Sandisk Technologies Inc.

Amir Segev

Shay Benisty

Elkana Richter

Yehuda Hahn

Amir Rozen

Nissim Elmaleh

Rotem Sela

When entering low power mode, any commands that are fetched and waiting to be executed are completed. When the data storage device exits low power mode, new commands are fetched. The completing and fetching takes time. Rather than completing all of the commands, some selected commands can be stored in a persistent memory region (PMR) when entering low power mode. The storing in PMR is faster than completing the commands. Furthermore, retrieving the commands from PMR is faster than retrieving new commands from a host device on exiting low power mode. Thus, using PMR results in more efficient use of data storage device bandwidth.

Pipeline optimization in low power mode using PMR

Instead of entering lanes into an unused power state, enter unactive lanes into an unconnected power state to save more current during low power states. Using a small control logic will allow a controller to control unactive lanes in low power mode. When a lane is in an unactive power state or in an unused power state, an unactive lane controller (ULC) uses side-band signaling to place the unactive lane into either unused power state or unconnected power state. When a lane is in unused power state, then the ULC places the lane in unconnected power state. A single ULC is able to controller multiple lanes or you can have multiple ULC's, each controlling a single lane.

PHY lanes disabling for power efficiency

A data storage device includes a memory device and a controller coupled to the memory device. When a command is received by the controller from a host device, the controller determines whether the command size is greater than a threshold size. If the command is not greater than the threshold size, the command is sent to a first queue, otherwise, the command is sent to a second queue. Commands are executed from the first queue until a command size tracker value, which increases by a size representative of each command executed from the first queue, equals or exceeds a threshold value. When the command size tracker value equals or exceeds the threshold value, a command from the second queue is executed and the command size tracker value decreases by a size representative of the command from the second queue. Completion messages are sent at specific intervals based on the executing.

Write completion pacing for unbalanced command length

Instead of maximizing the possible bandwidth of device, utilize time slice credits (TSC), to ensure bandwidth average over a sliding window. When the average is ensured over a sliding window, the device should not care when the host decides to sample a 100 mSec for example, as the average will always be correct. By utilizing set percentage of predetermined allotment for the average bandwidth requirement, the system can give out credit on a predetermined interval. The credit is given out based on usage and once credit is depleted, data cannot be sent until more credit is accumulated. When data is not sent, the system is given a chance to accumulate credit to increase the amount of data sent. Once credit is at a level high enough to send data the device will send the data, but not at a speed that will surpass the average bandwidth requirement.

Bandwidth averaging in a stochastic system

The present disclosure generally relates improved key-per IO (KIPO) processing for multiple tenants. Rather than when a tenant requests a key change to stop tenants from working, indirect-double-indexing can be used to prevent bandwidth loss in tenants during adaptions for other tenants. When a tenant requests to manipulate the key-index table, the system will keep working. The current key index list will be duplicated. While the duplicated key-index list is manipulated according to the request, all tenants may still work on their current key-index tables until the request is complete. Once the request is complete, the tenant with the request will switch to the new table, while the old table is updated. Once the old table is updated, the tenant will switch to the updated table for continued work. No tenant, including the tenant that makes the request, continues working as the request is completed.

Key-per-IO multiple tenant isolation

Instead of the firmware (FW) monitoring for both an overlap read command and a write protected (WP) overlapped write command, the FW can monitor the overlap table only. The controller will receive a read command or an overlap read command. If the command is not an overlap read command, then the controller will process the command normally. If the command is an overlap read command, then the controller will look for the write command that is being overlapped. When passing the write overlap of write protect (WP) into the data path instead of the exception queue, the FW is limited to monitoring the overlap table. As such, when the WP command is completed the entry is removed by the data path. As a result, the FW can limit monitoring to when the data path clears the overlap table.

Write protect HW acceleration

Instead of an arbitration over the link not considering bursts, a smart scheduler in a solid state drive (SSD) host interface is burst aware. The scheduler considers the type of transactions that are going to be sent over the interface. The scheduler sends the transactions in the most efficient way while maximizing the efficiency over the host DRAM. The schedulers may be calibrated from time to time on-the-fly to find the optimal configurations adapted to the current workload. The scheduler will organize the packets selected by the arbitration module so that the data transfers are sent in a burst of a predetermined sized to the host for optimum performance. For further optimization other packet types are sent in bursts as well.

Burst awareness scheduler over host interface

The present disclosure generally relates to read and write operations utilizing barrier commands. Using barrier commands and a snapshot of doorbell states of submission queues (SQs), the necessary write commands to perform a read may be identified and executed to reduce any wait time of the host. As such, host delays during reads and writes are reduced. In absence of a barrier command, the host needs to wait for writes to complete before performing a read. When a barrier command is used, the host needs to wait for the barrier command to complete before performing a read. The controller will execute the post barrier reads only after completing the pre-barrier writes. As will be discussed herein, the controller completes the barrier command as soon as a doorbell snapshot is taken even though the pre-barrier writes may not yet be completed.

