Very preliminary: remove from BattFS the LRU of free blocks; now the old blocks are...
[bertos.git] / bertos / fs / battfs.c
index aeacb9cd40b2cf294c1d9d90efc4142342eff32e..eb2523ea6d94eea818a539e4eb49e4d5f58e2cb7 100644 (file)
@@ -54,7 +54,7 @@
  */
 INLINE void battfs_to_disk(struct BattFsPageHeader *hdr, uint8_t *buf)
 {
-       STATIC_ASSERT(BATTFS_HEADER_LEN == 12);
+       STATIC_ASSERT(BATTFS_HEADER_LEN == 10);
        buf[0] = hdr->inode;
 
        buf[1] = hdr->fill;
@@ -64,33 +64,25 @@ INLINE void battfs_to_disk(struct BattFsPageHeader *hdr, uint8_t *buf)
        buf[4] = hdr->pgoff >> 8;
 
        /*
-        * Mark is at least 1 bit longer than page address.
+        * Sequence number is at least 1 bit longer than page address.
         * Needed to take care of wraparonds.
         */
-       buf[5] = hdr->mark;
-       buf[6] = hdr->mark >> 8;
+       buf[5] = hdr->seq;
+       buf[6] = hdr->seq >> 8;
 
        /*
-        * First bit used by mark, last 2 bits used by seq.
-        * Since only 2 pages with the same inode and pgoff
-        * can exist at the same time, 2 bit for seq are enough.
+        * First bit used by seq.
         * Unused bits are set to 1.
         */
-       buf[7] = ((hdr->mark >> 16) & 0x01) | (hdr->seq << 6) | ~(BV(7) | BV(6) | BV(0));
-
-       /*
-        * This field must be the before the last one!
-        */
-       buf[8] = hdr->fcs_free;
-       buf[9] = hdr->fcs_free >> 8;
+       buf[7] = (hdr->seq >> 16) ? 0xFF : 0xFE;
 
        /*
         * This field must be the last one!
         * This is needed because if the page is only partially
         * written, we can use this to detect it.
         */
-       buf[10] = hdr->fcs;
-       buf[11] = hdr->fcs >> 8;
+       buf[8] = hdr->fcs;
+       buf[9] = hdr->fcs >> 8;
 }
 
 /**
@@ -99,14 +91,12 @@ INLINE void battfs_to_disk(struct BattFsPageHeader *hdr, uint8_t *buf)
  */
 INLINE void disk_to_battfs(uint8_t *buf, struct BattFsPageHeader *hdr)
 {
-       STATIC_ASSERT(BATTFS_HEADER_LEN == 12);
+       STATIC_ASSERT(BATTFS_HEADER_LEN == 10);
        hdr->inode = buf[0];
        hdr->fill = buf[2] << 8 | buf[1];
        hdr->pgoff = buf[4] << 8 | buf[3];
-       hdr->mark = (mark_t)(buf[7] & 0x01) << 16 | buf[6] << 8 | buf[5];
-       hdr->seq = buf[7] >> 6;
-       hdr->fcs_free = buf[9] << 8 | buf[8];
-       hdr->fcs = buf[11] << 8 | buf[10];
+       hdr->seq = (seq_t)(buf[7] & 0x01) << 16 | buf[6] << 8 | buf[5];
+       hdr->fcs = buf[9] << 8 | buf[8];
 }
 
 /**
@@ -124,21 +114,6 @@ static fcs_t computeFcs(struct BattFsPageHeader *hdr)
        return cks;
 }
 
-/**
- * Compute the fcs of the header marked as free.
- */
-static fcs_t computeFcsFree(struct BattFsPageHeader *hdr)
-{
-       uint8_t buf[BATTFS_HEADER_LEN];
-       fcs_t cks;
-
-       battfs_to_disk(hdr, buf);
-       rotating_init(&cks);
-       /* fcs_free is just before fcs of whole header */
-       rotating_update(buf, BATTFS_HEADER_LEN - 2 * sizeof(fcs_t), &cks);
-       return cks;
-}
-
 
