04-12 17:07:32 +0200299 unsigned int drops = 0, sched = 0;
300 void *frames[CPUMAP_BATCH];
Jesper Dangaard Brouer8f0504a92019-04-12 17:07:43 +0200301 void *skbs[CPUMAP_BATCH];
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200302 int i, n, m, nframes;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200303
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200304 /* Release CPU reschedule checks */
305 if (__ptr_ring_empty(rcpu->queue)) {
Jesper Dangaard Brouer31749462017-10-23 19:39:28 +0200306 set_current_state(TASK_INTERRUPTIBLE);
307 /* Recheck to avoid lost wake-up */
308 if (__ptr_ring_empty(rcpu->queue)) {
309 schedule();
310 sched = 1;
311 } else {
312 __set_current_state(TASK_RUNNING);
313 }
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200314 } else {
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200315 sched = cond_resched();
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200316 }
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200317
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200318 /*
319 * The bpf_cpu_map_entry is single consumer, with this
320 * kthread CPU pinned. Lockless access to ptr_ring
321 * consume side valid as no-resize allowed of queue.
322 */
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200323 n = __ptr_ring_consume_batched(rcpu->queue, frames,
324 CPUMAP_BATCH);
Jesper Dangaard Brouer86d23142019-04-12 17:07:48 +0200325 for (i = 0; i < n; i++) {
326 void *f = frames[i];
327 struct page *page = virt_to_page(f);
328
329 /* Bring struct page memory area to curr CPU. Read by
330 * build_skb_around via page_is_pfmemalloc(), and when
331 * freed written by page_frag_free call.
332 */
333 prefetchw(page);
334 }
335
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200336 /* Support running another XDP prog on this CPU */
337 nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, n, &stats);
338 if (nframes) {
339 m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, nframes, skbs);
340 if (unlikely(m == 0)) {
341 for (i = 0; i < nframes; i++)
342 skbs[i] = NULL; /* effect: xdp_return_frame */
343 drops += nframes;
344 }
Jesper Dangaard Brouer8f0504a92019-04-12 17:07:43 +0200345 }
Jesper Dangaard Brouer77361822019-04-12 17:07:32 +0200346
347 local_bh_disable();
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200348 for (i = 0; i < nframes; i++) {
Jesper Dangaard Brouer77361822019-04-12 17:07:32 +0200349 struct xdp_frame *xdpf = frames[i];
Jesper Dangaard Brouer8f0504a92019-04-12 17:07:43 +0200350 struct sk_buff *skb = skbs[i];
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200351 int ret;
352
Lorenzo Bianconiefa90b52020-09-28 13:24:57 +0200353 skb = cpu_map_build_skb(xdpf, skb);
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200354 if (!skb) {
Jesper Dangaard Brouer03993092018-04-17 16:46:32 +0200355 xdp_return_frame(xdpf);
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200356 continue;
357 }
358
359 /* Inject into network stack */
360 ret = netif_receive_skb_core(skb);
361 if (ret == NET_RX_DROP)
362 drops++;
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200363 }
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200364 /* Feedback loop via tracepoint */
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200365 trace_xdp_cpumap_kthread(rcpu->map_id, n, drops, sched, &stats);
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200366
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200367 local_bh_enable(); /* resched point, may call do_softirq() */
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200368 }
369 __set_current_state(TASK_RUNNING);
370
371 put_cpu_map_entry(rcpu);
372 return 0;
373}
374
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200375bool cpu_map_prog_allowed(struct bpf_map *map)
376{
377 return map->map_type == BPF_MAP_TYPE_CPUMAP &&
378 map->value_size != offsetofend(struct bpf_cpumap_val, qsize);
379}
380
381static int __cpu_map_load_bpf_program(struct bpf_cpu_map_entry *rcpu, int fd)
382{
383 struct bpf_prog *prog;
384
385 prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
386 if (IS_ERR(prog))
387 return PTR_ERR(prog);
388
