FIGS. 3A(i)-3A(iii) illustrate schematically the mapping between a logical group and a metablock, according to a preferred embodiment of the present invention.
The host 10 accesses the memory 200 when running an application under a file system or operating system. Typically, the host system addresses data in units of logical sectors where, for example, each sector may contain 512 bytes of data. Also, it is usual for the host to read or write to the memory system in unit of logical clusters, each consisting of one or more logical sectors. In some host systems, an optional host-side memory manager may exist to perform lower level memory management at the host. In most cases during read or write operations, the host 10 essentially issues a command to the memory system 20 to read or write a segment containing a string of logical sectors of data with contiguous addresses.
A memory-side memory manager is implemented in the controller 100 of the memory system 20 to manage the storage and retrieval of the data of host logical sectors among metablocks of the flash memory 200. In the preferred embodiment, the memory manager contains a number of software modules for managing erase, read and write operations of the metablocks. The memory manager also maintains system control and directory data associated with its operations among the flash memory 200 and the controller RAM 130.
FIGS. 3A(i)-3A(iii) illustrate schematically the mapping between a logical group and a metablock, according to a preferred embodiment of the present invention. The metablock of the physical memory has N physical sectors for storing N logical sectors of data of a logical group. FIG. 3A(i) shows the data from a logical group LGi, where the logical sectors are in contiguous logical order 0, 1, . . . , N−1. FIG. 3A(ii) shows the same data being stored in the metablock in the same logical order. The metablock when stored in this manner is said to be “sequential.” In general, the metablock may have data stored in a different order, in which case the metablock is said to be “non-sequential” or “chaotic.”
There may be an offset between the lowest address of a logical group and the lowest address of the metablock to which it is mapped. In this case, logical sector address wraps round as a loop from bottom back to top of the logical group within the metablock. For example, in FIG. 3A(iii), the metablock stores in its first location beginning with the data of logical sector k. When the last logical sector N−1 is reached, it wraps around to sector 0 and finally storing data associated with logical sector k−1 in its last physical sector. In the preferred embodiment, a page tag is used to identify any offset, such as identifying the starting logical sector address of the data stored in the first physical sector of the metablock. Two blocks will be considered to have their logical sectors stored in similar order when they only differ by a page tag.
Other types of logical group to metablock mapping are also comtemplated. For example, metablocks with variable size are disclosed in co-pending and co-owned United States Patent application, entitled, “Adaptive Metablocks,” filed by Alan Sinclair, on the same day as the present application. The entire disclosure of the co-pending application is hereby incorporated herein by reference.
One feature of the invention is that the system operates with a single logical partition, and groups of logical sectors throughout the logical address range of the memory system are treated identically. For example, sectors containing system data and sectors containing user data can be distributed anywhere among the logical address space.
Unlike prior art systems, there is no special partitioning or zoning of system sectors (i.e., sectors relating to file allocation tables, directories or sub-directories) in order to localize in logical address space sectors that are likely to contain data with high-frequency and small-size updates. Instead, the present scheme of updating logical groups of sectors will efficiently handle the patterns of access that are typical of system sectors, as well as those typical of file data.
The metablock represents, at the system level, a group of memory locations, e.g., sectors that are erasable together. The physical address space of the flash memory is treated as a set of metablocks, with a metablock being the minimum unit of erasure. Within this specification, the terms “metablock” and “block” are used synonymously to define the minimum unit of erasure at the system level for media management, and the term “minimum erase unit” or MEU is used to denote the minimum unit of erasure of flash memory.
In order to maximize programming speed and erase speed, parallelism is exploited as much as possible by arranging for multiple pages of information, located in multiple MEUs, to be programmed in parallel, and for multiple MEUs to be erased in parallel.
In flash memory, a page is a grouping of memory cells that may be programmed together in a single operation. A page may comprise one or more sector. Also, a memory array may be partitioned into more than one plane, where only one MEU within a plane may be programmed or erased at a time. Finally, the planes may be distributed among one or more memory chips.
In flash memory, the MEUs may comprise one or more page. MEUs within a flash memory chip may be organized in planes. Since one MEU from each plane may be programmed or erased concurrently, it is expedient to form a multiple MEU metablock by selecting one MEU from each plane (see
The linking and re-linking of MEUs into metablocks is also disclosed in co-pending and co-owned United States Patent application, entitled “Adaptive Deterministic Grouping of Blocks into Multi-Block Structures,” filed by Carlos Gonzales et al, on the same day as the present application. The entire disclosure of the co-pending application is hereby incorporated herein by reference.
The interface 110 allows the metablock management system to interface with a host system. The logical to physical address translation module 140 maps the logical address from the host to a physical memory location. The update block Manager module 150 manages data update operations in memory for a given logical group of data. The erased block manager 160 manages the erase operation of the metablocks and their allocation for storage of new information. A metablock link manager 170 manages the linking of subgroups of minimum erasable blocks of sectors to constitute a given metablock. Detailed description of these modules will be given in their respective sections.
During operation the metablock management system generates and works with control data such as addresses, control and status information. Since much of the control data tends to be frequently changing data of small size, it can not be readily stored and maintained efficiently in a flash memory with a large block structure. A hierarchical and distributed scheme is employed to store the more static control data in the nonvolatile flash memory while locating the smaller amount of the more varying control data in controller RAM for more efficient update and access. In the event of a power shutdown or failure, the scheme allows the control data in the volatile controller RAM to be rebuilt quickly by scanning a small set of control data in the nonvolatile memory. This is possible because the invention restricts the number of blocks associated with the possible activity of a given logical group of data. In this way, the scanning is confined. In addition, some of the control data that requires persistence are stored in a nonvolatile metablock that can be updated sector-by-sector, with each update resulting in a new sector being recorded that supercedes a previous one. A sector indexing scheme is employed for control data to keep track of the sector-by-sector updates in a metablock.
The non-volatile flash memory 200 stores the bulk of control data that are relatively static. This includes group address tables (GAT) 210, chaotic block indices (CBI) 220, erased block lists (EBL) 230 and MAP 240. The GAT 210 keeps track of the mapping between logical groups of sectors and their corresponding metablocks. The mappings do not change except for those undergoing updates. The CBI 220 keeps track of the mapping of logically non-sequential sectors during an update. The EBL 230 keeps track of the pool of metablocks that have been erased. MAP 240 is a bitmap showing the erase status of all metablocks in the flash memory.
The volatile controller RAM 130 stores a small portion of control data that are frequently changing and accessed. This includes an allocation block list (ABL) 134 and a cleared block list (CBL) 136. The ABL 134 keeps track of the allocation of metablocks for recording update data while the CBL 136 keeps track of metablocks that have been deallocated and erased. In the preferred embodiment, the RAM 130 acts as a cache for control data stored in flash memory 200.
The update block manager 150 (shown in
Data of a complete logical group of sectors is preferably stored in logically sequential order in a single metablock. In this way, the index to the stored logical sectors is predefined. When the metablock has in store all the sectors of a given logical group in a predefined order it is said to be “intact.” As for an update block, when it eventually fills up with update data in logically sequential order, then the update block will become an updated intact metablock that readily replace the original metablock. On the other hand, if the update block fills up with update data in a logically different order from that of the intact block, the update block is a non-sequential or chaotic update block and the out of order segments must be further processed so that eventually the update data of the logical group is stored in the same order as that of the intact block. In the preferred case, it is in logically sequential order in a single metablock. The further processing involves consolidating the updated sectors in the update block with unchanged sectors in the original block into yet another update metablock. The consolidated update block will then be in logically sequential order and can be used to replace the original block. Under some predetermined condition, the consolidation process is preceded by one or more compaction processes. The compaction process simply re-records the sectors of the chaotic update block into a replacing chaotic update block while eliminating any duplicate logical sector that has been rendered obsolete by a subsequent update of the same logical sector.
The update scheme allows for multiple update threads running concurrently, up to a predefined maximum. Each thread is a logical group undergoing updates using its dedicated update metablock.
When data belonging to a logical group is first updated, a metablock is allocated and dedicated as an update block for the update data of the logical group. The update block is allocated when a command is received from the host to write a segment of one or more sectors of the logical group for which an existing metablock has been storing all its sectors intact. For the first host write operation, a first segment of data is recorded on the update block. Since each host write is a segment of one or more sector with contiguous logical address, it follows that the first update is always sequential in nature. In subsequent host writes, update segments within the same logical group are recorded in the update block in the order received from the host. A block continues to be managed as a sequential update block whilst sectors updated by the host within the associated logical group remain logically sequential. All sectors updated in this logical group are written to this sequential update block, until the block is either closed or converted to a chaotic update block.
For expediency, the first sector to be updated in the logical group is recorded in the dedicated update block starting from the first physical sector location. In general, the first logical sector to be updated is not necessarily the logical first sector of the group, and there may therefore be an offset between the start of the logical group and the start of the update block. This offset is known as page tag as described previously in connection with
In host write operation #2, the segment of data in the logical sectors LS9-LS12 are being updated. The updated data as LS9′-LS12′ are recorded in the dedicated update block in a location directly following where the last write ends. It can be seen that the two host writes are such that the update data has been recorded in the update block in logically sequential order, namely LS5′-LS12′. The update block is regarded as a sequential update block since it has been filled in logically sequential order. The update data recorded in the update block obsoletes the corresponding ones in the original block.
