Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Wed, 23 Oct 2019 16:12:35 -0500
From:   Bjorn Helgaas <helgaas@kernel.org>
To:     Jian-Hong Pan <jian-hong@endlessm.com>
Cc:     Matthew Wilcox <willy@linux.intel.com>, linux@endlesssm.com,
        linux-pci@vger.kernel.org, linux-kernel@vger.kernel.org,
        linux-nvme@lists.infradead.org
Subject: Re: [PATCH] PCI/MSI: Fix restoring of MSI-X vector control's mask bit
Message-ID: <20191023211235.GA236419@google.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20191008034238.2503-1-jian-hong@endlessm.com>
User-Agent: Mutt/1.10.1 (2018-07-13)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Tue, Oct 08, 2019 at 11:42:39AM +0800, Jian-Hong Pan wrote:
> MSI-X vector control's bit 0 is the mask bit, which masks the
> corresponding interrupt request, or not. Other reserved bits might be
> used for other purpose by device vendors. For example, the values of
> Kingston NVMe SSD's MSI-X vector control are neither 0, nor 1, but other
> values [1].
> 
> The original restoring logic in system resuming uses the whole MSI-X
> vector control value as the flag to set/clear the mask bit. However,
> this logic conflicts the idea mentioned above. It may mislead system to
> disable the MSI-X vector entries. That makes system get no interrupt
> from Kingston NVMe SSD after resume and usually get NVMe I/O timeout
> error.
> 
> [  174.715534] nvme nvme0: I/O 978 QID 3 timeout, completion polled
> 
> This patch takes only the mask bit of original MSI-X vector control
> value as the flag to fix this issue.
> 
> [1] https://bugzilla.kernel.org/show_bug.cgi?id=204887#c8
> 
> Buglink: https://bugzilla.kernel.org/show_bug.cgi?id=204887
> Fixed: f2440d9acbe8 ("PCI MSI: Refactor interrupt masking code")
> Signed-off-by: Jian-Hong Pan <jian-hong@endlessm.com>
> ---
>  drivers/pci/msi.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/pci/msi.c b/drivers/pci/msi.c
> index 0884bedcfc7a..deae3d5acaf6 100644
> --- a/drivers/pci/msi.c
> +++ b/drivers/pci/msi.c
> @@ -433,6 +433,7 @@ static void __pci_restore_msi_state(struct pci_dev *dev)
>  static void __pci_restore_msix_state(struct pci_dev *dev)
>  {
>  	struct msi_desc *entry;
> +	u32 flag;
>  
>  	if (!dev->msix_enabled)
>  		return;
> @@ -444,8 +445,10 @@ static void __pci_restore_msix_state(struct pci_dev *dev)
>  				PCI_MSIX_FLAGS_ENABLE | PCI_MSIX_FLAGS_MASKALL);
>  
>  	arch_restore_msi_irqs(dev);
> -	for_each_pci_msi_entry(entry, dev)
> -		msix_mask_irq(entry, entry->masked);
> +	for_each_pci_msi_entry(entry, dev) {
> +		flag = entry->masked & PCI_MSIX_ENTRY_CTRL_MASKBIT;
> +		msix_mask_irq(entry, flag);

This makes good sense: before your patch, when we restore, we set the
mask bit if *any* bits in the Vector Control register are set.

There are other paths leading to __pci_msix_desc_mask_irq(), so I
think it would be better to do the masking there, e.g.,

  if (flag & PCI_MSIX_ENTRY_CTRL_MASKBIT)
    mask_bits |= PCI_MSIX_ENTRY_CTRL_MASKBIT;

I think the other paths all pass either 0 or 1, so they're all safe
today.  But doing the masking in __pci_msix_desc_mask_irq() removes
that assumption from the callers.

I applied the patch below to pci/msi, let me know if it doesn't work
for you.

> +	}
>  
>  	pci_msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_MASKALL, 0);
>  }

commit 1a828a554650
Author: Jian-Hong Pan <jian-hong@endlessm.com>
Date:   Tue Oct 8 11:42:39 2019 +0800

    PCI/MSI: Fix incorrect MSI-X masking on resume
    
    When a driver enables MSI-X, msix_program_entries() reads the MSI-X Vector
    Control register for each vector and saves it in desc->masked.  Each
    register is 32 bits and bit 0 is the actual Mask bit.
    
    When we restored these registers during resume, we previously set the Mask
    bit if *any* bit in desc->masked was set instead of when the Mask bit
    itself was set:
    
      pci_restore_state
        pci_restore_msi_state
          __pci_restore_msix_state
            for_each_pci_msi_entry
              msix_mask_irq(entry, entry->masked)   <-- entire u32 word
                __pci_msix_desc_mask_irq(desc, flag)
                  mask_bits = desc->masked & ~PCI_MSIX_ENTRY_CTRL_MASKBIT
                  if (flag)       <-- testing entire u32, not just bit 0
                    mask_bits |= PCI_MSIX_ENTRY_CTRL_MASKBIT
                  writel(mask_bits, desc_addr + PCI_MSIX_ENTRY_VECTOR_CTRL)
    
    This means that after resume, MSI-X vectors were masked when they shouldn't
    be, which leads to timeouts like this:
    
      nvme nvme0: I/O 978 QID 3 timeout, completion polled
    
    On resume, set the Mask bit only when the saved Mask bit from suspend was
    set.
    
    This should remove the need for 19ea025e1d28 ("nvme: Add quirk for Kingston
    NVME SSD running FW E8FK11.T").
    
    [bhelgaas: commit log, move fix to __pci_msix_desc_mask_irq()]
    Fixes: f2440d9acbe8 ("PCI MSI: Refactor interrupt masking code")
    Link: https://bugzilla.kernel.org/show_bug.cgi?id=204887
    Link: https://lore.kernel.org/r/20191008034238.2503-1-jian-hong@endlessm.com
    Signed-off-by: Jian-Hong Pan <jian-hong@endlessm.com>
    Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>

diff --git a/drivers/pci/msi.c b/drivers/pci/msi.c
index 0884bedcfc7a..771041784e64 100644
--- a/drivers/pci/msi.c
+++ b/drivers/pci/msi.c
@@ -213,12 +213,13 @@ u32 __pci_msix_desc_mask_irq(struct msi_desc *desc, u32 flag)
 
 	if (pci_msi_ignore_mask)
 		return 0;
+
 	desc_addr = pci_msix_desc_addr(desc);
 	if (!desc_addr)
 		return 0;
 
 	mask_bits &= ~PCI_MSIX_ENTRY_CTRL_MASKBIT;
-	if (flag)
+	if (flag & PCI_MSIX_ENTRY_CTRL_MASKBIT)
 		mask_bits |= PCI_MSIX_ENTRY_CTRL_MASKBIT;
 
 	writel(mask_bits, desc_addr + PCI_MSIX_ENTRY_VECTOR_CTRL);