1 files changed, 426 insertions, 0 deletions

diff --git a/drivers/md/dm-vdo/indexer/open-chapter.c b/drivers/md/dm-vdo/indexer/open-chapter.c
new file mode 100644
index 0000000000..4a67bcadaa
--- /dev/null
+++ b/drivers/md/dm-vdo/indexer/open-chapter.c
@@ -0,0 +1,426 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright 2023 Red Hat
+ */
+
+#include "open-chapter.h"
+
+#include <linux/log2.h>
+
+#include "logger.h"
+#include "memory-alloc.h"
+#include "numeric.h"
+#include "permassert.h"
+
+#include "config.h"
+#include "hash-utils.h"
+
+/*
+ * Each index zone has a dedicated open chapter zone structure which gets an equal share of the
+ * open chapter space. Records are assigned to zones based on their record name. Within each zone,
+ * records are stored in an array in the order they arrive. Additionally, a reference to each
+ * record is stored in a hash table to help determine if a new record duplicates an existing one.
+ * If new metadata for an existing name arrives, the record is altered in place. The array of
+ * records is 1-based so that record number 0 can be used to indicate an unused hash slot.
+ *
+ * Deleted records are marked with a flag rather than actually removed to simplify hash table
+ * management. The array of deleted flags overlays the array of hash slots, but the flags are
+ * indexed by record number instead of by record name. The number of hash slots will always be a
+ * power of two that is greater than the number of records to be indexed, guaranteeing that hash
+ * insertion cannot fail, and that there are sufficient flags for all records.
+ *
+ * Once any open chapter zone fills its available space, the chapter is closed. The records from
+ * each zone are interleaved to attempt to preserve temporal locality and assigned to record pages.
+ * Empty or deleted records are replaced by copies of a valid record so that the record pages only
+ * contain valid records. The chapter then constructs a delta index which maps each record name to
+ * the record page on which that record can be found, which is split into index pages. These
+ * structures are then passed to the volume to be recorded on storage.
+ *
+ * When the index is saved, the open chapter records are saved in a single array, once again
+ * interleaved to attempt to preserve temporal locality. When the index is reloaded, there may be a
+ * different number of zones than previously, so the records must be parcelled out to their new
+ * zones. In addition, depending on the distribution of record names, a new zone may have more
+ * records than it has space. In this case, the latest records for that zone will be discarded.
+ */
+
+static const u8 OPEN_CHAPTER_MAGIC[] = "ALBOC";
+static const u8 OPEN_CHAPTER_VERSION[] = "02.00";
+
+#define OPEN_CHAPTER_MAGIC_LENGTH (sizeof(OPEN_CHAPTER_MAGIC) - 1)
+#define OPEN_CHAPTER_VERSION_LENGTH (sizeof(OPEN_CHAPTER_VERSION) - 1)
+#define LOAD_RATIO 2
+
+static inline size_t records_size(const struct open_chapter_zone *open_chapter)
+{
+	return sizeof(struct uds_volume_record) * (1 + open_chapter->capacity);
+}
+
+static inline size_t slots_size(size_t slot_count)
+{
+	return sizeof(struct open_chapter_zone_slot) * slot_count;
+}
+
+int uds_make_open_chapter(const struct index_geometry *geometry, unsigned int zone_count,
+			  struct open_chapter_zone **open_chapter_ptr)
+{
+	int result;
+	struct open_chapter_zone *open_chapter;
+	size_t capacity = geometry->records_per_chapter / zone_count;
+	size_t slot_count = (1 << bits_per(capacity * LOAD_RATIO));
+
+	result = vdo_allocate_extended(struct open_chapter_zone, slot_count,
+				       struct open_chapter_zone_slot, "open chapter",
+				       &open_chapter);
+	if (result != VDO_SUCCESS)
+		return result;
+
+	open_chapter->slot_count = slot_count;
+	open_chapter->capacity = capacity;
+	result = vdo_allocate_cache_aligned(records_size(open_chapter), "record pages",
+					    &open_chapter->records);
+	if (result != VDO_SUCCESS) {
+		uds_free_open_chapter(open_chapter);
+		return result;
+	}
+
+	*open_chapter_ptr = open_chapter;
+	return UDS_SUCCESS;
+}
+