 /**
  * Read header of page \a page.
@@ -221,21 +196,17 @@ static void movePages(struct BattFsSuper *disk, pgcnt_t src, int offset)
        }
 }
 
+#if 0
+
 /**
- * Insert \a page into page allocation array of \a disk,
- * using  \a mark to compute position.
+ * Insert \a page at the bottom of page allocation array of \a disk.
  */
-static void insertFreePage(struct BattFsSuper *disk, mark_t mark, pgcnt_t page)
+static void insertFreePage(struct BattFsSuper *disk, pgcnt_t page)
 {
-       ASSERT(mark - disk->free_start < disk->free_next - disk->free_start);
-
-       pgcnt_t free_pos = disk->page_count - disk->free_next + mark;
-       ASSERT(free_pos < disk->page_count);
-
-       TRACEMSG("mark:%u, page:%u, free_start:%u, free_next:%u, free_pos:%u\n",
-               mark, page, disk->free_start, disk->free_next, free_pos);
-
+       pgcnt_t free_pos = disk->page_count - 1;
        ASSERT(disk->page_array[free_pos] == PAGE_UNSET_SENTINEL);
+       ASSERT(page <= free_pos);
+
        disk->page_array[free_pos] = page;
 }
 
@@ -350,6 +321,7 @@ static void findFreeStartNext(struct BattFsSuper *disk, mark_t minl, mark_t maxl
        TRACEMSG("Free markers:\n minl %u\n maxl %u\n minh %u\n maxh %u\n free_start %u\n free_next %u\n",
                minl, maxl, minh, maxh, disk->free_start, disk->free_next);
 }
+#endif
 
 /**
  * Count number of pages per file on \a disk.
@@ -363,14 +335,7 @@ static void findFreeStartNext(struct BattFsSuper *disk, mark_t minl, mark_t maxl
 static bool countDiskFilePages(struct BattFsSuper *disk, pgoff_t *filelen_table)
 {
        BattFsPageHeader hdr;
-       mark_t minl, maxl, minh, maxh;
-
-       /* Initialize min and max counters to keep trace od free blocks */
-       minl = MAX_PAGE_ADDR;
-       maxl = 0;
-       minh = MAX_PAGE_ADDR | MARK_HALF_SIZE;
-       maxh = 0 | MARK_HALF_SIZE;
-
+       disk->free_page_start = 0;
 
        /* Count the number of disk page per file */
        for (pgcnt_t page = 0; page < disk->page_count; page++)
@@ -384,8 +349,6 @@ static bool countDiskFilePages(struct BattFsSuper *disk, pgoff_t *filelen_table)
                /* Check header FCS */
                if (hdr.fcs == computeFcs(&hdr))
                {
-                       ASSERT(hdr.mark == MARK_PAGE_VALID);
-                       ASSERT(hdr.fcs_free == FCS_FREE_VALID);
                        ASSERT(hdr.fill <= disk->page_size - BATTFS_HEADER_LEN);
 
                        /* Page is valid and is owned by a file */
@@ -393,32 +356,10 @@ static bool countDiskFilePages(struct BattFsSuper *disk, pgoff_t *filelen_table)
 
                        /* Keep trace of free space */
                        disk->free_bytes -= hdr.fill;
-               }
-               else
-               {
-                       /* Check if page is marked free */
-                       if (hdr.fcs_free == computeFcsFree(&hdr))
-                       {
-                               /*
-                                * This page is a valid and marked free page.
-                                * Update min and max free page markers.
-                                */
-                               if (hdr.mark < MARK_HALF_SIZE)
-                               {
-                                       minl = MIN(minl, hdr.mark);
-                                       maxl = MAX(maxl, hdr.mark);
-                               }
-                               else
-                               {
-                                       minh = MIN(minh, hdr.mark);
-                                       maxh = MAX(maxh, hdr.mark);
-                               }
-                       }
-                       else
-                               TRACEMSG("page [%d] invalid, keeping as free\n", page);
+                       disk->free_page_start++;
                }
        }
-       findFreeStartNext(disk, minl, maxl, minh, maxh);
+
        return true;
 }
 
