1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * amx tests
4  *
5  * Copyright (C) 2021, Intel, Inc.
6  *
7  * Tests for amx #NM exception and save/restore.
8  */
9 
10 #define _GNU_SOURCE /* for program_invocation_short_name */
11 #include <fcntl.h>
12 #include <stdio.h>
13 #include <stdlib.h>
14 #include <string.h>
15 #include <sys/ioctl.h>
16 #include <sys/syscall.h>
17 
18 #include "test_util.h"
19 
20 #include "kvm_util.h"
21 #include "processor.h"
22 #include "vmx.h"
23 
24 #ifndef __x86_64__
25 # error This test is 64-bit only
26 #endif
27 
28 #define NUM_TILES			8
29 #define TILE_SIZE			1024
30 #define XSAVE_SIZE			((NUM_TILES * TILE_SIZE) + PAGE_SIZE)
31 
32 /* Tile configuration associated: */
33 #define MAX_TILES			16
34 #define RESERVED_BYTES			14
35 
36 #define XFEATURE_XTILECFG		17
37 #define XFEATURE_XTILEDATA		18
38 #define XFEATURE_MASK_XTILECFG		(1 << XFEATURE_XTILECFG)
39 #define XFEATURE_MASK_XTILEDATA		(1 << XFEATURE_XTILEDATA)
40 #define XFEATURE_MASK_XTILE		(XFEATURE_MASK_XTILECFG | XFEATURE_MASK_XTILEDATA)
41 
42 #define XSAVE_HDR_OFFSET		512
43 
44 struct xsave_data {
45 	u8 area[XSAVE_SIZE];
46 } __aligned(64);
47 
48 struct tile_config {
49 	u8  palette_id;
50 	u8  start_row;
51 	u8  reserved[RESERVED_BYTES];
52 	u16 colsb[MAX_TILES];
53 	u8  rows[MAX_TILES];
54 };
55 
56 struct tile_data {
57 	u8 data[NUM_TILES * TILE_SIZE];
58 };
59 
60 struct xtile_info {
61 	u16 bytes_per_tile;
62 	u16 bytes_per_row;
63 	u16 max_names;
64 	u16 max_rows;
65 	u32 xsave_offset;
66 	u32 xsave_size;
67 };
68 
69 static struct xtile_info xtile;
70 
71 static inline u64 __xgetbv(u32 index)
72 {
73 	u32 eax, edx;
74 
75 	asm volatile("xgetbv;"
76 		     : "=a" (eax), "=d" (edx)
77 		     : "c" (index));
78 	return eax + ((u64)edx << 32);
79 }
80 
81 static inline void __xsetbv(u32 index, u64 value)
82 {
83 	u32 eax = value;
84 	u32 edx = value >> 32;
85 
86 	asm volatile("xsetbv" :: "a" (eax), "d" (edx), "c" (index));
87 }
88 
89 static inline void __ldtilecfg(void *cfg)
90 {
91 	asm volatile(".byte 0xc4,0xe2,0x78,0x49,0x00"
92 		     : : "a"(cfg));
93 }
94 
95 static inline void __tileloadd(void *tile)
96 {
97 	asm volatile(".byte 0xc4,0xe2,0x7b,0x4b,0x04,0x10"
98 		     : : "a"(tile), "d"(0));
99 }
100 
101 static inline void __tilerelease(void)
102 {
103 	asm volatile(".byte 0xc4, 0xe2, 0x78, 0x49, 0xc0" ::);
104 }
105 
106 static inline void __xsavec(struct xsave_data *data, uint64_t rfbm)
107 {
108 	uint32_t rfbm_lo = rfbm;
109 	uint32_t rfbm_hi = rfbm >> 32;
110 
111 	asm volatile("xsavec (%%rdi)"
112 		     : : "D" (data), "a" (rfbm_lo), "d" (rfbm_hi)
113 		     : "memory");
