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.
23 #include <sys/param.h>
29 #include "interrupt.h"
33 #include <sys/socket.h>
34 #include <sys/utsname.h>
36 #include <sys/types.h>
38 /* #include <sys/sysinfo.h> */
43 #include "x86-validate.h"
44 size_t os_vm_page_size;
52 /* Early versions of Linux don't support the mmap(..) functionality
58 major_version = atoi(name.release);
59 if (major_version < 2) {
60 lose("linux major version=%d (can't run in version < 2.0.0)",
65 os_vm_page_size = getpagesize();
67 SET_FPU_CONTROL_WORD(0x1372|4|8|16|32); /* no interrupts */
70 /* KLUDGE: As of kernel 2.2.14 on Red Hat 6.2, there's code in the
71 * <sys/ucontext.h> file to define symbolic names for offsets into
72 * gregs[], but it's conditional on __USE_GNU and not defined, so
73 * we need to do this nasty absolute index magic number thing
76 os_context_register_addr(os_context_t *context, int offset)
79 case 0: return &context->uc_mcontext.gregs[11]; /* EAX */
80 case 2: return &context->uc_mcontext.gregs[10]; /* ECX */
81 case 4: return &context->uc_mcontext.gregs[9]; /* EDX */
82 case 6: return &context->uc_mcontext.gregs[8]; /* EBX */
83 case 8: return &context->uc_mcontext.gregs[7]; /* ESP */
84 case 10: return &context->uc_mcontext.gregs[6]; /* EBP */
85 case 12: return &context->uc_mcontext.gregs[5]; /* ESI */
86 case 14: return &context->uc_mcontext.gregs[4]; /* EDI */
91 os_context_pc_addr(os_context_t *context)
93 return &context->uc_mcontext.gregs[14];
96 os_context_sp_addr(os_context_t *context)
98 return &context->uc_mcontext.gregs[17];
102 os_context_sigmask_addr(os_context_t *context)
104 return &context->uc_sigmask;
107 /* In Debian CMU CL ca. 2.4.9, it was possible to get an infinite
108 * cascade of errors from do_mmap(..). This variable is a counter to
109 * prevent that; when it counts down to zero, an error in do_mmap
110 * causes the low-level monitor to be called. */
111 int n_do_mmap_ignorable_errors = 3;
113 /* Return 0 for success. */
115 do_mmap(os_vm_address_t *addr, os_vm_size_t len, int flags)
117 /* We *must* have the memory where we want it. */
118 os_vm_address_t old_addr=*addr;
120 *addr = mmap(*addr, len, OS_VM_PROT_ALL, flags, -1, 0);
121 if (*addr == MAP_FAILED ||
122 ((old_addr != NULL) && (*addr != old_addr))) {
124 "error in allocating memory from the OS\n"
125 "(addr=%lx, len=%lx, flags=%lx)\n",
129 if (n_do_mmap_ignorable_errors > 0) {
130 --n_do_mmap_ignorable_errors;
132 lose("too many errors in allocating memory from the OS");
141 os_validate(os_vm_address_t addr, os_vm_size_t len)
144 int flags = MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED;
145 os_vm_address_t base_addr = addr;
147 /* KLUDGE: It looks as though this code allocates memory
148 * in chunks of size no larger than 'magic', but why? What
149 * is the significance of 0x1000000 here? Also, can it be
150 * right that if the first few 'do_mmap' calls succeed,
151 * then one fails, we leave the memory allocated by the
152 * first few in place even while we return a code for
153 * complete failure? -- WHN 19991020
155 * Peter Van Eynde writes (20000211)
156 * This was done because the kernel would only check for
157 * overcommit for every allocation seperately. So if you
158 * had 16MB of free mem+swap you could allocate 16M. And
159 * again, and again, etc.
160 * This in [Linux] 2.X could be bad as they changed the memory
161 * system. A side effect was/is (I don't really know) that
162 * programs with a lot of memory mappings run slower. But
163 * of course for 2.2.2X we now have the NO_RESERVE flag that
166 * FIXME: The logic is also flaky w.r.t. failed
167 * allocations. If we make one or more successful calls to
168 * do_mmap(..) before one fails, then we've allocated
169 * memory, and we should ensure that it gets deallocated
170 * sometime somehow. If this function's response to any
171 * failed do_mmap(..) is to give up and return NULL (as in
172 * sbcl-0.6.7), then any failed do_mmap(..) after any
173 * successful do_mmap(..) causes a memory leak. */
174 int magic = 0x1000000;
176 if (do_mmap(&addr, len, flags)) {
181 if (do_mmap(&addr, magic, flags)) {
190 int flags = MAP_PRIVATE | MAP_ANONYMOUS;
191 if (do_mmap(&addr, len, flags)) {
200 os_invalidate(os_vm_address_t addr, os_vm_size_t len)
202 if (munmap(addr,len) == -1) {
208 os_map(int fd, int offset, os_vm_address_t addr, os_vm_size_t len)
210 addr = mmap(addr, len,
212 MAP_PRIVATE | MAP_FILE | MAP_FIXED,
215 if(addr == MAP_FAILED) {
217 lose("unexpected mmap(..) failure");
224 os_flush_icache(os_vm_address_t address, os_vm_size_t length)
229 os_protect(os_vm_address_t address, os_vm_size_t length, os_vm_prot_t prot)
231 if (mprotect(address, length, prot) == -1) {
236 /* FIXME: Now that FOO_END, rather than FOO_SIZE, is the fundamental
237 * description of a space, we could probably punt this and just do
238 * (FOO_START <= x && x < FOO_END) everywhere it's called. */
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_SPACE_START , DYNAMIC_SPACE_SIZE) ||
255 in_range_p(addr, CONTROL_STACK_START , CONTROL_STACK_SIZE) ||
256 in_range_p(addr, BINDING_STACK_START , BINDING_STACK_SIZE);
260 * any OS-dependent special low-level handling for signals
266 os_install_interrupt_handlers(void)
272 * The GENCGC needs to be hooked into whatever signal is raised for
273 * page fault on this OS.
276 sigsegv_handler(int signal, siginfo_t *info, void* void_context)
278 os_context_t *context = (os_context_t*)void_context;
279 void* fault_addr = (void*)context->uc_mcontext.cr2;
280 if (!gencgc_handle_wp_violation(fault_addr)) {
281 interrupt_handle_now(signal, info, void_context);
285 os_install_interrupt_handlers(void)
287 interrupt_install_low_level_handler(SIGSEGV, sigsegv_handler);