@@ -440,6 +381,7 @@ static bool countDiskFilePages(struct BattFsSuper *disk, pgoff_t *filelen_table)
 static bool fillPageArray(struct BattFsSuper *disk, pgoff_t *filelen_table)
 {
        BattFsPageHeader hdr;
+       pgcnt_t curr_free_page = disk->free_page_start;
        /* Fill page allocation array */
        for (pgcnt_t page = 0; page < disk->page_count; page++)
        {
@@ -449,87 +391,246 @@ static bool fillPageArray(struct BattFsSuper *disk, pgoff_t *filelen_table)
                /* Check header FCS */
                if (hdr.fcs == computeFcs(&hdr))
                {
-                       /* Page is valid and is owned by a file */
-                       ASSERT(hdr.mark == MARK_PAGE_VALID);
-                       ASSERT(hdr.fcs_free == FCS_FREE_VALID);
-
                        /* Compute array position */
-                       pgcnt_t array_pos = countPages(filelen_table, hdr.inode);
-                       array_pos += hdr.pgoff;
+                       pgcnt_t array_pos_start = countPages(filelen_table, hdr.inode);
+                       pgcnt_t array_pos = array_pos_start + hdr.pgoff;
+
+                       /* Find the first free position */
+                       while (disk->page_array[array_pos] != PAGE_UNSET_SENTINEL)
+                       {
+                               ASSERT(array_pos < array_pos_start + filelen_table[hdr.inode + 1]);
+                               array_pos++;
+                       }
+
+                       disk->page_array[array_pos] = page;
+               }
+               else
+               {
+                       /* Invalid page, keep as free */
+                       ASSERT(disk->page_array[curr_free_page] == PAGE_UNSET_SENTINEL);
+                       LOG_INFO("Page %d invalid, keeping as free\n", page);
+                       disk->page_array[curr_free_page++] = page;
+               }
+       }
+       return true;
+}
+
+/**
+ * Find the latest version of a page, starting from the
+ * page supplied by \a page_array.
+ * The pages are read from the disk until a different
+ * inode or page offset is found.
+ * The lastest version of the page is moved in the first
+ * position of \a page_array.
+ * \return the number of old versions of the page or PAGE_ERROR
+ *         on disk read errors.
+ */
+static pgcnt_t findLastVersion(pgcnt_t *page_array)
+{
+       pgcnt_t *array_start = page_array;
+       BattFsPageHeader hdr;
+       if (!battfs_readHeader(disk, *page_array++, &hdr))
+               return PAGE_ERROR;
+
+       /* Free space: early bailout */
+       if (hdr.fcs != computeFcs(&hdr))
+               return 0;
+
+       /*
+        * If the first page is valid,
+        * inode and pg_off in the array are taken
+        * as the current page markers.
+        */
+       inode_t curr_inode = hdr.inode;
+       pgoff_t curr_pgoff = hdr.pgoff;
+
+       /* Temps used to find the sequence number range */
+       seq_t minl = HALF_SEQ - 1;
+       seq_t maxl = 0;
+       seq_t minh = FULL_SEQ;
+       seq_t maxh = HALF_SEQ;
+       pgcnt_t lpos = 0, hpos = 0, dup_cnt = 0;
+
+       /*
+        * Find min and max values for the two
+        * half of seq_num range.
+        * With this we can find seqnum wraparounds.
+        * seq_t is a type that has at least 1 bit more than
+        * pgaddr_t. So all version of a page blocks can be numbered using
+        * at most half numbers of a seq_t type.
+        * The sequence number algorithm increments by 1 the previous seq_num
+        * every time a page is rewritten. So the sequence is
+        * guaranteed to be countiguous.
+        * Only wrap arounds may happen, but due to half size sequence limitation,
+        * there are only 4 possible situations:
+        *
+        * \verbatim
+        *    |------lower half------|-------upper half-------|
+        *
+        * 1) |------minl*****maxl---|------------------------|
+        * 2) |------minl********maxl|minh******maxh----------|
+        * 3) |----------------------|----minh*******maxh-----|
+        * 4) |minl******maxl--------|------------minh****maxh|
+        * \endverbatim
+        *
+        * Situations 1 and 3 are easy to detect, while 2 and 4 require more care.
+        */
+       do
+       {
+               if (hdr.seq < SEQ_HALF_SIZE)
+               {
+                       minl = MIN(minl, hdr.seq);
+                       if (hdr.seq > maxl)
+                       {
+                               maxl = hdr.seq;
+                               lpos = dup_cnt;
+                       }
+               }
+               else
+               {
+                       minh = MIN(minh, hdr.seq);
+                       if (hdr.seq > maxh)
+                       {
+                               maxh = hdr.seq;
+                               hpos = dup_cnt;
+                       }
+               }
+
+               if (!battfs_readHeader(disk, *page_array++, &hdr))
+                       return PAGE_ERROR;
+               dup_cnt++;
+       }
+       while (curr_inode == hdr.inode && curr_pgoff == hdr.pgoff && hdr.fcs == computeFcs(&hdr))
+
+
+       /* Return early if there is only one version of the current page */
+       if (dup_cnt == 1)
+               return 0;
+
+       /* Find the position in the array of the last version of the page */
+       pgcnt_t last_ver = hpos;
+       if (maxl >= minl)
+       {
+               /* Valid interval found in lower half */
+               if (maxh >= minh)
+               {
+                       /* Valid interval also found in upper half */
+                       if (maxl != minh - 1)
+                       {
+                               /* Interval starts in upper half and ends in lower */
+                               ASSERT(minl == 0);
+                               ASSERT(maxh == FULL_SEQ);
+
+                               last_ver = lpos;
+                       }
+               }
+               else
+                       /*
+                        * Upper interval is invalid.
+                        * Use lower values.
+                        */
+                       last_ver = lpos;
+       }
+
+       /* Put last page version at array start position */
+       SWAP(array_start[0], array_start[last_ver]);
+
+       return dup_cnt - 1;
+}
 
