net/virtio: support TSO
[dpdk.git] / drivers / net / virtio / virtio_user / vhost_user.c
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright(c) 2010-2016 Intel Corporation. All rights reserved.
5  *   All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Intel Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33
34 #include <sys/socket.h>
35 #include <sys/types.h>
36 #include <sys/stat.h>
37 #include <unistd.h>
38 #include <fcntl.h>
39 #include <sys/un.h>
40 #include <string.h>
41 #include <errno.h>
42
43 #include "vhost.h"
44
45 static int
46 vhost_user_write(int fd, void *buf, int len, int *fds, int fd_num)
47 {
48         int r;
49         struct msghdr msgh;
50         struct iovec iov;
51         size_t fd_size = fd_num * sizeof(int);
52         char control[CMSG_SPACE(fd_size)];
53         struct cmsghdr *cmsg;
54
55         memset(&msgh, 0, sizeof(msgh));
56         memset(control, 0, sizeof(control));
57
58         iov.iov_base = (uint8_t *)buf;
59         iov.iov_len = len;
60
61         msgh.msg_iov = &iov;
62         msgh.msg_iovlen = 1;
63         msgh.msg_control = control;
64         msgh.msg_controllen = sizeof(control);
65
66         cmsg = CMSG_FIRSTHDR(&msgh);
67         cmsg->cmsg_len = CMSG_LEN(fd_size);
68         cmsg->cmsg_level = SOL_SOCKET;
69         cmsg->cmsg_type = SCM_RIGHTS;
70         memcpy(CMSG_DATA(cmsg), fds, fd_size);
71
72         do {
73                 r = sendmsg(fd, &msgh, 0);
74         } while (r < 0 && errno == EINTR);
75
76         return r;
77 }
78
79 static int
80 vhost_user_read(int fd, struct vhost_user_msg *msg)
81 {
82         uint32_t valid_flags = VHOST_USER_REPLY_MASK | VHOST_USER_VERSION;
83         int ret, sz_hdr = VHOST_USER_HDR_SIZE, sz_payload;
84
85         ret = recv(fd, (void *)msg, sz_hdr, 0);
86         if (ret < sz_hdr) {
87                 PMD_DRV_LOG(ERR, "Failed to recv msg hdr: %d instead of %d.",
88                             ret, sz_hdr);
89                 goto fail;
90         }
91
92         /* validate msg flags */
93         if (msg->flags != (valid_flags)) {
94                 PMD_DRV_LOG(ERR, "Failed to recv msg: flags %x instead of %x.",
95                             msg->flags, valid_flags);
96                 goto fail;
97         }
98
99         sz_payload = msg->size;
100         if (sz_payload) {
101                 ret = recv(fd, (void *)((char *)msg + sz_hdr), sz_payload, 0);
102                 if (ret < sz_payload) {
103                         PMD_DRV_LOG(ERR,
104                                 "Failed to recv msg payload: %d instead of %d.",
105                                 ret, msg->size);
106                         goto fail;
107                 }
108         }
109
110         return 0;
111
112 fail:
113         return -1;
114 }
115
116 struct hugepage_file_info {
117         uint64_t addr;            /**< virtual addr */
118         size_t   size;            /**< the file size */
119         char     path[PATH_MAX];  /**< path to backing file */
120 };
121
122 /* Two possible options:
123  * 1. Match HUGEPAGE_INFO_FMT to find the file storing struct hugepage_file
124  * array. This is simple but cannot be used in secondary process because
125  * secondary process will close and munmap that file.
126  * 2. Match HUGEFILE_FMT to find hugepage files directly.
127  *
128  * We choose option 2.