+void uds_reset_open_chapter(struct open_chapter_zone *open_chapter)
+{
+	open_chapter->size = 0;
+	open_chapter->deletions = 0;
+
+	memset(open_chapter->records, 0, records_size(open_chapter));
+	memset(open_chapter->slots, 0, slots_size(open_chapter->slot_count));
+}
+
+static unsigned int probe_chapter_slots(struct open_chapter_zone *open_chapter,
+					const struct uds_record_name *name)
+{
+	struct uds_volume_record *record;
+	unsigned int slot_count = open_chapter->slot_count;
+	unsigned int slot = uds_name_to_hash_slot(name, slot_count);
+	unsigned int record_number;
+	unsigned int attempts = 1;
+
+	while (true) {
+		record_number = open_chapter->slots[slot].record_number;
+
+		/*
+		 * If the hash slot is empty, we've reached the end of a chain without finding the
+		 * record and should terminate the search.
+		 */
+		if (record_number == 0)
+			return slot;
+
+		/*
+		 * If the name of the record referenced by the slot matches and has not been
+		 * deleted, then we've found the requested name.
+		 */
+		record = &open_chapter->records[record_number];
+		if ((memcmp(&record->name, name, UDS_RECORD_NAME_SIZE) == 0) &&
+		    !open_chapter->slots[record_number].deleted)
+			return slot;
+
+		/*
+		 * Quadratic probing: advance the probe by 1, 2, 3, etc. and try again. This
+		 * performs better than linear probing and works best for 2^N slots.
+		 */
+		slot = (slot + attempts++) % slot_count;
+	}
+}
+
+void uds_search_open_chapter(struct open_chapter_zone *open_chapter,
+			     const struct uds_record_name *name,
+			     struct uds_record_data *metadata, bool *found)
+{
+	unsigned int slot;
+	unsigned int record_number;
+
+	slot = probe_chapter_slots(open_chapter, name);
+	record_number = open_chapter->slots[slot].record_number;
+	if (record_number == 0) {
+		*found = false;
+	} else {
+		*found = true;
+		*metadata = open_chapter->records[record_number].data;
+	}
+}
+
+/* Add a record to the open chapter zone and return the remaining space. */
+int uds_put_open_chapter(struct open_chapter_zone *open_chapter,
+			 const struct uds_record_name *name,
+			 const struct uds_record_data *metadata)
+{
+	unsigned int slot;
+	unsigned int record_number;
+	struct uds_volume_record *record;
+
+	if (open_chapter->size >= open_chapter->capacity)
+		return 0;
+
+	slot = probe_chapter_slots(open_chapter, name);
+	record_number = open_chapter->slots[slot].record_number;
+
+	if (record_number == 0) {
+		record_number = ++open_chapter->size;
+		open_chapter->slots[slot].record_number = record_number;
+	}
+
+	record = &open_chapter->records[record_number];
+	record->name = *name;
+	record->data = *metadata;
+
+	return open_chapter->capacity - open_chapter->size;
+}
+
+void uds_remove_from_open_chapter(struct open_chapter_zone *open_chapter,
+				  const struct uds_record_name *name)
+{
+	unsigned int slot;
+	unsigned int record_number;
+
+	slot = probe_chapter_slots(open_chapter, name);
+	record_number = open_chapter->slots[slot].record_number;
+
+	if (record_number > 0) {
+		open_chapter->slots[record_number].deleted = true;
+		open_chapter->deletions += 1;
+	}
+}
+
+void uds_free_open_chapter(struct open_chapter_zone *open_chapter)
+{
+	if (open_chapter != NULL) {
+		vdo_free(open_chapter->records);
+		vdo_free(open_chapter);
+	}
+}
+
+/* Map each record name to its record page number in the delta chapter index. */
+static int fill_delta_chapter_index(struct open_chapter_zone **chapter_zones,
+				    unsigned int zone_count,
+				    struct open_chapter_index *index,
+				    struct uds_volume_record *collated_records)
+{
+	int result;
+	unsigned int records_per_chapter;
+	unsigned int records_per_page;
+	unsigned int record_index;
+	unsigned int records = 0;
+	u32 page_number;
+	unsigned int z;
+	int overflow_count = 0;
+	struct uds_volume_record *fill_record = NULL;
+
+	/*
+	 * The record pages should not have any empty space, so find a record with which to fill
+	 * the chapter zone if it was closed early, and also to replace any deleted records. The
+	 * last record in any filled zone is guaranteed to not have been deleted, so use one of
+	 * those.
