[Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[Alexey Kardashevskiy]

This is for passing through NVIDIA V100 GPUs on POWER9 systems.

This implements a subdriver for NVIDIA V100 GPU with coherent memory and
NPU/ATS support available in the POWER9 CPU.

1/3 is not strictly related but since new memory also needs to be mapped
to the 64bit DMA window and it is located quite high in the address space,
some adjustments are needed.


This is based on dwg/ppc-for-4.0 sha1 a0a8bff and requires headers update
from v5.0-rc1 staged by Paolo already.

Please comment. Thanks.



Alexey Kardashevskiy (3):
  vfio/spapr: Fix indirect levels calculation
  vfio: Make vfio_get_region_info_cap public
  spapr: Support NVIDIA V100 GPU with NVLink2

 hw/vfio/pci.h                 |   2 +
 include/hw/pci-host/spapr.h   |   9 +
 include/hw/ppc/spapr.h        |   3 +-
 include/hw/vfio/vfio-common.h |   2 +
 hw/ppc/spapr.c                |  25 ++-
 hw/ppc/spapr_pci.c            | 333 +++++++++++++++++++++++++++++++++-
 hw/vfio/common.c              |   2 +-
 hw/vfio/pci-quirks.c          | 120 ++++++++++++
 hw/vfio/pci.c                 |  14 ++
 hw/vfio/spapr.c               |  38 +++-
 hw/vfio/trace-events          |   6 +-
 11 files changed, 539 insertions(+), 15 deletions(-)

-- 
2.17.1

[Qemu-devel] [PATCH qemu 1/3] vfio/spapr: Fix indirect levels calculation

[Alexey Kardashevskiy]

The current code assumes that we can address more bits on a PCI bus
for DMA than we really can but there is no way knowing the actual limit.

This makes a better guess for the number of levels and if the kernel
fails to allocate that, this increases the level numbers till succeeded
or reached the 64bit limit.

This adds levels to the trace point.

This may cause the kernel to warn about failed allocation:
   [65122.837458] Failed to allocate a TCE memory, level shift=28
which might happen if MAX_ORDER is not large enough as it can vary:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/powerpc/Kconfig?h=v5.0-rc2#n727

Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
---
 hw/vfio/spapr.c      | 38 +++++++++++++++++++++++++++++---------
 hw/vfio/trace-events |  2 +-
 2 files changed, 30 insertions(+), 10 deletions(-)

diff --git a/hw/vfio/spapr.c b/hw/vfio/spapr.c
index becf71a..2675bf5 100644
--- a/hw/vfio/spapr.c
+++ b/hw/vfio/spapr.c
@@ -146,7 +146,7 @@ int vfio_spapr_create_window(VFIOContainer *container,
     int ret;
     IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
     uint64_t pagesize = memory_region_iommu_get_min_page_size(iommu_mr);
-    unsigned entries, pages;
+    unsigned entries, bits_total, bits_per_level, max_levels;
     struct vfio_iommu_spapr_tce_create create = { .argsz = sizeof(create) };
     long systempagesize = qemu_getrampagesize();
 
@@ -176,16 +176,35 @@ int vfio_spapr_create_window(VFIOContainer *container,
     create.window_size = int128_get64(section->size);
     create.page_shift = ctz64(pagesize);
     /*
-     * SPAPR host supports multilevel TCE tables, there is some
-     * heuristic to decide how many levels we want for our table:
-     * 0..64 = 1; 65..4096 = 2; 4097..262144 = 3; 262145.. = 4
+     * SPAPR host supports multilevel TCE tables. We try to guess optimal
+     * levels number and if this fails (for example due to the host memory
+     * fragmentation), we increase levels. The DMA address structure is:
+     * rrrrrrrr rxxxxxxx xxxxxxxx xxxxxxxx  xxxxxxxx xxxxxxxx xxxxxxxx iiiiiiii
+     * where:
+     *   r = reserved (bits >= 55 are reserved in the existing hardware)
+     *   i = IOMMU page offset (64K in this example)
+     *   x = bits to index a TCE which can be split to equal chunks to index
+     *      within the level.
+     * The aim is to split "x" to smaller possible number of levels.
      */
     entries = create.window_size >> create.page_shift;
-    pages = MAX((entries * sizeof(uint64_t)) / getpagesize(), 1);
-    pages = MAX(pow2ceil(pages), 1); /* Round up */
-    create.levels = ctz64(pages) / 6 + 1;
-
-    ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_CREATE, &create);
+    /* bits_total is number of "x" needed */
+    bits_total = ctz64((entries * sizeof(uint64_t)));
+    /*
+     * bits_per_level is a safe guess of how much we can allocate per level:
+     * 8 is the current minimum for CONFIG_FORCE_MAX_ZONEORDER and MAX_ORDER
+     * is usually bigger than that.
+     */
+    bits_per_level = ctz64(systempagesize) + 8;
+    create.levels = bits_total / bits_per_level;
+    create.levels = MAX(1, create.levels);
+    max_levels = (64 - create.page_shift) / ctz64(systempagesize);
+    for ( ; create.levels <= max_levels; ++create.levels) {
+        ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_CREATE, &create);
+        if (!ret) {
+            break;
+        }
+    }
     if (ret) {
         error_report("Failed to create a window, ret = %d (%m)", ret);
         return -errno;
@@ -200,6 +219,7 @@ int vfio_spapr_create_window(VFIOContainer *container,
         return -EINVAL;
     }
     trace_vfio_spapr_create_window(create.page_shift,
+                                   create.levels,
                                    create.window_size,
                                    create.start_addr);
     *pgsize = pagesize;
diff --git a/hw/vfio/trace-events b/hw/vfio/trace-events
index 0f9f810..adfa75e 100644
--- a/hw/vfio/trace-events
+++ b/hw/vfio/trace-events
@@ -129,6 +129,6 @@ vfio_prereg_listener_region_add_skip(uint64_t start, uint64_t end) "0x%"PRIx64"
 vfio_prereg_listener_region_del_skip(uint64_t start, uint64_t end) "0x%"PRIx64" - 0x%"PRIx64
 vfio_prereg_register(uint64_t va, uint64_t size, int ret) "va=0x%"PRIx64" size=0x%"PRIx64" ret=%d"
 vfio_prereg_unregister(uint64_t va, uint64_t size, int ret) "va=0x%"PRIx64" size=0x%"PRIx64" ret=%d"
-vfio_spapr_create_window(int ps, uint64_t ws, uint64_t off) "pageshift=0x%x winsize=0x%"PRIx64" offset=0x%"PRIx64
+vfio_spapr_create_window(int ps, unsigned int levels, uint64_t ws, uint64_t off) "pageshift=0x%x levels=%u winsize=0x%"PRIx64" offset=0x%"PRIx64
 vfio_spapr_remove_window(uint64_t off) "offset=0x%"PRIx64
 vfio_spapr_group_attach(int groupfd, int tablefd) "Attached groupfd %d to liobn fd %d"
-- 
2.17.1

Re: [Qemu-devel] [PATCH qemu 1/3] vfio/spapr: Fix indirect levels calculation

[David Gibson]

[-- Attachment #1: Type: text/plain, Size: 5875 bytes --]

On Thu, Jan 17, 2019 at 01:51:13PM +1100, Alexey Kardashevskiy wrote:
> The current code assumes that we can address more bits on a PCI bus
> for DMA than we really can but there is no way knowing the actual limit.
> 
> This makes a better guess for the number of levels and if the kernel
> fails to allocate that, this increases the level numbers till succeeded
> or reached the 64bit limit.
> 
> This adds levels to the trace point.
> 
> This may cause the kernel to warn about failed allocation:
>    [65122.837458] Failed to allocate a TCE memory, level shift=28
> which might happen if MAX_ORDER is not large enough as it can vary:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/powerpc/Kconfig?h=v5.0-rc2#n727
> 
> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
> ---
>  hw/vfio/spapr.c      | 38 +++++++++++++++++++++++++++++---------
>  hw/vfio/trace-events |  2 +-
>  2 files changed, 30 insertions(+), 10 deletions(-)
> 
> diff --git a/hw/vfio/spapr.c b/hw/vfio/spapr.c
> index becf71a..2675bf5 100644
> --- a/hw/vfio/spapr.c
> +++ b/hw/vfio/spapr.c
> @@ -146,7 +146,7 @@ int vfio_spapr_create_window(VFIOContainer *container,
>      int ret;
>      IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
>      uint64_t pagesize = memory_region_iommu_get_min_page_size(iommu_mr);
> -    unsigned entries, pages;
> +    unsigned entries, bits_total, bits_per_level, max_levels;
>      struct vfio_iommu_spapr_tce_create create = { .argsz = sizeof(create) };
>      long systempagesize = qemu_getrampagesize();

So, "systempagesize" is a bad name, since this is *not* the system
page size, but rather the (minimum) page size we're using to map guest
RAM...

> @@ -176,16 +176,35 @@ int vfio_spapr_create_window(VFIOContainer *container,
>      create.window_size = int128_get64(section->size);
>      create.page_shift = ctz64(pagesize);
>      /*
> -     * SPAPR host supports multilevel TCE tables, there is some
> -     * heuristic to decide how many levels we want for our table:
> -     * 0..64 = 1; 65..4096 = 2; 4097..262144 = 3; 262145.. = 4
> +     * SPAPR host supports multilevel TCE tables. We try to guess optimal
> +     * levels number and if this fails (for example due to the host memory
> +     * fragmentation), we increase levels. The DMA address structure is:
> +     * rrrrrrrr rxxxxxxx xxxxxxxx xxxxxxxx  xxxxxxxx xxxxxxxx xxxxxxxx iiiiiiii
> +     * where:
> +     *   r = reserved (bits >= 55 are reserved in the existing hardware)
> +     *   i = IOMMU page offset (64K in this example)
> +     *   x = bits to index a TCE which can be split to equal chunks to index
> +     *      within the level.
> +     * The aim is to split "x" to smaller possible number of levels.
>       */
>      entries = create.window_size >> create.page_shift;
> -    pages = MAX((entries * sizeof(uint64_t)) / getpagesize(), 1);
> -    pages = MAX(pow2ceil(pages), 1); /* Round up */
> -    create.levels = ctz64(pages) / 6 + 1;
> -
> -    ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_CREATE, &create);
> +    /* bits_total is number of "x" needed */
> +    bits_total = ctz64((entries * sizeof(uint64_t)));
> +    /*
> +     * bits_per_level is a safe guess of how much we can allocate per level:
> +     * 8 is the current minimum for CONFIG_FORCE_MAX_ZONEORDER and MAX_ORDER
> +     * is usually bigger than that.
> +     */
> +    bits_per_level = ctz64(systempagesize) + 8;

..which is relevant, because here you really do want the system page
size, since you're guestimating what the kernel will be able to put on
a page, entirely unrelated to the guest page size.

> +    create.levels = bits_total / bits_per_level;

Don't you need a div-rounded-up here?

> +    create.levels = MAX(1, create.levels);
> +    max_levels = (64 - create.page_shift) / ctz64(systempagesize);

And I think you want the real system page size here again.

> +    for ( ; create.levels <= max_levels; ++create.levels) {
> +        ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_CREATE, &create);
> +        if (!ret) {
> +            break;
> +        }
> +    }
>      if (ret) {
>          error_report("Failed to create a window, ret = %d (%m)", ret);
>          return -errno;
> @@ -200,6 +219,7 @@ int vfio_spapr_create_window(VFIOContainer *container,
>          return -EINVAL;
>      }
>      trace_vfio_spapr_create_window(create.page_shift,
> +                                   create.levels,
>                                     create.window_size,
>                                     create.start_addr);
>      *pgsize = pagesize;
> diff --git a/hw/vfio/trace-events b/hw/vfio/trace-events
> index 0f9f810..adfa75e 100644
> --- a/hw/vfio/trace-events
> +++ b/hw/vfio/trace-events
> @@ -129,6 +129,6 @@ vfio_prereg_listener_region_add_skip(uint64_t start, uint64_t end) "0x%"PRIx64"
>  vfio_prereg_listener_region_del_skip(uint64_t start, uint64_t end) "0x%"PRIx64" - 0x%"PRIx64
>  vfio_prereg_register(uint64_t va, uint64_t size, int ret) "va=0x%"PRIx64" size=0x%"PRIx64" ret=%d"
>  vfio_prereg_unregister(uint64_t va, uint64_t size, int ret) "va=0x%"PRIx64" size=0x%"PRIx64" ret=%d"
> -vfio_spapr_create_window(int ps, uint64_t ws, uint64_t off) "pageshift=0x%x winsize=0x%"PRIx64" offset=0x%"PRIx64
> +vfio_spapr_create_window(int ps, unsigned int levels, uint64_t ws, uint64_t off) "pageshift=0x%x levels=%u winsize=0x%"PRIx64" offset=0x%"PRIx64
>  vfio_spapr_remove_window(uint64_t off) "offset=0x%"PRIx64
>  vfio_spapr_group_attach(int groupfd, int tablefd) "Attached groupfd %d to liobn fd %d"

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

[Qemu-devel] [PATCH qemu 2/3] vfio: Make vfio_get_region_info_cap public

[Alexey Kardashevskiy]

This makes vfio_get_region_info_cap() to be used in quirks.

Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
---
 include/hw/vfio/vfio-common.h | 2 ++
 hw/vfio/common.c              | 2 +-
 2 files changed, 3 insertions(+), 1 deletion(-)

diff --git a/include/hw/vfio/vfio-common.h b/include/hw/vfio/vfio-common.h
index 1b434d0..abc82f9 100644
--- a/include/hw/vfio/vfio-common.h
+++ b/include/hw/vfio/vfio-common.h
@@ -189,6 +189,8 @@ int vfio_get_region_info(VFIODevice *vbasedev, int index,
 int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type,
                              uint32_t subtype, struct vfio_region_info **info);
 bool vfio_has_region_cap(VFIODevice *vbasedev, int region, uint16_t cap_type);
+struct vfio_info_cap_header *
+vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id);
 #endif
 extern const MemoryListener vfio_prereg_listener;
 
diff --git a/hw/vfio/common.c b/hw/vfio/common.c
index 7c185e5a..f58b487 100644
--- a/hw/vfio/common.c
+++ b/hw/vfio/common.c
@@ -711,7 +711,7 @@ static void vfio_listener_release(VFIOContainer *container)
     }
 }
 
-static struct vfio_info_cap_header *
+struct vfio_info_cap_header *
 vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id)
 {
     struct vfio_info_cap_header *hdr;
-- 
2.17.1

[Qemu-devel] [PATCH qemu 3/3] spapr: Support NVIDIA V100 GPU with NVLink2

[Alexey Kardashevskiy]

NVIDIA V100 GPUs have on-board RAM which is mapped into the host memory
space and accessible as normal RAM via an NVLink bus. The VFIO-PCI driver
implements special regions for such GPUs and emulates an NVLink bridge.
NVLink2-enabled POWER9 CPUs also provide address translation services
which includes an ATS shootdown (ATSD) register exported via the NVLink
bridge device.

This adds a quirk to VFIO to map the GPU memory and create an MR;
the new MR is stored in a PCI device as a QOM link. The sPAPR PCI uses
this to get the MR and map it to the system address space.
Another quirk does the same for ATSD.

This adds 4 additional steps to the FDT builder in spapr-pci:

1. Search for specific GPUs and NPUs and collect findings in
sPAPRPHBState::nvgpus;

2. Add properties in the device tree such as "ibm,npu", "ibm,gpu",
"memory-block" and others to advertise the NVLink2 function to the guest;

3. Add new memory blocks with an extra "linux,memory-usable" to prevent
the guest OS from accessing the new memory until it is online by the GPU
driver in the guest;

4. Add a npuphb# node representing an NPU unit for every vPHB as
the GPU driver uses it to detect NPU2 hardware and discover links; this
is not backed by any QEMU device as it does need to.

This allocates space for GPU RAM and ATSD like we do for MMIOs by
adding 2 new parameters to the phb_placement() hook. Older machine types
set these to zero.

This puts new memory nodes in a separate NUMA node to replicate the host
system setup as the GPU driver relies on this.

This adds requirement similar to EEH - one IOMMU group per vPHB.
The reason for this is that ATSD registers belong to a physical NPU
so they cannot invalidate translations on GPUs attached to another NPU.
It is guaranteed by the host platform as it does not mix NVLink bridges
or GPUs from different NPU in the same IOMMU group. If more than one
IOMMU group is detected on a vPHB, this disables ATSD support for that
vPHB and prints a warning.

Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
---

The example command line for redbud system:

pbuild/qemu-aiku1804le-ppc64/ppc64-softmmu/qemu-system-ppc64 \
-nodefaults \
-chardev stdio,id=STDIO0,signal=off,mux=on \
-device spapr-vty,id=svty0,reg=0x71000110,chardev=STDIO0 \
-mon id=MON0,chardev=STDIO0,mode=readline -nographic -vga none \
-enable-kvm -m 384G \
-chardev socket,id=SOCKET0,server,nowait,host=localhost,port=40000 \
-mon chardev=SOCKET0,mode=control \
-smp 80,sockets=1,threads=4 \
-netdev "tap,id=TAP0,helper=/home/aik/qemu-bridge-helper --br=br0" \
-device "virtio-net-pci,id=vnet0,mac=52:54:00:12:34:56,netdev=TAP0" \
img/vdisk0.img \
-device "vfio-pci,id=vfio0004_04_00_0,host=0004:04:00.0" \
-device "vfio-pci,id=vfio0006_00_00_0,host=0006:00:00.0" \
-device "vfio-pci,id=vfio0006_00_00_1,host=0006:00:00.1" \
-device "vfio-pci,id=vfio0006_00_00_2,host=0006:00:00.2" \
-device "vfio-pci,id=vfio0004_05_00_0,host=0004:05:00.0" \
-device "vfio-pci,id=vfio0006_00_01_0,host=0006:00:01.0" \
-device "vfio-pci,id=vfio0006_00_01_1,host=0006:00:01.1" \
-device "vfio-pci,id=vfio0006_00_01_2,host=0006:00:01.2" \
-device spapr-pci-host-bridge,id=phb1,index=1 \
-device "vfio-pci,id=vfio0035_03_00_0,host=0035:03:00.0" \
-device "vfio-pci,id=vfio0007_00_00_0,host=0007:00:00.0" \
-device "vfio-pci,id=vfio0007_00_00_1,host=0007:00:00.1" \
-device "vfio-pci,id=vfio0007_00_00_2,host=0007:00:00.2" \
-device "vfio-pci,id=vfio0035_04_00_0,host=0035:04:00.0" \
-device "vfio-pci,id=vfio0007_00_01_0,host=0007:00:01.0" \
-device "vfio-pci,id=vfio0007_00_01_1,host=0007:00:01.1" \
-device "vfio-pci,id=vfio0007_00_01_2,host=0007:00:01.2" -snapshot \
-machine pseries \
-L /home/aik/t/qemu-ppc64-bios/ -d guest_errors

Note that QEMU attaches PCI devices to the last added vPHB so first
8 devices - 4:04:00.0 till 6:00:01.2 - go to the default vPHB, and
35:03:00.0..7:00:01.2 to the vPHB with id=phb1.
---
 hw/vfio/pci.h               |   2 +
 include/hw/pci-host/spapr.h |   9 +
 include/hw/ppc/spapr.h      |   3 +-
 hw/ppc/spapr.c              |  25 ++-
 hw/ppc/spapr_pci.c          | 333 +++++++++++++++++++++++++++++++++++-
 hw/vfio/pci-quirks.c        | 120 +++++++++++++
 hw/vfio/pci.c               |  14 ++
 hw/vfio/trace-events        |   4 +
 8 files changed, 506 insertions(+), 4 deletions(-)

diff --git a/hw/vfio/pci.h b/hw/vfio/pci.h
index f4c5fb6..fd7703a 100644
--- a/hw/vfio/pci.h
+++ b/hw/vfio/pci.h
@@ -195,6 +195,8 @@ int vfio_populate_vga(VFIOPCIDevice *vdev, Error **errp);
 int vfio_pci_igd_opregion_init(VFIOPCIDevice *vdev,
                                struct vfio_region_info *info,
                                Error **errp);
+int vfio_pci_nvidia_v100_ram_init(VFIOPCIDevice *vdev, Error **errp);
+int vfio_pci_nvlink2_init(VFIOPCIDevice *vdev, Error **errp);
 
 void vfio_display_reset(VFIOPCIDevice *vdev);
 int vfio_display_probe(VFIOPCIDevice *vdev, Error **errp);
diff --git a/include/hw/pci-host/spapr.h b/include/hw/pci-host/spapr.h
index 4eb3a2c..bd5e43f 100644
--- a/include/hw/pci-host/spapr.h
+++ b/include/hw/pci-host/spapr.h
@@ -87,6 +87,9 @@ struct sPAPRPHBState {
     uint32_t mig_liobn;
     hwaddr mig_mem_win_addr, mig_mem_win_size;
     hwaddr mig_io_win_addr, mig_io_win_size;
+    hwaddr nv2_gpa_win_addr;
+    hwaddr nv2_atsd_win_addr;
+    struct spapr_phb_pci_nvgpu_config *nvgpus;
 };
 
 #define SPAPR_PCI_MEM_WIN_BUS_OFFSET 0x80000000ULL
@@ -106,6 +109,12 @@ struct sPAPRPHBState {
 
 #define SPAPR_PCI_MSI_WINDOW         0x40000000000ULL
 
+#define SPAPR_PCI_NV2RAM64_WIN_BASE  0x10000000000ULL /* 1 TiB */
+#define SPAPR_PCI_NV2RAM64_WIN_SIZE  0x10000000000ULL /* 1 TiB for all 6xGPUs */
+
+#define SPAPR_PCI_NV2ATSD_WIN_BASE   SPAPR_PCI_LIMIT
+#define SPAPR_PCI_NV2ATSD_WIN_SIZE   (16 * 0x10000)
+
 static inline qemu_irq spapr_phb_lsi_qirq(struct sPAPRPHBState *phb, int pin)
 {
     sPAPRMachineState *spapr = SPAPR_MACHINE(qdev_get_machine());
diff --git a/include/hw/ppc/spapr.h b/include/hw/ppc/spapr.h
index 9e01a5a..21a3ed3 100644
--- a/include/hw/ppc/spapr.h
+++ b/include/hw/ppc/spapr.h
@@ -111,7 +111,8 @@ struct sPAPRMachineClass {
     void (*phb_placement)(sPAPRMachineState *spapr, uint32_t index,
                           uint64_t *buid, hwaddr *pio, 
                           hwaddr *mmio32, hwaddr *mmio64,
-                          unsigned n_dma, uint32_t *liobns, Error **errp);
+                          unsigned n_dma, uint32_t *liobns, hwaddr *nv2gpa,
+                          hwaddr *nv2atsd, Error **errp);
     sPAPRResizeHPT resize_hpt_default;
     sPAPRCapabilities default_caps;
     sPAPRIrq *irq;
diff --git a/hw/ppc/spapr.c b/hw/ppc/spapr.c
index e86a4bf..1803bdb 100644
--- a/hw/ppc/spapr.c
+++ b/hw/ppc/spapr.c
@@ -3838,7 +3838,8 @@ static const CPUArchIdList *spapr_possible_cpu_arch_ids(MachineState *machine)
 static void spapr_phb_placement(sPAPRMachineState *spapr, uint32_t index,
                                 uint64_t *buid, hwaddr *pio,
                                 hwaddr *mmio32, hwaddr *mmio64,
-                                unsigned n_dma, uint32_t *liobns, Error **errp)
+                                unsigned n_dma, uint32_t *liobns,
+                                hwaddr *nv2gpa, hwaddr *nv2atsd, Error **errp)
 {
     /*
      * New-style PHB window placement.
@@ -3883,6 +3884,9 @@ static void spapr_phb_placement(sPAPRMachineState *spapr, uint32_t index,
     *pio = SPAPR_PCI_BASE + index * SPAPR_PCI_IO_WIN_SIZE;
     *mmio32 = SPAPR_PCI_BASE + (index + 1) * SPAPR_PCI_MEM32_WIN_SIZE;
     *mmio64 = SPAPR_PCI_BASE + (index + 1) * SPAPR_PCI_MEM64_WIN_SIZE;
+
+    *nv2gpa = SPAPR_PCI_NV2RAM64_WIN_BASE + index * SPAPR_PCI_NV2RAM64_WIN_SIZE;
+    *nv2atsd = SPAPR_PCI_NV2ATSD_WIN_BASE + index * SPAPR_PCI_NV2ATSD_WIN_SIZE;
 }
 
 static ICSState *spapr_ics_get(XICSFabric *dev, int irq)
@@ -4084,6 +4088,18 @@ DEFINE_SPAPR_MACHINE(4_0, "4.0", true);
 /*
  * pseries-3.1
  */
+static void phb_placement_3_1(sPAPRMachineState *spapr, uint32_t index,
+                              uint64_t *buid, hwaddr *pio,
+                              hwaddr *mmio32, hwaddr *mmio64,
+                              unsigned n_dma, uint32_t *liobns,
+                              hwaddr *nv2gpa, hwaddr *nv2atsd, Error **errp)
+{
+    spapr_phb_placement(spapr, index, buid, pio, mmio32, mmio64, n_dma, liobns,
+                        nv2gpa, nv2atsd, errp);
+    *nv2gpa = 0;
+    *nv2atsd = 0;
+}
+
 static void spapr_machine_3_1_class_options(MachineClass *mc)
 {
     sPAPRMachineClass *smc = SPAPR_MACHINE_CLASS(mc);
@@ -4092,6 +4108,7 @@ static void spapr_machine_3_1_class_options(MachineClass *mc)
     compat_props_add(mc->compat_props, hw_compat_3_1, hw_compat_3_1_len);
     mc->default_cpu_type = POWERPC_CPU_TYPE_NAME("power8_v2.0");
     smc->update_dt_enabled = false;
+    smc->phb_placement = phb_placement_3_1;
 }
 
