[PATCHv2] nvme-pci: do not directly handle subsys reset fallout

[Nilay Shroff]

Hi Keith,

On 6/24/24 22:56, Keith Busch wrote:
> From: Keith Busch <kbusch at kernel.org>
> 
> Scheduling reset_work after a nvme subsystem reset is expected to fail
> on pcie, but this also prevents potential handling from the platform's
> pcie services may provide that may successfully recovering the link
> without re-enumeration. Such examples include AER, DPC, and power's EEH.
> 
> Provide a pci specific operation that safely initiates a subsystem
> reset, and instead of scheduling reset work, read back the status
> register to trigger a pcie read error.
> 
> Since this only affects pci, the other fabrics drivers subscribe to a
> generic nvmf subsystem reset that is exactly the same as before. The
> loop fabric doesn't use it because nvmet doesn't support setting that
> property anyway.
> 
> And since we're using the magic NSSR value in two places now, provide a
> symbolic define for it.
> 
> Reported-by: Nilay Shroff <nilay at linux.ibm.com>
> Signed-off-by: Keith Busch <kbusch at kernel.org>
> ---
> Changes from v1, applied comments from Christoph:
> 
>   Split out and require new ctrl subsystem reset op
> 
>   Comments updated for correctness and clarity
> 
>   Use a symbolic name for the special reset value
> 
>  drivers/nvme/host/fabrics.c | 15 +++++++++++++++
>  drivers/nvme/host/fabrics.h |  1 +
>  drivers/nvme/host/fc.c      |  1 +
>  drivers/nvme/host/nvme.h    | 14 +++-----------
>  drivers/nvme/host/pci.c     | 35 +++++++++++++++++++++++++++++++++++
>  drivers/nvme/host/rdma.c    |  1 +
>  drivers/nvme/host/tcp.c     |  1 +
>  include/linux/nvme.h        |  3 +++
>  8 files changed, 60 insertions(+), 11 deletions(-)
> 
>  
>  /*
> diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
> index 102a9fb0c65ff..6cb7acef05576 100644
> --- a/drivers/nvme/host/pci.c
> +++ b/drivers/nvme/host/pci.c
> @@ -1143,6 +1143,40 @@ static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
>  	spin_unlock(&nvmeq->sq_lock);
>  }
>  
> +static int nvme_pci_subsystem_reset(struct nvme_ctrl *ctrl)
> +{
> +	struct nvme_dev *dev = to_nvme_dev(ctrl);
> +	int ret = 0;
> +
> +	/*
> +	 * Taking the shutdown_lock ensures the BAR mapping is not being
> +	 * altered by reset_work. Holding this lock before the RESETTING state
> +	 * change, if successful, also ensures nvme_remove won't be able to
> +	 * proceed to iounmap until we're done.
> +	 */
> +	mutex_lock(&dev->shutdown_lock);
> +	if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING)) {
> +		ret = -EBUSY;
> +		goto unlock;
> +	}
> +
> +	if (!dev->bar_mapped_size) {
> +		ret = -ENODEV;
> +		goto unlock;
> +	}
> +
> +	writel(NVME_SUBSYS_RESET, dev->bar + NVME_REG_NSSR);
> +
> +	/*
> +	 * Read controller status to flush the previous write and trigger a
> +	 * pcie read error.
> +	 */
> +	readl(dev->bar + NVME_REG_CSTS);
> +unlock:
> +	mutex_unlock(&dev->shutdown_lock);
> +	return ret;
> +}
I noticed that in your earlier proposal we were changing the NVMe controller 
state from RESETTING to LIVE just after write to NVME_REG_NSSR. In this patch,
you removed that code and due to it now pci error recovery code (nvme_error_detected()) 
couldn't change the state of controller to RESETTING and that causes the pci error 
recovery to immediately bail out without recovering the NVMe adapter. 

I think we need to change state of controller to LIVE after write to NVME_REG_NSSR.
So was there any specific reason you removed that code or was it missed out by mistake?

Thanks,
--Nilay