[Qemu-devel] [PATCH 0/3] vfio/pci: Add NVIDIA GPUDirect P2P clique support

[Qemu-devel] [PATCH 0/3] vfio/pci: Add NVIDIA GPUDirect P2P clique support

[Alex Williamson]

NVIDIA has a specification for exposing a virtual vendor capability
which provides a hint to guest drivers as to which sets of GPUs can
support direct peer-to-peer DMA.  Devices with the same clique ID are
expected to support this.  The user can specify a clique ID for an
NVIDIA graphics device using the new vfio-pci x-nv-gpudirect-clique=
option, where valid clique IDs are a 4-bit integer.  It's entirely the
user's responsibility to specify sets of devices for which P2P works
correctly and provides some benefit.  This is only useful for DMA
between NVIDIA GPUs, therefore it's only useful to specify cliques
comprised of more than one GPU.  Furthermore, this does not enable DMA
between VMs, there is no change to VM DMA mapping, this only exposes
hints about existing DMA paths to the guest driver.  Thanks,

Alex

---

Alex Williamson (3):
      vfio/pci: Do not unwind on error
      vfio/pci: Add virtual capabilities quirk infrastructure
      vfio/pci: Add NVIDIA GPUDirect Cliques support


 hw/vfio/pci-quirks.c |  114 ++++++++++++++++++++++++++++++++++++++++++++++++++
 hw/vfio/pci.c        |   17 +++++++
 hw/vfio/pci.h        |    4 ++
 3 files changed, 133 insertions(+), 2 deletions(-)

[Qemu-devel] [PATCH 1/3] vfio/pci: Do not unwind on error

[Alex Williamson]

If vfio_add_std_cap() errors then going to out prepends irrelevant
errors for capabilities we haven't attempted to add as we unwind our
recursive stack.  Just return error.

Fixes: 7ef165b9a8d9 ("vfio/pci: Pass an error object to vfio_add_capabilities")
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
---
 hw/vfio/pci.c |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
index 31e1edf44745..916d365dfab3 100644
--- a/hw/vfio/pci.c
+++ b/hw/vfio/pci.c
@@ -1826,7 +1826,7 @@ static int vfio_add_std_cap(VFIOPCIDevice *vdev, uint8_t pos, Error **errp)
     if (next) {
         ret = vfio_add_std_cap(vdev, next, errp);
         if (ret) {
-            goto out;
+            return ret;
         }
     } else {
         /* Begin the rebuild, use QEMU emulated list bits */
@@ -1862,7 +1862,7 @@ static int vfio_add_std_cap(VFIOPCIDevice *vdev, uint8_t pos, Error **errp)
         ret = pci_add_capability(pdev, cap_id, pos, size, errp);
         break;
     }
-out:
+
     if (ret < 0) {
         error_prepend(errp,
                       "failed to add PCI capability 0x%x[0x%x]@0x%x: ",

[Qemu-devel] [PATCH 2/3] vfio/pci: Add virtual capabilities quirk infrastructure

[Alex Williamson]

If the hypervisor needs to add purely virtual capabilties, give us a
hook through quirks to do that.  Note that we determine the maximum
size for a capability based on the physical device, if we insert a
virtual capability, that can change.  Therefore if maximum size is
smaller after added virt capabilities, use that.

Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
---
 hw/vfio/pci-quirks.c |    4 ++++
 hw/vfio/pci.c        |    8 ++++++++
 hw/vfio/pci.h        |    1 +
 3 files changed, 13 insertions(+)

diff --git a/hw/vfio/pci-quirks.c b/hw/vfio/pci-quirks.c
index 349085ea12bc..40aaae76feb9 100644
--- a/hw/vfio/pci-quirks.c
+++ b/hw/vfio/pci-quirks.c
@@ -1850,3 +1850,7 @@ void vfio_setup_resetfn_quirk(VFIOPCIDevice *vdev)
         break;
     }
 }
+int vfio_add_virt_caps(VFIOPCIDevice *vdev, Error **errp)
+{
+    return 0;
+}
diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
index 916d365dfab3..bfeaaef22d00 100644
--- a/hw/vfio/pci.c
+++ b/hw/vfio/pci.c
@@ -1833,8 +1833,16 @@ static int vfio_add_std_cap(VFIOPCIDevice *vdev, uint8_t pos, Error **errp)
         pdev->config[PCI_CAPABILITY_LIST] = 0;
         vdev->emulated_config_bits[PCI_CAPABILITY_LIST] = 0xff;
         vdev->emulated_config_bits[PCI_STATUS] |= PCI_STATUS_CAP_LIST;
+
+        ret = vfio_add_virt_caps(vdev, errp);
+        if (ret) {
+            return ret;
+        }
     }
 