 DEFINE_SPAPR_MACHINE(3_1, "3.1", false);
@@ -4239,7 +4256,8 @@ DEFINE_SPAPR_MACHINE(2_8, "2.8", false);
 static void phb_placement_2_7(sPAPRMachineState *spapr, uint32_t index,
                               uint64_t *buid, hwaddr *pio,
                               hwaddr *mmio32, hwaddr *mmio64,
-                              unsigned n_dma, uint32_t *liobns, Error **errp)
+                              unsigned n_dma, uint32_t *liobns,
+                              hwaddr *nv2gpa, hwaddr *nv2atsd, Error **errp)
 {
     /* Legacy PHB placement for pseries-2.7 and earlier machine types */
     const uint64_t base_buid = 0x800000020000000ULL;
@@ -4283,6 +4301,9 @@ static void phb_placement_2_7(sPAPRMachineState *spapr, uint32_t index,
      * fallback behaviour of automatically splitting a large "32-bit"
      * window into contiguous 32-bit and 64-bit windows
      */
+
+    *nv2gpa = 0;
+    *nv2atsd = 0;
 }
 
 static void spapr_machine_2_7_class_options(MachineClass *mc)
diff --git a/hw/ppc/spapr_pci.c b/hw/ppc/spapr_pci.c
index b74f263..b5b9259 100644
--- a/hw/ppc/spapr_pci.c
+++ b/hw/ppc/spapr_pci.c
@@ -50,6 +50,46 @@
 #include "sysemu/hostmem.h"
 #include "sysemu/numa.h"
 
+#define PHANDLE_PCIDEV(phb, pdev)    (0x12000000 | \
+                                     (((phb)->index) << 16) | ((pdev)->devfn))
+#define PHANDLE_GPURAM(phb, n)       (0x110000FF | ((n) << 8) | \
+                                     (((phb)->index) << 16))
+/* NVLink2 wants a separate NUMA node for its RAM */
+#define GPURAM_ASSOCIATIVITY(phb, n) (255 - ((phb)->index * 3 + (n)))
+#define PHANDLE_NVLINK(phb, gn, nn)  (0x00130000 | (((phb)->index) << 8) | \
+                                     ((gn) << 4) | (nn))
+
+/* Max number of these GPUs per a physical box */
+#define NVGPU_MAX_NUM                6
+
+/* Max number of NVLinks per GPU in any physical box */
+#define NVGPU_MAX_LINKS              3
+
+/*
+ * One NVLink bridge provides one ATSD register so it should be 18.
+ * In practice though since we allow only one group per vPHB which equals
+ * to an NPU2 which has maximum 6 NVLink bridges.
+ */
+#define NVGPU_MAX_ATSD               6
+
+struct spapr_phb_pci_nvgpu_config {
+    uint64_t nv2_ram_current;
+    uint64_t nv2_atsd_current;
+    int num; /* number of valid entries in gpus[] */
+    int gpunum; /* number of entries in gpus[] with gpdev!=NULL */
+    struct {
+        int links;
+        uint64_t tgt;
+        uint64_t gpa;
+        PCIDevice *gpdev;
+        uint64_t atsd_gpa[NVGPU_MAX_LINKS];
+        PCIDevice *npdev[NVGPU_MAX_LINKS];
+        uint32_t link_speed[NVGPU_MAX_LINKS];
+    } gpus[NVGPU_MAX_NUM];
+    uint64_t atsd_prop[NVGPU_MAX_ATSD]; /* Big Endian for DT */
+    int atsd_num;
+};
+
 /* Copied from the kernel arch/powerpc/platforms/pseries/msi.c */
 #define RTAS_QUERY_FN           0
 #define RTAS_CHANGE_FN          1
@@ -1249,6 +1289,7 @@ static uint32_t spapr_phb_get_pci_drc_index(sPAPRPHBState *phb,
 static void spapr_populate_pci_child_dt(PCIDevice *dev, void *fdt, int offset,
                                        sPAPRPHBState *sphb)
 {
+    int i, j;
     ResourceProps rp;
     bool is_bridge = false;
     int pci_status;
@@ -1349,6 +1390,60 @@ static void spapr_populate_pci_child_dt(PCIDevice *dev, void *fdt, int offset,
     if (sphb->pcie_ecs && pci_is_express(dev)) {
         _FDT(fdt_setprop_cell(fdt, offset, "ibm,pci-config-space-type", 0x1));
     }
+
+    if (sphb->nvgpus) {
+        for (i = 0; i < sphb->nvgpus->num; ++i) {
+            PCIDevice *gpdev = sphb->nvgpus->gpus[i].gpdev;
+
+            if (!gpdev) {
+                continue;
+            }
+            if (dev == gpdev) {
+                uint32_t npus[sphb->nvgpus->gpus[i].links];
+
+                for (j = 0; j < sphb->nvgpus->gpus[i].links; ++j) {
+                    PCIDevice *npdev = sphb->nvgpus->gpus[i].npdev[j];
+
+                    npus[j] = cpu_to_be32(PHANDLE_PCIDEV(sphb, npdev));
+                }
+                _FDT(fdt_setprop(fdt, offset, "ibm,npu", npus,
+                                 j * sizeof(npus[0])));
+                _FDT((fdt_setprop_cell(fdt, offset, "phandle",
+                                       PHANDLE_PCIDEV(sphb, dev))));
+                continue;
+            }
+            for (j = 0; j < sphb->nvgpus->gpus[i].links; ++j) {
+                if (dev != sphb->nvgpus->gpus[i].npdev[j]) {
+                    continue;
+                }
+
+                _FDT((fdt_setprop_cell(fdt, offset, "phandle",
+                                       PHANDLE_PCIDEV(sphb, dev))));
+
+                _FDT(fdt_setprop_cell(fdt, offset, "ibm,gpu",
+                                      PHANDLE_PCIDEV(sphb, gpdev)));
+
+                _FDT((fdt_setprop_cell(fdt, offset, "ibm,nvlink",
+                                       PHANDLE_NVLINK(sphb, i, j))));
+
+                /*
+                 * If we ever want to emulate GPU RAM at the same location as on
+                 * the host - here is the encoding GPA->TGT:
+                 *
+                 * gta  = ((sphb->nv2_gpa >> 42) & 0x1) << 42;
+                 * gta |= ((sphb->nv2_gpa >> 45) & 0x3) << 43;
+                 * gta |= ((sphb->nv2_gpa >> 49) & 0x3) << 45;
+                 * gta |= sphb->nv2_gpa & ((1UL << 43) - 1);
+                 */
+                _FDT(fdt_setprop_cell(fdt, offset, "memory-region",
+                                      PHANDLE_GPURAM(sphb, i)));
+                _FDT(fdt_setprop_u64(fdt, offset, "ibm,device-tgt-addr",
+                                     sphb->nvgpus->gpus[i].tgt));
+                _FDT(fdt_setprop_cell(fdt, offset, "ibm,nvlink-speed",
+                                      sphb->nvgpus->gpus[i].link_speed[j]));
+            }
+        }
+    }
 }
 
 /* create OF node for pci device and required OF DT properties */
@@ -1590,7 +1685,9 @@ static void spapr_phb_realize(DeviceState *dev, Error **errp)
         smc->phb_placement(spapr, sphb->index,
                            &sphb->buid, &sphb->io_win_addr,
                            &sphb->mem_win_addr, &sphb->mem64_win_addr,
-                           windows_supported, sphb->dma_liobn, &local_err);
+                           windows_supported, sphb->dma_liobn,
+                           &sphb->nv2_gpa_win_addr,
+                           &sphb->nv2_atsd_win_addr, &local_err);
         if (local_err) {
             error_propagate(errp, local_err);
             return;
@@ -1837,6 +1934,8 @@ static Property spapr_phb_properties[] = {
                      pre_2_8_migration, false),
     DEFINE_PROP_BOOL("pcie-extended-configuration-space", sPAPRPHBState,
                      pcie_ecs, true),
+    DEFINE_PROP_UINT64("gpa", sPAPRPHBState, nv2_gpa_win_addr, 0),
+    DEFINE_PROP_UINT64("atsd", sPAPRPHBState, nv2_atsd_win_addr, 0),
     DEFINE_PROP_END_OF_LIST(),
 };
 
@@ -2066,6 +2165,80 @@ static void spapr_phb_pci_enumerate(sPAPRPHBState *phb)
 
 }
 
+static void spapr_phb_pci_find_nvgpu(PCIBus *bus, PCIDevice *pdev, void *opaque)
+{
+    struct spapr_phb_pci_nvgpu_config *nvgpus = opaque;
+    PCIBus *sec_bus;
+    Object *mr_gpu, *mr_npu;
+    uint64_t tgt = 0, gpa, atsd = 0;
+    int i;
+
+    mr_gpu = object_property_get_link(OBJECT(pdev), "nvlink2-mr[0]", NULL);
+    mr_npu = object_property_get_link(OBJECT(pdev), "nvlink2-atsd-mr[0]", NULL);
+    if (mr_gpu) {
+        gpa = nvgpus->nv2_ram_current;
+        nvgpus->nv2_ram_current += memory_region_size(MEMORY_REGION(mr_gpu));
+    } else if (mr_npu) {
+        if (nvgpus->atsd_num == ARRAY_SIZE(nvgpus->atsd_prop)) {
+            warn_report("Found too many ATSD registers per vPHB");
+        } else {
+            atsd = nvgpus->nv2_atsd_current;
+            nvgpus->atsd_prop[nvgpus->atsd_num] = cpu_to_be64(atsd);
+            ++nvgpus->atsd_num;
+            nvgpus->nv2_atsd_current +=
+                memory_region_size(MEMORY_REGION(mr_npu));
+        }
+    }
+
+    tgt = object_property_get_uint(OBJECT(pdev), "tgt", NULL);
+    if (tgt) {
+        for (i = 0; i < nvgpus->num; ++i) {
+            if (nvgpus->gpus[i].tgt == tgt) {
+                break;
+            }
+        }
+
+        if (i == nvgpus->num) {
+            if (nvgpus->num == ARRAY_SIZE(nvgpus->gpus)) {
+                warn_report("Found too many NVLink bridges per GPU");
+                return;
+            }
+            ++nvgpus->num;
+        }
+
+        nvgpus->gpus[i].tgt = tgt;
+        if (mr_gpu) {
+            g_assert(!nvgpus->gpus[i].gpdev);
+            nvgpus->gpus[i].gpdev = pdev;
+            nvgpus->gpus[i].gpa = gpa;
+            ++nvgpus->gpunum;
+        } else {
+            int j = nvgpus->gpus[i].links;
+
+            ++nvgpus->gpus[i].links;
+
+            g_assert(!nvgpus->gpus[i].npdev[j]);
+            nvgpus->gpus[i].npdev[j] = pdev;
+            nvgpus->gpus[i].atsd_gpa[j] = atsd;
+            nvgpus->gpus[i].link_speed[j] =
+                    object_property_get_uint(OBJECT(pdev), "link_speed", NULL);
+        }
+    }
+
+    if ((pci_default_read_config(pdev, PCI_HEADER_TYPE, 1) !=
+         PCI_HEADER_TYPE_BRIDGE)) {
+        return;
+    }
+
+    sec_bus = pci_bridge_get_sec_bus(PCI_BRIDGE(pdev));
+    if (!sec_bus) {
+        return;
+    }
+
+    pci_for_each_device(sec_bus, pci_bus_num(sec_bus),
+                        spapr_phb_pci_find_nvgpu, opaque);
+}
+
 int spapr_populate_pci_dt(sPAPRPHBState *phb, uint32_t xics_phandle, void *fdt,
                           uint32_t nr_msis)
 {
@@ -2184,6 +2357,69 @@ int spapr_populate_pci_dt(sPAPRPHBState *phb, uint32_t xics_phandle, void *fdt,
     spapr_phb_pci_enumerate(phb);
     _FDT(fdt_setprop_cell(fdt, bus_off, "qemu,phb-enumerated", 0x1));
 