Fast execution of barrier command

The present disclosure generally relates to more effective utilization of write and read bandwidth in submission queues (SQs). The data storage device treats a SQ as two separate SQs: one write SQ and one read SQ. Rather than a single fetch pointer for the entire SQ, the write SQ has a write fetch pointer (WFP) while the read SQ has a separate read fetch pointer (RFP). So long as the individual pointers are less than a queue pointer (QP), the data storage device can still process commands for either read or write SQ even if the other SQ has run out of credits. In so doing, read and write bandwidths can be effectively utilized.

Bandwidth balancing for a single namespace tenant in multi-function nonvolatile memory express devices

There is a tradeoff between the amount of power consumption decreased and the latency needed to return a data storage device back to an operational power mode. When the data storage device receives a wake up indication from a host device, a controller of the data storage device initiates a counter in order to determine a host exit latency. Based on the host exit latency, the controller determines a group of low power state entrance actions from a plurality of groups to perform during a next entrance into a firmware active idle state based on an associated completion wake up time and the host exit latency. The controller selects the group whose completion wake up time is closest to the host exit latency and less than or equal to the host exit latency. The controller performs the selected groups low power state entrance actions during a next entrance into the firmware active idle state.

Low power optimization based upon host exit latency

Instead of using programmable block size aggregation, a lower multiple of page, and down to a page size aggregation is used. A bad block prediction unit in a controller is able to predict when a programmable block has a bad page. The bad block prediction unit can lower the aggregation size of a programmable block by monitoring the life cycle of the programmable block through bad block statistic collection. When the accumulation size passes a threshold, the bad block prediction unit lowers the aggregation size. The bad block prediction unit can also predict when to lower aggregation size based on the number of reconstructions. An aggregate size level is set at a page boundary, and once the number of reconstructions reaches that page boundary, the bad block prediction unit lowers the aggregation size to page aggregation. The bad block prediction unit is able to predict both life cycle threshold changes and reconstructions changes.

Write aggregation based on NAND wear level

A data storage device includes a memory device and a controller coupled to the memory device. The controller is configured to receive a shutdown notification, fetch one or more command identifiers from a submission queue of a host device, generate error indications for the one or more command identifiers, and send a completion message, including the generated error indication, for each of the one or more command identifiers to the host device. The controller is further configured to push non-processed pending commands to a completion finite state machine, where the controller generates an error indication for each of the non-processed pending commands and sends a completion message, including the generated error indication, for each of the non-processed pending commands to the host device. While the controller is fetching command identifiers and pushing non-process commands, the controller is configured to continue processing processed commands in parallel.

Instant submission queue release

In order to ensure that a bandwidth allocated to each tenant of a plurality of tenants of a data storage device is maintained, a controller of the data storage device may split a large read command, received from a host device, into a plurality of chunks, where each chunk corresponds to a distinct portion of the split large read command. Because the allocated bandwidth for a tenant is static, one or more chunks of the plurality of chunks, up to the allocated bandwidth, are executed, such that the bandwidth required to perform the one or more chunks does not exceed the allocated bandwidth for the particular tenant. Split information is added to the plurality of chunks in order to maintain coherency when executing the one or more chunks. Therefore, the agreed-upon allocated bandwidth for each tenant is maintained while performing large read commands requiring more bandwidth than allocated.

Multi-tenant device read command service balancing

The present disclosure generally relates to improved address translation. Rather than fetching translated addresses using ATS/ATC, a HIM address translation search engine (HATS) is used through implementing the ATC in a layer above per an NVMe command. The HATS is an engine that will monitor pointers with untranslated addresses and will fetch the translated addresses for the pointers. Once the translated addresses are fetched for the pointer, the HATS will overwrite the untranslated address with the translated address. The HATS will then update the status of the pointers. When a translation request fails, the device will use PRI to request the translated address. During a translation request fail the device will drain any incoming requests while skipping the data transfer phase. The device will not block any other requests in a queue. Once that translated address is received through the PRI flow, the status of the pointer will be updated. After PRI completes, the device will return to using the HATS to get all the translated addresses for the pointers with an untranslated address status. After getting the translated addresses, the device shall use the translated addresses to perform DMA operations.

ATS PRI support with implicit cache

The present disclosure generally relates to memory management during SGL fetching. When a data storage device is required to fetch an SGL from a host device, the data storage device cannot determine how much memory will be required to be allocated. The disclosure herein reduces the impact of the problem of under or over allocating memory and over-fetching, thereby reducing performance of the device during transfers. The disclosure provides guidance on how to implement an adaptive learning process based upon statistic collection of SGL fetches. By maintaining a table of statistics, the data storage device controller may learn and more closely predict an amount of memory to allocate for SGL fetching.

Growing community of inventors

Meiter, Israel

Amir Segev