+	 */
+	for (z = 0; z < zone_count; z++) {
+		struct open_chapter_zone *zone = chapter_zones[z];
+
+		if (zone->size == zone->capacity) {
+			fill_record = &zone->records[zone->size];
+			break;
+		}
+	}
+
+	records_per_chapter = index->geometry->records_per_chapter;
+	records_per_page = index->geometry->records_per_page;
+
+	for (records = 0; records < records_per_chapter; records++) {
+		struct uds_volume_record *record = &collated_records[records];
+		struct open_chapter_zone *open_chapter;
+
+		/* The record arrays in the zones are 1-based. */
+		record_index = 1 + (records / zone_count);
+		page_number = records / records_per_page;
+		open_chapter = chapter_zones[records % zone_count];
+
+		/* Use the fill record in place of an unused record. */
+		if (record_index > open_chapter->size ||
+		    open_chapter->slots[record_index].deleted) {
+			*record = *fill_record;
+			continue;
+		}
+
+		*record = open_chapter->records[record_index];
+		result = uds_put_open_chapter_index_record(index, &record->name,
+							   page_number);
+		switch (result) {
+		case UDS_SUCCESS:
+			break;
+		case UDS_OVERFLOW:
+			overflow_count++;
+			break;
+		default:
+			vdo_log_error_strerror(result,
+					       "failed to build open chapter index");
+			return result;
+		}
+	}
+
+	if (overflow_count > 0)
+		vdo_log_warning("Failed to add %d entries to chapter index",
+				overflow_count);
+
+	return UDS_SUCCESS;
+}
+
+int uds_close_open_chapter(struct open_chapter_zone **chapter_zones,
+			   unsigned int zone_count, struct volume *volume,
+			   struct open_chapter_index *chapter_index,
+			   struct uds_volume_record *collated_records,
+			   u64 virtual_chapter_number)
+{
+	int result;
+
+	uds_empty_open_chapter_index(chapter_index, virtual_chapter_number);
+	result = fill_delta_chapter_index(chapter_zones, zone_count, chapter_index,
+					  collated_records);
+	if (result != UDS_SUCCESS)
+		return result;
+
+	return uds_write_chapter(volume, chapter_index, collated_records);
+}
+
+int uds_save_open_chapter(struct uds_index *index, struct buffered_writer *writer)
+{
+	int result;
+	struct open_chapter_zone *open_chapter;
+	struct uds_volume_record *record;
+	u8 record_count_data[sizeof(u32)];
+	u32 record_count = 0;
+	unsigned int record_index;
+	unsigned int z;
+
+	result = uds_write_to_buffered_writer(writer, OPEN_CHAPTER_MAGIC,
+					      OPEN_CHAPTER_MAGIC_LENGTH);
+	if (result != UDS_SUCCESS)
+		return result;
+
+	result = uds_write_to_buffered_writer(writer, OPEN_CHAPTER_VERSION,
+					      OPEN_CHAPTER_VERSION_LENGTH);
+	if (result != UDS_SUCCESS)
+		return result;
+
+	for (z = 0; z < index->zone_count; z++) {
+		open_chapter = index->zones[z]->open_chapter;
+		record_count += open_chapter->size - open_chapter->deletions;
+	}
+
+	put_unaligned_le32(record_count, record_count_data);
+	result = uds_write_to_buffered_writer(writer, record_count_data,
+					      sizeof(record_count_data));
+	if (result != UDS_SUCCESS)