+    /* If there are existing NVLink2 MRs, unmap those before recreating */
+    if (phb->nvgpus) {
+        for (i = 0; i < phb->nvgpus->num; ++i) {
+            PCIDevice *gpdev = phb->nvgpus->gpus[i].gpdev;
+            Object *nv_mrobj = object_property_get_link(OBJECT(gpdev),
+                                                        "nvlink2-mr[0]", NULL);
+
+            if (nv_mrobj) {
+                memory_region_del_subregion(get_system_memory(),
+                                            MEMORY_REGION(nv_mrobj));
+            }
+            for (j = 0; j < phb->nvgpus->gpus[i].links; ++j) {
+                PCIDevice *npdev = phb->nvgpus->gpus[i].npdev[j];
+                Object *atsd_mrobj;
+                atsd_mrobj = object_property_get_link(OBJECT(npdev),
+                                                         "nvlink2-atsd-mr[0]",
+                                                         NULL);
+                if (atsd_mrobj) {
+                    memory_region_del_subregion(get_system_memory(),
+                                                MEMORY_REGION(atsd_mrobj));
+                }
+            }
+        }
+        g_free(phb->nvgpus);
+        phb->nvgpus = NULL;
+    }
+
+    /* Search for GPUs and NPUs */
+    if (phb->nv2_gpa_win_addr && phb->nv2_atsd_win_addr) {
+        phb->nvgpus = g_new0(struct spapr_phb_pci_nvgpu_config, 1);
+        phb->nvgpus->nv2_ram_current = phb->nv2_gpa_win_addr;
+        phb->nvgpus->nv2_atsd_current = phb->nv2_atsd_win_addr;
+
+        pci_for_each_device(bus, pci_bus_num(bus),
+                            spapr_phb_pci_find_nvgpu, phb->nvgpus);
+
+        if (!phb->nvgpus->gpunum) {
+            /* We did not find any interesting GPU */
+            g_free(phb->nvgpus);
+            phb->nvgpus = NULL;
+        } else {
+            /*
+             * ibm,mmio-atsd contains ATSD registers; these belong to an NPU PHB
+             * which we do not emulate as a separate device. Instead we put
+             * ibm,mmio-atsd to the vPHB with GPU and make sure that we do not
+             * put GPUs from different IOMMU groups to the same vPHB to ensure
+             * that the guest will use ATSDs from the corresponding NPU.
+             */
+            if (!spapr_phb_eeh_available(phb)) {
+                warn_report("ATSD requires a separate vPHB per GPU IOMMU group");
+            } else if (!phb->nvgpus->atsd_num) {
+                warn_report("No ATSD registers found");
+            } else if (phb->nvgpus->atsd_num > 8) {
+                warn_report("Bogus ATSD configuration is found");
+            } else {
+                _FDT((fdt_setprop(fdt, bus_off, "ibm,mmio-atsd",
+                                  phb->nvgpus->atsd_prop,
+                                  phb->nvgpus->atsd_num *
+                                  sizeof(phb->nvgpus->atsd_prop[0]))));
+            }
+        }
+    }
+
     /* Populate tree nodes with PCI devices attached */
     s_fdt.fdt = fdt;
     s_fdt.node_off = bus_off;
@@ -2198,6 +2434,101 @@ int spapr_populate_pci_dt(sPAPRPHBState *phb, uint32_t xics_phandle, void *fdt,
         return ret;
     }
 
+    if (phb->nvgpus) {
+        /* Add memory nodes for GPU RAM and mark them unusable */
+        for (i = 0; i < phb->nvgpus->num; ++i) {
+            PCIDevice *gpdev = phb->nvgpus->gpus[i].gpdev;
+            Object *nv_mrobj = object_property_get_link(OBJECT(gpdev),
+                                                        "nvlink2-mr[0]", NULL);
+            char *mem_name;
+            int off;
+            uint32_t at = cpu_to_be32(GPURAM_ASSOCIATIVITY(phb, i));
+            uint32_t associativity[] = { cpu_to_be32(0x4), at, at, at, at };
+            uint64_t nv2_size = object_property_get_uint(nv_mrobj,
+                                                         "size", NULL);
+            uint64_t mem_reg_property[2] = {
+                cpu_to_be64(phb->nvgpus->gpus[i].gpa), cpu_to_be64(nv2_size) };
+
+            mem_name = g_strdup_printf("memory@%lx", phb->nvgpus->gpus[i].gpa);
+            off = fdt_add_subnode(fdt, 0, mem_name);
+            _FDT(off);
+            _FDT((fdt_setprop_string(fdt, off, "device_type", "memory")));
+            _FDT((fdt_setprop(fdt, off, "reg", mem_reg_property,
+                              sizeof(mem_reg_property))));
+            _FDT((fdt_setprop(fdt, off, "ibm,associativity", associativity,
+                              sizeof(associativity))));
+
+            _FDT((fdt_setprop_string(fdt, off, "compatible",
+                                     "ibm,coherent-device-memory")));
+            mem_reg_property[1] = 0;
+            _FDT((fdt_setprop(fdt, off, "linux,usable-memory", mem_reg_property,
+                              sizeof(mem_reg_property))));
+            _FDT((fdt_setprop_cell(fdt, off, "phandle",
+                                   PHANDLE_GPURAM(phb, i))));
+
+            g_free(mem_name);
+        }
+
+        /* Add NPUs with links underneath */
+        if (phb->nvgpus->num) {
+            char *npuname = g_strdup_printf("npuphb%d", phb->index);
+            int npuoff = fdt_add_subnode(fdt, 0, npuname);
+            int linkidx = 0;
+
+            _FDT(npuoff);
+            _FDT(fdt_setprop_cell(fdt, npuoff, "#address-cells", 1));
+            _FDT(fdt_setprop_cell(fdt, npuoff, "#size-cells", 0));
+            /* Advertise NPU as POWER9 so the guest can enable NPU2 contexts */
+            _FDT((fdt_setprop_string(fdt, npuoff, "compatible",
+                                     "ibm,power9-npu")));
+            g_free(npuname);
+
+            for (i = 0; i < phb->nvgpus->num; ++i) {
+                for (j = 0; j < phb->nvgpus->gpus[i].links; ++j) {
+                    char *linkname = g_strdup_printf("link@%d", linkidx);
+                    int off = fdt_add_subnode(fdt, npuoff, linkname);
+
+                    _FDT(off);
+                    _FDT((fdt_setprop_cell(fdt, off, "reg", linkidx)));
+                    _FDT((fdt_setprop_string(fdt, off, "compatible",
+                                             "ibm,npu-link")));
+                    _FDT((fdt_setprop_cell(fdt, off, "phandle",
+                                           PHANDLE_NVLINK(phb, i, j))));
+                    _FDT((fdt_setprop_cell(fdt, off, "ibm,npu-link-index",
+                                           linkidx)));
+                    g_free(linkname);
+                    ++linkidx;
+                }
+            }
+        }
+
+        /* Finally, when we finished with the device tree, map subregions */
+        for (i = 0; i < phb->nvgpus->num; ++i) {
+            PCIDevice *gpdev = phb->nvgpus->gpus[i].gpdev;
+            Object *nv_mrobj = object_property_get_link(OBJECT(gpdev),
+                                                        "nvlink2-mr[0]", NULL);
+
+            if (nv_mrobj) {
+                memory_region_add_subregion(get_system_memory(),
+                                            phb->nvgpus->gpus[i].gpa,
+                                            MEMORY_REGION(nv_mrobj));
+            }
+            for (j = 0; j < phb->nvgpus->gpus[i].links; ++j) {
+                PCIDevice *npdev = phb->nvgpus->gpus[i].npdev[j];
+                Object *atsd_mrobj;
+                atsd_mrobj = object_property_get_link(OBJECT(npdev),
+                                                    "nvlink2-atsd-mr[0]",
+                                                    NULL);
+                if (atsd_mrobj) {
+                    hwaddr atsd_gpa = phb->nvgpus->gpus[i].atsd_gpa[j];
+
+                    memory_region_add_subregion(get_system_memory(), atsd_gpa,
+                                                MEMORY_REGION(atsd_mrobj));
+                }
+            }
+        }
+    } /* if (phb->nvgpus) */
+
     return 0;
 }
 
diff --git a/hw/vfio/pci-quirks.c b/hw/vfio/pci-quirks.c
index 2796837..c7c1250 100644
--- a/hw/vfio/pci-quirks.c
+++ b/hw/vfio/pci-quirks.c
@@ -2206,3 +2206,123 @@ int vfio_add_virt_caps(VFIOPCIDevice *vdev, Error **errp)
 
     return 0;
 }
+
+static void vfio_pci_nvlink2_get_tgt(Object *obj, Visitor *v,
+                                     const char *name,
+                                     void *opaque, Error **errp)
+{
+    uint64_t tgt = (uint64_t) opaque;
+    visit_type_uint64(v, name, &tgt, errp);
+}
+
+static void vfio_pci_nvlink2_get_link_speed(Object *obj, Visitor *v,
+                                                 const char *name,
+                                                 void *opaque, Error **errp)
+{
+    uint32_t link_speed = (uint32_t)(uint64_t) opaque;
+    visit_type_uint32(v, name, &link_speed, errp);
+}
+
+int vfio_pci_nvidia_v100_ram_init(VFIOPCIDevice *vdev, Error **errp)
+{
+    int ret;
+    void *p;
+    struct vfio_region_info *nv2region = NULL;
+    struct vfio_info_cap_header *hdr;
+    MemoryRegion *nv2mr = g_malloc0(sizeof(*nv2mr));
+
+    ret = vfio_get_dev_region_info(&vdev->vbasedev,
+                                   VFIO_REGION_TYPE_PCI_VENDOR_TYPE |
+                                   PCI_VENDOR_ID_NVIDIA,
+                                   VFIO_REGION_SUBTYPE_NVIDIA_NVLINK2_RAM,
+                                   &nv2region);
+    if (ret) {
+        return ret;
+    }
+
+    p = mmap(NULL, nv2region->size, PROT_READ | PROT_WRITE | PROT_EXEC,
+             MAP_SHARED, vdev->vbasedev.fd, nv2region->offset);
+
+    if (!p) {
+        return -errno;
+    }
+
+    memory_region_init_ram_ptr(nv2mr, OBJECT(vdev), "nvlink2-mr",
+                               nv2region->size, p);
+
+    hdr = vfio_get_region_info_cap(nv2region,
+                                   VFIO_REGION_INFO_CAP_NVLINK2_SSATGT);
+    if (hdr) {
+        struct vfio_region_info_cap_nvlink2_ssatgt *cap = (void *) hdr;
+
+        object_property_add(OBJECT(vdev), "tgt", "uint64",
+                            vfio_pci_nvlink2_get_tgt, NULL, NULL,
+                            (void *) cap->tgt, NULL);
+        trace_vfio_pci_nvidia_gpu_setup_quirk(vdev->vbasedev.name, cap->tgt,
+                                              nv2region->size);
+    }
+    g_free(nv2region);
+
+    return 0;
+}
+
+int vfio_pci_nvlink2_init(VFIOPCIDevice *vdev, Error **errp)
+{
+    int ret;
+    void *p;
+    struct vfio_region_info *atsd_region = NULL;
+    struct vfio_info_cap_header *hdr;
+
+    ret = vfio_get_dev_region_info(&vdev->vbasedev,
+                                   VFIO_REGION_TYPE_PCI_VENDOR_TYPE |
+                                   PCI_VENDOR_ID_IBM,
+                                   VFIO_REGION_SUBTYPE_IBM_NVLINK2_ATSD,
+                                   &atsd_region);
+    if (ret) {
+        return ret;
+    }
+
+    /* Some NVLink bridges come without assigned ATSD, skip MR part */
+    if (atsd_region->size) {
+        MemoryRegion *atsd_mr = g_malloc0(sizeof(*atsd_mr));
+
+        p = mmap(NULL, atsd_region->size, PROT_READ | PROT_WRITE | PROT_EXEC,
+                 MAP_SHARED, vdev->vbasedev.fd, atsd_region->offset);
+
+        if (!p) {
+            return -errno;
+        }
+
+        memory_region_init_ram_device_ptr(atsd_mr, OBJECT(vdev),
+                                          "nvlink2-atsd-mr",
+                                          atsd_region->size,
+                                          p);
+    }
+
+    hdr = vfio_get_region_info_cap(atsd_region,
+                                   VFIO_REGION_INFO_CAP_NVLINK2_SSATGT);
+    if (hdr) {
+        struct vfio_region_info_cap_nvlink2_ssatgt *cap = (void *) hdr;
+
+        object_property_add(OBJECT(vdev), "tgt", "uint64",
+                            vfio_pci_nvlink2_get_tgt, NULL, NULL,
+                            (void *) cap->tgt, NULL);
+        trace_vfio_pci_nvlink2_setup_quirk_ssatgt(vdev->vbasedev.name, cap->tgt,
+                                                  atsd_region->size);
+    }
+
+    hdr = vfio_get_region_info_cap(atsd_region,
+                                   VFIO_REGION_INFO_CAP_NVLINK2_LNKSPD);
+    if (hdr) {
+        struct vfio_region_info_cap_nvlink2_lnkspd *cap = (void *) hdr;
+
+        object_property_add(OBJECT(vdev), "link_speed", "uint32",
+                            vfio_pci_nvlink2_get_link_speed, NULL, NULL,
+                            (void *) (uint64_t) cap->link_speed, NULL);
+        trace_vfio_pci_nvlink2_setup_quirk_lnkspd(vdev->vbasedev.name,
+                                                  cap->link_speed);
+    }
+    g_free(atsd_region);
+
+    return 0;
+}
diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c
index 2265fe7..25d1cb5 100644
--- a/hw/vfio/pci.c
+++ b/hw/vfio/pci.c
@@ -3069,6 +3069,20 @@ static void vfio_realize(PCIDevice *pdev, Error **errp)
         goto out_teardown;
     }
 
