powerpc/powernv/npu: Rework TCE Kill handling
authorAlexey Kardashevskiy <aik@ozlabs.ru>
Fri, 29 Apr 2016 08:55:23 +0000 (18:55 +1000)
committerMichael Ellerman <mpe@ellerman.id.au>
Wed, 11 May 2016 11:54:31 +0000 (21:54 +1000)
The pnv_ioda_pe struct keeps an array of peers. At the moment it is only
used to link GPU and NPU for 2 purposes:

1. Access NPU quickly when configuring DMA for GPU - this was addressed
in the previos patch by removing use of it as DMA setup is not what
the kernel would constantly do.

2. Invalidate TCE cache for NPU when it is invalidated for GPU.
GPU and NPU are in different PE. There is already a mechanism to
attach multiple iommu_table_group to the same iommu_table (used for VFIO),
we can reuse it here so does this patch.

This gets rid of peers[] array and PNV_IODA_PE_PEER flag as they are
not needed anymore.

While we are here, add TCE cache invalidation after enabling bypass.

Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-By: Alistair Popple <alistair@popple.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
arch/powerpc/platforms/powernv/npu-dma.c
arch/powerpc/platforms/powernv/pci-ioda.c
arch/powerpc/platforms/powernv/pci.h

index 800d70f..cb2d1da 100644 (file)
@@ -136,22 +136,17 @@ static struct pnv_ioda_pe *get_gpu_pci_dev_and_pe(struct pnv_ioda_pe *npe,
        struct pnv_ioda_pe *pe;
        struct pci_dn *pdn;
 
-       if (npe->flags & PNV_IODA_PE_PEER) {
-               pe = npe->peers[0];
-               pdev = pe->pdev;
-       } else {
-               pdev = pnv_pci_get_gpu_dev(npe->pdev);
-               if (!pdev)
-                       return NULL;
-
-               pdn = pci_get_pdn(pdev);
-               if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
-                       return NULL;
+       pdev = pnv_pci_get_gpu_dev(npe->pdev);
+       if (!pdev)
+               return NULL;
 
-               hose = pci_bus_to_host(pdev->bus);
-               phb = hose->private_data;
-               pe = &phb->ioda.pe_array[pdn->pe_number];
-       }
+       pdn = pci_get_pdn(pdev);
+       if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
+               return NULL;
+
+       hose = pci_bus_to_host(pdev->bus);
+       phb = hose->private_data;
+       pe = &phb->ioda.pe_array[pdn->pe_number];
 
        if (gpdev)
                *gpdev = pdev;
@@ -186,6 +181,10 @@ static long pnv_npu_set_window(struct pnv_ioda_pe *npe,
        }
        pnv_pci_ioda2_tce_invalidate_entire(phb, false);
 
+       /* Add the table to the list so its TCE cache will get invalidated */
+       pnv_pci_link_table_and_group(phb->hose->node, 0,
+                       tbl, &npe->table_group);
+
        return 0;
 }
 
@@ -206,43 +205,10 @@ static long pnv_npu_unset_window(struct pnv_ioda_pe *npe)
        }
        pnv_pci_ioda2_tce_invalidate_entire(phb, false);
 
-       return 0;
-}
-
-void pnv_npu_init_dma_pe(struct pnv_ioda_pe *npe)
-{
-       struct pnv_ioda_pe *gpe;
-       struct pci_dev *gpdev;
-       int i, avail = -1;
-
-       if (!npe->pdev || !(npe->flags & PNV_IODA_PE_DEV))
-               return;
-
-       gpe = get_gpu_pci_dev_and_pe(npe, &gpdev);
-       if (!gpe)
-               return;
-
-       for (i = 0; i < PNV_IODA_MAX_PEER_PES; i++) {
-               /* Nothing to do if the PE is already connected. */
-               if (gpe->peers[i] == npe)
-                       return;
-
-               if (!gpe->peers[i])
-                       avail = i;
-       }
+       pnv_pci_unlink_table_and_group(npe->table_group.tables[0],
+                       &npe->table_group);
 
-       if (WARN_ON(avail < 0))
-               return;
-
-       gpe->peers[avail] = npe;
-       gpe->flags |= PNV_IODA_PE_PEER;
-
-       /*
-        * We assume that the NPU devices only have a single peer PE
-        * (the GPU PCIe device PE).
-        */
-       npe->peers[0] = gpe;
-       npe->flags |= PNV_IODA_PE_PEER;
+       return 0;
 }
 
 /*
@@ -302,6 +268,9 @@ static int pnv_npu_dma_set_bypass(struct pnv_ioda_pe *npe)
                        npe->pe_number, npe->pe_number,
                        0 /* bypass base */, top);
 
