* Use is subject to license terms.
*/
/*
- * Copyright (c) 2012, 2016 by Delphix. All rights reserved.
+ * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
*/
#include <sys/zfs_context.h>
#include <sys/zfeature.h>
/*
+ * Note on space map block size:
+ *
* The data for a given space map can be kept on blocks of any size.
- * Larger blocks entail fewer i/o operations, but they also cause the
- * DMU to keep more data in-core, and also to waste more i/o bandwidth
+ * Larger blocks entail fewer I/O operations, but they also cause the
+ * DMU to keep more data in-core, and also to waste more I/O bandwidth
* when only a few blocks have changed since the last transaction group.
*/
-int space_map_blksz = (1 << 12);
+
+/*
+ * Enabled whenever we want to stress test the use of double-word
+ * space map entries.
+ */
+boolean_t zfs_force_some_double_word_sm_entries = B_FALSE;
+
+/*
+ * Override the default indirect block size of 128K, instead use 16K for
+ * spacemaps (2^14 bytes). This dramatically reduces write inflation since
+ * appending to a spacemap typically has to write one data block (4KB) and one
+ * or two indirect blocks (16K-32K, rather than 128K).
+ */
+int space_map_ibs = 14;
+
+boolean_t
+sm_entry_is_debug(uint64_t e)
+{
+ return (SM_PREFIX_DECODE(e) == SM_DEBUG_PREFIX);
+}
+
+boolean_t
+sm_entry_is_single_word(uint64_t e)
+{
+ uint8_t prefix = SM_PREFIX_DECODE(e);
+ return (prefix != SM_DEBUG_PREFIX && prefix != SM2_PREFIX);
+}
+
+boolean_t
+sm_entry_is_double_word(uint64_t e)
+{
+ return (SM_PREFIX_DECODE(e) == SM2_PREFIX);
+}
/*
* Iterate through the space map, invoking the callback on each (non-debug)
- * space map entry.
+ * space map entry. Stop after reading 'end' bytes of the space map.
*/
int
-space_map_iterate(space_map_t *sm, sm_cb_t callback, void *arg)
+space_map_iterate(space_map_t *sm, uint64_t end, sm_cb_t callback, void *arg)
{
- uint64_t *entry, *entry_map, *entry_map_end;
- uint64_t bufsize, size, offset, end;
+ uint64_t blksz = sm->sm_blksz;
+
+ ASSERT3U(blksz, !=, 0);
+ ASSERT3U(end, <=, space_map_length(sm));
+ ASSERT0(P2PHASE(end, sizeof (uint64_t)));
+
+ dmu_prefetch(sm->sm_os, space_map_object(sm), 0, 0, end,
+ ZIO_PRIORITY_SYNC_READ);
+
int error = 0;
+ for (uint64_t block_base = 0; block_base < end && error == 0;
+ block_base += blksz) {
+ dmu_buf_t *db;
+ error = dmu_buf_hold(sm->sm_os, space_map_object(sm),
+ block_base, FTAG, &db, DMU_READ_PREFETCH);
+ if (error != 0)
+ return (error);
+
+ uint64_t *block_start = db->db_data;
+ uint64_t block_length = MIN(end - block_base, blksz);
+ uint64_t *block_end = block_start +
+ (block_length / sizeof (uint64_t));
+
+ VERIFY0(P2PHASE(block_length, sizeof (uint64_t)));
+ VERIFY3U(block_length, !=, 0);
+ ASSERT3U(blksz, ==, db->db_size);
+
+ for (uint64_t *block_cursor = block_start;
+ block_cursor < block_end && error == 0; block_cursor++) {
+ uint64_t e = *block_cursor;
+
+ if (sm_entry_is_debug(e)) /* Skip debug entries */
+ continue;
+
+ uint64_t raw_offset, raw_run, vdev_id;
+ maptype_t type;
+ if (sm_entry_is_single_word(e)) {
+ type = SM_TYPE_DECODE(e);
+ vdev_id = SM_NO_VDEVID;
+ raw_offset = SM_OFFSET_DECODE(e);
+ raw_run = SM_RUN_DECODE(e);
+ } else {
+ /* it is a two-word entry */
+ ASSERT(sm_entry_is_double_word(e));
+ raw_run = SM2_RUN_DECODE(e);
+ vdev_id = SM2_VDEV_DECODE(e);
+
+ /* move on to the second word */
+ block_cursor++;
+ e = *block_cursor;
+ VERIFY3P(block_cursor, <=, block_end);
+
+ type = SM2_TYPE_DECODE(e);
+ raw_offset = SM2_OFFSET_DECODE(e);
+ }
- end = space_map_length(sm);
+ uint64_t entry_offset = (raw_offset << sm->sm_shift) +
+ sm->sm_start;
+ uint64_t entry_run = raw_run << sm->sm_shift;
+
+ VERIFY0(P2PHASE(entry_offset, 1ULL << sm->sm_shift));
+ VERIFY0(P2PHASE(entry_run, 1ULL << sm->sm_shift));
+ ASSERT3U(entry_offset, >=, sm->sm_start);
+ ASSERT3U(entry_offset, <, sm->sm_start + sm->sm_size);
+ ASSERT3U(entry_run, <=, sm->sm_size);
+ ASSERT3U(entry_offset + entry_run, <=,
+ sm->sm_start + sm->sm_size);
+
+ space_map_entry_t sme = {
+ .sme_type = type,
+ .sme_vdev = vdev_id,
+ .sme_offset = entry_offset,
+ .sme_run = entry_run
+ };
+ error = callback(&sme, arg);
+ }
+ dmu_buf_rele(db, FTAG);
+ }
+ return (error);
+}
- bufsize = MAX(sm->sm_blksz, SPA_MINBLOCKSIZE);
- entry_map = vmem_alloc(bufsize, KM_SLEEP);
+/*
+ * Reads the entries from the last block of the space map into
+ * buf in reverse order. Populates nwords with number of words
+ * in the last block.
