On 10/12/24 11:41AM, Bart Van Assche wrote:
On 12/9/24 6:20 PM, Martin K. Petersen wrote:
What would be the benefit of submitting these operations concurrently?
I expect that submitting the two copy operations concurrently would
result in lower latency for NVMe devices because the REQ_OP_COPY_DST
operation can be submitted without waiting for the REQ_OP_COPY_SRC
result.
As I have explained, it adds substantial complexity and object lifetime
issues throughout the stack. To what end?
I think the approach of embedding the ROD token in the bio payload would
add complexity in the block layer. The token-based copy offload approach
involves submitting at least the following commands to the SCSI device:
* POPULATE TOKEN with a list identifier and source data ranges as
parameters to send the source data ranges to the device.
* RECEIVE ROD TOKEN INFORMATION with a list identifier as parameter to
receive the ROD token.
* WRITE USING TOKEN with the ROD token and the destination ranges as
parameters to tell the device to start the copy operation.
If the block layer would have to manage the ROD token, how would the ROD
token be provided to the block layer? Bidirectional commands have been
removed from the Linux kernel a while ago so the REQ_OP_COPY_IN
parameter data would have to be used to pass parameters to the SCSI
driver and also to pass the ROD token back to the block layer. A
possible approach is to let the SCSI core allocate memory for the ROD
token with kmalloc and to pass that pointer back to the block layer
by writing that pointer into the REQ_OP_COPY_IN parameter data. While
this can be implemented, I'm not sure that we should integrate support
in the block layer for managing ROD tokens since ROD tokens are a
concept that is specific to the SCSI protocol.
Block layer can allocate a buffer and send this as part of copy
operation.
Driver can store token/custom info inside the buffer sent along with
REQ_OP_COPY_SRC and expect that block layer sends back this info/buffer
again in REQ_OP_COPY_DST ?
This will reduce the effort for block layer to manage the lifetime issues.
Is there any reason, why we cant store the info inside this buffer in
driver ?
This scheme will require sequential submission of SRC and DST
bio's. This might increase in latency, but allows to have simpler design.
Main use case for copy is GC, which is mostly a background operation.
--
Nitesh Shetty
