Loading...
Note: File does not exist in v6.2.
1// SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause)
2
3/*
4 * AF_XDP user-space access library.
5 *
6 * Copyright(c) 2018 - 2019 Intel Corporation.
7 *
8 * Author(s): Magnus Karlsson <magnus.karlsson@intel.com>
9 */
10
11#include <errno.h>
12#include <stdlib.h>
13#include <string.h>
14#include <unistd.h>
15#include <arpa/inet.h>
16#include <asm/barrier.h>
17#include <linux/compiler.h>
18#include <linux/ethtool.h>
19#include <linux/filter.h>
20#include <linux/if_ether.h>
21#include <linux/if_packet.h>
22#include <linux/if_xdp.h>
23#include <linux/sockios.h>
24#include <net/if.h>
25#include <sys/ioctl.h>
26#include <sys/mman.h>
27#include <sys/socket.h>
28#include <sys/types.h>
29
30#include "bpf.h"
31#include "libbpf.h"
32#include "libbpf_internal.h"
33#include "xsk.h"
34
35#ifndef SOL_XDP
36 #define SOL_XDP 283
37#endif
38
39#ifndef AF_XDP
40 #define AF_XDP 44
41#endif
42
43#ifndef PF_XDP
44 #define PF_XDP AF_XDP
45#endif
46
47struct xsk_umem {
48 struct xsk_ring_prod *fill;
49 struct xsk_ring_cons *comp;
50 char *umem_area;
51 struct xsk_umem_config config;
52 int fd;
53 int refcount;
54};
55
56struct xsk_socket {
57 struct xsk_ring_cons *rx;
58 struct xsk_ring_prod *tx;
59 __u64 outstanding_tx;
60 struct xsk_umem *umem;
61 struct xsk_socket_config config;
62 int fd;
63 int ifindex;
64 int prog_fd;
65 int xsks_map_fd;
66 __u32 queue_id;
67 char ifname[IFNAMSIZ];
68};
69
70struct xsk_nl_info {
71 bool xdp_prog_attached;
72 int ifindex;
73 int fd;
74};
75
76int xsk_umem__fd(const struct xsk_umem *umem)
77{
78 return umem ? umem->fd : -EINVAL;
79}
80
81int xsk_socket__fd(const struct xsk_socket *xsk)
82{
83 return xsk ? xsk->fd : -EINVAL;
84}
85
86static bool xsk_page_aligned(void *buffer)
87{
88 unsigned long addr = (unsigned long)buffer;
89
90 return !(addr & (getpagesize() - 1));
91}
92
93static void xsk_set_umem_config(struct xsk_umem_config *cfg,
94 const struct xsk_umem_config *usr_cfg)
95{
96 if (!usr_cfg) {
97 cfg->fill_size = XSK_RING_PROD__DEFAULT_NUM_DESCS;
98 cfg->comp_size = XSK_RING_CONS__DEFAULT_NUM_DESCS;
99 cfg->frame_size = XSK_UMEM__DEFAULT_FRAME_SIZE;
100 cfg->frame_headroom = XSK_UMEM__DEFAULT_FRAME_HEADROOM;
101 cfg->flags = XSK_UMEM__DEFAULT_FLAGS;
102 return;
103 }
104
105 cfg->fill_size = usr_cfg->fill_size;
106 cfg->comp_size = usr_cfg->comp_size;
107 cfg->frame_size = usr_cfg->frame_size;
108 cfg->frame_headroom = usr_cfg->frame_headroom;
109 cfg->flags = usr_cfg->flags;
110}
111
112static int xsk_set_xdp_socket_config(struct xsk_socket_config *cfg,
113 const struct xsk_socket_config *usr_cfg)
114{
115 if (!usr_cfg) {
116 cfg->rx_size = XSK_RING_CONS__DEFAULT_NUM_DESCS;
117 cfg->tx_size = XSK_RING_PROD__DEFAULT_NUM_DESCS;
118 cfg->libbpf_flags = 0;
119 cfg->xdp_flags = 0;
120 cfg->bind_flags = 0;
121 return 0;
122 }
123
124 if (usr_cfg->libbpf_flags & ~XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD)
125 return -EINVAL;
126
127 cfg->rx_size = usr_cfg->rx_size;
128 cfg->tx_size = usr_cfg->tx_size;
129 cfg->libbpf_flags = usr_cfg->libbpf_flags;
130 cfg->xdp_flags = usr_cfg->xdp_flags;
