xref: /linux/tools/testing/selftests/kvm/lib/memstress.c (revision 9f2c9170934eace462499ba0bfe042cc72900173)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2020, Google LLC.
4  */
5 #define _GNU_SOURCE
6 
7 #include <inttypes.h>
8 
9 #include "kvm_util.h"
10 #include "memstress.h"
11 #include "processor.h"
12 
13 struct memstress_args memstress_args;
14 
15 /*
16  * Guest virtual memory offset of the testing memory slot.
17  * Must not conflict with identity mapped test code.
18  */
19 static uint64_t guest_test_virt_mem = DEFAULT_GUEST_TEST_MEM;
20 
21 struct vcpu_thread {
22 	/* The index of the vCPU. */
23 	int vcpu_idx;
24 
25 	/* The pthread backing the vCPU. */
26 	pthread_t thread;
27 
28 	/* Set to true once the vCPU thread is up and running. */
29 	bool running;
30 };
31 
32 /* The vCPU threads involved in this test. */
33 static struct vcpu_thread vcpu_threads[KVM_MAX_VCPUS];
34 
35 /* The function run by each vCPU thread, as provided by the test. */
36 static void (*vcpu_thread_fn)(struct memstress_vcpu_args *);
37 
38 /* Set to true once all vCPU threads are up and running. */
39 static bool all_vcpu_threads_running;
40 
41 static struct kvm_vcpu *vcpus[KVM_MAX_VCPUS];
42 
43 /*
44  * Continuously write to the first 8 bytes of each page in the
45  * specified region.
46  */
47 void memstress_guest_code(uint32_t vcpu_idx)
48 {
49 	struct memstress_args *args = &memstress_args;
50 	struct memstress_vcpu_args *vcpu_args = &args->vcpu_args[vcpu_idx];
51 	struct guest_random_state rand_state;
52 	uint64_t gva;
53 	uint64_t pages;
54 	uint64_t addr;
55 	uint64_t page;
56 	int i;
57 
58 	rand_state = new_guest_random_state(args->random_seed + vcpu_idx);
59 
60 	gva = vcpu_args->gva;
61 	pages = vcpu_args->pages;
62 
63 	/* Make sure vCPU args data structure is not corrupt. */
64 	GUEST_ASSERT(vcpu_args->vcpu_idx == vcpu_idx);
65 
66 	while (true) {
67 		for (i = 0; i < pages; i++) {
68 			if (args->random_access)
69 				page = guest_random_u32(&rand_state) % pages;
70 			else
71 				page = i;
72 
73 			addr = gva + (page * args->guest_page_size);
74 
75 			if (guest_random_u32(&rand_state) % 100 < args->write_percent)
76 				*(uint64_t *)addr = 0x0123456789ABCDEF;
77 			else
78 				READ_ONCE(*(uint64_t *)addr);
79 		}
80 
81 		GUEST_SYNC(1);
82 	}
83 }
84 
85 void memstress_setup_vcpus(struct kvm_vm *vm, int nr_vcpus,
86 			   struct kvm_vcpu *vcpus[],
87 			   uint64_t vcpu_memory_bytes,
88 			   bool partition_vcpu_memory_access)
89 {
90 	struct memstress_args *args = &memstress_args;
91 	struct memstress_vcpu_args *vcpu_args;
92 	int i;
93 
94 	for (i = 0; i < nr_vcpus; i++) {
95 		vcpu_args = &args->vcpu_args[i];
96 
97 		vcpu_args->vcpu = vcpus[i];
98 		vcpu_args->vcpu_idx = i;
99 
100 		if (partition_vcpu_memory_access) {
101 			vcpu_args->gva = guest_test_virt_mem +
102 					 (i * vcpu_memory_bytes);
103 			vcpu_args->pages = vcpu_memory_bytes /
104 					   args->guest_page_size;
105 			vcpu_args->gpa = args->gpa + (i * vcpu_memory_bytes);
106 		} else {
107 			vcpu_args->gva = guest_test_virt_mem;
108 			vcpu_args->pages = (nr_vcpus * vcpu_memory_bytes) /
109 					   args->guest_page_size;
110 			vcpu_args->gpa = args->gpa;
111 		}
112 
113 		vcpu_args_set(vcpus[i], 1, i);
114 
115 		pr_debug("Added VCPU %d with test mem gpa [%lx, %lx)\n",
116 			 i, vcpu_args->gpa, vcpu_args->gpa +
117 			 (vcpu_args->pages * args->guest_page_size));
118 	}
119 }
120 
121 struct kvm_vm *memstress_create_vm(enum vm_guest_mode mode, int nr_vcpus,
122 				   uint64_t vcpu_memory_bytes, int slots,
123 				   enum vm_mem_backing_src_type backing_src,
124 				   bool partition_vcpu_memory_access)
125 {
126 	struct memstress_args *args = &memstress_args;
127 	struct kvm_vm *vm;
128 	uint64_t guest_num_pages, slot0_pages = 0;
129 	uint64_t backing_src_pagesz = get_backing_src_pagesz(backing_src);
130 	uint64_t region_end_gfn;
131 	int i;
132 
133 	pr_info("Testing guest mode: %s\n", vm_guest_mode_string(mode));
134 
135 	/* By default vCPUs will write to memory. */
136 	args->write_percent = 100;
137 
138 	/*
139 	 * Snapshot the non-huge page size.  This is used by the guest code to
140 	 * access/dirty pages at the logging granularity.
