Re: [PATCH 3/3] qemu-kvm: device assignment: emulate MSI-X mask bits

["Michael S. Tsirkin" <mst@redhat.com>]

On Thu, Nov 04, 2010 at 02:18:21PM +0800, Sheng Yang wrote:
> This patch emulated MSI-X per vector mask bit on assigned device.
> 
> Signed-off-by: Sheng Yang <sheng@linux.intel.com>
> ---
>  hw/device-assignment.c |  161 ++++++++++++++++++++++++++++++++++++++++++++++--
>  1 files changed, 155 insertions(+), 6 deletions(-)
> 
> diff --git a/hw/device-assignment.c b/hw/device-assignment.c
> index 8a98876..639aa0b 100644
> --- a/hw/device-assignment.c
> +++ b/hw/device-assignment.c
> @@ -62,6 +62,11 @@ static void assigned_dev_load_option_rom(AssignedDevice *dev);
>  
>  static void assigned_dev_unregister_msix_mmio(AssignedDevice *dev);
>  
> +static uint32_t calc_assigned_dev_id(uint16_t seg, uint8_t bus, uint8_t devfn)
> +{
> +    return (uint32_t)seg << 16 | (uint32_t)bus << 8 | (uint32_t)devfn;
> +}
> +
>  static uint32_t assigned_dev_ioport_rw(AssignedDevRegion *dev_region,
>                                         uint32_t addr, int len, uint32_t *val)
>  {
> @@ -264,6 +269,9 @@ static void assigned_dev_iomem_map(PCIDevice *pci_dev, int region_num,
>      AssignedDevRegion *region = &r_dev->v_addrs[region_num];
>      PCIRegion *real_region = &r_dev->real_device.regions[region_num];
>      int ret = 0;
> +#ifdef KVM_CAP_DEVICE_MSIX_MASK
> +    struct kvm_assigned_msix_mmio msix_mmio;
> +#endif
>  
>      DEBUG("e_phys=%08" FMT_PCIBUS " r_virt=%p type=%d len=%08" FMT_PCIBUS " region_num=%d \n",
>            e_phys, region->u.r_virtbase, type, e_size, region_num);
> @@ -282,6 +290,16 @@ static void assigned_dev_iomem_map(PCIDevice *pci_dev, int region_num,
>  
>              cpu_register_physical_memory(e_phys + offset,
>                      TARGET_PAGE_SIZE, r_dev->mmio_index);
> +#ifdef KVM_CAP_DEVICE_MSIX_MASK
> +	    memset(&msix_mmio, 0, sizeof(struct kvm_assigned_msix_mmio));
> +	    msix_mmio.assigned_dev_id = calc_assigned_dev_id(r_dev->h_segnr,
> +			    r_dev->h_busnr, r_dev->h_devfn);
> +	    msix_mmio.base_addr = e_phys + offset;
> +            /* We required kernel MSI-X support */
> +	    ret = kvm_assign_reg_msix_mmio(kvm_context, &msix_mmio);
> +	    if (ret)
> +                fprintf(stderr, "fail to register in-kernel msix_mmio!\n");
> +#endif
>          }
>      }
>  
> @@ -824,11 +842,6 @@ static void free_assigned_device(AssignedDevice *dev)
>      }
>  }
>  
> -static uint32_t calc_assigned_dev_id(uint16_t seg, uint8_t bus, uint8_t devfn)
> -{
> -    return (uint32_t)seg << 16 | (uint32_t)bus << 8 | (uint32_t)devfn;
> -}
> -
>  static void assign_failed_examine(AssignedDevice *dev)
>  {
>      char name[PATH_MAX], dir[PATH_MAX], driver[PATH_MAX] = {}, *ns;
> @@ -1123,6 +1136,27 @@ static int get_msix_entries_max_nr(AssignedDevice *adev)
>      return entries_max_nr;
>  }
>  
> +#ifdef KVM_CAP_DEVICE_MSIX_MASK
> +static int assigned_dev_msix_entry_masked(AssignedDevice *adev, int entry)
> +{
> +    struct kvm_assigned_msix_entry msix_entry;
> +    int r;
> +
> +    memset(&msix_entry, 0, sizeof msix_entry);
> +    msix_entry.assigned_dev_id = calc_assigned_dev_id(adev->h_segnr,
> +            adev->h_busnr, (uint8_t)adev->h_devfn);
> +    msix_entry.entry = entry;
> +    msix_entry.flags = KVM_MSIX_FLAG_QUERY_MASK;
> +    r = kvm_assign_get_msix_entry(kvm_context, &msix_entry);
> +    if (r) {
> +        fprintf(stderr, "assigned_dev_msix_entry_masked: "
> +			"Fail to get mask bit of entry %d\n", entry);
> +        return 1;

