eal: factorise unsupported option handling
[dpdk.git] / lib / librte_eal / linuxapp / eal / eal.c
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
5  *   Copyright(c) 2012-2014 6WIND S.A.
6  *   All rights reserved.
7  *
8  *   Redistribution and use in source and binary forms, with or without
9  *   modification, are permitted provided that the following conditions
10  *   are met:
11  *
12  *     * Redistributions of source code must retain the above copyright
13  *       notice, this list of conditions and the following disclaimer.
14  *     * Redistributions in binary form must reproduce the above copyright
15  *       notice, this list of conditions and the following disclaimer in
16  *       the documentation and/or other materials provided with the
17  *       distribution.
18  *     * Neither the name of Intel Corporation nor the names of its
19  *       contributors may be used to endorse or promote products derived
20  *       from this software without specific prior written permission.
21  *
22  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
25  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
26  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
27  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
28  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
32  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33  */
34
35 #include <stdio.h>
36 #include <stdlib.h>
37 #include <stdint.h>
38 #include <string.h>
39 #include <stdarg.h>
40 #include <unistd.h>
41 #include <pthread.h>
42 #include <syslog.h>
43 #include <getopt.h>
44 #include <sys/file.h>
45 #include <fcntl.h>
46 #include <dlfcn.h>
47 #include <stddef.h>
48 #include <errno.h>
49 #include <limits.h>
50 #include <errno.h>
51 #include <sys/mman.h>
52 #include <sys/queue.h>
53 #include <sys/io.h>
54
55 #include <rte_common.h>
56 #include <rte_debug.h>
57 #include <rte_memory.h>
58 #include <rte_memzone.h>
59 #include <rte_launch.h>
60 #include <rte_tailq.h>
61 #include <rte_eal.h>
62 #include <rte_eal_memconfig.h>
63 #include <rte_per_lcore.h>
64 #include <rte_lcore.h>
65 #include <rte_log.h>
66 #include <rte_random.h>
67 #include <rte_cycles.h>
68 #include <rte_string_fns.h>
69 #include <rte_cpuflags.h>
70 #include <rte_interrupts.h>
71 #include <rte_pci.h>
72 #include <rte_devargs.h>
73 #include <rte_common.h>
74 #include <rte_version.h>
75 #include <rte_atomic.h>
76 #include <malloc_heap.h>
77 #include <rte_eth_ring.h>
78 #include <rte_dev.h>
79
80 #include "eal_private.h"
81 #include "eal_thread.h"
82 #include "eal_internal_cfg.h"
83 #include "eal_filesystem.h"
84 #include "eal_hugepages.h"
85
86 #define OPT_HUGE_DIR    "huge-dir"
87 #define OPT_PROC_TYPE   "proc-type"
88 #define OPT_NO_SHCONF   "no-shconf"
89 #define OPT_NO_HPET     "no-hpet"
90 #define OPT_VMWARE_TSC_MAP   "vmware-tsc-map"
91 #define OPT_NO_PCI      "no-pci"
92 #define OPT_NO_HUGE     "no-huge"
93 #define OPT_FILE_PREFIX "file-prefix"
94 #define OPT_SOCKET_MEM  "socket-mem"
95 #define OPT_PCI_WHITELIST "pci-whitelist"
96 #define OPT_PCI_BLACKLIST "pci-blacklist"
97 #define OPT_VDEV        "vdev"
98 #define OPT_SYSLOG      "syslog"
99 #define OPT_LOG_LEVEL   "log-level"
100 #define OPT_BASE_VIRTADDR   "base-virtaddr"
101 #define OPT_XEN_DOM0    "xen-dom0"
102 #define OPT_CREATE_UIO_DEV "create-uio-dev"
103 #define OPT_VFIO_INTR    "vfio-intr"
104
105 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
106
107 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
108
109 #define HIGHEST_RPL 3
110
111 #define BITS_PER_HEX 4
112
113 /* Allow the application to print its usage message too if set */
114 static rte_usage_hook_t rte_application_usage_hook = NULL;
115
116 TAILQ_HEAD(shared_driver_list, shared_driver);
117
118 /* Definition for shared object drivers. */
119 struct shared_driver {
120         TAILQ_ENTRY(shared_driver) next;
121
122         char    name[PATH_MAX];
123         void*   lib_handle;
124 };
125
126 /* List of external loadable drivers */
127 static struct shared_driver_list solib_list =
128 TAILQ_HEAD_INITIALIZER(solib_list);
129
130 /* early configuration structure, when memory config is not mmapped */
131 static struct rte_mem_config early_mem_config;
132
133 /* define fd variable here, because file needs to be kept open for the
134  * duration of the program, as we hold a write lock on it in the primary proc */
135 static int mem_cfg_fd = -1;
136
137 static struct flock wr_lock = {
138                 .l_type = F_WRLCK,
139                 .l_whence = SEEK_SET,
140                 .l_start = offsetof(struct rte_mem_config, memseg),
141                 .l_len = sizeof(early_mem_config.memseg),
142 };
143
144 /* Address of global and public configuration */
145 static struct rte_config rte_config = {
146                 .mem_config = &early_mem_config,
147 };
148
149 /* internal configuration (per-core) */
150 struct lcore_config lcore_config[RTE_MAX_LCORE];
151
152 /* internal configuration */
153 struct internal_config internal_config;
154
155 /* used by rte_rdtsc() */
156 int rte_cycles_vmware_tsc_map;
157
158 /* Return a pointer to the configuration structure */
159 struct rte_config *
160 rte_eal_get_configuration(void)
161 {
162         return &rte_config;
163 }
164
165 /* parse a sysfs (or other) file containing one integer value */
166 int
167 eal_parse_sysfs_value(const char *filename, unsigned long *val)
168 {
169         FILE *f;
170         char buf[BUFSIZ];
171         char *end = NULL;
172
173         if ((f = fopen(filename, "r")) == NULL) {
174                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
175                         __func__, filename);
176                 return -1;
177         }
178
179         if (fgets(buf, sizeof(buf), f) == NULL) {
180                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
181                         __func__, filename);
182                 fclose(f);
183                 return -1;
184         }
185         *val = strtoul(buf, &end, 0);
186         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
187                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
188                                 __func__, filename);
189                 fclose(f);
190                 return -1;
191         }
192         fclose(f);
193         return 0;
194 }
195
196
197 /* create memory configuration in shared/mmap memory. Take out
198  * a write lock on the memsegs, so we can auto-detect primary/secondary.
