Question about ioctl interface - IO passthru

Question about ioctl interface - IO passthru

[Eyal BenDavid]

Hi,

I wrote a small program to perform IOs based on the ioctl interface
from linux_ioctl.h and code from nvme-cli.

Specifically

Question about ioctl interface - IO passthru

[Eyal BenDavid]

Hi,

I wrote a small program to perform IOs based on the ioctl interface
from linux_ioctl.h and code from nvme-cli.

Specifically I wrote this function for write zero (code is used in
stress and verification testing).
Note that this is C++ (only difference is the default values for
function arguments)

====
int nvme_io_passthru(int      fd,
                     uint32_t nsid,
                     uint8_t  opcode,
                     uint64_t slba,
                     uint16_t nblocks,
                     void*    data = NULL,
                     uint32_t dlen = 0,
                     uint8_t  flags = 0,
                     uint32_t timeout = 30000,
                     uint16_t control = 0,
                     uint32_t reftag = 0,
                     uint16_t apptag = 0,
                     uint16_t appmask = 0)
{
    struct nvme_passthru_cmd cmd;
    memset(&cmd, 0, sizeof(cmd));

    cmd.opcode     = opcode;
    cmd.flags      = flags;
    cmd.nsid       = nsid;
    cmd.addr       = (uint64_t)(uintptr_t) data;
    cmd.data_len   = dlen;
    cmd.cdw10      = (slba & 0xffffffff);
    cmd.cdw11      = (slba >> 32);
    cmd.cdw12      = (nblocks | (control << 16));
    cmd.cdw14      = reftag;
    cmd.cdw15      = (apptag | (appmask << 16));
    cmd.timeout_ms = timeout;

    return ioctl(fd, NVME_IOCTL_IO_CMD, &cmd);
}
====

To my surprise the function worked fine not only with write zero but
with other IO opcodes (read, write, compare)
The advantage is now I control the timeout value in IO commands!

What is the difference from using "nvme_user_io" structure?
Is it safe to use this version of IO commands?

Thanks!
Eyal BD.

Question about ioctl interface - IO passthru

[Keith Busch]

On Thu, Jan 31, 2019@04:09:18PM +0200, Eyal BenDavid wrote:
> What is the difference from using "nvme_user_io" structure?
> Is it safe to use this version of IO commands?

The passthru structure is good to use for any command. The
nvme_user_io structure is a legacy feature from very early nvme days
when read/write/compare were the only IO commands. It didn't extend to
other IO commands, and deprecating it wouldn't be a bad thing.

Question about ioctl interface - IO passthru

[Eyal BenDavid]

On Thu, Jan 31, 2019@4:56 PM Keith Busch <keith.busch@intel.com> wrote:
>
> On Thu, Jan 31, 2019@04:09:18PM +0200, Eyal BenDavid wrote:
> > What is the difference from using "nvme_user_io" structure?
> > Is it safe to use this version of IO commands?
>
> The passthru structure is good to use for any command. The
> nvme_user_io structure is a legacy feature from very early nvme days
> when read/write/compare were the only IO commands. It didn't extend to
> other IO commands, and deprecating it wouldn't be a bad thing.

Thanks!

A good start is to rewrite nvme read/write/compare commands using the
passthru structure instead of nvme_user_io.

---
Eyal

Question about ioctl interface - IO passthru

[Chaitanya Kulkarni]

One of the issue with the current nvme_user_io structure is that the new commands
such as NVMe Zone append?which returns result will not be able to use the nvme_user_io
since it lacks?that field (e.g. result field) and had to use the passthru?command or add a new
command definition.

So I think it is a good start that to use the nvme_passthru command and make it more generic.

Keith any comments on this ?

or should we extend the nvme_user_io command with
the new addition such as result field (less likely in my?opinion, may break IOCTL)??

From: Linux-nvme <linux-nvme-bounces@lists.infradead.org> on behalf of Eyal BenDavid <bdeyal@gmail.com>
Sent: Thursday, January 31, 2019 10:08 AM
To: Keith Busch
Cc: linux-nvme
Subject: Re: Question about ioctl interface - IO passthru
?
On Thu, Jan 31, 2019@4:56 PM Keith Busch <keith.busch@intel.com> wrote:
>
> On Thu, Jan 31, 2019@04:09:18PM +0200, Eyal BenDavid wrote:
> > What is the difference from using "nvme_user_io" structure?
> > Is it safe to use this version of IO commands?
>
> The passthru structure is good to use for any command. The
> nvme_user_io structure is a legacy feature from very early nvme days
> when read/write/compare were the only IO commands. It didn't extend to
> other IO commands, and deprecating it wouldn't be a bad thing.

Thanks!

A good start is to rewrite nvme read/write/compare commands using the
passthru structure instead of nvme_user_io.

---
Eyal

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

Question about ioctl interface - IO passthru

[Keith Busch]

On Thu, Jan 31, 2019@06:39:28PM +0000, Chaitanya Kulkarni wrote:
> So I think it is a good start that to use the nvme_passthru command and make it more generic.
> 
> Keith any comments on this ?
> 
> or should we extend the nvme_user_io command with
> the new addition such as result field (less likely in my?opinion, may break IOCTL)??

Yeah, we can't break the ioctl ABI. :(

The only reason I haven't moved nvme-cli from user_io to passthru is
that user_io works on every nvme capable kernel, and it also supported
metadata long before passthru did. I think it's overdue, though, I'll
just add a new param for ancient-kernel compatibility mode to fallback
to the legacy user_io.