1 #include <unistd.h> // pread, pwrite
3 #include "mc_page_store.h"
5 #include "mc_private.h"
6 #include "mc_snapshot.h"
8 #include <xbt/mmalloc.h>
10 #define SOFT_DIRTY_BIT_NUMBER 55
11 #define SOFT_DIRTY (((uint64_t)1) << SOFT_DIRTY_BIT_NUMBER)
15 // ***** Region management:
17 /** @brief Take a per-page snapshot of a region
19 * @param data The start of the region (must be at the beginning of a page)
20 * @param pag_count Number of pages of the region
21 * @param pagemap Linux kernel pagemap values fot this region (or NULL)
22 * @param reference_pages Snapshot page numbers of the previous soft_dirty_reset (or NULL)
23 * @return Snapshot page numbers of this new snapshot
25 size_t* mc_take_page_snapshot_region(mc_process_t process,
26 void* data, size_t page_count, uint64_t* pagemap, size_t* reference_pages)
28 size_t* pagenos = (size_t*) malloc(page_count * sizeof(size_t));
30 const bool is_self = MC_process_is_self(process);
34 temp = malloc(xbt_pagebits);
36 for (size_t i=0; i!=page_count; ++i) {
37 bool softclean = pagemap && !(pagemap[i] & SOFT_DIRTY);
38 if (softclean && reference_pages) {
39 // The page is softclean, it is the same page as the reference page:
40 pagenos[i] = reference_pages[i];
41 mc_model_checker->pages->ref_page(reference_pages[i]);
43 // Otherwise, we need to store the page the hard way
44 // (by reading its content):
45 void* page = (char*) data + (i << xbt_pagebits);
46 xbt_assert(mc_page_offset(page)==0, "Not at the beginning of a page");
51 /* Adding another copy (and a syscall) will probably slow things a lot.
52 TODO, optimize this somehow (at least by grouping the syscalls)
54 - reduce the number of syscalls;
55 - let the application snapshot itself;
56 - move the segments in shared memory (this will break `fork` however).
59 MC_process_read(process, temp, page, xbt_pagesize);
61 pagenos[i] = mc_model_checker->pages->store_page(page_data);
69 void mc_free_page_snapshot_region(size_t* pagenos, size_t page_count)
71 for (size_t i=0; i!=page_count; ++i) {
72 mc_model_checker->pages->unref_page(pagenos[i]);
76 /** @brief Restore a snapshot of a region
78 * If possible, the restoration will be incremental
79 * (the modified pages will not be touched).
82 * @param page_count Number of pages of the region
84 * @param pagemap Linux kernel pagemap values fot this region (or NULL)
85 * @param reference_pages Snapshot page numbers of the previous soft_dirty_reset (or NULL)
87 void mc_restore_page_snapshot_region(mc_process_t process,
88 void* start_addr, size_t page_count, size_t* pagenos, uint64_t* pagemap, size_t* reference_pagenos)
90 for (size_t i=0; i!=page_count; ++i) {
92 bool softclean = pagemap && !(pagemap[i] & SOFT_DIRTY);
93 if (softclean && reference_pagenos && pagenos[i] == reference_pagenos[i]) {
94 // The page is softclean and is the same as the reference one:
95 // the page is already in the target state.
99 // Otherwise, copy the page:
100 void* target_page = mc_page_from_number(start_addr, i);
101 const void* source_page = mc_model_checker->pages->get_page(pagenos[i]);
102 MC_process_write(process, source_page, target_page, xbt_pagesize);
106 // ***** Soft dirty tracking
108 /** @brief Like pread() but without partial reads */
109 static size_t pread_whole(int fd, void* buf, size_t count, off_t offset) {
112 char* data = (char*) buf;
114 ssize_t n = pread(fd, buf, count, offset);
127 // It might be a partial read:
137 static inline __attribute__ ((always_inline))
138 void mc_ensure_fd(int* fd, const char* path, int flags) {
141 *fd = open(path, flags);
143 xbt_die("Could not open file %s", path);
147 /** @brief Reset the soft-dirty bits
149 * This is done after checkpointing and after checkpoint restoration
150 * (if per page checkpoiting is used) in order to know which pages were
153 * See https://www.kernel.org/doc/Documentation/vm/soft-dirty.txt
155 void mc_softdirty_reset() {
156 mc_ensure_fd(&mc_model_checker->fd_clear_refs, "/proc/self/clear_refs", O_WRONLY|O_CLOEXEC);
157 if( ::write(mc_model_checker->fd_clear_refs, "4\n", 2) != 2) {
158 xbt_die("Could not reset softdirty bits");
162 /** @brief Read memory page informations
164 * For each virtual memory page of the process,
165 * /proc/self/pagemap provides a 64 bit field of information.
