TODO list 2011-12-09

Here's the current TODO list that I have. Thuoght it might be useful for people who are wonder what still needs doing. If you have anything else (public) that you'd like to see changed in the driver, let me know and I'll add it to the list.

Also, if you happen to be working on any of these features, feel free to let me know so I can make a note. Keith's working on #2, which prompted me to actually post this list.

While these items are numbered ... it's not a strict priority list.

1. Write an error handler
2. Add a character device for /dev/nvmeN for devices with no visible namespaces (Keith Busch)
3. Submit Async Event Request commands
4. Better CPU -> queue mapping
5. Submit multiple partial I/Os with a single doorbell write
6. Handle block/VM queue congestion
7. Make kthread per CPU
8. Use the async mechanism for namespaces
9. MODULE_PARM_DESC for module parameters (Reported-by: Randy Dunlap)
10. Handle ioremap returning NULL
11. Add support for SCSI ioctls
12. Decide on a sysfs representation for NVMe devices
13. Figure out how to support the Range Type partitioning scheme
14. Use a mempool for nvme requests
15. Call completion handlers at submission time
16. Experiment with interrupt mitigation settings
17. Let user tune interrupt mitigation through sysfs
18. Use an ida for the 'instance' code
19. Implement DSM Deallocate (ie trim)
20. Get a bi_rw incompressible bit, set it in ecryptfs & dm_crypt
21. Write PCI error handling code
22. PCI Suspend/Resume
23. Handle CPU hotplug (allocate/free queues)
24. Add support for data integrity (DIX)
25. Investigate amortising SQ and CQ doorbell writes
26. Experiment with NUMA locality (SQ / CQ local to CPU or to device)
27. Sanity check for namespace count?
28. make_request may sleep -- should the NVMe driver's do that?
29. Allow the Arbitration feature to be adjusted
30. Implement Security Send / Receive
31. Investigate other DSM fields
32. Implement NVMe Power Management
33. Allow for controlling Volatile Write Cache
34. Expose Write Atomicity to block layer
35. Implement Software Progress
36. Remove 'process_cq did something' message
37. Add tracepoint support
38. Allow maximum transfer size and queue depth to be controlled

Mathew,

Regarding the character device for /dev/nvmeN, can we just create a block device namespace 0 with 0 capacity ?
Ie. nvme0n0.

-Kwok

-----Original Message-----
From: [email protected] [mailto:linux-nvme-bounces at lists.infradead.org] On Behalf Of Matthew Wilcox
Sent: Friday, December 09, 2011 1:38 PM
To: linux-nvme at lists.infradead.org
Subject: TODO list 2011-12-09

Here's the current TODO list that I have. Thuoght it might be useful for people who are wonder what still needs doing. If you have anything else (public) that you'd like to see changed in the driver, let me know and I'll add it to the list.

Also, if you happen to be working on any of these features, feel free to let me know so I can make a note. Keith's working on #2, which prompted me to actually post this list.

While these items are numbered ... it's not a strict priority list.

1. Write an error handler
2. Add a character device for /dev/nvmeN for devices with no visible namespaces (Keith Busch)
3. Submit Async Event Request commands
4. Better CPU -> queue mapping
5. Submit multiple partial I/Os with a single doorbell write 6. Handle
block/VM queue congestion 7. Make kthread per CPU 8. Use the async mechanism for namespaces 9. MODULE_PARM_DESC for module parameters (Reported-by: Randy Dunlap) 10. Handle ioremap returning NULL 11. Add support for SCSI ioctls 12. Decide on a sysfs representation for NVMe devices 13. Figure out how to support the Range Type partitioning scheme
14. Use a mempool for nvme requests 15. Call completion handlers at submission time 16. Experiment with interrupt mitigation settings 17. Let user tune interrupt mitigation through sysfs 18. Use an ida for the 'instance' code 19. Implement DSM Deallocate (ie trim) 20. Get a bi_rw incompressible bit, set it in ecryptfs & dm_crypt 21. Write PCI error handling code 22. PCI Suspend/Resume 23. Handle CPU hotplug (allocate/free queues) 24. Add support for data integrity (DIX) 25. Investigate amortising SQ and CQ doorbell writes 26. Experiment with NUMA locality (SQ / CQ local to CPU or to device) 27. Sanity check for namespace count?
28. make_request may sleep -- should the NVMe driver's do that?
29. Allow the Arbitration feature to be adjusted 30. Implement Security Send / Receive 31. Investigate other DSM fields 32. Implement NVMe Power Management 33. Allow for controlling Volatile Write Cache 34. Expose Write Atomicity to block layer 35. Implement Software Progress 36. Remove 'process_cq did something' message 37. Add tracepoint support 38. Allow maximum transfer size and queue depth to be controlled