-                       /* Check if position is already used by another page of the same file */
-                       if (LIKELY(disk->page_array[array_pos] == PAGE_UNSET_SENTINEL))
-                               disk->page_array[array_pos] = page;
+/**
+ * Collect old pages, removing empty spaces from \a pg_array, for a maximum len of \a pg_len.
+ * Once the collect task is completed, copy \a old_cnt pages from \a old_pages at the
+ * end of free space in pg_array.
+ */
+void collectOldPages(pgcnt_t *pg_array, pgcnt_t pg_len, pgcnt_t *old_pages, pgcnt_t old_cnt)
+{
+       bool copy = false;
+       pgcnt_t gap = 0;
+
+       for (pgcnt_t curr_page = 0; curr_page < pg_len; pg_len++)
+       {
+               if (!copy)
+               {
+                       if (pg_array[curr_page] == PAGE_UNSET_SENTINEL)
+                               gap++;
                        else
                        {
-                               BattFsPageHeader hdr_old;
-
-                               if (!battfs_readHeader(disk, disk->page_array[array_pos], &hdr_old))
-                                       return false;
-
-                               /* Check header FCS */
-                               ASSERT(hdr_old.fcs == computeFcs(&hdr_old));
-
-                               /* Only the very same page with a different seq number can be here */
-                               ASSERT(hdr.inode == hdr_old.inode);
-                               ASSERT(hdr.pgoff == hdr_old.pgoff);
-                               ASSERT(hdr.mark == hdr_old.mark);
-                               ASSERT(hdr.fcs_free == hdr_old.fcs_free);
-                               ASSERT(hdr.seq != hdr_old.seq);
-
-                               pgcnt_t new_page, old_page;
-                               fill_t old_fill;
-
-                               /* Fancy check to handle seq wraparound (2 bits only) */
-                               if (((hdr.seq - hdr_old.seq) & 0x03) < 2)
-                               {
-                                       /* Current header is newer than the previuos one */
-                                       old_page = disk->page_array[array_pos];
-                                       new_page = page;
-                                       old_fill = hdr_old.fill;
-                               }
-                               else
-                               {
-                                       /* Previous header is newer than the current one */
-                                       old_page = page;
-                                       new_page = disk->page_array[array_pos];
-                                       old_fill = hdr.fill;
-                               }
-
-                               /* Set new page */
-                               disk->page_array[array_pos] = new_page;
-
-                               /* Add free space */
-                               disk->free_bytes += old_fill;
-
-                               /* Shift all array one position to the left, overwriting duplicate page */
-                               array_pos -= hdr.pgoff;
-                               array_pos += filelen_table[hdr.inode];
-                               movePages(disk, array_pos, -1);
-
-                               /* Decrease file page count */
-                               filelen_table[hdr.inode]--;
-
-                               /* Add old page to free pages pool */
-                               if (!battfs_markFree(disk, &hdr, old_page))
-                                       return false;
-
-                               insertFreePage(disk, hdr.mark, old_page);
+                               pg_array[curr_page - gap] = pg_array[curr_page];
+                               copy = true;
                        }
                }
                else
                {
-                       /* Check if page is free */
-                       if (hdr.fcs_free != computeFcsFree(&hdr))
-                               /* Page is not a valid marked page, insert at list beginning */
-                               hdr.mark = --disk->free_start;
+                       if (pg_array[curr_page] != PAGE_UNSET_SENTINEL)
+                               pg_array[curr_page - gap] = pg_array[curr_page];
+                       else
+                       {
+                               gap++;
+                               copy = false;
+                       }
+               }
+       }
+       ASSERT(gap == old_cnt);
+       pg_array += pg_len - old_cnt;
 