+       if (rc == OPAL_SUCCESS)
+               pnv_pci_ioda2_tce_invalidate_entire(phb, false);
+
        return rc;
 }
 
index b2831c0..ef31218 100644 (file)
@@ -1845,23 +1845,12 @@ static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe *pe)
        /* 01xb - invalidate TCEs that match the specified PE# */
        unsigned long val = TCE_KILL_INVAL_PE | (pe->pe_number & 0xFF);
        struct pnv_phb *phb = pe->phb;
-       struct pnv_ioda_pe *npe;
-       int i;
 
        if (!phb->ioda.tce_inval_reg)
                return;
 
        mb(); /* Ensure above stores are visible */
        __raw_writeq(cpu_to_be64(val), phb->ioda.tce_inval_reg);
-
-       if (pe->flags & PNV_IODA_PE_PEER)
-               for (i = 0; i < PNV_IODA_MAX_PEER_PES; i++) {
-                       npe = pe->peers[i];
-                       if (!npe || npe->phb->type != PNV_PHB_NPU)
-                               continue;
-
-                       pnv_pci_ioda2_tce_invalidate_entire(npe->phb, false);
-               }
 }
 
 static void pnv_pci_ioda2_do_tce_invalidate(unsigned pe_number, bool rm,
@@ -1896,33 +1885,24 @@ static void pnv_pci_ioda2_tce_invalidate(struct iommu_table *tbl,
        struct iommu_table_group_link *tgl;
 
        list_for_each_entry_rcu(tgl, &tbl->it_group_list, next) {
-               struct pnv_ioda_pe *npe;
                struct pnv_ioda_pe *pe = container_of(tgl->table_group,
                                struct pnv_ioda_pe, table_group);
                __be64 __iomem *invalidate = rm ?
                        (__be64 __iomem *)pe->phb->ioda.tce_inval_reg_phys :
                        pe->phb->ioda.tce_inval_reg;
-               int i;
-
-               pnv_pci_ioda2_do_tce_invalidate(pe->pe_number, rm,
-                       invalidate, tbl->it_page_shift,
-                       index, npages);
 
-               if (pe->flags & PNV_IODA_PE_PEER)
+               if (pe->phb->type == PNV_PHB_NPU) {
                        /*
                         * The NVLink hardware does not support TCE kill
                         * per TCE entry so we have to invalidate
                         * the entire cache for it.
                         */
-                       for (i = 0; i < PNV_IODA_MAX_PEER_PES; i++) {
-                               npe = pe->peers[i];
-                               if (!npe || npe->phb->type != PNV_PHB_NPU ||
-                                               !npe->phb->ioda.tce_inval_reg)
-                                       continue;
-
-                               pnv_pci_ioda2_tce_invalidate_entire(npe->phb,
-                                               rm);
-                       }
+                       pnv_pci_ioda2_tce_invalidate_entire(pe->phb, rm);
+                       continue;
+               }
+               pnv_pci_ioda2_do_tce_invalidate(pe->pe_number, rm,
+                       invalidate, tbl->it_page_shift,
+                       index, npages);
        }
 }
 
@@ -3156,31 +3136,6 @@ static void pnv_pci_ioda_create_dbgfs(void)
 #endif /* CONFIG_DEBUG_FS */
 }
 
-static void pnv_npu_ioda_fixup(void)
-{
-       bool enable_bypass;
-       struct pci_controller *hose, *tmp;
-       struct pnv_phb *phb;
-       struct pnv_ioda_pe *pe;
-       unsigned int weight;
-
-       list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
-               phb = hose->private_data;
-               if (phb->type != PNV_PHB_NPU)
-                       continue;
-
-               list_for_each_entry(pe, &phb->ioda.pe_list, list) {
-                       weight = pnv_pci_ioda_pe_dma_weight(pe);
-                       if (WARN_ON(!weight))
-                               continue;
-
-                       enable_bypass = dma_get_mask(&pe->pdev->dev) ==
-                               DMA_BIT_MASK(64);
-                       pnv_npu_init_dma_pe(pe);
-               }
-       }
-}
-
 static void pnv_pci_ioda_fixup(void)
 {
        pnv_pci_ioda_setup_PEs();
@@ -3193,9 +3148,6 @@ static void pnv_pci_ioda_fixup(void)
        eeh_init();
        eeh_addr_cache_build();
 #endif
-
-       /* Link NPU IODA tables to their PCI devices. */
-       pnv_npu_ioda_fixup();
 }
 
 /*
index b426614..2b35606 100644 (file)
@@ -24,7 +24,6 @@ enum pnv_phb_model {
 #define PNV_IODA_PE_MASTER     (1 << 3)        /* Master PE in compound case   */
 #define PNV_IODA_PE_SLAVE      (1 << 4)        /* Slave PE in compound case    */
 #define PNV_IODA_PE_VF         (1 << 5)        /* PE for one VF                */
-#define PNV_IODA_PE_PEER       (1 << 6)        /* PE has peers                 */
 
 /* Data associated with a PE, including IOMMU tracking etc.. */
 struct pnv_phb;
@@ -32,9 +31,6 @@ struct pnv_ioda_pe {
        unsigned long           flags;
        struct pnv_phb          *phb;
 
-#define PNV_IODA_MAX_PEER_PES  8
-       struct pnv_ioda_pe      *peers[PNV_IODA_MAX_PEER_PES];
-
        /* A PE can be associated with a single device or an
         * entire bus (& children). In the former case, pdev
         * is populated, in the later case, pbus is.
@@ -227,8 +223,6 @@ extern void pe_level_printk(const struct pnv_ioda_pe *pe, const char *level,
        pe_level_printk(pe, KERN_INFO, fmt, ##__VA_ARGS__)
 
 /* Nvlink functions */
-extern void pnv_npu_init_dma_pe(struct pnv_ioda_pe *npe);
-extern void pnv_npu_setup_dma_pe(struct pnv_ioda_pe *npe);
 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);