+ *
+ * Refer to block comment within space_map_incremental_destroy()
+ * to understand why this function is needed.
+ */
+static int
+space_map_reversed_last_block_entries(space_map_t *sm, uint64_t *buf,
+ uint64_t bufsz, uint64_t *nwords)
+{
+ int error = 0;
+ dmu_buf_t *db;
- if (end > bufsize) {
- dmu_prefetch(sm->sm_os, space_map_object(sm), 0, bufsize,
- end - bufsize, ZIO_PRIORITY_SYNC_READ);
+ /*
+ * Find the offset of the last word in the space map and use
+ * that to read the last block of the space map with
+ * dmu_buf_hold().
+ */
+ uint64_t last_word_offset =
+ sm->sm_phys->smp_length - sizeof (uint64_t);
+ error = dmu_buf_hold(sm->sm_os, space_map_object(sm), last_word_offset,
+ FTAG, &db, DMU_READ_NO_PREFETCH);
+ if (error != 0)
+ return (error);
+
+ ASSERT3U(sm->sm_object, ==, db->db_object);
+ ASSERT3U(sm->sm_blksz, ==, db->db_size);
+ ASSERT3U(bufsz, >=, db->db_size);
+ ASSERT(nwords != NULL);
+
+ uint64_t *words = db->db_data;
+ *nwords =
+ (sm->sm_phys->smp_length - db->db_offset) / sizeof (uint64_t);
+
+ ASSERT3U(*nwords, <=, bufsz / sizeof (uint64_t));
+
+ uint64_t n = *nwords;
+ uint64_t j = n - 1;
+ for (uint64_t i = 0; i < n; i++) {
+ uint64_t entry = words[i];
+ if (sm_entry_is_double_word(entry)) {
+ /*
+ * Since we are populating the buffer backwards
+ * we have to be extra careful and add the two
+ * words of the double-word entry in the right
+ * order.
+ */
+ ASSERT3U(j, >, 0);
+ buf[j - 1] = entry;
+
+ i++;
+ ASSERT3U(i, <, n);
+ entry = words[i];
+ buf[j] = entry;
+ j -= 2;
+ } else {
+ ASSERT(sm_entry_is_debug(entry) ||
+ sm_entry_is_single_word(entry));
+ buf[j] = entry;
+ j--;
+ }
}
- for (offset = 0; offset < end && error == 0; offset += bufsize) {
- size = MIN(end - offset, bufsize);
- VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
- VERIFY(size != 0);
- ASSERT3U(sm->sm_blksz, !=, 0);
+ /*
+ * Assert that we wrote backwards all the
+ * way to the beginning of the buffer.
+ */
+ ASSERT3S(j, ==, -1);
+
+ dmu_buf_rele(db, FTAG);
+ return (error);
+}
+
+/*
+ * Note: This function performs destructive actions - specifically
+ * it deletes entries from the end of the space map. Thus, callers
+ * should ensure that they are holding the appropriate locks for
+ * the space map that they provide.
+ */
+int
+space_map_incremental_destroy(space_map_t *sm, sm_cb_t callback, void *arg,
+ dmu_tx_t *tx)
+{
+ uint64_t bufsz = MAX(sm->sm_blksz, SPA_MINBLOCKSIZE);
+ uint64_t *buf = zio_buf_alloc(bufsz);
- dprintf("object=%llu offset=%llx size=%llx\n",
- space_map_object(sm), offset, size);
+ dmu_buf_will_dirty(sm->sm_dbuf, tx);
- error = dmu_read(sm->sm_os, space_map_object(sm), offset, size,
- entry_map, DMU_READ_PREFETCH);
+ /*
+ * Ideally we would want to iterate from the beginning of the
+ * space map to the end in incremental steps. The issue with this
+ * approach is that we don't have any field on-disk that points
+ * us where to start between each step. We could try zeroing out
+ * entries that we've destroyed, but this doesn't work either as
+ * an entry that is 0 is a valid one (ALLOC for range [0x0:0x200]).
