FROMLIST: virt: gunyah: Enable demand paging

Tell resource manager to enable demand paging and wire vCPU faults to
provide the backing folio when a guestmemfd is bound to the faulting
access.

Bug: 338347082
Link: https://lore.kernel.org/all/20240222-gunyah-v17-28-1e9da6763d38@quicinc.com/
Change-Id: I4e6decdc34d8afc90e7c95e91b30526cf59d0914
Signed-off-by: Elliot Berman <quic_eberman@quicinc.com>
Signed-off-by: Elliot Berman <elliot.berman@oss.qualcomm.com>

drivers/virt/gunyah/guest_memfd.c

@@ -891,3 +891,97 @@ int gunyah_gmem_reclaim_parcel(struct gunyah_vm *ghvm,
 
 	return 0;
 }
+
+int gunyah_gmem_setup_demand_paging(struct gunyah_vm *ghvm)
+{
+	struct gunyah_rm_mem_entry *entries;
+	struct gunyah_gmem_binding *b;
+	unsigned long index = 0;
+	u32 count = 0, i;
+	int ret = 0;
+
+	down_read(&ghvm->bindings_lock);
+	mt_for_each(&ghvm->bindings, b, index, ULONG_MAX)
+		if (gunyah_guest_mem_is_lend(ghvm, b->flags))
+			count++;
+
+	if (!count)
+		goto out;
+
+	entries = kcalloc(count, sizeof(*entries), GFP_KERNEL);
+	if (!entries) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	index = i = 0;
+	mt_for_each(&ghvm->bindings, b, index, ULONG_MAX) {
+		if (!gunyah_guest_mem_is_lend(ghvm, b->flags))
+			continue;
+		entries[i].phys_addr = cpu_to_le64(gunyah_gfn_to_gpa(b->gfn));
+		entries[i].size = cpu_to_le64(b->nr << PAGE_SHIFT);
+		if (++i == count)
+			break;
+	}
+
+	ret = gunyah_rm_vm_set_demand_paging(ghvm->rm, ghvm->vmid, i, entries);
+	kfree(entries);
+out:
+	up_read(&ghvm->bindings_lock);
+	return ret;
+}
+
+int gunyah_gmem_demand_page(struct gunyah_vm *ghvm, u64 gpa, bool write)
+{
+	unsigned long gfn = gunyah_gpa_to_gfn(gpa);
+	struct gunyah_gmem_binding *b;
+	struct folio *folio;
+	int ret;
+
+	down_read(&ghvm->bindings_lock);
+	b = mtree_load(&ghvm->bindings, gfn);
+	if (!b) {
+		ret = -ENOENT;
+		goto unlock;
+	}
+
+	if (write && !(b->flags & GUNYAH_MEM_ALLOW_WRITE)) {
+		ret = -EPERM;
+		goto unlock;
+	}
+
+	folio = gunyah_gmem_get_folio(file_inode(b->file),
+				      gunyah_gfn_to_off(b, gfn));
+	if (!folio) {
+		ret = -ENOMEM;
+		pr_err_ratelimited("Failed to obtain memory for guest addr %016llx\n", gpa);
+		goto unlock;
+	}
+
+	/**
+	 * the folio covers the requested guest address, but the folio may not
+	 * start at the requested guest address. recompute the gfn based on the
+	 * folio itself.
+	 */
+	gfn = gunyah_off_to_gfn(b, folio_index(folio));
+
+	filemap_invalidate_lock_shared(b->file->f_mapping);
+	ret = gunyah_vm_provide_folio(ghvm, folio, gfn,
+				      !gunyah_guest_mem_is_lend(ghvm, b->flags),
+				      !!(b->flags & GUNYAH_MEM_ALLOW_WRITE));
+	filemap_invalidate_unlock_shared(b->file->f_mapping);
+	if (ret) {
+		if (ret != -EAGAIN)
+			pr_err_ratelimited(
+				"Failed to provide folio for guest addr: %016llx: %d\n",
+				gpa, ret);
+		goto out;
+	}
+out:
+	folio_unlock(folio);
+	folio_put(folio);
+unlock:
+	up_read(&ghvm->bindings_lock);
+	return ret;
+}
+EXPORT_SYMBOL_GPL(gunyah_gmem_demand_page);

drivers/virt/gunyah/gunyah_vcpu.c

@@ -89,29 +89,44 @@ static irqreturn_t gunyah_vcpu_irq_handler(int irq, void *data)
 	return IRQ_HANDLED;
 }
 