114 }
115 
116 static inline void check_cpuid_xsave(void)
117 {
118 	GUEST_ASSERT(this_cpu_has(X86_FEATURE_XSAVE));
119 	GUEST_ASSERT(this_cpu_has(X86_FEATURE_OSXSAVE));
120 }
121 
122 static bool check_xsave_supports_xtile(void)
123 {
124 	return __xgetbv(0) & XFEATURE_MASK_XTILE;
125 }
126 
127 static void check_xtile_info(void)
128 {
129 	GUEST_ASSERT(this_cpu_has_p(X86_PROPERTY_XSTATE_MAX_SIZE_XCR0));
130 	GUEST_ASSERT(this_cpu_property(X86_PROPERTY_XSTATE_MAX_SIZE_XCR0) <= XSAVE_SIZE);
131 
132 	xtile.xsave_offset = this_cpu_property(X86_PROPERTY_XSTATE_TILE_OFFSET);
133 	GUEST_ASSERT(xtile.xsave_offset == 2816);
134 	xtile.xsave_size = this_cpu_property(X86_PROPERTY_XSTATE_TILE_SIZE);
135 	GUEST_ASSERT(xtile.xsave_size == 8192);
136 	GUEST_ASSERT(sizeof(struct tile_data) >= xtile.xsave_size);
137 
138 	GUEST_ASSERT(this_cpu_has_p(X86_PROPERTY_AMX_NR_TILE_REGS));
139 	xtile.max_names = this_cpu_property(X86_PROPERTY_AMX_NR_TILE_REGS);
140 	GUEST_ASSERT(xtile.max_names == 8);
141 	xtile.bytes_per_tile = this_cpu_property(X86_PROPERTY_AMX_BYTES_PER_TILE);
142 	GUEST_ASSERT(xtile.bytes_per_tile == 1024);
143 	xtile.bytes_per_row = this_cpu_property(X86_PROPERTY_AMX_BYTES_PER_ROW);
144 	GUEST_ASSERT(xtile.bytes_per_row == 64);
145 	xtile.max_rows = this_cpu_property(X86_PROPERTY_AMX_MAX_ROWS);
146 	GUEST_ASSERT(xtile.max_rows == 16);
147 }
148 
149 static void set_tilecfg(struct tile_config *cfg)
150 {
151 	int i;
152 
153 	/* Only palette id 1 */
154 	cfg->palette_id = 1;
155 	for (i = 0; i < xtile.max_names; i++) {
156 		cfg->colsb[i] = xtile.bytes_per_row;
157 		cfg->rows[i] = xtile.max_rows;
158 	}
159 }
160 
161 static void set_xstatebv(void *data, uint64_t bv)
162 {
163 	*(uint64_t *)(data + XSAVE_HDR_OFFSET) = bv;
164 }
165 
166 static u64 get_xstatebv(void *data)
167 {
168 	return *(u64 *)(data + XSAVE_HDR_OFFSET);
169 }
170 
171 static void init_regs(void)
172 {
173 	uint64_t cr4, xcr0;
174 
175 	/* turn on CR4.OSXSAVE */
176 	cr4 = get_cr4();
177 	cr4 |= X86_CR4_OSXSAVE;
178 	set_cr4(cr4);
179 
180 	xcr0 = __xgetbv(0);
181 	xcr0 |= XFEATURE_MASK_XTILE;
182 	__xsetbv(0x0, xcr0);
183 }
184 
185 static void __attribute__((__flatten__)) guest_code(struct tile_config *amx_cfg,
186 						    struct tile_data *tiledata,
187 						    struct xsave_data *xsave_data)
188 {
189 	init_regs();
190 	check_cpuid_xsave();
191 	check_xsave_supports_xtile();
192 	check_xtile_info();
193 	GUEST_SYNC(1);
194 
195 	/* xfd=0, enable amx */
196 	wrmsr(MSR_IA32_XFD, 0);
197 	GUEST_SYNC(2);
198 	GUEST_ASSERT(rdmsr(MSR_IA32_XFD) == 0);
199 	set_tilecfg(amx_cfg);
200 	__ldtilecfg(amx_cfg);
201 	GUEST_SYNC(3);
