pre_fault_memory_test.c

  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Copyright (C) 2024, Intel, Inc
  4 *
  5 * Author:
  6 * Isaku Yamahata <isaku.yamahata at gmail.com>
  7 */
  8#include <linux/sizes.h>
  9
 10#include <test_util.h>
 11#include <kvm_util.h>
 12#include <processor.h>
 13#include <pthread.h>
 14
 15/* Arbitrarily chosen values */
 16#define TEST_SIZE		(SZ_2M + PAGE_SIZE)
 17#define TEST_NPAGES		(TEST_SIZE / PAGE_SIZE)
 18#define TEST_SLOT		10
 19
 20static void guest_code(uint64_t base_gva)
 21{
 22	volatile uint64_t val __used;
 23	int i;
 24
 25	for (i = 0; i < TEST_NPAGES; i++) {
 26		uint64_t *src = (uint64_t *)(base_gva + i * PAGE_SIZE);
 27
 28		val = *src;
 29	}
 30
 31	GUEST_DONE();
 32}
 33
 34struct slot_worker_data {
 35	struct kvm_vm *vm;
 36	u64 gpa;
 37	uint32_t flags;
 38	bool worker_ready;
 39	bool prefault_ready;
 40	bool recreate_slot;
 41};
 42
 43static void *delete_slot_worker(void *__data)
 44{
 45	struct slot_worker_data *data = __data;
 46	struct kvm_vm *vm = data->vm;
 47
 48	WRITE_ONCE(data->worker_ready, true);
 49
 50	while (!READ_ONCE(data->prefault_ready))
 51		cpu_relax();
 52
 53	vm_mem_region_delete(vm, TEST_SLOT);
 54
 55	while (!READ_ONCE(data->recreate_slot))
 56		cpu_relax();
 57
 58	vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS, data->gpa,
 59				    TEST_SLOT, TEST_NPAGES, data->flags);
 60
 61	return NULL;
 62}
 63
 64static void pre_fault_memory(struct kvm_vcpu *vcpu, u64 base_gpa, u64 offset,
 65			     u64 size, u64 expected_left, bool private)
 66{
 67	struct kvm_pre_fault_memory range = {
 68		.gpa = base_gpa + offset,
 69		.size = size,
 70		.flags = 0,
 71	};
 72	struct slot_worker_data data = {
 73		.vm = vcpu->vm,
 74		.gpa = base_gpa,
 75		.flags = private ? KVM_MEM_GUEST_MEMFD : 0,
 76	};
 77	bool slot_recreated = false;
 78	pthread_t slot_worker;
 79	int ret, save_errno;
 80	u64 prev;
 81
 82	/*
 83	 * Concurrently delete (and recreate) the slot to test KVM's handling
 84	 * of a racing memslot deletion with prefaulting.
 85	 */
 86	pthread_create(&slot_worker, NULL, delete_slot_worker, &data);
 87
 88	while (!READ_ONCE(data.worker_ready))
 89		cpu_relax();
 90
 91	WRITE_ONCE(data.prefault_ready, true);
 92
 93	for (;;) {
 94		prev = range.size;
 95		ret = __vcpu_ioctl(vcpu, KVM_PRE_FAULT_MEMORY, &range);
 96		save_errno = errno;
 97		TEST_ASSERT((range.size < prev) ^ (ret < 0),
 98			    "%sexpecting range.size to change on %s",
 99			    ret < 0 ? "not " : "",
100			    ret < 0 ? "failure" : "success");
101
102		/*
103		 * Immediately retry prefaulting if KVM was interrupted by an
104		 * unrelated signal/event.
105		 */
106		if (ret < 0 && save_errno == EINTR)
107			continue;
108
109		/*
110		 * Tell the worker to recreate the slot in order to complete
111		 * prefaulting (if prefault didn't already succeed before the
112		 * slot was deleted) and/or to prepare for the next testcase.
113		 * Wait for the worker to exit so that the next invocation of
114		 * prefaulting is guaranteed to complete (assuming no KVM bugs).
115		 */
116		if (!slot_recreated) {
117			WRITE_ONCE(data.recreate_slot, true);
118			pthread_join(slot_worker, NULL);
119			slot_recreated = true;
120
121			/*
122			 * Retry prefaulting to get a stable result, i.e. to
123			 * avoid seeing random EAGAIN failures.  Don't retry if
124			 * prefaulting already succeeded, as KVM disallows
125			 * prefaulting with size=0, i.e. blindly retrying would
126			 * result in test failures due to EINVAL.  KVM should
127			 * always return success if all bytes are prefaulted,
128			 * i.e. there is no need to guard against EAGAIN being
129			 * returned.
