eal_memory.c revision 32e04ea0
1/*-
2 *   BSD LICENSE
3 *
4 *   Copyright(c) 2010-2014 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/*   BSD LICENSE
34 *
35 *   Copyright(c) 2013 6WIND.
36 *
37 *   Redistribution and use in source and binary forms, with or without
38 *   modification, are permitted provided that the following conditions
39 *   are met:
40 *
41 *     * Redistributions of source code must retain the above copyright
42 *       notice, this list of conditions and the following disclaimer.
43 *     * Redistributions in binary form must reproduce the above copyright
44 *       notice, this list of conditions and the following disclaimer in
45 *       the documentation and/or other materials provided with the
46 *       distribution.
47 *     * Neither the name of 6WIND S.A. nor the names of its
48 *       contributors may be used to endorse or promote products derived
49 *       from this software without specific prior written permission.
50 *
51 *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
52 *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
53 *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
54 *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
55 *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
56 *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
57 *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
58 *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
59 *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
60 *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
61 *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
62 */
63
64#define _FILE_OFFSET_BITS 64
65#include <errno.h>
66#include <stdarg.h>
67#include <stdlib.h>
68#include <stdio.h>
69#include <stdint.h>
70#include <inttypes.h>
71#include <string.h>
72#include <stdarg.h>
73#include <sys/mman.h>
74#include <sys/types.h>
75#include <sys/stat.h>
76#include <sys/queue.h>
77#include <sys/file.h>
78#include <unistd.h>
79#include <limits.h>
80#include <errno.h>
81#include <sys/ioctl.h>
82#include <sys/time.h>
83#include <signal.h>
84#include <setjmp.h>
85
86#include <rte_log.h>
87#include <rte_memory.h>
88#include <rte_memzone.h>
89#include <rte_launch.h>
90#include <rte_eal.h>
91#include <rte_eal_memconfig.h>
92#include <rte_per_lcore.h>
93#include <rte_lcore.h>
94#include <rte_common.h>
95#include <rte_string_fns.h>
96
97#include "eal_private.h"
98#include "eal_internal_cfg.h"
99#include "eal_filesystem.h"
100#include "eal_hugepages.h"
101
102#define PFN_MASK_SIZE	8
103
104#ifdef RTE_LIBRTE_XEN_DOM0
105int rte_xen_dom0_supported(void)
106{
107	return internal_config.xen_dom0_support;
108}
109#endif
110
111/**
112 * @file
113 * Huge page mapping under linux
114 *
115 * To reserve a big contiguous amount of memory, we use the hugepage
116 * feature of linux. For that, we need to have hugetlbfs mounted. This
117 * code will create many files in this directory (one per page) and
118 * map them in virtual memory. For each page, we will retrieve its
119 * physical address and remap it in order to have a virtual contiguous
120 * zone as well as a physical contiguous zone.
121 */
122
123static uint64_t baseaddr_offset;
124
125static unsigned proc_pagemap_readable;
126
127#define RANDOMIZE_VA_SPACE_FILE "/proc/sys/kernel/randomize_va_space"
128
129static void
130test_proc_pagemap_readable(void)
131{
132	int fd = open("/proc/self/pagemap", O_RDONLY);
133
134	if (fd < 0) {
135		RTE_LOG(ERR, EAL,
136			"Cannot open /proc/self/pagemap: %s. "
137			"virt2phys address translation will not work\n",
138			strerror(errno));
139		return;
140	}
141
142	/* Is readable */
143	close(fd);
144	proc_pagemap_readable = 1;
145}
146
147/* Lock page in physical memory and prevent from swapping. */
148int
149rte_mem_lock_page(const void *virt)
150{
151	unsigned long virtual = (unsigned long)virt;
152	int page_size = getpagesize();
153	unsigned long aligned = (virtual & ~ (page_size - 1));
154	return mlock((void*)aligned, page_size);
155}
156
157/*
158 * Get physical address of any mapped virtual address in the current process.
159 */
160phys_addr_t
161rte_mem_virt2phy(const void *virtaddr)
162{
163	int fd, retval;
164	uint64_t page, physaddr;
165	unsigned long virt_pfn;
166	int page_size;
167	off_t offset;
168
169	/* when using dom0, /proc/self/pagemap always returns 0, check in
170	 * dpdk memory by browsing the memsegs */
171	if (rte_xen_dom0_supported()) {
172		struct rte_mem_config *mcfg;
173		struct rte_memseg *memseg;
174		unsigned i;
175
176		mcfg = rte_eal_get_configuration()->mem_config;
177		for (i = 0; i < RTE_MAX_MEMSEG; i++) {
178			memseg = &mcfg->memseg[i];
179			if (memseg->addr == NULL)
180				break;
181			if (virtaddr > memseg->addr &&
182					virtaddr < RTE_PTR_ADD(memseg->addr,
183						memseg->len)) {
184				return memseg->phys_addr +
185					RTE_PTR_DIFF(virtaddr, memseg->addr);
186			}
187		}
188
189		return RTE_BAD_PHYS_ADDR;
190	}
191
192	/* Cannot parse /proc/self/pagemap, no need to log errors everywhere */
193	if (!proc_pagemap_readable)
194		return RTE_BAD_PHYS_ADDR;
195
196	/* standard page size */
197	page_size = getpagesize();
198
199	fd = open("/proc/self/pagemap", O_RDONLY);
200	if (fd < 0) {
201		RTE_LOG(ERR, EAL, "%s(): cannot open /proc/self/pagemap: %s\n",
202			__func__, strerror(errno));
203		return RTE_BAD_PHYS_ADDR;
204	}
205
206	virt_pfn = (unsigned long)virtaddr / page_size;
207	offset = sizeof(uint64_t) * virt_pfn;
208	if (lseek(fd, offset, SEEK_SET) == (off_t) -1) {
209		RTE_LOG(ERR, EAL, "%s(): seek error in /proc/self/pagemap: %s\n",
210				__func__, strerror(errno));
211		close(fd);
212		return RTE_BAD_PHYS_ADDR;
213	}
214
215	retval = read(fd, &page, PFN_MASK_SIZE);
216	close(fd);
217	if (retval < 0) {
218		RTE_LOG(ERR, EAL, "%s(): cannot read /proc/self/pagemap: %s\n",
219				__func__, strerror(errno));
220		return RTE_BAD_PHYS_ADDR;
221	} else if (retval != PFN_MASK_SIZE) {
222		RTE_LOG(ERR, EAL, "%s(): read %d bytes from /proc/self/pagemap "
223				"but expected %d:\n",
224				__func__, retval, PFN_MASK_SIZE);
225		return RTE_BAD_PHYS_ADDR;
226	}
227
228	/*
229	 * the pfn (page frame number) are bits 0-54 (see
230	 * pagemap.txt in linux Documentation)
231	 */
232	physaddr = ((page & 0x7fffffffffffffULL) * page_size)
233		+ ((unsigned long)virtaddr % page_size);
234
235	return physaddr;
236}
237
238/*
239 * For each hugepage in hugepg_tbl, fill the physaddr value. We find
240 * it by browsing the /proc/self/pagemap special file.