Chaotic update block management may be initiated for an existing sequential update block when any sector updated by the host within the associated logical group is logically non-sequential. A chaotic update block is a form of data update block in which logical sectors within an associated logical group may be updated in any order and with any amount of repetition. It is created by conversion from a sequential update block when a sector written by a host is logically non-sequential to the previously written sector within the logical group being updated. All sectors subsequently updated in this logical group are written in the next available sector location in the chaotic update block, whatever their logical sector address within the group.
STEP 260: The memory is organized into blocks, each block partitioned into memory units that are erasable together, each memory unit for storing a logical unit of data.
STEP 262: The data is organized into logical groups, each logical group partitioned into logical units.
STEP 264: In the standard case, all logical units of a logical group is stored among the memory units of an original block according to a first prescribed order, preferably, in logically sequential order. In this way, the index for accessing the individual logical units in the block is known.
STEP 270: For a given logical group (e.g., LGX) of data, a request is made to update a logical unit within LGX. (A logical unit update is given as an example. In general the update will be a segment of one or more contiguous logical units within LGx.)
STEP 272: The requested update logical unit is to be stored in a second block, dedicated to recording the updates of LGX. The recording order is according to a second order, typically, the order the updates are requested. One feature of the invention allows an update block to be set up initially generic to recording data in logically sequential or chaotic order. So depending on the second order, the second block can be a sequential one or a chaotic one.
STEP 274: The second block continues to have requested logical units recorded as the process loops back to STEP 270. The second block will be closed to receiving further update when a predetermined condition for closure materializes. In that case, the process proceeds to STEP 276.
STEP 276: Determination is made whether or not the closed, second block has its update logical units recorded in a similar order as that of the original block. The two blocks are considered to have similar order when they recorded logical units differ by only a page tag, as described in connection with
STEP 280: Since the second block has the same order as the first block, it is used to replace the original, first block. The update process then ends at STEP 299.
STEP 290: The latest version of each logical units of the given logical group are gathered from among the second block (update block) and the first block (original block). The consolidated logical units of the given logical group are then written to a third block in an order similar to the first block.
STEP 292: Since the third block (consolidated block) has a similar order to the first block, it is used to replace the original, first block. The update process then ends at STEP 299.
STEP 299: When a closeout process creates an intact update block, it becomes the new standard block for the given logical group. The update thread for the logical group will be terminated.
STEP 310: For a given logical group (e.g., LGX) of data, a request is made to update a logical sector within LGX. (A sector update is given as an example. In general the update will be a segment of one or more contiguous logical sectors within LG.)
STEP 312: If an update block dedicated to LGX does not already exist, proceed to STEP 410 to initiate a new update thread for the logical group. This will be accomplished by allocating an update block dedicated to recording update data of the logical group. If there is already an update block open, proceed to STEP 314 to begin recording the update sector onto the update block.
STEP 314: If the current update block is already chaotic (i.e., non-sequential) then simply proceed to STEP 510 for recording the requested update sector onto the chaotic update block. If the current update block is sequential, proceed to STEP 316 for processing of a sequential update block.
STEP 316: One feature of the invention allows an update block to be set up initially generic to recording data in logically sequential or chaotic order. However, since the logical group ultimately has its data stored in a metablock in a logically sequential order, it is desirable to keep the update block sequential as far as possible. Less processing will then be required when an update block is closed to further updates as garbage collection will not be needed.
Thus determination is made whether the requested update will follow the current sequential order of the update block. If the update follows sequentially, then proceed to STEP 510 to perform a sequential update, and the update block will remain sequential. On the other hand, if the update does not follow sequentially (chaotic update), it will convert the sequential update block to a chaotic one if no other actions are taken.
In one embodiment, nothing more is done to salvage the situation and the process proceeds directly to STEP 370 where the update is allowed to turn the update block into a chaotic one.
In another embodiment, a forced sequential process STEP 320 is optionally performed to preserve the sequential update block as far as possible in view of a pending chaotic update. There are two situations, both of which require copying missing sectors from the original block to maintain the sequential order of logical sectors recorded on the update block. The first situation is where the update creates a short address jump. The second situation is to prematurely close out an update block in order to keep it sequential. The forced sequential process STEP 320 comprises the following substeps:
STEP 330: If the update creates a logical address jump not greater a predetermined amount, CB, the process proceeds to a forced sequential update process in STEP 350, otherwise the process proceeds to STEP 340 to consider if it qualifies for a forced sequential closeout.
STEP 340: If the number of unfilled physical sectors exceeds a predetermined design parameter, CC, whose typical value is half of the size of the update block, then the update block is relatively unused and will not be prematurely closed. The process proceeds to STEP 370 and the update block will become chaotic. On the other hand, if the update block is substantially filled, it is considered to have been well utilized already and therefore is directed to STEP 360 for forced sequential closeout.
STEP 350: Forced sequential update allows current sequential update block to remain sequential as long as the address jump does not exceed a predetermined amount, CB. Essentially, sectors from the update block's associated original block are copied to fill the gap spanned by the address jump. Thus, the sequential update block will be padded with data in the intervening addresses before proceeding to STEP 510 to record the current update sequentially.
STEP 360: Forced sequential closeout allows the currently sequential update block to be closed out if it is already substantially filled rather than converted to a chaotic one by the pending chaotic update. A chaotic or non-sequential update is defined as one with a forward address transition not covered by the address jump exception described above, a backward address transition, or an address repetition. To prevent a sequential update block to be converted by a chaotic update, the unwritten sector locations of the update block are filled by copying sectors from the update block's associated original partly-obsolete block. The original block is then fully obsolete and may be erased. The current update block now has the full set of logical sectors and is then closed out as an intact metablock replacing the original metablock. The process then proceeds to STEP 430 to have a new update block allocated in its place to accept the recording of the pending sector update that was first requested in STEP 310.
STEP 370: When the pending update is not in sequential order and optionally, if the forced sequential conditions are not satisfied, the sequential update block is allowed to be converted to a chaotic one by virtue of allowing the pending update sector, with non-sequential address, to be recorded on the update block when the process proceeds to STEP 510. If the maximum number of chaotic update blocks exist, it is necessary to close the least recently accessed chaotic update block before allowing the conversion to proceed; thus preventing the maximum number of chaotic blocks from being exceeded. The identification of the least recently accessed chaotic update block is the same as the general case described in STEP 420, but is constrained to chaotic update blocks only. Closing a chaotic update block at this time is achieved by consolidation as described in STEP 550.
STEP 410: The process of allocating an erase metablock as an update block begins with the determination whether a predetermined system limitation is exceeded or not. Due to finite resources, the memory management system typically allows a predetermined maximum number of update blocks, UMAX, to exist concurrently. This limit is the aggregate of sequential update blocks and chaotic update blocks, and is a design parameter. In a preferred embodiment, the limit is, for example, a maximum of 8 update blocks. Also, due to the higher demand on system resources, there may also be a corresponding predetermined limit on the maximum number of chaotic update blocks that can be open concurrently (e.g., 4.)
Thus, when UMAX update blocks have already been allocated, then the next allocation request could only be satisfied after closing one of the existing allocated ones. The process proceeds to STEP 420. When the number of open update blocks is less than CA, the process proceeds directly to STEP 430.
STEP 420: In the event the maximum number of update blocks, CA, is exceeded, the least-recently accessed update block is closed and garbage collection is performed. The least recently accessed update block is identified as the update block associated with the logical block that has been accessed least recently. For the purpose of determining the least recently accessed blocks, an access includes writes and optionally reads of logical sectors. A list of open update blocks is maintained in order of access; at initialization, no access order is assumed. The closure of an update block follows along the similar process described in connection with STEP 360 and STEP 530 when the update block is sequential, and in connection with STEP 540 when the update block is chaotic. The closure makes room for the allocation of a new update block in STEP 430.
STEP 430: The allocation request is fulfilled with the allocation of a new metablock as an update block dedicated to the given logical group LGx. The process then proceeds to STEP 510.
STEP 510: The requested update sector is recorded onto next available physical location of the update block. The process then proceeds to STEP 520 to determine if the update block is ripe for closeout.
STEP 520: If the update block still has room for accepting additional updates, proceed to STEP 570. Otherwise proceed to STEP 522 to closeout the update block. There are two possible implementations of filling up an update block when the current requested write attempts to write more logical sectors than the block has room for. In the first implementation, the write request is split into two portions, with the first portion writing up to the last physical sector of the block. The block is then closed and the second portion of the write will be treated as the next requested write. In the other implementation, the requested write is withheld while the block has it remaining sectors padded and is then closed. The requested write will be treated as the next requested write.
STEP 522: If the update block is sequential, proceed to STEP 530 for sequential closure. If the update block is chaotic, proceed to STEP 540 for chaotic closure.
