Re: [PATCH 05/15] hw/nvme: Add support for SR-IOV

[Lukasz Maniak]

On Tue, Nov 02, 2021 at 03:33:15PM +0100, Klaus Jensen wrote:
> On Oct  7 18:23, Lukasz Maniak wrote:
> > This patch implements initial support for Single Root I/O Virtualization
> > on an NVMe device.
> > 
> > Essentially, it allows to define the maximum number of virtual functions
> > supported by the NVMe controller via sriov_max_vfs parameter.
> > 
> > Passing a non-zero value to sriov_max_vfs triggers reporting of SR-IOV
> > capability by a physical controller and ARI capability by both the
> > physical and virtual function devices.
> > 
> > NVMe controllers created via virtual functions mirror functionally
> > the physical controller, which may not entirely be the case, thus
> > consideration would be needed on the way to limit the capabilities of
> > the VF.
> > 
> > NVMe subsystem is required for the use of SR-IOV.
> > 
> > Signed-off-by: Lukasz Maniak <lukasz.maniak@linux.intel.com>
> > ---
> >  hw/nvme/ctrl.c           | 74 ++++++++++++++++++++++++++++++++++++++--
> >  hw/nvme/nvme.h           |  1 +
> >  include/hw/pci/pci_ids.h |  1 +
> >  3 files changed, 73 insertions(+), 3 deletions(-)
> > 
> > diff --git a/hw/nvme/ctrl.c b/hw/nvme/ctrl.c
> > index 6a571d18cf..ad79ff0c00 100644
> > --- a/hw/nvme/ctrl.c
> > +++ b/hw/nvme/ctrl.c
> > @@ -6361,8 +6406,12 @@ static int nvme_init_pci(NvmeCtrl *n, PCIDevice 
> > *pci_dev, Error **errp)
> >                            n->reg_size);
> >      memory_region_add_subregion(&n->bar0, 0, &n->iomem);
> >  
> > -    pci_register_bar(pci_dev, 0, PCI_BASE_ADDRESS_SPACE_MEMORY |
> > -                     PCI_BASE_ADDRESS_MEM_TYPE_64, &n->bar0);
> > +    if (pci_is_vf(pci_dev)) {
> > +        pcie_sriov_vf_register_bar(pci_dev, 0, &n->bar0);
> > +    } else {
> > +        pci_register_bar(pci_dev, 0, PCI_BASE_ADDRESS_SPACE_MEMORY |
> > +                         PCI_BASE_ADDRESS_MEM_TYPE_64, &n->bar0);
> > +    }
> 
> I assume that the assert we are seeing means that the pci_register_bars
> in nvme_init_cmb and nvme_init_pmr must be changed similarly to this.

Assert will only arise for CMB as VF params are initialized with PF
params.

@@ -6532,6 +6585,15 @@ static void nvme_realize(PCIDevice *pci_dev, Error 
**errp)
     NvmeCtrl *n = NVME(pci_dev);
     NvmeNamespace *ns;
     Error *local_err = NULL;
+    NvmeCtrl *pn = NVME(pcie_sriov_get_pf(pci_dev));
+
+    if (pci_is_vf(pci_dev)) {
+        /* VFs derive settings from the parent. PF's lifespan exceeds
+         * that of VF's, so it's safe to share params.serial.
+         */
+        memcpy(&n->params, &pn->params, sizeof(NvmeParams));
+        n->subsys = pn->subsys;
+    }
 
     nvme_check_constraints(n, &local_err);
     if (local_err) {

The following simple fix will both fix assert and also allow
each VF to have its own CMB of the size defined for PF.

---
 hw/nvme/ctrl.c | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/hw/nvme/ctrl.c b/hw/nvme/ctrl.c
index 19b32dd4da..99daa6290c 100644
--- a/hw/nvme/ctrl.c
+++ b/hw/nvme/ctrl.c
@@ -6837,10 +6837,15 @@ static void nvme_init_cmb(NvmeCtrl *n, PCIDevice 
*pci_dev)
     n->cmb.buf = g_malloc0(cmb_size);
     memory_region_init_io(&n->cmb.mem, OBJECT(n), &nvme_cmb_ops, n,
                           "nvme-cmb", cmb_size);
-    pci_register_bar(pci_dev, NVME_CMB_BIR,
-                     PCI_BASE_ADDRESS_SPACE_MEMORY |
-                     PCI_BASE_ADDRESS_MEM_TYPE_64 |
-                     PCI_BASE_ADDRESS_MEM_PREFETCH, &n->cmb.mem);
+
+    if (pci_is_vf(pci_dev)) {
+        pcie_sriov_vf_register_bar(pci_dev, NVME_CMB_BIR, &n->cmb.mem);
+    } else {
+        pci_register_bar(pci_dev, NVME_CMB_BIR,
+                        PCI_BASE_ADDRESS_SPACE_MEMORY |
+                        PCI_BASE_ADDRESS_MEM_TYPE_64 |
+                        PCI_BASE_ADDRESS_MEM_PREFETCH, &n->cmb.mem);
+    }
 
     NVME_CAP_SET_CMBS(cap, 1);
     stq_le_p(&n->bar.cap, cap);

As for PMR, it is currently only available on PF, as only PF is capable
of specifying the memory-backend-file object to use with PMR.
Otherwise, either VFs would have to share the PMR with its PF, or there
would be a requirement to define a memory-backend-file object for each VF.