166 * We are interested in the soft-dirty bit: with this we can track which
167 * pages were modified between snapshots/restorations and avoid
168 * copying data which was not modified.
170 * See https://www.kernel.org/doc/Documentation/vm/pagemap.txt
172 * @param pagemap Output buffer for pagemap informations
173 * @param start_addr Address of the first page
174 * @param page_count Number of pages
176 static void mc_read_pagemap(uint64_t* pagemap, size_t page_start, size_t page_count)
178 mc_ensure_fd(&mc_model_checker->fd_pagemap, "/proc/self/pagemap", O_RDONLY|O_CLOEXEC);
179 size_t bytesize = sizeof(uint64_t) * page_count;
180 off_t offset = sizeof(uint64_t) * page_start;
181 if (pread_whole(mc_model_checker->fd_pagemap, pagemap, bytesize, offset) != bytesize) {
182 xbt_die("Could not read pagemap");
186 // ***** High level API
188 mc_mem_region_t mc_region_new_sparse(mc_region_type_t region_type,
189 void *start_addr, void* permanent_addr, size_t size,
190 mc_mem_region_t ref_reg)
192 mc_process_t process = &mc_model_checker->process;
194 mc_mem_region_t region = xbt_new(s_mc_mem_region_t, 1);
195 region->region_type = region_type;
196 region->storage_type = MC_REGION_STORAGE_TYPE_CHUNKED;
197 region->start_addr = start_addr;
198 region->permanent_addr = permanent_addr;
201 xbt_assert((((uintptr_t)start_addr) & (xbt_pagesize-1)) == 0,
202 "Not at the beginning of a page");
203 xbt_assert((((uintptr_t)permanent_addr) & (xbt_pagesize-1)) == 0,
204 "Not at the beginning of a page");
205 size_t page_count = mc_page_count(size);
207 uint64_t* pagemap = NULL;
208 if (_sg_mc_soft_dirty && mc_model_checker->parent_snapshot &&
209 MC_process_is_self(process)) {
210 pagemap = (uint64_t*) mmalloc_no_memset(mc_heap, sizeof(uint64_t) * page_count);
211 mc_read_pagemap(pagemap, mc_page_number(NULL, permanent_addr), page_count);
214 size_t* reg_page_numbers = NULL;
215 if (ref_reg!=NULL && ref_reg->storage_type == MC_REGION_STORAGE_TYPE_CHUNKED)
216 reg_page_numbers = ref_reg->chunked.page_numbers;
218 // Take incremental snapshot:
219 region->chunked.page_numbers = mc_take_page_snapshot_region(process,
220 permanent_addr, page_count, pagemap, reg_page_numbers);
223 mfree(mc_heap, pagemap);
228 void mc_region_restore_sparse(mc_process_t process, mc_mem_region_t reg, mc_mem_region_t ref_reg)
230 xbt_assert((((uintptr_t)reg->permanent_addr) & (xbt_pagesize-1)) == 0,
231 "Not at the beginning of a page");
232 size_t page_count = mc_page_count(reg->size);
234 uint64_t* pagemap = NULL;
236 // Read soft-dirty bits if necessary in order to know which pages have changed:
237 if (_sg_mc_soft_dirty && mc_model_checker->parent_snapshot
238 && MC_process_is_self(process)) {
239 pagemap = (uint64_t*) mmalloc_no_memset(mc_heap, sizeof(uint64_t) * page_count);
240 mc_read_pagemap(pagemap, mc_page_number(NULL, reg->permanent_addr), page_count);
243 // Incremental per-page snapshot restoration:s
244 size_t* reg_page_numbers = NULL;
245 if (ref_reg && ref_reg->storage_type == MC_REGION_STORAGE_TYPE_CHUNKED)
246 reg_page_numbers = ref_reg->chunked.page_numbers;
248 mc_restore_page_snapshot_region(process,
249 reg->permanent_addr, page_count, reg->chunked.page_numbers,
250 pagemap, reg_page_numbers);