199  * This means we never close the file while running (auto-close on exit).
200  * We also don't lock the whole file, so that in future we can use read-locks
201  * on other parts, e.g. memzones, to detect if there are running secondary
202  * processes. */
203 static void
204 rte_eal_config_create(void)
205 {
206         void *rte_mem_cfg_addr;
207         int retval;
208
209         const char *pathname = eal_runtime_config_path();
210
211         if (internal_config.no_shconf)
212                 return;
213
214         /* map the config before hugepage address so that we don't waste a page */
215         if (internal_config.base_virtaddr != 0)
216                 rte_mem_cfg_addr = (void *)
217                         RTE_ALIGN_FLOOR(internal_config.base_virtaddr -
218                         sizeof(struct rte_mem_config), sysconf(_SC_PAGE_SIZE));
219         else
220                 rte_mem_cfg_addr = NULL;
221
222         if (mem_cfg_fd < 0){
223                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
224                 if (mem_cfg_fd < 0)
225                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
226         }
227
228         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
229         if (retval < 0){
230                 close(mem_cfg_fd);
231                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
232         }
233
234         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
235         if (retval < 0){
236                 close(mem_cfg_fd);
237                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
238                                 "process running?\n", pathname);
239         }
240
241         rte_mem_cfg_addr = mmap(rte_mem_cfg_addr, sizeof(*rte_config.mem_config),
242                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
243
244         if (rte_mem_cfg_addr == MAP_FAILED){
245                 rte_panic("Cannot mmap memory for rte_config\n");
246         }
247         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
248         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
249
250         /* store address of the config in the config itself so that secondary
251          * processes could later map the config into this exact location */
252         rte_config.mem_config->mem_cfg_addr = (uintptr_t) rte_mem_cfg_addr;
253
254 }
255
256 /* attach to an existing shared memory config */
257 static void
258 rte_eal_config_attach(void)
259 {
260         struct rte_mem_config *mem_config;
261
262         const char *pathname = eal_runtime_config_path();
263
264         if (internal_config.no_shconf)
265                 return;
266
267         if (mem_cfg_fd < 0){
268                 mem_cfg_fd = open(pathname, O_RDWR);
269                 if (mem_cfg_fd < 0)
270                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
271         }
272
273         /* map it as read-only first */
274         mem_config = (struct rte_mem_config *) mmap(NULL, sizeof(*mem_config),
275                         PROT_READ, MAP_SHARED, mem_cfg_fd, 0);
276         if (mem_config == MAP_FAILED)
277                 rte_panic("Cannot mmap memory for rte_config\n");
278
279         rte_config.mem_config = mem_config;
280 }
281
282 /* reattach the shared config at exact memory location primary process has it */
283 static void
284 rte_eal_config_reattach(void)
285 {
286         struct rte_mem_config *mem_config;
287         void *rte_mem_cfg_addr;
288
289         if (internal_config.no_shconf)
290                 return;
291
292         /* save the address primary process has mapped shared config to */
293         rte_mem_cfg_addr = (void *) (uintptr_t) rte_config.mem_config->mem_cfg_addr;
294
295         /* unmap original config */
296         munmap(rte_config.mem_config, sizeof(struct rte_mem_config));
297
298         /* remap the config at proper address */
299         mem_config = (struct rte_mem_config *) mmap(rte_mem_cfg_addr,
300                         sizeof(*mem_config), PROT_READ | PROT_WRITE, MAP_SHARED,
301                         mem_cfg_fd, 0);
302         close(mem_cfg_fd);
303         if (mem_config == MAP_FAILED || mem_config != rte_mem_cfg_addr)
304                 rte_panic("Cannot mmap memory for rte_config\n");
305
306         rte_config.mem_config = mem_config;
307 }
308
309 /* Detect if we are a primary or a secondary process */
310 static enum rte_proc_type_t
311 eal_proc_type_detect(void)
312 {
313         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
314         const char *pathname = eal_runtime_config_path();
315
316         /* if we can open the file but not get a write-lock we are a secondary
317          * process. NOTE: if we get a file handle back, we keep that open
318          * and don't close it to prevent a race condition between multiple opens */
319         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
320                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
321                 ptype = RTE_PROC_SECONDARY;
322
323         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
324                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
325
326         return ptype;
327 }
328
329 /* Sets up rte_config structure with the pointer to shared memory config.*/
330 static void
331 rte_config_init(void)
332 {
333         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