STEP 530: Since the update block is sequential and fully filled, the logical group stored in it is intact. The metablock is intact and replaces the original one. At this time, the original block is fully obsolete and may be erased. The process then proceeds to STEP 570 where the update thread for the given logical group ends.
STEP 540: Since the update block is non-sequentially filled and may contain multiple updates of some logical sectors, garbage collection is performed to salvage the valid data in it. The chaotic update block will either be compacted or consolidated. Which process to perform will be determined in STEP 542.
STEP 542: To perform compaction or consolidation will depend on the degeneracy of the update block. If a logical sector is updated multiple times, its logical address is highly degenerate. There will be multiple versions of the same logical sector recorded on the update block and only the last recorded version is the valid one for that logical sector. In an update block containing logical sectors with multiple versions, the number of distinct logical sectors will be much less than that of a logical group.
In the preferred embodiment, when the number of distinct logical sectors in the update block exceeds a predetermined design parameter, CD, whose typical value is half of the size of a logical group, the closeout process will perform a consolidation in STEP 550, otherwise the process will proceed to compaction in STEP 560.
STEP 550: If the chaotic update block is to be consolidated, the original block and the update block will be replaced by a new standard metablock containing the consolidated data. After consolidation the update thread will end in STEP 570.
STEP 560: If the chaotic update block is to be compacted, it will be replaced by a new update block carrying the compacted data. After compaction the processing of the compacted update block will end in STEP 570. Alternatively, compaction can be delayed until the update block is written to again, thus removing the possibility of compaction being followed by consolidation without intervening updates. The new update block will then be used in further updating of the given logical block when a next request for update in LGX appears in STEP 502.
STEP 570: When a closeout process creates an intact update block, it becomes the new standard block for the given logical group. The update thread for the logical group will be terminated. When a closeout process creates a new update block replacing an existing one, the new update block will be used to record the next update requested for the given logical group. When an update block is not closed out, the processing will continue when a next request for update in LGX appears in STEP 310.
As can be seen from the process described above, when a chaotic update block is closed, the update data recorded on it is further processed. In particular its valid data is garbage collected either by a process of compaction to another chaotic block, or by a process of consolidation with its associated original block to form a new standard sequential block.
STEP 551: When a chaotic update block is being closed, a new metablock replacing it will be allocated.
STEP 552: Gather the latest version of each logical sector among the chaotic update block and its associated original block, ignoring all the obsolete sectors.
STEP 554: Record the gathered valid sectors onto the new metablock in logically sequential order to form an intact block, i.e., a block with all the logical sectors of a logical group recorded in sequential order.
STEP 556: Replace the original block with the new intact block.
STEP 558: Erase the closed out update block and the original block.
STEP 561: When a chaotic update block is being compacted, a new metablock replacing it will be allocated.
STEP 562: Gather the latest version of each logical sector among the existing chaotic update block to be compacted.
STEP 564: Record the gathered sectors onto the new update block to form a new update block having compacted sectors.
STEP 566: Replace the existing update block with the new update block having compacted sectors.
STEP 568: Erase the closed out update block
2. Unwritten: No logical sector in the Logical Group has ever been written. The Logical Group is marked as unwritten in a group address table and has no allocated metablock. A predefined data pattern is returned in response to a host read for every sector within this group.
4. Chaotic Update: Some sectors within the Logical Group have been written in logically non-sequential order in a metablock, possibly using page tag, so that they supersede the corresponding logical sectors from any previous Intact state of the group. A sector within the group may be written more than once, with the latest version superseding all previous versions.
2. Sequential Update: The metablock is partially written with sectors in logically sequential order, possibly using page tag. All the sectors belong to the same Logical Group.
3. Chaotic Update: The metablock is partially or fully written with sectors in logically non-sequential order. Any sector can be written more than once. All sectors belong to the same Logical Group.
The open update block list 614 is the set of block entries in the ABL with the attributes of Open Update Block. The open update block list has one entry for each data update block currently open. Each entry holds the following information. LG is the logical group address the current update metablock is dedicated to. Sequential/Chaotic is a status indicating whether the update block has been filled with sequential or chaotic update data. MB is the metablock address of the update block. Page tag is the starting logical sector recorded at the first physical location of the update block. Number of sectors written indicates the number of sectors currently written onto the update block. MB0 is the metablock address of the associated original block. Page Tag0 is the page tag of the associated original block.
The closed update block list 616 is a subset of the Allocation Block List (ABL). It is the set of block entries in the ABL with the attributes of Closed Update Block. The closed update block list has one entry for each data update block which has been closed, but whose entry has not been updated in a logical to a main physical directory. Each entry holds the following information. LG is the logical group address the current update block is dedicated to. MB is the metablock address of the update block. Page tag is the starting logical sector recorded at the first physical location of the update block. MB0 is the metablock address of the associated original block.
A sequential update block has the data stored in logically sequential order, thus any logical sector among the block can be located easily. A chaotic update block has its logical sectors stored out of order and may also store multiple update generations of a logical sector. Additional information must be maintained to keep track of where each valid logical sector is located in the chaotic update block.
In the preferred embodiment, chaotic block indexing data structures allow tracking and fast access of all valid sectors in a chaotic block. Chaotic block indexing independently manages small regions of logical address space, and efficiently handles system data and hot regions of user data. The indexing data structures essentially allow indexing information to be maintained in flash memory with infrequent update requirement so that performance is not significantly impacted. On the other hand, lists of recently written sectors in chaotic blocks are held in a chaotic sector list in controller RAM. Also, a cache of index information from flash memory is held in controller RAM in order to minimize the number of flash sector accesses for address translation. Indexes for each chaotic block are stored in chaotic block index (CBI) sectors in flash memory.
The chaotic block index field within a CBI sector contains an index entry for each logical sector within a logical group or sub-group mapped to a chaotic update block. Each index entry signifies an offset within the chaotic update block at which valid data for the corresponding logical sector is located. A reserved index value indicates that no valid data for the logical sector exists in the chaotic update block, and that the corresponding sector in the associated original block is valid. A cache of some chaotic block index field entries is held in controller RAM.
The chaotic block info field within a CBI sector contains one entry for each chaotic update block that exists in the system, recording address parameter information for the block. Information in this field is only valid in the last written sector in the CBI block. This information is also present in data structures in RAM.
The entry for each chaotic update block includes three address parameters. The first is the logical address of the logical group (or logical group number) associated with the chaotic update block. The second is the metablock address of the chaotic update block. The third is the physical address offset of the last sector written in the chaotic update block. The offset information sets the start point for scanning of the chaotic update block during initialization, to rebuild data structures in RAM.
The sector index field contains an entry for each valid CBI sector in the CBI block. It defines the offsets within the CBI block at which the most recently written CBI sectors relating to each permitted chaotic update block are located. A reserved value of an offset in the index indicates that a permitted chaotic update block does not exist.
STEP 650: Begin locating a given logical sector of a given logical group.
STEP 652: Locate last written CBI sector in the CBI block
STEP 654: Locate the chaotic update block or original block associated with the given logical group by looking up the Chaotic Block Info field of the last written CBI sector. This step can be performed any time just before STEP 662.
STEP 658: If the last written CBI sector is directed to the given logical group, the CBI sector is located. Proceed to STEP 662. Otherwise, proceed to STEP 660.
STEP 660: Locate the CBI sector for the given logical group by looking up the sector index field of the last written CBI sector.
STEP 662: Locate the given logical sector among either the chaotic block or the original block by looking up the Chaotic Block Index field of the located CBI sector.
In the preferred embodiment, an indirect indexing scheme is employed to facilitate management of the index. Each entry of the sector index has direct and indirect fields.
The direct sector index defines the offsets within the CBI block at which all possible CBI sectors relating to a specific chaotic update block are located. Information in this field is only valid in the last written CBI sector relating to that specific chaotic update block. A reserved value of an offset in the index indicates that the CBI sector does not exist because the corresponding logical subgroup relating to the chaotic update block either does not exist, or has not been updated since the update block was allocated.
The indirect sector index defines the offsets within the CBI block at which the most recently written CBI sectors relating to each permitted chaotic update block are located. A reserved value of an offset in the index indicates that a permitted chaotic update block does not exist.
STEP 670: Partition each Logical Group into multiple subgroups and assign a CBI sector to each subgroup
STEP 680: Begin locating a given logical sector of a given subgroup of a given logical group.
STEP 682: Locate the last written CBI sector in the CBI block.
STEP 684: Locate the chaotic update block or original block associated with the given subgroup by looking up the Chaotic Block Info field of the last written CBI sector. This step can be performed any time just before STEP 696.
STEP 686: If the last written CBI sector is directed to the given logical group, proceed to STEP 691. Otherwise, proceed to STEP 690.
STEP 690: Locate the last written of the multiple CBI sectors for the given logical group by looking up the Indirect Sector Index field of the last written CBI sector.
STEP 691: At least a CBI sector associate with one of the subgroups for the given logical group has been located. Continue.
STEP 692: If the located CBI sector directed to the given subgroup, the CBI sector for the given subgroup is located. Proceed to STEP 696. Otherwise, proceed to STEP 694.
STEP 694: Locate the CBI sector for the given subgroup by looking up the direct sector index field of the currently located CBI sector.
