On Thu, Jul 30, 2015 at 11:15:01AM +1000, Gavin Shan wrote:

On Wed, Jul 29, 2015 at 03:22:07PM +0800, Wei Yang wrote:

quoted

In current implementation, when VF BAR is bigger than 64MB, it uses 4 M64
BAR in Single PE mode to cover the number of VFs required to be enabled.
By doing so, several VFs would be in one VF Group and leads to interference
between VFs in the same group.

This patch changes the design by using one M64 BAR in Single PE mode for
one VF BAR. This gives absolute isolation for VFs.

Signed-off-by: Wei Yang <redacted>
---
arch/powerpc/include/asm/pci-bridge.h     |    5 +-
arch/powerpc/platforms/powernv/pci-ioda.c |  104 +++++------------------------
2 files changed, 18 insertions(+), 91 deletions(-)

questions regarding this:

(1) When M64 BAR is running in single-PE-mode for VFs, the alignment for one
   particular IOV BAR still have to be (IOV_BAR_size * max_vf_number), or
   M64 segment size of last BAR (0x10000000) is fine? If the later one is fine,
   more M64 space would be saved. On the other hand, if the IOV BAR size
   (for all VFs) is less than 256MB, will the allocated resource conflict
   with the M64 segments in last BAR?

Not need to be IOV BAR size aligned, be individual VF BAR size aligned is fine.

IOV BAR size = VF BAR size * expended_num_vfs

(2) When M64 BAR is in single-PE-mode, the PE numbers allocated for VFs need
   continuous or not.

No, not need.

(3) Each PF could have 6 IOV BARs and there're 15 available M64 BAR. It means
   only two VFs can be enabled in the extreme case. Would it be a problem?

Yes, you are right.

Based on Alexey's mail, full isolation is more important than more VFs.

quoted

diff --git a/arch/powerpc/include/asm/pci-bridge.h b/arch/powerpc/include/asm/pci-bridge.h
index 712add5..1997e5d 100644
--- a/arch/powerpc/include/asm/pci-bridge.h
+++ b/arch/powerpc/include/asm/pci-bridge.h

@@ -214,10 +214,9 @@ struct pci_dn {

	u16     vfs_expanded;		/* number of VFs IOV BAR expanded */
	u16     num_vfs;		/* number of VFs enabled*/
	int     offset;			/* PE# for the first VF PE */
-#define M64_PER_IOV 4
-	int     m64_per_iov;
+#define MAX_M64_WINDOW  16
#define IODA_INVALID_M64        (-1)
-	int     m64_wins[PCI_SRIOV_NUM_BARS][M64_PER_IOV];
+	int     m64_wins[PCI_SRIOV_NUM_BARS][MAX_M64_WINDOW];
#endif /* CONFIG_PCI_IOV */
#endif

The "m64_wins" would be renamed to "m64_map". Also, it would have dynamic size:

- When the IOV BAR is extended to 256 segments, its size is sizeof(int) * PCI_SRIOV_NUM_BARS;
- When the IOV BAR is extended to max_vf_num, its size is sizeof(int) * max_vf_num;

quoted

	struct list_head child_list;

diff --git a/arch/powerpc/platforms/powernv/pci-ioda.c b/arch/powerpc/platforms/powernv/pci-ioda.c
index 5738d31..b3e7909 100644
--- a/arch/powerpc/platforms/powernv/pci-ioda.c
+++ b/arch/powerpc/platforms/powernv/pci-ioda.c

@@ -1168,7 +1168,7 @@ static int pnv_pci_vf_release_m64(struct pci_dev *pdev)

	pdn = pci_get_pdn(pdev);

	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++)
-		for (j = 0; j < M64_PER_IOV; j++) {
+		for (j = 0; j < MAX_M64_WINDOW; j++) {
			if (pdn->m64_wins[i][j] == IODA_INVALID_M64)
				continue;
			opal_pci_phb_mmio_enable(phb->opal_id,

@@ -1193,8 +1193,7 @@ static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)

	int                    total_vfs;
	resource_size_t        size, start;
	int                    pe_num;
-	int                    vf_groups;
-	int                    vf_per_group;
+	int                    m64s;

"m64s" could have better name. For example, "vfs_per_m64_bar"...

m64s is used to represent number of M64 BARs necessary to enable num_vfs.
vfs_per_m64_bar may be misleading.