This error handling seems pretty useless. assert?

> +    }
> +    return (msix_entry.flags & KVM_MSIX_FLAG_MASK);
> +}
> +#endif
> +
>  static int get_msix_valid_entries_nr(AssignedDevice *adev,
>  				     uint16_t entries_max_nr)
>  {
> @@ -1136,7 +1170,11 @@ static int get_msix_valid_entries_nr(AssignedDevice *adev,
>          memcpy(&msg_ctrl, va + i * 16 + 12, 4);
>          memcpy(&msg_data, va + i * 16 + 8, 4);
>          /* Ignore unused entry even it's unmasked */
> +#ifdef KVM_CAP_DEVICE_MSIX_MASK
> +        if (assigned_dev_msix_entry_masked(adev, i))
> +#else
>          if (msg_data == 0)
> +#endif

So, we are replacing msg_data == 0 check with masked check?
If yes why not do this for non-KVM_CAP_DEVICE_MSIX_MASK too?


>              continue;
>          entries_nr ++;
>      }
> @@ -1165,6 +1203,8 @@ static int assigned_dev_update_msix_mmio(PCIDevice *pci_dev,
>      }
>  
>      free_dev_irq_entries(adev);
> +    memset(pci_dev->msix_entry_used, 0, KVM_MAX_MSIX_PER_DEV *
> +                                        sizeof(*pci_dev->msix_entry_used));
>      adev->irq_entries_nr = entries_nr;
>      adev->entry = calloc(entries_nr, sizeof(struct kvm_irq_routing_entry));
>      if (!adev->entry) {
> @@ -1179,7 +1219,11 @@ static int assigned_dev_update_msix_mmio(PCIDevice *pci_dev,
>              break;
>          memcpy(&msg_ctrl, va + i * 16 + 12, 4);
>          memcpy(&msg_data, va + i * 16 + 8, 4);
> +#ifdef KVM_CAP_DEVICE_MSIX_MASK
> +        if (assigned_dev_msix_entry_masked(adev, i))
> +#else
>          if (msg_data == 0)
> +#endif

You can't use ifdef to check that kernel supports an ioctl.
You must check this at runtime.

>              continue;
>  
>          memcpy(&msg_addr, va + i * 16, 4);
> @@ -1200,6 +1244,7 @@ static int assigned_dev_update_msix_mmio(PCIDevice *pci_dev,
>  
>          msix_entry.gsi = adev->entry[entries_nr].gsi;
>          msix_entry.entry = i;
> +        pci_dev->msix_entry_used[i] = 1;
>          r = kvm_assign_set_msix_entry(kvm_context, &msix_entry);
>          if (r) {
>              fprintf(stderr, "fail to set MSI-X entry! %s\n", strerror(-r));
> @@ -1243,6 +1288,8 @@ static void assigned_dev_update_msix(PCIDevice *pci_dev, int enable_msix)
>              perror("assigned_dev_update_msix: deassign irq");
>  
>          assigned_dev->irq_requested_type = 0;
> +        memset(pci_dev->msix_entry_used, 0, KVM_MAX_MSIX_PER_DEV *
> +                                        sizeof(*pci_dev->msix_entry_used));
>      }
>  
>      entries_max_nr = get_msix_entries_max_nr(assigned_dev);
> @@ -1250,10 +1297,16 @@ static void assigned_dev_update_msix(PCIDevice *pci_dev, int enable_msix)
>          fprintf(stderr, "assigned_dev_update_msix: MSI-X entries_max_nr == 0");
>          return;
>      }
> +    /*
> +     * Guest may try to enable MSI-X before setting MSI-X entry done, so
> +     * let's wait until guest unmask the entries.
> +     */