129  */
130 static int
131 get_hugepage_file_info(struct hugepage_file_info huges[], int max)
132 {
133         int idx;
134         FILE *f;
135         char buf[BUFSIZ], *tmp, *tail;
136         char *str_underline, *str_start;
137         int huge_index;
138         uint64_t v_start, v_end;
139
140         f = fopen("/proc/self/maps", "r");
141         if (!f) {
142                 PMD_DRV_LOG(ERR, "cannot open /proc/self/maps");
143                 return -1;
144         }
145
146         idx = 0;
147         while (fgets(buf, sizeof(buf), f) != NULL) {
148                 if (sscanf(buf, "%" PRIx64 "-%" PRIx64, &v_start, &v_end) < 2) {
149                         PMD_DRV_LOG(ERR, "Failed to parse address");
150                         goto error;
151                 }
152
153                 tmp = strchr(buf, ' ') + 1; /** skip address */
154                 tmp = strchr(tmp, ' ') + 1; /** skip perm */
155                 tmp = strchr(tmp, ' ') + 1; /** skip offset */
156                 tmp = strchr(tmp, ' ') + 1; /** skip dev */
157                 tmp = strchr(tmp, ' ') + 1; /** skip inode */
158                 while (*tmp == ' ')         /** skip spaces */
159                         tmp++;
160                 tail = strrchr(tmp, '\n');  /** remove newline if exists */
161                 if (tail)
162                         *tail = '\0';
163
164                 /* Match HUGEFILE_FMT, aka "%s/%smap_%d",
165                  * which is defined in eal_filesystem.h
166                  */
167                 str_underline = strrchr(tmp, '_');
168                 if (!str_underline)
169                         continue;
170
171                 str_start = str_underline - strlen("map");
172                 if (str_start < tmp)
173                         continue;
174
175                 if (sscanf(str_start, "map_%d", &huge_index) != 1)
176                         continue;
177
178                 if (idx >= max) {
179                         PMD_DRV_LOG(ERR, "Exceed maximum of %d", max);
180                         goto error;
181                 }
182                 huges[idx].addr = v_start;
183                 huges[idx].size = v_end - v_start;
184                 snprintf(huges[idx].path, PATH_MAX, "%s", tmp);
185                 idx++;
186         }
187
188         fclose(f);
189         return idx;
190
191 error:
192         fclose(f);
193         return -1;
194 }
195
196 static int
197 prepare_vhost_memory_user(struct vhost_user_msg *msg, int fds[])
198 {
199         int i, num;
200         struct hugepage_file_info huges[VHOST_MEMORY_MAX_NREGIONS];
201         struct vhost_memory_region *mr;
202
203         num = get_hugepage_file_info(huges, VHOST_MEMORY_MAX_NREGIONS);
204         if (num < 0) {
205                 PMD_INIT_LOG(ERR, "Failed to prepare memory for vhost-user");
206                 return -1;
207         }
208
209         for (i = 0; i < num; ++i) {
210                 mr = &msg->payload.memory.regions[i];
211                 mr->guest_phys_addr = huges[i].addr; /* use vaddr! */
212                 mr->userspace_addr = huges[i].addr;
213                 mr->memory_size = huges[i].size;
214                 mr->mmap_offset = 0;
215                 fds[i] = open(huges[i].path, O_RDWR);
216         }
217
218         msg->payload.memory.nregions = num;
219         msg->payload.memory.padding = 0;
220
221         return 0;
222 }
223
224 static struct vhost_user_msg m;
225
226 static const char * const vhost_msg_strings[] = {
227         [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
228         [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
229         [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
230         [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
231         [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
232         [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
233         [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
234         [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
235         [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
236         [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
237         [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
238         [VHOST_USER_SET_VRING_ENABLE] = "VHOST_USER_SET_VRING_ENABLE",
239         NULL,
240 };
241
242 int
243 vhost_user_sock(int vhostfd, enum vhost_user_request req, void *arg)
244 {
245         struct vhost_user_msg msg;
246         struct vhost_vring_file *file = 0;
247         int need_reply = 0;
248         int fds[VHOST_MEMORY_MAX_NREGIONS];
249         int fd_num = 0;
250         int i, len;
251
252         RTE_SET_USED(m);
253         RTE_SET_USED(vhost_msg_strings);
254
255         PMD_DRV_LOG(INFO, "%s", vhost_msg_strings[req]);
256
257         msg.request = req;
258         msg.flags = VHOST_USER_VERSION;
259         msg.size = 0;
260
261         switch (req) {
262         case VHOST_USER_GET_FEATURES:
263                 need_reply = 1;
264                 break;
265
266         case VHOST_USER_SET_FEATURES:
267         case VHOST_USER_SET_LOG_BASE:
268                 msg.payload.u64 = *((__u64 *)arg);
269                 msg.size = sizeof(m.payload.u64);
270                 break;
271
272         case VHOST_USER_SET_OWNER:
273         case VHOST_USER_RESET_OWNER:
274                 break;
275
276         case VHOST_USER_SET_MEM_TABLE:
277                 if (prepare_vhost_memory_user(&msg, fds) < 0)
278                         return -1;