+		return result;
+
+	record_index = 1;
+	while (record_count > 0) {
+		for (z = 0; z < index->zone_count; z++) {
+			open_chapter = index->zones[z]->open_chapter;
+			if (record_index > open_chapter->size)
+				continue;
+
+			if (open_chapter->slots[record_index].deleted)
+				continue;
+
+			record = &open_chapter->records[record_index];
+			result = uds_write_to_buffered_writer(writer, (u8 *) record,
+							      sizeof(*record));
+			if (result != UDS_SUCCESS)
+				return result;
+
+			record_count--;
+		}
+
+		record_index++;
+	}
+
+	return uds_flush_buffered_writer(writer);
+}
+
+u64 uds_compute_saved_open_chapter_size(struct index_geometry *geometry)
+{
+	unsigned int records_per_chapter = geometry->records_per_chapter;
+
+	return OPEN_CHAPTER_MAGIC_LENGTH + OPEN_CHAPTER_VERSION_LENGTH + sizeof(u32) +
+		records_per_chapter * sizeof(struct uds_volume_record);
+}
+
+static int load_version20(struct uds_index *index, struct buffered_reader *reader)
+{
+	int result;
+	u32 record_count;
+	u8 record_count_data[sizeof(u32)];
+	struct uds_volume_record record;
+
+	/*
+	 * Track which zones cannot accept any more records. If the open chapter had a different
+	 * number of zones previously, some new zones may have more records than they have space
+	 * for. These overflow records will be discarded.
+	 */
+	bool full_flags[MAX_ZONES] = {
+		false,
+	};
+
+	result = uds_read_from_buffered_reader(reader, (u8 *) &record_count_data,
+					       sizeof(record_count_data));
+	if (result != UDS_SUCCESS)
+		return result;
+
+	record_count = get_unaligned_le32(record_count_data);
+	while (record_count-- > 0) {
+		unsigned int zone = 0;
+
+		result = uds_read_from_buffered_reader(reader, (u8 *) &record,
+						       sizeof(record));
+		if (result != UDS_SUCCESS)
+			return result;
+
+		if (index->zone_count > 1)
+			zone = uds_get_volume_index_zone(index->volume_index,
+							 &record.name);
+
+		if (!full_flags[zone]) {
+			struct open_chapter_zone *open_chapter;
+			unsigned int remaining;
+
+			open_chapter = index->zones[zone]->open_chapter;
+			remaining = uds_put_open_chapter(open_chapter, &record.name,
+							 &record.data);
+			/* Do not allow any zone to fill completely. */
+			full_flags[zone] = (remaining <= 1);
+		}
+	}
+
+	return UDS_SUCCESS;
+}
+
+int uds_load_open_chapter(struct uds_index *index, struct buffered_reader *reader)
+{
+	u8 version[OPEN_CHAPTER_VERSION_LENGTH];
+	int result;
+
+	result = uds_verify_buffered_data(reader, OPEN_CHAPTER_MAGIC,
+					  OPEN_CHAPTER_MAGIC_LENGTH);
+	if (result != UDS_SUCCESS)
+		return result;
+
+	result = uds_read_from_buffered_reader(reader, version, sizeof(version));
+	if (result != UDS_SUCCESS)
+		return result;
+
+	if (memcmp(OPEN_CHAPTER_VERSION, version, sizeof(version)) != 0) {
+		return vdo_log_error_strerror(UDS_CORRUPT_DATA,
+					      "Invalid open chapter version: %.*s",
+					      (int) sizeof(version), version);
+	}
+
+	return load_version20(index, reader);
+}