-static void gunyah_handle_page_fault(
+static bool gunyah_handle_page_fault(
 	struct gunyah_vcpu *vcpu,
 	const struct gunyah_hypercall_vcpu_run_resp *vcpu_run_resp)
 {
 	u64 addr = vcpu_run_resp->state_data[0];
+	bool write = !!vcpu_run_resp->state_data[1];
+	int ret = 0;
+
+	ret = gunyah_gmem_demand_page(vcpu->ghvm, addr, write);
+	if (!ret || ret == -EAGAIN)
+		return true;
 
 	vcpu->vcpu_run->page_fault.resume_action = GUNYAH_VCPU_RESUME_FAULT;
-	vcpu->vcpu_run->page_fault.attempt = 0;
+	vcpu->vcpu_run->page_fault.attempt = ret;
 	vcpu->vcpu_run->page_fault.phys_addr = addr;
 	vcpu->vcpu_run->exit_reason = GUNYAH_VCPU_EXIT_PAGE_FAULT;
+	return false;
 }
 
-static void
-gunyah_handle_mmio(struct gunyah_vcpu *vcpu,
+static bool
+gunyah_handle_mmio(struct gunyah_vcpu *vcpu, unsigned long resume_data[3],
 		   const struct gunyah_hypercall_vcpu_run_resp *vcpu_run_resp)
 {
 	u64 addr = vcpu_run_resp->state_data[0],
 	    len = vcpu_run_resp->state_data[1],
 	    data = vcpu_run_resp->state_data[2];
+	int ret;
 
 	if (WARN_ON(len > sizeof(u64)))
 		len = sizeof(u64);
 
+	ret = gunyah_gmem_demand_page(vcpu->ghvm, addr,
+				      vcpu->vcpu_run->mmio.is_write);
+	if (!ret || ret == -EAGAIN) {
+		resume_data[1] = GUNYAH_ADDRSPACE_VMMIO_ACTION_RETRY;
+		return true;
+	}
+
 	if (vcpu_run_resp->state == GUNYAH_VCPU_ADDRSPACE_VMMIO_READ) {
 		vcpu->vcpu_run->mmio.is_write = 0;
 		/* Record that we need to give vCPU user's supplied value next gunyah_vcpu_run() */
@@ -128,11 +143,15 @@ gunyah_handle_mmio(struct gunyah_vcpu *vcpu,
 	vcpu->mmio_addr = vcpu->vcpu_run->mmio.phys_addr = addr;
 	vcpu->vcpu_run->mmio.len = len;
 	vcpu->vcpu_run->exit_reason = GUNYAH_VCPU_EXIT_MMIO;
+
+	return false;
 }
 