+    /* Scale down size, esp in case virt caps were added above */
+    size = MIN(size, vfio_std_cap_max_size(pdev, pos));
+
     /* Use emulated next pointer to allow dropping caps */
     pci_set_byte(vdev->emulated_config_bits + pos + PCI_CAP_LIST_NEXT, 0xff);
 
diff --git a/hw/vfio/pci.h b/hw/vfio/pci.h
index a8366bb2a74a..958cee058b3b 100644
--- a/hw/vfio/pci.h
+++ b/hw/vfio/pci.h
@@ -160,6 +160,7 @@ void vfio_bar_quirk_setup(VFIOPCIDevice *vdev, int nr);
 void vfio_bar_quirk_exit(VFIOPCIDevice *vdev, int nr);
 void vfio_bar_quirk_finalize(VFIOPCIDevice *vdev, int nr);
 void vfio_setup_resetfn_quirk(VFIOPCIDevice *vdev);
+int vfio_add_virt_caps(VFIOPCIDevice *vdev, Error **errp);
 
 int vfio_populate_vga(VFIOPCIDevice *vdev, Error **errp);
 

[Qemu-devel] [PATCH 3/3] vfio/pci: Add NVIDIA GPUDirect Cliques support

[Alex Williamson]

NVIDIA has defined a specification for creating GPUDirect "cliques",
where devices with the same clique ID support direct peer-to-peer DMA.
When running on bare-metal, tools like NVIDIA's p2pBandwidthLatencyTest
(part of cuda-samples) determine which GPUs can support peer-to-peer
based on chipset and topology.  When running in a VM, these tools have
no visibility to the physical hardware support or topology.  This
option allows the user to specify hints via a vendor defined
capability.  For instance:

  <qemu:commandline>
    <qemu:arg value='-set'/>
    <qemu:arg value='device.hostdev0.x-nv-gpudirect-clique=0'/>
    <qemu:arg value='-set'/>
    <qemu:arg value='device.hostdev1.x-nv-gpudirect-clique=1'/>
    <qemu:arg value='-set'/>
    <qemu:arg value='device.hostdev2.x-nv-gpudirect-clique=1'/>
  </qemu:commandline>

This enables two cliques.  The first is a singleton clique with ID 0,
for the first hostdev defined in the XML (note that since cliques
define peer-to-peer sets, singleton clique offer no benefit).  The
subsequent two hostdevs are both added to clique ID 1, indicating
peer-to-peer is possible between these devices.

QEMU only provides validation that the clique ID is valid and applied
to an NVIDIA graphics device, any validation that the resulting
cliques are functional and valid is the user's responsibility.  The
NVIDIA specification allows a 4-bit clique ID, thus valid values are
0-15.

Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
---
 hw/vfio/pci-quirks.c |  110 ++++++++++++++++++++++++++++++++++++++++++++++++++
 hw/vfio/pci.c        |    5 ++
 hw/vfio/pci.h        |    3 +
 3 files changed, 118 insertions(+)

