1 /* ----------------------------------------------------------------------------
2 Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
3 This is free software; you can redistribute it and/or modify it under the
4 terms of the MIT license. A copy of the license can be found in the file
5 "LICENSE" at the root of this distribution.
6 -----------------------------------------------------------------------------*/
7 #pragma once
8 #ifndef MIMALLOC_PRIM_H
9 #define MIMALLOC_PRIM_H
10
11
12 // --------------------------------------------------------------------------
13 // This file specifies the primitive portability API.
14 // Each OS/host needs to implement these primitives, see `src/prim`
15 // for implementations on Window, macOS, WASI, and Linux/Unix.
16 //
17 // note: on all primitive functions, we always have result parameters != NUL, and:
18 // addr != NULL and page aligned
19 // size > 0 and page aligned
20 // return value is an error code an int where 0 is success.
21 // --------------------------------------------------------------------------
22
23 // OS memory configuration
24 typedef struct mi_os_mem_config_s {
25 size_t page_size; // 4KiB
26 size_t large_page_size; // 2MiB
27 size_t alloc_granularity; // smallest allocation size (on Windows 64KiB)
28 bool has_overcommit; // can we reserve more memory than can be actually committed?
29 bool must_free_whole; // must allocated blocks be freed as a whole (false for mmap, true for VirtualAlloc)
30 bool has_virtual_reserve; // supports virtual address space reservation? (if true we can reserve virtual address space without using commit or physical memory)
31 } mi_os_mem_config_t;
32
33 // Initialize
34 void _mi_prim_mem_init( mi_os_mem_config_t* config );
35
36 // Free OS memory
37 int _mi_prim_free(void* addr, size_t size );
38
39 // Allocate OS memory. Return NULL on error.
40 // The `try_alignment` is just a hint and the returned pointer does not have to be aligned.
41 // If `commit` is false, the virtual memory range only needs to be reserved (with no access)
42 // which will later be committed explicitly using `_mi_prim_commit`.
43 // `is_zero` is set to true if the memory was zero initialized (as on most OS's)
44 // pre: !commit => !allow_large
45 // try_alignment >= _mi_os_page_size() and a power of 2
46 int _mi_prim_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, bool* is_zero, void** addr);
47
48 // Commit memory. Returns error code or 0 on success.
49 // For example, on Linux this would make the memory PROT_READ|PROT_WRITE.
50 // `is_zero` is set to true if the memory was zero initialized (e.g. on Windows)
51 int _mi_prim_commit(void* addr, size_t size, bool* is_zero);
52
53 // Decommit memory. Returns error code or 0 on success. The `needs_recommit` result is true
54 // if the memory would need to be re-committed. For example, on Windows this is always true,
55 // but on Linux we could use MADV_DONTNEED to decommit which does not need a recommit.
56 // pre: needs_recommit != NULL
57 int _mi_prim_decommit(void* addr, size_t size, bool* needs_recommit);
58
59 // Reset memory. The range keeps being accessible but the content might be reset.
60 // Returns error code or 0 on success.
61 int _mi_prim_reset(void* addr, size_t size);
62
63 // Protect memory. Returns error code or 0 on success.
64 int _mi_prim_protect(void* addr, size_t size, bool protect);
65
66 // Allocate huge (1GiB) pages possibly associated with a NUMA node.
67 // `is_zero` is set to true if the memory was zero initialized (as on most OS's)
68 // pre: size > 0 and a multiple of 1GiB.
69 // numa_node is either negative (don't care), or a numa node number.
70 int _mi_prim_alloc_huge_os_pages(void* hint_addr, size_t size, int numa_node, bool* is_zero, void** addr);
71
72 // Return the current NUMA node
73 size_t _mi_prim_numa_node(void);
74
75 // Return the number of logical NUMA nodes
76 size_t _mi_prim_numa_node_count(void);
77
78 // Clock ticks
79 mi_msecs_t _mi_prim_clock_now(void);
80
81 // Return process information (only for statistics)
82 typedef struct mi_process_info_s {
83 mi_msecs_t elapsed;
84 mi_msecs_t utime;
85 mi_msecs_t stime;
86 size_t current_rss;
87 size_t peak_rss;
88 size_t current_commit;
89 size_t peak_commit;
90 size_t page_faults;
91 } mi_process_info_t;
92
93 void _mi_prim_process_info(mi_process_info_t* pinfo);
94
95 // Default stderr output. (only for warnings etc. with verbose enabled)
96 // msg != NULL && _mi_strlen(msg) > 0
97 void _mi_prim_out_stderr( const char* msg );
98
99 // Get an environment variable. (only for options)
100 // name != NULL, result != NULL, result_size >= 64
101 bool _mi_prim_getenv(const char* name, char* result, size_t result_size);
102
103
104 // Fill a buffer with strong randomness; return `false` on error or if
105 // there is no strong randomization available.