279                 fd_num = msg.payload.memory.nregions;
280                 msg.size = sizeof(m.payload.memory.nregions);
281                 msg.size += sizeof(m.payload.memory.padding);
282                 msg.size += fd_num * sizeof(struct vhost_memory_region);
283                 break;
284
285         case VHOST_USER_SET_LOG_FD:
286                 fds[fd_num++] = *((int *)arg);
287                 break;
288
289         case VHOST_USER_SET_VRING_NUM:
290         case VHOST_USER_SET_VRING_BASE:
291         case VHOST_USER_SET_VRING_ENABLE:
292                 memcpy(&msg.payload.state, arg, sizeof(msg.payload.state));
293                 msg.size = sizeof(m.payload.state);
294                 break;
295
296         case VHOST_USER_GET_VRING_BASE:
297                 memcpy(&msg.payload.state, arg, sizeof(msg.payload.state));
298                 msg.size = sizeof(m.payload.state);
299                 need_reply = 1;
300                 break;
301
302         case VHOST_USER_SET_VRING_ADDR:
303                 memcpy(&msg.payload.addr, arg, sizeof(msg.payload.addr));
304                 msg.size = sizeof(m.payload.addr);
305                 break;
306
307         case VHOST_USER_SET_VRING_KICK:
308         case VHOST_USER_SET_VRING_CALL:
309         case VHOST_USER_SET_VRING_ERR:
310                 file = arg;
311                 msg.payload.u64 = file->index & VHOST_USER_VRING_IDX_MASK;
312                 msg.size = sizeof(m.payload.u64);
313                 if (file->fd > 0)
314                         fds[fd_num++] = file->fd;
315                 else
316                         msg.payload.u64 |= VHOST_USER_VRING_NOFD_MASK;
317                 break;
318
319         default:
320                 PMD_DRV_LOG(ERR, "trying to send unhandled msg type");
321                 return -1;
322         }
323
324         len = VHOST_USER_HDR_SIZE + msg.size;
325         if (vhost_user_write(vhostfd, &msg, len, fds, fd_num) < 0) {
326                 PMD_DRV_LOG(ERR, "%s failed: %s",
327                             vhost_msg_strings[req], strerror(errno));
328                 return -1;
329         }
330
331         if (req == VHOST_USER_SET_MEM_TABLE)
332                 for (i = 0; i < fd_num; ++i)
333                         close(fds[i]);
334
335         if (need_reply) {
336                 if (vhost_user_read(vhostfd, &msg) < 0) {
337                         PMD_DRV_LOG(ERR, "Received msg failed: %s",
338                                     strerror(errno));
339                         return -1;
340                 }
341
342                 if (req != msg.request) {
343                         PMD_DRV_LOG(ERR, "Received unexpected msg type");
344                         return -1;
345                 }
346
347                 switch (req) {
348                 case VHOST_USER_GET_FEATURES:
349                         if (msg.size != sizeof(m.payload.u64)) {
350                                 PMD_DRV_LOG(ERR, "Received bad msg size");
351                                 return -1;
352                         }
353                         *((__u64 *)arg) = msg.payload.u64;
354                         break;
355                 case VHOST_USER_GET_VRING_BASE:
356                         if (msg.size != sizeof(m.payload.state)) {
357                                 PMD_DRV_LOG(ERR, "Received bad msg size");
358                                 return -1;
359                         }
360                         memcpy(arg, &msg.payload.state,
361                                sizeof(struct vhost_vring_state));
362                         break;
363                 default:
364                         PMD_DRV_LOG(ERR, "Received unexpected msg type");
365                         return -1;
366                 }
367         }
368
369         return 0;
370 }
371
372 /**
373  * Set up environment to talk with a vhost user backend.
374  * @param path
375  *   - The path to vhost user unix socket file.
376  *
377  * @return
378  *   - (-1) if fail to set up;
379  *   - (>=0) if successful, and it is the fd to vhostfd.
380  */
381 int
382 vhost_user_setup(const char *path)
383 {
384         int fd;
385         int flag;
386         struct sockaddr_un un;
387
388         fd = socket(AF_UNIX, SOCK_STREAM, 0);
389         if (fd < 0) {
390                 PMD_DRV_LOG(ERR, "socket() error, %s", strerror(errno));
391                 return -1;
392         }
393
394         flag = fcntl(fd, F_GETFD);
395         if (fcntl(fd, F_SETFD, flag | FD_CLOEXEC) < 0)
396                 PMD_DRV_LOG(WARNING, "fcntl failed, %s", strerror(errno));
397
398         memset(&un, 0, sizeof(un));
399         un.sun_family = AF_UNIX;
400         snprintf(un.sun_path, sizeof(un.sun_path), "%s", path);
401         if (connect(fd, (struct sockaddr *)&un, sizeof(un)) < 0) {
402                 PMD_DRV_LOG(ERR, "connect error, %s", strerror(errno));
403                 close(fd);
404                 return -1;
405         }
406
407         return fd;
408 }
409
410 int
411 vhost_user_enable_queue_pair(int vhostfd, uint16_t pair_idx, int enable)
412 {
413         int i;
414
415         for (i = 0; i < 2; ++i) {
416                 struct vhost_vring_state state = {
417                         .index = pair_idx * 2 + i,
418                         .num   = enable,
419                 };
420
421                 if (vhost_user_sock(vhostfd,
422                                     VHOST_USER_SET_VRING_ENABLE, &state))
423                         return -1;
424         }
425
426         return 0;
427 }