How about "m64_bars" ?

quoted

	bus = pdev->bus;
	hose = pci_bus_to_host(bus);

@@ -1204,17 +1203,13 @@ static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)

	/* Initialize the m64_wins to IODA_INVALID_M64 */
	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++)
-		for (j = 0; j < M64_PER_IOV; j++)
+		for (j = 0; j < MAX_M64_WINDOW; j++)
			pdn->m64_wins[i][j] = IODA_INVALID_M64;

-	if (pdn->m64_per_iov == M64_PER_IOV) {
-		vf_groups = (num_vfs <= M64_PER_IOV) ? num_vfs: M64_PER_IOV;
-		vf_per_group = (num_vfs <= M64_PER_IOV)? 1:
-			roundup_pow_of_two(num_vfs) / pdn->m64_per_iov;
-	} else {
-		vf_groups = 1;
-		vf_per_group = 1;
-	}
+	if (pdn->vfs_expanded != phb->ioda.total_pe)
+		m64s = num_vfs;
+	else
+		m64s = 1;

The condition (pdn->vfs_expanded != phb->ioda.total_pe) isn't precise enough as
explained below.

quoted

	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
		res = &pdev->resource[i + PCI_IOV_RESOURCES];

@@ -1224,7 +1219,7 @@ static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)

		if (!pnv_pci_is_mem_pref_64(res->flags))
			continue;

-		for (j = 0; j < vf_groups; j++) {
+		for (j = 0; j < m64s; j++) {
			do {
				win = find_next_zero_bit(&phb->ioda.m64_bar_alloc,
						phb->ioda.m64_bar_idx + 1, 0);

@@ -1235,10 +1230,9 @@ static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)

			pdn->m64_wins[i][j] = win;

-			if (pdn->m64_per_iov == M64_PER_IOV) {
+			if (pdn->vfs_expanded != phb->ioda.total_pe) {
				size = pci_iov_resource_size(pdev,
							PCI_IOV_RESOURCES + i);
-				size = size * vf_per_group;
				start = res->start + size * j;
			} else {
				size = resource_size(res);

@@ -1246,7 +1240,7 @@ static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)

			}

			/* Map the M64 here */
-			if (pdn->m64_per_iov == M64_PER_IOV) {
+			if (pdn->vfs_expanded != phb->ioda.total_pe) {
				pe_num = pdn->offset + j;
				rc = opal_pci_map_pe_mmio_window(phb->opal_id,
						pe_num, OPAL_M64_WINDOW_TYPE,

@@ -1267,7 +1261,7 @@ static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)

				goto m64_failed;
			}

-			if (pdn->m64_per_iov == M64_PER_IOV)
+			if (pdn->vfs_expanded != phb->ioda.total_pe)
				rc = opal_pci_phb_mmio_enable(phb->opal_id,
				     OPAL_M64_WINDOW_TYPE, pdn->m64_wins[i][j], 2);
			else

@@ -1311,15 +1305,13 @@ static void pnv_pci_ioda2_release_dma_pe(struct pci_dev *dev, struct pnv_ioda_pe

	iommu_free_table(tbl, of_node_full_name(dev->dev.of_node));
}

-static void pnv_ioda_release_vf_PE(struct pci_dev *pdev, u16 num_vfs)
+static void pnv_ioda_release_vf_PE(struct pci_dev *pdev)
{
	struct pci_bus        *bus;
	struct pci_controller *hose;
	struct pnv_phb        *phb;
	struct pnv_ioda_pe    *pe, *pe_n;
	struct pci_dn         *pdn;
-	u16                    vf_index;
-	int64_t                rc;

	bus = pdev->bus;
	hose = pci_bus_to_host(bus);

@@ -1329,35 +1321,6 @@ static void pnv_ioda_release_vf_PE(struct pci_dev *pdev, u16 num_vfs)

	if (!pdev->is_physfn)
		return;

-	if (pdn->m64_per_iov == M64_PER_IOV && num_vfs > M64_PER_IOV) {
-		int   vf_group;
-		int   vf_per_group;
-		int   vf_index1;
-
-		vf_per_group = roundup_pow_of_two(num_vfs) / pdn->m64_per_iov;
-
-		for (vf_group = 0; vf_group < M64_PER_IOV; vf_group++)
-			for (vf_index = vf_group * vf_per_group;
-				vf_index < (vf_group + 1) * vf_per_group &&
-				vf_index < num_vfs;
-				vf_index++)
-				for (vf_index1 = vf_group * vf_per_group;
-					vf_index1 < (vf_group + 1) * vf_per_group &&
-					vf_index1 < num_vfs;
-					vf_index1++){
-
-					rc = opal_pci_set_peltv(phb->opal_id,
-						pdn->offset + vf_index,
-						pdn->offset + vf_index1,
-						OPAL_REMOVE_PE_FROM_DOMAIN);
-
-					if (rc)
-					    dev_warn(&pdev->dev, "%s: Failed to unlink same group PE#%d(%lld)\n",
-						__func__,
-						pdn->offset + vf_index1, rc);
-				}
-	}
-
	list_for_each_entry_safe(pe, pe_n, &phb->ioda.pe_list, list) {
		if (pe->parent_dev != pdev)
			continue;

@@ -1392,10 +1355,10 @@ void pnv_pci_sriov_disable(struct pci_dev *pdev)

	num_vfs = pdn->num_vfs;