106 bool _mi_prim_random_buf(void* buf, size_t buf_len);
107
108 // Called on the first thread start, and should ensure `_mi_thread_done` is called on thread termination.
109 void _mi_prim_thread_init_auto_done(void);
110
111 // Called on process exit and may take action to clean up resources associated with the thread auto done.
112 void _mi_prim_thread_done_auto_done(void);
113
114 // Called when the default heap for a thread changes
115 void _mi_prim_thread_associate_default_heap(mi_heap_t* heap);
116
117
118 //-------------------------------------------------------------------
119 // Thread id: `_mi_prim_thread_id()`
120 //
121 // Getting the thread id should be performant as it is called in the
122 // fast path of `_mi_free` and we specialize for various platforms as
123 // inlined definitions. Regular code should call `init.c:_mi_thread_id()`.
124 // We only require _mi_prim_thread_id() to return a unique id
125 // for each thread (unequal to zero).
126 //-------------------------------------------------------------------
127
128 // defined in `init.c`; do not use these directly
129 extern mi_decl_thread mi_heap_t* _mi_heap_default; // default heap to allocate from
130 extern bool _mi_process_is_initialized; // has mi_process_init been called?
131
132 static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept;
133
134 #ifdef MI_PRIM_THREAD_ID
135
_mi_prim_thread_id(void)136 static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
137 return MI_PRIM_THREAD_ID();
138 }
139
140 #elif defined(_WIN32)
141
142 #define WIN32_LEAN_AND_MEAN
143 #include <windows.h>
_mi_prim_thread_id(void)144 static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
145 // Windows: works on Intel and ARM in both 32- and 64-bit
146 return (uintptr_t)NtCurrentTeb();
147 }
148
149 // We use assembly for a fast thread id on the main platforms. The TLS layout depends on
150 // both the OS and libc implementation so we use specific tests for each main platform.
151 // If you test on another platform and it works please send a PR :-)
152 // see also https://akkadia.org/drepper/tls.pdf for more info on the TLS register.
153 #elif defined(__GNUC__) && ( \
154 (defined(__GLIBC__) && (defined(__x86_64__) || defined(__i386__) || (defined(__arm__) && __ARM_ARCH >= 7) || defined(__aarch64__))) \
155 || (defined(__APPLE__) && (defined(__x86_64__) || defined(__aarch64__))) \
156 || (defined(__BIONIC__) && (defined(__x86_64__) || defined(__i386__) || (defined(__arm__) && __ARM_ARCH >= 7) || defined(__aarch64__))) \
157 || (defined(__FreeBSD__) && (defined(__x86_64__) || defined(__i386__) || defined(__aarch64__))) \
158 || (defined(__OpenBSD__) && (defined(__x86_64__) || defined(__i386__) || defined(__aarch64__))) \
159 )
160
mi_prim_tls_slot(size_t slot)161 static inline void* mi_prim_tls_slot(size_t slot) mi_attr_noexcept {
162 void* res;
163 const size_t ofs = (slot*sizeof(void*));
164 #if defined(__i386__)
165 __asm__("movl %%gs:%1, %0" : "=r" (res) : "m" (*((void**)ofs)) : ); // x86 32-bit always uses GS
166 #elif defined(__APPLE__) && defined(__x86_64__)
167 __asm__("movq %%gs:%1, %0" : "=r" (res) : "m" (*((void**)ofs)) : ); // x86_64 macOSX uses GS
168 #elif defined(__x86_64__) && (MI_INTPTR_SIZE==4)
169 __asm__("movl %%fs:%1, %0" : "=r" (res) : "m" (*((void**)ofs)) : ); // x32 ABI
170 #elif defined(__x86_64__)
171 __asm__("movq %%fs:%1, %0" : "=r" (res) : "m" (*((void**)ofs)) : ); // x86_64 Linux, BSD uses FS
172 #elif defined(__arm__)
173 void** tcb; MI_UNUSED(ofs);
174 __asm__ volatile ("mrc p15, 0, %0, c13, c0, 3\nbic %0, %0, #3" : "=r" (tcb));
175 res = tcb[slot];
