1 #include "mc_page_store.h"
4 #include <xbt/mmalloc.h>
6 #define SOFT_DIRTY_BIT_NUMBER 55
7 #define SOFT_DIRTY (((uint64_t)1) << SOFT_DIRTY_BIT_NUMBER)
11 // ***** Region management:
13 /** @brief Take a per-page snapshot of a region
15 * @param data The start of the region (must be at the beginning of a page)
16 * @param pag_count Number of pages of the region
17 * @param pagemap Linux kernel pagemap values fot this region (or NULL)
18 * @param reference_pages Snapshot page numbers of the previous soft_dirty_reset (or NULL)
19 * @return Snapshot page numbers of this new snapshot
21 size_t* mc_take_page_snapshot_region(void* data, size_t page_count, uint64_t* pagemap, size_t* reference_pages)
23 size_t* pagenos = (size_t*) malloc(page_count * sizeof(size_t));
25 for (size_t i=0; i!=page_count; ++i) {
26 bool softclean = pagemap && !(pagemap[i] & SOFT_DIRTY);
27 if (softclean && reference_pages) {
28 // The page is softclean, it is the same page as the reference page:
29 pagenos[i] = reference_pages[i];
30 mc_model_checker->pages->ref_page(reference_pages[i]);
32 // Otherwise, we need to store the page the hard hard
33 // (by reading its content):
34 void* page = (char*) data + (i << xbt_pagebits);
35 pagenos[i] = mc_model_checker->pages->store_page(page);
42 void mc_free_page_snapshot_region(size_t* pagenos, size_t page_count)
44 for (size_t i=0; i!=page_count; ++i) {
45 mc_model_checker->pages->unref_page(pagenos[i]);
49 /** @brief Restore a snapshot of a region
51 * If possible, the restoration will be incremental
52 * (the modified pages will not be touched).
54 * @param data The start of the region (must be at the beginning of a page)
55 * @param pag_count Number of pages of the region
56 * @param pagemap Linux kernel pagemap values fot this region (or NULL)
57 * @param reference_pages Snapshot page numbers of the previous soft_dirty_reset (or NULL)
59 void mc_restore_page_snapshot_region(mc_mem_region_t region, size_t page_count, uint64_t* pagemap, mc_mem_region_t reference_region)
61 for (size_t i=0; i!=page_count; ++i) {
63 bool softclean = pagemap && !(pagemap[i] & SOFT_DIRTY);
64 if (softclean && reference_region && reference_region->page_numbers[i] == region->page_numbers[i]) {
65 // The page is softclean and is the same as the reference one:
66 // the page is already in the target state.
70 // Otherwise, copy the page:
71 void* target_page = mc_page_from_number(region->start_addr, i);
72 const void* source_page = mc_model_checker->pages->get_page(region->page_numbers[i]);
73 memcpy(target_page, source_page, xbt_pagesize);
77 // ***** Soft dirty tracking
79 /** @brief Like pread() but without partial reads */
80 static size_t pread_whole(int fd, void* buf, size_t count, off_t offset) {
83 char* data = (char*) buf;
85 ssize_t n = pread(fd, buf, count, offset);
98 // It might be a partial read:
108 static inline void mc_ensure_fd(int* fd, const char* path, int flags) {
111 *fd = open(path, flags);
113 xbt_die("Could not open file %s", path);
117 /** @brief Reset the soft-dirty bits
119 * This is done after checkpointing and after checkpoint restoration
120 * (if per page checkpoiting is used) in order to know which pages were
123 * See https://www.kernel.org/doc/Documentation/vm/soft-dirty.txt
125 void mc_softdirty_reset() {
126 mc_ensure_fd(&mc_model_checker->fd_clear_refs, "/proc/self/clear_refs", O_WRONLY|O_CLOEXEC);
127 if( ::write(mc_model_checker->fd_clear_refs, "4\n", 2) != 2) {
128 xbt_die("Could not reset softdirty bits");
132 /** @brief Read memory page informations
134 * For each virtual memory page of the process,
135 * /proc/self/pagemap provides a 64 bit field of information.
136 * We are interested in the soft-dirty bit: with this we can track which
137 * pages were modified between snapshots/restorations and avoid
138 * copying data which was not modified.
140 * See https://www.kernel.org/doc/Documentation/vm/pagemap.txt
142 * @param pagemap Output buffer for pagemap informations
143 * @param start_addr Address of the first page
144 * @param page_count Number of pages
146 static void mc_read_pagemap(uint64_t* pagemap, size_t page_start, size_t page_count)
148 mc_ensure_fd(&mc_model_checker->fd_pagemap, "/proc/self/pagemap", O_RDONLY|O_CLOEXEC);
149 size_t bytesize = sizeof(uint64_t) * page_count;
150 off_t offset = sizeof(uint64_t) * page_start;
151 if (pread_whole(mc_model_checker->fd_pagemap, pagemap, bytesize, offset) != bytesize) {
152 xbt_die("Could not read pagemap");
156 // ***** High level API
158 mc_mem_region_t mc_region_new_sparse(int type, void *start_addr, size_t size, mc_mem_region_t ref_reg)
160 mc_mem_region_t new_reg = xbt_new(s_mc_mem_region_t, 1);
162 new_reg->start_addr = start_addr;
163 new_reg->data = NULL;
164 new_reg->size = size;
165 new_reg->page_numbers = NULL;
167 xbt_assert((((uintptr_t)start_addr) & (xbt_pagesize-1)) == 0,
168 "Not at the beginning of a page");
169 size_t page_count = mc_page_count(size);
171 uint64_t* pagemap = NULL;
172 if (_sg_mc_soft_dirty && mc_model_checker->parent_snapshot) {
173 pagemap = (uint64_t*) mmalloc_no_memset((xbt_mheap_t) mc_heap, sizeof(uint64_t) * page_count);
174 mc_read_pagemap(pagemap, mc_page_number(NULL, start_addr), page_count);
177 // Take incremental snapshot:
178 new_reg->page_numbers = mc_take_page_snapshot_region(start_addr, page_count, pagemap,
179 ref_reg==NULL ? NULL : ref_reg->page_numbers);
182 mfree((xbt_mheap_t) mc_heap, pagemap);
187 void mc_region_restore_sparse(mc_mem_region_t reg, mc_mem_region_t ref_reg)
189 xbt_assert((((uintptr_t)reg->start_addr) & (xbt_pagesize-1)) == 0,
190 "Not at the beginning of a page");
191 size_t page_count = mc_page_count(reg->size);
193 uint64_t* pagemap = NULL;
195 // Read soft-dirty bits if necessary in order to know which pages have changed:
196 if (_sg_mc_soft_dirty && mc_model_checker->parent_snapshot) {
197 pagemap = (uint64_t*) mmalloc_no_memset((xbt_mheap_t) mc_heap, sizeof(uint64_t) * page_count);
198 mc_read_pagemap(pagemap, mc_page_number(NULL, reg->start_addr), page_count);
201 // Incremental per-page snapshot restoration:
202 mc_restore_page_snapshot_region(reg, page_count, pagemap, ref_reg);
204 // This is funny, the restoration can restore the state of the current heap,
205 // if this happen free(pagemap) would free from the wrong heap:
207 mfree((xbt_mheap_t) mc_heap, pagemap);