131 cfg->bind_flags = usr_cfg->bind_flags;
132
133 return 0;
134}
135
136int xsk_umem__create_v0_0_4(struct xsk_umem **umem_ptr, void *umem_area,
137 __u64 size, struct xsk_ring_prod *fill,
138 struct xsk_ring_cons *comp,
139 const struct xsk_umem_config *usr_config)
140{
141 struct xdp_mmap_offsets off;
142 struct xdp_umem_reg mr;
143 struct xsk_umem *umem;
144 socklen_t optlen;
145 void *map;
146 int err;
147
148 if (!umem_area || !umem_ptr || !fill || !comp)
149 return -EFAULT;
150 if (!size && !xsk_page_aligned(umem_area))
151 return -EINVAL;
152
153 umem = calloc(1, sizeof(*umem));
154 if (!umem)
155 return -ENOMEM;
156
157 umem->fd = socket(AF_XDP, SOCK_RAW, 0);
158 if (umem->fd < 0) {
159 err = -errno;
160 goto out_umem_alloc;
161 }
162
163 umem->umem_area = umem_area;
164 xsk_set_umem_config(&umem->config, usr_config);
165
166 mr.addr = (uintptr_t)umem_area;
167 mr.len = size;
168 mr.chunk_size = umem->config.frame_size;
169 mr.headroom = umem->config.frame_headroom;
170 mr.flags = umem->config.flags;
171
172 err = setsockopt(umem->fd, SOL_XDP, XDP_UMEM_REG, &mr, sizeof(mr));
173 if (err) {
174 err = -errno;
175 goto out_socket;
176 }
177 err = setsockopt(umem->fd, SOL_XDP, XDP_UMEM_FILL_RING,
178 &umem->config.fill_size,
179 sizeof(umem->config.fill_size));
180 if (err) {
181 err = -errno;
182 goto out_socket;
183 }
184 err = setsockopt(umem->fd, SOL_XDP, XDP_UMEM_COMPLETION_RING,
185 &umem->config.comp_size,
186 sizeof(umem->config.comp_size));
187 if (err) {
188 err = -errno;
189 goto out_socket;
190 }
191
192 optlen = sizeof(off);
193 err = getsockopt(umem->fd, SOL_XDP, XDP_MMAP_OFFSETS, &off, &optlen);
194 if (err) {
195 err = -errno;
196 goto out_socket;
197 }
198
199 map = mmap(NULL, off.fr.desc + umem->config.fill_size * sizeof(__u64),
200 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, umem->fd,
201 XDP_UMEM_PGOFF_FILL_RING);
202 if (map == MAP_FAILED) {
203 err = -errno;
204 goto out_socket;
205 }
206
207 umem->fill = fill;
208 fill->mask = umem->config.fill_size - 1;
209 fill->size = umem->config.fill_size;
210 fill->producer = map + off.fr.producer;
211 fill->consumer = map + off.fr.consumer;
212 fill->flags = map + off.fr.flags;
213 fill->ring = map + off.fr.desc;
214 fill->cached_cons = umem->config.fill_size;
215
216 map = mmap(NULL, off.cr.desc + umem->config.comp_size * sizeof(__u64),
217 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, umem->fd,
218 XDP_UMEM_PGOFF_COMPLETION_RING);
219 if (map == MAP_FAILED) {
220 err = -errno;
221 goto out_mmap;
222 }
223
224 umem->comp = comp;
225 comp->mask = umem->config.comp_size - 1;
226 comp->size = umem->config.comp_size;
227 comp->producer = map + off.cr.producer;
228 comp->consumer = map + off.cr.consumer;
229 comp->flags = map + off.cr.flags;
230 comp->ring = map + off.cr.desc;
231
232 *umem_ptr = umem;
233 return 0;
234
235out_mmap:
236 munmap(map, off.fr.desc + umem->config.fill_size * sizeof(__u64));
237out_socket:
238 close(umem->fd);
239out_umem_alloc:
240 free(umem);
241 return err;
242}
243
244struct xsk_umem_config_v1 {
245 __u32 fill_size;
246 __u32 comp_size;
247 __u32 frame_size;
248 __u32 frame_headroom;
249};
250