-                       insertFreePage(disk, hdr.mark, page);
+       memcpy(pg_array, old_pages, old_cnt * sizeof(pgcnt_t));
+}
+
+/**
+ * This function scan the page array of \a disk looking for
+ * old versions of the same page.
+ *
+ * Only the last version is kept as valid, the old ones are inserted
+ * in the free blocks heap.
+ * \return true if ok, false on disk read errors.
+ * \note The whole disk is scanned once.
+ */
+static bool dropOldPages(struct BattFsSuper *disk)
+{
+       #define OLD_PAGE_BUFLEN 64
+       pgcnt_t old_pages[OLD_PAGE_BUFLEN];
+       pgcnt_t old_cnt = 0;
+
+       pgcnt_t *curr_page = disk->page_array;
+       pgcnt_t *collect_start = disk->page_array;
+       pgcnt_t collect_len = disk->page_count;
+       pgcnt_t dup_pages;
+
+       do
+       {
+               dup_pages = findLastVersion(curr_page);
+               if (dup_pages == PAGE_ERROR)
+                       return false;
+               /* The first page is the last version */
+               curr_page++;
+               while (dup_pages--)
+               {
+                       if (old_cnt >= OLD_PAGE_BUFLEN)
+                       {
+                               collectOldPages(collect_start, collect_len, old_pages, old_cnt);
+                               collect_len -= old_cnt;
+                               disk->free_bytes += old_cnt * (disk->page_size - BATTFS_HEADER_LEN);
+                               disk->free_page_start -= old_cnt;
+                               curr_page -= old_cnt;
+                               collect_start = curr_page;
+                               old_cnt = 0;
+                       }
+
+                       old_pages[old_cnt++] = *curr_page;
+                       *curr_page++ = PAGE_UNSET_SENTINEL;
                }
        }
+       while (curr_page < disk->page_array + disk->free_page_start);
+
+       collectOldPages(collect_start, collect_len, old_pages, old_cnt);
+       disk->free_bytes += old_cnt * (disk->page_size - BATTFS_HEADER_LEN);
+       disk->free_page_start -= old_cnt;
+
        return true;
 }
 
+
 /**
  * Initialize and mount disk described by
  * \a disk.
@@ -584,6 +685,12 @@ bool battfs_init(struct BattFsSuper *disk)
                return false;
        }
 
+       if (!dropOldPages(disk))
+       {
+               LOG_ERR("error dropping old pages\n");
+               return false;
+       }
+
        /* Init list for opened files. */
        LIST_INIT(&disk->file_opened_list);
        return true;