241 */
242static int
243find_physaddrs(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
244{
245	unsigned i;
246	phys_addr_t addr;
247
248	for (i = 0; i < hpi->num_pages[0]; i++) {
249		addr = rte_mem_virt2phy(hugepg_tbl[i].orig_va);
250		if (addr == RTE_BAD_PHYS_ADDR)
251			return -1;
252		hugepg_tbl[i].physaddr = addr;
253	}
254	return 0;
255}
256
257/*
258 * Check whether address-space layout randomization is enabled in
259 * the kernel. This is important for multi-process as it can prevent
260 * two processes mapping data to the same virtual address
261 * Returns:
262 *    0 - address space randomization disabled
263 *    1/2 - address space randomization enabled
264 *    negative error code on error
265 */
266static int
267aslr_enabled(void)
268{
269	char c;
270	int retval, fd = open(RANDOMIZE_VA_SPACE_FILE, O_RDONLY);
271	if (fd < 0)
272		return -errno;
273	retval = read(fd, &c, 1);
274	close(fd);
275	if (retval < 0)
276		return -errno;
277	if (retval == 0)
278		return -EIO;
279	switch (c) {
280		case '0' : return 0;
281		case '1' : return 1;
282		case '2' : return 2;
283		default: return -EINVAL;
284	}
285}
286
287/*
288 * Try to mmap *size bytes in /dev/zero. If it is successful, return the
289 * pointer to the mmap'd area and keep *size unmodified. Else, retry
290 * with a smaller zone: decrease *size by hugepage_sz until it reaches
291 * 0. In this case, return NULL. Note: this function returns an address
292 * which is a multiple of hugepage size.
293 */
294static void *
295get_virtual_area(size_t *size, size_t hugepage_sz)
296{
297	void *addr;
298	int fd;
299	long aligned_addr;
300
301	if (internal_config.base_virtaddr != 0) {
302		addr = (void*) (uintptr_t) (internal_config.base_virtaddr +
303				baseaddr_offset);
304	}
305	else addr = NULL;
306
307	RTE_LOG(DEBUG, EAL, "Ask a virtual area of 0x%zx bytes\n", *size);
308
309	fd = open("/dev/zero", O_RDONLY);
310	if (fd < 0){
311		RTE_LOG(ERR, EAL, "Cannot open /dev/zero\n");
312		return NULL;
313	}
314	do {
315		addr = mmap(addr,
316				(*size) + hugepage_sz, PROT_READ, MAP_PRIVATE, fd, 0);
317		if (addr == MAP_FAILED)
318			*size -= hugepage_sz;
319	} while (addr == MAP_FAILED && *size > 0);
320
321	if (addr == MAP_FAILED) {
322		close(fd);
323		RTE_LOG(ERR, EAL, "Cannot get a virtual area: %s\n",
324			strerror(errno));
325		return NULL;
326	}
327
328	munmap(addr, (*size) + hugepage_sz);
329	close(fd);
330
331	/* align addr to a huge page size boundary */
332	aligned_addr = (long)addr;
333	aligned_addr += (hugepage_sz - 1);
334	aligned_addr &= (~(hugepage_sz - 1));
335	addr = (void *)(aligned_addr);
336
337	RTE_LOG(DEBUG, EAL, "Virtual area found at %p (size = 0x%zx)\n",
338		addr, *size);
339
340	/* increment offset */
341	baseaddr_offset += *size;
342
343	return addr;
344}
345
346static sigjmp_buf huge_jmpenv;
347
348static void huge_sigbus_handler(int signo __rte_unused)
349{
350	siglongjmp(huge_jmpenv, 1);
351}
352
353/* Put setjmp into a wrap method to avoid compiling error. Any non-volatile,
354 * non-static local variable in the stack frame calling sigsetjmp might be
355 * clobbered by a call to longjmp.
356 */
357static int huge_wrap_sigsetjmp(void)
358{
359	return sigsetjmp(huge_jmpenv, 1);
360}
361
362/*
363 * Mmap all hugepages of hugepage table: it first open a file in
364 * hugetlbfs, then mmap() hugepage_sz data in it. If orig is set, the
365 * virtual address is stored in hugepg_tbl[i].orig_va, else it is stored
366 * in hugepg_tbl[i].final_va. The second mapping (when orig is 0) tries to
367 * map continguous physical blocks in contiguous virtual blocks.
368 */
369static unsigned
370map_all_hugepages(struct hugepage_file *hugepg_tbl,
371		struct hugepage_info *hpi, int orig)
372{
373	int fd;
374	unsigned i;
375	void *virtaddr;
376	void *vma_addr = NULL;
377	size_t vma_len = 0;
378
379#ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
380	RTE_SET_USED(vma_len);
381#endif
382
383	for (i = 0; i < hpi->num_pages[0]; i++) {
384		uint64_t hugepage_sz = hpi->hugepage_sz;
385
386		if (orig) {
387			hugepg_tbl[i].file_id = i;
388			hugepg_tbl[i].size = hugepage_sz;
389#ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
390			eal_get_hugefile_temp_path(hugepg_tbl[i].filepath,
391					sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
392					hugepg_tbl[i].file_id);
393#else
394			eal_get_hugefile_path(hugepg_tbl[i].filepath,
395					sizeof(hugepg_tbl[i].filepath), hpi->hugedir,
396					hugepg_tbl[i].file_id);
397#endif
398			hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
399		}
400#ifndef RTE_ARCH_64
401		/* for 32-bit systems, don't remap 1G and 16G pages, just reuse
402		 * original map address as final map address.
403		 */
404		else if ((hugepage_sz == RTE_PGSIZE_1G)
405			|| (hugepage_sz == RTE_PGSIZE_16G)) {
406			hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
407			hugepg_tbl[i].orig_va = NULL;
408			continue;
409		}
410#endif
411
412#ifndef RTE_EAL_SINGLE_FILE_SEGMENTS
413		else if (vma_len == 0) {
414			unsigned j, num_pages;
415
416			/* reserve a virtual area for next contiguous
417			 * physical block: count the number of
418			 * contiguous physical pages. */
419			for (j = i+1; j < hpi->num_pages[0] ; j++) {
420#ifdef RTE_ARCH_PPC_64
421				/* The physical addresses are sorted in
422				 * descending order on PPC64 */
423				if (hugepg_tbl[j].physaddr !=
424				    hugepg_tbl[j-1].physaddr - hugepage_sz)
425					break;
426#else
427				if (hugepg_tbl[j].physaddr !=
428				    hugepg_tbl[j-1].physaddr + hugepage_sz)
429					break;
430#endif
431			}
432			num_pages = j - i;
433			vma_len = num_pages * hugepage_sz;
434
435			/* get the biggest virtual memory area up to
436			 * vma_len. If it fails, vma_addr is NULL, so
437			 * let the kernel provide the address. */
438			vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
439			if (vma_addr == NULL)
440				vma_len = hugepage_sz;
441		}
442#endif
443
444		/* try to create hugepage file */
445		fd = open(hugepg_tbl[i].filepath, O_CREAT | O_RDWR, 0755);
446		if (fd < 0) {
447			RTE_LOG(DEBUG, EAL, "%s(): open failed: %s\n", __func__,
448					strerror(errno));
449			return i;
450		}
451
452		/* map the segment, and populate page tables,
453		 * the kernel fills this segment with zeros */
454		virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
455				MAP_SHARED | MAP_POPULATE, fd, 0);
456		if (virtaddr == MAP_FAILED) {
457			RTE_LOG(DEBUG, EAL, "%s(): mmap failed: %s\n", __func__,
458					strerror(errno));
459			close(fd);
460			return i;
461		}
462
463		if (orig) {
464			hugepg_tbl[i].orig_va = virtaddr;
465		}
466		else {
467			hugepg_tbl[i].final_va = virtaddr;
468		}
469
470		if (orig) {
471			/* In linux, hugetlb limitations, like cgroup, are
472			 * enforced at fault time instead of mmap(), even
473			 * with the option of MAP_POPULATE. Kernel will send
474			 * a SIGBUS signal. To avoid to be killed, save stack
475			 * environment here, if SIGBUS happens, we can jump
476			 * back here.
