xref: /php-src/ext/opcache/shared_alloc_mmap.c (revision 6fb8b9d7)
1 /*
2    +----------------------------------------------------------------------+
3    | Zend OPcache                                                         |
4    +----------------------------------------------------------------------+
5    | Copyright (c) The PHP Group                                          |
6    +----------------------------------------------------------------------+
7    | This source file is subject to version 3.01 of the PHP license,      |
8    | that is bundled with this package in the file LICENSE, and is        |
9    | available through the world-wide-web at the following url:           |
10    | https://www.php.net/license/3_01.txt                                 |
11    | If you did not receive a copy of the PHP license and are unable to   |
12    | obtain it through the world-wide-web, please send a note to          |
13    | license@php.net so we can mail you a copy immediately.               |
14    +----------------------------------------------------------------------+
15    | Authors: Andi Gutmans <andi@php.net>                                 |
16    |          Zeev Suraski <zeev@php.net>                                 |
17    |          Stanislav Malyshev <stas@zend.com>                          |
18    |          Dmitry Stogov <dmitry@php.net>                              |
19    +----------------------------------------------------------------------+
20 */
21 
22 #include "zend_shared_alloc.h"
23 
24 #ifdef USE_MMAP
25 
26 #include <sys/types.h>
27 #include <sys/stat.h>
28 #include <stdio.h>
29 #include <stdlib.h>
30 #include <sys/mman.h>
31 
32 #ifdef __APPLE__
33 #include <mach/vm_statistics.h>
34 #endif
35 
36 #include "zend_execute.h"
37 #ifdef HAVE_PROCCTL
38 #include <sys/procctl.h>
39 #endif
40 
41 #if defined(MAP_ANON) && !defined(MAP_ANONYMOUS)
42 # define MAP_ANONYMOUS MAP_ANON
43 #endif
44 #if defined(MAP_ALIGNED_SUPER)
45 # include <sys/types.h>
46 # include <sys/sysctl.h>
47 # include <sys/user.h>
48 # define MAP_HUGETLB MAP_ALIGNED_SUPER
49 #endif
50 
51 #if (defined(__linux__) || defined(__FreeBSD__)) && (defined(__x86_64__) || defined (__aarch64__)) && !defined(__SANITIZE_ADDRESS__)
find_prefered_mmap_base(size_t requested_size)52 static void *find_prefered_mmap_base(size_t requested_size)
53 {
54 	size_t huge_page_size = 2 * 1024 * 1024;
55 	uintptr_t last_free_addr = huge_page_size;
56 	uintptr_t last_candidate = (uintptr_t)MAP_FAILED;
57 	uintptr_t start, end, text_start = 0;
58 #if defined(__linux__)
59 	FILE *f;
60 	char buffer[MAXPATHLEN];
61 
62 	f = fopen("/proc/self/maps", "r");
63 	if (!f) {
64 		return MAP_FAILED;
65 	}
66 
67 	while (fgets(buffer, MAXPATHLEN, f) && sscanf(buffer, "%lx-%lx", &start, &end) == 2) {
68 		/* Don't place the segment directly before or after the heap segment. Due to an selinux bug,
69 		 * a segment directly preceding or following the heap is interpreted as heap memory, which
70 		 * will result in an execheap violation for the JIT.
71 		 * See https://bugzilla.kernel.org/show_bug.cgi?id=218258. */
72 		bool heap_segment = strstr(buffer, "[heap]") != NULL;
73 		if (heap_segment) {
74 			uintptr_t start_base = start & ~(huge_page_size - 1);
75 			if (last_free_addr + requested_size >= start_base) {
76 				last_free_addr = ZEND_MM_ALIGNED_SIZE_EX(end + huge_page_size, huge_page_size);
77 				continue;
78 			}
79 		}
80 		if ((uintptr_t)execute_ex >= start) {
81 			/* the current segment lays before PHP .text segment or PHP .text segment itself */
82 			/*Search for candidates at the end of the free segment near the .text segment
83 			  to prevent candidates from being missed due to large hole*/
84 			if (last_free_addr + requested_size <= start) {
85 				last_candidate = ZEND_MM_ALIGNED_SIZE_EX(start - requested_size, huge_page_size);
86 				if (last_candidate + requested_size > start) {
87 					last_candidate -= huge_page_size;
88 				}
89 			}
90 			if ((uintptr_t)execute_ex < end) {
91 				/* the current segment is PHP .text segment itself */
92 				if (last_candidate != (uintptr_t)MAP_FAILED) {
93 					if (end - last_candidate < UINT32_MAX) {
94 						/* we have found a big enough hole before the text segment */
95 						break;
96 					}
97 					last_candidate = (uintptr_t)MAP_FAILED;
98 				}
99 				text_start = start;
100 			}
101 		} else {
102 			/* the current segment lays after PHP .text segment */