_______________________________________________
Linux-nvme mailing list
Linux-nvme at lists.infradead.org
http://merlin.infradead.org/mailman/listinfo/linux-nvme

Matthew,

I have a proposal on the module parameters. Please let me know what you think. We at IDT can implement this.

The following module parameters should be configurable:

Name min/max/default NVMe spec
reference
Admin_sub_queue_size 2/4096/128 AQA.ASQS
Admin_cpl_queue_size 2/4096/128 AQA.ACQS
Io_sub_queue_size 2/64K/1K Create
I/O submission Queue
Io_cpl_queue_size 2/64K/1K Create
I/O Completion Queue
Max_transfer_size 4K/1M/128K ECN12 -
Maximum Data Transfer Size
- Actual Max transfer size = min(max_transfer_size, ECN12.MDTS)
Intr_coalescing_threshold 1/256/8 Set Features - Interrupt
Coalescing
Intr_coalescing_time 0/255/10 (in 100us) Set Features - Interrupt
coalescing
Io_command_ID_size 2/64K/io_sub_queue_size (Number of Command ID
entries per io sub queue.)
Max_io_request 1/io_sub_queue_size/io_sub_queue_size (max
outstanding io request per io queue)

Are there any parameters that we should add ?

Thanks

-Kwok

-----Original Message-----
From: [email protected] [mailto:linux-nvme-bounces at lists.infradead.org] On Behalf Of Matthew Wilcox
Sent: Friday, December 09, 2011 1:38 PM
To: linux-nvme at lists.infradead.org
Subject: TODO list 2011-12-09

Here's the current TODO list that I have. Thuoght it might be useful for people who are wonder what still needs doing. If you have anything else (public) that you'd like to see changed in the driver, let me know and I'll add it to the list.

Also, if you happen to be working on any of these features, feel free to let me know so I can make a note. Keith's working on #2, which prompted me to actually post this list.

While these items are numbered ... it's not a strict priority list.

1. Write an error handler
2. Add a character device for /dev/nvmeN for devices with no visible namespaces (Keith Busch)
3. Submit Async Event Request commands
4. Better CPU -> queue mapping
5. Submit multiple partial I/Os with a single doorbell write 6. Handle
block/VM queue congestion 7. Make kthread per CPU 8. Use the async mechanism for namespaces 9. MODULE_PARM_DESC for module parameters (Reported-by: Randy Dunlap) 10. Handle ioremap returning NULL 11. Add support for SCSI ioctls 12. Decide on a sysfs representation for NVMe devices 13. Figure out how to support the Range Type partitioning scheme
14. Use a mempool for nvme requests 15. Call completion handlers at submission time 16. Experiment with interrupt mitigation settings 17. Let user tune interrupt mitigation through sysfs 18. Use an ida for the 'instance' code 19. Implement DSM Deallocate (ie trim) 20. Get a bi_rw incompressible bit, set it in ecryptfs & dm_crypt 21. Write PCI error handling code 22. PCI Suspend/Resume 23. Handle CPU hotplug (allocate/free queues) 24. Add support for data integrity (DIX) 25. Investigate amortising SQ and CQ doorbell writes 26. Experiment with NUMA locality (SQ / CQ local to CPU or to device) 27. Sanity check for namespace count?
28. make_request may sleep -- should the NVMe driver's do that?
29. Allow the Arbitration feature to be adjusted 30. Implement Security Send / Receive 31. Investigate other DSM fields 32. Implement NVMe Power Management 33. Allow for controlling Volatile Write Cache 34. Expose Write Atomicity to block layer 35. Implement Software Progress 36. Remove 'process_cq did something' message 37. Add tracepoint support 38. Allow maximum transfer size and queue depth to be controlled

_______________________________________________
Linux-nvme mailing list
Linux-nvme at lists.infradead.org
http://merlin.infradead.org/mailman/listinfo/linux-nvme

Matthew,

What about host add/remove a namespace ?

