On Wed, Apr 06, 2016 at 01:40:36PM +1000, Alexey Kardashevskiy wrote:
> On 04/06/2016 12:41 PM, David Gibson wrote:
> >On Wed, Apr 06, 2016 at 11:45:34AM +1000, Alexey Kardashevskiy wrote:
> >>Ping?
> >>
> >>On 03/24/2016 12:42 PM, Alexey Kardashevskiy wrote:
> >>>IBM POWER8 NVlink systems come with Tesla K40-ish GPUs each of which
> >>>also has a couple of fast speed links (NVLink). The interface to links
> >>>is exposed as an emulated PCI bridge which is included into the same
> >>>IOMMU group as the corresponding GPU.
> >>>
> >>>In the kernel, NPUs get a separate PHB of the PNV_PHB_NPU type and a PE.
> >>>
> >>>In order to make these links work when GPU is passed to the guest,
> >>>these bridges need to be passed as well; otherwise performance will
> >>>degrade.
> >>>
> >>>This implements and exports API to manage NPU state in regard to VFIO;
> >>>it replicates iommu_table_group_ops.
> >>>
> >>>This defines a new pnv_pci_ioda2_npu_ops which is assigned to
> >>>the IODA2 bridge if there are NPUs for a GPU on the bridge.
> >>>The new callbacks call the default IODA2 callbacks plus new NPU API.
> >>>This adds a gpe_table_group_to_npe() helper to find NPU PE for the IODA2
> >>>table_group, it is not expected to fail as the helper is only called
> >>>from the pnv_pci_ioda2_npu_ops.
> >>>
> >>>This adds a pnv_pci_npu_setup_iommu() helper which adds NPUs to
> >>>the GPU group if any found. The helper uses helpers to look for
> >>>the "ibm,gpu" property in the device tree which is a phandle of
> >>>the corresponding GPU.
> >>>
> >>>This adds an additional loop over PEs in pnv_ioda_setup_dma() as the main
> >>>loop skips NPU PEs as they do not have 32bit DMA segments.
> >>>
> >>>Signed-off-by: Alexey Kardashevskiy <a...@ozlabs.ru>
> >
> >Looks correct conceptually, a few smallish queries below.
> >
> >
> >>>---
> >>>Changes:
> >>>v2:
> >>>* reimplemented to support NPU + GPU in the same group
> >>>* merged "powerpc/powernv/npu: Add NPU devices to IOMMU group" and
> >>>"powerpc/powernv/npu: Enable passing through via VFIO" into this patch
> >>>
> >>>---
> >>>
> >>>The rest of the series is the same, I only merged 2 patches into one and
> >>>reworked it to have GPU and NPU in the same IOMMU group like:
> >>>
> >>>aik@g86L:~$ lspci | grep -e '\(NVIDIA\|IBM Device 04ea\)'
> >>>0002:01:00.0 3D controller: NVIDIA Corporation Device 15ff (rev a1)
> >>>0003:01:00.0 3D controller: NVIDIA Corporation Device 15ff (rev a1)
> >>>0006:01:00.0 3D controller: NVIDIA Corporation Device 15ff (rev a1)
> >>>0007:01:00.0 3D controller: NVIDIA Corporation Device 15ff (rev a1)
> >>>0008:00:00.0 Bridge: IBM Device 04ea
> >>>0008:00:00.1 Bridge: IBM Device 04ea
> >>>0008:00:01.0 Bridge: IBM Device 04ea
> >>>0008:00:01.1 Bridge: IBM Device 04ea
> >>>0009:00:00.0 Bridge: IBM Device 04ea
> >>>0009:00:00.1 Bridge: IBM Device 04ea
> >>>0009:00:01.0 Bridge: IBM Device 04ea
> >>>0009:00:01.1 Bridge: IBM Device 04ea
> >>>aik@g86L:~$ ls /sys/bus/pci/devices/0002\:01\:00.0/iommu_group/devices/
> >>>0002:01:00.0 0008:00:01.0 0008:00:01.1
> >>>aik@g86L:~$ ls /sys/bus/pci/devices/0003\:01\:00.0/iommu_group/devices/
> >>>0003:01:00.0 0008:00:00.0 0008:00:00.1
> >>>aik@g86L:~$ ls /sys/bus/pci/devices/0006\:01\:00.0/iommu_group/devices/
> >>>0006:01:00.0 0009:00:01.0 0009:00:01.1
> >>>aik@g86L:~$ ls /sys/bus/pci/devices/0007\:01\:00.0/iommu_group/devices/
> >>>0007:01:00.0 0009:00:00.0 0009:00:00.1
> >>>
> >>>
> >>>Please comment. If this one is ok, I'll repost the whole thing. Thanks!
