2 * the Linux incarnation of OS-dependent routines
4 * This file (along with os.h) exports an OS-independent interface to
5 * the operating system VM facilities. Surprise surprise, this
6 * interface looks a lot like the Mach interface (but simpler in some
7 * places). For some operating systems, a subset of these functions
8 * will have to be emulated.
12 * This software is part of the SBCL system. See the README file for
15 * This software is derived from the CMU CL system, which was
16 * written at Carnegie Mellon University and released into the
17 * public domain. The software is in the public domain and is
18 * provided with absolutely no warranty. See the COPYING and CREDITS
19 * files for more information.
27 #include <sys/param.h>
33 #include "interrupt.h"
36 #include <sys/socket.h>
37 #include <sys/utsname.h>
39 #include <sys/types.h>
41 /* #include <sys/sysinfo.h> */
46 #include "x86-validate.h"
47 size_t os_vm_page_size;
55 /* Early versions of Linux don't support the mmap(..) functionality
61 major_version = atoi(name.release);
62 if (major_version < 2) {
63 lose("linux major version=%d (can't run in version < 2.0.0)",
68 os_vm_page_size = getpagesize();
70 SET_FPU_CONTROL_WORD(0x1372|4|8|16|32); /* no interrupts */
73 /* KLUDGE: As of kernel 2.2.14 on Red Hat 6.2, there's code in the
74 * <sys/ucontext.h> file to define symbolic names for offsets into
75 * gregs[], but it's conditional on __USE_GNU and not defined, so
76 * we need to do this nasty absolute index magic number thing
79 os_context_register_addr(os_context_t *context, int offset)
82 case 0: return &context->uc_mcontext.gregs[11]; /* EAX */
83 case 2: return &context->uc_mcontext.gregs[10]; /* ECX */
84 case 4: return &context->uc_mcontext.gregs[9]; /* EDX */
85 case 6: return &context->uc_mcontext.gregs[8]; /* EBX */
86 case 8: return &context->uc_mcontext.gregs[7]; /* ESP */
87 case 10: return &context->uc_mcontext.gregs[6]; /* EBP */
88 case 12: return &context->uc_mcontext.gregs[5]; /* ESI */
89 case 14: return &context->uc_mcontext.gregs[4]; /* EDI */
94 os_context_pc_addr(os_context_t *context)
96 return &context->uc_mcontext.gregs[14];
99 os_context_sp_addr(os_context_t *context)
101 return &context->uc_mcontext.gregs[17];
105 os_context_sigmask_addr(os_context_t *context)
107 return &context->uc_sigmask;
110 /* In Debian CMU CL ca. 2.4.9, it was possible to get an infinite
111 * cascade of errors from do_mmap(..). This variable is a counter to
112 * prevent that; when it counts down to zero, an error in do_mmap
113 * causes the low-level monitor to be called. */
114 int n_do_mmap_ignorable_errors = 3;
116 /* Return 0 for success. */
118 do_mmap(os_vm_address_t *addr, os_vm_size_t len, int flags)
120 /* We *must* have the memory where we want it. */
121 os_vm_address_t old_addr=*addr;
123 *addr = mmap(*addr, len, OS_VM_PROT_ALL, flags, -1, 0);
124 if (*addr == MAP_FAILED ||
125 ((old_addr != NULL) && (*addr != old_addr))) {
127 "error in allocating memory from the OS\n"
128 "(addr=%lx, len=%lx, flags=%lx)\n",
132 if (n_do_mmap_ignorable_errors > 0) {
133 --n_do_mmap_ignorable_errors;
135 lose("too many errors in allocating memory from the OS");
144 os_validate(os_vm_address_t addr, os_vm_size_t len)
147 int flags = MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED;
148 os_vm_address_t base_addr = addr;
150 /* KLUDGE: It looks as though this code allocates memory
151 * in chunks of size no larger than 'magic', but why? What
152 * is the significance of 0x1000000 here? Also, can it be
153 * right that if the first few 'do_mmap' calls succeed,
154 * then one fails, we leave the memory allocated by the
155 * first few in place even while we return a code for
156 * complete failure? -- WHN 19991020
158 * Peter Van Eynde writes (20000211)
159 * This was done because the kernel would only check for
160 * overcommit for every allocation seperately. So if you
161 * had 16MB of free mem+swap you could allocate 16M. And
162 * again, and again, etc.
163 * This in [Linux] 2.X could be bad as they changed the memory
164 * system. A side effect was/is (I don't really know) that
165 * programs with a lot of memory mappings run slower. But
166 * of course for 2.2.2X we now have the NO_RESERVE flag that
169 * FIXME: The logic is also flaky w.r.t. failed
170 * allocations. If we make one or more successful calls to
171 * do_mmap(..) before one fails, then we've allocated
172 * memory, and we should ensure that it gets deallocated
173 * sometime somehow. If this function's response to any
174 * failed do_mmap(..) is to give up and return NULL (as in
175 * sbcl-0.6.7), then any failed do_mmap(..) after any
176 * successful do_mmap(..) causes a memory leak. */
177 int magic = 0x1000000;
179 if (do_mmap(&addr, len, flags)) {
184 if (do_mmap(&addr, magic, flags)) {
193 int flags = MAP_PRIVATE | MAP_ANONYMOUS;
194 if (do_mmap(&addr, len, flags)) {
203 os_invalidate(os_vm_address_t addr, os_vm_size_t len)
205 if (munmap(addr,len) == -1) {
211 os_map(int fd, int offset, os_vm_address_t addr, os_vm_size_t len)
213 addr = mmap(addr, len,
215 MAP_PRIVATE | MAP_FILE | MAP_FIXED,
218 if(addr == MAP_FAILED) {
220 lose("unexpected mmap(..) failure");
227 os_flush_icache(os_vm_address_t address, os_vm_size_t length)
232 os_protect(os_vm_address_t address, os_vm_size_t length, os_vm_prot_t prot)
234 if (mprotect(address, length, prot) == -1) {
240 in_range_p(os_vm_address_t a, lispobj sbeg, size_t slen)
242 char* beg = (char*)sbeg;
243 char* end = (char*)sbeg + slen;
244 char* adr = (char*)a;
245 return (adr >= beg && adr < end);
249 is_valid_lisp_addr(os_vm_address_t addr)
252 in_range_p(addr, READ_ONLY_SPACE_START, READ_ONLY_SPACE_SIZE) ||
253 in_range_p(addr, STATIC_SPACE_START , STATIC_SPACE_SIZE) ||
254 in_range_p(addr, DYNAMIC_0_SPACE_START, DYNAMIC_SPACE_SIZE) ||
255 in_range_p(addr, DYNAMIC_1_SPACE_START, DYNAMIC_SPACE_SIZE) ||
256 in_range_p(addr, CONTROL_STACK_START , CONTROL_STACK_SIZE) ||
257 in_range_p(addr, BINDING_STACK_START , BINDING_STACK_SIZE);
261 * any OS-dependent special low-level handling for signals
267 os_install_interrupt_handlers(void)
273 * The GENCGC needs to be hooked into whatever signal is raised for
274 * page fault on this OS.
277 sigsegv_handler(int signal, siginfo_t *info, void* void_context)
279 os_context_t *context = (os_context_t*)void_context;
280 void* fault_addr = (void*)context->uc_mcontext.cr2;
281 if (!gencgc_handle_wp_violation(fault_addr)) {
282 interrupt_handle_now(signal, info, void_context);
286 os_install_interrupt_handlers(void)
288 interrupt_install_low_level_handler(SIGSEGV, sigsegv_handler);