141 	 */
142 	args->guest_page_size = vm_guest_mode_params[mode].page_size;
143 
144 	guest_num_pages = vm_adjust_num_guest_pages(mode,
145 				(nr_vcpus * vcpu_memory_bytes) / args->guest_page_size);
146 
147 	TEST_ASSERT(vcpu_memory_bytes % getpagesize() == 0,
148 		    "Guest memory size is not host page size aligned.");
149 	TEST_ASSERT(vcpu_memory_bytes % args->guest_page_size == 0,
150 		    "Guest memory size is not guest page size aligned.");
151 	TEST_ASSERT(guest_num_pages % slots == 0,
152 		    "Guest memory cannot be evenly divided into %d slots.",
153 		    slots);
154 
155 	/*
156 	 * If using nested, allocate extra pages for the nested page tables and
157 	 * in-memory data structures.
158 	 */
159 	if (args->nested)
160 		slot0_pages += memstress_nested_pages(nr_vcpus);
161 
162 	/*
163 	 * Pass guest_num_pages to populate the page tables for test memory.
164 	 * The memory is also added to memslot 0, but that's a benign side
165 	 * effect as KVM allows aliasing HVAs in meslots.
166 	 */
167 	vm = __vm_create_with_vcpus(mode, nr_vcpus, slot0_pages + guest_num_pages,
168 				    memstress_guest_code, vcpus);
169 
170 	args->vm = vm;
171 
172 	/* Put the test region at the top guest physical memory. */
173 	region_end_gfn = vm->max_gfn + 1;
174 
175 #ifdef __x86_64__
176 	/*
177 	 * When running vCPUs in L2, restrict the test region to 48 bits to
178 	 * avoid needing 5-level page tables to identity map L2.
179 	 */
180 	if (args->nested)
181 		region_end_gfn = min(region_end_gfn, (1UL << 48) / args->guest_page_size);
182 #endif
183 	/*
184 	 * If there should be more memory in the guest test region than there
185 	 * can be pages in the guest, it will definitely cause problems.