> >>>
> >>>
> >>>---
> >>> arch/powerpc/platforms/powernv/npu-dma.c | 129
> >>> ++++++++++++++++++++++++++++++
> >>> arch/powerpc/platforms/powernv/pci-ioda.c | 101 +++++++++++++++++++++++
> >>> arch/powerpc/platforms/powernv/pci.h | 6 ++
> >>> 3 files changed, 236 insertions(+)
> >>>
> >>>diff --git a/arch/powerpc/platforms/powernv/npu-dma.c
> >>>b/arch/powerpc/platforms/powernv/npu-dma.c
> >>>index 8e70221..d048e0e 100644
> >>>--- a/arch/powerpc/platforms/powernv/npu-dma.c
> >>>+++ b/arch/powerpc/platforms/powernv/npu-dma.c
> >>>@@ -262,3 +262,132 @@ void pnv_npu_try_dma_set_bypass(struct pci_dev
> >>>*gpdev, bool bypass)
> >>> }
> >>> }
> >>> }
> >>>+
> >>>+long pnv_npu_set_window(struct pnv_ioda_pe *npe, int num,
> >>>+ struct iommu_table *tbl)
> >>>+{
> >>>+ struct pnv_phb *phb = npe->phb;
> >>>+ int64_t rc;
> >>>+ const unsigned long size = tbl->it_indirect_levels ?
> >>>+ tbl->it_level_size : tbl->it_size;
> >>>+ const __u64 start_addr = tbl->it_offset << tbl->it_page_shift;
> >>>+ const __u64 win_size = tbl->it_size << tbl->it_page_shift;
> >>>+
> >>>+ pe_info(npe, "Setting up window#%d %llx..%llx pg=%lx\n", num,
> >>>+ start_addr, start_addr + win_size - 1,
> >>>+ IOMMU_PAGE_SIZE(tbl));
> >>>+
> >>>+ rc = opal_pci_map_pe_dma_window(phb->opal_id,
> >>>+ npe->pe_number,
> >>>+ npe->pe_number,
> >>>+ tbl->it_indirect_levels + 1,
> >>>+ __pa(tbl->it_base),
> >>>+ size << 3,
> >>>+ IOMMU_PAGE_SIZE(tbl));
> >>>+ if (rc) {
> >>>+ pe_err(npe, "Failed to configure TCE table, err %lld\n", rc);
> >>>+ return rc;
> >>>+ }
> >>>+
> >>>+ pnv_pci_link_table_and_group(phb->hose->node, num,
> >>>+ tbl, &npe->table_group);
> >>>+ pnv_pci_ioda2_tce_invalidate_entire(npe->phb, false);
> >>>+
> >>>+ return rc;
> >>>+}
> >>>+
> >>>+long pnv_npu_unset_window(struct pnv_ioda_pe *npe, int num)
> >>>+{
> >>>+ struct pnv_phb *phb = npe->phb;
> >>>+ long ret;
> >>>+
> >>>+ pe_info(npe, "Removing DMA window #%d\n", num);
> >>>+
> >>>+ ret = opal_pci_map_pe_dma_window(phb->opal_id, npe->pe_number,
> >>>+ npe->pe_number,
> >>>+ 0/* levels */, 0/* table address */,
> >>>+ 0/* table size */, 0/* page size */);
> >>>+ if (ret)
> >>>+ pe_warn(npe, "Unmapping failed, ret = %ld\n", ret);
> >>>+ else
> >>>+ pnv_pci_ioda2_tce_invalidate_entire(npe->phb, false);
> >>>+
> >>>+ pnv_pci_unlink_table_and_group(npe->table_group.tables[num],
> >>>+ &npe->table_group);
> >>>+
> >>>+ return ret;
> >>>+}
> >>>+
> >>>+/* Switch ownership from platform code to external user (e.g. VFIO) */
> >>>+void pnv_npu_take_ownership(struct pnv_ioda_pe *npe)
> >>>+{
> >>>+ struct pnv_phb *phb = npe->phb;
> >>>+ int64_t ret;
> >>>+
> >>>+ if (npe->table_group.tables[0]) {
> >>>+ pnv_pci_unlink_table_and_group(npe->table_group.tables[0],
> >>>+ &npe->table_group);
> >>>+ npe->table_group.tables[0] = NULL;
> >>>+ ret = opal_pci_map_pe_dma_window(phb->opal_id, npe->pe_number,
> >>>+ npe->pe_number,
> >>>+ 0/* levels */, 0/* table address */,
> >>>+ 0/* table size */, 0/* page size */);
> >>>+ } else {
> >>>+ ret = opal_pci_map_pe_dma_window_real(phb->opal_id,
> >>>+ npe->pe_number, npe->pe_number,
> >>>+ 0 /* bypass base */, 0);
> >
> >I take it this call disables iommu bypass on the NPU?
