[PATCH v6 07/20] hw/block/nvme: fix pin-based interrupt behavior

Klaus Jensen posted 20 patches 5 years, 9 months ago
Maintainers: Max Reitz <mreitz@redhat.com>, Kevin Wolf <kwolf@redhat.com>, Keith Busch <kbusch@kernel.org>
There is a newer version of this series
[PATCH v6 07/20] hw/block/nvme: fix pin-based interrupt behavior
Posted by Klaus Jensen 5 years, 9 months ago
From: Klaus Jensen <k.jensen@samsung.com>

First, since the device only supports MSI-X or pin-based interrupt, if
MSI-X is not enabled, it should not accept interrupt vectors different
from 0 when creating completion queues.

Secondly, the irq_status NvmeCtrl member is meant to be compared to the
INTMS register, so it should only be 32 bits wide. And it is really only
useful when used with multi-message MSI.

Third, since we do not force a 1-to-1 correspondence between cqid and
interrupt vector, the irq_status register should not have bits set
according to cqid, but according to the associated interrupt vector.

Fix these issues, but keep irq_status available so we can easily support
multi-message MSI down the line.

Fixes: 5e9aa92eb1a5 ("hw/block: Fix pin-based interrupt behaviour of NVMe")
Cc: "Michael S. Tsirkin" <mst@redhat.com>
Cc: Marcel Apfelbaum <marcel.apfelbaum@gmail.com>
Signed-off-by: Klaus Jensen <k.jensen@samsung.com>
---
 hw/block/nvme.c | 12 ++++++++----
 hw/block/nvme.h |  2 +-
 2 files changed, 9 insertions(+), 5 deletions(-)

diff --git a/hw/block/nvme.c b/hw/block/nvme.c
index 623a88be93dc..c9d10df1f763 100644
--- a/hw/block/nvme.c
+++ b/hw/block/nvme.c
@@ -138,8 +138,8 @@ static void nvme_irq_assert(NvmeCtrl *n, NvmeCQueue *cq)
             msix_notify(&(n->parent_obj), cq->vector);
         } else {
             trace_pci_nvme_irq_pin();
-            assert(cq->cqid < 64);
-            n->irq_status |= 1 << cq->cqid;
+            assert(cq->vector < 32);
+            n->irq_status |= 1 << cq->vector;
             nvme_irq_check(n);
         }
     } else {
@@ -153,8 +153,8 @@ static void nvme_irq_deassert(NvmeCtrl *n, NvmeCQueue *cq)
         if (msix_enabled(&(n->parent_obj))) {
             return;
         } else {
-            assert(cq->cqid < 64);
-            n->irq_status &= ~(1 << cq->cqid);
+            assert(cq->vector < 32);
+            n->irq_status &= ~(1 << cq->vector);
             nvme_irq_check(n);
         }
     }
@@ -653,6 +653,10 @@ static uint16_t nvme_create_cq(NvmeCtrl *n, NvmeCmd *cmd)
         trace_pci_nvme_err_invalid_create_cq_addr(prp1);
         return NVME_INVALID_FIELD | NVME_DNR;
     }
+    if (unlikely(!msix_enabled(&n->parent_obj) && vector)) {
+        trace_pci_nvme_err_invalid_create_cq_vector(vector);
+        return NVME_INVALID_IRQ_VECTOR | NVME_DNR;
+    }
     if (unlikely(vector > n->params.num_queues)) {
         trace_pci_nvme_err_invalid_create_cq_vector(vector);
         return NVME_INVALID_IRQ_VECTOR | NVME_DNR;
diff --git a/hw/block/nvme.h b/hw/block/nvme.h
index 9df244c93c02..91f16c812582 100644
--- a/hw/block/nvme.h
+++ b/hw/block/nvme.h
@@ -84,7 +84,7 @@ typedef struct NvmeCtrl {
     uint32_t    cmbsz;
     uint32_t    cmbloc;
     uint8_t     *cmbuf;
-    uint64_t    irq_status;
+    uint32_t    irq_status;
     uint64_t    host_timestamp;                 /* Timestamp sent by the host */
     uint64_t    timestamp_set_qemu_clock_ms;    /* QEMU clock time */
 