diff --git a/hw/vfio/pci-quirks.c b/hw/vfio/pci-quirks.c
index 40aaae76feb9..14291c2a16b2 100644
--- a/hw/vfio/pci-quirks.c
+++ b/hw/vfio/pci-quirks.c
@@ -14,6 +14,7 @@
 #include "qemu/error-report.h"
 #include "qemu/range.h"
 #include "qapi/error.h"
+#include "qapi/visitor.h"
 #include "hw/nvram/fw_cfg.h"
 #include "pci.h"
 #include "trace.h"
@@ -1850,7 +1851,116 @@ void vfio_setup_resetfn_quirk(VFIOPCIDevice *vdev)
         break;
     }
 }
+
+/*
+ * The NVIDIA GPUDirect P2P Vendor capability allows the user to specify
+ * devices as a member of a clique.  Devices within the same clique ID
+ * are capable of direct P2P.  It's the user's responsibility that this
+ * is correct.  The spec says that this may reside at any unused config
+ * offset, but reserves and recommends hypervisors place this at C8h.
+ * The spec also states that the hypervisor should place this capability
+ * at the end of the capability list, thus next is defined as 0h.
+ *
+ * +----------------+----------------+----------------+----------------+
+ * | sig 7:0 ('P')  |  vndr len (8h) |    next (0h)   |   cap id (9h)  |
+ * +----------------+----------------+----------------+----------------+
+ * | rsvd 15:7(0h),id 6:3,ver 2:0(0h)|          sig 23:8 ('P2')        |
+ * +---------------------------------+---------------------------------+
+ *
+ * https://lists.gnu.org/archive/html/qemu-devel/2017-08/pdfUda5iEpgOS.pdf
+ */
+static void get_nv_gpudirect_clique_id(Object *obj, Visitor *v,
+                                       const char *name, void *opaque,
+                                       Error **errp)
+{
+    DeviceState *dev = DEVICE(obj);
+    Property *prop = opaque;
+    uint8_t *ptr = qdev_get_prop_ptr(dev, prop);
+
+    visit_type_uint8(v, name, ptr, errp);
+}
+
+static void set_nv_gpudirect_clique_id(Object *obj, Visitor *v,
+                                       const char *name, void *opaque,
+                                       Error **errp)
+{
+    DeviceState *dev = DEVICE(obj);
+    Property *prop = opaque;
+    uint8_t value, *ptr = qdev_get_prop_ptr(dev, prop);
+    Error *local_err = NULL;
+
+    if (dev->realized) {
+        qdev_prop_set_after_realize(dev, name, errp);
+        return;
+    }
+
+    visit_type_uint8(v, name, &value, &local_err);
+    if (local_err) {
+        error_propagate(errp, local_err);
+        return;
+    }
+
+    if (value & ~0xF) {
+        error_setg(errp, "Property %s: valid range 0-15", name);
+        return;
+    }
+
+    *ptr = value;
+}
+
+const PropertyInfo qdev_prop_nv_gpudirect_clique = {
+    .name = "uint4",
+    .description = "NVIDIA GPUDirect Clique ID (0 - 15)",
+    .get = get_nv_gpudirect_clique_id,
+    .set = set_nv_gpudirect_clique_id,
+};
+
+static int vfio_add_nv_gpudirect_cap(VFIOPCIDevice *vdev, Error **errp)
+{
+    PCIDevice *pdev = &vdev->pdev;
+    int ret, pos = 0xC8;
+
+    if (vdev->nv_gpudirect_clique == 0xFF) {
+        return 0;
+    }
+
+    if (!vfio_pci_is(vdev, PCI_VENDOR_ID_NVIDIA, PCI_ANY_ID)) {
+        error_setg(errp, "NVIDIA GPUDirect Clique ID: invalid device vendor");
+        return -EINVAL;
+    }
+
+    if (pci_get_byte(pdev->config + PCI_CLASS_DEVICE + 1) !=
+        PCI_BASE_CLASS_DISPLAY) {
+        error_setg(errp, "NVIDIA GPUDirect Clique ID: unsupported PCI class");
+        return -EINVAL;
+    }
+
+    ret = pci_add_capability(pdev, PCI_CAP_ID_VNDR, pos, 8, errp);
+    if (ret < 0) {
+        error_prepend(errp, "Failed to add NVIDIA GPUDirect cap: ");
+        return ret;
+    }
+
+    memset(vdev->emulated_config_bits + pos, 0xFF, 8);
+    pos += PCI_CAP_FLAGS;
+    pci_set_byte(pdev->config + pos++, 8);
+    pci_set_byte(pdev->config + pos++, 'P');
+    pci_set_byte(pdev->config + pos++, '2');
+    pci_set_byte(pdev->config + pos++, 'P');
+    pci_set_byte(pdev->config + pos++, vdev->nv_gpudirect_clique << 3);
+    pci_set_byte(pdev->config + pos, 0);
+
+    return 0;
+}
+
 int vfio_add_virt_caps(VFIOPCIDevice *vdev, Error **errp)
 {
+    int ret;
+
+    ret = vfio_add_nv_gpudirect_cap(vdev, errp);
+    if (ret) {
+        return ret;
+    }
+
     return 0;
 }
diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
index bfeaaef22d00..9e86db7c3b6d 100644
--- a/hw/vfio/pci.c
+++ b/hw/vfio/pci.c
@@ -2970,6 +2970,8 @@ static void vfio_instance_init(Object *obj)
     vdev->host.bus = ~0U;
     vdev->host.slot = ~0U;
     vdev->host.function = ~0U;
+
+    vdev->nv_gpudirect_clique = 0xFF;
 }
 