>
>
> Yes.
>
>
> >
> >>>+ }
> >>>+
> >>>+ if (ret != OPAL_SUCCESS)
> >>>+ pe_err(npe, "Failed to remove DMA window");
> >>>+ else
> >>>+ pnv_pci_ioda2_tce_invalidate_entire(npe->phb, false);
> >>>+}
> >>>+
> >>>+/* Switch ownership from external user (e.g. VFIO) back to core */
> >>>+void pnv_npu_release_ownership(struct pnv_ioda_pe *npe)
> >>>+{
> >>>+ struct pnv_phb *phb = npe->phb;
> >>>+ int64_t ret;
> >>>+
> >>>+ ret = opal_pci_map_pe_dma_window(phb->opal_id, npe->pe_number,
> >>>+ npe->pe_number,
> >>>+ 0/* levels */, 0/* table address */,
> >>>+ 0/* table size */, 0/* page size */);
> >>>+ if (ret != OPAL_SUCCESS)
> >>>+ pe_err(npe, "Failed to remove DMA window");
> >>>+ else
> >>>+ pnv_pci_ioda2_tce_invalidate_entire(npe->phb, false);
> >>>+}
> >>>+
> >>>+struct pnv_ioda_pe *pnv_pci_npu_setup_iommu(struct pnv_ioda_pe *npe)
> >>>+{
> >>>+ struct iommu_table *tbl;
> >>>+ struct pnv_phb *phb = npe->phb;
> >>>+ struct pci_bus *pbus = phb->hose->bus;
> >>>+ struct pci_dev *npdev, *gpdev = NULL, *gptmp;
> >>>+ struct pnv_ioda_pe *gpe = get_gpu_pci_dev_and_pe(npe, &gpdev);
> >>>+
> >>>+ if (!gpe || !gpdev)
> >>>+ return NULL;
> >>>+
> >>>+ list_for_each_entry(npdev, &pbus->devices, bus_list) {
> >>>+ gptmp = pnv_pci_get_gpu_dev(npdev);
> >>>+
> >>>+ if (gptmp != gpdev)
> >>>+ continue;
> >>>+ /*
> >>>+ * The iommu_add_device() picks an IOMMU group from
> >>>+ * the first IOMMU group attached to the iommu_table
> >>>+ * so we need to pretend that there is a table so
> >>>+ * iommu_add_device() can complete the job.
> >>>+ * We unlink the tempopary table from the group afterwards.
> >>>+ */
> >>>+ tbl = get_iommu_table_base(&gpdev->dev);
> >>>+ set_iommu_table_base(&npdev->dev, tbl);
> >>>+ iommu_add_device(&npdev->dev);
> >>>+ set_iommu_table_base(&npdev->dev, NULL);
> >
> >As far as I can tell, this is correct, but it's a bit icky,
> >temporarily changing the state just so iommu_add_device() can use it -
> >and I worry a bit about races.
>
> This is a boot time thing, not racy afaict.
>
>
> >I'd prefer to see a new variant of iommu_add_device() which takes an
> >explicit table instead of getting it from get_iommu_table_base(). The
> >regular iommu_add_device() could be implemented in terms of that
> >variant, obviously.
>
>
> Hm. After reading the code, I could actually just call the generic
> iommu_group_add_device() directly here as I do not really care about the
> checks which iommu_add_device() performs (may be except
> device_is_registered()) in addition to just calling
> iommu_group_add_device(), would not this be better?
Yes, that sounds like a better idea.