389 if (prog->expected_attach_type != BPF_XDP_CPUMAP) {
390 bpf_prog_put(prog);
391 return -EINVAL;
392 }
393
394 rcpu->value.bpf_prog.id = prog->aux->id;
395 rcpu->prog = prog;
396
397 return 0;
398}
399
Lorenzo Bianconi644bfe52020-07-14 15:56:37 +0200400static struct bpf_cpu_map_entry *
Roman Gushchine88cc052020-12-01 13:58:35 -0800401__cpu_map_entry_alloc(struct bpf_map *map, struct bpf_cpumap_val *value,
402 u32 cpu)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200403{
Lorenzo Bianconi92164772020-07-14 15:56:38 +0200404 int numa, err, i, fd = value->bpf_prog.fd;
Jason Wang7fc17e92018-02-14 22:17:34 +0800405 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200406 struct bpf_cpu_map_entry *rcpu;
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200407 struct xdp_bulk_queue *bq;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200408
409 /* Have map->numa_node, but choose node of redirect target CPU */
410 numa = cpu_to_node(cpu);
411
Roman Gushchine88cc052020-12-01 13:58:35 -0800412 rcpu = bpf_map_kmalloc_node(map, sizeof(*rcpu), gfp | __GFP_ZERO, numa);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200413 if (!rcpu)
414 return NULL;
415
416 /* Alloc percpu bulkq */
Roman Gushchine88cc052020-12-01 13:58:35 -0800417 rcpu->bulkq = bpf_map_alloc_percpu(map, sizeof(*rcpu->bulkq),
418 sizeof(void *), gfp);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200419 if (!rcpu->bulkq)
420 goto free_rcu;
421
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200422 for_each_possible_cpu(i) {
423 bq = per_cpu_ptr(rcpu->bulkq, i);
424 bq->obj = rcpu;
425 }
426
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200427 /* Alloc queue */
Roman Gushchine88cc052020-12-01 13:58:35 -0800428 rcpu->queue = bpf_map_kmalloc_node(map, sizeof(*rcpu->queue), gfp,
429 numa);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200430 if (!rcpu->queue)
431 goto free_bulkq;
432
Lorenzo Bianconi644bfe52020-07-14 15:56:37 +0200433 err = ptr_ring_init(rcpu->queue, value->qsize, gfp);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200434 if (err)
435 goto free_queue;
436
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200437 rcpu->cpu = cpu;
Roman Gushchine88cc052020-12-01 13:58:35 -0800438 rcpu->map_id = map->id;
Lorenzo Bianconi644bfe52020-07-14 15:56:37 +0200439 rcpu->value.qsize = value->qsize;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200440
Lorenzo Bianconic576b9c2020-07-19 17:52:41 +0200441 if (fd > 0 && __cpu_map_load_bpf_program(rcpu, fd))
442 goto free_ptr_ring;
443
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200444 /* Setup kthread */
445 rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
Roman Gushchine88cc052020-12-01 13:58:35 -0800446 "cpumap/%d/map:%d", cpu,
447 map->id);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200448 if (IS_ERR(rcpu->kthread))
Lorenzo Bianconic576b9c2020-07-19 17:52:41 +0200449 goto free_prog;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200450
451 get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
452 get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
453
454 /* Make sure kthread runs on a single CPU */
455 kthread_bind(rcpu->kthread, cpu);
456 wake_up_process(rcpu->kthread);
457
458 return rcpu;
459
Lorenzo Bianconic576b9c2020-07-19 17:52:41 +0200460free_prog:
461 if (rcpu->prog)
462 bpf_prog_put(rcpu->prog);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200463free_ptr_ring:
464 ptr_ring_cleanup(rcpu->queue, NULL);
465free_queue:
466 kfree(rcpu->queue);
467free_bulkq:
468 free_percpu(rcpu->bulkq);
469free_rcu:
470 kfree(rcpu);
471 return NULL;
472}
473
Wei Yongjun0fe875c2018-01-16 11:27:05 +0000474static void __cpu_map_entry_free(struct rcu_head *rcu)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200475{
476 struct bpf_cpu_map_entry *rcpu;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200477
478 /* This cpu_map_entry have been disconnected from map and one
Björn Töpelfb5aacd2019-12-19 07:10:01 +0100479 * RCU grace-period have elapsed. Thus, XDP cannot queue any
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200480 * new packets and cannot change/set flush_needed that can
481 * find this entry.