251int xsk_umem__create_v0_0_2(struct xsk_umem **umem_ptr, void *umem_area,
252 __u64 size, struct xsk_ring_prod *fill,
253 struct xsk_ring_cons *comp,
254 const struct xsk_umem_config *usr_config)
255{
256 struct xsk_umem_config config;
257
258 memcpy(&config, usr_config, sizeof(struct xsk_umem_config_v1));
259 config.flags = 0;
260
261 return xsk_umem__create_v0_0_4(umem_ptr, umem_area, size, fill, comp,
262 &config);
263}
264COMPAT_VERSION(xsk_umem__create_v0_0_2, xsk_umem__create, LIBBPF_0.0.2)
265DEFAULT_VERSION(xsk_umem__create_v0_0_4, xsk_umem__create, LIBBPF_0.0.4)
266
267static int xsk_load_xdp_prog(struct xsk_socket *xsk)
268{
269 static const int log_buf_size = 16 * 1024;
270 char log_buf[log_buf_size];
271 int err, prog_fd;
272
273 /* This is the C-program:
274 * SEC("xdp_sock") int xdp_sock_prog(struct xdp_md *ctx)
275 * {
276 * int index = ctx->rx_queue_index;
277 *
278 * // A set entry here means that the correspnding queue_id
279 * // has an active AF_XDP socket bound to it.
280 * if (bpf_map_lookup_elem(&xsks_map, &index))
281 * return bpf_redirect_map(&xsks_map, index, 0);
282 *
283 * return XDP_PASS;
284 * }
285 */
286 struct bpf_insn prog[] = {
287 /* r1 = *(u32 *)(r1 + 16) */
288 BPF_LDX_MEM(BPF_W, BPF_REG_1, BPF_REG_1, 16),
289 /* *(u32 *)(r10 - 4) = r1 */
290 BPF_STX_MEM(BPF_W, BPF_REG_10, BPF_REG_1, -4),
291 BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),
292 BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4),
293 BPF_LD_MAP_FD(BPF_REG_1, xsk->xsks_map_fd),
294 BPF_EMIT_CALL(BPF_FUNC_map_lookup_elem),
295 BPF_MOV64_REG(BPF_REG_1, BPF_REG_0),
296 BPF_MOV32_IMM(BPF_REG_0, 2),
297 /* if r1 == 0 goto +5 */
298 BPF_JMP_IMM(BPF_JEQ, BPF_REG_1, 0, 5),
299 /* r2 = *(u32 *)(r10 - 4) */
300 BPF_LD_MAP_FD(BPF_REG_1, xsk->xsks_map_fd),
301 BPF_LDX_MEM(BPF_W, BPF_REG_2, BPF_REG_10, -4),
302 BPF_MOV32_IMM(BPF_REG_3, 0),
303 BPF_EMIT_CALL(BPF_FUNC_redirect_map),
304 /* The jumps are to this instruction */
305 BPF_EXIT_INSN(),
306 };
307 size_t insns_cnt = sizeof(prog) / sizeof(struct bpf_insn);
308
309 prog_fd = bpf_load_program(BPF_PROG_TYPE_XDP, prog, insns_cnt,
310 "LGPL-2.1 or BSD-2-Clause", 0, log_buf,
311 log_buf_size);
312 if (prog_fd < 0) {
313 pr_warning("BPF log buffer:\n%s", log_buf);
314 return prog_fd;
315 }
316
317 err = bpf_set_link_xdp_fd(xsk->ifindex, prog_fd, xsk->config.xdp_flags);
318 if (err) {
319 close(prog_fd);
320 return err;
321 }
322
323 xsk->prog_fd = prog_fd;
324 return 0;
325}
326
327static int xsk_get_max_queues(struct xsk_socket *xsk)
328{
329 struct ethtool_channels channels = { .cmd = ETHTOOL_GCHANNELS };
330 struct ifreq ifr = {};
331 int fd, err, ret;
332
333 fd = socket(AF_INET, SOCK_DGRAM, 0);
334 if (fd < 0)
335 return -errno;
336
337 ifr.ifr_data = (void *)&channels;
338 memcpy(ifr.ifr_name, xsk->ifname, IFNAMSIZ - 1);
339 ifr.ifr_name[IFNAMSIZ - 1] = '\0';
340 err = ioctl(fd, SIOCETHTOOL, &ifr);
341 if (err && errno != EOPNOTSUPP) {
342 ret = -errno;
343 goto out;
344 }
345
346 if (err || channels.max_combined == 0)
347 /* If the device says it has no channels, then all traffic
348 * is sent to a single stream, so max queues = 1.