+ *
+ * As a result, we destroy its entries incrementally starting from
+ * the end after applying the callback to each of them.
+ *
+ * The problem with this approach is that we cannot literally
+ * iterate through the words in the space map backwards as we
+ * can't distinguish two-word space map entries from their second
+ * word. Thus we do the following:
+ *
+ * 1] We get all the entries from the last block of the space map
+ * and put them into a buffer in reverse order. This way the
+ * last entry comes first in the buffer, the second to last is
+ * second, etc.
+ * 2] We iterate through the entries in the buffer and we apply
+ * the callback to each one. As we move from entry to entry we
+ * we decrease the size of the space map, deleting effectively
+ * each entry.
+ * 3] If there are no more entries in the space map or the callback
+ * returns a value other than 0, we stop iterating over the
+ * space map. If there are entries remaining and the callback
+ * returned 0, we go back to step [1].
+ */
+ int error = 0;
+ while (space_map_length(sm) > 0 && error == 0) {
+ uint64_t nwords = 0;
+ error = space_map_reversed_last_block_entries(sm, buf, bufsz,
+ &nwords);
if (error != 0)
break;
- entry_map_end = entry_map + (size / sizeof (uint64_t));
- for (entry = entry_map; entry < entry_map_end && error == 0;
- entry++) {
- uint64_t e = *entry;
- uint64_t offset, size;
+ ASSERT3U(nwords, <=, bufsz / sizeof (uint64_t));
- if (SM_DEBUG_DECODE(e)) /* Skip debug entries */
+ for (uint64_t i = 0; i < nwords; i++) {
+ uint64_t e = buf[i];
+
+ if (sm_entry_is_debug(e)) {
+ sm->sm_phys->smp_length -= sizeof (uint64_t);
continue;
+ }
- offset = (SM_OFFSET_DECODE(e) << sm->sm_shift) +
- sm->sm_start;
- size = SM_RUN_DECODE(e) << sm->sm_shift;
+ int words = 1;
+ uint64_t raw_offset, raw_run, vdev_id;
+ maptype_t type;
+ if (sm_entry_is_single_word(e)) {
+ type = SM_TYPE_DECODE(e);
+ vdev_id = SM_NO_VDEVID;
+ raw_offset = SM_OFFSET_DECODE(e);
+ raw_run = SM_RUN_DECODE(e);
+ } else {
+ ASSERT(sm_entry_is_double_word(e));
+ words = 2;
+
+ raw_run = SM2_RUN_DECODE(e);
+ vdev_id = SM2_VDEV_DECODE(e);
+
+ /* move to the second word */
+ i++;
+ e = buf[i];
+
+ ASSERT3P(i, <=, nwords);
+
+ type = SM2_TYPE_DECODE(e);
+ raw_offset = SM2_OFFSET_DECODE(e);
+ }
- VERIFY0(P2PHASE(offset, 1ULL << sm->sm_shift));
- VERIFY0(P2PHASE(size, 1ULL << sm->sm_shift));
- VERIFY3U(offset, >=, sm->sm_start);
- VERIFY3U(offset + size, <=, sm->sm_start + sm->sm_size);
- error = callback(SM_TYPE_DECODE(e), offset, size, arg);
+ uint64_t entry_offset =
+ (raw_offset << sm->sm_shift) + sm->sm_start;
+ uint64_t entry_run = raw_run << sm->sm_shift;
+
+ VERIFY0(P2PHASE(entry_offset, 1ULL << sm->sm_shift));
+ VERIFY0(P2PHASE(entry_run, 1ULL << sm->sm_shift));
+ VERIFY3U(entry_offset, >=, sm->sm_start);
+ VERIFY3U(entry_offset, <, sm->sm_start + sm->sm_size);
+ VERIFY3U(entry_run, <=, sm->sm_size);
+ VERIFY3U(entry_offset + entry_run, <=,
+ sm->sm_start + sm->sm_size);
+
+ space_map_entry_t sme = {
+ .sme_type = type,
+ .sme_vdev = vdev_id,
+ .sme_offset = entry_offset,
+ .sme_run = entry_run
+ };
+ error = callback(&sme, arg);
+ if (error != 0)
+ break;
+
+ if (type == SM_ALLOC)
+ sm->sm_phys->smp_alloc -= entry_run;
+ else
+ sm->sm_phys->smp_alloc += entry_run;
+ sm->sm_phys->smp_length -= words * sizeof (uint64_t);
}
}
- vmem_free(entry_map, bufsize);
+ if (space_map_length(sm) == 0) {
+ ASSERT0(error);
+ ASSERT0(space_map_allocated(sm));
+ }
+
+ zio_buf_free(buf, bufsz);
return (error);
}
} space_map_load_arg_t;
static int
-space_map_load_callback(maptype_t type, uint64_t offset, uint64_t size,
- void *arg)
+space_map_load_callback(space_map_entry_t *sme, void *arg)
{
space_map_load_arg_t *smla = arg;
- if (type == smla->smla_type) {
- VERIFY3U(range_tree_space(smla->smla_rt) + size, <=,
+ if (sme->sme_type == smla->smla_type) {
+ VERIFY3U(range_tree_space(smla->smla_rt) + sme->sme_run, <=,
smla->smla_sm->sm_size);
- range_tree_add(smla->smla_rt, offset, size);
+ range_tree_add(smla->smla_rt, sme->sme_offset, sme->sme_run);
} else {
- range_tree_remove(smla->smla_rt, offset, size);
+ range_tree_remove(smla->smla_rt, sme->sme_offset, sme->sme_run);
}
return (0);
}
/*
- * Load the space map disk into the specified range tree. Segments of maptype
- * are added to the range tree, other segment types are removed.