 static int gunyah_handle_mmio_resume(struct gunyah_vcpu *vcpu,
 				     unsigned long resume_data[3])
 {
+	bool write = vcpu->state == GUNYAH_VCPU_RUN_STATE_MMIO_WRITE;
+
 	switch (vcpu->vcpu_run->mmio.resume_action) {
 	case GUNYAH_VCPU_RESUME_HANDLED:
 		if (vcpu->state == GUNYAH_VCPU_RUN_STATE_MMIO_READ) {
@@ -148,6 +167,8 @@ static int gunyah_handle_mmio_resume(struct gunyah_vcpu *vcpu,
 		resume_data[1] = GUNYAH_ADDRSPACE_VMMIO_ACTION_FAULT;
 		break;
 	case GUNYAH_VCPU_RESUME_RETRY:
+		/* userspace probably added a memory binding */
+		gunyah_gmem_demand_page(vcpu->ghvm, vcpu->mmio_addr, write);
 		resume_data[1] = GUNYAH_ADDRSPACE_VMMIO_ACTION_RETRY;
 		break;
 	default:
@@ -310,11 +331,15 @@ static int gunyah_vcpu_run(struct gunyah_vcpu *vcpu)
 				break;
 			case GUNYAH_VCPU_ADDRSPACE_VMMIO_READ:
 			case GUNYAH_VCPU_ADDRSPACE_VMMIO_WRITE:
-				gunyah_handle_mmio(vcpu, &vcpu_run_resp);
-				goto out;
+				if (!gunyah_handle_mmio(vcpu, resume_data,
+							&vcpu_run_resp))
+					goto out;
+				break;
 			case GUNYAH_VCPU_ADDRSPACE_PAGE_FAULT:
-				gunyah_handle_page_fault(vcpu, &vcpu_run_resp);
-				goto out;
+				if (!gunyah_handle_page_fault(vcpu,
+							      &vcpu_run_resp))
+					goto out;
+				break;
 			default:
 				pr_warn_ratelimited(
 					"Unknown vCPU state: %llx\n",

drivers/virt/gunyah/vm_mgr.c

@@ -474,6 +474,23 @@ static int gunyah_vm_start(struct gunyah_vm *ghvm)
 		goto err;
 	}
 
+	ret = gunyah_gmem_setup_demand_paging(ghvm);
+	if (ret) {
+		dev_warn(ghvm->parent,
+			 "Failed to set up gmem demand paging: %d\n", ret);
+		goto err;
+	}
+
+	ret = gunyah_rm_vm_set_address_layout(
+		ghvm->rm, ghvm->vmid, GUNYAH_RM_RANGE_ID_IMAGE,
+		ghvm->dtb.parcel_start << PAGE_SHIFT,
+		ghvm->dtb.parcel_pages << PAGE_SHIFT);
+	if (ret) {
+		dev_warn(ghvm->parent,
+			 "Failed to set location of DTB mem parcel: %d\n", ret);
+		goto err;
+	}
+
 	ret = gunyah_rm_vm_init(ghvm->rm, ghvm->vmid);
 	if (ret) {
 		ghvm->vm_status = GUNYAH_RM_VM_STATUS_INIT_FAILED;
@@ -501,8 +518,20 @@ static int gunyah_vm_start(struct gunyah_vm *ghvm)
 		gunyah_vm_add_resource(ghvm, ghrsc);
 	}
 
+	ret = gunyah_vm_parcel_to_paged(ghvm, &ghvm->dtb.parcel,
+					ghvm->dtb.parcel_start,
+					ghvm->dtb.parcel_pages);
+	if (ret)
+		goto err;
+
 	ret = gunyah_rm_vm_start(ghvm->rm, ghvm->vmid);
 	if (ret) {
+		/**
+		 * need to rollback parcel_to_paged because RM is still
+		 * tracking the parcel
+		 */
+		gunyah_vm_mm_erase_range(ghvm, ghvm->dtb.parcel_start,
+					 ghvm->dtb.parcel_pages);
 		dev_warn(ghvm->parent, "Failed to start VM: %d\n", ret);
 		goto err;
 	}

drivers/virt/gunyah/vm_mgr.h

@@ -199,4 +199,7 @@ int gunyah_gmem_reclaim_parcel(struct gunyah_vm *ghvm,
 			       struct gunyah_rm_mem_parcel *parcel, u64 gfn,
 			       u64 nr);
 
+int gunyah_gmem_setup_demand_paging(struct gunyah_vm *ghvm);
+int gunyah_gmem_demand_page(struct gunyah_vm *ghvm, u64 gpa, bool write);
+
 #endif