 static Property vfio_pci_dev_properties[] = {
@@ -2994,6 +2996,9 @@ static Property vfio_pci_dev_properties[] = {
     DEFINE_PROP_UINT32("x-pci-sub-device-id", VFIOPCIDevice,
                        sub_device_id, PCI_ANY_ID),
     DEFINE_PROP_UINT32("x-igd-gms", VFIOPCIDevice, igd_gms, 0),
+    DEFINE_PROP_UNSIGNED_NODEFAULT("x-nv-gpudirect-clique", VFIOPCIDevice,
+                                   nv_gpudirect_clique,
+                                   qdev_prop_nv_gpudirect_clique, uint8_t),
     /*
      * TODO - support passed fds... is this necessary?
      * DEFINE_PROP_STRING("vfiofd", VFIOPCIDevice, vfiofd_name),
diff --git a/hw/vfio/pci.h b/hw/vfio/pci.h
index 958cee058b3b..502a5755b944 100644
--- a/hw/vfio/pci.h
+++ b/hw/vfio/pci.h
@@ -135,6 +135,7 @@ typedef struct VFIOPCIDevice {
     int32_t bootindex;
     uint32_t igd_gms;
     uint8_t pm_cap;
+    uint8_t nv_gpudirect_clique;
     bool pci_aer;
     bool req_enabled;
     bool has_flr;
@@ -162,6 +163,8 @@ void vfio_bar_quirk_finalize(VFIOPCIDevice *vdev, int nr);
 void vfio_setup_resetfn_quirk(VFIOPCIDevice *vdev);
 int vfio_add_virt_caps(VFIOPCIDevice *vdev, Error **errp);
 
+extern const PropertyInfo qdev_prop_nv_gpudirect_clique;
+
 int vfio_populate_vga(VFIOPCIDevice *vdev, Error **errp);
 
 int vfio_pci_igd_opregion_init(VFIOPCIDevice *vdev,

Re: [Qemu-devel] [PATCH 0/3] vfio/pci: Add NVIDIA GPUDirect P2P clique support

[Bob Chen]

There seem to be some bugs in these patches, causing my VM failed to boot.

Test case:

0. Merge these 3 patches in to release 2.10.1

1. qemu-system-x86_64_2.10.1  ... \
-device vfio-pci,host=04:00.0 \
-device vfio-pci,host=05:00.0 \
-device vfio-pci,host=08:00.0 \
-device vfio-pci,host=09:00.0 \
-device vfio-pci,host=85:00.0 \
-device vfio-pci,host=86:00.0 \
-device vfio-pci,host=89:00.0 \
-device vfio-pci,host=8a:00.0 ...

The guest was able to boot up.

2. qemu-system-x86_64_2.10.1  ... \
-device vfio-pci,host=04:00.0,x-nv-gpudirect-clique=0 \
-device vfio-pci,host=05:00.0,x-nv-gpudirect-clique=0 \
-device vfio-pci,host=08:00.0,x-nv-gpudirect-clique=0 \
-device vfio-pci,host=09:00.0,x-nv-gpudirect-clique=0 \
-device vfio-pci,host=85:00.0,x-nv-gpudirect-clique=8 \
-device vfio-pci,host=86:00.0,x-nv-gpudirect-clique=8 \
-device vfio-pci,host=89:00.0,x-nv-gpudirect-clique=8 \
-device vfio-pci,host=8a:00.0,x-nv-gpudirect-clique=8 \

Hang. VNC couldn't connect.


My personal patch used to work, although it was done by straightforward
hacking and not that friendly to read.

--- a/hw/vfio/pci.c
+++ b/hw/vfio/pci.c

 @@static int vfio_initfn(PCIDevice *pdev)

    vfio_add_emulated_long(vdev, 0xc8, 0x50080009, ~0);
    if (count < 4)
    {
        vfio_add_emulated_long(vdev, 0xcc, 0x00005032, ~0);
    }
    else
    {
        vfio_add_emulated_long(vdev, 0xcc, 0x00085032, ~0);
    }
    vfio_add_emulated_word(vdev, 0x78, 0xc810, ~0);

2017-08-30 6:05 GMT+08:00 Alex Williamson <alex.williamson@redhat.com>:

> NVIDIA has a specification for exposing a virtual vendor capability
> which provides a hint to guest drivers as to which sets of GPUs can
> support direct peer-to-peer DMA.  Devices with the same clique ID are
> expected to support this.  The user can specify a clique ID for an
> NVIDIA graphics device using the new vfio-pci x-nv-gpudirect-clique=
> option, where valid clique IDs are a 4-bit integer.  It's entirely the
> user's responsibility to specify sets of devices for which P2P works
> correctly and provides some benefit.  This is only useful for DMA
> between NVIDIA GPUs, therefore it's only useful to specify cliques
> comprised of more than one GPU.  Furthermore, this does not enable DMA
> between VMs, there is no change to VM DMA mapping, this only exposes
> hints about existing DMA paths to the guest driver.  Thanks,
>
> Alex
>
> ---
>
> Alex Williamson (3):
>       vfio/pci: Do not unwind on error
>       vfio/pci: Add virtual capabilities quirk infrastructure
>       vfio/pci: Add NVIDIA GPUDirect Cliques support
>
>
>  hw/vfio/pci-quirks.c |  114 ++++++++++++++++++++++++++++++
> ++++++++++++++++++++
>  hw/vfio/pci.c        |   17 +++++++
>  hw/vfio/pci.h        |    4 ++
>  3 files changed, 133 insertions(+), 2 deletions(-)
>