+ * Load the spacemap into the rangetree, like space_map_load. But only
+ * read the first 'length' bytes of the spacemap.
*/
int
-space_map_load(space_map_t *sm, range_tree_t *rt, maptype_t maptype)
+space_map_load_length(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
+ uint64_t length)
{
- uint64_t space;
- int err;
space_map_load_arg_t smla;
VERIFY0(range_tree_space(rt));
- space = space_map_allocated(sm);
- if (maptype == SM_FREE) {
+ if (maptype == SM_FREE)
range_tree_add(rt, sm->sm_start, sm->sm_size);
- space = sm->sm_size - space;
- }
smla.smla_rt = rt;
smla.smla_sm = sm;
smla.smla_type = maptype;
- err = space_map_iterate(sm, space_map_load_callback, &smla);
+ int err = space_map_iterate(sm, length,
+ space_map_load_callback, &smla);
- if (err == 0) {
- VERIFY3U(range_tree_space(rt), ==, space);
- } else {
+ if (err != 0)
range_tree_vacate(rt, NULL, NULL);
- }
return (err);
}
+/*
+ * Load the space map disk into the specified range tree. Segments of maptype
+ * are added to the range tree, other segment types are removed.
+ */
+int
+space_map_load(space_map_t *sm, range_tree_t *rt, maptype_t maptype)
+{
+ return (space_map_load_length(sm, rt, maptype, space_map_length(sm)));
+}
+
void
space_map_histogram_clear(space_map_t *sm)
{
}
}
-uint64_t
-space_map_entries(space_map_t *sm, range_tree_t *rt)
+static void
+space_map_write_intro_debug(space_map_t *sm, maptype_t maptype, dmu_tx_t *tx)
{
- avl_tree_t *t = &rt->rt_root;
- range_seg_t *rs;
- uint64_t size, entries;
+ dmu_buf_will_dirty(sm->sm_dbuf, tx);
+
+ uint64_t dentry = SM_PREFIX_ENCODE(SM_DEBUG_PREFIX) |
+ SM_DEBUG_ACTION_ENCODE(maptype) |
+ SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(tx->tx_pool->dp_spa)) |
+ SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
+
+ dmu_write(sm->sm_os, space_map_object(sm), sm->sm_phys->smp_length,
+ sizeof (dentry), &dentry, tx);
+
+ sm->sm_phys->smp_length += sizeof (dentry);
+}
+
+/*
+ * Writes one or more entries given a segment.
+ *
+ * Note: The function may release the dbuf from the pointer initially
+ * passed to it, and return a different dbuf. Also, the space map's
+ * dbuf must be dirty for the changes in sm_phys to take effect.
+ */
+static void
+space_map_write_seg(space_map_t *sm, range_seg_t *rs, maptype_t maptype,
+ uint64_t vdev_id, uint8_t words, dmu_buf_t **dbp, void *tag, dmu_tx_t *tx)
+{
+ ASSERT3U(words, !=, 0);
+ ASSERT3U(words, <=, 2);
+
+ /* ensure the vdev_id can be represented by the space map */
+ ASSERT3U(vdev_id, <=, SM_NO_VDEVID);
+
+ /*
+ * if this is a single word entry, ensure that no vdev was
+ * specified.