103 			if (last_free_addr + requested_size - text_start > UINT32_MAX) {
104 				/* the current segment and the following segments lay too far from PHP .text segment */
105 				break;
106 			}
107 			if (last_free_addr + requested_size <= start) {
108 				last_candidate = last_free_addr;
109 				break;
110 			}
111 		}
112 		last_free_addr = ZEND_MM_ALIGNED_SIZE_EX(end, huge_page_size);
113 		if (heap_segment) {
114 			last_free_addr += huge_page_size;
115 		}
116 	}
117 	fclose(f);
118 #elif defined(__FreeBSD__)
119 	size_t s = 0;
120 	int mib[4] = {CTL_KERN, KERN_PROC, KERN_PROC_VMMAP, getpid()};
121 	if (sysctl(mib, 4, NULL, &s, NULL, 0) == 0) {
122 		s = s * 4 / 3;
123 		void *addr = mmap(NULL, s, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANON, -1, 0);
124 		if (addr != MAP_FAILED) {
125 			if (sysctl(mib, 4, addr, &s, NULL, 0) == 0) {
126 				start = (uintptr_t)addr;
127 				end = start + s;
128 				while (start < end) {
129 					struct kinfo_vmentry *entry = (struct kinfo_vmentry *)start;
130 					size_t sz = entry->kve_structsize;
131 					if (sz == 0) {
132 						break;
133 					}
134 					uintptr_t e_start = entry->kve_start;
135 					uintptr_t e_end = entry->kve_end;
136 					if ((uintptr_t)execute_ex >= e_start) {
137 						/* the current segment lays before PHP .text segment or PHP .text segment itself */
138 						if (last_free_addr + requested_size <= e_start) {
139 							last_candidate = ZEND_MM_ALIGNED_SIZE_EX(e_start - requested_size, huge_page_size);
140 							if (last_candidate + requested_size > e_start) {
141 								last_candidate -= huge_page_size;
142 							}
143 						}
144 						if ((uintptr_t)execute_ex < e_end) {
145 							/* the current segment is PHP .text segment itself */
146 							if (last_candidate != (uintptr_t)MAP_FAILED) {
147 								if (e_end - last_candidate < UINT32_MAX) {
148 									/* we have found a big enough hole before the text segment */
149 									break;
150 								}
151 								last_candidate = (uintptr_t)MAP_FAILED;
152 							}
153 							text_start = e_start;
154 						}
155 					} else {
156 						/* the current segment lays after PHP .text segment */
157 						if (last_free_addr + requested_size - text_start > UINT32_MAX) {
158 							/* the current segment and the following segments lay too far from PHP .text segment */
159 							break;
160 						}
161 						if (last_free_addr + requested_size <= e_start) {
162 							last_candidate = last_free_addr;
163 							break;
164 						}
165 					}
166 					last_free_addr = ZEND_MM_ALIGNED_SIZE_EX(e_end, huge_page_size);
167 					start += sz;
168 				}
169 			}
170 			munmap(addr, s);
171 		}
172 	}
173 #endif
174 
175 	return (void*)last_candidate;
176 }
177 #endif
178 
create_segments(size_t requested_size,zend_shared_segment *** shared_segments_p,int * shared_segments_count,const char ** error_in)179 static int create_segments(size_t requested_size, zend_shared_segment ***shared_segments_p, int *shared_segments_count, const char **error_in)
180 {
181 	zend_shared_segment *shared_segment;
182 	int flags = PROT_READ | PROT_WRITE, fd = -1;
183 	void *p;
184 #if defined(HAVE_PROCCTL) && defined(PROC_WXMAP_CTL)
185 	int enable_wxmap = PROC_WX_MAPPINGS_PERMIT;
186 	if (procctl(P_PID, getpid(), PROC_WXMAP_CTL, &enable_wxmap) == -1) {
187 		return ALLOC_FAILURE;
188 	}
189 #endif
190 #ifdef PROT_MPROTECT
191 	flags |= PROT_MPROTECT(PROT_EXEC);
192 #endif
193 #ifdef VM_MAKE_TAG
194 	/* allows tracking segments via tools such as vmmap */
195 	fd = VM_MAKE_TAG(251U);
196 #endif
197 #ifdef PROT_MAX
198 	flags |= PROT_MAX(PROT_READ | PROT_WRITE | PROT_EXEC);
199 #endif
200 #if (defined(__linux__) || defined(__FreeBSD__)) && (defined(__x86_64__) || defined (__aarch64__)) && !defined(__SANITIZE_ADDRESS__)
201 	void *hint = find_prefered_mmap_base(requested_size);
202 	if (hint != MAP_FAILED) {
203 # ifdef MAP_HUGETLB
204 		size_t huge_page_size = 2 * 1024 * 1024;
205 		if (requested_size >= huge_page_size && requested_size % huge_page_size == 0) {
206 			p = mmap(hint, requested_size, flags, MAP_SHARED|MAP_ANONYMOUS|MAP_HUGETLB|MAP_FIXED, -1, 0);
207 			if (p != MAP_FAILED) {
208 				goto success;
209 			}
210 		}
211 #endif
212 		p = mmap(hint, requested_size, flags, MAP_SHARED|MAP_ANONYMOUS|MAP_FIXED, -1, 0);
213 		if (p != MAP_FAILED) {
214 			goto success;
215 		}
216 	}
217 #endif
218 #ifdef MAP_HUGETLB
219 	size_t huge_page_size = 2 * 1024 * 1024;
220 
221 	/* Try to allocate huge pages first to reduce dTLB misses.
