nvme: add support for abort command#2825
nvme: add support for abort command#2825ikegami-t wants to merge 2 commits intolinux-nvme:masterfrom
Conversation
The abort command can be submitted by the admin-passthru command. But better to support by the nvme abort command directly then add. Signed-off-by: Tokunori Ikegami <[email protected]>
Since actually the message output by the error case. Signed-off-by: Tokunori Ikegami <[email protected]>
|
The changes depended on the PR: linux-nvme/libnvme#1017. |
|
No no no! Userspace doesn't know anything about the command ids or queue contexts that are used, there is no way this interface could be safely or consistently used! |
|
Just sent the commit below to get the SQID and CID informtion from the driver before the PR but is this still not safe? By the way just thought to change the driver to abort a command with the opcode or etc., used by the userspace also. |
Absolutely not. Outside broken controllers, that sysfs interface proposal just displays a racy snapshot, and I don't think it is appropriate for sysfs anyway as it looks like a debug interface rather than a device attribute. Nvme devices do millions of IOPs, there's no way the output from reading a sysfs file is accurate before the text even hits the output steam. |
|
The only reason the kernel is able to do this at all is because it holds a reference on the targeted qid:cid pair that locks the context of the outstanding IO. It can't possibly be a different context in this path, we know exactly what we're aborting. You can't do that from user space because it doesn't own the hardware abstractions. There is zero reason to make this an easily reachable command, just like we don't support create/delete cq/sq. |
|
I could understood well so let me close the PRs. Thank you for your explanation.
Yes the sysfs interface is for the debugging purpose so I will do consider more later if any other way to confirm the command not completed. |
2 participants
The abort command can be submitted by the admin-passthru command. But better to support by the nvme abort command directly then add.