477			 */
478			if (huge_wrap_sigsetjmp()) {
479				RTE_LOG(DEBUG, EAL, "SIGBUS: Cannot mmap more "
480					"hugepages of size %u MB\n",
481					(unsigned)(hugepage_sz / 0x100000));
482				munmap(virtaddr, hugepage_sz);
483				close(fd);
484				unlink(hugepg_tbl[i].filepath);
485				return i;
486			}
487			*(int *)virtaddr = 0;
488		}
489
490
491		/* set shared flock on the file. */
492		if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
493			RTE_LOG(DEBUG, EAL, "%s(): Locking file failed:%s \n",
494				__func__, strerror(errno));
495			close(fd);
496			return i;
497		}
498
499		close(fd);
500
501		vma_addr = (char *)vma_addr + hugepage_sz;
502		vma_len -= hugepage_sz;
503	}
504
505	return i;
506}
507
508#ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
509
510/*
511 * Remaps all hugepages into single file segments
512 */
513static int
514remap_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
515{
516	int fd;
517	unsigned i = 0, j, num_pages, page_idx = 0;
518	void *vma_addr = NULL, *old_addr = NULL, *page_addr = NULL;
519	size_t vma_len = 0;
520	size_t hugepage_sz = hpi->hugepage_sz;
521	size_t total_size, offset;
522	char filepath[MAX_HUGEPAGE_PATH];
523	phys_addr_t physaddr;
524	int socket;
525
526	while (i < hpi->num_pages[0]) {
527
528#ifndef RTE_ARCH_64
529		/* for 32-bit systems, don't remap 1G pages and 16G pages,
530		 * just reuse original map address as final map address.
531		 */
532		if ((hugepage_sz == RTE_PGSIZE_1G)
533			|| (hugepage_sz == RTE_PGSIZE_16G)) {
534			hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
535			hugepg_tbl[i].orig_va = NULL;
536			i++;
537			continue;
538		}
539#endif
540
541		/* reserve a virtual area for next contiguous
542		 * physical block: count the number of
543		 * contiguous physical pages. */
544		for (j = i+1; j < hpi->num_pages[0] ; j++) {
545#ifdef RTE_ARCH_PPC_64
546			/* The physical addresses are sorted in descending
547			 * order on PPC64 */
548			if (hugepg_tbl[j].physaddr !=
549				hugepg_tbl[j-1].physaddr - hugepage_sz)
550				break;
551#else
552			if (hugepg_tbl[j].physaddr !=
553				hugepg_tbl[j-1].physaddr + hugepage_sz)
554				break;
555#endif
556		}
557		num_pages = j - i;
558		vma_len = num_pages * hugepage_sz;
559
560		socket = hugepg_tbl[i].socket_id;
561
562		/* get the biggest virtual memory area up to
563		 * vma_len. If it fails, vma_addr is NULL, so
564		 * let the kernel provide the address. */
565		vma_addr = get_virtual_area(&vma_len, hpi->hugepage_sz);
566
567		/* If we can't find a big enough virtual area, work out how many pages
568		 * we are going to get */
569		if (vma_addr == NULL)
570			j = i + 1;
571		else if (vma_len != num_pages * hugepage_sz) {
572			num_pages = vma_len / hugepage_sz;
573			j = i + num_pages;
574
575		}
576
577		hugepg_tbl[page_idx].file_id = page_idx;
578		eal_get_hugefile_path(filepath,
579				sizeof(filepath),
580				hpi->hugedir,
581				hugepg_tbl[page_idx].file_id);
582
583		/* try to create hugepage file */
584		fd = open(filepath, O_CREAT | O_RDWR, 0755);
585		if (fd < 0) {
586			RTE_LOG(ERR, EAL, "%s(): open failed: %s\n", __func__, strerror(errno));
587			return -1;
588		}
589
590		total_size = 0;
591		for (;i < j; i++) {
592
593			/* unmap current segment */
594			if (total_size > 0)
595				munmap(vma_addr, total_size);
596
597			/* unmap original page */
598			munmap(hugepg_tbl[i].orig_va, hugepage_sz);
599			unlink(hugepg_tbl[i].filepath);
600
601			total_size += hugepage_sz;
602
603			old_addr = vma_addr;
604
605			/* map new, bigger segment, and populate page tables,
606			 * the kernel fills this segment with zeros */
607			vma_addr = mmap(vma_addr, total_size,
608					PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, 0);
609
610			if (vma_addr == MAP_FAILED || vma_addr != old_addr) {
611				RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__, strerror(errno));
612				close(fd);
613				return -1;
614			}
615		}
616
617		/* set shared flock on the file. */
618		if (flock(fd, LOCK_SH | LOCK_NB) == -1) {
619			RTE_LOG(ERR, EAL, "%s(): Locking file failed:%s \n",
620				__func__, strerror(errno));
621			close(fd);
622			return -1;
623		}
624
625		snprintf(hugepg_tbl[page_idx].filepath, MAX_HUGEPAGE_PATH, "%s",
626				filepath);
627
628		physaddr = rte_mem_virt2phy(vma_addr);
629
630		if (physaddr == RTE_BAD_PHYS_ADDR)
631			return -1;
632
633		hugepg_tbl[page_idx].final_va = vma_addr;
634
635		hugepg_tbl[page_idx].physaddr = physaddr;
636
637		hugepg_tbl[page_idx].repeated = num_pages;
638
639		hugepg_tbl[page_idx].socket_id = socket;
640
641		close(fd);
642
643		/* verify the memory segment - that is, check that every VA corresponds
644		 * to the physical address we expect to see
645		 */
646		for (offset = 0; offset < vma_len; offset += hugepage_sz) {
647			uint64_t expected_physaddr;
648
649			expected_physaddr = hugepg_tbl[page_idx].physaddr + offset;
650			page_addr = RTE_PTR_ADD(vma_addr, offset);
651			physaddr = rte_mem_virt2phy(page_addr);
652
653			if (physaddr != expected_physaddr) {
654				RTE_LOG(ERR, EAL, "Segment sanity check failed: wrong physaddr "
655						"at %p (offset 0x%" PRIx64 ": 0x%" PRIx64
656						" (expected 0x%" PRIx64 ")\n",
657						page_addr, offset, physaddr, expected_physaddr);
658				return -1;
659			}
660		}
661
662		page_idx++;
663	}
664
665	/* zero out the rest */
666	memset(&hugepg_tbl[page_idx], 0, (hpi->num_pages[0] - page_idx) * sizeof(struct hugepage_file));
667	return page_idx;
668}
669#else/* RTE_EAL_SINGLE_FILE_SEGMENTS=n */
670
671/* Unmap all hugepages from original mapping */
672static int
673unmap_all_hugepages_orig(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
674{
675        unsigned i;
676        for (i = 0; i < hpi->num_pages[0]; i++) {
677                if (hugepg_tbl[i].orig_va) {
678                        munmap(hugepg_tbl[i].orig_va, hpi->hugepage_sz);
679                        hugepg_tbl[i].orig_va = NULL;
680                }
681        }
682        return 0;
683}
684#endif /* RTE_EAL_SINGLE_FILE_SEGMENTS */
685
686/*
687 * Parse /proc/self/numa_maps to get the NUMA socket ID for each huge
688 * page.