+    if (vdev->vendor_id == PCI_VENDOR_ID_NVIDIA) {
+        ret = vfio_pci_nvidia_v100_ram_init(vdev, errp);
+        if (ret && ret != -ENODEV) {
+            error_report("Failed to setup NVIDIA V100 GPU RAM");
+        }
+    }
+
+    if (vdev->vendor_id == PCI_VENDOR_ID_IBM) {
+        ret = vfio_pci_nvlink2_init(vdev, errp);
+        if (ret && ret != -ENODEV) {
+            error_report("Failed to setup NVlink2 bridge");
+        }
+    }
+
     vfio_register_err_notifier(vdev);
     vfio_register_req_notifier(vdev);
     vfio_setup_resetfn_quirk(vdev);
diff --git a/hw/vfio/trace-events b/hw/vfio/trace-events
index adfa75e..31deb55 100644
--- a/hw/vfio/trace-events
+++ b/hw/vfio/trace-events
@@ -88,6 +88,10 @@ vfio_pci_igd_opregion_enabled(const char *name) "%s"
 vfio_pci_igd_host_bridge_enabled(const char *name) "%s"
 vfio_pci_igd_lpc_bridge_enabled(const char *name) "%s"
 
+vfio_pci_nvidia_gpu_setup_quirk(const char *name, uint64_t tgt, uint64_t size) "%s tgt=0x%"PRIx64" size=0x%"PRIx64
+vfio_pci_nvlink2_setup_quirk_ssatgt(const char *name, uint64_t tgt, uint64_t size) "%s tgt=0x%"PRIx64" size=0x%"PRIx64
+vfio_pci_nvlink2_setup_quirk_lnkspd(const char *name, uint32_t link_speed) "%s link_speed=0x%x"
+
 # hw/vfio/common.c
 vfio_region_write(const char *name, int index, uint64_t addr, uint64_t data, unsigned size) " (%s:region%d+0x%"PRIx64", 0x%"PRIx64 ", %d)"
 vfio_region_read(char *name, int index, uint64_t addr, unsigned size, uint64_t data) " (%s:region%d+0x%"PRIx64", %d) = 0x%"PRIx64
-- 
2.17.1

Re: [Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[Alexey Kardashevskiy]

On 17/01/2019 13:51, Alexey Kardashevskiy wrote:
> This is for passing through NVIDIA V100 GPUs on POWER9 systems.
> 
> This implements a subdriver for NVIDIA V100 GPU with coherent memory and
> NPU/ATS support available in the POWER9 CPU.
> 
> 1/3 is not strictly related but since new memory also needs to be mapped
> to the 64bit DMA window and it is located quite high in the address space,
> some adjustments are needed.
> 
> 
> This is based on dwg/ppc-for-4.0 sha1 a0a8bff and requires headers update
> from v5.0-rc1 staged by Paolo already.
> 
> Please comment. Thanks.


Ping?

> 
> 
> 
> Alexey Kardashevskiy (3):
>   vfio/spapr: Fix indirect levels calculation
>   vfio: Make vfio_get_region_info_cap public
>   spapr: Support NVIDIA V100 GPU with NVLink2
> 
>  hw/vfio/pci.h                 |   2 +
>  include/hw/pci-host/spapr.h   |   9 +
>  include/hw/ppc/spapr.h        |   3 +-
>  include/hw/vfio/vfio-common.h |   2 +
>  hw/ppc/spapr.c                |  25 ++-
>  hw/ppc/spapr_pci.c            | 333 +++++++++++++++++++++++++++++++++-
>  hw/vfio/common.c              |   2 +-
>  hw/vfio/pci-quirks.c          | 120 ++++++++++++
>  hw/vfio/pci.c                 |  14 ++
>  hw/vfio/spapr.c               |  38 +++-
>  hw/vfio/trace-events          |   6 +-
>  11 files changed, 539 insertions(+), 15 deletions(-)
> 

-- 
Alexey

Re: [Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[Daniel Henrique Barboza]

Based on this series, I've sent a Libvirt patch to allow a QEMU process
to inherit IPC_LOCK when using VFIO passthrough with the Tesla V100
GPU:

https://www.redhat.com/archives/libvir-list/2019-February/msg00219.html


In that thread, Alex raised concerns about allowing QEMU to freely lock
all the memory it wants. Is this an issue to be considered in the review
of this series here?

Reading the patches, specially patch 3/3, it seems to me that QEMU is
going to lock the KVM memory to populate the NUMA node with memory
of the GPU itself, so at first there is no risk of not taking over the 
host RAM.
Am I missing something?


Thanks,


DHB


On 1/17/19 12:51 AM, Alexey Kardashevskiy wrote:
> This is for passing through NVIDIA V100 GPUs on POWER9 systems.
>
> This implements a subdriver for NVIDIA V100 GPU with coherent memory and
> NPU/ATS support available in the POWER9 CPU.
>
> 1/3 is not strictly related but since new memory also needs to be mapped
> to the 64bit DMA window and it is located quite high in the address space,
> some adjustments are needed.
>
>
> This is based on dwg/ppc-for-4.0 sha1 a0a8bff and requires headers update
> from v5.0-rc1 staged by Paolo already.
>
> Please comment. Thanks.
>
>
>
> Alexey Kardashevskiy (3):
>    vfio/spapr: Fix indirect levels calculation
>    vfio: Make vfio_get_region_info_cap public
>    spapr: Support NVIDIA V100 GPU with NVLink2
>
>   hw/vfio/pci.h                 |   2 +
>   include/hw/pci-host/spapr.h   |   9 +
>   include/hw/ppc/spapr.h        |   3 +-
>   include/hw/vfio/vfio-common.h |   2 +
>   hw/ppc/spapr.c                |  25 ++-
>   hw/ppc/spapr_pci.c            | 333 +++++++++++++++++++++++++++++++++-
>   hw/vfio/common.c              |   2 +-
>   hw/vfio/pci-quirks.c          | 120 ++++++++++++
>   hw/vfio/pci.c                 |  14 ++
>   hw/vfio/spapr.c               |  38 +++-
>   hw/vfio/trace-events          |   6 +-
>   11 files changed, 539 insertions(+), 15 deletions(-)
>

Re: [Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[Alexey Kardashevskiy]

On 07/02/2019 04:22, Daniel Henrique Barboza wrote:
> Based on this series, I've sent a Libvirt patch to allow a QEMU process
> to inherit IPC_LOCK when using VFIO passthrough with the Tesla V100
> GPU:
> 
> https://www.redhat.com/archives/libvir-list/2019-February/msg00219.html
> 
> 
> In that thread, Alex raised concerns about allowing QEMU to freely lock
> all the memory it wants. Is this an issue to be considered in the review
> of this series here?
> 
> Reading the patches, specially patch 3/3, it seems to me that QEMU is
> going to lock the KVM memory to populate the NUMA node with memory
> of the GPU itself, so at first there is no risk of not taking over the
> host RAM.
> Am I missing something?


The GPU memory belongs to the device and not visible to the host as
memory blocks and not covered by page structs, for the host it is more
like MMIO which is passed through to the guest without that locked
accounting, I'd expect libvirt to keep working as usual except that:

when libvirt calculates the amount of memory needed for TCE tables
(which is guestRAM/64k*8), now it needs to use the end of the last GPU
RAM window as a guest RAM size. For example, in QEMU HMP "info mtree -f":

FlatView #2
 AS "memory", root: system
 AS "cpu-memory-0", root: system
 Root memory region: system
  0000000000000000-000000007fffffff (prio 0, ram): ppc_spapr.ram
  0000010000000000-0000011fffffffff (prio 0, ram): nvlink2-mr

So previously the DMA window would cover 0x7fffffff+1, now it has to
cover 0x11fffffffff+1.


> 
> 
> Thanks,
> 
> 
> DHB
> 
> 
> On 1/17/19 12:51 AM, Alexey Kardashevskiy wrote:
>> This is for passing through NVIDIA V100 GPUs on POWER9 systems.
>>
>> This implements a subdriver for NVIDIA V100 GPU with coherent memory and
>> NPU/ATS support available in the POWER9 CPU.
>>
>> 1/3 is not strictly related but since new memory also needs to be mapped
>> to the 64bit DMA window and it is located quite high in the address space,
>> some adjustments are needed.
>>
>>
>> This is based on dwg/ppc-for-4.0 sha1 a0a8bff and requires headers update
>> from v5.0-rc1 staged by Paolo already.
>>
>> Please comment. Thanks.
>>
>>
>>
>> Alexey Kardashevskiy (3):
>>   vfio/spapr: Fix indirect levels calculation
>>   vfio: Make vfio_get_region_info_cap public
>>   spapr: Support NVIDIA V100 GPU with NVLink2
>>
>>  hw/vfio/pci.h                 |   2 +
>>  include/hw/pci-host/spapr.h   |   9 +
>>  include/hw/ppc/spapr.h        |   3 +-
>>  include/hw/vfio/vfio-common.h |   2 +
>>  hw/ppc/spapr.c                |  25 ++-
>>  hw/ppc/spapr_pci.c            | 333 +++++++++++++++++++++++++++++++++-
>>  hw/vfio/common.c              |   2 +-
>>  hw/vfio/pci-quirks.c          | 120 ++++++++++++
>>  hw/vfio/pci.c                 |  14 ++
>>  hw/vfio/spapr.c               |  38 +++-
>>  hw/vfio/trace-events          |   6 +-
>>  11 files changed, 539 insertions(+), 15 deletions(-)
>>
> 

-- 
Alexey

Re: [Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[Alex Williamson]

On Thu, 7 Feb 2019 15:43:18 +1100
Alexey Kardashevskiy <aik@ozlabs.ru> wrote:

> On 07/02/2019 04:22, Daniel Henrique Barboza wrote:
> > Based on this series, I've sent a Libvirt patch to allow a QEMU process
> > to inherit IPC_LOCK when using VFIO passthrough with the Tesla V100
> > GPU:
> > 
> > https://www.redhat.com/archives/libvir-list/2019-February/msg00219.html
> > 
> > 
> > In that thread, Alex raised concerns about allowing QEMU to freely lock
> > all the memory it wants. Is this an issue to be considered in the review
> > of this series here?
> > 
> > Reading the patches, specially patch 3/3, it seems to me that QEMU is
> > going to lock the KVM memory to populate the NUMA node with memory
> > of the GPU itself, so at first there is no risk of not taking over the
> > host RAM.
> > Am I missing something?  
> 
> 
> The GPU memory belongs to the device and not visible to the host as
> memory blocks and not covered by page structs, for the host it is more
> like MMIO which is passed through to the guest without that locked
> accounting, I'd expect libvirt to keep working as usual except that:
> 
> when libvirt calculates the amount of memory needed for TCE tables
> (which is guestRAM/64k*8), now it needs to use the end of the last GPU
> RAM window as a guest RAM size. For example, in QEMU HMP "info mtree -f":
> 
> FlatView #2
>  AS "memory", root: system
>  AS "cpu-memory-0", root: system
>  Root memory region: system
>   0000000000000000-000000007fffffff (prio 0, ram): ppc_spapr.ram
>   0000010000000000-0000011fffffffff (prio 0, ram): nvlink2-mr
> 
> So previously the DMA window would cover 0x7fffffff+1, now it has to
> cover 0x11fffffffff+1.

This looks like a chicken and egg problem, you're saying libvirt needs
to query mtree to understand the extent of the GPU layout, but we need
to specify the locked memory limits in order for QEMU to start?  Is
libvirt supposed to start the VM with unlimited locked memory and fix
it at some indeterminate point in the future?  Run a dummy VM with
unlimited locked memory in order to determine the limits for the real
VM?  Neither of these sound practical.  Thanks,

Alex

Re: [Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[Alexey Kardashevskiy]

On 08/02/2019 02:18, Alex Williamson wrote:
> On Thu, 7 Feb 2019 15:43:18 +1100
> Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> 
>> On 07/02/2019 04:22, Daniel Henrique Barboza wrote:
>>> Based on this series, I've sent a Libvirt patch to allow a QEMU process
>>> to inherit IPC_LOCK when using VFIO passthrough with the Tesla V100
>>> GPU:
>>>
>>> https://www.redhat.com/archives/libvir-list/2019-February/msg00219.html
>>>
>>>
>>> In that thread, Alex raised concerns about allowing QEMU to freely lock
>>> all the memory it wants. Is this an issue to be considered in the review
>>> of this series here?
>>>
>>> Reading the patches, specially patch 3/3, it seems to me that QEMU is
>>> going to lock the KVM memory to populate the NUMA node with memory
>>> of the GPU itself, so at first there is no risk of not taking over the
>>> host RAM.
>>> Am I missing something?  
>>
>>
>> The GPU memory belongs to the device and not visible to the host as
>> memory blocks and not covered by page structs, for the host it is more
>> like MMIO which is passed through to the guest without that locked
>> accounting, I'd expect libvirt to keep working as usual except that:
>>
>> when libvirt calculates the amount of memory needed for TCE tables
>> (which is guestRAM/64k*8), now it needs to use the end of the last GPU
>> RAM window as a guest RAM size. For example, in QEMU HMP "info mtree -f":
>>
>> FlatView #2
>>  AS "memory", root: system
>>  AS "cpu-memory-0", root: system
>>  Root memory region: system
>>   0000000000000000-000000007fffffff (prio 0, ram): ppc_spapr.ram
>>   0000010000000000-0000011fffffffff (prio 0, ram): nvlink2-mr
>>
>> So previously the DMA window would cover 0x7fffffff+1, now it has to
>> cover 0x11fffffffff+1.
> 
> This looks like a chicken and egg problem, you're saying libvirt needs
> to query mtree to understand the extent of the GPU layout, but we need
> to specify the locked memory limits in order for QEMU to start?  Is
> libvirt supposed to start the VM with unlimited locked memory and fix
> it at some indeterminate point in the future?  Run a dummy VM with
> unlimited locked memory in order to determine the limits for the real
> VM?  Neither of these sound practical.  Thanks,


QEMU maps GPU RAM at known locations (which only depends on the vPHB's
index or can be set explicitely) and libvirt knows how many GPUs are
passed so it is quite easy to calculate the required amount of memory.