+ */
+ IMPLY(words == 1, vdev_id == SM_NO_VDEVID);
+
+ dmu_buf_t *db = *dbp;
+ ASSERT3U(db->db_size, ==, sm->sm_blksz);
+
+ uint64_t *block_base = db->db_data;
+ uint64_t *block_end = block_base + (sm->sm_blksz / sizeof (uint64_t));
+ uint64_t *block_cursor = block_base +
+ (sm->sm_phys->smp_length - db->db_offset) / sizeof (uint64_t);
+
+ ASSERT3P(block_cursor, <=, block_end);
+
+ uint64_t size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
+ uint64_t start = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
+ uint64_t run_max = (words == 2) ? SM2_RUN_MAX : SM_RUN_MAX;
+
+ ASSERT3U(rs->rs_start, >=, sm->sm_start);
+ ASSERT3U(rs->rs_start, <, sm->sm_start + sm->sm_size);
+ ASSERT3U(rs->rs_end - rs->rs_start, <=, sm->sm_size);
+ ASSERT3U(rs->rs_end, <=, sm->sm_start + sm->sm_size);
+
+ while (size != 0) {
+ ASSERT3P(block_cursor, <=, block_end);
+
+ /*
+ * If we are at the end of this block, flush it and start
+ * writing again from the beginning.
+ */
+ if (block_cursor == block_end) {
+ dmu_buf_rele(db, tag);
+
+ uint64_t next_word_offset = sm->sm_phys->smp_length;
+ VERIFY0(dmu_buf_hold(sm->sm_os,
+ space_map_object(sm), next_word_offset,
+ tag, &db, DMU_READ_PREFETCH));
+ dmu_buf_will_dirty(db, tx);
+
+ /* update caller's dbuf */
+ *dbp = db;
+
+ ASSERT3U(db->db_size, ==, sm->sm_blksz);
+ block_base = db->db_data;
+ block_cursor = block_base;
+ block_end = block_base +
+ (db->db_size / sizeof (uint64_t));
+ }
+
+ /*
+ * If we are writing a two-word entry and we only have one
+ * word left on this block, just pad it with an empty debug
+ * entry and write the two-word entry in the next block.
+ */
+ uint64_t *next_entry = block_cursor + 1;
+ if (next_entry == block_end && words > 1) {
+ ASSERT3U(words, ==, 2);
+ *block_cursor = SM_PREFIX_ENCODE(SM_DEBUG_PREFIX) |
+ SM_DEBUG_ACTION_ENCODE(0) |
+ SM_DEBUG_SYNCPASS_ENCODE(0) |
+ SM_DEBUG_TXG_ENCODE(0);
+ block_cursor++;
+ sm->sm_phys->smp_length += sizeof (uint64_t);
+ ASSERT3P(block_cursor, ==, block_end);
+ continue;
+ }
+
+ uint64_t run_len = MIN(size, run_max);
+ switch (words) {
+ case 1:
+ *block_cursor = SM_OFFSET_ENCODE(start) |
+ SM_TYPE_ENCODE(maptype) |
+ SM_RUN_ENCODE(run_len);
+ block_cursor++;
+ break;
+ case 2:
+ /* write the first word of the entry */
+ *block_cursor = SM_PREFIX_ENCODE(SM2_PREFIX) |
+ SM2_RUN_ENCODE(run_len) |
+ SM2_VDEV_ENCODE(vdev_id);
+ block_cursor++;
+
+ /* move on to the second word of the entry */
+ ASSERT3P(block_cursor, <, block_end);
+ *block_cursor = SM2_TYPE_ENCODE(maptype) |
+ SM2_OFFSET_ENCODE(start);
+ block_cursor++;
+ break;
+ default:
+ panic("%d-word space map entries are not supported",
+ words);
+ break;
+ }
+ sm->sm_phys->smp_length += words * sizeof (uint64_t);
+
+ start += run_len;
+ size -= run_len;
+ }
+ ASSERT0(size);
+
+}
+
+/*
+ * Note: The space map's dbuf must be dirty for the changes in sm_phys to
+ * take effect.
+ */
+static void
+space_map_write_impl(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
+ uint64_t vdev_id, dmu_tx_t *tx)
+{
+ spa_t *spa = tx->tx_pool->dp_spa;
+ dmu_buf_t *db;
+
+ space_map_write_intro_debug(sm, maptype, tx);
+
+#ifdef DEBUG
/*
- * All space_maps always have a debug entry so account for it here.
+ * We do this right after we write the intro debug entry
+ * because the estimate does not take it into account.
*/
- entries = 1;
+ uint64_t initial_objsize = sm->sm_phys->smp_length;
+ uint64_t estimated_growth =
+ space_map_estimate_optimal_size(sm, rt, SM_NO_VDEVID);
+ uint64_t estimated_final_objsize = initial_objsize + estimated_growth;
+#endif
/*
- * Traverse the range tree and calculate the number of space map
- * entries that would be required to write out the range tree.
+ * Find the offset right after the last word in the space map
+ * and use that to get a hold of the last block, so we can
+ * start appending to it.
*/
- for (rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
- size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
- entries += howmany(size, SM_RUN_MAX);
+ uint64_t next_word_offset = sm->sm_phys->smp_length;
+ VERIFY0(dmu_buf_hold(sm->sm_os, space_map_object(sm),
+ next_word_offset, FTAG, &db, DMU_READ_PREFETCH));
+ ASSERT3U(db->db_size, ==, sm->sm_blksz);
+
+ dmu_buf_will_dirty(db, tx);
+
+ avl_tree_t *t = &rt->rt_root;
+ for (range_seg_t *rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
+ uint64_t offset = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
+ uint64_t length = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
+ uint8_t words = 1;
+
+ /*
+ * We only write two-word entries when both of the following
+ * are true:
+ *
+ * [1] The feature is enabled.