STEP 696: Locate the given logical sector among either the chaotic block or the original block by looking up the Chaotic Block Index field of the CBI sector for the given subgroup.
In order to locate the ith sector in the subgroup B, the last written CBI sector in the CBI block 620 is first located. The chaotic block info field of the last written CBI sector provides the address to locate the chaotic update block 704 for the given logical group. At the same time it provides the location of the last sector written in the chaotic block. This information is useful in the event of scanning and rebuilding indices.
If the last written CBI sector turns out to be one of the four CBI sectors of the given logical group, it will be further determined if it is exactly the CBI sector for the given subgroup B that contains the ith logical sector. If it is, then the CBI sector's chaotic block index will point to the metablock location for storing the data for the ith logical sector. The sector location could be either in the chaotic update block 704 or the original block 702.
If the last written CBI sector turns out to be one of the four CBI sectors of the given logical group but is not exactly for the subgroup B, then its direct sector index is looked up to locate the CBI sector for the subgroup B. Once this exact CBI sector is located, its chaotic block index is looked up to locate the ith logical sector among the chaotic update block 704 and the original block 702.
If the last written CBI sector turns out not to be anyone of the four CBI sectors of the given logical group, its indirect sector index is looked up to locate one of the four. In the example shown in
Similar consideration applies to locating the jth logical sector in subgroup C of the given logical group. The example shows that the last written CBI sector turns out not to be any one of the four CBI sectors of the given logical group. Its indirect sector index points to one of the four CBI sectors for the given group. The last written of four pointed to also turns out to be exactly the CBI sector for the subgroup C. When its chaotic block index is looked up, the jth logical sector is found to be located at a designated location in the chaotic update block 704.
A list of chaotic sectors exists in controller RAM for each chaotic update block in the system. Each list contains a record of sectors written in the chaotic update block since a related CBI sector was last updated in flash memory. The number of logical sector addresses for a specific chaotic update block, which can be held in a chaotic sector list, is a design parameter with a typical value of 8 to 16. The optimum size of the list is determined as a tradeoff between its effects on overhead for chaotic data-write operations and sector scanning time during initialization.
During system initialization, each chaotic update block is scanned as necessary to identify valid sectors written since the previous update of one of its associated CBI sectors. A chaotic sector list in controller RAM for each chaotic update block is constructed. Each block need only be scanned from the last sector address defined in its chaotic block info field in the last written CBI sector.
When a chaotic update block is allocated, a CBI sector is written to correspond to all updated logical sub-groups. The logical and physical addresses for the chaotic update block are written in an available chaotic block info field in the sector, with null entries in the chaotic block index field. A chaotic sector list is opened in controller RAM.
When a chaotic update block is closed, a CBI sector is written with the logical and physical addresses of the block removed from the chaotic block info field in the sector. The corresponding chaotic sector list in RAM becomes unused.
The corresponding chaotic sector list in controller RAM is modified to include records of sectors written to a chaotic update block. When a chaotic sector list in controller RAM has no available space for records of further sector writes to a chaotic update block, updated CBI sectors are written for logical sub-groups relating to sectors in the list, and the list is cleared.
When the CBI block 620 becomes full, valid CBI sectors are copied to an allocated erased block, and the previous CBI block is erased.
The logical to physical address translation module 140 shown in
The hierarchy of address records for logical groups includes the open update block list, the closed update block list in RAM and the group address table (GAT) maintained in flash memory.
The open update block list is a list in controller RAM of data update blocks which are currently open for writing updated host sector data. The entry for a block is moved to the closed update block list when the block is closed. The closed update block list is a list in controller RAM of data update blocks which have been closed. A subset of the entries in the list is moved to a sector in the Group Address Table during a control write operation.
The Group Address Table (GAT) is a list of metablock addresses for all logical groups of host data in the memory system. The GAT contains one entry for each logical group, ordered sequentially according to logical address. The nth entry in the GAT contains the metablock address for the logical group with address n. In the preferred embodiment, it is a table in flash memory, comprising a set of sectors (referred to as GAT sectors) with entries defining metablock addresses for every logical group in the memory system. The GAT sectors are located in one or more dedicated control blocks (referred to as GAT blocks) in flash memory.
As described earlier, a GAT block contains entries for a logically contiguous set of groups in a region of logical address space. GAT sectors within a GAT block each contain logical to physical mapping information for 128 contiguous logical groups. The number of GAT sectors required to store entries for all logical groups within the address range spanned by a GAT block occupy only a fraction of the total sector positions in the block. A GAT sector may therefore be updated by writing it at the next available sector position in the block. An index of all valid GAT sectors and their position in the GAT block is maintained in an index field in the most recently written GAT sector. The fraction of the total sectors in a GAT block occupied by valid GAT sectors is a system design parameter, which is typically 25%. However, there is a maximum of 64 valid GAT sectors per GAT block. In systems with large logical capacity, it may be necessary to store GAT sectors in more than one GAT block. In this case, each GAT block is associated with a fixed range of logical groups.
A GAT update is performed as part of a control write operation, which is triggered when the ABL runs out of blocks for allocation (see
A GAT rewrite operation occurs during a control write operation when no sector location is available for an updated GAT sector. A new GAT block is allocated, and valid GAT sectors as defined by the GAT index are copied in sequential order from the full GAT block. The full GAT block is then erased.
A GAT cache is a copy in controller RAM 130 of entries in a subdivision of the 128 entries in a GAT sector. The number of GAT cache entries is a system design parameter, with typical value 32. A GAT cache for the relevant sector subdivision is created each time an entry is read from a GAT sector. Multiple GAT caches are maintained. The number is a design parameter with a typical value of 4. A GAT cache is overwritten with entries for a different sector subdivision on a least-recently-used basis.
The erase block manager 160 shown in
In the preferred embodiment, the controller RAM 130 holds the allocation block list (ABL) 610 and a cleared block list (CBL) 740. As described earlier in connection with
The allocation block list (ABL) keeps track of a pool of erased metablocks and the allocation of the erased metablocks to be an update block. Thus, each of these metablocks that may be described by an attribute designating whether it is an erased block in the ABL pending allocation, an open update block, or a closed update block.
The MAP block 750 is a metablock dedicated to storing erase management records in flash memory 200. The MAP block stores a time series of MAP block sectors, with each MAP sector being either an erase block management (EBM) sector 760 or a MAP sector 780. As erased blocks are used up in allocation and recycled when a metablock is retired, the associated control and directory data is preferably contained in a logical sector which may be updated in the MAP block, with each instance of update data being recorded to a new block sector. Multiple copies of EBM sectors 760 and MAP sectors 780 may exist in the MAP block 750, with only the latest version being valid. An index to the positions of valid MAP sectors is contained in a field in the EMB block. A valid EMB sector is always written last in the MAP block during a control write operation. When the MAP block 750 is full, it is compacted during a control write operation by rewriting all valid sectors to a new block location. The full block is then erased.
Each EBM sector 760 contains erased block lists (EBL) 770, which are lists of addresses of a subset of the population of erased blocks. The erased block lists (EBL) 770 act as a buffer containing erased metablock numbers, from which metablock numbers are periodically taken to re-fill the ABL, and to which metablock numbers are periodically added to re-empty the CBL. The EBL 770 serves as buffers for the available block buffer (ABB) 772, the erased block buffer (EBB) 774 and the cleared block buffer (CBB) 776.
The available block buffer (ABB) 772 contains a copy of the entries in the ABL 610 immediately following the previous ABL fill operation. It is in effect a backup copy of the ABL just after an ABL fill operation.
The erased block buffer (EBB) 774 contains erased block addresses which have been previously transferred either from MAP sectors 780 or from the CBB list 776 (described below), and which are available for transfer to the ABL 610 during an ABL fill operation.
The cleared block buffer (CBB) 776 contains addresses of erased blocks which have been transferred from the CBL 740 during a CBL empty operation and which will be subsequently transferred to MAP sectors 780 or to the EBB list 774.
Each of the MAP sectors 780 contains a bitmap structure referred to as MAP. The MAP uses one bit for each metablock in flash memory, which is used to indicate the erase status of each block. Bits corresponding to block addresses listed in the ABL, CBL, or erased block lists in the EBM sector are not set to the erased state in the MAP.
Any block which does not contain valid data structures and which is not designated as an erased block within the MAP, erased block lists, ABL or CBL is never used by the block allocation algorithm and is therefore inaccessible for storage of host or control data structures. This provides a simple mechanism for excluding blocks with defective locations from the accessible flash memory address space.
The hierarchy shown in
The algorithms adopted for updating the hierarchy of erased metablock records results in erased blocks being allocated for use in an order which interleaves bursts of blocks in address order from the MAP block 750 with bursts of block addresses from the CBL 740 which reflect the order blocks were updated by the host. For most metablock sizes and system memory capacities, a single MAP sector can provide a bitmap for all metablocks in the system. In this case, erased blocks are always allocated for use in address order as recorded in this MAP sector.