Here is the window start calculation:
https://github.com/aik/qemu/commit/7073cad3ae7708d657e01672bcf53092808b54fb#diff-662409c2a5a150fe231d07ea8384b920R3812

We do not exactly know the GPU RAM window size until QEMU reads it from
VFIO/nvlink2 but we know that all existing hardware has a window of
128GB (the adapters I have access to only have 16/32GB on board).


-- 
Alexey

Re: [Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[Alex Williamson]

On Fri, 8 Feb 2019 13:29:37 +1100
Alexey Kardashevskiy <aik@ozlabs.ru> wrote:

> On 08/02/2019 02:18, Alex Williamson wrote:
> > On Thu, 7 Feb 2019 15:43:18 +1100
> > Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> >   
> >> On 07/02/2019 04:22, Daniel Henrique Barboza wrote:  
> >>> Based on this series, I've sent a Libvirt patch to allow a QEMU process
> >>> to inherit IPC_LOCK when using VFIO passthrough with the Tesla V100
> >>> GPU:
> >>>
> >>> https://www.redhat.com/archives/libvir-list/2019-February/msg00219.html
> >>>
> >>>
> >>> In that thread, Alex raised concerns about allowing QEMU to freely lock
> >>> all the memory it wants. Is this an issue to be considered in the review
> >>> of this series here?
> >>>
> >>> Reading the patches, specially patch 3/3, it seems to me that QEMU is
> >>> going to lock the KVM memory to populate the NUMA node with memory
> >>> of the GPU itself, so at first there is no risk of not taking over the
> >>> host RAM.
> >>> Am I missing something?    
> >>
> >>
> >> The GPU memory belongs to the device and not visible to the host as
> >> memory blocks and not covered by page structs, for the host it is more
> >> like MMIO which is passed through to the guest without that locked
> >> accounting, I'd expect libvirt to keep working as usual except that:
> >>
> >> when libvirt calculates the amount of memory needed for TCE tables
> >> (which is guestRAM/64k*8), now it needs to use the end of the last GPU
> >> RAM window as a guest RAM size. For example, in QEMU HMP "info mtree -f":
> >>
> >> FlatView #2
> >>  AS "memory", root: system
> >>  AS "cpu-memory-0", root: system
> >>  Root memory region: system
> >>   0000000000000000-000000007fffffff (prio 0, ram): ppc_spapr.ram
> >>   0000010000000000-0000011fffffffff (prio 0, ram): nvlink2-mr
> >>
> >> So previously the DMA window would cover 0x7fffffff+1, now it has to
> >> cover 0x11fffffffff+1.  
> > 
> > This looks like a chicken and egg problem, you're saying libvirt needs
> > to query mtree to understand the extent of the GPU layout, but we need
> > to specify the locked memory limits in order for QEMU to start?  Is
> > libvirt supposed to start the VM with unlimited locked memory and fix
> > it at some indeterminate point in the future?  Run a dummy VM with
> > unlimited locked memory in order to determine the limits for the real
> > VM?  Neither of these sound practical.  Thanks,  
> 
> 
> QEMU maps GPU RAM at known locations (which only depends on the vPHB's
> index or can be set explicitely) and libvirt knows how many GPUs are
> passed so it is quite easy to calculate the required amount of memory.
> 
> Here is the window start calculation:
> https://github.com/aik/qemu/commit/7073cad3ae7708d657e01672bcf53092808b54fb#diff-662409c2a5a150fe231d07ea8384b920R3812
> 
> We do not exactly know the GPU RAM window size until QEMU reads it from
> VFIO/nvlink2 but we know that all existing hardware has a window of
> 128GB (the adapters I have access to only have 16/32GB on board).

So you're asking that libvirt add 128GB per GPU with magic nvlink
properties, which may be 8x what's actually necessary and libvirt
determines which GPUs to apply this to how?  Does libvirt need to sort
through device tree properties for this?  Thanks,

Alex

Re: [Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[David Gibson]

[-- Attachment #1: Type: text/plain, Size: 3915 bytes --]

On Thu, Feb 07, 2019 at 08:26:20PM -0700, Alex Williamson wrote:
> On Fri, 8 Feb 2019 13:29:37 +1100
> Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> 
> > On 08/02/2019 02:18, Alex Williamson wrote:
> > > On Thu, 7 Feb 2019 15:43:18 +1100
> > > Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> > >   
> > >> On 07/02/2019 04:22, Daniel Henrique Barboza wrote:  
> > >>> Based on this series, I've sent a Libvirt patch to allow a QEMU process
> > >>> to inherit IPC_LOCK when using VFIO passthrough with the Tesla V100
> > >>> GPU:
> > >>>
> > >>> https://www.redhat.com/archives/libvir-list/2019-February/msg00219.html
> > >>>
> > >>>
> > >>> In that thread, Alex raised concerns about allowing QEMU to freely lock
> > >>> all the memory it wants. Is this an issue to be considered in the review
> > >>> of this series here?
> > >>>
> > >>> Reading the patches, specially patch 3/3, it seems to me that QEMU is
> > >>> going to lock the KVM memory to populate the NUMA node with memory
> > >>> of the GPU itself, so at first there is no risk of not taking over the
> > >>> host RAM.
> > >>> Am I missing something?    
> > >>
> > >>
> > >> The GPU memory belongs to the device and not visible to the host as
> > >> memory blocks and not covered by page structs, for the host it is more
> > >> like MMIO which is passed through to the guest without that locked
> > >> accounting, I'd expect libvirt to keep working as usual except that:
> > >>
> > >> when libvirt calculates the amount of memory needed for TCE tables
> > >> (which is guestRAM/64k*8), now it needs to use the end of the last GPU
> > >> RAM window as a guest RAM size. For example, in QEMU HMP "info mtree -f":
> > >>
> > >> FlatView #2
> > >>  AS "memory", root: system
> > >>  AS "cpu-memory-0", root: system
> > >>  Root memory region: system
> > >>   0000000000000000-000000007fffffff (prio 0, ram): ppc_spapr.ram
> > >>   0000010000000000-0000011fffffffff (prio 0, ram): nvlink2-mr
> > >>
> > >> So previously the DMA window would cover 0x7fffffff+1, now it has to
> > >> cover 0x11fffffffff+1.  
> > > 
> > > This looks like a chicken and egg problem, you're saying libvirt needs
> > > to query mtree to understand the extent of the GPU layout, but we need
> > > to specify the locked memory limits in order for QEMU to start?  Is
> > > libvirt supposed to start the VM with unlimited locked memory and fix
> > > it at some indeterminate point in the future?  Run a dummy VM with
> > > unlimited locked memory in order to determine the limits for the real
> > > VM?  Neither of these sound practical.  Thanks,  
> > 
> > 
> > QEMU maps GPU RAM at known locations (which only depends on the vPHB's
> > index or can be set explicitely) and libvirt knows how many GPUs are
> > passed so it is quite easy to calculate the required amount of memory.
> > 
> > Here is the window start calculation:
> > https://github.com/aik/qemu/commit/7073cad3ae7708d657e01672bcf53092808b54fb#diff-662409c2a5a150fe231d07ea8384b920R3812
> > 
> > We do not exactly know the GPU RAM window size until QEMU reads it from
> > VFIO/nvlink2 but we know that all existing hardware has a window of
> > 128GB (the adapters I have access to only have 16/32GB on board).
> 
> So you're asking that libvirt add 128GB per GPU with magic nvlink
> properties, which may be 8x what's actually necessary and libvirt
> determines which GPUs to apply this to how?  Does libvirt need to sort
> through device tree properties for this?  Thanks,

Hm.  If the GPU memory is really separate from main RAM, which it
sounds like, I don't think it makes sense to account it against the
same locked memory limit as regular RAM.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

Re: [Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[Alex Williamson]

On Fri, 8 Feb 2019 16:28:50 +1100
David Gibson <david@gibson.dropbear.id.au> wrote:

> On Thu, Feb 07, 2019 at 08:26:20PM -0700, Alex Williamson wrote:
> > On Fri, 8 Feb 2019 13:29:37 +1100
> > Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> >   
> > > On 08/02/2019 02:18, Alex Williamson wrote:  
> > > > On Thu, 7 Feb 2019 15:43:18 +1100
> > > > Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> > > >     
> > > >> On 07/02/2019 04:22, Daniel Henrique Barboza wrote:    
> > > >>> Based on this series, I've sent a Libvirt patch to allow a QEMU process
> > > >>> to inherit IPC_LOCK when using VFIO passthrough with the Tesla V100
> > > >>> GPU:
> > > >>>
> > > >>> https://www.redhat.com/archives/libvir-list/2019-February/msg00219.html
> > > >>>
> > > >>>
> > > >>> In that thread, Alex raised concerns about allowing QEMU to freely lock
> > > >>> all the memory it wants. Is this an issue to be considered in the review
> > > >>> of this series here?
> > > >>>
> > > >>> Reading the patches, specially patch 3/3, it seems to me that QEMU is
> > > >>> going to lock the KVM memory to populate the NUMA node with memory
> > > >>> of the GPU itself, so at first there is no risk of not taking over the
> > > >>> host RAM.
> > > >>> Am I missing something?      
> > > >>
> > > >>
> > > >> The GPU memory belongs to the device and not visible to the host as
> > > >> memory blocks and not covered by page structs, for the host it is more
> > > >> like MMIO which is passed through to the guest without that locked
> > > >> accounting, I'd expect libvirt to keep working as usual except that:
> > > >>
> > > >> when libvirt calculates the amount of memory needed for TCE tables
> > > >> (which is guestRAM/64k*8), now it needs to use the end of the last GPU
> > > >> RAM window as a guest RAM size. For example, in QEMU HMP "info mtree -f":
> > > >>
> > > >> FlatView #2
> > > >>  AS "memory", root: system
> > > >>  AS "cpu-memory-0", root: system
> > > >>  Root memory region: system
> > > >>   0000000000000000-000000007fffffff (prio 0, ram): ppc_spapr.ram
> > > >>   0000010000000000-0000011fffffffff (prio 0, ram): nvlink2-mr
> > > >>
> > > >> So previously the DMA window would cover 0x7fffffff+1, now it has to
> > > >> cover 0x11fffffffff+1.    
> > > > 
> > > > This looks like a chicken and egg problem, you're saying libvirt needs
> > > > to query mtree to understand the extent of the GPU layout, but we need
> > > > to specify the locked memory limits in order for QEMU to start?  Is
> > > > libvirt supposed to start the VM with unlimited locked memory and fix
> > > > it at some indeterminate point in the future?  Run a dummy VM with
> > > > unlimited locked memory in order to determine the limits for the real
> > > > VM?  Neither of these sound practical.  Thanks,    
> > > 
> > > 
> > > QEMU maps GPU RAM at known locations (which only depends on the vPHB's
> > > index or can be set explicitely) and libvirt knows how many GPUs are
> > > passed so it is quite easy to calculate the required amount of memory.
> > > 
> > > Here is the window start calculation:
> > > https://github.com/aik/qemu/commit/7073cad3ae7708d657e01672bcf53092808b54fb#diff-662409c2a5a150fe231d07ea8384b920R3812
> > > 
> > > We do not exactly know the GPU RAM window size until QEMU reads it from
> > > VFIO/nvlink2 but we know that all existing hardware has a window of
> > > 128GB (the adapters I have access to only have 16/32GB on board).  
> > 
> > So you're asking that libvirt add 128GB per GPU with magic nvlink
> > properties, which may be 8x what's actually necessary and libvirt
> > determines which GPUs to apply this to how?  Does libvirt need to sort
> > through device tree properties for this?  Thanks,  
> 
> Hm.  If the GPU memory is really separate from main RAM, which it
> sounds like, I don't think it makes sense to account it against the
> same locked memory limit as regular RAM.