Well it can also set up any number of entries, enable msix then
set up more entries. Now what?

>      entries_nr = get_msix_valid_entries_nr(assigned_dev, entries_max_nr);
>      if (entries_nr == 0) {
> +#ifndef KVM_CAP_DEVICE_MSIX_MASK
>          if (enable_msix)
>              fprintf(stderr, "MSI-X entry number is zero!\n");

And what happens then?

> +#endif
>          return;
>      }
>      if (enable_msix) {
> @@ -1297,7 +1350,8 @@ static void assigned_device_pci_cap_write_config(PCIDevice *pci_dev, uint32_t ad
>          if (address <= ctrl_pos && address + len > ctrl_pos) {
>              ctrl_pos--; /* control is word long */
>              ctrl_word = (uint16_t *)(pci_dev->config + ctrl_pos);
> -            assigned_dev_update_msix(pci_dev, (*ctrl_word & PCI_MSIX_ENABLE));
> +            assigned_dev_update_msix(pci_dev,
> +                    (*ctrl_word & PCI_MSIX_ENABLE) && !(*ctrl_word & PCI_MSIX_MASK));
>  	}
>          pos += PCI_CAPABILITY_CONFIG_MSIX_LENGTH;
>      }
> @@ -1395,10 +1449,101 @@ static void msix_mmio_writel(void *opaque,
>      AssignedDevice *adev = opaque;
>      unsigned int offset = addr & 0xfff;
>      void *page = adev->msix_table_page;
> +#ifdef KVM_CAP_DEVICE_MSIX_MASK
> +    int pos, ctrl_word, index;
> +    struct kvm_irq_routing_entry new_entry = {};
> +    int entry_idx, entries_max_nr, r = 0, i;
> +    uint32_t msg_ctrl, msg_data, msg_upper_addr, msg_addr;
> +#endif
>  
>      DEBUG("write to MSI-X entry table mmio offset 0x%lx, val 0x%x\n",
>  		    addr, val);
>      memcpy((void *)((char *)page + offset), &val, 4);
> +
> +#ifdef KVM_CAP_DEVICE_MSIX_MASK
> +    index = offset / 16;
> +
> +    /* Check if mask bit is being accessed */
> +    memcpy(&msg_addr, (char *)page + index * 16, 4);
> +    memcpy(&msg_upper_addr, (char *)page + index * 16 + 4, 4);
> +    memcpy(&msg_data, (char *)page + index * 16 + 8, 4);
> +    memcpy(&msg_ctrl, (char *)page + index * 16 + 12, 4);
> +    DEBUG("MSI-X entries index %d: "
> +            "msg_addr 0x%x, msg_upper_addr 0x%x, msg_data 0x%x, vec_ctl 0x%x\n",
> +            index, msg_addr, msg_upper_addr, msg_data, msg_ctrl);
> +
> +    if (adev->cap.available & ASSIGNED_DEVICE_CAP_MSI)
> +        pos = adev->dev.cap.start + PCI_CAPABILITY_CONFIG_MSI_LENGTH;
> +    else
> +        pos = adev->dev.cap.start;
> +
> +    ctrl_word = *(uint16_t *)(adev->dev.config + pos + 2);
> +
> +    if (!((ctrl_word & PCI_MSIX_ENABLE) && !(ctrl_word & PCI_MSIX_MASK)))
> +        return;
> +
> +    if (!assigned_dev_msix_entry_masked(adev, index)) {
> +        if (!adev->dev.msix_entry_used[index]) {
> +            DEBUG("Try to modify unenabled MSI-X entry %d's mask. "
> +                    "Reenable MSI-X.\n",
> +                    index);