689 */
690static int
691find_numasocket(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
692{
693	int socket_id;
694	char *end, *nodestr;
695	unsigned i, hp_count = 0;
696	uint64_t virt_addr;
697	char buf[BUFSIZ];
698	char hugedir_str[PATH_MAX];
699	FILE *f;
700
701	f = fopen("/proc/self/numa_maps", "r");
702	if (f == NULL) {
703		RTE_LOG(NOTICE, EAL, "cannot open /proc/self/numa_maps,"
704				" consider that all memory is in socket_id 0\n");
705		return 0;
706	}
707
708	snprintf(hugedir_str, sizeof(hugedir_str),
709			"%s/%s", hpi->hugedir, internal_config.hugefile_prefix);
710
711	/* parse numa map */
712	while (fgets(buf, sizeof(buf), f) != NULL) {
713
714		/* ignore non huge page */
715		if (strstr(buf, " huge ") == NULL &&
716				strstr(buf, hugedir_str) == NULL)
717			continue;
718
719		/* get zone addr */
720		virt_addr = strtoull(buf, &end, 16);
721		if (virt_addr == 0 || end == buf) {
722			RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
723			goto error;
724		}
725
726		/* get node id (socket id) */
727		nodestr = strstr(buf, " N");
728		if (nodestr == NULL) {
729			RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
730			goto error;
731		}
732		nodestr += 2;
733		end = strstr(nodestr, "=");
734		if (end == NULL) {
735			RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
736			goto error;
737		}
738		end[0] = '\0';
739		end = NULL;
740
741		socket_id = strtoul(nodestr, &end, 0);
742		if ((nodestr[0] == '\0') || (end == NULL) || (*end != '\0')) {
743			RTE_LOG(ERR, EAL, "%s(): error in numa_maps parsing\n", __func__);
744			goto error;
745		}
746
747		/* if we find this page in our mappings, set socket_id */
748		for (i = 0; i < hpi->num_pages[0]; i++) {
749			void *va = (void *)(unsigned long)virt_addr;
750			if (hugepg_tbl[i].orig_va == va) {
751				hugepg_tbl[i].socket_id = socket_id;
752				hp_count++;
753			}
754		}
755	}
756
757	if (hp_count < hpi->num_pages[0])
758		goto error;
759
760	fclose(f);
761	return 0;
762
763error:
764	fclose(f);
765	return -1;
766}
767
768static int
769cmp_physaddr(const void *a, const void *b)
770{
771#ifndef RTE_ARCH_PPC_64
772	const struct hugepage_file *p1 = (const struct hugepage_file *)a;
773	const struct hugepage_file *p2 = (const struct hugepage_file *)b;
774#else
775	/* PowerPC needs memory sorted in reverse order from x86 */
776	const struct hugepage_file *p1 = (const struct hugepage_file *)b;
777	const struct hugepage_file *p2 = (const struct hugepage_file *)a;
778#endif
779	if (p1->physaddr < p2->physaddr)
780		return -1;
781	else if (p1->physaddr > p2->physaddr)
782		return 1;
783	else
784		return 0;
785}
786
787/*
788 * Uses mmap to create a shared memory area for storage of data
789 * Used in this file to store the hugepage file map on disk
790 */
791static void *
792create_shared_memory(const char *filename, const size_t mem_size)
793{
794	void *retval;
795	int fd = open(filename, O_CREAT | O_RDWR, 0666);
796	if (fd < 0)
797		return NULL;
798	if (ftruncate(fd, mem_size) < 0) {
799		close(fd);
800		return NULL;
801	}
802	retval = mmap(NULL, mem_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
803	close(fd);
804	return retval;
805}
806
807/*
808 * this copies *active* hugepages from one hugepage table to another.
809 * destination is typically the shared memory.
810 */
811static int
812copy_hugepages_to_shared_mem(struct hugepage_file * dst, int dest_size,
813		const struct hugepage_file * src, int src_size)
814{
815	int src_pos, dst_pos = 0;
816
817	for (src_pos = 0; src_pos < src_size; src_pos++) {
818		if (src[src_pos].final_va != NULL) {
819			/* error on overflow attempt */
820			if (dst_pos == dest_size)
821				return -1;
822			memcpy(&dst[dst_pos], &src[src_pos], sizeof(struct hugepage_file));
823			dst_pos++;
824		}
825	}
826	return 0;
827}
828
829static int
830unlink_hugepage_files(struct hugepage_file *hugepg_tbl,
831		unsigned num_hp_info)
832{
833	unsigned socket, size;
834	int page, nrpages = 0;
835
836	/* get total number of hugepages */
837	for (size = 0; size < num_hp_info; size++)
838		for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
839			nrpages +=
840			internal_config.hugepage_info[size].num_pages[socket];
841
842	for (page = 0; page < nrpages; page++) {
843		struct hugepage_file *hp = &hugepg_tbl[page];
844
845		if (hp->final_va != NULL && unlink(hp->filepath)) {
846			RTE_LOG(WARNING, EAL, "%s(): Removing %s failed: %s\n",
847				__func__, hp->filepath, strerror(errno));
848		}
849	}
850	return 0;
851}
852
853/*
854 * unmaps hugepages that are not going to be used. since we originally allocate
855 * ALL hugepages (not just those we need), additional unmapping needs to be done.
856 */
857static int
858unmap_unneeded_hugepages(struct hugepage_file *hugepg_tbl,
859		struct hugepage_info *hpi,
860		unsigned num_hp_info)
861{
862	unsigned socket, size;
863	int page, nrpages = 0;
864
865	/* get total number of hugepages */
866	for (size = 0; size < num_hp_info; size++)
867		for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++)
868			nrpages += internal_config.hugepage_info[size].num_pages[socket];
869
870	for (size = 0; size < num_hp_info; size++) {
871		for (socket = 0; socket < RTE_MAX_NUMA_NODES; socket++) {
872			unsigned pages_found = 0;
873
874			/* traverse until we have unmapped all the unused pages */
875			for (page = 0; page < nrpages; page++) {
876				struct hugepage_file *hp = &hugepg_tbl[page];
877
878#ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
879				/* if this page was already cleared */
880				if (hp->final_va == NULL)
881					continue;
882#endif
883
884				/* find a page that matches the criteria */
885				if ((hp->size == hpi[size].hugepage_sz) &&
886						(hp->socket_id == (int) socket)) {
887
888					/* if we skipped enough pages, unmap the rest */
889					if (pages_found == hpi[size].num_pages[socket]) {
890						uint64_t unmap_len;
891
892#ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
893						unmap_len = hp->size * hp->repeated;
894#else
895						unmap_len = hp->size;
896#endif
897
898						/* get start addr and len of the remaining segment */
899						munmap(hp->final_va, (size_t) unmap_len);
900
901						hp->final_va = NULL;
902						if (unlink(hp->filepath) == -1) {
903							RTE_LOG(ERR, EAL, "%s(): Removing %s failed: %s\n",
904									__func__, hp->filepath, strerror(errno));
905							return -1;
906						}
907					}
908#ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
909					/* else, check how much do we need to map */
910					else {
911						int nr_pg_left =
912								hpi[size].num_pages[socket] - pages_found;
913
914						/* if we need enough memory to fit into the segment */
915						if (hp->repeated <= nr_pg_left) {
916							pages_found += hp->repeated;
917						}
918						/* truncate the segment */
919						else {
920							uint64_t final_size = nr_pg_left * hp->size;
921							uint64_t seg_size = hp->repeated * hp->size;
922
923							void * unmap_va = RTE_PTR_ADD(hp->final_va,
924									final_size);
925							int fd;
926
927							munmap(unmap_va, seg_size - final_size);
928
929							fd = open(hp->filepath, O_RDWR);
930							if (fd < 0) {
931								RTE_LOG(ERR, EAL, "Cannot open %s: %s\n",
932										hp->filepath, strerror(errno));
933								return -1;
934							}
935							if (ftruncate(fd, final_size) < 0) {
936								RTE_LOG(ERR, EAL, "Cannot truncate %s: %s\n",
937										hp->filepath, strerror(errno));
938								return -1;
939							}
940							close(fd);
941
942							pages_found += nr_pg_left;
943							hp->repeated = nr_pg_left;
944						}
945					}
946#else
947					/* else, lock the page and skip */
948					else
949						pages_found++;
950#endif
951
952				} /* match page */
953			} /* foreach page */
954		} /* foreach socket */
955	} /* foreach pagesize */
956
957	return 0;
958}
959
960static inline uint64_t
961get_socket_mem_size(int socket)
962{
963	uint64_t size = 0;
964	unsigned i;
965
966	for (i = 0; i < internal_config.num_hugepage_sizes; i++){
967		struct hugepage_info *hpi = &internal_config.hugepage_info[i];
968		if (hpi->hugedir != NULL)
969			size += hpi->hugepage_sz * hpi->num_pages[socket];
970	}
971
972	return size;
973}
974
975/*
976 * This function is a NUMA-aware equivalent of calc_num_pages.