As described earlier, the ABL 610 is a list with address entries for erased metablocks which may be allocated for use, and metablocks which have recently been allocated as data update blocks. The actual number of block addresses in the ABL lies between maximum and minimum limits, which are system design variables. The number of ABL entries formatted during manufacturing is a function of the card type and capacity. In addition, the number of entries in the ABL may be reduced near the end of life of the system, as the number of available erased blocks is reduced by failure of blocks during life. For example, after a fill operation, entries in the ABL may designate blocks available for the following purposes. Entries for Partially written data update blocks with one entry per block, not exceeding a system limit for a maximum of concurrently opened update blocks. Between one to twenty entries for Erased blocks for allocation as data update blocks. Four entries for erased blocks for allocation as control blocks.
As the ABL 610 becomes depleted through allocations, it will need to be refilled. An operation to fill the ABL occurs during a control write operation. This is triggered when a block must be allocated, but the ABL contains insufficient erased block entries available for allocation as a data update block, or for some other control data update block. During a control write, the ABL fill operation is concurrent with a GAT update operation.
The following actions occur during an ABL fill operation.
1. ABL entries with attributes of current data update blocks are retained.
2. ABL entries with attributes of closed data update blocks are retained, unless an entry for the block is being written in the concurrent GAT update operation, in which case the entry is removed from the ABL.
3. ABL entries for unallocated erase blocks are retained.
4. The ABL is compacted to remove gaps created by removal of entries, maintaining the order of entries.
5. The ABL is completely filled by appending the next available entries from the EBB list.
6. The ABB list is over-written with the current entries in the ABL.
The CBL is a list of erased block addresses in controller RAM with the same limitation on the number of erased block entries as the ABL. An operation to empty the CBL occurs during a control write operation. It is therefore concurrent with an ABL fill/GAT update operations, or CBI block write operations. In a CBL empty operation, entries are removed from the CBL 740 and written to the CBB list 776.
A MAP exchange operation between the erase block information in the MAP sectors 780 and the EBM sectors 760 may occur periodically during a control write operation, when the EBB list 774 is empty. If all erased metablocks in the system are recorded in the EBM sector 760, no MAP sector 780 exists and no MAP exchange is performed. During a MAP exchange operation, a MAP sector feeding the EBB 774 with erased blocks is regarded as a source MAP sector 782. Conversely, a MAP sector receiving erased blocks from the CBB 776 is regarded as a destination MAP sector 784. If only one MAP sector exists, it acts as both source and destination MAP sector, as defined below.
The following actions are performed during a MAP exchange.
1. A source MAP sector is selected, on the basis of an incremental pointer.
2. A destination MAP sector is selected, on the basis of the block address in the first CBB entry that is not in the source MAP sector.
3. The destination MAP sector is updated, as defined by relevant entries in the CBB, and the entries are removed from the CBB.
4. The updated destination MAP sector is written in the MAP block, unless no separate source MAP sector exists.
5. The source MAP sector is updated, as defined by relevant entries in the CBB, and the entries are removed from the CBB.
6. Remaining entries in the CBB are appended to the EBB.
7. The EBB is filled to the extent possible with erased block addresses defined from the source MAP sector.
8. The updated source MAP sector is written in the MAP block.
9. An updated EBM sector is written in the MAP block.
[C] When an ABL entry is created with Open Update Block attributes, an Associated Original Block field is added to the entry to record the original metablock address for the logical group being updated. This information is obtained from the GAT.
[O] Subsequent to [N] during a MAP exchange operation, entries other than those moved in [M] are moved from the MAP source sector to fill the EBB list, if possible.
To locate a logical sector's physical location in flash memory, the logical to physical address translation module 140 shown in
STEP 800: A logical sector address is given.
STEP 810: Look up given logical address in the open update blocks list 614 (see
STEP 820: Look up given logical address in the closed update block list 616. If lookup fails, the given logical address is not part of any update process; proceed to STEP 870 for GAT address translation. Otherwise proceed to STEP 860 for closed update block address translation.
STEP 830: If the update block containing the given logical address is sequential, proceed to STEP 840 for sequential update block address translation. Otherwise proceed to STEP 850 for chaotic update block address translation.
STEP 840: Obtain the metablock address using sequential update block address translation. Proceed to STEP 880.
STEP 850: Obtain the metablock address using chaotic update block address translation. Proceed to STEP 880.
STEP 860: Obtain the metablock address using closed update block address translation. Proceed to STEP 880.
STEP 870: Obtain the metablock address using group address table (GAT) translation. Proceed to STEP 880.
STEP 880: Convert the Metablock Address to a physical address. The translation method depends on whether the metablock has been relinked.
STEP 890: Physical sector address obtained.
The various address translation processes are described in more detail as follows:
Address translation for a target logical sector address in a logical group associated with a sequential update block can be accomplished directly from information in the open update block list 614 (
The address translation sequence for a target logical sector address in a logical group associated with a chaotic update block is as follows.
1. If it is determined from the chaotic sector list in RAM that the sector is a recently written sector, address translation may be accomplished directly from its position in this list.
2. The most recently written sector in the CBI block contains, within its chaotic block data field, the physical address of the chaotic update block relevant to the target logical sector address. It also contains, within its indirect sector index field, the offset within the CBI block of the last written CBI sector relating to this chaotic update block (see
3. The information in these fields is cached in RAM, eliminating the need to read the sector during subsequent address translation.
4. The CBI sector identified by the indirect sector index field at step 3 is read.
5. The direct sector index field for the most recently accessed chaotic update sub-group is cached in RAM, eliminating the need to perform the read at step 4 for repeated accesses to the same chaotic update block.
6. The direct sector index field read at step 4 or step 5 identifies in turn the CBI sector relating to the logical sub-group containing the target logical sector address.
7. The chaotic block index entry for the target logical sector address is read from the CBI sector identified in step 6.
8. The most recently read chaotic block index field may be cached in controller RAM, eliminating the need to perform the reads at step 4 and step 7 for repeated accesses to the same logical sub-group.
9. The chaotic block index entry defines the location of the target logical sector either in the chaotic update block or in the associated original block. If the valid copy of the target logical sector is in the original block, it is located by use of the original metablock and page tag information.
Address translation for a target logical sector address in a logical group associated with a closed update block can be accomplished directly from information in the closed block update list (see
1. The metablock address assigned to the target logical group is read from the list.
2. The sector address within the metablock is determined from the “page tag” field in the list.
If a logical group is not referenced by either the open or closed block update lists, its entry in the GAT is valid. The address translation sequence for a target logical sector address in a logical group referenced by the GAT is as follows.
1. The ranges of the available GAT caches in RAM are evaluated to determine if an entry for the target logical group is contained in a GAT cache.
2. If the target logical group is found in step 1, the GAT cache contains full group address information, including both metablock address and page tag, allowing translation of the target logical sector address.
3. If the target address is not in a GAT cache, the GAT index must be read for the target GAT block, to identify the location of the GAT sector relating to the target logical group address.
4. The GAT index for the last accessed GAT block is held in controller RAM, and may be accessed without need to read a sector from flash memory.
5. A list of metablock addresses for every GAT block, and the number of sectors written in each GAT block, is held in controller RAM. If the required GAT index is not available at step 4, it may therefore be read immediately from flash memory.
6. The GAT sector relating to the target logical group address is read from the sector location in the GAT block defined by the GAT index obtained at step 4 or step 6. A GAT cache is updated with the subdivision of the sector containing the target entry.
7. The target sector address is obtained from the metablock address and “page tag” fields within the target GAT entry.
If a flag associated with the metablock address indicates that the metablock has been re-linked, the relevant LT sector is read from the BLM block, to determine the erase block address for the target sector address. Otherwise, the erase block address is determined directly from the metablock address.
In another embodiment, a scratch pad block (“SPB”) is implemented to buffer data written to an update block. Update data to a non-volatile memory may be recorded in at least two interleaving streams such as either into an update block or a scratch pad block depending on a predetermined condition. The scratch pad block is used to buffered update data that are ultimately destined for the update block. In a preferred embodiment, an index (“SPBI/CBI”) of the data stored in the scratch pad block as well that stored in the update block is kept in the data structures of the controller RAM. These data structures allow tracking and fast access of all valid sectors in the scratch pad block (SPB) and the chaotic blocks. At appropriate time, as described below, the SPBI/CBI data will be saved in an unused portion of a page of the scratch pad block.
In this embodiment, the address translation shown in
Implementation of scratch pad block has been described in U.S. Application Publication No. US-2006-0155922-A1 published Jul. 13, 2006, entitled “Non-Volatile Memory And Method With Improved Indexing For Scratch Pad And Update Blocks”, by Gorobets et al.
Data update management operations are performed in RAM on the ABL, the CBL and the Scratch Pad Sector List/Chaotic Sector List. The ABL is updated when an erased block is allocated as an update block or a control block, or when an update block is closed. The CBL is updated when a control block is erased or when an entry for a closed update block is written to the GAT. The Scratch Pad Sector List is updated when sectors are written to a scratch pad block. The update chaotic sector list is updated when a sector is written to a chaotic update block. It will be understood, the sector here is an example of a unit of write data, which is also referred to as a page.