334                         eal_proc_type_detect() : /* for auto, detect the type */
335                         internal_config.process_type; /* otherwise use what's already set */
336
337         switch (rte_config.process_type){
338         case RTE_PROC_PRIMARY:
339                 rte_eal_config_create();
340                 break;
341         case RTE_PROC_SECONDARY:
342                 rte_eal_config_attach();
343                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
344                 rte_eal_config_reattach();
345                 break;
346         case RTE_PROC_AUTO:
347         case RTE_PROC_INVALID:
348                 rte_panic("Invalid process type\n");
349         }
350 }
351
352 /* Unlocks hugepage directories that were locked by eal_hugepage_info_init */
353 static void
354 eal_hugedirs_unlock(void)
355 {
356         int i;
357
358         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
359         {
360                 /* skip uninitialized */
361                 if (internal_config.hugepage_info[i].lock_descriptor < 0)
362                         continue;
363                 /* unlock hugepage file */
364                 flock(internal_config.hugepage_info[i].lock_descriptor, LOCK_UN);
365                 close(internal_config.hugepage_info[i].lock_descriptor);
366                 /* reset the field */
367                 internal_config.hugepage_info[i].lock_descriptor = -1;
368         }
369 }
370
371 /* display usage */
372 static void
373 eal_usage(const char *prgname)
374 {
375         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
376                "[--proc-type primary|secondary|auto] \n\n"
377                "EAL options:\n"
378                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
379                "  -n NUM       : Number of memory channels\n"
380                    "  -v           : Display version information on startup\n"
381                "  -d LIB.so    : add driver (can be used multiple times)\n"
382                "  -m MB        : memory to allocate (see also --"OPT_SOCKET_MEM")\n"
383                "  -r NUM       : force number of memory ranks (don't detect)\n"
384                "  --"OPT_XEN_DOM0" : support application running on Xen Domain0 "
385                            "without hugetlbfs\n"
386                "  --"OPT_SYSLOG"     : set syslog facility\n"
387                "  --"OPT_LOG_LEVEL"  : set default log level\n"
388                "  --"OPT_SOCKET_MEM" : memory to allocate on specific\n"
389                    "                 sockets (use comma separated values)\n"
390                "  --"OPT_HUGE_DIR"   : directory where hugetlbfs is mounted\n"
391                "  --"OPT_PROC_TYPE"  : type of this process\n"
392                "  --"OPT_FILE_PREFIX": prefix for hugepage filenames\n"
393                "  --"OPT_PCI_BLACKLIST", -b: add a PCI device in black list.\n"
394                "               Prevent EAL from using this PCI device. The argument\n"
395                "               format is <domain:bus:devid.func>.\n"
396                "  --"OPT_PCI_WHITELIST", -w: add a PCI device in white list.\n"
397                "               Only use the specified PCI devices. The argument format\n"
398                "               is <[domain:]bus:devid.func>. This option can be present\n"
399                "               several times (once per device).\n"
400                "               [NOTE: PCI whitelist cannot be used with -b option]\n"
401                "  --"OPT_VDEV": add a virtual device.\n"
402                "               The argument format is <driver><id>[,key=val,...]\n"
403                "               (ex: --vdev=eth_pcap0,iface=eth2).\n"
404                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of native RDTSC\n"
405                "  --"OPT_BASE_VIRTADDR": specify base virtual address\n"
406                "  --"OPT_VFIO_INTR": specify desired interrupt mode for VFIO "
407                            "(legacy|msi|msix)\n"
408                "  --"OPT_CREATE_UIO_DEV": create /dev/uioX (usually done by hotplug)\n"
409                "\nEAL options for DEBUG use only:\n"
410                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
411                "  --"OPT_NO_PCI"   : disable pci\n"
412                "  --"OPT_NO_HPET"  : disable hpet\n"
413                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
414                "\n",
415                prgname);
416         /* Allow the application to print its usage message too if hook is set */
417         if ( rte_application_usage_hook ) {
418                 printf("===== Application Usage =====\n\n");
419                 rte_application_usage_hook(prgname);
420         }
421 }
422
423 /* Set a per-application usage message */
424 rte_usage_hook_t
425 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
426 {
427         rte_usage_hook_t        old_func;
428
429         /* Will be NULL on the first call to denote the last usage routine. */
430         old_func                                        = rte_application_usage_hook;
431         rte_application_usage_hook      = usage_func;
432
433         return old_func;
434 }
435
436 /*
437  * Parse the coremask given as argument (hexadecimal string) and fill
438  * the global configuration (core role and core count) with the parsed
439  * value.