222 	 * OSes has to be configured properly
223 	 * on Linux
224 	 * (e.g. https://wiki.debian.org/Hugepages#Enabling_HugeTlbPage)
225 	 * You may verify huge page usage with the following command:
226 	 * `grep "Huge" /proc/meminfo`
227 	 * on FreeBSD
228 	 * sysctl vm.pmap.pg_ps_enabled entry
229 	 * (boot time config only, but enabled by default on most arches).
230 	 */
231 	if (requested_size >= huge_page_size && requested_size % huge_page_size == 0) {
232 # if defined(__x86_64__) && defined(MAP_32BIT)
233 		/* to got HUGE PAGES in low 32-bit address we have to reserve address
234 		   space and then remap it using MAP_HUGETLB */
235 
236 		p = mmap(NULL, requested_size, flags, MAP_SHARED|MAP_ANONYMOUS|MAP_32BIT, fd, 0);
237 		if (p != MAP_FAILED) {
238 			munmap(p, requested_size);
239 			p = (void*)(ZEND_MM_ALIGNED_SIZE_EX((ptrdiff_t)p, huge_page_size));
240 			p = mmap(p, requested_size, flags, MAP_SHARED|MAP_ANONYMOUS|MAP_32BIT|MAP_HUGETLB|MAP_FIXED, -1, 0);
241 			if (p != MAP_FAILED) {
242 				goto success;
243 			} else {
244 				p = mmap(NULL, requested_size, flags, MAP_SHARED|MAP_ANONYMOUS|MAP_32BIT, fd, 0);
245 				if (p != MAP_FAILED) {
246 					goto success;
247 				}
248 			}
249 		}
250 # endif
251 		p = mmap(0, requested_size, flags, MAP_SHARED|MAP_ANONYMOUS|MAP_HUGETLB, fd, 0);
252 		if (p != MAP_FAILED) {
253 			goto success;
254 		}
255 	}
256 #elif defined(PREFER_MAP_32BIT) && defined(__x86_64__) && defined(MAP_32BIT)
257 	p = mmap(NULL, requested_size, flags, MAP_SHARED|MAP_ANONYMOUS|MAP_32BIT, fd, 0);
258 	if (p != MAP_FAILED) {
259 		goto success;
260 	}
261 #endif
262 
263 	p = mmap(0, requested_size, flags, MAP_SHARED|MAP_ANONYMOUS, fd, 0);
264 	if (p == MAP_FAILED) {
265 		*error_in = "mmap";
266 		return ALLOC_FAILURE;
267 	}
268 
269 success: ZEND_ATTRIBUTE_UNUSED;
270 	*shared_segments_count = 1;
271 	*shared_segments_p = (zend_shared_segment **) calloc(1, sizeof(zend_shared_segment) + sizeof(void *));
272 	if (!*shared_segments_p) {
273 		munmap(p, requested_size);
274 		*error_in = "calloc";
275 		return ALLOC_FAILURE;
276 	}
277 	shared_segment = (zend_shared_segment *)((char *)(*shared_segments_p) + sizeof(void *));
278 	(*shared_segments_p)[0] = shared_segment;
279 
280 	shared_segment->p = p;
281 	shared_segment->pos = 0;
282 	shared_segment->size = requested_size;
283 
284 	return ALLOC_SUCCESS;
285 }
286 
detach_segment(zend_shared_segment * shared_segment)287 static int detach_segment(zend_shared_segment *shared_segment)
288 {
289 	munmap(shared_segment->p, shared_segment->size);
290 	return 0;
291 }
292 
segment_type_size(void)293 static size_t segment_type_size(void)
294 {
295 	return sizeof(zend_shared_segment);
296 }
297 
298 const zend_shared_memory_handlers zend_alloc_mmap_handlers = {
299 	create_segments,
300 	detach_segment,
301 	segment_type_size
302 };
303 
304 #endif /* USE_MMAP */
305