+ * [2] The offset or run is too big for a single-word entry,
+ * or the vdev_id is set (meaning not equal to
+ * SM_NO_VDEVID).
+ *
+ * Note that for purposes of testing we've added the case that
+ * we write two-word entries occasionally when the feature is
+ * enabled and zfs_force_some_double_word_sm_entries has been
+ * set.
+ */
+ if (spa_feature_is_active(spa, SPA_FEATURE_SPACEMAP_V2) &&
+ (offset >= (1ULL << SM_OFFSET_BITS) ||
+ length > SM_RUN_MAX ||
+ vdev_id != SM_NO_VDEVID ||
+ (zfs_force_some_double_word_sm_entries &&
+ spa_get_random(100) == 0)))
+ words = 2;
+
+ space_map_write_seg(sm, rs, maptype, vdev_id, words,
+ &db, FTAG, tx);
}
- return (entries);
+
+ dmu_buf_rele(db, FTAG);
+
+#ifdef DEBUG
+ /*
+ * We expect our estimation to be based on the worst case
+ * scenario [see comment in space_map_estimate_optimal_size()].
+ * Therefore we expect the actual objsize to be equal or less
+ * than whatever we estimated it to be.
+ */
+ ASSERT3U(estimated_final_objsize, >=, sm->sm_phys->smp_length);
+#endif
}
+/*
+ * Note: This function manipulates the state of the given space map but
+ * does not hold any locks implicitly. Thus the caller is responsible
+ * for synchronizing writes to the space map.
+ */
void
space_map_write(space_map_t *sm, range_tree_t *rt, maptype_t maptype,
- dmu_tx_t *tx)
+ uint64_t vdev_id, dmu_tx_t *tx)
{
- objset_t *os = sm->sm_os;
- spa_t *spa = dmu_objset_spa(os);
- avl_tree_t *t = &rt->rt_root;
- range_seg_t *rs;
- uint64_t size, total, rt_space, nodes;
- uint64_t *entry, *entry_map, *entry_map_end;
- uint64_t expected_entries, actual_entries = 1;
-
- ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
+ ASSERT(dsl_pool_sync_context(dmu_objset_pool(sm->sm_os)));
VERIFY3U(space_map_object(sm), !=, 0);
+
dmu_buf_will_dirty(sm->sm_dbuf, tx);
/*
*/
sm->sm_phys->smp_object = sm->sm_object;
- if (range_tree_space(rt) == 0) {
+ if (range_tree_is_empty(rt)) {
VERIFY3U(sm->sm_object, ==, sm->sm_phys->smp_object);
return;
}
else
sm->sm_phys->smp_alloc -= range_tree_space(rt);
- expected_entries = space_map_entries(sm, rt);
+ uint64_t nodes = avl_numnodes(&rt->rt_root);
+ uint64_t rt_space = range_tree_space(rt);
- entry_map = vmem_alloc(sm->sm_blksz, KM_SLEEP);
- entry_map_end = entry_map + (sm->sm_blksz / sizeof (uint64_t));
- entry = entry_map;
-
- *entry++ = SM_DEBUG_ENCODE(1) |
- SM_DEBUG_ACTION_ENCODE(maptype) |
- SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
- SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
-
- total = 0;
- nodes = avl_numnodes(&rt->rt_root);
- rt_space = range_tree_space(rt);
- for (rs = avl_first(t); rs != NULL; rs = AVL_NEXT(t, rs)) {
- uint64_t start;
-
- size = (rs->rs_end - rs->rs_start) >> sm->sm_shift;
- start = (rs->rs_start - sm->sm_start) >> sm->sm_shift;
-
- total += size << sm->sm_shift;
-
- while (size != 0) {
- uint64_t run_len;
-
- run_len = MIN(size, SM_RUN_MAX);
-
- if (entry == entry_map_end) {
- dmu_write(os, space_map_object(sm),
- sm->sm_phys->smp_objsize, sm->sm_blksz,
- entry_map, tx);
- sm->sm_phys->smp_objsize += sm->sm_blksz;
- entry = entry_map;
- }
-
- *entry++ = SM_OFFSET_ENCODE(start) |
- SM_TYPE_ENCODE(maptype) |
- SM_RUN_ENCODE(run_len);
-
- start += run_len;
- size -= run_len;
- actual_entries++;
- }
- }
-
- if (entry != entry_map) {
- size = (entry - entry_map) * sizeof (uint64_t);
- dmu_write(os, space_map_object(sm), sm->sm_phys->smp_objsize,
- size, entry_map, tx);
- sm->sm_phys->smp_objsize += size;
- }
- ASSERT3U(expected_entries, ==, actual_entries);
+ space_map_write_impl(sm, rt, maptype, vdev_id, tx);
/*
* Ensure that the space_map's accounting wasn't changed
*/