176 #elif defined(__aarch64__)
177 void** tcb; MI_UNUSED(ofs);
178 #if defined(__APPLE__) // M1, issue #343
179 __asm__ volatile ("mrs %0, tpidrro_el0\nbic %0, %0, #7" : "=r" (tcb));
180 #else
181 __asm__ volatile ("mrs %0, tpidr_el0" : "=r" (tcb));
182 #endif
183 res = tcb[slot];
184 #endif
185 return res;
186 }
187
188 // setting a tls slot is only used on macOS for now
mi_prim_tls_slot_set(size_t slot,void * value)189 static inline void mi_prim_tls_slot_set(size_t slot, void* value) mi_attr_noexcept {
190 const size_t ofs = (slot*sizeof(void*));
191 #if defined(__i386__)
192 __asm__("movl %1,%%gs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : ); // 32-bit always uses GS
193 #elif defined(__APPLE__) && defined(__x86_64__)
194 __asm__("movq %1,%%gs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : ); // x86_64 macOS uses GS
195 #elif defined(__x86_64__) && (MI_INTPTR_SIZE==4)
196 __asm__("movl %1,%%fs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : ); // x32 ABI
197 #elif defined(__x86_64__)
198 __asm__("movq %1,%%fs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : ); // x86_64 Linux, BSD uses FS
199 #elif defined(__arm__)
200 void** tcb; MI_UNUSED(ofs);
201 __asm__ volatile ("mrc p15, 0, %0, c13, c0, 3\nbic %0, %0, #3" : "=r" (tcb));
202 tcb[slot] = value;
203 #elif defined(__aarch64__)
204 void** tcb; MI_UNUSED(ofs);
205 #if defined(__APPLE__) // M1, issue #343
206 __asm__ volatile ("mrs %0, tpidrro_el0\nbic %0, %0, #7" : "=r" (tcb));
207 #else
208 __asm__ volatile ("mrs %0, tpidr_el0" : "=r" (tcb));
209 #endif
210 tcb[slot] = value;
211 #endif
212 }
213
_mi_prim_thread_id(void)214 static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
215 #if defined(__BIONIC__)
216 // issue #384, #495: on the Bionic libc (Android), slot 1 is the thread id
217 // see: https://github.com/aosp-mirror/platform_bionic/blob/c44b1d0676ded732df4b3b21c5f798eacae93228/libc/platform/bionic/tls_defines.h#L86
218 return (uintptr_t)mi_prim_tls_slot(1);
219 #else
220 // in all our other targets, slot 0 is the thread id
221 // glibc: https://sourceware.org/git/?p=glibc.git;a=blob_plain;f=sysdeps/x86_64/nptl/tls.h
222 // apple: https://github.com/apple/darwin-xnu/blob/main/libsyscall/os/tsd.h#L36
223 return (uintptr_t)mi_prim_tls_slot(0);
224 #endif
225 }
226
227 #else
228
229 // otherwise use portable C, taking the address of a thread local variable (this is still very fast on most platforms).
_mi_prim_thread_id(void)230 static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
231 return (uintptr_t)&_mi_heap_default;
232 }
233
234 #endif
235
236
237
238 /* ----------------------------------------------------------------------------------------
239 The thread local default heap: `_mi_prim_get_default_heap()`
240 This is inlined here as it is on the fast path for allocation functions.
241
242 On most platforms (Windows, Linux, FreeBSD, NetBSD, etc), this just returns a
243 __thread local variable (`_mi_heap_default`). With the initial-exec TLS model this ensures
244 that the storage will always be available (allocated on the thread stacks).
245
246 On some platforms though we cannot use that when overriding `malloc` since the underlying
247 TLS implementation (or the loader) will call itself `malloc` on a first access and recurse.
248 We try to circumvent this in an efficient way:
249 - macOSX : we use an unused TLS slot from the OS allocated slots (MI_TLS_SLOT). On OSX, the
250 loader itself calls `malloc` even before the modules are initialized.
251 - OpenBSD: we use an unused slot from the pthread block (MI_TLS_PTHREAD_SLOT_OFS).