482 */
483 rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
484
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200485 free_percpu(rcpu->bulkq);
486 /* Cannot kthread_stop() here, last put free rcpu resources */
487 put_cpu_map_entry(rcpu);
488}
489
490/* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
491 * ensure any driver rcu critical sections have completed, but this
492 * does not guarantee a flush has happened yet. Because driver side
493 * rcu_read_lock/unlock only protects the running XDP program. The
494 * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
495 * pending flush op doesn't fail.
496 *
497 * The bpf_cpu_map_entry is still used by the kthread, and there can
498 * still be pending packets (in queue and percpu bulkq). A refcnt
499 * makes sure to last user (kthread_stop vs. call_rcu) free memory
500 * resources.
501 *
502 * The rcu callback __cpu_map_entry_free flush remaining packets in
503 * percpu bulkq to queue. Due to caller map_delete_elem() disable
504 * preemption, cannot call kthread_stop() to make sure queue is empty.
505 * Instead a work_queue is started for stopping kthread,
Björn Töpelfb5aacd2019-12-19 07:10:01 +0100506 * cpu_map_kthread_stop, which waits for an RCU grace period before
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200507 * stopping kthread, emptying the queue.
508 */
Wei Yongjun0fe875c2018-01-16 11:27:05 +0000509static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
510 u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200511{
512 struct bpf_cpu_map_entry *old_rcpu;
513
514 old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu);
515 if (old_rcpu) {
516 call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
517 INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
518 schedule_work(&old_rcpu->kthread_stop_wq);
519 }
520}
521
Wei Yongjun0fe875c2018-01-16 11:27:05 +0000522static int cpu_map_delete_elem(struct bpf_map *map, void *key)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200523{
524 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
525 u32 key_cpu = *(u32 *)key;
526
527 if (key_cpu >= map->max_entries)
528 return -EINVAL;
529
530 /* notice caller map_delete_elem() use preempt_disable() */
531 __cpu_map_entry_replace(cmap, key_cpu, NULL);
532 return 0;
533}
534
Wei Yongjun0fe875c2018-01-16 11:27:05 +0000535static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
536 u64 map_flags)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200537{
538 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
Lorenzo Bianconi644bfe52020-07-14 15:56:37 +0200539 struct bpf_cpumap_val cpumap_value = {};
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200540 struct bpf_cpu_map_entry *rcpu;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200541 /* Array index key correspond to CPU number */
542 u32 key_cpu = *(u32 *)key;
Lorenzo Bianconi644bfe52020-07-14 15:56:37 +0200543
544 memcpy(&cpumap_value, value, map->value_size);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200545
546 if (unlikely(map_flags > BPF_EXIST))
547 return -EINVAL;
548 if (unlikely(key_cpu >= cmap->map.max_entries))
549 return -E2BIG;
550 if (unlikely(map_flags == BPF_NOEXIST))
551 return -EEXIST;
Lorenzo Bianconi644bfe52020-07-14 15:56:37 +0200552 if (unlikely(cpumap_value.qsize > 16384)) /* sanity limit on qsize */
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200553 return -EOVERFLOW;
554
555 /* Make sure CPU is a valid possible cpu */
Toke Høiland-Jørgensenbc23d0e2020-04-16 10:31:20 +0200556 if (key_cpu >= nr_cpumask_bits || !cpu_possible(key_cpu))
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200557 return -ENODEV;
558
Lorenzo Bianconi644bfe52020-07-14 15:56:37 +0200559 if (cpumap_value.qsize == 0) {
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200560 rcpu = NULL; /* Same as deleting */
561 } else {
562 /* Updating qsize cause re-allocation of bpf_cpu_map_entry */
Roman Gushchine88cc052020-12-01 13:58:35 -0800563 rcpu = __cpu_map_entry_alloc(map, &cpumap_value, key_cpu);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200564 if (!rcpu)
565 return -ENOMEM;
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200566 rcpu->cmap = cmap;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200567 }
568 rcu_read_lock();
569 __cpu_map_entry_replace(cmap, key_cpu, rcpu);
570 rcu_read_unlock();
571 return 0;
572}
573
Wei Yongjun0fe875c2018-01-16 11:27:05 +0000574static void cpu_map_free(struct bpf_map *map)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200575{
576 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200577 u32 i;
578
579 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
580 * so the bpf programs (can be more than one that used this map) were
581 * disconnected from events. Wait for outstanding critical sections in
582 * these programs to complete. The rcu critical section only guarantees
583 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
584 * It does __not__ ensure pending flush operations (if any) are
585 * complete.