That's true, if the user owns the device and the device provides the
memory backing, it doesn't seem like it should count against the user's
locked memory limit.  That'd make things easy for libvirt.  Thanks,

Alex

Re: [Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[Daniel Henrique Barboza]

On 2/8/19 1:52 PM, Alex Williamson wrote:
>> Hm.  If the GPU memory is really separate from main RAM, which it
>> sounds like, I don't think it makes sense to account it against the
>> same locked memory limit as regular RAM.
> That's true, if the user owns the device and the device provides the
> memory backing, it doesn't seem like it should count against the user's
> locked memory limit.  That'd make things easy for libvirt.  Thanks,

Sounds good.

Is this a QEMU (or even kernel) side change? If so, and we choose to go
through with it, I'll wait for a new spin of this series to see what must
be done in Libvirt to proper support it.


Thanks,


DHB

Re: [Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[Alexey Kardashevskiy]

On 08/02/2019 16:28, David Gibson wrote:
> On Thu, Feb 07, 2019 at 08:26:20PM -0700, Alex Williamson wrote:
>> On Fri, 8 Feb 2019 13:29:37 +1100
>> Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
>>
>>> On 08/02/2019 02:18, Alex Williamson wrote:
>>>> On Thu, 7 Feb 2019 15:43:18 +1100
>>>> Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
>>>>   
>>>>> On 07/02/2019 04:22, Daniel Henrique Barboza wrote:  
>>>>>> Based on this series, I've sent a Libvirt patch to allow a QEMU process
>>>>>> to inherit IPC_LOCK when using VFIO passthrough with the Tesla V100
>>>>>> GPU:
>>>>>>
>>>>>> https://www.redhat.com/archives/libvir-list/2019-February/msg00219.html
>>>>>>
>>>>>>
>>>>>> In that thread, Alex raised concerns about allowing QEMU to freely lock
>>>>>> all the memory it wants. Is this an issue to be considered in the review
>>>>>> of this series here?
>>>>>>
>>>>>> Reading the patches, specially patch 3/3, it seems to me that QEMU is
>>>>>> going to lock the KVM memory to populate the NUMA node with memory
>>>>>> of the GPU itself, so at first there is no risk of not taking over the
>>>>>> host RAM.
>>>>>> Am I missing something?    
>>>>>
>>>>>
>>>>> The GPU memory belongs to the device and not visible to the host as
>>>>> memory blocks and not covered by page structs, for the host it is more
>>>>> like MMIO which is passed through to the guest without that locked
>>>>> accounting, I'd expect libvirt to keep working as usual except that:
>>>>>
>>>>> when libvirt calculates the amount of memory needed for TCE tables
>>>>> (which is guestRAM/64k*8), now it needs to use the end of the last GPU
>>>>> RAM window as a guest RAM size. For example, in QEMU HMP "info mtree -f":
>>>>>
>>>>> FlatView #2
>>>>>  AS "memory", root: system
>>>>>  AS "cpu-memory-0", root: system
>>>>>  Root memory region: system
>>>>>   0000000000000000-000000007fffffff (prio 0, ram): ppc_spapr.ram
>>>>>   0000010000000000-0000011fffffffff (prio 0, ram): nvlink2-mr
>>>>>
>>>>> So previously the DMA window would cover 0x7fffffff+1, now it has to
>>>>> cover 0x11fffffffff+1.  
>>>>
>>>> This looks like a chicken and egg problem, you're saying libvirt needs
>>>> to query mtree to understand the extent of the GPU layout, but we need
>>>> to specify the locked memory limits in order for QEMU to start?  Is
>>>> libvirt supposed to start the VM with unlimited locked memory and fix
>>>> it at some indeterminate point in the future?  Run a dummy VM with
>>>> unlimited locked memory in order to determine the limits for the real
>>>> VM?  Neither of these sound practical.  Thanks,  
>>>
>>>
>>> QEMU maps GPU RAM at known locations (which only depends on the vPHB's
>>> index or can be set explicitely) and libvirt knows how many GPUs are
>>> passed so it is quite easy to calculate the required amount of memory.
>>>
>>> Here is the window start calculation:
>>> https://github.com/aik/qemu/commit/7073cad3ae7708d657e01672bcf53092808b54fb#diff-662409c2a5a150fe231d07ea8384b920R3812
>>>
>>> We do not exactly know the GPU RAM window size until QEMU reads it from
>>> VFIO/nvlink2 but we know that all existing hardware has a window of
>>> 128GB (the adapters I have access to only have 16/32GB on board).
>>
>> So you're asking that libvirt add 128GB per GPU with magic nvlink
>> properties, which may be 8x what's actually necessary and libvirt
>> determines which GPUs to apply this to how?  Does libvirt need to sort
>> through device tree properties for this?  Thanks,
> 
> Hm.  If the GPU memory is really separate from main RAM, which it
> sounds like, I don't think it makes sense to account it against the
> same locked memory limit as regular RAM.


This is accounting for TCE table to cover GPU RAM, not for GPU RAM itself.

So I am asking libvirt to add 128GB/64k*8=16MB to the locked_vm. It
already does so for the guest RAM.


-- 
Alexey

Re: [Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[Alex Williamson]

On Mon, 11 Feb 2019 14:49:49 +1100
Alexey Kardashevskiy <aik@ozlabs.ru> wrote:

> On 08/02/2019 16:28, David Gibson wrote:
> > On Thu, Feb 07, 2019 at 08:26:20PM -0700, Alex Williamson wrote:  
> >> On Fri, 8 Feb 2019 13:29:37 +1100
> >> Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> >>  
> >>> On 08/02/2019 02:18, Alex Williamson wrote:  
> >>>> On Thu, 7 Feb 2019 15:43:18 +1100
> >>>> Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> >>>>     
> >>>>> On 07/02/2019 04:22, Daniel Henrique Barboza wrote:    
> >>>>>> Based on this series, I've sent a Libvirt patch to allow a QEMU process
> >>>>>> to inherit IPC_LOCK when using VFIO passthrough with the Tesla V100
> >>>>>> GPU:
> >>>>>>
> >>>>>> https://www.redhat.com/archives/libvir-list/2019-February/msg00219.html
> >>>>>>
> >>>>>>
> >>>>>> In that thread, Alex raised concerns about allowing QEMU to freely lock
> >>>>>> all the memory it wants. Is this an issue to be considered in the review
> >>>>>> of this series here?
> >>>>>>
> >>>>>> Reading the patches, specially patch 3/3, it seems to me that QEMU is
> >>>>>> going to lock the KVM memory to populate the NUMA node with memory
> >>>>>> of the GPU itself, so at first there is no risk of not taking over the
> >>>>>> host RAM.
> >>>>>> Am I missing something?      
> >>>>>
> >>>>>
> >>>>> The GPU memory belongs to the device and not visible to the host as
> >>>>> memory blocks and not covered by page structs, for the host it is more
> >>>>> like MMIO which is passed through to the guest without that locked
> >>>>> accounting, I'd expect libvirt to keep working as usual except that:
> >>>>>
> >>>>> when libvirt calculates the amount of memory needed for TCE tables
> >>>>> (which is guestRAM/64k*8), now it needs to use the end of the last GPU
> >>>>> RAM window as a guest RAM size. For example, in QEMU HMP "info mtree -f":
> >>>>>
> >>>>> FlatView #2
> >>>>>  AS "memory", root: system
> >>>>>  AS "cpu-memory-0", root: system
> >>>>>  Root memory region: system
> >>>>>   0000000000000000-000000007fffffff (prio 0, ram): ppc_spapr.ram
> >>>>>   0000010000000000-0000011fffffffff (prio 0, ram): nvlink2-mr
> >>>>>
> >>>>> So previously the DMA window would cover 0x7fffffff+1, now it has to
> >>>>> cover 0x11fffffffff+1.    
> >>>>
> >>>> This looks like a chicken and egg problem, you're saying libvirt needs
> >>>> to query mtree to understand the extent of the GPU layout, but we need
> >>>> to specify the locked memory limits in order for QEMU to start?  Is
> >>>> libvirt supposed to start the VM with unlimited locked memory and fix
> >>>> it at some indeterminate point in the future?  Run a dummy VM with
> >>>> unlimited locked memory in order to determine the limits for the real
> >>>> VM?  Neither of these sound practical.  Thanks,    
> >>>
> >>>
> >>> QEMU maps GPU RAM at known locations (which only depends on the vPHB's
> >>> index or can be set explicitely) and libvirt knows how many GPUs are
> >>> passed so it is quite easy to calculate the required amount of memory.
> >>>
> >>> Here is the window start calculation:
> >>> https://github.com/aik/qemu/commit/7073cad3ae7708d657e01672bcf53092808b54fb#diff-662409c2a5a150fe231d07ea8384b920R3812
> >>>
> >>> We do not exactly know the GPU RAM window size until QEMU reads it from
> >>> VFIO/nvlink2 but we know that all existing hardware has a window of
> >>> 128GB (the adapters I have access to only have 16/32GB on board).  
> >>
> >> So you're asking that libvirt add 128GB per GPU with magic nvlink
> >> properties, which may be 8x what's actually necessary and libvirt
> >> determines which GPUs to apply this to how?  Does libvirt need to sort
> >> through device tree properties for this?  Thanks,  
> > 
> > Hm.  If the GPU memory is really separate from main RAM, which it
> > sounds like, I don't think it makes sense to account it against the
> > same locked memory limit as regular RAM.  
> 
> 
> This is accounting for TCE table to cover GPU RAM, not for GPU RAM itself.
> 
> So I am asking libvirt to add 128GB/64k*8=16MB to the locked_vm. It
> already does so for the guest RAM.

Why do host internal data structures count against the user's locked
memory limit?  We don't include IOMMU page tables or type1 accounting
structures on other archs.  Thanks,

Alex

Re: [Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[Alexey Kardashevskiy]

On 11/02/2019 17:07, Alex Williamson wrote:
> On Mon, 11 Feb 2019 14:49:49 +1100
> Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> 
>> On 08/02/2019 16:28, David Gibson wrote:
>>> On Thu, Feb 07, 2019 at 08:26:20PM -0700, Alex Williamson wrote:  
>>>> On Fri, 8 Feb 2019 13:29:37 +1100
>>>> Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
>>>>  
>>>>> On 08/02/2019 02:18, Alex Williamson wrote:  
>>>>>> On Thu, 7 Feb 2019 15:43:18 +1100
>>>>>> Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
>>>>>>     
>>>>>>> On 07/02/2019 04:22, Daniel Henrique Barboza wrote:    
>>>>>>>> Based on this series, I've sent a Libvirt patch to allow a QEMU process
>>>>>>>> to inherit IPC_LOCK when using VFIO passthrough with the Tesla V100
>>>>>>>> GPU:
>>>>>>>>
>>>>>>>> https://www.redhat.com/archives/libvir-list/2019-February/msg00219.html
>>>>>>>>
>>>>>>>>
>>>>>>>> In that thread, Alex raised concerns about allowing QEMU to freely lock
>>>>>>>> all the memory it wants. Is this an issue to be considered in the review
>>>>>>>> of this series here?
>>>>>>>>
>>>>>>>> Reading the patches, specially patch 3/3, it seems to me that QEMU is
>>>>>>>> going to lock the KVM memory to populate the NUMA node with memory
>>>>>>>> of the GPU itself, so at first there is no risk of not taking over the
>>>>>>>> host RAM.
>>>>>>>> Am I missing something?      
>>>>>>>
>>>>>>>
>>>>>>> The GPU memory belongs to the device and not visible to the host as
>>>>>>> memory blocks and not covered by page structs, for the host it is more
>>>>>>> like MMIO which is passed through to the guest without that locked
>>>>>>> accounting, I'd expect libvirt to keep working as usual except that:
>>>>>>>
>>>>>>> when libvirt calculates the amount of memory needed for TCE tables
>>>>>>> (which is guestRAM/64k*8), now it needs to use the end of the last GPU
>>>>>>> RAM window as a guest RAM size. For example, in QEMU HMP "info mtree -f":
>>>>>>>
>>>>>>> FlatView #2
>>>>>>>  AS "memory", root: system
>>>>>>>  AS "cpu-memory-0", root: system
>>>>>>>  Root memory region: system
>>>>>>>   0000000000000000-000000007fffffff (prio 0, ram): ppc_spapr.ram
>>>>>>>   0000010000000000-0000011fffffffff (prio 0, ram): nvlink2-mr
>>>>>>>
>>>>>>> So previously the DMA window would cover 0x7fffffff+1, now it has to
>>>>>>> cover 0x11fffffffff+1.    
>>>>>>
>>>>>> This looks like a chicken and egg problem, you're saying libvirt needs
>>>>>> to query mtree to understand the extent of the GPU layout, but we need
>>>>>> to specify the locked memory limits in order for QEMU to start?  Is
>>>>>> libvirt supposed to start the VM with unlimited locked memory and fix
>>>>>> it at some indeterminate point in the future?  Run a dummy VM with
>>>>>> unlimited locked memory in order to determine the limits for the real
>>>>>> VM?  Neither of these sound practical.  Thanks,    
>>>>>
>>>>>
>>>>> QEMU maps GPU RAM at known locations (which only depends on the vPHB's
>>>>> index or can be set explicitely) and libvirt knows how many GPUs are
>>>>> passed so it is quite easy to calculate the required amount of memory.
>>>>>
>>>>> Here is the window start calculation:
>>>>> https://github.com/aik/qemu/commit/7073cad3ae7708d657e01672bcf53092808b54fb#diff-662409c2a5a150fe231d07ea8384b920R3812
>>>>>
>>>>> We do not exactly know the GPU RAM window size until QEMU reads it from
>>>>> VFIO/nvlink2 but we know that all existing hardware has a window of
>>>>> 128GB (the adapters I have access to only have 16/32GB on board).  
>>>>
>>>> So you're asking that libvirt add 128GB per GPU with magic nvlink
>>>> properties, which may be 8x what's actually necessary and libvirt
>>>> determines which GPUs to apply this to how?  Does libvirt need to sort
>>>> through device tree properties for this?  Thanks,  
>>>
>>> Hm.  If the GPU memory is really separate from main RAM, which it
>>> sounds like, I don't think it makes sense to account it against the
>>> same locked memory limit as regular RAM.  
>>
>>
>> This is accounting for TCE table to cover GPU RAM, not for GPU RAM itself.
>>
>> So I am asking libvirt to add 128GB/64k*8=16MB to the locked_vm. It
>> already does so for the guest RAM.
> 
> Why do host internal data structures count against the user's locked
> memory limit?  We don't include IOMMU page tables or type1 accounting
> structures on other archs.  Thanks,