440  */
441 static int xdigit2val(unsigned char c)
442 {
443         int val;
444         if(isdigit(c))
445                 val = c - '0';
446         else if(isupper(c))
447                 val = c - 'A' + 10;
448         else
449                 val = c - 'a' + 10;
450         return val;
451 }
452 static int
453 eal_parse_coremask(const char *coremask)
454 {
455         struct rte_config *cfg = rte_eal_get_configuration();
456         int i, j, idx = 0 ;
457         unsigned count = 0;
458         char c;
459         int val;
460
461         if (coremask == NULL)
462                 return -1;
463         /* Remove all blank characters ahead and after .
464          * Remove 0x/0X if exists.
465          */
466         while (isblank(*coremask))
467                 coremask++;
468         if (coremask[0] == '0' && ((coremask[1] == 'x')
469                 ||  (coremask[1] == 'X')) )
470                 coremask += 2;
471         i = strnlen(coremask, PATH_MAX);
472         while ((i > 0) && isblank(coremask[i - 1]))
473                 i--;
474         if (i == 0)
475                 return -1;
476
477         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
478                 c = coremask[i];
479                 if (isxdigit(c) == 0) {
480                         /* invalid characters */
481                         return (-1);
482                 }
483                 val = xdigit2val(c);
484                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
485                         if((1 << j) & val) {
486                                 if (!lcore_config[idx].detected) {
487                                         RTE_LOG(ERR, EAL, "lcore %u "
488                                                 "unavailable\n", idx);
489                                         return -1;
490                                 }
491                                 cfg->lcore_role[idx] = ROLE_RTE;
492                                 if(count == 0)
493                                         cfg->master_lcore = idx;
494                                 count++;
495                         } else  {
496                                 cfg->lcore_role[idx] = ROLE_OFF;
497                         }
498                 }
499         }
500         for(; i >= 0; i--)
501                 if(coremask[i] != '0')
502                         return -1;
503         for(; idx < RTE_MAX_LCORE; idx++)
504                 cfg->lcore_role[idx] = ROLE_OFF;
505         if(count == 0)
506                 return -1;
507         /* Update the count of enabled logical cores of the EAL configuration */
508         cfg->lcore_count = count;
509         return 0;
510 }
511
512 static int
513 eal_parse_syslog(const char *facility)
514 {
515         int i;
516         static struct {
517                 const char *name;
518                 int value;
519         } map[] = {
520                 { "auth", LOG_AUTH },
521                 { "cron", LOG_CRON },
522                 { "daemon", LOG_DAEMON },
523                 { "ftp", LOG_FTP },
524                 { "kern", LOG_KERN },
525                 { "lpr", LOG_LPR },
526                 { "mail", LOG_MAIL },
527                 { "news", LOG_NEWS },
528                 { "syslog", LOG_SYSLOG },
529                 { "user", LOG_USER },
530                 { "uucp", LOG_UUCP },
531                 { "local0", LOG_LOCAL0 },
532                 { "local1", LOG_LOCAL1 },
533                 { "local2", LOG_LOCAL2 },
534                 { "local3", LOG_LOCAL3 },
535                 { "local4", LOG_LOCAL4 },
536                 { "local5", LOG_LOCAL5 },
537                 { "local6", LOG_LOCAL6 },
538                 { "local7", LOG_LOCAL7 },
539                 { NULL, 0 }
540         };
541
542         for (i = 0; map[i].name; i++) {
543                 if (!strcmp(facility, map[i].name)) {
544                         internal_config.syslog_facility = map[i].value;
545                         return 0;
546                 }
547         }
548         return -1;
549 }
550
551 static int
552 eal_parse_log_level(const char *level, uint32_t *log_level)
553 {
554         char *end;
555         unsigned long tmp;
556
557         errno = 0;
558         tmp = strtoul(level, &end, 0);
559
560         /* check for errors */
561         if ((errno != 0) || (level[0] == '\0') ||
562             end == NULL || (*end != '\0'))
563                 return -1;
564
565         /* log_level is a uint32_t */
566         if (tmp >= UINT32_MAX)
567                 return -1;
568
569         *log_level = tmp;
570         return 0;
571 }
572
573 static int
574 eal_parse_socket_mem(char *socket_mem)
575 {
576         char * arg[RTE_MAX_NUMA_NODES];
577         char *end;
578         int arg_num, i, len;
579         uint64_t total_mem = 0;
580
581         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
582         if (len == SOCKET_MEM_STRLEN) {
583                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
584                 return -1;
585         }
586
587         /* all other error cases will be caught later */
588         if (!isdigit(socket_mem[len-1]))
589                 return -1;
590
591         /* split the optarg into separate socket values */
592         arg_num = rte_strsplit(socket_mem, len,
593                         arg, RTE_MAX_NUMA_NODES, ',');
594
595         /* if split failed, or 0 arguments */
596         if (arg_num <= 0)
597                 return -1;
598
599         internal_config.force_sockets = 1;
600
601         /* parse each defined socket option */
602         errno = 0;
603         for (i = 0; i < arg_num; i++) {
604                 end = NULL;
605                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
606
607                 /* check for invalid input */
608                 if ((errno != 0)  ||
609                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
610                         return -1;
611                 internal_config.socket_mem[i] *= 1024ULL;
612                 internal_config.socket_mem[i] *= 1024ULL;
613                 total_mem += internal_config.socket_mem[i];
614         }
615
616         /* check if we have a positive amount of total memory */
617         if (total_mem == 0)
618                 return -1;
619
620         return 0;
621 }
622
623 static int
624 eal_parse_base_virtaddr(const char *arg)
625 {
626         char *end;
627         uint64_t addr;
628
629         errno = 0;
630         addr = strtoull(arg, &end, 16);
631
632         /* check for errors */
