[+cc Keith, Jens, Christoph, Sagi, linux-nvme, LKML]
On Mon, Mar 11, 2019 at 04:31:09PM +0300, Sergey Miroshnichenko wrote:
> Hotplugged devices can affect the existing ones by moving their BARs.
> PCI subsystem will inform the NVME driver about this by invoking
> reset_prepare()+reset_done(), then iounmap()+ioremap() must be called.
Do you mean the PCI core will invoke ->rescan_prepare() and
->rescan_done() (as opposed to *reset*)?
> Signed-off-by: Sergey Miroshnichenko <s.miroshniche...@yadro.com>
> ---
> drivers/nvme/host/pci.c | 29 +++++++++++++++++++++++++++--
> 1 file changed, 27 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
> index 92bad1c810ac..ccea3033a67a 100644
> --- a/drivers/nvme/host/pci.c
> +++ b/drivers/nvme/host/pci.c
> @@ -106,6 +106,7 @@ struct nvme_dev {
> unsigned int num_vecs;
> int q_depth;
> u32 db_stride;
> + resource_size_t current_phys_bar;
> void __iomem *bar;
> unsigned long bar_mapped_size;
> struct work_struct remove_work;
> @@ -1672,13 +1673,16 @@ static int nvme_remap_bar(struct nvme_dev *dev,
> unsigned long size)
> {
> struct pci_dev *pdev = to_pci_dev(dev->dev);
>
> - if (size <= dev->bar_mapped_size)
> + if (dev->bar &&
> + dev->current_phys_bar == pci_resource_start(pdev, 0) &&
> + size <= dev->bar_mapped_size)
> return 0;
> if (size > pci_resource_len(pdev, 0))
> return -ENOMEM;
> if (dev->bar)
> iounmap(dev->bar);
> - dev->bar = ioremap(pci_resource_start(pdev, 0), size);
> + dev->current_phys_bar = pci_resource_start(pdev, 0);
> + dev->bar = ioremap(dev->current_phys_bar, size);
dev->current_phys_bar is different from pci_resource_start() in the
case where the PCI core has moved the nvme BAR, but nvme has not yet
remapped it.
I'm not sure it's worth keeping track of current_phys_bar, as opposed
to always unmapping and remapping. Is this a performance path? I
think there are advantages to always exercising the same code path,
regardless of whether the BAR happened to be moved, e.g., if there's a
bug in the "BAR moved" path, it may be a heisenbug because whether we
exercise that path depends on the current configuration.
If you do need to cache current_phys_bar, maybe this, so it's a little
easier to see that you're not changing the ioremap() itself:
dev->bar = ioremap(pci_resource_start(pdev, 0), size);
dev->current_phys_bar = pci_resource_start(pdev, 0);
> if (!dev->bar) {
> dev->bar_mapped_size = 0;
> return -ENOMEM;
> @@ -2504,6 +2508,8 @@ static void nvme_reset_work(struct work_struct *work)
> if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING))
> goto out;
>
> + nvme_remap_bar(dev, db_bar_size(dev, 0));
How is this change connected to rescan? This looks reset-related.
> /*
> * If we're called to reset a live controller first shut it down before
> * moving on.
> @@ -2910,6 +2916,23 @@ static void nvme_error_resume(struct pci_dev *pdev)
> flush_work(&dev->ctrl.reset_work);
> }
>
> +void nvme_rescan_prepare(struct pci_dev *pdev)
> +{
> + struct nvme_dev *dev = pci_get_drvdata(pdev);
> +
> + nvme_dev_disable(dev, false);
> + nvme_dev_unmap(dev);
> + dev->bar = NULL;
> +}
> +
> +void nvme_rescan_done(struct pci_dev *pdev)
> +{
> + struct nvme_dev *dev = pci_get_drvdata(pdev);
> +
> + nvme_dev_map(dev);
> + nvme_reset_ctrl_sync(&dev->ctrl);
> +}
> +
> static const struct pci_error_handlers nvme_err_handler = {
> .error_detected = nvme_error_detected,
> .slot_reset = nvme_slot_reset,
> @@ -2974,6 +2997,8 @@ static struct pci_driver nvme_driver = {
> },
> .sriov_configure = pci_sriov_configure_simple,
> .err_handler = &nvme_err_handler,
> + .rescan_prepare = nvme_rescan_prepare,
> + .rescan_done = nvme_rescan_done,
> };
>
> static int __init nvme_init(void)
> --
> 2.20.1
>