Re: [Qemu-devel] [PATCH 0/3] vfio/pci: Add NVIDIA GPUDirect P2P clique support

[Bob Chen]

It's a mistake, please ignore. This patch is able to work.

2017-10-26 18:45 GMT+08:00 Bob Chen <a175818323@gmail.com>:

> There seem to be some bugs in these patches, causing my VM failed to boot.
>
> Test case:
>
> 0. Merge these 3 patches in to release 2.10.1
>
> 1. qemu-system-x86_64_2.10.1  ... \
> -device vfio-pci,host=04:00.0 \
> -device vfio-pci,host=05:00.0 \
> -device vfio-pci,host=08:00.0 \
> -device vfio-pci,host=09:00.0 \
> -device vfio-pci,host=85:00.0 \
> -device vfio-pci,host=86:00.0 \
> -device vfio-pci,host=89:00.0 \
> -device vfio-pci,host=8a:00.0 ...
>
> The guest was able to boot up.
>
> 2. qemu-system-x86_64_2.10.1  ... \
> -device vfio-pci,host=04:00.0,x-nv-gpudirect-clique=0 \
> -device vfio-pci,host=05:00.0,x-nv-gpudirect-clique=0 \
> -device vfio-pci,host=08:00.0,x-nv-gpudirect-clique=0 \
> -device vfio-pci,host=09:00.0,x-nv-gpudirect-clique=0 \
> -device vfio-pci,host=85:00.0,x-nv-gpudirect-clique=8 \
> -device vfio-pci,host=86:00.0,x-nv-gpudirect-clique=8 \
> -device vfio-pci,host=89:00.0,x-nv-gpudirect-clique=8 \
> -device vfio-pci,host=8a:00.0,x-nv-gpudirect-clique=8 \
>
> Hang. VNC couldn't connect.
>
>
> My personal patch used to work, although it was done by straightforward
> hacking and not that friendly to read.
>
> --- a/hw/vfio/pci.c
> +++ b/hw/vfio/pci.c
>
>  @@static int vfio_initfn(PCIDevice *pdev)
>
>     vfio_add_emulated_long(vdev, 0xc8, 0x50080009, ~0);
>     if (count < 4)
>     {
>         vfio_add_emulated_long(vdev, 0xcc, 0x00005032, ~0);
>     }
>     else
>     {
>         vfio_add_emulated_long(vdev, 0xcc, 0x00085032, ~0);
>     }
>     vfio_add_emulated_word(vdev, 0x78, 0xc810, ~0);
>
> 2017-08-30 6:05 GMT+08:00 Alex Williamson <alex.williamson@redhat.com>:
>
>> NVIDIA has a specification for exposing a virtual vendor capability
>> which provides a hint to guest drivers as to which sets of GPUs can
>> support direct peer-to-peer DMA.  Devices with the same clique ID are
>> expected to support this.  The user can specify a clique ID for an
>> NVIDIA graphics device using the new vfio-pci x-nv-gpudirect-clique=
>> option, where valid clique IDs are a 4-bit integer.  It's entirely the
>> user's responsibility to specify sets of devices for which P2P works
>> correctly and provides some benefit.  This is only useful for DMA
>> between NVIDIA GPUs, therefore it's only useful to specify cliques
>> comprised of more than one GPU.  Furthermore, this does not enable DMA
>> between VMs, there is no change to VM DMA mapping, this only exposes
>> hints about existing DMA paths to the guest driver.  Thanks,
>>
>> Alex
>>
>> ---
>>
>> Alex Williamson (3):
>>       vfio/pci: Do not unwind on error
>>       vfio/pci: Add virtual capabilities quirk infrastructure
>>       vfio/pci: Add NVIDIA GPUDirect Cliques support
>>
>>
>>  hw/vfio/pci-quirks.c |  114 ++++++++++++++++++++++++++++++
>> ++++++++++++++++++++
>>  hw/vfio/pci.c        |   17 +++++++
>>  hw/vfio/pci.h        |    4 ++
>>  3 files changed, 133 insertions(+), 2 deletions(-)
>>
>
>