633         if ((errno != 0) || (arg[0] == '\0') || end == NULL || (*end != '\0'))
634                 return -1;
635
636         /* make sure we don't exceed 32-bit boundary on 32-bit target */
637 #ifndef RTE_ARCH_X86_64
638         if (addr >= UINTPTR_MAX)
639                 return -1;
640 #endif
641
642         /* align the addr on 2M boundary */
643         internal_config.base_virtaddr = RTE_PTR_ALIGN_CEIL((uintptr_t)addr,
644                                                            RTE_PGSIZE_2M);
645
646         return 0;
647 }
648
649 static int
650 eal_parse_vfio_intr(const char *mode)
651 {
652         unsigned i;
653         static struct {
654                 const char *name;
655                 enum rte_intr_mode value;
656         } map[] = {
657                 { "legacy", RTE_INTR_MODE_LEGACY },
658                 { "msi", RTE_INTR_MODE_MSI },
659                 { "msix", RTE_INTR_MODE_MSIX },
660         };
661
662         for (i = 0; i < RTE_DIM(map); i++) {
663                 if (!strcmp(mode, map[i].name)) {
664                         internal_config.vfio_intr_mode = map[i].value;
665                         return 0;
666                 }
667         }
668         return -1;
669 }
670
671 static inline size_t
672 eal_get_hugepage_mem_size(void)
673 {
674         uint64_t size = 0;
675         unsigned i, j;
676
677         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
678                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
679                 if (hpi->hugedir != NULL) {
680                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
681                                 size += hpi->hugepage_sz * hpi->num_pages[j];
682                         }
683                 }
684         }
685
686         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
687 }
688
689 static enum rte_proc_type_t
690 eal_parse_proc_type(const char *arg)
691 {
692         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
693                 return RTE_PROC_PRIMARY;
694         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
695                 return RTE_PROC_SECONDARY;
696         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
697                 return RTE_PROC_AUTO;
698
699         return RTE_PROC_INVALID;
700 }
701
702 /* Parse the argument given in the command line of the application */
703 static int
704 eal_parse_args(int argc, char **argv)
705 {
706         int opt, ret, i;
707         char **argvopt;
708         int option_index;
709         int coremask_ok = 0;
710         char *prgname = argv[0];
711         static struct option lgopts[] = {
712                 {OPT_NO_HUGE, 0, 0, 0},
713                 {OPT_NO_PCI, 0, 0, 0},
714                 {OPT_NO_HPET, 0, 0, 0},
715                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
716                 {OPT_HUGE_DIR, 1, 0, 0},
717                 {OPT_NO_SHCONF, 0, 0, 0},
718                 {OPT_PROC_TYPE, 1, 0, 0},
719                 {OPT_FILE_PREFIX, 1, 0, 0},
720                 {OPT_SOCKET_MEM, 1, 0, 0},
721                 {OPT_PCI_WHITELIST, 1, 0, 0},
722                 {OPT_PCI_BLACKLIST, 1, 0, 0},
723                 {OPT_VDEV, 1, 0, 0},
724                 {OPT_SYSLOG, 1, NULL, 0},
725                 {OPT_LOG_LEVEL, 1, NULL, 0},
726                 {OPT_VFIO_INTR, 1, NULL, 0},
727                 {OPT_BASE_VIRTADDR, 1, 0, 0},
728                 {OPT_XEN_DOM0, 0, 0, 0},
729                 {OPT_CREATE_UIO_DEV, 1, NULL, 0},
730                 {0, 0, 0, 0}
731         };
732         struct shared_driver *solib;
733
734         argvopt = argv;
735
736         internal_config.memory = 0;
737         internal_config.force_nrank = 0;
738         internal_config.force_nchannel = 0;
739         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
740         internal_config.hugepage_dir = NULL;
741         internal_config.force_sockets = 0;
742         internal_config.syslog_facility = LOG_DAEMON;
743         /* default value from build option */
744         internal_config.log_level = RTE_LOG_LEVEL;
745         internal_config.xen_dom0_support = 0;
746         /* if set to NONE, interrupt mode is determined automatically */
747         internal_config.vfio_intr_mode = RTE_INTR_MODE_NONE;
748 #ifdef RTE_LIBEAL_USE_HPET
749         internal_config.no_hpet = 0;
750 #else
751         internal_config.no_hpet = 1;
752 #endif
753         /* zero out the NUMA config */
754         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
755                 internal_config.socket_mem[i] = 0;
756
757         /* zero out hugedir descriptors */
758         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
759                 internal_config.hugepage_info[i].lock_descriptor = -1;
760
761         internal_config.vmware_tsc_map = 0;
762         internal_config.base_virtaddr = 0;
763
764         while ((opt = getopt_long(argc, argvopt, "b:w:c:d:m:n:r:v",
765                                   lgopts, &option_index)) != EOF) {
766
767                 switch (opt) {
768                 /* blacklist */
769                 case 'b':
770                         if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
771                                         optarg) < 0) {
772                                 eal_usage(prgname);
773                                 return (-1);
774                         }
775                         break;
776                 /* whitelist */
777                 case 'w':
778                         if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
779                                         optarg) < 0) {
780                                 eal_usage(prgname);
781                                 return -1;
782                         }
783                         break;
784                 /* coremask */
785                 case 'c':
786                         if (eal_parse_coremask(optarg) < 0) {
787                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
788                                 eal_usage(prgname);
789                                 return -1;
790                         }
791                         coremask_ok = 1;
792                         break;
793                 /* force loading of external driver */
794                 case 'd':
795                         solib = malloc(sizeof(*solib));
796                         if (solib == NULL) {
797                                 RTE_LOG(ERR, EAL, "malloc(solib) failed\n");
798                                 return -1;