> >>>+ }
> >>>+
> >>>+ return gpe;
> >>>+}
> >>>diff --git a/arch/powerpc/platforms/powernv/pci-ioda.c
> >>>b/arch/powerpc/platforms/powernv/pci-ioda.c
> >>>index e765870..fa6278b 100644
> >>>--- a/arch/powerpc/platforms/powernv/pci-ioda.c
> >>>+++ b/arch/powerpc/platforms/powernv/pci-ioda.c
> >>>@@ -2299,6 +2299,96 @@ static struct iommu_table_group_ops
> >>>pnv_pci_ioda2_ops = {
> >>> .take_ownership = pnv_ioda2_take_ownership,
> >>> .release_ownership = pnv_ioda2_release_ownership,
> >>> };
> >>>+
> >>>+static int gpe_table_group_to_npe_cb(struct device *dev, void *opaque)
> >>>+{
> >>>+ struct pnv_ioda_pe **ptmppe = opaque;
> >>>+ struct pci_dev *pdev = container_of(dev, struct pci_dev, dev);
> >>>+ struct pci_controller *hose = pci_bus_to_host(pdev->bus);
> >>>+ struct pnv_phb *phb = hose->private_data;
> >>>+ struct pci_dn *pdn = pci_get_pdn(pdev);
> >>>+ struct pnv_ioda_pe *pe;
> >>>+
> >>>+ if (!pdn || pdn->pe_number == IODA_INVALID_PE)
> >>>+ return 0;
> >>>+
> >>>+ pe = &phb->ioda.pe_array[pdn->pe_number];
> >>>+ if (pe == *ptmppe)
> >>>+ return 0;
> >
> >I'm not quite sure what the test above is for. It looks like you're
> >excluding devices in the PE of the GPU device, because you're trying
> >to find the NPU device instead, but shouldn't that be covered by the
> >test below?
>
>
> Just an extra check which I'll remove. And I won't have to rely on the
> initial content of *ptmppe which is cleaner.
Ok, good.
> >>>+ if (phb->type != PNV_PHB_NPU)
> >>>+ return 0;
> >>>+
> >>>+ *ptmppe = pe;
> >>>+ return 1;
> >>>+}
> >>>+
> >>>+/*
> >>>+ * This returns PE of associated NPU.
> >>>+ * This assumes that NPU is in the same IOMMU group with GPU and there is
> >>>+ * no other PEs.
> >>>+ */
> >>>+static struct pnv_ioda_pe *gpe_table_group_to_npe(
> >>>+ struct iommu_table_group *table_group)
> >>>+{
> >>>+ struct pnv_ioda_pe *npe = container_of(table_group, struct pnv_ioda_pe,
> >>>+ table_group);
> >>>+ int ret = iommu_group_for_each_dev(table_group->group, &npe,
> >>>+ gpe_table_group_to_npe_cb);
> >>>+
> >>>+ BUG_ON(!ret || !npe);
> >>>+
> >>>+ return npe;
> >>>+}
> >
> >Am I correct in thinking that this logic (and indeed, this interface)
> >depends on there being at most one NPU per iommu group? That's not
> >necessarily a problem, just trying to make sure I understand
> >correctly.
>
> One NPU PE per a GPU PE. Each NPU PE has 2 NPU devices (i.e. 2 links).
Ah, yes, sorry. I meant one NPU PE per GPU PE.