977 * It takes in the list of hugepage sizes and the
978 * number of pages thereof, and calculates the best number of
979 * pages of each size to fulfill the request for <memory> ram
980 */
981static int
982calc_num_pages_per_socket(uint64_t * memory,
983		struct hugepage_info *hp_info,
984		struct hugepage_info *hp_used,
985		unsigned num_hp_info)
986{
987	unsigned socket, j, i = 0;
988	unsigned requested, available;
989	int total_num_pages = 0;
990	uint64_t remaining_mem, cur_mem;
991	uint64_t total_mem = internal_config.memory;
992
993	if (num_hp_info == 0)
994		return -1;
995
996	/* if specific memory amounts per socket weren't requested */
997	if (internal_config.force_sockets == 0) {
998		int cpu_per_socket[RTE_MAX_NUMA_NODES];
999		size_t default_size, total_size;
1000		unsigned lcore_id;
1001
1002		/* Compute number of cores per socket */
1003		memset(cpu_per_socket, 0, sizeof(cpu_per_socket));
1004		RTE_LCORE_FOREACH(lcore_id) {
1005			cpu_per_socket[rte_lcore_to_socket_id(lcore_id)]++;
1006		}
1007
1008		/*
1009		 * Automatically spread requested memory amongst detected sockets according
1010		 * to number of cores from cpu mask present on each socket
1011		 */
1012		total_size = internal_config.memory;
1013		for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
1014
1015			/* Set memory amount per socket */
1016			default_size = (internal_config.memory * cpu_per_socket[socket])
1017			                / rte_lcore_count();
1018
1019			/* Limit to maximum available memory on socket */
1020			default_size = RTE_MIN(default_size, get_socket_mem_size(socket));
1021
1022			/* Update sizes */
1023			memory[socket] = default_size;
1024			total_size -= default_size;
1025		}
1026
1027		/*
1028		 * If some memory is remaining, try to allocate it by getting all
1029		 * available memory from sockets, one after the other
1030		 */
1031		for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_size != 0; socket++) {
1032			/* take whatever is available */
1033			default_size = RTE_MIN(get_socket_mem_size(socket) - memory[socket],
1034			                       total_size);
1035
1036			/* Update sizes */
1037			memory[socket] += default_size;
1038			total_size -= default_size;
1039		}
1040	}
1041
1042	for (socket = 0; socket < RTE_MAX_NUMA_NODES && total_mem != 0; socket++) {
1043		/* skips if the memory on specific socket wasn't requested */
1044		for (i = 0; i < num_hp_info && memory[socket] != 0; i++){
1045			hp_used[i].hugedir = hp_info[i].hugedir;
1046			hp_used[i].num_pages[socket] = RTE_MIN(
1047					memory[socket] / hp_info[i].hugepage_sz,
1048					hp_info[i].num_pages[socket]);
1049
1050			cur_mem = hp_used[i].num_pages[socket] *
1051					hp_used[i].hugepage_sz;
1052
1053			memory[socket] -= cur_mem;
1054			total_mem -= cur_mem;
1055
1056			total_num_pages += hp_used[i].num_pages[socket];
1057
1058			/* check if we have met all memory requests */
1059			if (memory[socket] == 0)
1060				break;
1061
1062			/* check if we have any more pages left at this size, if so
1063			 * move on to next size */
1064			if (hp_used[i].num_pages[socket] == hp_info[i].num_pages[socket])
1065				continue;
1066			/* At this point we know that there are more pages available that are
1067			 * bigger than the memory we want, so lets see if we can get enough
1068			 * from other page sizes.
1069			 */
1070			remaining_mem = 0;
1071			for (j = i+1; j < num_hp_info; j++)
1072				remaining_mem += hp_info[j].hugepage_sz *
1073				hp_info[j].num_pages[socket];
1074
1075			/* is there enough other memory, if not allocate another page and quit */
1076			if (remaining_mem < memory[socket]){
1077				cur_mem = RTE_MIN(memory[socket],
1078						hp_info[i].hugepage_sz);
1079				memory[socket] -= cur_mem;
1080				total_mem -= cur_mem;
1081				hp_used[i].num_pages[socket]++;
1082				total_num_pages++;
1083				break; /* we are done with this socket*/
1084			}
1085		}
1086		/* if we didn't satisfy all memory requirements per socket */
1087		if (memory[socket] > 0) {
1088			/* to prevent icc errors */
1089			requested = (unsigned) (internal_config.socket_mem[socket] /
1090					0x100000);
1091			available = requested -
1092					((unsigned) (memory[socket] / 0x100000));
1093			RTE_LOG(ERR, EAL, "Not enough memory available on socket %u! "
1094					"Requested: %uMB, available: %uMB\n", socket,
1095					requested, available);
1096			return -1;
1097		}
1098	}
1099
1100	/* if we didn't satisfy total memory requirements */
1101	if (total_mem > 0) {
1102		requested = (unsigned) (internal_config.memory / 0x100000);
1103		available = requested - (unsigned) (total_mem / 0x100000);
1104		RTE_LOG(ERR, EAL, "Not enough memory available! Requested: %uMB,"
1105				" available: %uMB\n", requested, available);
1106		return -1;
1107	}
1108	return total_num_pages;
1109}
1110
1111static inline size_t
1112eal_get_hugepage_mem_size(void)
1113{
1114	uint64_t size = 0;
1115	unsigned i, j;
1116
1117	for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
1118		struct hugepage_info *hpi = &internal_config.hugepage_info[i];
1119		if (hpi->hugedir != NULL) {
1120			for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1121				size += hpi->hugepage_sz * hpi->num_pages[j];
1122			}
1123		}
1124	}
1125
1126	return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
1127}
1128
1129static struct sigaction huge_action_old;
1130static int huge_need_recover;
1131
1132static void
1133huge_register_sigbus(void)
1134{
1135	sigset_t mask;
1136	struct sigaction action;
1137
1138	sigemptyset(&mask);
1139	sigaddset(&mask, SIGBUS);
1140	action.sa_flags = 0;
1141	action.sa_mask = mask;
1142	action.sa_handler = huge_sigbus_handler;
1143
1144	huge_need_recover = !sigaction(SIGBUS, &action, &huge_action_old);
1145}
1146
1147static void
1148huge_recover_sigbus(void)
1149{
1150	if (huge_need_recover) {
1151		sigaction(SIGBUS, &huge_action_old, NULL);
1152		huge_need_recover = 0;
1153	}
1154}
1155
1156/*
1157 * Prepare physical memory mapping: fill configuration structure with
1158 * these infos, return 0 on success.
