powerpc/powernv: Group VF PE when IOV BAR is big on PHB3
authorWei Yang <weiyang@linux.vnet.ibm.com>
Wed, 25 Mar 2015 08:23:59 +0000 (16:23 +0800)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Tue, 31 Mar 2015 02:02:38 +0000 (13:02 +1100)
When IOV BAR is big, each is covered by 4 M64 windows.  This leads to
several VF PE sits in one PE in terms of M64.

Group VF PEs according to the M64 allocation.

[bhelgaas: use dev_printk() when possible]
Signed-off-by: Wei Yang <weiyang@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/include/asm/pci-bridge.h
arch/powerpc/platforms/powernv/pci-ioda.c

index 415df8509f5207b7463ec06877c867f0b7d218de..560c739964742782def9dc5ab2eb8b6bc277848e 100644 (file)
@@ -185,7 +185,7 @@ struct pci_dn {
 #define M64_PER_IOV 4
        int     m64_per_iov;
 #define IODA_INVALID_M64        (-1)
-       int     m64_wins[PCI_SRIOV_NUM_BARS];
+       int     m64_wins[PCI_SRIOV_NUM_BARS][M64_PER_IOV];
 #endif /* CONFIG_PCI_IOV */
 #endif
        struct list_head child_list;
index b63925f483fc36d19c6ed7c0af5eb4281c6424e7..33088f6f732848747ca100b9f75267d3e06888cd 100644 (file)
@@ -1156,26 +1156,27 @@ static int pnv_pci_vf_release_m64(struct pci_dev *pdev)
        struct pci_controller *hose;
        struct pnv_phb        *phb;
        struct pci_dn         *pdn;
-       int                    i;
+       int                    i, j;
 
        bus = pdev->bus;
        hose = pci_bus_to_host(bus);
        phb = hose->private_data;
        pdn = pci_get_pdn(pdev);
 
-       for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
-               if (pdn->m64_wins[i] == IODA_INVALID_M64)
-                       continue;
-               opal_pci_phb_mmio_enable(phb->opal_id,
-                               OPAL_M64_WINDOW_TYPE, pdn->m64_wins[i], 0);
-               clear_bit(pdn->m64_wins[i], &phb->ioda.m64_bar_alloc);
-               pdn->m64_wins[i] = IODA_INVALID_M64;
-       }
+       for (i = 0; i < PCI_SRIOV_NUM_BARS; i++)
+               for (j = 0; j < M64_PER_IOV; j++) {
+                       if (pdn->m64_wins[i][j] == IODA_INVALID_M64)
+                               continue;
+                       opal_pci_phb_mmio_enable(phb->opal_id,
+                               OPAL_M64_WINDOW_TYPE, pdn->m64_wins[i][j], 0);
+                       clear_bit(pdn->m64_wins[i][j], &phb->ioda.m64_bar_alloc);
+                       pdn->m64_wins[i][j] = IODA_INVALID_M64;
+               }
 
        return 0;
 }
 
-static int pnv_pci_vf_assign_m64(struct pci_dev *pdev)
+static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)
 {
        struct pci_bus        *bus;
        struct pci_controller *hose;
@@ -1183,17 +1184,33 @@ static int pnv_pci_vf_assign_m64(struct pci_dev *pdev)
        struct pci_dn         *pdn;
        unsigned int           win;
        struct resource       *res;
-       int                    i;
+       int                    i, j;
        int64_t                rc;
+       int                    total_vfs;
+       resource_size_t        size, start;
+       int                    pe_num;
+       int                    vf_groups;
+       int                    vf_per_group;
 
        bus = pdev->bus;
        hose = pci_bus_to_host(bus);
        phb = hose->private_data;
        pdn = pci_get_pdn(pdev);
+       total_vfs = pci_sriov_get_totalvfs(pdev);
 
        /* Initialize the m64_wins to IODA_INVALID_M64 */
        for (i = 0; i < PCI_SRIOV_NUM_BARS; i++)
-               pdn->m64_wins[i] = IODA_INVALID_M64;
+               for (j = 0; j < M64_PER_IOV; 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;
+       }
 
        for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
                res = &pdev->resource[i + PCI_IOV_RESOURCES];
@@ -1203,35 +1220,61 @@ static int pnv_pci_vf_assign_m64(struct pci_dev *pdev)
                if (!pnv_pci_is_mem_pref_64(res->flags))
                        continue;
 
-               do {
-                       win = find_next_zero_bit(&phb->ioda.m64_bar_alloc,
-                                       phb->ioda.m64_bar_idx + 1, 0);
-
-                       if (win >= phb->ioda.m64_bar_idx + 1)
-                               goto m64_failed;
-               } while (test_and_set_bit(win, &phb->ioda.m64_bar_alloc));
+               for (j = 0; j < vf_groups; j++) {
+                       do {
+                               win = find_next_zero_bit(&phb->ioda.m64_bar_alloc,
+                                               phb->ioda.m64_bar_idx + 1, 0);
+
+                               if (win >= phb->ioda.m64_bar_idx + 1)
+                                       goto m64_failed;
+                       } while (test_and_set_bit(win, &phb->ioda.m64_bar_alloc));
+
+                       pdn->m64_wins[i][j] = win;
+
+                       if (pdn->m64_per_iov == M64_PER_IOV) {
+                               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);
+                               start = res->start;
+                       }
 