349 */
350 ret = 1;
351 else
352 ret = channels.max_combined;
353
354out:
355 close(fd);
356 return ret;
357}
358
359static int xsk_create_bpf_maps(struct xsk_socket *xsk)
360{
361 int max_queues;
362 int fd;
363
364 max_queues = xsk_get_max_queues(xsk);
365 if (max_queues < 0)
366 return max_queues;
367
368 fd = bpf_create_map_name(BPF_MAP_TYPE_XSKMAP, "xsks_map",
369 sizeof(int), sizeof(int), max_queues, 0);
370 if (fd < 0)
371 return fd;
372
373 xsk->xsks_map_fd = fd;
374
375 return 0;
376}
377
378static void xsk_delete_bpf_maps(struct xsk_socket *xsk)
379{
380 bpf_map_delete_elem(xsk->xsks_map_fd, &xsk->queue_id);
381 close(xsk->xsks_map_fd);
382}
383
384static int xsk_lookup_bpf_maps(struct xsk_socket *xsk)
385{
386 __u32 i, *map_ids, num_maps, prog_len = sizeof(struct bpf_prog_info);
387 __u32 map_len = sizeof(struct bpf_map_info);
388 struct bpf_prog_info prog_info = {};
389 struct bpf_map_info map_info;
390 int fd, err;
391
392 err = bpf_obj_get_info_by_fd(xsk->prog_fd, &prog_info, &prog_len);
393 if (err)
394 return err;
395
396 num_maps = prog_info.nr_map_ids;
397
398 map_ids = calloc(prog_info.nr_map_ids, sizeof(*map_ids));
399 if (!map_ids)
400 return -ENOMEM;
401
402 memset(&prog_info, 0, prog_len);
403 prog_info.nr_map_ids = num_maps;
404 prog_info.map_ids = (__u64)(unsigned long)map_ids;
405
406 err = bpf_obj_get_info_by_fd(xsk->prog_fd, &prog_info, &prog_len);
407 if (err)
408 goto out_map_ids;
409
410 xsk->xsks_map_fd = -1;
411
412 for (i = 0; i < prog_info.nr_map_ids; i++) {
413 fd = bpf_map_get_fd_by_id(map_ids[i]);
414 if (fd < 0)
415 continue;
416
417 err = bpf_obj_get_info_by_fd(fd, &map_info, &map_len);
418 if (err) {
419 close(fd);
420 continue;
421 }
422
423 if (!strcmp(map_info.name, "xsks_map")) {
424 xsk->xsks_map_fd = fd;
425 continue;
426 }
427
428 close(fd);
429 }
430
431 err = 0;
432 if (xsk->xsks_map_fd == -1)
433 err = -ENOENT;
434
435out_map_ids:
436 free(map_ids);
437 return err;
438}
439
440static int xsk_set_bpf_maps(struct xsk_socket *xsk)
441{
442 return bpf_map_update_elem(xsk->xsks_map_fd, &xsk->queue_id,
443 &xsk->fd, 0);
444}
445
446static int xsk_setup_xdp_prog(struct xsk_socket *xsk)
447{
448 __u32 prog_id = 0;
449 int err;
450
451 err = bpf_get_link_xdp_id(xsk->ifindex, &prog_id,
452 xsk->config.xdp_flags);
453 if (err)
454 return err;
455
456 if (!prog_id) {
457 err = xsk_create_bpf_maps(xsk);
458 if (err)
459 return err;
460
461 err = xsk_load_xdp_prog(xsk);
462 if (err) {
463 xsk_delete_bpf_maps(xsk);
464 return err;
465 }
466 } else {
467 xsk->prog_fd = bpf_prog_get_fd_by_id(prog_id);
468 err = xsk_lookup_bpf_maps(xsk);
469 if (err) {
470 close(xsk->prog_fd);
471 return err;
472 }
473 }
474