186 	 */
187 	TEST_ASSERT(guest_num_pages < region_end_gfn,
188 		    "Requested more guest memory than address space allows.\n"
189 		    "    guest pages: %" PRIx64 " max gfn: %" PRIx64
190 		    " nr_vcpus: %d wss: %" PRIx64 "]\n",
191 		    guest_num_pages, region_end_gfn - 1, nr_vcpus, vcpu_memory_bytes);
192 
193 	args->gpa = (region_end_gfn - guest_num_pages - 1) * args->guest_page_size;
194 	args->gpa = align_down(args->gpa, backing_src_pagesz);
195 #ifdef __s390x__
196 	/* Align to 1M (segment size) */
197 	args->gpa = align_down(args->gpa, 1 << 20);
198 #endif
199 	args->size = guest_num_pages * args->guest_page_size;
200 	pr_info("guest physical test memory: [0x%lx, 0x%lx)\n",
201 		args->gpa, args->gpa + args->size);
202 
203 	/* Add extra memory slots for testing */
204 	for (i = 0; i < slots; i++) {
205 		uint64_t region_pages = guest_num_pages / slots;
206 		vm_paddr_t region_start = args->gpa + region_pages * args->guest_page_size * i;
207 
208 		vm_userspace_mem_region_add(vm, backing_src, region_start,
209 					    MEMSTRESS_MEM_SLOT_INDEX + i,
210 					    region_pages, 0);
211 	}
212 
213 	/* Do mapping for the demand paging memory slot */
214 	virt_map(vm, guest_test_virt_mem, args->gpa, guest_num_pages);
215 
216 	memstress_setup_vcpus(vm, nr_vcpus, vcpus, vcpu_memory_bytes,
217 			      partition_vcpu_memory_access);
218 
219 	if (args->nested) {
220 		pr_info("Configuring vCPUs to run in L2 (nested).\n");
221 		memstress_setup_nested(vm, nr_vcpus, vcpus);
222 	}
223 
224 	/* Export the shared variables to the guest. */
225 	sync_global_to_guest(vm, memstress_args);
226 
227 	return vm;
228 }
229 
230 void memstress_destroy_vm(struct kvm_vm *vm)
231 {
232 	kvm_vm_free(vm);
233 }
234 
235 void memstress_set_write_percent(struct kvm_vm *vm, uint32_t write_percent)
236 {
237 	memstress_args.write_percent = write_percent;
238 	sync_global_to_guest(vm, memstress_args.write_percent);
239 }
240 
241 void memstress_set_random_seed(struct kvm_vm *vm, uint32_t random_seed)
242 {
243 	memstress_args.random_seed = random_seed;
244 	sync_global_to_guest(vm, memstress_args.random_seed);
245 }
246 
247 void memstress_set_random_access(struct kvm_vm *vm, bool random_access)
248 {
249 	memstress_args.random_access = random_access;
250 	sync_global_to_guest(vm, memstress_args.random_access);
251 }
252 
253 uint64_t __weak memstress_nested_pages(int nr_vcpus)
254 {
255 	return 0;
256 }
257 
258 void __weak memstress_setup_nested(struct kvm_vm *vm, int nr_vcpus, struct kvm_vcpu **vcpus)
259 {
260 	pr_info("%s() not support on this architecture, skipping.\n", __func__);
261 	exit(KSFT_SKIP);
262 }
263 
264 static void *vcpu_thread_main(void *data)
265 {
266 	struct vcpu_thread *vcpu = data;
267 	int vcpu_idx = vcpu->vcpu_idx;
268 
269 	if (memstress_args.pin_vcpus)
270 		kvm_pin_this_task_to_pcpu(memstress_args.vcpu_to_pcpu[vcpu_idx]);
271 
272 	WRITE_ONCE(vcpu->running, true);
273 
274 	/*
275 	 * Wait for all vCPU threads to be up and running before calling the test-
276 	 * provided vCPU thread function. This prevents thread creation (which
277 	 * requires taking the mmap_sem in write mode) from interfering with the
278 	 * guest faulting in its memory.
279 	 */
280 	while (!READ_ONCE(all_vcpu_threads_running))
281 		;
282 
283 	vcpu_thread_fn(&memstress_args.vcpu_args[vcpu_idx]);
284 
285 	return NULL;
286 }
287 
288 void memstress_start_vcpu_threads(int nr_vcpus,
289 				  void (*vcpu_fn)(struct memstress_vcpu_args *))
290 {
291 	int i;
292 
293 	vcpu_thread_fn = vcpu_fn;
294 	WRITE_ONCE(all_vcpu_threads_running, false);
295 	WRITE_ONCE(memstress_args.stop_vcpus, false);
296 
297 	for (i = 0; i < nr_vcpus; i++) {
298 		struct vcpu_thread *vcpu = &vcpu_threads[i];
299 
300 		vcpu->vcpu_idx = i;
301 		WRITE_ONCE(vcpu->running, false);
302 
303 		pthread_create(&vcpu->thread, NULL, vcpu_thread_main, vcpu);
304 	}
305 
306 	for (i = 0; i < nr_vcpus; i++) {
307 		while (!READ_ONCE(vcpu_threads[i].running))
308 			;
309 	}
310 
311 	WRITE_ONCE(all_vcpu_threads_running, true);
312 }
313 
314 void memstress_join_vcpu_threads(int nr_vcpus)
315 {
316 	int i;
317 
318 	WRITE_ONCE(memstress_args.stop_vcpus, true);
319 
320 	for (i = 0; i < nr_vcpus; i++)
321 		pthread_join(vcpu_threads[i].thread, NULL);
322 }
323