xbt_assert(mc_page_offset(page)==0, "Not at the beginning of a page");
xbt_assert(top_index_ <= this->capacity_, "top_index is not consistent");
- // Search the page in the snapshot pages:
+ // First, we check if a page with the same content is already in the page
+ // store:
+ // 1. compute the hash of the page;
+ // 2. find pages with the same hash using `hash_index_`;
+ // 3. find a page with the same content.
uint64_t hash = mc_hash_page(page);
page_set_type& page_set = this->hash_index_[hash];
BOOST_FOREACH (size_t pageno, page_set) {
const void* snapshot_page = this->get_page(pageno);
if (memcmp(page, snapshot_page, xbt_pagesize) == 0) {
- // Page found, reuse it:
+
+ // If a page with the same content is already in the page store it is
+ // reused and its reference count is incremented.
page_counts_[pageno]++;
return pageno;
+
}
}
- // Allocate a new page for this page:
+ // Otherwise, a new page is allocated in the page store and the content
+ // of the page is `memcpy()`-ed to this new page.
size_t pageno = alloc_page();
void* snapshot_page = (void*) this->get_page(pageno);
memcpy(snapshot_page, page, xbt_pagesize);
#ifndef MC_PAGE_SNAPSHOT_H
#define MC_PAGE_SNAPSHOT_H
-/** @brief Manager for snapshot pages
+/** @brief Storage for snapshot memory pages
+ *
+ * The first (lower) layer of the per-page snapshot mechanism is a page
+ * store: it's responsibility is to store immutable shareable
+ * reference-counted memory pages independently of the snapshoting
+ * logic. Snapshot management and representation, soft-dirty tracking is
+ * handled to an higher layer. READMORE
+ *
+ * Data structure:
+ *
+ * * A pointer (`memory_`) to a (currently anonymous) `mmap()`ed memory
+ * region holding the memory pages (the address of the first page).
+ *
+ * We want to keep this memory region aligned on the memory pages (so
+ * that we might be able to create non-linear memory mappings on those
+ * pages in the future) and be able to expand it without coyping the
+ * data (there will be a lot of pages here): we will be able to
+ * efficiently expand the memory mapping using `mremap()`, moving it
+ * to another virtual address if necessary.
+ *
+ * Because we will move this memory mapping on the virtual address
+ * space, only the index of the page will be stored in the snapshots
+ * and the page will always be looked up by going through `memory`:
+ *
+ * void* page = (char*) page_store->memory + page_index << pagebits;
+ *
+ * * The number of pages mapped in virtual memory (`capacity_`). Once all
+ * those pages are used, we need to expand the page store with
+ * `mremap()`.
+ *
+ * * A reference count for each memory page `page_counts_`. Each time a
+ * snapshot references a page, the counter is incremented. If a
+ * snapshot is freed, the reference count is decremented. When the
+ * reference count, of a page reaches 0 it is added to a list of available
+ * pages (`free_pages_`).
+ *
+ * * A list of free pages `free_pages_` which can be reused. This avoids having
+ * to scan the reference count list to find a free page.
+ *
+ * * When we are expanding the memory map we do not want to add thousand of page
+ * to the `free_pages_` list and remove them just afterwards. The `top_index_`
+ * field is an index after which all pages are free and are not in the `free_pages_`
+ * list.
+ *
+ * * When we are adding a page, we need to check if a page with the same
+ * content is already in the page store in order to reuse it. For this
+ * reason, we maintain an index (`hash_index_`) mapping the hash of a
+ * page to the list of page indices with this hash.
+ * We use a fast (non cryptographic) hash so there may be conflicts:
+ * we must be able to store multiple indices for the same hash.
*
- * Page management: the free pages are stored as a simple linked-list.
- * The number of the first page if stored in `free_pages`.
- * Each free page store the number of the next free page.
*/
struct s_mc_pages_store {
private: // Types
public: // Methods
- /** @brief Decrement the refcount for a given page */
+ /** @brief Decrement the reference count for a given page
+ *
+ * Decrement the reference count of this page. Used when a snapshot is
+ * destroyed.
+ *
+ * If the reference count reaches zero, the page is recycled:
+ * it is added to the `free_pages_` list and removed from the `hash_index_`.
+ *
+ * */
void unref_page(size_t pageno) {
if ((--this->page_counts_[pageno]) == 0) {
this->remove_page(pageno);
}
}
- /** @brief Increment the refcount for a given page */
+ /** @brief Increment the refcount for a given page
+ *
+ * This method used to increase a reference count of a page if we know
+ * that the content of a page is the same as a page already in the page
+ * store.
+ *
+ * This will be the case if a page if soft clean: we know that is has not
+ * changed since the previous cnapshot/restoration and we can avoid
+ * hashing the page, comparing byte-per-byte to candidates.
+ * */
void ref_page(size_t pageno) {
++this->page_counts_[pageno];
}
- /** @brief Store a page
- *
- * Either allocate a new page in the store or reuse
- * a shared page if is is already in the page store.
- */
+ /** @brief Store a page in the page store */
size_t store_page(void* page);
/** @brief Get a page from its page number