-               pdn->m64_wins[i] = win;
+                       /* Map the M64 here */
+                       if (pdn->m64_per_iov == M64_PER_IOV) {
+                               pe_num = pdn->offset + j;
+                               rc = opal_pci_map_pe_mmio_window(phb->opal_id,
+                                               pe_num, OPAL_M64_WINDOW_TYPE,
+                                               pdn->m64_wins[i][j], 0);
+                       }
 
-               /* Map the M64 here */
-               rc = opal_pci_set_phb_mem_window(phb->opal_id,
+                       rc = opal_pci_set_phb_mem_window(phb->opal_id,
                                                 OPAL_M64_WINDOW_TYPE,
-                                                pdn->m64_wins[i],
-                                                res->start,
+                                                pdn->m64_wins[i][j],
+                                                start,
                                                 0, /* unused */
-                                                resource_size(res));
-               if (rc != OPAL_SUCCESS) {
-                       dev_err(&pdev->dev, "Failed to map M64 window #%d: %lld\n",
-                               win, rc);
-                       goto m64_failed;
-               }
+                                                size);
 
-               rc = opal_pci_phb_mmio_enable(phb->opal_id,
-                               OPAL_M64_WINDOW_TYPE, pdn->m64_wins[i], 1);
-               if (rc != OPAL_SUCCESS) {
-                       dev_err(&pdev->dev, "Failed to enable M64 window #%d: %llx\n",
-                               win, rc);
-                       goto m64_failed;
+
+                       if (rc != OPAL_SUCCESS) {
+                               dev_err(&pdev->dev, "Failed to map M64 window #%d: %lld\n",
+                                       win, rc);
+                               goto m64_failed;
+                       }
+
+                       if (pdn->m64_per_iov == M64_PER_IOV)
+                               rc = opal_pci_phb_mmio_enable(phb->opal_id,
+                                    OPAL_M64_WINDOW_TYPE, pdn->m64_wins[i][j], 2);
+                       else
+                               rc = opal_pci_phb_mmio_enable(phb->opal_id,
+                                    OPAL_M64_WINDOW_TYPE, pdn->m64_wins[i][j], 1);
+
+                       if (rc != OPAL_SUCCESS) {
+                               dev_err(&pdev->dev, "Failed to enable M64 window #%d: %llx\n",
+                                       win, rc);
+                               goto m64_failed;
+                       }
                }
        }
        return 0;
@@ -1273,22 +1316,53 @@ static void pnv_pci_ioda2_release_dma_pe(struct pci_dev *dev, struct pnv_ioda_pe
        pe->tce32_table = NULL;
 }
 
-static void pnv_ioda_release_vf_PE(struct pci_dev *pdev)
+static void pnv_ioda_release_vf_PE(struct pci_dev *pdev, u16 num_vfs)
 {
        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);
        phb = hose->private_data;
+       pdn = pci_get_pdn(pdev);
 
        if (!pdev->is_physfn)
                return;
 
-       pdn = pci_get_pdn(pdev);
+       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;
@@ -1323,10 +1397,11 @@ void pnv_pci_sriov_disable(struct pci_dev *pdev)
        num_vfs = pdn->num_vfs;
 
        /* Release VF PEs */
-       pnv_ioda_release_vf_PE(pdev);
+       pnv_ioda_release_vf_PE(pdev, num_vfs);
 
        if (phb->type == PNV_PHB_IODA2) {
-               pnv_pci_vf_resource_shift(pdev, -pdn->offset);
+               if (pdn->m64_per_iov == 1)
+                       pnv_pci_vf_resource_shift(pdev, -pdn->offset);
 
                /* Release M64 windows */
                pnv_pci_vf_release_m64(pdev);
@@ -1348,6 +1423,7 @@ 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);
@@ -1396,6 +1472,37 @@ 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)
@@ -1428,7 +1535,7 @@ int pnv_pci_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
                mutex_unlock(&phb->ioda.pe_alloc_mutex);
 
                /* Assign M64 window accordingly */
-               ret = pnv_pci_vf_assign_m64(pdev);
+               ret = pnv_pci_vf_assign_m64(pdev, num_vfs);
                if (ret) {
                        dev_info(&pdev->dev, "Not enough M64 window resources\n");
                        goto m64_failed;
@@ -1439,9 +1546,11 @@ 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.
                 */
-               ret = pnv_pci_vf_resource_shift(pdev, pdn->offset);
-               if (ret)
-                       goto m64_failed;
+               if (pdn->m64_per_iov == 1) {
+                       ret = pnv_pci_vf_resource_shift(pdev, pdn->offset);
+                       if (ret)
+                               goto m64_failed;
+               }
        }
 
        /* Setup VF PEs */