Re: [PATCH 23/32] KVM: s390: pci: handle refresh of PCI translations

On Tue, 2021-12-07 at 15:57 -0500, Matthew Rosato wrote:

Hmm, this looks like a virtual vs physical address mixup to me that is
currently not a problem because kernel virtual addresses are equal to
their physical address. Here page_to_virt(page) gives us a virtual
address but the entries in the I/O translation table have to be
physical (aka absolute) addresses.

With my commit "s390/pci: use physical addresses in DMA tables"
currently in the s390 feature branch this is also reflected in the
argument types taken by set_pt_pfaa() below so gaddr should have type
phys_addr_t not void *. That should also remove the need for the cast
to unsigned long for the duplicate check.

+	}
+
+	if (pt_entry_isvalid(*entry)) {
+		/* Either we are invalidating, replacing or no-op */
+		if (gaddr) {
+			if ((*entry & ZPCI_PTE_ADDR_MASK) ==
+			    (unsigned long)gaddr) {
+				/* Duplicate */
+				kvm_release_pfn_dirty(*entry >> PAGE_SHIFT);
+			} else {
+				/* Replace */
+				pfn = (*entry >> PAGE_SHIFT);
+				invalidate_pt_entry(entry);
+				set_pt_pfaa(entry, gaddr);
+				validate_pt_entry(entry);
+				kvm_release_pfn_dirty(pfn);
+				rc = 1;
+			}
+		} else {
+			/* Invalidate */
+			pfn = (*entry >> PAGE_SHIFT);
+			invalidate_pt_entry(entry);
+			kvm_release_pfn_dirty(pfn);
+			rc = 1;
+		}
+	} else if (gaddr) {
+		/* New Entry */
+		set_pt_pfaa(entry, gaddr);
+		validate_pt_entry(entry);
+	}
+
+	return rc;
+}
+
+unsigned long *dma_walk_guest_cpu_trans(struct kvm_vcpu *vcpu,
+					struct kvm_zdev_ioat *ioat,
+					dma_addr_t dma_addr)
+{
+	unsigned long *rto, *sto, *pto;
+	unsigned int rtx, rts, sx, px, idx;
+	struct page *page;
+	gpa_t addr;
+	int i;
+
+	/* Pin guest segment table if needed */
+	rtx = calc_rtx(dma_addr);
+	rto = ioat->head[(rtx / ZPCI_TABLE_ENTRIES_PER_PAGE)];
+	rts = rtx * ZPCI_TABLE_PAGES;
+	if (!ioat->seg[rts]) {
+		if (!reg_entry_isvalid(rto[rtx % ZPCI_TABLE_ENTRIES_PER_PAGE]))
+			return NULL;
+		sto = get_rt_sto(rto[rtx % ZPCI_TABLE_ENTRIES_PER_PAGE]);
+		addr = ((u64)sto & ZPCI_RTE_ADDR_MASK);
+		idx = srcu_read_lock(&vcpu->kvm->srcu);
+		for (i = 0; i < ZPCI_TABLE_PAGES; i++) {
+			page = gfn_to_page(vcpu->kvm, gpa_to_gfn(addr));
+			if (is_error_page(page)) {
+				srcu_read_unlock(&vcpu->kvm->srcu, idx);
+				return NULL;
+			}
+			ioat->seg[rts + i] = page_to_virt(page) +
+					     (addr & ~PAGE_MASK);

Here on the other hand I think the page_to_virt() is correct since you
want the virtual addresses to be able to derference it, correct? 

+			addr += PAGE_SIZE;
+		}
+		srcu_read_unlock(&vcpu->kvm->srcu, idx);
+	}
+
+	/* Allocate pin pointers for another segment table if needed */
+	if (!ioat->pt[rtx]) {
+		ioat->pt[rtx] = kcalloc(ZPCI_TABLE_ENTRIES,
+					(sizeof(unsigned long *)), GFP_KERNEL);
+		if (!ioat->pt[rtx])
+			return NULL;
+	}
+	/* Pin guest page table if needed */
+	sx = calc_sx(dma_addr);
+	sto = ioat->seg[(rts + (sx / ZPCI_TABLE_ENTRIES_PER_PAGE))];
+	if (!ioat->pt[rtx][sx]) {
+		if (!reg_entry_isvalid(sto[sx % ZPCI_TABLE_ENTRIES_PER_PAGE]))
+			return NULL;
+		pto = get_st_pto(sto[sx % ZPCI_TABLE_ENTRIES_PER_PAGE]);
+		if (!pto)
+			return NULL;
+		addr = ((u64)pto & ZPCI_STE_ADDR_MASK);
+		idx = srcu_read_lock(&vcpu->kvm->srcu);
+		page = gfn_to_page(vcpu->kvm, gpa_to_gfn(addr));
+		srcu_read_unlock(&vcpu->kvm->srcu, idx);
+		if (is_error_page(page))
+			return NULL;
+		ioat->pt[rtx][sx] = page_to_virt(page) + (addr & ~PAGE_MASK);

Same as above.

+	}
+	pto = ioat->pt[rtx][sx];
+
+	/* Return guest PTE */
+	px = calc_px(dma_addr);
+	return &pto[px];
+}
+

---8<---