Because pseries guests create DMA windows dynamically and the userspace
can pass multiple devices to a guest, placing each on its own vPHB each
of which most likely will create an additional 64bit DMA window which is
backed with an IOMMU table => the userspace triggers these allocations.
We account guest RAM once as it is shared among vPHBs but not the IOMMU
tables.


-- 
Alexey

Re: [Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[David Gibson]

[-- Attachment #1: Type: text/plain, Size: 5717 bytes --]

On Mon, Feb 11, 2019 at 06:46:32PM +1100, Alexey Kardashevskiy wrote:
> 
> 
> On 11/02/2019 17:07, Alex Williamson wrote:
> > On Mon, 11 Feb 2019 14:49:49 +1100
> > Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> > 
> >> On 08/02/2019 16:28, David Gibson wrote:
> >>> On Thu, Feb 07, 2019 at 08:26:20PM -0700, Alex Williamson wrote:  
> >>>> On Fri, 8 Feb 2019 13:29:37 +1100
> >>>> Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> >>>>  
> >>>>> On 08/02/2019 02:18, Alex Williamson wrote:  
> >>>>>> On Thu, 7 Feb 2019 15:43:18 +1100
> >>>>>> Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> >>>>>>     
> >>>>>>> On 07/02/2019 04:22, Daniel Henrique Barboza wrote:    
> >>>>>>>> Based on this series, I've sent a Libvirt patch to allow a QEMU process
> >>>>>>>> to inherit IPC_LOCK when using VFIO passthrough with the Tesla V100
> >>>>>>>> GPU:
> >>>>>>>>
> >>>>>>>> https://www.redhat.com/archives/libvir-list/2019-February/msg00219.html
> >>>>>>>>
> >>>>>>>>
> >>>>>>>> In that thread, Alex raised concerns about allowing QEMU to freely lock
> >>>>>>>> all the memory it wants. Is this an issue to be considered in the review
> >>>>>>>> of this series here?
> >>>>>>>>
> >>>>>>>> Reading the patches, specially patch 3/3, it seems to me that QEMU is
> >>>>>>>> going to lock the KVM memory to populate the NUMA node with memory
> >>>>>>>> of the GPU itself, so at first there is no risk of not taking over the
> >>>>>>>> host RAM.
> >>>>>>>> Am I missing something?      
> >>>>>>>
> >>>>>>>
> >>>>>>> The GPU memory belongs to the device and not visible to the host as
> >>>>>>> memory blocks and not covered by page structs, for the host it is more
> >>>>>>> like MMIO which is passed through to the guest without that locked
> >>>>>>> accounting, I'd expect libvirt to keep working as usual except that:
> >>>>>>>
> >>>>>>> when libvirt calculates the amount of memory needed for TCE tables
> >>>>>>> (which is guestRAM/64k*8), now it needs to use the end of the last GPU
> >>>>>>> RAM window as a guest RAM size. For example, in QEMU HMP "info mtree -f":
> >>>>>>>
> >>>>>>> FlatView #2
> >>>>>>>  AS "memory", root: system
> >>>>>>>  AS "cpu-memory-0", root: system
> >>>>>>>  Root memory region: system
> >>>>>>>   0000000000000000-000000007fffffff (prio 0, ram): ppc_spapr.ram
> >>>>>>>   0000010000000000-0000011fffffffff (prio 0, ram): nvlink2-mr
> >>>>>>>
> >>>>>>> So previously the DMA window would cover 0x7fffffff+1, now it has to
> >>>>>>> cover 0x11fffffffff+1.    
> >>>>>>
> >>>>>> This looks like a chicken and egg problem, you're saying libvirt needs
> >>>>>> to query mtree to understand the extent of the GPU layout, but we need
> >>>>>> to specify the locked memory limits in order for QEMU to start?  Is
> >>>>>> libvirt supposed to start the VM with unlimited locked memory and fix
> >>>>>> it at some indeterminate point in the future?  Run a dummy VM with
> >>>>>> unlimited locked memory in order to determine the limits for the real
> >>>>>> VM?  Neither of these sound practical.  Thanks,    
> >>>>>
> >>>>>
> >>>>> QEMU maps GPU RAM at known locations (which only depends on the vPHB's
> >>>>> index or can be set explicitely) and libvirt knows how many GPUs are
> >>>>> passed so it is quite easy to calculate the required amount of memory.
> >>>>>
> >>>>> Here is the window start calculation:
> >>>>> https://github.com/aik/qemu/commit/7073cad3ae7708d657e01672bcf53092808b54fb#diff-662409c2a5a150fe231d07ea8384b920R3812
> >>>>>
> >>>>> We do not exactly know the GPU RAM window size until QEMU reads it from
> >>>>> VFIO/nvlink2 but we know that all existing hardware has a window of
> >>>>> 128GB (the adapters I have access to only have 16/32GB on board).  
> >>>>
> >>>> So you're asking that libvirt add 128GB per GPU with magic nvlink
> >>>> properties, which may be 8x what's actually necessary and libvirt
> >>>> determines which GPUs to apply this to how?  Does libvirt need to sort
> >>>> through device tree properties for this?  Thanks,  
> >>>
> >>> Hm.  If the GPU memory is really separate from main RAM, which it
> >>> sounds like, I don't think it makes sense to account it against the
> >>> same locked memory limit as regular RAM.  
> >>
> >>
> >> This is accounting for TCE table to cover GPU RAM, not for GPU RAM itself.
> >>
> >> So I am asking libvirt to add 128GB/64k*8=16MB to the locked_vm. It
> >> already does so for the guest RAM.
> > 
> > Why do host internal data structures count against the user's locked
> > memory limit?  We don't include IOMMU page tables or type1 accounting
> > structures on other archs.  Thanks,
> 
> Because pseries guests create DMA windows dynamically and the userspace
> can pass multiple devices to a guest, placing each on its own vPHB each
> of which most likely will create an additional 64bit DMA window which is
> backed with an IOMMU table => the userspace triggers these allocations.
> We account guest RAM once as it is shared among vPHBs but not the IOMMU
> tables.

Uh.. I think that's missing the point.

The real reason is that on x86, IOMMU page tables (IIUC) live within
guest memory, so they're essentially already accounted for.

Under PAPR, the IOMMU tables live outside the guest accessed via
hypercalls.  So, they consume hypervisor memory that the guest can
cause to be allocated.  We need to account that somewhere, so the
guest can't cause the hypervisor to allocate arbitrary amounts of
space.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

Re: [Qemu-devel] [PATCH qemu 0/3] spapr_pci, vfio: NVIDIA V100 + P9 passthrough

[David Gibson]

[-- Attachment #1: Type: text/plain, Size: 4521 bytes --]

On Mon, Feb 11, 2019 at 02:49:49PM +1100, Alexey Kardashevskiy wrote:
> 
> 
> On 08/02/2019 16:28, David Gibson wrote:
> > On Thu, Feb 07, 2019 at 08:26:20PM -0700, Alex Williamson wrote:
> >> On Fri, 8 Feb 2019 13:29:37 +1100
> >> Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> >>
> >>> On 08/02/2019 02:18, Alex Williamson wrote:
> >>>> On Thu, 7 Feb 2019 15:43:18 +1100
> >>>> Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> >>>>   
> >>>>> On 07/02/2019 04:22, Daniel Henrique Barboza wrote:  
> >>>>>> Based on this series, I've sent a Libvirt patch to allow a QEMU process
> >>>>>> to inherit IPC_LOCK when using VFIO passthrough with the Tesla V100
> >>>>>> GPU:
> >>>>>>
> >>>>>> https://www.redhat.com/archives/libvir-list/2019-February/msg00219.html
> >>>>>>
> >>>>>>
> >>>>>> In that thread, Alex raised concerns about allowing QEMU to freely lock
> >>>>>> all the memory it wants. Is this an issue to be considered in the review
> >>>>>> of this series here?
> >>>>>>
> >>>>>> Reading the patches, specially patch 3/3, it seems to me that QEMU is
> >>>>>> going to lock the KVM memory to populate the NUMA node with memory
> >>>>>> of the GPU itself, so at first there is no risk of not taking over the
> >>>>>> host RAM.
> >>>>>> Am I missing something?    
> >>>>>
> >>>>>
> >>>>> The GPU memory belongs to the device and not visible to the host as
> >>>>> memory blocks and not covered by page structs, for the host it is more
> >>>>> like MMIO which is passed through to the guest without that locked
> >>>>> accounting, I'd expect libvirt to keep working as usual except that:
> >>>>>
> >>>>> when libvirt calculates the amount of memory needed for TCE tables
> >>>>> (which is guestRAM/64k*8), now it needs to use the end of the last GPU
> >>>>> RAM window as a guest RAM size. For example, in QEMU HMP "info mtree -f":
> >>>>>
> >>>>> FlatView #2
> >>>>>  AS "memory", root: system
> >>>>>  AS "cpu-memory-0", root: system
> >>>>>  Root memory region: system
> >>>>>   0000000000000000-000000007fffffff (prio 0, ram): ppc_spapr.ram
> >>>>>   0000010000000000-0000011fffffffff (prio 0, ram): nvlink2-mr
> >>>>>
> >>>>> So previously the DMA window would cover 0x7fffffff+1, now it has to
> >>>>> cover 0x11fffffffff+1.  
> >>>>
> >>>> This looks like a chicken and egg problem, you're saying libvirt needs
> >>>> to query mtree to understand the extent of the GPU layout, but we need
> >>>> to specify the locked memory limits in order for QEMU to start?  Is
> >>>> libvirt supposed to start the VM with unlimited locked memory and fix
> >>>> it at some indeterminate point in the future?  Run a dummy VM with
> >>>> unlimited locked memory in order to determine the limits for the real
> >>>> VM?  Neither of these sound practical.  Thanks,  
> >>>
> >>>
> >>> QEMU maps GPU RAM at known locations (which only depends on the vPHB's
> >>> index or can be set explicitely) and libvirt knows how many GPUs are
> >>> passed so it is quite easy to calculate the required amount of memory.
> >>>
> >>> Here is the window start calculation:
> >>> https://github.com/aik/qemu/commit/7073cad3ae7708d657e01672bcf53092808b54fb#diff-662409c2a5a150fe231d07ea8384b920R3812
> >>>
> >>> We do not exactly know the GPU RAM window size until QEMU reads it from
> >>> VFIO/nvlink2 but we know that all existing hardware has a window of
> >>> 128GB (the adapters I have access to only have 16/32GB on board).
> >>
> >> So you're asking that libvirt add 128GB per GPU with magic nvlink
> >> properties, which may be 8x what's actually necessary and libvirt
> >> determines which GPUs to apply this to how?  Does libvirt need to sort
> >> through device tree properties for this?  Thanks,
> > 
> > Hm.  If the GPU memory is really separate from main RAM, which it
> > sounds like, I don't think it makes sense to account it against the
> > same locked memory limit as regular RAM.
> 
> This is accounting for TCE table to cover GPU RAM, not for GPU RAM itself.

Ah, ok, that makes sense then

> So I am asking libvirt to add 128GB/64k*8=16MB to the locked_vm. It
> already does so for the guest RAM.

That seems reasonable.  IIRC we already have some slop in the amount
of locked vm that libvirt allocates; not sure if it'll be enough as
is.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]