586 */
Daniel Borkmannf6069b92018-08-17 23:26:14 +0200587
588 bpf_clear_redirect_map(map);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200589 synchronize_rcu();
590
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200591 /* For cpu_map the remote CPUs can still be using the entries
592 * (struct bpf_cpu_map_entry).
593 */
594 for (i = 0; i < cmap->map.max_entries; i++) {
595 struct bpf_cpu_map_entry *rcpu;
596
597 rcpu = READ_ONCE(cmap->cpu_map[i]);
598 if (!rcpu)
599 continue;
600
Björn Töpelfb5aacd2019-12-19 07:10:01 +0100601 /* bq flush and cleanup happens after RCU grace-period */
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200602 __cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
603 }
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200604 bpf_map_area_free(cmap->cpu_map);
605 kfree(cmap);
606}
607
608struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
609{
610 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
611 struct bpf_cpu_map_entry *rcpu;
612
613 if (key >= map->max_entries)
614 return NULL;
615
616 rcpu = READ_ONCE(cmap->cpu_map[key]);
617 return rcpu;
618}
619
620static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
621{
622 struct bpf_cpu_map_entry *rcpu =
623 __cpu_map_lookup_elem(map, *(u32 *)key);
624
Lorenzo Bianconi644bfe52020-07-14 15:56:37 +0200625 return rcpu ? &rcpu->value : NULL;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200626}
627
628static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
629{
630 struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
631 u32 index = key ? *(u32 *)key : U32_MAX;
632 u32 *next = next_key;
633
634 if (index >= cmap->map.max_entries) {
635 *next = 0;
636 return 0;
637 }
638
639 if (index == cmap->map.max_entries - 1)
640 return -ENOENT;
641 *next = index + 1;
642 return 0;
643}
644
Andrey Ignatov2872e9a2020-06-19 14:11:44 -0700645static int cpu_map_btf_id;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200646const struct bpf_map_ops cpu_map_ops = {
Martin KaFai Lauf4d05252020-08-27 18:18:06 -0700647 .map_meta_equal = bpf_map_meta_equal,
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200648 .map_alloc = cpu_map_alloc,
649 .map_free = cpu_map_free,
650 .map_delete_elem = cpu_map_delete_elem,
651 .map_update_elem = cpu_map_update_elem,
652 .map_lookup_elem = cpu_map_lookup_elem,
653 .map_get_next_key = cpu_map_get_next_key,
Daniel Borkmanne8d2bec2018-08-12 01:59:17 +0200654 .map_check_btf = map_check_no_btf,
Andrey Ignatov2872e9a2020-06-19 14:11:44 -0700655 .map_btf_name = "bpf_cpu_map",
656 .map_btf_id = &cpu_map_btf_id,
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200657};
658
Björn Töpelebc4ecd2020-09-01 10:39:28 +0200659static void bq_flush_to_queue(struct xdp_bulk_queue *bq)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200660{
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200661 struct bpf_cpu_map_entry *rcpu = bq->obj;
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200662 unsigned int processed = 0, drops = 0;
663 const int to_cpu = rcpu->cpu;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200664 struct ptr_ring *q;
665 int i;
666
667 if (unlikely(!bq->count))
Björn Töpelebc4ecd2020-09-01 10:39:28 +0200668 return;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200669
670 q = rcpu->queue;
671 spin_lock(&q->producer_lock);
672
673 for (i = 0; i < bq->count; i++) {
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200674 struct xdp_frame *xdpf = bq->q[i];
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200675 int err;
676
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200677 err = __ptr_ring_produce(q, xdpf);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200678 if (err) {
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200679 drops++;
Björn Töpel4bc188c2019-12-19 07:10:00 +0100680 xdp_return_frame_rx_napi(xdpf);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200681 }
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200682 processed++;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200683 }
684 bq->count = 0;
685 spin_unlock(&q->producer_lock);
686
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200687 __list_del_clearprev(&bq->flush_node);
688
Jesper Dangaard Brouerf9419f72017-10-16 12:19:44 +0200689 /* Feedback loop via tracepoints */
690 trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200691}
692
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200693/* Runs under RCU-read-side, plus in softirq under NAPI protection.