	/* Release VF PEs */
-	pnv_ioda_release_vf_PE(pdev, num_vfs);
+	pnv_ioda_release_vf_PE(pdev);

	if (phb->type == PNV_PHB_IODA2) {
-		if (pdn->m64_per_iov == 1)
+		if (pdn->vfs_expanded == phb->ioda.total_pe)
			pnv_pci_vf_resource_shift(pdev, -pdn->offset);

		/* Release M64 windows */

@@ -1418,7 +1381,6 @@ static void pnv_ioda_setup_vf_PE(struct pci_dev *pdev, u16 num_vfs)

	int                    pe_num;
	u16                    vf_index;
	struct pci_dn         *pdn;
-	int64_t                rc;

	bus = pdev->bus;
	hose = pci_bus_to_host(bus);

@@ -1463,37 +1425,6 @@ static void pnv_ioda_setup_vf_PE(struct pci_dev *pdev, u16 num_vfs)

		pnv_pci_ioda2_setup_dma_pe(phb, pe);
	}
-
-	if (pdn->m64_per_iov == M64_PER_IOV && num_vfs > M64_PER_IOV) {
-		int   vf_group;
-		int   vf_per_group;
-		int   vf_index1;
-
-		vf_per_group = roundup_pow_of_two(num_vfs) / pdn->m64_per_iov;
-
-		for (vf_group = 0; vf_group < M64_PER_IOV; vf_group++) {
-			for (vf_index = vf_group * vf_per_group;
-			     vf_index < (vf_group + 1) * vf_per_group &&
-			     vf_index < num_vfs;
-			     vf_index++) {
-				for (vf_index1 = vf_group * vf_per_group;
-				     vf_index1 < (vf_group + 1) * vf_per_group &&
-				     vf_index1 < num_vfs;
-				     vf_index1++) {
-
-					rc = opal_pci_set_peltv(phb->opal_id,
-						pdn->offset + vf_index,
-						pdn->offset + vf_index1,
-						OPAL_ADD_PE_TO_DOMAIN);
-
-					if (rc)
-					    dev_warn(&pdev->dev, "%s: Failed to link same group PE#%d(%lld)\n",
-						__func__,
-						pdn->offset + vf_index1, rc);
-				}
-			}
-		}
-	}
}

int pnv_pci_sriov_enable(struct pci_dev *pdev, u16 num_vfs)

@@ -1537,7 +1468,7 @@ int pnv_pci_sriov_enable(struct pci_dev *pdev, u16 num_vfs)

		 * the IOV BAR according to the PE# allocated to the VFs.
		 * Otherwise, the PE# for the VF will conflict with others.
		 */
-		if (pdn->m64_per_iov == 1) {
+		if (pdn->vfs_expanded == phb->ioda.total_pe) {

This condition isn't precise enough. When PF occasionally supports 256 VFs
and the summed size of all IOV BARs (explained below) exceeds 64MB, we're
expecting to use singole-pe-mode M64 BARs, not shared-mode.

Yes, you are right. The vfs_expanded is not reliable.

quoted

			ret = pnv_pci_vf_resource_shift(pdev, pdn->offset);
			if (ret)
				goto m64_failed;

@@ -1570,8 +1501,7 @@ int pcibios_sriov_enable(struct pci_dev *pdev, u16 num_vfs)

	/* Allocate PCI data */
	add_dev_pci_data(pdev);

-	pnv_pci_sriov_enable(pdev, num_vfs);
-	return 0;
+	return pnv_pci_sriov_enable(pdev, num_vfs);
}
#endif /* CONFIG_PCI_IOV */

@@ -2766,7 +2696,6 @@ static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev *pdev)

	pdn->vfs_expanded = 0;

	total_vfs = pci_sriov_get_totalvfs(pdev);
-	pdn->m64_per_iov = 1;
	mul = phb->ioda.total_pe;

	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {

@@ -2785,7 +2714,6 @@ static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev *pdev)

		if (size > (1 << 26)) {

Actually, the condition isn't precise enough. In theory, every PF can have 6 IOV BARs.
If all of their size are 64MB, we will have 256 extended VFs. The total MMIO size needed
is: 96GB = (6 * 64MB * 256), which exceeds 64GB. The original idea would be to have
the scheme other than extending to 256 VFs when the sum of all IOV BARs is bigger
than 64MB, not single M64 BAR. It's different issue and you can fix it up in another
patch if you want.

I didn't get your point here.

You mean it is necessary to check the sum of IOV BAR instead of a single one?

quoted

			dev_info(&pdev->dev, "PowerNV: VF BAR%d: %pR IOV size is bigger than 64M, roundup power2\n",
				 i, res);
-			pdn->m64_per_iov = M64_PER_IOV;
			mul = roundup_pow_of_two(total_vfs);
			break;
		}
-- 
1.7.9.5

-- 
Richard Yang
Help you, Help me