1159 *  1. map N huge pages in separate files in hugetlbfs
1160 *  2. find associated physical addr
1161 *  3. find associated NUMA socket ID
1162 *  4. sort all huge pages by physical address
1163 *  5. remap these N huge pages in the correct order
1164 *  6. unmap the first mapping
1165 *  7. fill memsegs in configuration with contiguous zones
1166 */
1167int
1168rte_eal_hugepage_init(void)
1169{
1170	struct rte_mem_config *mcfg;
1171	struct hugepage_file *hugepage = NULL, *tmp_hp = NULL;
1172	struct hugepage_info used_hp[MAX_HUGEPAGE_SIZES];
1173
1174	uint64_t memory[RTE_MAX_NUMA_NODES];
1175
1176	unsigned hp_offset;
1177	int i, j, new_memseg;
1178	int nr_hugefiles, nr_hugepages = 0;
1179	void *addr;
1180#ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1181	int new_pages_count[MAX_HUGEPAGE_SIZES];
1182#endif
1183
1184	test_proc_pagemap_readable();
1185
1186	memset(used_hp, 0, sizeof(used_hp));
1187
1188	/* get pointer to global configuration */
1189	mcfg = rte_eal_get_configuration()->mem_config;
1190
1191	/* hugetlbfs can be disabled */
1192	if (internal_config.no_hugetlbfs) {
1193		addr = mmap(NULL, internal_config.memory, PROT_READ | PROT_WRITE,
1194				MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
1195		if (addr == MAP_FAILED) {
1196			RTE_LOG(ERR, EAL, "%s: mmap() failed: %s\n", __func__,
1197					strerror(errno));
1198			return -1;
1199		}
1200		mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
1201		mcfg->memseg[0].addr = addr;
1202		mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
1203		mcfg->memseg[0].len = internal_config.memory;
1204		mcfg->memseg[0].socket_id = 0;
1205		return 0;
1206	}
1207
1208/* check if app runs on Xen Dom0 */
1209	if (internal_config.xen_dom0_support) {
1210#ifdef RTE_LIBRTE_XEN_DOM0
1211		/* use dom0_mm kernel driver to init memory */
1212		if (rte_xen_dom0_memory_init() < 0)
1213			return -1;
1214		else
1215			return 0;
1216#endif
1217	}
1218
1219	/* calculate total number of hugepages available. at this point we haven't
1220	 * yet started sorting them so they all are on socket 0 */
1221	for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1222		/* meanwhile, also initialize used_hp hugepage sizes in used_hp */
1223		used_hp[i].hugepage_sz = internal_config.hugepage_info[i].hugepage_sz;
1224
1225		nr_hugepages += internal_config.hugepage_info[i].num_pages[0];
1226	}
1227
1228	/*
1229	 * allocate a memory area for hugepage table.
1230	 * this isn't shared memory yet. due to the fact that we need some
1231	 * processing done on these pages, shared memory will be created
1232	 * at a later stage.
1233	 */
1234	tmp_hp = malloc(nr_hugepages * sizeof(struct hugepage_file));
1235	if (tmp_hp == NULL)
1236		goto fail;
1237
1238	memset(tmp_hp, 0, nr_hugepages * sizeof(struct hugepage_file));
1239
1240	hp_offset = 0; /* where we start the current page size entries */
1241
1242	huge_register_sigbus();
1243
1244	/* map all hugepages and sort them */
1245	for (i = 0; i < (int)internal_config.num_hugepage_sizes; i ++){
1246		unsigned pages_old, pages_new;
1247		struct hugepage_info *hpi;
1248
1249		/*
1250		 * we don't yet mark hugepages as used at this stage, so
1251		 * we just map all hugepages available to the system
1252		 * all hugepages are still located on socket 0
1253		 */
1254		hpi = &internal_config.hugepage_info[i];
1255
1256		if (hpi->num_pages[0] == 0)
1257			continue;
1258
1259		/* map all hugepages available */
1260		pages_old = hpi->num_pages[0];
1261		pages_new = map_all_hugepages(&tmp_hp[hp_offset], hpi, 1);
1262		if (pages_new < pages_old) {
1263#ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1264			RTE_LOG(ERR, EAL,
1265				"%d not %d hugepages of size %u MB allocated\n",
1266				pages_new, pages_old,
1267				(unsigned)(hpi->hugepage_sz / 0x100000));
1268			goto fail;
1269#else
1270			RTE_LOG(DEBUG, EAL,
1271				"%d not %d hugepages of size %u MB allocated\n",
1272				pages_new, pages_old,
1273				(unsigned)(hpi->hugepage_sz / 0x100000));
1274
1275			int pages = pages_old - pages_new;
1276
1277			nr_hugepages -= pages;
1278			hpi->num_pages[0] = pages_new;
1279			if (pages_new == 0)
1280				continue;
1281#endif
1282		}
1283
1284		/* find physical addresses and sockets for each hugepage */
1285		if (find_physaddrs(&tmp_hp[hp_offset], hpi) < 0){
1286			RTE_LOG(DEBUG, EAL, "Failed to find phys addr for %u MB pages\n",
1287					(unsigned)(hpi->hugepage_sz / 0x100000));
1288			goto fail;
1289		}
1290
1291		if (find_numasocket(&tmp_hp[hp_offset], hpi) < 0){
1292			RTE_LOG(DEBUG, EAL, "Failed to find NUMA socket for %u MB pages\n",
1293					(unsigned)(hpi->hugepage_sz / 0x100000));
1294			goto fail;
1295		}
1296
1297		qsort(&tmp_hp[hp_offset], hpi->num_pages[0],
1298		      sizeof(struct hugepage_file), cmp_physaddr);
1299
1300#ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1301		/* remap all hugepages into single file segments */
1302		new_pages_count[i] = remap_all_hugepages(&tmp_hp[hp_offset], hpi);
1303		if (new_pages_count[i] < 0){
1304			RTE_LOG(DEBUG, EAL, "Failed to remap %u MB pages\n",
1305					(unsigned)(hpi->hugepage_sz / 0x100000));
1306			goto fail;
1307		}
1308
1309		/* we have processed a num of hugepages of this size, so inc offset */
1310		hp_offset += new_pages_count[i];
1311#else
1312		/* remap all hugepages */
1313		if (map_all_hugepages(&tmp_hp[hp_offset], hpi, 0) !=
1314		    hpi->num_pages[0]) {
1315			RTE_LOG(ERR, EAL, "Failed to remap %u MB pages\n",
1316					(unsigned)(hpi->hugepage_sz / 0x100000));
1317			goto fail;
1318		}
1319
1320		/* unmap original mappings */
1321		if (unmap_all_hugepages_orig(&tmp_hp[hp_offset], hpi) < 0)
1322			goto fail;
1323
1324		/* we have processed a num of hugepages of this size, so inc offset */
1325		hp_offset += hpi->num_pages[0];
1326#endif
1327	}
1328
1329	huge_recover_sigbus();
1330
1331	if (internal_config.memory == 0 && internal_config.force_sockets == 0)
1332		internal_config.memory = eal_get_hugepage_mem_size();
1333
1334#ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1335	nr_hugefiles = 0;
1336	for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1337		nr_hugefiles += new_pages_count[i];
1338	}
1339#else