> >>>+static long pnv_pci_ioda2_npu_set_window(struct iommu_table_group
> >>>*table_group,
> >>>+ int num, struct iommu_table *tbl)
> >>>+{
> >>>+ long ret = pnv_pci_ioda2_set_window(table_group, num, tbl);
> >>>+
> >>>+ if (ret)
> >>>+ return ret;
> >>>+
> >>>+ ret = pnv_npu_set_window(gpe_table_group_to_npe(table_group), num, tbl);
> >>>+ if (ret)
> >>>+ pnv_pci_ioda2_unset_window(table_group, num);
> >>>+
> >>>+ return ret;
> >>>+}
> >>>+
> >>>+static long pnv_pci_ioda2_npu_unset_window(
> >>>+ struct iommu_table_group *table_group,
> >>>+ int num)
> >>>+{
> >>>+ long ret = pnv_pci_ioda2_unset_window(table_group, num);
> >>>+
> >>>+ if (ret)
> >>>+ return ret;
> >>>+
> >>>+ return pnv_npu_unset_window(gpe_table_group_to_npe(table_group), num);
> >>>+}
> >>>+
> >>>+static void pnv_ioda2_npu_take_ownership(struct iommu_table_group
> >>>*table_group)
> >>>+{
> >>>+ pnv_ioda2_take_ownership(table_group);
> >>>+ pnv_npu_take_ownership(gpe_table_group_to_npe(table_group));
> >>>+}
> >>>+
> >>>+static void pnv_ioda2_npu_release_ownership(
> >>>+ struct iommu_table_group *table_group)
> >>>+{
> >>>+ pnv_npu_release_ownership(gpe_table_group_to_npe(table_group));
> >>>+ pnv_ioda2_release_ownership(table_group);
> >>>+}
> >>>+
> >>>+static struct iommu_table_group_ops pnv_pci_ioda2_npu_ops = {
> >>>+ .get_table_size = pnv_pci_ioda2_get_table_size,
> >>>+ .create_table = pnv_pci_ioda2_create_table,
> >>>+ .set_window = pnv_pci_ioda2_npu_set_window,
> >>>+ .unset_window = pnv_pci_ioda2_npu_unset_window,
> >>>+ .take_ownership = pnv_ioda2_npu_take_ownership,
> >>>+ .release_ownership = pnv_ioda2_npu_release_ownership,
> >>>+};
> >>> #endif
> >>>
> >>> static void pnv_pci_ioda_setup_opal_tce_kill(struct pnv_phb *phb)
> >>>@@ -2563,6 +2653,17 @@ static void pnv_ioda_setup_dma(struct pnv_phb *phb)
> >>> remaining -= segs;
> >>> base += segs;
> >>> }
> >>>+ /*
> >>>+ * Create an IOMMU group and add devices to it.
> >>>+ * DMA setup is done via GPU's dma_set_mask().
> >>>+ */
> >>>+ if (phb->type == PNV_PHB_NPU) {
> >>>+ list_for_each_entry(pe, &phb->ioda.pe_dma_list, dma_link) {
> >>>+ struct pnv_ioda_pe *gpe = pnv_pci_npu_setup_iommu(pe);
> >>>+ if (gpe)
> >>>+ gpe->table_group.ops = &pnv_pci_ioda2_npu_ops;
> >>>+ }
> >>>+ }
> >
> >It looks like this relies on running *after* the setup for the
> >ordinary GPU devices has been done, so it can modify the ops pointer
> >to add the NPU stuff. Does anything ensure that the bridge discovery
> >happens in that order?
>
>
> Good point. No, it is possible (at least in theory) that NPU PHBs appear
> first and GPU later, this will fail. I'll move this part after all PHBs are
> discovered, to pnv_pci_ioda_fixup().
O.
> Thanks for the review!
>
>
> >>> }
> >>>
> >>> #ifdef CONFIG_PCI_MSI
> >>>diff --git a/arch/powerpc/platforms/powernv/pci.h
> >>>b/arch/powerpc/platforms/powernv/pci.h
> >>>index f9c3aca..4200bb9 100644
> >>>--- a/arch/powerpc/platforms/powernv/pci.h
> >>>+++ b/arch/powerpc/platforms/powernv/pci.h
> >>>@@ -250,5 +250,11 @@ extern void pe_level_printk(const struct pnv_ioda_pe
> >>>*pe, const char *level,
> >>> /* Nvlink functions */
> >>> extern void pnv_npu_try_dma_set_bypass(struct pci_dev *gpdev, bool
> >>> bypass);
> >>> extern void pnv_pci_ioda2_tce_invalidate_entire(struct pnv_phb *phb,
> >>> bool rm);
> >>>+extern struct pnv_ioda_pe *pnv_pci_npu_setup_iommu(struct pnv_ioda_pe
> >>>*npe);
> >>>+extern long pnv_npu_set_window(struct pnv_ioda_pe *npe, int num,
> >>>+ struct iommu_table *tbl);
> >>>+extern long pnv_npu_unset_window(struct pnv_ioda_pe *npe, int num);
> >>>+extern void pnv_npu_take_ownership(struct pnv_ioda_pe *npe);
> >>>+extern void pnv_npu_release_ownership(struct pnv_ioda_pe *npe);
> >>>
> >>> #endif /* __POWERNV_PCI_H */
> >>>
> >>
> >>
> >
>
>
--
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
signature.asc
Description: PGP signature
_______________________________________________ Linuxppc-dev mailing list Linuxppc-dev@lists.ozlabs.org https://lists.ozlabs.org/listinfo/linuxppc-dev