A control write operation causes information from control data structures in RAM to be written to control data structures in flash memory, with consequent update of other supporting control data structures in flash memory and RAM, if necessary. It is triggered either when the ABL contains no further entries for erased blocks to be allocated as update blocks, or when the SP block is rewritten.
In the preferred embodiment, the ABL fill operation, the CBL empty operation and the EBM sector update operation are performed during every control write operation. When the MAP block containing the EBM sector becomes full, valid EBM and MAP sectors are copied to an allocated erased block, and the previous MAP block is erased.
One GAT sector is written, and the Closed Update Block List is modified accordingly, during every control write operation.
A GAT block rewrite takes place when a GAT block becomes full and the data in the full block will be relocated to an allocated erased block.
A SPBI/CBI sector is written, after certain chaotic sector write operations.
A SPB block rewrite takes place when the SPBI/CBI block becomes full. Valid SPBI/CBI sectors are copied to an allocated erased block, and the previous SPB block is erased.
A MAP exchange operation, as described earlier, is performed when there are no further erased block entries in the EBB list in the EBM sector.
A MAP Block rewrite takes place when the MAP block becomes full and valid EBM and MAP sectors are copied to an allocated erased block, and the previous MAP block is erased.
A Boot sector is written in a current Boot block on each occasion the MAP block is moved.
A Boot Block rewrite takes place when the boot block becomes full. The valid Boot sector is copied from the current version of the Boot block to the backup version, which then becomes the current version. The previous current version is erased and becomes the backup version, and the valid Boot sector is written back to it.
Example of control data are the directory information and block allocation information associated with the memory block management system, such as those described in connection with
As described in connection with
There are certain memory applications where the data is intended to be committed to memory not to be updated again. Therefore memory devices for these applications, referred to as one-time programmable memory or OTP memory devices, need not provide erase and reprogram facilities. OTP memory devices can have simplified block management system, thereby reducing complexity and overheads.
The block management system described herein is compatible with the implementation of a OTP memory device. Essentially for the OTP memory, each block is treated as a unit of memory storage. The difference with the erasable block management system described is that the blocks are not erased. However, the techniques of pre-emptive relocation of data from one block to another is equally applicable to OTP memory.
OTP memory systems has been described in U.S. Application Publication No. US-2006-0047920-A1 published Mar. 2, 2006, entitled “Method and Apparatus for Using a One-Time or Few-Time Programmable Memory with a Host Device designated for Erasable/Rewriteable Memory.”
It has also been described earlier that a hierarchy of control data exists, with the ones in the lower hierarchy being updated more often than those higher up. For example, assuming that every control block has N control sectors to write, the following sequence of control updates and control block relocations, normally happens. Referring to
Since the hierarchy is formed by the BOOT control data at the top followed by MAP and then GAT, thus, in some instances after a GAT update there will be a “cascade control update”, where all of the GAT, MAP and BOOT blocks would be relocated. In the case when GAT update is caused by a Chaotic or Sequential Update block closure as a result of a host write, there will also be a garbage collection operation (i.e., relocation or rewrite.) In that case of Chaotic Update Block garbage collection, a SPBI/CBI index would be updated, and that can also trigger a SP block relocation. Thus, in this extreme situation, a large number of metablocks need be garbage collected at the same time.
In can be seen that each control data block of the hierarchy has its own periodicity in terms of getting filled and being relocated. If each proceeds normally, there will be times when the phases of a large number of the blocks will line up and trigger a massive relocation or garbage collection involving all those blocks at the same time. Relocation of many control blocks will take a long time and should be avoided as some hosts do not tolerate long delays caused by such massive control operations.
For example, this undesirable situation can happen when updating control data used for controlling the operation of the block management system. A hierarchy of control data type can exist with varying degree of update frequencies, resulting in their associated update blocks requiring garbage collection or relocation at different rates. There will be certain times that the garbage collection operations of more than one control data types coincide. In the extreme situation, the relocation phases of the blocks for all control data types could line up, resulting in all of the blocks requiring relocation or rewrite at the same time.
One solution to avoid cascade relocation of data has been described in U.S. Application Publication No. US-2005-0144365-A1 published Jun. 30, 2005, entitled “Non-Volatile Memory and Method with Control Data Management,” by Gorobets et al. In a nonvolatile memory with a block management system, a preemptive relocation of a memory block or controlled rewrite is implemented to avoid the situation where a large number of control update blocks all happen to need relocation at the same time. This undesirable situation is avoided by whenever a current host operation can also accommodate a housekeeping operation, a preemptive relocation of a control block takes place in advance the block being totally filled. In particular, priority is given to the block with a data type having the fastest fill rate. The method can be regarded as introducing some sort of dithering to the overall mix of things in order to avoid alignment of the phases of the various blocks in question. Thus, whenever an opportunity arises, a fast-filling block that has a slight margin from being totally filled is to be relocated preemptively.
As described earlier, rewrite operations will be necessary where a block containing control data is full. After undergoing a series of updates, the filled block typically contains valid data as well as obsolete data. The valid data will be copied to another block with empty space. This relocation operation is a garbage collection operation where the full block is erased and recycled after its valid data are salvaged and copied to another block. Another reason for relocation is when a defect has been encountered in a block, rendering the block unusable. This is particular true for those defects that requires excessive error correction by a built-in error correction code or that simply cannot be corrected. Yet another reason for relocation is the need to ensure uniform usage of all blocks in the memory so that no block gets excessive erase/program cycling to wear out prematurely.
The relocation operations mentioned above are all examples of a system housekeeping operation. Relocation of data from one block to another is typically relatively time consuming as it involves reading and writing substantial amount of data. The housekeeping operations can be performed in the background when a host is not actively engaging the memory. However, while it is ongoing, a host is excluded from sending a command to the memory and may even power down the memory thereby interrupting the ongoing housekeeping operation. A preferred way is to perform the housekeeping operations in the foreground, contemporaneously with the memory executing a host command.
U.S. Application Publication No. US-2006-0161728-A1 published Jul. 20, 2006, entitled “Scheduling of Housekeeping Operations in Flash Memory Systems,” by Bennett et al discloses execution of a host command performed together with one or more such housekeeping operations within a time budget established for executing the particular host command. In particular, one such host command is to write data being received to the memory. One such housekeeping operation is to level out the wear of the individual blocks that accumulates through repetitive erasing and re-programming.
According to the present invention, an improved scheme is provided to avoid possible lengthy cascade updates of the control data. This is accomplished by setting a block margin for each type of control data and rewriting the block at the earliest opportunity when the block margin has been reached. In particular, the margin is set just sufficient to accommodate data accumulated in a predetermined interval before the rewrite can take place so as not to totally fill the block before the rewrite can take place. The predetermined interval is determined, among other things, by considering a host write pattern that yields a worst-case interval before the rewrite can take place. Other considerations for setting the margin include the time required for each control block rewrite and the time available for control block rewrites based on the configuration of the update blocks for storing host data, the time required in the foreground host operation and the host write latency.
The improvement also makes allowance for multiple program errors per the cascade control update, so that it is able to handle more than one ECC or program error occurring one soon after another within the timing limitation. This feature is particularly important for one-time programmable (“OTP”) memory since the risk is quite high if the defects are not patched on the lower level. The improvement also enables a minimum of blocks to be reserved in a pool of update blocks for storing control data. The reserved blocks enable the memory control system to handle the worst cascade update where all control data blocks can potentially be filled at the same time, and must all be rewritten in the same busy period. If fewer blocks are required to be reserved for control data, more blocks will be available for host data updates.
The advantages of the invention include the following. An increased number of errors can be handled in the worst-case update sequence. A worst-case of a longest combination of garbage collections (GC) and control block compaction can be avoided. For example, Chaotic GC takes longer than Sequential GC, so by avoiding doing control updates at the same time as Chaotic GC the worst case command latency can be reduced. Optimized performance is obtained by optimum selection of the block margins (e.g., by selecting a fuller control block to compact) and scheduling of an internal operation to perform. Reduced number of reserved erased blocks is required to handle the worst case update sequence. Errors can be handled much quicker in the cases of pre-emptive internal operations as the error handling can be rescheduled. Partial error handling and schedule completion of the error handling is possible. It is possible to schedule ECC error handling during read operation, which has short latency, to be done later (e.g., during next write operation.)
In a typical host command such as write host data, the host specifies a timeout or write latency designed to accommodate the worst-case situation for the memory to complete the command. The actual duration for the memory to execute the command depends on the state of the memory block the data is being written to. In particular it depends on whether the writing includes additional time-consuming data relocation between blocks. These data relocation is caused by closure of a block in response to a new block being allocated. The closure of a block typically requires a garbage collection before being erased and recycled.
A block is typically closed after it is full or when data are no longer written to it for some reason. Another factor that affects the timing of a block getting closed is how a pool of update blocks opened for updates concurrently is configured. Since there is a limit on the number of blocks in the pool, an existing block must be closed if a new block is introduced into a fully populated pool.