VERIFY3U(nodes, ==, avl_numnodes(&rt->rt_root));
VERIFY3U(range_tree_space(rt), ==, rt_space);
- VERIFY3U(range_tree_space(rt), ==, total);
-
- vmem_free(entry_map, sm->sm_blksz);
}
static int
sm->sm_shift = shift;
sm->sm_os = os;
sm->sm_object = object;
- sm->sm_length = 0;
- sm->sm_alloc = 0;
sm->sm_blksz = 0;
sm->sm_dbuf = NULL;
sm->sm_phys = NULL;
space_map_close(sm);
return (error);
}
-
*smp = sm;
return (0);
}
void
-space_map_truncate(space_map_t *sm, dmu_tx_t *tx)
+space_map_truncate(space_map_t *sm, int blocksize, dmu_tx_t *tx)
{
objset_t *os = sm->sm_os;
spa_t *spa = dmu_objset_spa(os);
*/
if ((spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
doi.doi_bonus_size != sizeof (space_map_phys_t)) ||
- doi.doi_data_block_size != space_map_blksz) {
+ doi.doi_data_block_size != blocksize ||
+ doi.doi_metadata_block_size != 1 << space_map_ibs) {
zfs_dbgmsg("txg %llu, spa %s, sm %p, reallocating "
"object[%llu]: old bonus %u, old blocksz %u",
dmu_tx_get_txg(tx), spa_name(spa), sm, sm->sm_object,
space_map_free(sm, tx);
dmu_buf_rele(sm->sm_dbuf, sm);
- sm->sm_object = space_map_alloc(sm->sm_os, tx);
+ sm->sm_object = space_map_alloc(sm->sm_os, blocksize, tx);
VERIFY0(space_map_open_impl(sm));
} else {
VERIFY0(dmu_free_range(os, space_map_object(sm), 0, -1ULL, tx));
}
dmu_buf_will_dirty(sm->sm_dbuf, tx);
- sm->sm_phys->smp_objsize = 0;
+ sm->sm_phys->smp_length = 0;
sm->sm_phys->smp_alloc = 0;
}
-/*
- * Update the in-core space_map allocation and length values.
- */
-void
-space_map_update(space_map_t *sm)
-{
- if (sm == NULL)
- return;
-
- sm->sm_alloc = sm->sm_phys->smp_alloc;
- sm->sm_length = sm->sm_phys->smp_objsize;
-}
-
uint64_t
-space_map_alloc(objset_t *os, dmu_tx_t *tx)
+space_map_alloc(objset_t *os, int blocksize, dmu_tx_t *tx)
{
spa_t *spa = dmu_objset_spa(os);
uint64_t object;
bonuslen = SPACE_MAP_SIZE_V0;
}
- object = dmu_object_alloc(os,
- DMU_OT_SPACE_MAP, space_map_blksz,
- DMU_OT_SPACE_MAP_HEADER, bonuslen, tx);
+ object = dmu_object_alloc_ibs(os, DMU_OT_SPACE_MAP, blocksize,
+ space_map_ibs, DMU_OT_SPACE_MAP_HEADER, bonuslen, tx);
return (object);
}
sm->sm_object = 0;
}
+/*
+ * Given a range tree, it makes a worst-case estimate of how much
+ * space would the tree's segments take if they were written to
+ * the given space map.
+ */
+uint64_t
+space_map_estimate_optimal_size(space_map_t *sm, range_tree_t *rt,
+ uint64_t vdev_id)
+{
+ spa_t *spa = dmu_objset_spa(sm->sm_os);
+ uint64_t shift = sm->sm_shift;
+ uint64_t *histogram = rt->rt_histogram;
+ uint64_t entries_for_seg = 0;
+
+ /*
+ * In order to get a quick estimate of the optimal size that this
+ * range tree would have on-disk as a space map, we iterate through
+ * its histogram buckets instead of iterating through its nodes.
+ *
+ * Note that this is a highest-bound/worst-case estimate for the
+ * following reasons:
+ *
+ * 1] We assume that we always add a debug padding for each block
+ * we write and we also assume that we start at the last word
+ * of a block attempting to write a two-word entry.
+ * 2] Rounding up errors due to the way segments are distributed
+ * in the buckets of the range tree's histogram.
+ * 3] The activation of zfs_force_some_double_word_sm_entries
+ * (tunable) when testing.
+ *
+ * = Math and Rounding Errors =
+ *
+ * rt_histogram[i] bucket of a range tree represents the number
+ * of entries in [2^i, (2^(i+1))-1] of that range_tree. Given
+ * that, we want to divide the buckets into groups: Buckets that
+ * can be represented using a single-word entry, ones that can
+ * be represented with a double-word entry, and ones that can
+ * only be represented with multiple two-word entries.
