[RFC 02/13] nvme: wire-up support for async-passthru on

Kanchan Joshi joshiiitr at gmail.com
Wed Dec 22 05:46:35 PST 2021


On Wed, Dec 22, 2021 at 2:46 AM Clay Mayers <Clay.Mayers at kioxia.com> wrote:
>
> Message-ID: <20211220141734.12206-3-joshi.k at samsung.com>
>
> On 12/20/21 19:47:23 +0530, Kanchan Joshi wrote:
> > Introduce handlers for fops->async_cmd(), implementing async passthru on
> > char device (including the multipath one).
> > The handlers supports NVME_IOCTL_IO64_CMD.
> >
> I commented on these two issues below in more detail at
> https://github.com/joshkan/nvme-uring-pt/issues

That is on general/existing nvme ioctl (and not specific to this
series). You might want to open up a discussion in the nvme mailing
list.

> > +static void nvme_setup_uring_cmd_data(struct request *rq,
> > +             struct io_uring_cmd *ioucmd, void *meta,
> > +             void __user *meta_buffer, u32 meta_len, bool write) {
> > +     struct nvme_uring_cmd *cmd = nvme_uring_cmd(ioucmd);
> > +
> > +     /* to free bio on completion, as req->bio will be null at that time */
> > +     cmd->bio = rq->bio;
> > +     /* meta update is required only for read requests */
> > +     if (meta && !write) {
> > +             cmd->meta = meta;
> > +             cmd->meta_buffer = meta_buffer;
> > +             cmd->meta_len = meta_len;
> > +     } else {
> > +             cmd->meta = NULL;
> I believe that not saving meta in cmd->meta will leak it when it's a write.

Indeed. Will fix that up.

> But nvme_pt_task_cb also needs to change to copy to user when
> cmd->meta_buffer is set instead of cmd->meta.
>
> > +
> > +int nvme_ns_chr_async_cmd(struct io_uring_cmd *ioucmd,
> > +             enum io_uring_cmd_flags flags)
> > +{
> > +     struct nvme_ns *ns = container_of(file_inode(ioucmd->file)->i_cdev,
> > +                     struct nvme_ns, cdev);
> > +
> > +     return nvme_ns_async_ioctl(ns, ioucmd); }
> > +
> The uring cmd flags are not being passed to nvme_ns_async_ioctl - what if
> IO_URING_F_NONBLOCK Is set?  When it is, I think the nvme_alloc_request()
> call in nvme_submit_user_cmd() needs to pass in BLK_MQ_REQ_NOWAIT as
> the flags parameter or move to another thread.  Our proto-type does the former
> requiring user mode to retry on -EWOULDBLOCK and -EBUSY.

Right, this part is not handled. Need to get that sorted in the next
version. Thanks.



-- 
Joshi



More information about the Linux-nvme mailing list