The practical system limitation of supporting up to a maximum number of concurrently opened update blocks has been described earlier. For example, in one embodiment described in connection with
Alternative selection of which update blocks in a full pool to close has been disclosed in U.S. patent application Ser. No. 11/532,456 filed Sep. 15, 2006, entitled “Method For Class-Based Update Block Replacement Rules In Non-Volatile Memory,” by Jason Lin.
In the event that a new update block needs to be allocated, one of the existing update blocks in the update pool will need to be closed to make room. For example in the event when the host writes sequential data for a logical group of sectors not serviced by the existing update blocks in the pool, a new update block will need to be allocated for recording the data.
As a control block becomes full, an internal rewrite operation relocates valid data from it to a new block which replaces it in the pool. In some implementation, a number of erased blocks 1406 are reserved in the pool in case a cascade of rewrites takes place at the same time.
The worst-case cascade update is when the Boot block, the MAP block, the Scratch Pad block and the GAT block are rewritten in the same busy period. Compound to this, the cascade update could also coincide with an update block garbage collection during a host write. In order to avoid such cascade updates, when the control blocks are nearly full, they will be rewritten at the earliest available opportunity so that in a worst case scenario, there will always be enough time to rewrite the control blocks preemptively before being forced to rewrite them as a result of a host write with critical timing.
In the case where more than one control block becomes nearly full at the same time, control blocks will be rewritten in a predetermined order to ensure that there is always a free reserved block 1406 available for the update, and that an update to one control block will not trigger the rewrite of another control block, forcing the cascade.
In one implementation, when there are more than one control block rewrites pending, the one with a control data type that is more active is preferentially executed in the next available opportunity found in a host operation. In this way, a minimum of reserved blocks need be set aside as resource for the control block rewrites as only one control block rewrite will take place at a time.
By ensuring that the threshold is set to allow the worst case host write pattern to happen, a cascade update will be assuredly avoided at any time. The thresholds for each of the control data blocks (e.g., MAP, GAT, SPB and BOOT) are set with a margin of a predetermined number of pages from end. The exact margin for each of the blocks will be dependant on the cascade avoidance mechanism used.
The worst case scenario is compounded by the maximum amount of data pages to transfer during each control block rewrite and the worst case of host write pattern that results in the least opportunity for piggy-backing control block rewrite during the host write.
Using specific example memory systems given earlier, the worst case from the point of amount of data page to transfer during each control block rewrite are as follows. A MAP block rewrite involves copying a maximum of 8 MAP sectors and the EBM sector. If each sector is written in a page, there will be 9 pages to be copied to the new MAP block. A GAT block rewrite involves copying 64 GAT sectors as 16 pages copied to the new GAT block, plus an EBM update of 1 page to the MAP block, which amounts to 8 pages to be copied plus one page to be written. A Scratch Pad block rewrite involves copying 8 Scratch Pad pages (assume there are 8 pages in the update pool) of buffered host data to the new SP block, plus 1 Scratch Pad index update on the new SP block, and 1 EBM update of 1 page to the MAP block, which amounts to 8 pages to be copied and 2 pages to be written. Boot block rewrite involves copying 8 LT sectors, 8 SPBL sectors, and the Boot Sector, which amounts to 17 pages. If two copies of the Boot block are maintained in the memory, the copies are repeated.
Each of the four types of control data block rewrites will require significant time to complete. In an implementation where the update pool is less than 8, or the host data do not needed to be buffered as much, the SPB block rewrites may be faster than the others since there are relatively less pages to copy.
The ideal time to perform a pre-emptive control block rewrite is to “piggy-back” onto the foreground execution of a host command. This is especially desirable when the new host command itself does not trigger a garbage collection so that there will be more time to perform the control block rewrites within the host command's latency period. However, in many instances a host command such as a host write will be executed along with additional garbage collection (sequential block close, or chaotic block consolidation.) In these instances, there will be less or even insufficient time to piggy-back a control block rewrite.
The case studies below of specific memory system and configurations will show that in a worst-case host write pattern, it is possible to get a sequential block close with every host write (see
To guarantee that cascade updates are avoided, at least one pre-emptive control block rewrite must be allowed in conjunction with a garbage collection. One method would be to allow one control block rewrite in conjunction with a sequential close, but not with a chaotic block consolidation, since a sequential close is generally a shorter operation. Generally, it is not possible to trigger many consolidations in a row. When there is no garbage collection triggered by the host command operation, the operation time can support up to two control block rewrites.
The method relies on rewriting control blocks at a convenient time, before they become absolutely full. These case studies aim to find the worst sequence of commands with respect to the overheads of garbage collection, and control updates that they trigger. This can then be used to define the order in which control blocks should be rewritten, and how much space must be reserved before they are considered nearly full.
Example calculations for typical update pool configurations (see
(1) Every write is a single sector write to the Scratch Pad (only 1 busy period, and at least 1 control block write)
(2) Each sequential close triggers a Scratch Pad update. This happens if there was valid host data in the Scratch Pad for this block.
(3) Each chaotic consolidation always triggers a Scratch Pad update.
(4) Each sequential to chaotic conversion triggers a Scratch Pad update.
(5) During the worst run, 1 of the MAP updates will involve a MAP exchange
(6) The update pool or Blocklist is always full, so every request for a new erased metablock triggers a Blocklist release (GAT, and MAP update)
(7) All GAT updates are to the same GAT block since updating different GAT blocks slows down the rate at which GAT blocks fill.
The initial state of the update pool has all 7 update blocks open, with 3 of them being chaotic update blocks. The host write pattern is such the host writes chaotically to each sequential update block, repeatedly opens a new sequential block and, on the next write, makes it go chaotic.
Step 0: Initial state
Step 1: Chaotic block 1 is closed which needs a new metablock (GAT and MAP update), and a write to the Scratch Pad. The new command makes sequential block 4 go chaotic (Scratch Pad update), and then the host data could be written to the Scratch Pad. Total control data triggered: 1 GAT update, 1 MAP update, and 3 Scratch Pad updates. Closing the chaotic block results in the valid data on it being relocated (or consolidated) into another block. According to the method, with this consolidation overhead, no rewrite of control blocks will be piggy-backed in this step.
Step 2-4: As step 1. So no rewrite of control blocks is possible.
Step 5: This is the first opportunity to do a pre-emptive rewrite. The new command opens a new sequential update block. There is a spare update block, so another block need not be closed, but the host write could go to the Scratch Pad. By this point 4 MAP updates, 4 GAT update, and 13 Scratch Pad updates could have been made.
Step 6: As step 1
Step 7: As step 5. By this point 6 MAP pages, 6 GAT pages, and 16 Scratch Pad pages could have been written.
Thus, it is possible to guarantee that the various control blocks get rewritten in time to avoid cascade by rewriting the MAP and GAT blocks in step 5, and the Scratch Pad in step 7, if a margin is set with 6 free MAP pages for the MAP block, with 4 free GAT pages for the GAT block, and with 16 free Scratch Pad pages for the SP block.
The initial state of the update pool has all 7 update blocks open, with 3 of them being chaotic update blocks and full. The host write pattern is such that the host writes chaotically to full chaotic blocks, and then repeatedly opens a sequential update block.
Step 0: Initial state
Step 1: Write to chaotic block 1 which is already full. This triggers a consolidation which triggers a GAT and MAP update and a Scratch Pad update. A new sequential update block is opened and the host data is written to the Scratch Pad.
Steps 2 and 3: As step 1
Step 4: This is the first opportunity to do a pre-emptive rewrite. The new command needs a new update block, which closes an existing sequential update block. The close needs a Scratch Pad write, and the new block needs a GAT and MAP update. By this point, 4 MAP updates, 4 GAT updates, and 8 Scratch Pad updates could have been done.
Steps 5 onward: As in step 4, each triggering a sequential close, and allocating a new block triggering 1 GAT, 1 MAP and 2 Scratch Pad updates.
Assuming the MAP block is rewritten in step 4, the GAT block in step 5 and the Scratch Pad in step 6, then the margin need be set with 6 pages in the MAP block, 5 pages in the GAT block, and 12 pages in the Scratch Pad block.
The initial state of the update pool has all 3 update blocks open, with 1 of them being chaotic update block. The host write pattern is such the host writes chaotically to each sequential update block, repeatedly opens a new sequential block and, on the next write, makes it go chaotic.
Step 0: Initial state
Step 1: Chaotic block 1 is closed which needs a new metablock (GAT and MAP update), and a write to the Scratch Pad. The new command makes sequential block 2 go chaotic (Scratch Pad update), and then the host data could be written to the Scratch Pad. Total 1 GAT page, 1 MAP page, and 3 Scratch Pad pages.
Step 2: As step 1
Step 3: This is the first opportunity to do a pre-emptive rewrite. The new command opens a new sequential update block. There is a spare update block, so another block need not be closed, but the host write could go to the Scratch Pad. By this point 3 MAP updates, 3 GAT updates, and 7 Scratch Pad updates could have been made.
Step 4: As step 1
Step 5: As step 3. By this point 6 MAP updates, 6 GAT updates, and 11 Scratch Pad updates could have been made.