252 - DragonFly: defaults are working but seem slow compared to freeBSD (see PR #323)
253 ------------------------------------------------------------------------------------------- */
254
255 static inline mi_heap_t* mi_prim_get_default_heap(void);
256
257 #if defined(MI_MALLOC_OVERRIDE)
258 #if defined(__APPLE__) // macOS
259 #define MI_TLS_SLOT 89 // seems unused?
260 // #define MI_TLS_RECURSE_GUARD 1
261 // other possible unused ones are 9, 29, __PTK_FRAMEWORK_JAVASCRIPTCORE_KEY4 (94), __PTK_FRAMEWORK_GC_KEY9 (112) and __PTK_FRAMEWORK_OLDGC_KEY9 (89)
262 // see <https://github.com/rweichler/substrate/blob/master/include/pthread_machdep.h>
263 #elif defined(__OpenBSD__)
264 // use end bytes of a name; goes wrong if anyone uses names > 23 characters (ptrhread specifies 16)
265 // see <https://github.com/openbsd/src/blob/master/lib/libc/include/thread_private.h#L371>
266 #define MI_TLS_PTHREAD_SLOT_OFS (6*sizeof(int) + 4*sizeof(void*) + 24)
267 // #elif defined(__DragonFly__)
268 // #warning "mimalloc is not working correctly on DragonFly yet."
269 // #define MI_TLS_PTHREAD_SLOT_OFS (4 + 1*sizeof(void*)) // offset `uniqueid` (also used by gdb?) <https://github.com/DragonFlyBSD/DragonFlyBSD/blob/master/lib/libthread_xu/thread/thr_private.h#L458>
270 #elif defined(__ANDROID__)
271 // See issue #381
272 #define MI_TLS_PTHREAD
273 #endif
274 #endif
275
276
277 #if defined(MI_TLS_SLOT)
278
mi_prim_get_default_heap(void)279 static inline mi_heap_t* mi_prim_get_default_heap(void) {
280 mi_heap_t* heap = (mi_heap_t*)mi_prim_tls_slot(MI_TLS_SLOT);
281 if mi_unlikely(heap == NULL) {
282 #ifdef __GNUC__
283 __asm(""); // prevent conditional load of the address of _mi_heap_empty
284 #endif
285 heap = (mi_heap_t*)&_mi_heap_empty;
286 }
287 return heap;
288 }
289
290 #elif defined(MI_TLS_PTHREAD_SLOT_OFS)
291
mi_prim_tls_pthread_heap_slot(void)292 static inline mi_heap_t** mi_prim_tls_pthread_heap_slot(void) {
293 pthread_t self = pthread_self();
294 #if defined(__DragonFly__)
295 if (self==NULL) return NULL;
296 #endif
297 return (mi_heap_t**)((uint8_t*)self + MI_TLS_PTHREAD_SLOT_OFS);
298 }
299
mi_prim_get_default_heap(void)300 static inline mi_heap_t* mi_prim_get_default_heap(void) {
301 mi_heap_t** pheap = mi_prim_tls_pthread_heap_slot();
302 if mi_unlikely(pheap == NULL) return _mi_heap_main_get();
303 mi_heap_t* heap = *pheap;
304 if mi_unlikely(heap == NULL) return (mi_heap_t*)&_mi_heap_empty;
305 return heap;
306 }
307
308 #elif defined(MI_TLS_PTHREAD)
309
310 extern pthread_key_t _mi_heap_default_key;
mi_prim_get_default_heap(void)311 static inline mi_heap_t* mi_prim_get_default_heap(void) {
312 mi_heap_t* heap = (mi_unlikely(_mi_heap_default_key == (pthread_key_t)(-1)) ? _mi_heap_main_get() : (mi_heap_t*)pthread_getspecific(_mi_heap_default_key));
313 return (mi_unlikely(heap == NULL) ? (mi_heap_t*)&_mi_heap_empty : heap);
314 }
315
316 #else // default using a thread local variable; used on most platforms.
317
mi_prim_get_default_heap(void)318 static inline mi_heap_t* mi_prim_get_default_heap(void) {
319 #if defined(MI_TLS_RECURSE_GUARD)
320 if (mi_unlikely(!_mi_process_is_initialized)) return _mi_heap_main_get();
321 #endif
322 return _mi_heap_default;
323 }
324
325 #endif // mi_prim_get_default_heap()
326
327
328
329 #endif // MIMALLOC_PRIM_H
330