Thanks

-Kwok

-----Original Message-----
From: [email protected] [mailto:linux-nvme-bounces at lists.infradead.org] On Behalf Of Matthew Wilcox
Sent: Friday, December 09, 2011 1:38 PM
To: linux-nvme at lists.infradead.org
Subject: TODO list 2011-12-09

Here's the current TODO list that I have. Thuoght it might be useful for people who are wonder what still needs doing. If you have anything else (public) that you'd like to see changed in the driver, let me know and I'll add it to the list.

Also, if you happen to be working on any of these features, feel free to let me know so I can make a note. Keith's working on #2, which prompted me to actually post this list.

While these items are numbered ... it's not a strict priority list.

1. Write an error handler
2. Add a character device for /dev/nvmeN for devices with no visible namespaces (Keith Busch)
3. Submit Async Event Request commands
4. Better CPU -> queue mapping
5. Submit multiple partial I/Os with a single doorbell write 6. Handle
block/VM queue congestion 7. Make kthread per CPU 8. Use the async mechanism for namespaces 9. MODULE_PARM_DESC for module parameters (Reported-by: Randy Dunlap) 10. Handle ioremap returning NULL 11. Add support for SCSI ioctls 12. Decide on a sysfs representation for NVMe devices 13. Figure out how to support the Range Type partitioning scheme
14. Use a mempool for nvme requests 15. Call completion handlers at submission time 16. Experiment with interrupt mitigation settings 17. Let user tune interrupt mitigation through sysfs 18. Use an ida for the 'instance' code 19. Implement DSM Deallocate (ie trim) 20. Get a bi_rw incompressible bit, set it in ecryptfs & dm_crypt 21. Write PCI error handling code 22. PCI Suspend/Resume 23. Handle CPU hotplug (allocate/free queues) 24. Add support for data integrity (DIX) 25. Investigate amortising SQ and CQ doorbell writes 26. Experiment with NUMA locality (SQ / CQ local to CPU or to device) 27. Sanity check for namespace count?
28. make_request may sleep -- should the NVMe driver's do that?
29. Allow the Arbitration feature to be adjusted 30. Implement Security Send / Receive 31. Investigate other DSM fields 32. Implement NVMe Power Management 33. Allow for controlling Volatile Write Cache 34. Expose Write Atomicity to block layer 35. Implement Software Progress 36. Remove 'process_cq did something' message 37. Add tracepoint support 38. Allow maximum transfer size and queue depth to be controlled

_______________________________________________
Linux-nvme mailing list
Linux-nvme at lists.infradead.org
http://merlin.infradead.org/mailman/listinfo/linux-nvme

That doesn't feel very intuitive to me. What would be the advantage?

There is code in the driver to create devices for NS 1, 2, 3 ...N already. It seems to me that it is very easy to create a "controller" device name with NS=0. There is really no need to create new code to create a character device.

i.e. use the same function nvme_alloc_ns() to create NS0 and hence driver nvme0n0.
The device name becomes:

nvme0n0: controller
nvme0n1: NS 1
name0n2: NS 2
...
...
etc

I am ok with a character device "nvme0" or "nvme0n0" but why do it if a block device works and the code is there already.

What do you think ?

-Kwok

-----Original Message-----
From: Matthew Wilcox [mailto:[email protected]]
Sent: Monday, December 12, 2011 12:48 PM
To: Kong, Kwok
Cc: linux-nvme at lists.infradead.org
Subject: Re: TODO list 2011-12-09

That doesn't feel very intuitive to me. What would be the advantage?