Thus, it is possible to guarantee that the various control blocks get rewritten in time to avoid cascade by rewriting the MAP and GAT blocks in step 3, and the Scratch Pad in step 5, if a margin is set with 5 free MAP pages for the MAP block, with 3 free GAT pages for the GAT block, and with 11 free Scratch Pad pages for the SP block.
The initial state of the update pool has all 3 update blocks open, with 1 of them being chaotic update block and full. The host write pattern is such that the host writes chaotically to full chaotic blocks, and then repeatedly opens a sequential update block.
Step 0: Initial state
Step 1: Write to chaotic block 1 which is already full. This triggers a consolidation which triggers a GAT and MAP update and a Scratch Pad update. A new sequential update block is opened and the host data is written to the Scratch Pad. A total of 2 GAT updates, 2 MAP updates and 2 Scratch Pad updates could have been made by this point.
Step 2: This is the first opportunity to do a pre-emptive rewrite. The new command needs a new update block, which closes an existing sequential update block. The close needs a Scratch Pad write, and the new block needs a GAT and MAP update. By this point, 3 MAP updates, 3 GAT updates, and 6 Scratch Pad updates could have been done.
Steps 3 onward: As in step 2, each triggering a sequential close, and allocating a new block triggering 1 GAT, 1 MAP and 2 Scratch Pad updates.
Assuming the MAP block is rewritten in step 2, the GAT block in step 3 and the Scratch Pad in step 4, then the margin need be set with 5 pages in the MAP block, 4 pages in the GAT block, and 10 pages in the Scratch Pad block.
The initial state of the update pool has all 3 update blocks open, with 3 of them being chaotic update blocks. The host write pattern is such the host writes chaotically to each sequential update block, repeatedly opens a new sequential block and, on the next write, makes it go chaotic.
Step 0: Initial state
Step 1: Chaotic block 1 is closed which needs a new metablock (GAT and MAP update), and a write to the Scratch Pad. The new command needs a new metablock, (GAT and MAP updates), and goes to the Scratch Pad. A total of 2 GAT updates, 2 MAP updates, and 2 Scratch Pad updates could have been made by this point.
Steps 2 and 3: As step 1
Step 4: This is the first opportunity to do a pre-emptive rewrite. The new command opens a new sequential update block. There is a spare update block, so another block need not be closed, but the host write could go to the Scratch Pad. By this point 6 MAP updates, 6 GAT updates, and 8 Scratch Pad updates could have been done.
Steps 5 and 6: As step 4
Step 7: As step 1
Thus, it is possible to guarantee that the various control blocks get rewritten in time to avoid cascade by rewriting the MAP and GAT blocks in step 4, and the Scratch Pad in step 5, if the margin is set with 8 free MAP pages for the MAP block, with 6 free GAT pages for the GAT block, and with 10 free Scratch Pad pages for the SP block.
The initial state of the update pool has all 3 update blocks open, with 3 of them being chaotic update blocks and full. The host write pattern is such that the host writes chaotically to full chaotic blocks, and then repeatedly opens a sequential update block.
Step 0: Initial state
Step 1: Write to chaotic block 1 which is already full. This triggers a consolidation which triggers a GAT and MAP update and a Scratch Pad update. A new sequential update block is opened and the host data is written to the Scratch Pad. A total of 2 GAT updates, 2 MAP updates and 2 Scratch Pad updates could have been made by this point.
Steps 2 and 3: As step 1
Step 4: This is the first opportunity to do a pre-emptive rewrite. The new command needs a new update block, which closes an existing sequential update block. The close needs a Scratch Pad write, and the new block needs a GAT and MAP update. By this point, 7 MAP updates, 4 GAT updates, and 8 Scratch Pad updates could have been done.
Steps 5 on: As step 4, each triggering a sequential close, and allocating a new block triggering 1 GAT, 1 MAP and 2 Scratch Pad updates.
Thus, it is possible to guarantee that the various control blocks get rewritten in time to avoid cascade by writing the MAP block in step 4, the GAT block in step 5 and the Scratch Pad in step 6, if the margin is set with 9 pages in the MAP block, 5 pages in the GAT block, and 12 pages in the Scratch Pad block.
A program error during a data relocation operation is more critical since the time-consuming operation may need to be restarted again. One possible occurrence is during a chaotic block consolidation or a sequential block close triggered by a host command. Another possible occurrence is during a control block rewrite. The pre-emptive control block rewrite to avoid cascade will need to take these problems into consideration.
A program error during consolidation is handled in one of two ways. If the error happens near the start of the consolidation, then the consolidation is restarted using another block. If the error happens nearer the end of the consolidation, then the phased error block is used to store the remaining sectors. Phased program error handling has been disclosed in U.S. Application Publication No. US-2005-0166087-A1, published Jul. 28, 2005. If phased error is used, then the phased error block will be closed at the next convenient time and its data relocated to a non-defective block. This means that pre-emptive rewrites would be delayed. To account for this more sectors need to be reserved in the margin of each of the control blocks. A program error during sequential close is essentially handled in the same manner as that during a consolidation.
A program error may also occur during a control data update. One way of handling the error is to relocate the control data to a new control block. An alternative is to write the sector to the next available page in the control block. A flag could then be set so this block is rewritten at the next convenient time. This would require reserving an extra page in the margin of the control block.
A program error during a pre-emptive control block rewrite is handled by repeating the rewrite to another block. An alternative is to abandon the pre-emptive rewrite and attempt the rewrite again at the next convenient time.
In both cases, any other pending pre-emptive rewrites would be delayed. To account for this, extra sectors need to be reserved in the margin of each of the control blocks.
As mentioned earlier, a number of control block rewrite scheduling methods are possible depending on various timings of the host and memory systems. The following is some examples of the control block rewrite scheduling methods.
Method 1 is the method used to perform the calculations for the case studies illustrated in
1. Do no pre-emptive control block rewrites in the same busy period as a chaotic block consolidation.
2. Allow 1 pre-emptive control block rewrite in the same busy period as a sequential block close.
3. Allow 2 pre-emptive control block rewrites when there is no update block garbage collection.
Method 2 basically assumes that the host write latency allows sufficient time to do a short control block rewrite such as a Scratch Pad rewrite even if the host write has triggered a garbage collection.
1. Allow pre-emptive Scratch Pad rewrite at any time.
2. Allow 1 pre-emptive control block rewrite in same busy period as sequential close.
3. Allow 2 pre-emptive control block rewrites when there is no update block garbage collection.
Method 3 basically takes a more quantitatively approach by examining the amount of pages to relocate for each of the rewrites and if any of them could be executed within the remaining time set by the host write latency. This method will utilize the host write latency period most efficiently at the expense of micro-tracking the amount of relocation for each rewrite. The advantage is that the margins will be at a minimum.
1. Count work required for each control block rewrite (number of page copies).
2. Allow pre-emptive control block rewrites until total work done exceeds a defined threshold.
Method 4 is similar to Method 1 with the additional assumption that a chaotic close can be performed at the same speed as a sequential close.
1. Rewrite the control blocks in the order MAP->Scratch Pad->GAT->BB. This allows us to reserve 4 less pages in the SP
2. Allow 1 pre-emptive control block rewrite in the same busy period as either a sequential close or chaotic compaction.
3. Allow 2 pre-emptive control block rewrites when there is no update block garbage collection.
STEP 1410: Setting a margin of a number of empty memory units before a block is full for each type of data, wherein the margin is just sufficient to accommodate data accumulated in a predetermined interval before data in the block are allowed to relocate, and the predetermined interval is determined from a host write pattern that yields a worst-case interval before data in the block are allowed to relocate.
STEP 1430: In response to a block storing data reaching the margin for the data type, relocating data in the block to another block when allowed to do so. Go to STEP 1420 unless interrupted.
STEP 1410: Setting a margin of a number of empty memory units before a block is full for each type of data, wherein the margin is just sufficient to accommodate data accumulated in a predetermined interval before data in the block are allowed to relocate, and the predetermined interval is determined from a host write pattern that yields a worst-case interval before data in the block are allowed to relocate.
STEP 1430′: In response to a block storing data reaching the margin for the data type and having data type of a highest rank among any similar blocks, relocating data in the block to another block when allowed to do so. Go to STEP 1420 unless interrupted.
STEP 1410′: Setting a margin of a number of empty memory units before a block is full for each type of data, wherein the margin is just sufficient to accommodate data accumulated in a predetermined interval before data in the block are allowed to relocate, and the predetermined interval is determined from a host write pattern that yields a worst-case interval before the block of data is allowed to relocate and from the amount of data to relocate.
All patents, patent applications, articles, books, specifications, other publications, documents and things referenced herein are hereby incorporated herein by this reference in their entirety for all purposes. To the extent of any inconsistency or conflict in the definition or use of a term between any of the incorporated publications, documents or things and the text of the present document, the definition or use of the term in the present document shall prevail.
Although the various aspects of the present invention have been described with respect to certain embodiments, it is understood that the invention is entitled to protection within the full scope of the appended claims.
This application is also related to the following U.S. patent application: U.S. application Ser. No. ______, entitled “Method for Non-Volatile Memory With Worst-Case Control Data Management,” by Bennett et al, filed concurrently herewith, on Oct. 12, 2006.