475 err = xsk_set_bpf_maps(xsk);
476 if (err) {
477 xsk_delete_bpf_maps(xsk);
478 close(xsk->prog_fd);
479 return err;
480 }
481
482 return 0;
483}
484
485int xsk_socket__create(struct xsk_socket **xsk_ptr, const char *ifname,
486 __u32 queue_id, struct xsk_umem *umem,
487 struct xsk_ring_cons *rx, struct xsk_ring_prod *tx,
488 const struct xsk_socket_config *usr_config)
489{
490 void *rx_map = NULL, *tx_map = NULL;
491 struct sockaddr_xdp sxdp = {};
492 struct xdp_mmap_offsets off;
493 struct xsk_socket *xsk;
494 socklen_t optlen;
495 int err;
496
497 if (!umem || !xsk_ptr || !rx || !tx)
498 return -EFAULT;
499
500 if (umem->refcount) {
501 pr_warning("Error: shared umems not supported by libbpf.\n");
502 return -EBUSY;
503 }
504
505 xsk = calloc(1, sizeof(*xsk));
506 if (!xsk)
507 return -ENOMEM;
508
509 if (umem->refcount++ > 0) {
510 xsk->fd = socket(AF_XDP, SOCK_RAW, 0);
511 if (xsk->fd < 0) {
512 err = -errno;
513 goto out_xsk_alloc;
514 }
515 } else {
516 xsk->fd = umem->fd;
517 }
518
519 xsk->outstanding_tx = 0;
520 xsk->queue_id = queue_id;
521 xsk->umem = umem;
522 xsk->ifindex = if_nametoindex(ifname);
523 if (!xsk->ifindex) {
524 err = -errno;
525 goto out_socket;
526 }
527 memcpy(xsk->ifname, ifname, IFNAMSIZ - 1);
528 xsk->ifname[IFNAMSIZ - 1] = '\0';
529
530 err = xsk_set_xdp_socket_config(&xsk->config, usr_config);
531 if (err)
532 goto out_socket;
533
534 if (rx) {
535 err = setsockopt(xsk->fd, SOL_XDP, XDP_RX_RING,
536 &xsk->config.rx_size,
537 sizeof(xsk->config.rx_size));
538 if (err) {
539 err = -errno;
540 goto out_socket;
541 }
542 }
543 if (tx) {
544 err = setsockopt(xsk->fd, SOL_XDP, XDP_TX_RING,
545 &xsk->config.tx_size,
546 sizeof(xsk->config.tx_size));
547 if (err) {
548 err = -errno;
549 goto out_socket;
550 }
551 }
552
553 optlen = sizeof(off);
554 err = getsockopt(xsk->fd, SOL_XDP, XDP_MMAP_OFFSETS, &off, &optlen);
555 if (err) {
556 err = -errno;
557 goto out_socket;
558 }
559
560 if (rx) {
561 rx_map = mmap(NULL, off.rx.desc +
562 xsk->config.rx_size * sizeof(struct xdp_desc),
563 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE,
564 xsk->fd, XDP_PGOFF_RX_RING);
565 if (rx_map == MAP_FAILED) {
566 err = -errno;
567 goto out_socket;
568 }
569
570 rx->mask = xsk->config.rx_size - 1;
571 rx->size = xsk->config.rx_size;
572 rx->producer = rx_map + off.rx.producer;
573 rx->consumer = rx_map + off.rx.consumer;
574 rx->flags = rx_map + off.rx.flags;
575 rx->ring = rx_map + off.rx.desc;
576 }
577 xsk->rx = rx;
578
579 if (tx) {
580 tx_map = mmap(NULL, off.tx.desc +
581 xsk->config.tx_size * sizeof(struct xdp_desc),
582 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE,
583 xsk->fd, XDP_PGOFF_TX_RING);
584 if (tx_map == MAP_FAILED) {