799                         }
800                         memset(solib, 0, sizeof(*solib));
801                         strncpy(solib->name, optarg, PATH_MAX-1);
802                         solib->name[PATH_MAX-1] = 0;
803                         TAILQ_INSERT_TAIL(&solib_list, solib, next);
804                         break;
805                 /* size of memory */
806                 case 'm':
807                         internal_config.memory = atoi(optarg);
808                         internal_config.memory *= 1024ULL;
809                         internal_config.memory *= 1024ULL;
810                         break;
811                 /* force number of channels */
812                 case 'n':
813                         internal_config.force_nchannel = atoi(optarg);
814                         if (internal_config.force_nchannel == 0 ||
815                             internal_config.force_nchannel > 4) {
816                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
817                                 eal_usage(prgname);
818                                 return -1;
819                         }
820                         break;
821                 /* force number of ranks */
822                 case 'r':
823                         internal_config.force_nrank = atoi(optarg);
824                         if (internal_config.force_nrank == 0 ||
825                             internal_config.force_nrank > 16) {
826                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
827                                 eal_usage(prgname);
828                                 return -1;
829                         }
830                         break;
831                 case 'v':
832                         /* since message is explicitly requested by user, we
833                          * write message at highest log level so it can always be seen
834                          * even if info or warning messages are disabled */
835                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
836                         break;
837
838                 /* long options */
839                 case 0:
840                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
841                                 internal_config.no_hugetlbfs = 1;
842                         }
843                         if (!strcmp(lgopts[option_index].name, OPT_XEN_DOM0)) {
844                 #ifdef RTE_LIBRTE_XEN_DOM0
845                                 internal_config.xen_dom0_support = 1;
846                 #else
847                                 RTE_LOG(ERR, EAL, "Can't support DPDK app "
848                                         "running on Dom0, please configure"
849                                         " RTE_LIBRTE_XEN_DOM0=y\n");
850                                 return -1;
851                 #endif
852                         }
853                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
854                                 internal_config.no_pci = 1;
855                         }
856                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
857                                 internal_config.no_hpet = 1;
858                         }
859                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
860                                 internal_config.vmware_tsc_map = 1;
861                         }
862                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
863                                 internal_config.no_shconf = 1;
864                         }
865                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
866                                 internal_config.hugepage_dir = optarg;
867                         }
868                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
869                                 internal_config.process_type = eal_parse_proc_type(optarg);
870                         }
871                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
872                                 internal_config.hugefile_prefix = optarg;
873                         }
874                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
875                                 if (eal_parse_socket_mem(optarg) < 0) {
876                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
877                                                         OPT_SOCKET_MEM "\n");
878                                         eal_usage(prgname);
879                                         return -1;
880                                 }
881                         }
882                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_BLACKLIST)) {
883                                 if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
884                                                 optarg) < 0) {
885                                         eal_usage(prgname);
886                                         return -1;
887                                 }
888                         }
889                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_WHITELIST)) {
890                                 if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
891                                                 optarg) < 0) {
892                                         eal_usage(prgname);
893                                         return -1;
894                                 }
895                         }
896                         else if (!strcmp(lgopts[option_index].name, OPT_VDEV)) {
897                                 if (rte_eal_devargs_add(RTE_DEVTYPE_VIRTUAL,
898                                                 optarg) < 0) {
899                                         eal_usage(prgname);
900                                         return -1;
901                                 }
902                         }
903                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
904                                 if (eal_parse_syslog(optarg) < 0) {
905                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
906                                                         OPT_SYSLOG "\n");
907                                         eal_usage(prgname);
908                                         return -1;
909                                 }
910                         } else if (!strcmp(lgopts[option_index].name,
911                                          OPT_LOG_LEVEL)) {
912                                 uint32_t log;
913