694 * Thus, safe percpu variable access.
695 */
Björn Töpelebc4ecd2020-09-01 10:39:28 +0200696static void bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200697{
Björn Töpelcdfafe92019-12-19 07:10:04 +0100698 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200699 struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
700
701 if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
Björn Töpel4bc188c2019-12-19 07:10:00 +0100702 bq_flush_to_queue(bq);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200703
704 /* Notice, xdp_buff/page MUST be queued here, long enough for
705 * driver to code invoking us to finished, due to driver
706 * (e.g. ixgbe) recycle tricks based on page-refcnt.
707 *
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200708 * Thus, incoming xdp_frame is always queued here (else we race
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200709 * with another CPU on page-refcnt and remaining driver code).
710 * Queue time is very short, as driver will invoke flush
711 * operation, when completing napi->poll call.
712 */
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200713 bq->q[bq->count++] = xdpf;
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200714
715 if (!bq->flush_node.prev)
716 list_add(&bq->flush_node, flush_list);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200717}
718
Jesper Dangaard Brouer9c270af2017-10-16 12:19:34 +0200719int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
720 struct net_device *dev_rx)
721{
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200722 struct xdp_frame *xdpf;
Jesper Dangaard Brouer9c270af2017-10-16 12:19:34 +0200723
Lorenzo Bianconi1b698fa2020-05-28 22:47:29 +0200724 xdpf = xdp_convert_buff_to_frame(xdp);
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200725 if (unlikely(!xdpf))
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200726 return -EOVERFLOW;
Jesper Dangaard Brouer9c270af2017-10-16 12:19:34 +0200727
Jesper Dangaard Brouer1c601d82017-10-16 12:19:39 +0200728 /* Info needed when constructing SKB on remote CPU */
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200729 xdpf->dev_rx = dev_rx;
Jesper Dangaard Brouer9c270af2017-10-16 12:19:34 +0200730
Jesper Dangaard Brouer70280ed2018-04-17 16:45:57 +0200731 bq_enqueue(rcpu, xdpf);
Jesper Dangaard Brouer9c270af2017-10-16 12:19:34 +0200732 return 0;
733}
734
Björn Töpelcdfafe92019-12-19 07:10:04 +0100735void __cpu_map_flush(void)
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200736{
Björn Töpelcdfafe92019-12-19 07:10:04 +0100737 struct list_head *flush_list = this_cpu_ptr(&cpu_map_flush_list);
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200738 struct xdp_bulk_queue *bq, *tmp;
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200739
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200740 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
Björn Töpel4bc188c2019-12-19 07:10:00 +0100741 bq_flush_to_queue(bq);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200742
743 /* If already running, costs spin_lock_irqsave + smb_mb */
Toke Høiland-Jørgensend5df2832019-06-28 11:12:34 +0200744 wake_up_process(bq->obj->kthread);
Jesper Dangaard Brouer6710e112017-10-16 12:19:28 +0200745 }
746}
Björn Töpelcdfafe92019-12-19 07:10:04 +0100747
748static int __init cpu_map_init(void)
749{
750 int cpu;
751
752 for_each_possible_cpu(cpu)
753 INIT_LIST_HEAD(&per_cpu(cpu_map_flush_list, cpu));
754 return 0;
755}
756
757subsys_initcall(cpu_map_init);