130			 */
131			if (range.size)
132				continue;
133		}
134
135		/*
136		 * All done if there are no remaining bytes to prefault, or if
137		 * prefaulting failed (EINTR was handled above, and EAGAIN due
138		 * to prefaulting a memslot that's being actively deleted should
139		 * be impossible since the memslot has already been recreated).
140		 */
141		if (!range.size || ret < 0)
142			break;
143	}
144
145	TEST_ASSERT(range.size == expected_left,
146		    "Completed with %llu bytes left, expected %lu",
147		    range.size, expected_left);
148
149	/*
150	 * Assert success if prefaulting the entire range should succeed, i.e.
151	 * complete with no bytes remaining.  Otherwise prefaulting should have
152	 * failed due to ENOENT (due to RET_PF_EMULATE for emulated MMIO when
153	 * no memslot exists).
154	 */
155	if (!expected_left)
156		TEST_ASSERT_VM_VCPU_IOCTL(!ret, KVM_PRE_FAULT_MEMORY, ret, vcpu->vm);
157	else
158		TEST_ASSERT_VM_VCPU_IOCTL(ret && save_errno == ENOENT,
159					  KVM_PRE_FAULT_MEMORY, ret, vcpu->vm);
160}
161
162static void __test_pre_fault_memory(unsigned long vm_type, bool private)
163{
164	uint64_t gpa, gva, alignment, guest_page_size;
165	const struct vm_shape shape = {
166		.mode = VM_MODE_DEFAULT,
167		.type = vm_type,
168	};
169	struct kvm_vcpu *vcpu;
170	struct kvm_run *run;
171	struct kvm_vm *vm;
172	struct ucall uc;
173
174	vm = vm_create_shape_with_one_vcpu(shape, &vcpu, guest_code);
175
176	alignment = guest_page_size = vm_guest_mode_params[VM_MODE_DEFAULT].page_size;
177	gpa = (vm->max_gfn - TEST_NPAGES) * guest_page_size;
178#ifdef __s390x__
179	alignment = max(0x100000UL, guest_page_size);
180#else
181	alignment = SZ_2M;
182#endif
183	gpa = align_down(gpa, alignment);
184	gva = gpa & ((1ULL << (vm->va_bits - 1)) - 1);
185
186	vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS, gpa, TEST_SLOT,
187				    TEST_NPAGES, private ? KVM_MEM_GUEST_MEMFD : 0);
188	virt_map(vm, gva, gpa, TEST_NPAGES);
189
190	if (private)
191		vm_mem_set_private(vm, gpa, TEST_SIZE);
192
193	pre_fault_memory(vcpu, gpa, 0, SZ_2M, 0, private);
194	pre_fault_memory(vcpu, gpa, SZ_2M, PAGE_SIZE * 2, PAGE_SIZE, private);
195	pre_fault_memory(vcpu, gpa, TEST_SIZE, PAGE_SIZE, PAGE_SIZE, private);
196
197	vcpu_args_set(vcpu, 1, gva);
198	vcpu_run(vcpu);
199
200	run = vcpu->run;
201	TEST_ASSERT(run->exit_reason == KVM_EXIT_IO,
202		    "Wanted KVM_EXIT_IO, got exit reason: %u (%s)",
203		    run->exit_reason, exit_reason_str(run->exit_reason));
204
205	switch (get_ucall(vcpu, &uc)) {
206	case UCALL_ABORT:
207		REPORT_GUEST_ASSERT(uc);
208		break;
209	case UCALL_DONE:
210		break;
211	default:
212		TEST_FAIL("Unknown ucall 0x%lx.", uc.cmd);
213		break;
214	}
215
216	kvm_vm_free(vm);
217}
218
219static void test_pre_fault_memory(unsigned long vm_type, bool private)
220{
221	if (vm_type && !(kvm_check_cap(KVM_CAP_VM_TYPES) & BIT(vm_type))) {
222		pr_info("Skipping tests for vm_type 0x%lx\n", vm_type);
223		return;
224	}
225
226	__test_pre_fault_memory(vm_type, private);
227}
228
229int main(int argc, char *argv[])
230{
231	TEST_REQUIRE(kvm_check_cap(KVM_CAP_PRE_FAULT_MEMORY));
232
233	test_pre_fault_memory(0, false);
234#ifdef __x86_64__
235	test_pre_fault_memory(KVM_X86_SW_PROTECTED_VM, false);
236	test_pre_fault_memory(KVM_X86_SW_PROTECTED_VM, true);
237#endif
238	return 0;
239}