914                                 if (eal_parse_log_level(optarg, &log) < 0) {
915                                         RTE_LOG(ERR, EAL,
916                                                 "invalid parameters for --"
917                                                 OPT_LOG_LEVEL "\n");
918                                         eal_usage(prgname);
919                                         return -1;
920                                 }
921                                 internal_config.log_level = log;
922                         }
923                         else if (!strcmp(lgopts[option_index].name, OPT_BASE_VIRTADDR)) {
924                                 if (eal_parse_base_virtaddr(optarg) < 0) {
925                                         RTE_LOG(ERR, EAL, "invalid parameter for --"
926                                                         OPT_BASE_VIRTADDR "\n");
927                                         eal_usage(prgname);
928                                         return -1;
929                                 }
930                         }
931                         else if (!strcmp(lgopts[option_index].name, OPT_VFIO_INTR)) {
932                                 if (eal_parse_vfio_intr(optarg) < 0) {
933                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
934                                                         OPT_VFIO_INTR "\n");
935                                         eal_usage(prgname);
936                                         return -1;
937                                 }
938                         }
939                         else if (!strcmp(lgopts[option_index].name, OPT_CREATE_UIO_DEV)) {
940                                 internal_config.create_uio_dev = 1;
941                         } else {
942                                 RTE_LOG(ERR, EAL, "Option %s is not supported "
943                                         "on Linux\n",
944                                         lgopts[option_index].name);
945                                 eal_usage(prgname);
946                                 return -1;
947                         }
948                         break;
949
950                 default:
951                         eal_usage(prgname);
952                         return -1;
953                 }
954         }
955
956         /* sanity checks */
957         if (!coremask_ok) {
958                 RTE_LOG(ERR, EAL, "coremask not specified\n");
959                 eal_usage(prgname);
960                 return -1;
961         }
962         if (internal_config.process_type == RTE_PROC_AUTO){
963                 internal_config.process_type = eal_proc_type_detect();
964         }
965         if (internal_config.process_type == RTE_PROC_INVALID){
966                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
967                 eal_usage(prgname);
968                 return -1;
969         }
970         if (internal_config.process_type == RTE_PROC_PRIMARY &&
971                         internal_config.force_nchannel == 0) {
972                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
973                 eal_usage(prgname);
974                 return -1;
975         }
976         if (index(internal_config.hugefile_prefix,'%') != NULL){
977                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
978                 eal_usage(prgname);
979                 return -1;
980         }
981         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
982                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
983                                 "at the same time\n");
984                 eal_usage(prgname);
985                 return -1;
986         }
987         /* --no-huge doesn't make sense with either -m or --socket-mem */
988         if (internal_config.no_hugetlbfs &&
989                         (internal_config.memory > 0 ||
990                                         internal_config.force_sockets == 1)) {
991                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
992                                 "together with --no-huge!\n");
993                 eal_usage(prgname);
994                 return -1;
995         }
996         /* --xen-dom0 doesn't make sense with --socket-mem */
997         if (internal_config.xen_dom0_support && internal_config.force_sockets == 1) {
998                 RTE_LOG(ERR, EAL, "Options --socket-mem cannot be specified "
999                                         "together with --xen_dom0!\n");
1000                 eal_usage(prgname);
1001                 return -1;
1002         }
1003
1004         if (rte_eal_devargs_type_count(RTE_DEVTYPE_WHITELISTED_PCI) != 0 &&
1005                 rte_eal_devargs_type_count(RTE_DEVTYPE_BLACKLISTED_PCI) != 0) {
1006                 RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
1007                         "[-w] options cannot be used at the same time\n");
1008                 eal_usage(prgname);
1009                 return -1;
1010         }
1011
1012         if (optind >= 0)
1013                 argv[optind-1] = prgname;
1014
1015         /* if no memory amounts were requested, this will result in 0 and
1016          * will be overriden later, right after eal_hugepage_info_init() */
1017         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1018                 internal_config.memory += internal_config.socket_mem[i];
1019
1020         ret = optind-1;
1021         optind = 0; /* reset getopt lib */
1022         return ret;
1023 }
1024
1025 static void
1026 eal_check_mem_on_local_socket(void)
1027 {
1028         const struct rte_memseg *ms;
1029         int i, socket_id;
1030
1031         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
1032
1033         ms = rte_eal_get_physmem_layout();
1034
1035         for (i = 0; i < RTE_MAX_MEMSEG; i++)
1036                 if (ms[i].socket_id == socket_id &&
1037                                 ms[i].len > 0)
1038                         return;
1039
1040         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
1041                         "memory on local socket!\n");
1042 }
1043
1044 static int
1045 sync_func(__attribute__((unused)) void *arg)
1046 {
1047         return 0;
1048 }
1049
1050 inline static void
1051 rte_eal_mcfg_complete(void)
1052 {
1053         /* ALL shared mem_config related INIT DONE */
1054         if (rte_config.process_type == RTE_PROC_PRIMARY)
1055                 rte_config.mem_config->magic = RTE_MAGIC;
1056 }
1057
1058 /*
1059  * Request iopl privilege for all RPL, returns 0 on success
