1 #include "mc_page_store.h"
3 #include "mc_private.h"
4 #include "mc_snapshot.h"
6 #include <xbt/mmalloc.h>
8 #define SOFT_DIRTY_BIT_NUMBER 55
9 #define SOFT_DIRTY (((uint64_t)1) << SOFT_DIRTY_BIT_NUMBER)
13 // ***** Region management:
15 /** @brief Take a per-page snapshot of a region
17 * @param data The start of the region (must be at the beginning of a page)
18 * @param pag_count Number of pages of the region
19 * @param pagemap Linux kernel pagemap values fot this region (or NULL)
20 * @param reference_pages Snapshot page numbers of the previous soft_dirty_reset (or NULL)
21 * @return Snapshot page numbers of this new snapshot
23 size_t* mc_take_page_snapshot_region(void* data, size_t page_count, uint64_t* pagemap, size_t* reference_pages)
25 size_t* pagenos = (size_t*) malloc(page_count * sizeof(size_t));
27 for (size_t i=0; i!=page_count; ++i) {
28 bool softclean = pagemap && !(pagemap[i] & SOFT_DIRTY);
29 if (softclean && reference_pages) {
30 // The page is softclean, it is the same page as the reference page:
31 pagenos[i] = reference_pages[i];
32 mc_model_checker->pages->ref_page(reference_pages[i]);
34 // Otherwise, we need to store the page the hard way
35 // (by reading its content):
36 void* page = (char*) data + (i << xbt_pagebits);
37 pagenos[i] = mc_model_checker->pages->store_page(page);
44 void mc_free_page_snapshot_region(size_t* pagenos, size_t page_count)
46 for (size_t i=0; i!=page_count; ++i) {
47 mc_model_checker->pages->unref_page(pagenos[i]);
51 /** @brief Restore a snapshot of a region
53 * If possible, the restoration will be incremental
54 * (the modified pages will not be touched).
57 * @param page_count Number of pages of the region
59 * @param pagemap Linux kernel pagemap values fot this region (or NULL)
60 * @param reference_pages Snapshot page numbers of the previous soft_dirty_reset (or NULL)
62 void mc_restore_page_snapshot_region(void* start_addr, size_t page_count, size_t* pagenos, uint64_t* pagemap, size_t* reference_pagenos)
64 for (size_t i=0; i!=page_count; ++i) {
66 bool softclean = pagemap && !(pagemap[i] & SOFT_DIRTY);
67 if (softclean && reference_pagenos && pagenos[i] == reference_pagenos[i]) {
68 // The page is softclean and is the same as the reference one:
69 // the page is already in the target state.
73 // Otherwise, copy the page:
74 void* target_page = mc_page_from_number(start_addr, i);
75 const void* source_page = mc_model_checker->pages->get_page(pagenos[i]);
76 memcpy(target_page, source_page, xbt_pagesize);
80 // ***** Soft dirty tracking
82 /** @brief Like pread() but without partial reads */
83 static size_t pread_whole(int fd, void* buf, size_t count, off_t offset) {
86 char* data = (char*) buf;
88 ssize_t n = pread(fd, buf, count, offset);
101 // It might be a partial read:
111 static inline __attribute__ ((always_inline))
112 void mc_ensure_fd(int* fd, const char* path, int flags) {
115 *fd = open(path, flags);
117 xbt_die("Could not open file %s", path);
121 /** @brief Reset the soft-dirty bits
123 * This is done after checkpointing and after checkpoint restoration
124 * (if per page checkpoiting is used) in order to know which pages were
127 * See https://www.kernel.org/doc/Documentation/vm/soft-dirty.txt
129 void mc_softdirty_reset() {
130 mc_ensure_fd(&mc_model_checker->fd_clear_refs, "/proc/self/clear_refs", O_WRONLY|O_CLOEXEC);
131 if( ::write(mc_model_checker->fd_clear_refs, "4\n", 2) != 2) {
132 xbt_die("Could not reset softdirty bits");
136 /** @brief Read memory page informations
138 * For each virtual memory page of the process,
139 * /proc/self/pagemap provides a 64 bit field of information.
140 * We are interested in the soft-dirty bit: with this we can track which
141 * pages were modified between snapshots/restorations and avoid
142 * copying data which was not modified.
144 * See https://www.kernel.org/doc/Documentation/vm/pagemap.txt
146 * @param pagemap Output buffer for pagemap informations
147 * @param start_addr Address of the first page
148 * @param page_count Number of pages
150 static void mc_read_pagemap(uint64_t* pagemap, size_t page_start, size_t page_count)
152 mc_ensure_fd(&mc_model_checker->fd_pagemap, "/proc/self/pagemap", O_RDONLY|O_CLOEXEC);
153 size_t bytesize = sizeof(uint64_t) * page_count;
154 off_t offset = sizeof(uint64_t) * page_start;
155 if (pread_whole(mc_model_checker->fd_pagemap, pagemap, bytesize, offset) != bytesize) {
156 xbt_die("Could not read pagemap");
160 // ***** High level API
162 mc_mem_region_t mc_region_new_sparse(int type, void *start_addr, void* permanent_addr, size_t size, mc_mem_region_t ref_reg)
164 mc_mem_region_t new_reg = xbt_new(s_mc_mem_region_t, 1);
166 new_reg->start_addr = start_addr;
167 new_reg->permanent_addr = permanent_addr;
168 new_reg->data = NULL;
169 new_reg->size = size;
170 new_reg->page_numbers = NULL;
172 xbt_assert((((uintptr_t)start_addr) & (xbt_pagesize-1)) == 0,
173 "Not at the beginning of a page");
174 xbt_assert((((uintptr_t)permanent_addr) & (xbt_pagesize-1)) == 0,
175 "Not at the beginning of a page");
176 size_t page_count = mc_page_count(size);
178 uint64_t* pagemap = NULL;
179 if (_sg_mc_soft_dirty && mc_model_checker->parent_snapshot) {
180 pagemap = (uint64_t*) mmalloc_no_memset(mc_heap, sizeof(uint64_t) * page_count);
181 mc_read_pagemap(pagemap, mc_page_number(NULL, permanent_addr), page_count);
184 // Take incremental snapshot:
185 new_reg->page_numbers = mc_take_page_snapshot_region(permanent_addr, page_count, pagemap,
186 ref_reg==NULL ? NULL : ref_reg->page_numbers);
189 mfree(mc_heap, pagemap);
194 void mc_region_restore_sparse(mc_mem_region_t reg, mc_mem_region_t ref_reg)
196 xbt_assert((((uintptr_t)reg->permanent_addr) & (xbt_pagesize-1)) == 0,
197 "Not at the beginning of a page");
198 size_t page_count = mc_page_count(reg->size);
200 uint64_t* pagemap = NULL;
202 // Read soft-dirty bits if necessary in order to know which pages have changed:
203 if (_sg_mc_soft_dirty && mc_model_checker->parent_snapshot) {
204 pagemap = (uint64_t*) mmalloc_no_memset(mc_heap, sizeof(uint64_t) * page_count);
205 mc_read_pagemap(pagemap, mc_page_number(NULL, reg->permanent_addr), page_count);
208 // Incremental per-page snapshot restoration:
209 mc_restore_page_snapshot_region(reg->permanent_addr, page_count, reg->page_numbers,
210 pagemap, ref_reg ? ref_reg->page_numbers : NULL);