1340	nr_hugefiles = nr_hugepages;
1341#endif
1342
1343
1344	/* clean out the numbers of pages */
1345	for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++)
1346		for (j = 0; j < RTE_MAX_NUMA_NODES; j++)
1347			internal_config.hugepage_info[i].num_pages[j] = 0;
1348
1349	/* get hugepages for each socket */
1350	for (i = 0; i < nr_hugefiles; i++) {
1351		int socket = tmp_hp[i].socket_id;
1352
1353		/* find a hugepage info with right size and increment num_pages */
1354		const int nb_hpsizes = RTE_MIN(MAX_HUGEPAGE_SIZES,
1355				(int)internal_config.num_hugepage_sizes);
1356		for (j = 0; j < nb_hpsizes; j++) {
1357			if (tmp_hp[i].size ==
1358					internal_config.hugepage_info[j].hugepage_sz) {
1359#ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1360					internal_config.hugepage_info[j].num_pages[socket] +=
1361						tmp_hp[i].repeated;
1362#else
1363				internal_config.hugepage_info[j].num_pages[socket]++;
1364#endif
1365			}
1366		}
1367	}
1368
1369	/* make a copy of socket_mem, needed for number of pages calculation */
1370	for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
1371		memory[i] = internal_config.socket_mem[i];
1372
1373	/* calculate final number of pages */
1374	nr_hugepages = calc_num_pages_per_socket(memory,
1375			internal_config.hugepage_info, used_hp,
1376			internal_config.num_hugepage_sizes);
1377
1378	/* error if not enough memory available */
1379	if (nr_hugepages < 0)
1380		goto fail;
1381
1382	/* reporting in! */
1383	for (i = 0; i < (int) internal_config.num_hugepage_sizes; i++) {
1384		for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
1385			if (used_hp[i].num_pages[j] > 0) {
1386				RTE_LOG(DEBUG, EAL,
1387					"Requesting %u pages of size %uMB"
1388					" from socket %i\n",
1389					used_hp[i].num_pages[j],
1390					(unsigned)
1391					(used_hp[i].hugepage_sz / 0x100000),
1392					j);
1393			}
1394		}
1395	}
1396
1397	/* create shared memory */
1398	hugepage = create_shared_memory(eal_hugepage_info_path(),
1399			nr_hugefiles * sizeof(struct hugepage_file));
1400
1401	if (hugepage == NULL) {
1402		RTE_LOG(ERR, EAL, "Failed to create shared memory!\n");
1403		goto fail;
1404	}
1405	memset(hugepage, 0, nr_hugefiles * sizeof(struct hugepage_file));
1406
1407	/*
1408	 * unmap pages that we won't need (looks at used_hp).
1409	 * also, sets final_va to NULL on pages that were unmapped.
1410	 */
1411	if (unmap_unneeded_hugepages(tmp_hp, used_hp,
1412			internal_config.num_hugepage_sizes) < 0) {
1413		RTE_LOG(ERR, EAL, "Unmapping and locking hugepages failed!\n");
1414		goto fail;
1415	}
1416
1417	/*
1418	 * copy stuff from malloc'd hugepage* to the actual shared memory.
1419	 * this procedure only copies those hugepages that have final_va
1420	 * not NULL. has overflow protection.
1421	 */
1422	if (copy_hugepages_to_shared_mem(hugepage, nr_hugefiles,
1423			tmp_hp, nr_hugefiles) < 0) {
1424		RTE_LOG(ERR, EAL, "Copying tables to shared memory failed!\n");
1425		goto fail;
1426	}
1427
1428	/* free the hugepage backing files */
1429	if (internal_config.hugepage_unlink &&
1430		unlink_hugepage_files(tmp_hp, internal_config.num_hugepage_sizes) < 0) {
1431		RTE_LOG(ERR, EAL, "Unlinking hugepage files failed!\n");
1432		goto fail;
1433	}
1434
1435	/* free the temporary hugepage table */
1436	free(tmp_hp);
1437	tmp_hp = NULL;
1438
1439	/* find earliest free memseg - this is needed because in case of IVSHMEM,
1440	 * segments might have already been initialized */
1441	for (j = 0; j < RTE_MAX_MEMSEG; j++)
1442		if (mcfg->memseg[j].addr == NULL) {
1443			/* move to previous segment and exit loop */
1444			j--;
1445			break;
1446		}
1447
1448	for (i = 0; i < nr_hugefiles; i++) {
1449		new_memseg = 0;
1450
1451		/* if this is a new section, create a new memseg */
1452		if (i == 0)
1453			new_memseg = 1;
1454		else if (hugepage[i].socket_id != hugepage[i-1].socket_id)
1455			new_memseg = 1;
1456		else if (hugepage[i].size != hugepage[i-1].size)
1457			new_memseg = 1;
1458
1459#ifdef RTE_ARCH_PPC_64
1460		/* On PPC64 architecture, the mmap always start from higher
1461		 * virtual address to lower address. Here, both the physical
1462		 * address and virtual address are in descending order */
1463		else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
1464		    hugepage[i].size)
1465			new_memseg = 1;
1466		else if (((unsigned long)hugepage[i-1].final_va -
1467		    (unsigned long)hugepage[i].final_va) != hugepage[i].size)
1468			new_memseg = 1;
1469#else
1470		else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
1471		    hugepage[i].size)
1472			new_memseg = 1;
1473		else if (((unsigned long)hugepage[i].final_va -
1474		    (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
1475			new_memseg = 1;
1476#endif
1477
1478		if (new_memseg) {
1479			j += 1;
1480			if (j == RTE_MAX_MEMSEG)
1481				break;
1482
1483			mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1484			mcfg->memseg[j].addr = hugepage[i].final_va;
1485#ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1486			mcfg->memseg[j].len = hugepage[i].size * hugepage[i].repeated;
1487#else
1488			mcfg->memseg[j].len = hugepage[i].size;
1489#endif
1490			mcfg->memseg[j].socket_id = hugepage[i].socket_id;
1491			mcfg->memseg[j].hugepage_sz = hugepage[i].size;
1492		}
1493		/* continuation of previous memseg */
1494		else {
1495#ifdef RTE_ARCH_PPC_64
1496		/* Use the phy and virt address of the last page as segment
1497		 * address for IBM Power architecture */
1498			mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
1499			mcfg->memseg[j].addr = hugepage[i].final_va;
1500#endif
1501			mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
1502		}
1503		hugepage[i].memseg_id = j;
1504	}
1505
1506	if (i < nr_hugefiles) {
1507		RTE_LOG(ERR, EAL, "Can only reserve %d pages "
1508			"from %d requested\n"
1509			"Current %s=%d is not enough\n"
1510			"Please either increase it or request less amount "
1511			"of memory.\n",
1512			i, nr_hugefiles, RTE_STR(CONFIG_RTE_MAX_MEMSEG),
1513			RTE_MAX_MEMSEG);
1514		goto fail;
1515	}
1516
1517	munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1518
1519	return 0;
1520
1521fail:
1522	huge_recover_sigbus();
1523	free(tmp_hp);