> +            assigned_dev_update_msix(&adev->dev, 1);
> +        }
> +        return;
> +    }
> +
> +    if (!adev->dev.msix_entry_used[index])
> +        return;
> +
> +    /*
> +     * We're here only because guest want to modify MSI data/addr, and
> +     * kernel would filter those writing with mask bit unset.
> +     */
> +    entries_max_nr = get_msix_entries_max_nr(adev);
> +
> +    /*
> +     * Find the index of routing entry, it can be different from 'index' if
> +     * empty entry existed in between
> +     */
> +    entry_idx = -1;
> +    for (i = 0; i <= index; i++) {
> +        if (adev->dev.msix_entry_used[i])
> +            entry_idx ++;
> +    }
> +    if (entry_idx >= entries_max_nr || entry_idx == -1) {
> +        fprintf(stderr, "msix_mmio_writel: Entry idx %d exceed limit!\n",
> +			entry_idx);
> +        return;
> +    }
> +
> +    if (!assigned_dev_msix_entry_masked(adev, index)) {
> +        fprintf(stderr, "msix_mmio_writel: Trying write to unmasked entry!\n");
> +        return;
> +    }
> +
> +    new_entry.gsi = adev->entry[entry_idx].gsi;
> +    new_entry.type = KVM_IRQ_ROUTING_MSI;
> +    new_entry.flags = 0;
> +    new_entry.u.msi.address_lo = msg_addr;
> +    new_entry.u.msi.address_hi = msg_upper_addr;
> +    new_entry.u.msi.data = msg_data;
> +    if (memcmp(&adev->entry[entry_idx].u.msi, &new_entry.u.msi,
> +                sizeof new_entry.u.msi)) {
> +        r = kvm_update_routing_entry(&adev->entry[entry_idx], &new_entry);
> +        if (r) {
> +            perror("msix_mmio_writel: kvm_update_routing_entry failed\n");
> +            return;
> +        }
> +        r = kvm_commit_irq_routes();
> +        if (r) {
> +            perror("msix_mmio_writel: kvm_commit_irq_routes failed\n");
> +            return;
> +        }
> +    }
> +    adev->entry[entry_idx].u.msi.address_lo = msg_addr;
> +    adev->entry[entry_idx].u.msi.address_hi = msg_upper_addr;
> +    adev->entry[entry_idx].u.msi.data = msg_data;
> +#endif
>  }
>  
>  static void msix_mmio_writew(void *opaque,
> @@ -1436,6 +1581,8 @@ static int assigned_dev_register_msix_mmio(AssignedDevice *dev)
>      memset(dev->msix_table_page, 0, 0x1000);
>      dev->mmio_index = cpu_register_io_memory(
>                          msix_mmio_read, msix_mmio_write, dev);
> +    dev->dev.msix_entry_used = qemu_mallocz(KVM_MAX_MSIX_PER_DEV *
> +                                        sizeof *dev->dev.msix_entry_used);
>      return 0;
>  }
>  
> @@ -1452,6 +1599,8 @@ static void assigned_dev_unregister_msix_mmio(AssignedDevice *dev)
>                  strerror(errno));
>      }
>      dev->msix_table_page = NULL;
> +    free(dev->dev.msix_entry_used);
> +    dev->dev.msix_entry_used = NULL;
>  }
>  
>  static int assigned_initfn(struct PCIDevice *pci_dev)
> -- 
> 1.7.0.1