+ *
+ * [Note that if the new encoding feature is not enabled there
+ * are only two groups: single-word entry buckets and multiple
+ * single-word entry buckets. The information below assumes
+ * two-word entries enabled, but it can easily applied when
+ * the feature is not enabled]
+ *
+ * To find the highest bucket that can be represented with a
+ * single-word entry we look at the maximum run that such entry
+ * can have, which is 2^(SM_RUN_BITS + sm_shift) [remember that
+ * the run of a space map entry is shifted by sm_shift, thus we
+ * add it to the exponent]. This way, excluding the value of the
+ * maximum run that can be represented by a single-word entry,
+ * all runs that are smaller exist in buckets 0 to
+ * SM_RUN_BITS + shift - 1.
+ *
+ * To find the highest bucket that can be represented with a
+ * double-word entry, we follow the same approach. Finally, any
+ * bucket higher than that are represented with multiple two-word
+ * entries. To be more specific, if the highest bucket whose
+ * segments can be represented with a single two-word entry is X,
+ * then bucket X+1 will need 2 two-word entries for each of its
+ * segments, X+2 will need 4, X+3 will need 8, ...etc.
+ *
+ * With all of the above we make our estimation based on bucket
+ * groups. There is a rounding error though. As we mentioned in
+ * the example with the one-word entry, the maximum run that can
+ * be represented in a one-word entry 2^(SM_RUN_BITS + shift) is
+ * not part of bucket SM_RUN_BITS + shift - 1. Thus, segments of
+ * that length fall into the next bucket (and bucket group) where
+ * we start counting two-word entries and this is one more reason
+ * why the estimated size may end up being bigger than the actual
+ * size written.
+ */
+ uint64_t size = 0;
+ uint64_t idx = 0;
+
+ if (!spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2) ||
+ (vdev_id == SM_NO_VDEVID && sm->sm_size < SM_OFFSET_MAX)) {
+
+ /*
+ * If we are trying to force some double word entries just
+ * assume the worst-case of every single word entry being
+ * written as a double word entry.
+ */
+ uint64_t entry_size =
+ (spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2) &&
+ zfs_force_some_double_word_sm_entries) ?
+ (2 * sizeof (uint64_t)) : sizeof (uint64_t);
+
+ uint64_t single_entry_max_bucket = SM_RUN_BITS + shift - 1;
+ for (; idx <= single_entry_max_bucket; idx++)
+ size += histogram[idx] * entry_size;
+
+ if (!spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2)) {
+ for (; idx < RANGE_TREE_HISTOGRAM_SIZE; idx++) {
+ ASSERT3U(idx, >=, single_entry_max_bucket);
+ entries_for_seg =
+ 1ULL << (idx - single_entry_max_bucket);
+ size += histogram[idx] *
+ entries_for_seg * entry_size;
+ }
+ return (size);
+ }
+ }
+
+ ASSERT(spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_V2));
+
+ uint64_t double_entry_max_bucket = SM2_RUN_BITS + shift - 1;
+ for (; idx <= double_entry_max_bucket; idx++)
+ size += histogram[idx] * 2 * sizeof (uint64_t);
+
+ for (; idx < RANGE_TREE_HISTOGRAM_SIZE; idx++) {
+ ASSERT3U(idx, >=, double_entry_max_bucket);
+ entries_for_seg = 1ULL << (idx - double_entry_max_bucket);
+ size += histogram[idx] *
+ entries_for_seg * 2 * sizeof (uint64_t);
+ }
+
+ /*
+ * Assume the worst case where we start with the padding at the end
+ * of the current block and we add an extra padding entry at the end
+ * of all subsequent blocks.
+ */
+ size += ((size / sm->sm_blksz) + 1) * sizeof (uint64_t);
+
+ return (size);
+}
+
uint64_t
space_map_object(space_map_t *sm)
{
return (sm != NULL ? sm->sm_object : 0);
}
-/*
- * Returns the already synced, on-disk allocated space.
- */
-uint64_t
+int64_t
space_map_allocated(space_map_t *sm)
{
- return (sm != NULL ? sm->sm_alloc : 0);
+ return (sm != NULL ? sm->sm_phys->smp_alloc : 0);
}
-/*
- * Returns the already synced, on-disk length;
- */
uint64_t
space_map_length(space_map_t *sm)
{
- return (sm != NULL ? sm->sm_length : 0);
-}
-
-/*
- * Returns the allocated space that is currently syncing.
- */
-int64_t
-space_map_alloc_delta(space_map_t *sm)
-{
- if (sm == NULL)
- return (0);
- ASSERT(sm->sm_dbuf != NULL);
- return (sm->sm_phys->smp_alloc - space_map_allocated(sm));
+ return (sm != NULL ? sm->sm_phys->smp_length : 0);
}