585 err = -errno;
586 goto out_mmap_rx;
587 }
588
589 tx->mask = xsk->config.tx_size - 1;
590 tx->size = xsk->config.tx_size;
591 tx->producer = tx_map + off.tx.producer;
592 tx->consumer = tx_map + off.tx.consumer;
593 tx->flags = tx_map + off.tx.flags;
594 tx->ring = tx_map + off.tx.desc;
595 tx->cached_cons = xsk->config.tx_size;
596 }
597 xsk->tx = tx;
598
599 sxdp.sxdp_family = PF_XDP;
600 sxdp.sxdp_ifindex = xsk->ifindex;
601 sxdp.sxdp_queue_id = xsk->queue_id;
602 sxdp.sxdp_flags = xsk->config.bind_flags;
603
604 err = bind(xsk->fd, (struct sockaddr *)&sxdp, sizeof(sxdp));
605 if (err) {
606 err = -errno;
607 goto out_mmap_tx;
608 }
609
610 xsk->prog_fd = -1;
611
612 if (!(xsk->config.libbpf_flags & XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD)) {
613 err = xsk_setup_xdp_prog(xsk);
614 if (err)
615 goto out_mmap_tx;
616 }
617
618 *xsk_ptr = xsk;
619 return 0;
620
621out_mmap_tx:
622 if (tx)
623 munmap(tx_map, off.tx.desc +
624 xsk->config.tx_size * sizeof(struct xdp_desc));
625out_mmap_rx:
626 if (rx)
627 munmap(rx_map, off.rx.desc +
628 xsk->config.rx_size * sizeof(struct xdp_desc));
629out_socket:
630 if (--umem->refcount)
631 close(xsk->fd);
632out_xsk_alloc:
633 free(xsk);
634 return err;
635}
636
637int xsk_umem__delete(struct xsk_umem *umem)
638{
639 struct xdp_mmap_offsets off;
640 socklen_t optlen;
641 int err;
642
643 if (!umem)
644 return 0;
645
646 if (umem->refcount)
647 return -EBUSY;
648
649 optlen = sizeof(off);
650 err = getsockopt(umem->fd, SOL_XDP, XDP_MMAP_OFFSETS, &off, &optlen);
651 if (!err) {
652 munmap(umem->fill->ring - off.fr.desc,
653 off.fr.desc + umem->config.fill_size * sizeof(__u64));
654 munmap(umem->comp->ring - off.cr.desc,
655 off.cr.desc + umem->config.comp_size * sizeof(__u64));
656 }
657
658 close(umem->fd);
659 free(umem);
660
661 return 0;
662}
663
664void xsk_socket__delete(struct xsk_socket *xsk)
665{
666 size_t desc_sz = sizeof(struct xdp_desc);
667 struct xdp_mmap_offsets off;
668 socklen_t optlen;
669 int err;
670
671 if (!xsk)
672 return;
673
674 if (xsk->prog_fd != -1) {
675 xsk_delete_bpf_maps(xsk);
676 close(xsk->prog_fd);
677 }
678
679 optlen = sizeof(off);
680 err = getsockopt(xsk->fd, SOL_XDP, XDP_MMAP_OFFSETS, &off, &optlen);
681 if (!err) {
682 if (xsk->rx) {
683 munmap(xsk->rx->ring - off.rx.desc,
684 off.rx.desc + xsk->config.rx_size * desc_sz);
685 }
686 if (xsk->tx) {
687 munmap(xsk->tx->ring - off.tx.desc,
688 off.tx.desc + xsk->config.tx_size * desc_sz);
689 }
690
691 }
692
693 xsk->umem->refcount--;
694 /* Do not close an fd that also has an associated umem connected
695 * to it.
696 */
697 if (xsk->fd != xsk->umem->fd)
698 close(xsk->fd);
699 free(xsk);
700}