-- 
2.26.2


Re: [PATCH v6 07/20] hw/block/nvme: fix pin-based interrupt behavior
Posted by Keith Busch 5 years, 8 months ago
On Thu, May 14, 2020 at 06:45:58AM +0200, Klaus Jensen wrote:
> From: Klaus Jensen <k.jensen@samsung.com>
> 
> First, since the device only supports MSI-X or pin-based interrupt, if
> MSI-X is not enabled, it should not accept interrupt vectors different
> from 0 when creating completion queues.
> 
> Secondly, the irq_status NvmeCtrl member is meant to be compared to the
> INTMS register, so it should only be 32 bits wide. And it is really only
> useful when used with multi-message MSI.
> 
> Third, since we do not force a 1-to-1 correspondence between cqid and
> interrupt vector, the irq_status register should not have bits set
> according to cqid, but according to the associated interrupt vector.
> 
> Fix these issues, but keep irq_status available so we can easily support
> multi-message MSI down the line.

Looks good.

Reviewed-by: Keith Busch <kbusch@kernel.org>

Re: [PATCH v6 07/20] hw/block/nvme: fix pin-based interrupt behavior
Posted by Klaus Jensen 5 years, 8 months ago
On May 14 06:45, Klaus Jensen wrote:
> From: Klaus Jensen <k.jensen@samsung.com>
> 
> First, since the device only supports MSI-X or pin-based interrupt, if
> MSI-X is not enabled, it should not accept interrupt vectors different
> from 0 when creating completion queues.
> 
> Secondly, the irq_status NvmeCtrl member is meant to be compared to the
> INTMS register, so it should only be 32 bits wide. And it is really only
> useful when used with multi-message MSI.
> 
> Third, since we do not force a 1-to-1 correspondence between cqid and
> interrupt vector, the irq_status register should not have bits set
> according to cqid, but according to the associated interrupt vector.
> 
> Fix these issues, but keep irq_status available so we can easily support
> multi-message MSI down the line.
> 
> Fixes: 5e9aa92eb1a5 ("hw/block: Fix pin-based interrupt behaviour of NVMe")
> Cc: "Michael S. Tsirkin" <mst@redhat.com>
> Cc: Marcel Apfelbaum <marcel.apfelbaum@gmail.com>
> Signed-off-by: Klaus Jensen <k.jensen@samsung.com>
> ---
>  hw/block/nvme.c | 12 ++++++++----
>  hw/block/nvme.h |  2 +-
>  2 files changed, 9 insertions(+), 5 deletions(-)
> 

Gentle ping on this and [PATCH v6 08/20].


Thanks,
Klaus

Re: [PATCH v6 07/20] hw/block/nvme: fix pin-based interrupt behavior
Posted by Klaus Jensen 5 years, 8 months ago
On May 18 07:02, Klaus Jensen wrote:
> On May 14 06:45, Klaus Jensen wrote:
> > From: Klaus Jensen <k.jensen@samsung.com>
> > 
> > First, since the device only supports MSI-X or pin-based interrupt, if
> > MSI-X is not enabled, it should not accept interrupt vectors different
> > from 0 when creating completion queues.
> > 
> > Secondly, the irq_status NvmeCtrl member is meant to be compared to the
> > INTMS register, so it should only be 32 bits wide. And it is really only
> > useful when used with multi-message MSI.
> > 
> > Third, since we do not force a 1-to-1 correspondence between cqid and
> > interrupt vector, the irq_status register should not have bits set
> > according to cqid, but according to the associated interrupt vector.
> > 
> > Fix these issues, but keep irq_status available so we can easily support
> > multi-message MSI down the line.
> > 
> > Fixes: 5e9aa92eb1a5 ("hw/block: Fix pin-based interrupt behaviour of NVMe")
> > Cc: "Michael S. Tsirkin" <mst@redhat.com>
> > Cc: Marcel Apfelbaum <marcel.apfelbaum@gmail.com>
> > Signed-off-by: Klaus Jensen <k.jensen@samsung.com>
> > ---
> >  hw/block/nvme.c | 12 ++++++++----
> >  hw/block/nvme.h |  2 +-
> >  2 files changed, 9 insertions(+), 5 deletions(-)
> > 
> 
> Gentle ping on this and [PATCH v6 08/20].
> 
 
And another ping :)