1060  */
1061 static int
1062 rte_eal_iopl_init(void)
1063 {
1064         return iopl(HIGHEST_RPL);
1065 }
1066
1067 /* Launch threads, called at application init(). */
1068 int
1069 rte_eal_init(int argc, char **argv)
1070 {
1071         int i, fctret, ret;
1072         pthread_t thread_id;
1073         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
1074         struct shared_driver *solib = NULL;
1075         const char *logid;
1076
1077         if (!rte_atomic32_test_and_set(&run_once))
1078                 return -1;
1079
1080         logid = strrchr(argv[0], '/');
1081         logid = strdup(logid ? logid + 1: argv[0]);
1082
1083         thread_id = pthread_self();
1084
1085         if (rte_eal_log_early_init() < 0)
1086                 rte_panic("Cannot init early logs\n");
1087
1088         if (rte_eal_cpu_init() < 0)
1089                 rte_panic("Cannot detect lcores\n");
1090
1091         fctret = eal_parse_args(argc, argv);
1092         if (fctret < 0)
1093                 exit(1);
1094
1095         /* set log level as early as possible */
1096         rte_set_log_level(internal_config.log_level);
1097
1098         if (internal_config.no_hugetlbfs == 0 &&
1099                         internal_config.process_type != RTE_PROC_SECONDARY &&
1100                         internal_config.xen_dom0_support == 0 &&
1101                         eal_hugepage_info_init() < 0)
1102                 rte_panic("Cannot get hugepage information\n");
1103
1104         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
1105                 if (internal_config.no_hugetlbfs)
1106                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
1107                 else
1108                         internal_config.memory = eal_get_hugepage_mem_size();
1109         }
1110
1111         if (internal_config.vmware_tsc_map == 1) {
1112 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
1113                 rte_cycles_vmware_tsc_map = 1;
1114                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
1115                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
1116 #else
1117                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
1118                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
1119 #endif
1120         }
1121
1122         rte_srand(rte_rdtsc());
1123
1124         rte_config_init();
1125
1126         if (rte_eal_iopl_init() == 0)
1127                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
1128
1129         if (rte_eal_pci_init() < 0)
1130                 rte_panic("Cannot init PCI\n");
1131
1132 #ifdef RTE_LIBRTE_IVSHMEM
1133         if (rte_eal_ivshmem_init() < 0)
1134                 rte_panic("Cannot init IVSHMEM\n");
1135 #endif
1136
1137         if (rte_eal_memory_init() < 0)
1138                 rte_panic("Cannot init memory\n");
1139
1140         /* the directories are locked during eal_hugepage_info_init */
1141         eal_hugedirs_unlock();
1142
1143         if (rte_eal_memzone_init() < 0)
1144                 rte_panic("Cannot init memzone\n");
1145
1146         if (rte_eal_tailqs_init() < 0)
1147                 rte_panic("Cannot init tail queues for objects\n");
1148
1149 #ifdef RTE_LIBRTE_IVSHMEM
1150         if (rte_eal_ivshmem_obj_init() < 0)
1151                 rte_panic("Cannot init IVSHMEM objects\n");
1152 #endif
1153
1154         if (rte_eal_log_init(logid, internal_config.syslog_facility) < 0)
1155                 rte_panic("Cannot init logs\n");
1156
1157         if (rte_eal_alarm_init() < 0)
1158                 rte_panic("Cannot init interrupt-handling thread\n");
1159
1160         if (rte_eal_intr_init() < 0)
1161                 rte_panic("Cannot init interrupt-handling thread\n");
1162
1163         if (rte_eal_timer_init() < 0)
1164                 rte_panic("Cannot init HPET or TSC timers\n");
1165
1166         eal_check_mem_on_local_socket();
1167
1168         rte_eal_mcfg_complete();
1169
1170         TAILQ_FOREACH(solib, &solib_list, next) {
1171                 RTE_LOG(INFO, EAL, "open shared lib %s\n", solib->name);
1172                 solib->lib_handle = dlopen(solib->name, RTLD_NOW);
1173                 if (solib->lib_handle == NULL)
1174                         RTE_LOG(WARNING, EAL, "%s\n", dlerror());
1175         }
1176
1177         eal_thread_init_master(rte_config.master_lcore);
1178
1179         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
1180                 rte_config.master_lcore, (int)thread_id);
1181
1182         if (rte_eal_dev_init(PMD_INIT_PRE_PCI_PROBE) < 0)
1183                 rte_panic("Cannot init pmd devices\n");
1184
1185         RTE_LCORE_FOREACH_SLAVE(i) {
1186
1187                 /*
1188                  * create communication pipes between master thread
1189                  * and children
1190                  */
1191                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
1192                         rte_panic("Cannot create pipe\n");
1193                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
1194                         rte_panic("Cannot create pipe\n");
1195
1196                 lcore_config[i].state = WAIT;
1197
1198                 /* create a thread for each lcore */
1199                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
1200                                      eal_thread_loop, NULL);
1201                 if (ret != 0)
1202                         rte_panic("Cannot create thread\n");
1203         }
1204
1205         /*
1206          * Launch a dummy function on all slave lcores, so that master lcore
1207          * knows they are all ready when this function returns.
1208          */
1209         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
1210         rte_eal_mp_wait_lcore();
1211
1212         /* Probe & Initialize PCI devices */
1213         if (rte_eal_pci_probe())
1214                         rte_panic("Cannot probe PCI\n");
1215
1216         /* Initialize any outstanding devices */
1217         if (rte_eal_dev_init(PMD_INIT_POST_PCI_PROBE) < 0)
1218                 rte_panic("Cannot init pmd devices\n");
1219
1220         return fctret;
1221 }
1222
1223 /* get core role */
1224 enum rte_lcore_role_t
1225 rte_eal_lcore_role(unsigned lcore_id)
1226 {
1227         return (rte_config.lcore_role[lcore_id]);
1228 }
1229
1230 enum rte_proc_type_t
1231 rte_eal_process_type(void)
1232 {
1233         return (rte_config.process_type);
1234 }
1235
1236 int rte_eal_has_hugepages(void)
1237 {
1238         return ! internal_config.no_hugetlbfs;
1239 }