1524	if (hugepage != NULL)
1525		munmap(hugepage, nr_hugefiles * sizeof(struct hugepage_file));
1526
1527	return -1;
1528}
1529
1530/*
1531 * uses fstat to report the size of a file on disk
1532 */
1533static off_t
1534getFileSize(int fd)
1535{
1536	struct stat st;
1537	if (fstat(fd, &st) < 0)
1538		return 0;
1539	return st.st_size;
1540}
1541
1542/*
1543 * This creates the memory mappings in the secondary process to match that of
1544 * the server process. It goes through each memory segment in the DPDK runtime
1545 * configuration and finds the hugepages which form that segment, mapping them
1546 * in order to form a contiguous block in the virtual memory space
1547 */
1548int
1549rte_eal_hugepage_attach(void)
1550{
1551	const struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
1552	struct hugepage_file *hp = NULL;
1553	unsigned num_hp = 0;
1554	unsigned i, s = 0; /* s used to track the segment number */
1555	unsigned max_seg = RTE_MAX_MEMSEG;
1556	off_t size = 0;
1557	int fd, fd_zero = -1, fd_hugepage = -1;
1558
1559	if (aslr_enabled() > 0) {
1560		RTE_LOG(WARNING, EAL, "WARNING: Address Space Layout Randomization "
1561				"(ASLR) is enabled in the kernel.\n");
1562		RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory "
1563				"into secondary processes\n");
1564	}
1565
1566	test_proc_pagemap_readable();
1567
1568	if (internal_config.xen_dom0_support) {
1569#ifdef RTE_LIBRTE_XEN_DOM0
1570		if (rte_xen_dom0_memory_attach() < 0) {
1571			RTE_LOG(ERR, EAL, "Failed to attach memory segments of primary "
1572					"process\n");
1573			return -1;
1574		}
1575		return 0;
1576#endif
1577	}
1578
1579	fd_zero = open("/dev/zero", O_RDONLY);
1580	if (fd_zero < 0) {
1581		RTE_LOG(ERR, EAL, "Could not open /dev/zero\n");
1582		goto error;
1583	}
1584	fd_hugepage = open(eal_hugepage_info_path(), O_RDONLY);
1585	if (fd_hugepage < 0) {
1586		RTE_LOG(ERR, EAL, "Could not open %s\n", eal_hugepage_info_path());
1587		goto error;
1588	}
1589
1590	/* map all segments into memory to make sure we get the addrs */
1591	for (s = 0; s < RTE_MAX_MEMSEG; ++s) {
1592		void *base_addr;
1593
1594		/*
1595		 * the first memory segment with len==0 is the one that
1596		 * follows the last valid segment.
1597		 */
1598		if (mcfg->memseg[s].len == 0)
1599			break;
1600
1601#ifdef RTE_LIBRTE_IVSHMEM
1602		/*
1603		 * if segment has ioremap address set, it's an IVSHMEM segment and
1604		 * doesn't need mapping as it was already mapped earlier
1605		 */
1606		if (mcfg->memseg[s].ioremap_addr != 0)
1607			continue;
1608#endif
1609
1610		/*
1611		 * fdzero is mmapped to get a contiguous block of virtual
1612		 * addresses of the appropriate memseg size.
1613		 * use mmap to get identical addresses as the primary process.
1614		 */
1615		base_addr = mmap(mcfg->memseg[s].addr, mcfg->memseg[s].len,
1616				 PROT_READ, MAP_PRIVATE, fd_zero, 0);
1617		if (base_addr == MAP_FAILED ||
1618		    base_addr != mcfg->memseg[s].addr) {
1619			RTE_LOG(ERR, EAL, "Could not mmap %llu bytes "
1620				"in /dev/zero to requested address [%p]: '%s'\n",
1621				(unsigned long long)mcfg->memseg[s].len,
1622				mcfg->memseg[s].addr, strerror(errno));
1623			max_seg = s;
1624			if (base_addr != MAP_FAILED)
1625				munmap(base_addr, mcfg->memseg[s].len);
1626			if (aslr_enabled() > 0) {
1627				RTE_LOG(ERR, EAL, "It is recommended to "
1628					"disable ASLR in the kernel "
1629					"and retry running both primary "
1630					"and secondary processes\n");
1631			}
1632			goto error;
1633		}
1634	}
1635
1636	size = getFileSize(fd_hugepage);
1637	hp = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd_hugepage, 0);
1638	if (hp == MAP_FAILED) {
1639		RTE_LOG(ERR, EAL, "Could not mmap %s\n", eal_hugepage_info_path());
1640		goto error;
1641	}
1642
1643	num_hp = size / sizeof(struct hugepage_file);
1644	RTE_LOG(DEBUG, EAL, "Analysing %u files\n", num_hp);
1645
1646	s = 0;
1647	while (s < RTE_MAX_MEMSEG && mcfg->memseg[s].len > 0){
1648		void *addr, *base_addr;
1649		uintptr_t offset = 0;
1650		size_t mapping_size;
1651#ifdef RTE_LIBRTE_IVSHMEM
1652		/*
1653		 * if segment has ioremap address set, it's an IVSHMEM segment and
1654		 * doesn't need mapping as it was already mapped earlier
1655		 */
1656		if (mcfg->memseg[s].ioremap_addr != 0) {
1657			s++;
1658			continue;
1659		}
1660#endif
1661		/*
1662		 * free previously mapped memory so we can map the
1663		 * hugepages into the space
1664		 */
1665		base_addr = mcfg->memseg[s].addr;
1666		munmap(base_addr, mcfg->memseg[s].len);
1667
1668		/* find the hugepages for this segment and map them
1669		 * we don't need to worry about order, as the server sorted the
1670		 * entries before it did the second mmap of them */
1671		for (i = 0; i < num_hp && offset < mcfg->memseg[s].len; i++){
1672			if (hp[i].memseg_id == (int)s){
1673				fd = open(hp[i].filepath, O_RDWR);
1674				if (fd < 0) {
1675					RTE_LOG(ERR, EAL, "Could not open %s\n",
1676						hp[i].filepath);
1677					goto error;
1678				}
1679#ifdef RTE_EAL_SINGLE_FILE_SEGMENTS
1680				mapping_size = hp[i].size * hp[i].repeated;
1681#else
1682				mapping_size = hp[i].size;
1683#endif
1684				addr = mmap(RTE_PTR_ADD(base_addr, offset),
1685						mapping_size, PROT_READ | PROT_WRITE,
1686						MAP_SHARED, fd, 0);
1687				close(fd); /* close file both on success and on failure */
1688				if (addr == MAP_FAILED ||
1689						addr != RTE_PTR_ADD(base_addr, offset)) {
1690					RTE_LOG(ERR, EAL, "Could not mmap %s\n",
1691						hp[i].filepath);
1692					goto error;
1693				}
1694				offset+=mapping_size;
1695			}
1696		}
1697		RTE_LOG(DEBUG, EAL, "Mapped segment %u of size 0x%llx\n", s,
1698				(unsigned long long)mcfg->memseg[s].len);
1699		s++;
1700	}
1701	/* unmap the hugepage config file, since we are done using it */
1702	munmap(hp, size);
1703	close(fd_zero);
1704	close(fd_hugepage);
1705	return 0;
1706
1707error:
1708	for (i = 0; i < max_seg && mcfg->memseg[i].len > 0; i++)
1709		munmap(mcfg->memseg[i].addr, mcfg->memseg[i].len);
1710	if (hp != NULL && hp != MAP_FAILED)
1711		munmap(hp, size);
1712	if (fd_zero >= 0)
1713		close(fd_zero);
1714	if (fd_hugepage >= 0)
1715		close(fd_hugepage);
1716	return -1;
1717}
1718