202 	/* Check save/restore when trap to userspace */
203 	__tileloadd(tiledata);
204 	GUEST_SYNC(4);
205 	__tilerelease();
206 	GUEST_SYNC(5);
207 	/* bit 18 not in the XCOMP_BV after xsavec() */
208 	set_xstatebv(xsave_data, XFEATURE_MASK_XTILEDATA);
209 	__xsavec(xsave_data, XFEATURE_MASK_XTILEDATA);
210 	GUEST_ASSERT((get_xstatebv(xsave_data) & XFEATURE_MASK_XTILEDATA) == 0);
211 
212 	/* xfd=0x40000, disable amx tiledata */
213 	wrmsr(MSR_IA32_XFD, XFEATURE_MASK_XTILEDATA);
214 	GUEST_SYNC(6);
215 	GUEST_ASSERT(rdmsr(MSR_IA32_XFD) == XFEATURE_MASK_XTILEDATA);
216 	set_tilecfg(amx_cfg);
217 	__ldtilecfg(amx_cfg);
218 	/* Trigger #NM exception */
219 	__tileloadd(tiledata);
220 	GUEST_SYNC(10);
221 
222 	GUEST_DONE();
223 }
224 
225 void guest_nm_handler(struct ex_regs *regs)
226 {
227 	/* Check if #NM is triggered by XFEATURE_MASK_XTILEDATA */
228 	GUEST_SYNC(7);
229 	GUEST_ASSERT(rdmsr(MSR_IA32_XFD_ERR) == XFEATURE_MASK_XTILEDATA);
230 	GUEST_SYNC(8);
231 	GUEST_ASSERT(rdmsr(MSR_IA32_XFD_ERR) == XFEATURE_MASK_XTILEDATA);
232 	/* Clear xfd_err */
233 	wrmsr(MSR_IA32_XFD_ERR, 0);
234 	/* xfd=0, enable amx */
235 	wrmsr(MSR_IA32_XFD, 0);
236 	GUEST_SYNC(9);
237 }
238 
239 int main(int argc, char *argv[])
240 {
241 	struct kvm_regs regs1, regs2;
242 	struct kvm_vcpu *vcpu;
243 	struct kvm_vm *vm;
244 	struct kvm_run *run;
245 	struct kvm_x86_state *state;
246 	int xsave_restore_size;
247 	vm_vaddr_t amx_cfg, tiledata, xsavedata;
248 	struct ucall uc;
249 	u32 amx_offset;
250 	int stage, ret;
251 
252 	/*
253 	 * Note, all off-by-default features must be enabled before anything
254 	 * caches KVM_GET_SUPPORTED_CPUID, e.g. before using kvm_cpu_has().
255 	 */
256 	vm_xsave_require_permission(XSTATE_XTILE_DATA_BIT);
257 
258 	TEST_REQUIRE(kvm_cpu_has(X86_FEATURE_XFD));
259 	TEST_REQUIRE(kvm_cpu_has(X86_FEATURE_XSAVE));
260 	TEST_REQUIRE(kvm_cpu_has(X86_FEATURE_AMX_TILE));
261 	TEST_REQUIRE(kvm_cpu_has(X86_FEATURE_XTILECFG));
262 	TEST_REQUIRE(kvm_cpu_has(X86_FEATURE_XTILEDATA));
263 
264 	/* Create VM */
265 	vm = vm_create_with_one_vcpu(&vcpu, guest_code);
266 
267 	TEST_ASSERT(kvm_cpu_has_p(X86_PROPERTY_XSTATE_MAX_SIZE),
268 		    "KVM should enumerate max XSAVE size when XSAVE is supported");
269 	xsave_restore_size = kvm_cpu_property(X86_PROPERTY_XSTATE_MAX_SIZE);
270 
271 	run = vcpu->run;
272 	vcpu_regs_get(vcpu, &regs1);
273 
274 	/* Register #NM handler */
275 	vm_init_descriptor_tables(vm);
276 	vcpu_init_descriptor_tables(vcpu);
277 	vm_install_exception_handler(vm, NM_VECTOR, guest_nm_handler);
278 
279 	/* amx cfg for guest_code */
280 	amx_cfg = vm_vaddr_alloc_page(vm);
