1 #include "mc_page_store.h"
3 #include "mc_private.h"
5 #include <xbt/mmalloc.h>
7 #define SOFT_DIRTY_BIT_NUMBER 55
8 #define SOFT_DIRTY (((uint64_t)1) << SOFT_DIRTY_BIT_NUMBER)
12 // ***** Region management:
14 /** @brief Take a per-page snapshot of a region
16 * @param data The start of the region (must be at the beginning of a page)
17 * @param pag_count Number of pages of the region
18 * @param pagemap Linux kernel pagemap values fot this region (or NULL)
19 * @param reference_pages Snapshot page numbers of the previous soft_dirty_reset (or NULL)
20 * @return Snapshot page numbers of this new snapshot
22 size_t* mc_take_page_snapshot_region(void* data, size_t page_count, uint64_t* pagemap, size_t* reference_pages)
24 size_t* pagenos = (size_t*) malloc(page_count * sizeof(size_t));
26 for (size_t i=0; i!=page_count; ++i) {
27 bool softclean = pagemap && !(pagemap[i] & SOFT_DIRTY);
28 if (softclean && reference_pages) {
29 // The page is softclean, it is the same page as the reference page:
30 pagenos[i] = reference_pages[i];
31 mc_model_checker->pages->ref_page(reference_pages[i]);
33 // Otherwise, we need to store the page the hard way
34 // (by reading its content):
35 void* page = (char*) data + (i << xbt_pagebits);
36 pagenos[i] = mc_model_checker->pages->store_page(page);
43 void mc_free_page_snapshot_region(size_t* pagenos, size_t page_count)
45 for (size_t i=0; i!=page_count; ++i) {
46 mc_model_checker->pages->unref_page(pagenos[i]);
50 /** @brief Restore a snapshot of a region
52 * If possible, the restoration will be incremental
53 * (the modified pages will not be touched).
56 * @param page_count Number of pages of the region
58 * @param pagemap Linux kernel pagemap values fot this region (or NULL)
59 * @param reference_pages Snapshot page numbers of the previous soft_dirty_reset (or NULL)
61 void mc_restore_page_snapshot_region(void* start_addr, size_t page_count, size_t* pagenos, uint64_t* pagemap, size_t* reference_pagenos)
63 for (size_t i=0; i!=page_count; ++i) {
65 bool softclean = pagemap && !(pagemap[i] & SOFT_DIRTY);
66 if (softclean && reference_pagenos && pagenos[i] == reference_pagenos[i]) {
67 // The page is softclean and is the same as the reference one:
68 // the page is already in the target state.
72 // Otherwise, copy the page:
73 void* target_page = mc_page_from_number(start_addr, i);
74 const void* source_page = mc_model_checker->pages->get_page(pagenos[i]);
75 memcpy(target_page, source_page, xbt_pagesize);
79 // ***** Soft dirty tracking
81 /** @brief Like pread() but without partial reads */
82 static size_t pread_whole(int fd, void* buf, size_t count, off_t offset) {
85 char* data = (char*) buf;
87 ssize_t n = pread(fd, buf, count, offset);
100 // It might be a partial read:
110 static inline __attribute__ ((always_inline))
111 void mc_ensure_fd(int* fd, const char* path, int flags) {
114 *fd = open(path, flags);
116 xbt_die("Could not open file %s", path);
120 /** @brief Reset the soft-dirty bits
122 * This is done after checkpointing and after checkpoint restoration
123 * (if per page checkpoiting is used) in order to know which pages were
126 * See https://www.kernel.org/doc/Documentation/vm/soft-dirty.txt
128 void mc_softdirty_reset() {
129 mc_ensure_fd(&mc_model_checker->fd_clear_refs, "/proc/self/clear_refs", O_WRONLY|O_CLOEXEC);
130 if( ::write(mc_model_checker->fd_clear_refs, "4\n", 2) != 2) {
131 xbt_die("Could not reset softdirty bits");
135 /** @brief Read memory page informations
137 * For each virtual memory page of the process,
138 * /proc/self/pagemap provides a 64 bit field of information.
139 * We are interested in the soft-dirty bit: with this we can track which
140 * pages were modified between snapshots/restorations and avoid
141 * copying data which was not modified.
143 * See https://www.kernel.org/doc/Documentation/vm/pagemap.txt
145 * @param pagemap Output buffer for pagemap informations
146 * @param start_addr Address of the first page
147 * @param page_count Number of pages
149 static void mc_read_pagemap(uint64_t* pagemap, size_t page_start, size_t page_count)
151 mc_ensure_fd(&mc_model_checker->fd_pagemap, "/proc/self/pagemap", O_RDONLY|O_CLOEXEC);
152 size_t bytesize = sizeof(uint64_t) * page_count;
153 off_t offset = sizeof(uint64_t) * page_start;
154 if (pread_whole(mc_model_checker->fd_pagemap, pagemap, bytesize, offset) != bytesize) {
155 xbt_die("Could not read pagemap");
159 // ***** High level API
161 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)
163 mc_mem_region_t new_reg = xbt_new(s_mc_mem_region_t, 1);
165 new_reg->start_addr = start_addr;
166 new_reg->permanent_addr = permanent_addr;
167 new_reg->data = NULL;
168 new_reg->size = size;
169 new_reg->page_numbers = NULL;
171 xbt_assert((((uintptr_t)start_addr) & (xbt_pagesize-1)) == 0,
172 "Not at the beginning of a page");
173 xbt_assert((((uintptr_t)permanent_addr) & (xbt_pagesize-1)) == 0,
174 "Not at the beginning of a page");
175 size_t page_count = mc_page_count(size);
177 uint64_t* pagemap = NULL;
178 if (_sg_mc_soft_dirty && mc_model_checker->parent_snapshot) {
179 pagemap = (uint64_t*) mmalloc_no_memset((xbt_mheap_t) mc_heap, sizeof(uint64_t) * page_count);
180 mc_read_pagemap(pagemap, mc_page_number(NULL, permanent_addr), page_count);
183 // Take incremental snapshot:
184 new_reg->page_numbers = mc_take_page_snapshot_region(permanent_addr, page_count, pagemap,
185 ref_reg==NULL ? NULL : ref_reg->page_numbers);
188 mfree((xbt_mheap_t) mc_heap, pagemap);
193 void mc_region_restore_sparse(mc_mem_region_t reg, mc_mem_region_t ref_reg)
195 xbt_assert((((uintptr_t)reg->permanent_addr) & (xbt_pagesize-1)) == 0,
196 "Not at the beginning of a page");
197 size_t page_count = mc_page_count(reg->size);
199 uint64_t* pagemap = NULL;
201 // Read soft-dirty bits if necessary in order to know which pages have changed:
202 if (_sg_mc_soft_dirty && mc_model_checker->parent_snapshot) {
203 pagemap = (uint64_t*) mmalloc_no_memset((xbt_mheap_t) mc_heap, sizeof(uint64_t) * page_count);
204 mc_read_pagemap(pagemap, mc_page_number(NULL, reg->permanent_addr), page_count);
207 // Incremental per-page snapshot restoration:
208 mc_restore_page_snapshot_region(reg->permanent_addr, page_count, reg->page_numbers,
209 pagemap, ref_reg ? ref_reg->page_numbers : NULL);