281 	memset(addr_gva2hva(vm, amx_cfg), 0x0, getpagesize());
282 
283 	/* amx tiledata for guest_code */
284 	tiledata = vm_vaddr_alloc_pages(vm, 2);
285 	memset(addr_gva2hva(vm, tiledata), rand() | 1, 2 * getpagesize());
286 
287 	/* xsave data for guest_code */
288 	xsavedata = vm_vaddr_alloc_pages(vm, 3);
289 	memset(addr_gva2hva(vm, xsavedata), 0, 3 * getpagesize());
290 	vcpu_args_set(vcpu, 3, amx_cfg, tiledata, xsavedata);
291 
292 	for (stage = 1; ; stage++) {
293 		vcpu_run(vcpu);
294 		TEST_ASSERT(run->exit_reason == KVM_EXIT_IO,
295 			    "Stage %d: unexpected exit reason: %u (%s),\n",
296 			    stage, run->exit_reason,
297 			    exit_reason_str(run->exit_reason));
298 
299 		switch (get_ucall(vcpu, &uc)) {
300 		case UCALL_ABORT:
301 			REPORT_GUEST_ASSERT(uc);
302 			/* NOT REACHED */
303 		case UCALL_SYNC:
304 			switch (uc.args[1]) {
305 			case 1:
306 			case 2:
307 			case 3:
308 			case 5:
309 			case 6:
310 			case 7:
311 			case 8:
312 				fprintf(stderr, "GUEST_SYNC(%ld)\n", uc.args[1]);
313 				break;
314 			case 4:
315 			case 10:
316 				fprintf(stderr,
317 				"GUEST_SYNC(%ld), check save/restore status\n", uc.args[1]);
318 
319 				/* Compacted mode, get amx offset by xsave area
320 				 * size subtract 8K amx size.
321 				 */
322 				amx_offset = xsave_restore_size - NUM_TILES*TILE_SIZE;
323 				state = vcpu_save_state(vcpu);
324 				void *amx_start = (void *)state->xsave + amx_offset;
325 				void *tiles_data = (void *)addr_gva2hva(vm, tiledata);
326 				/* Only check TMM0 register, 1 tile */
327 				ret = memcmp(amx_start, tiles_data, TILE_SIZE);
328 				TEST_ASSERT(ret == 0, "memcmp failed, ret=%d\n", ret);
329 				kvm_x86_state_cleanup(state);
330 				break;
331 			case 9:
332 				fprintf(stderr,
333 				"GUEST_SYNC(%ld), #NM exception and enable amx\n", uc.args[1]);
334 				break;
335 			}
336 			break;
337 		case UCALL_DONE:
338 			fprintf(stderr, "UCALL_DONE\n");
339 			goto done;
340 		default:
341 			TEST_FAIL("Unknown ucall %lu", uc.cmd);
342 		}
343 
344 		state = vcpu_save_state(vcpu);
345 		memset(&regs1, 0, sizeof(regs1));
346 		vcpu_regs_get(vcpu, &regs1);
347 
348 		kvm_vm_release(vm);
349 
350 		/* Restore state in a new VM.  */
351 		vcpu = vm_recreate_with_one_vcpu(vm);
352 		vcpu_load_state(vcpu, state);
353 		run = vcpu->run;
354 		kvm_x86_state_cleanup(state);
355 
356 		memset(&regs2, 0, sizeof(regs2));
357 		vcpu_regs_get(vcpu, &regs2);
358 		TEST_ASSERT(!memcmp(&regs1, &regs2, sizeof(regs2)),
359 			    "Unexpected register values after vcpu_load_state; rdi: %lx rsi: %lx",
360 			    (ulong) regs2.rdi, (ulong) regs2.rsi);
361 	}
362 done:
363 	kvm_vm_free(vm);
364 }
365