target/arm: Implement bfloat16 matrix multiply accumulate
[qemu.git] / target / arm / arch_dump.c
blob01848453109137b711bda9ab86d33cd6b8831429
1 /* Support for writing ELF notes for ARM architectures
3 * Copyright (C) 2015 Red Hat Inc.
5 * Author: Andrew Jones <drjones@redhat.com>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (at your option) any later version.
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License along
18 * with this program; if not, see <http://www.gnu.org/licenses/>.
21 #include "qemu/osdep.h"
22 #include "cpu.h"
23 #include "elf.h"
24 #include "sysemu/dump.h"
26 /* struct user_pt_regs from arch/arm64/include/uapi/asm/ptrace.h */
27 struct aarch64_user_regs {
28 uint64_t regs[31];
29 uint64_t sp;
30 uint64_t pc;
31 uint64_t pstate;
32 } QEMU_PACKED;
34 QEMU_BUILD_BUG_ON(sizeof(struct aarch64_user_regs) != 272);
36 /* struct elf_prstatus from include/uapi/linux/elfcore.h */
37 struct aarch64_elf_prstatus {
38 char pad1[32]; /* 32 == offsetof(struct elf_prstatus, pr_pid) */
39 uint32_t pr_pid;
40 char pad2[76]; /* 76 == offsetof(struct elf_prstatus, pr_reg) -
41 offsetof(struct elf_prstatus, pr_ppid) */
42 struct aarch64_user_regs pr_reg;
43 uint32_t pr_fpvalid;
44 char pad3[4];
45 } QEMU_PACKED;
47 QEMU_BUILD_BUG_ON(sizeof(struct aarch64_elf_prstatus) != 392);
49 /* struct user_fpsimd_state from arch/arm64/include/uapi/asm/ptrace.h
51 * While the vregs member of user_fpsimd_state is of type __uint128_t,
52 * QEMU uses an array of uint64_t, where the high half of the 128-bit
53 * value is always in the 2n+1'th index. Thus we also break the 128-
54 * bit values into two halves in this reproduction of user_fpsimd_state.
56 struct aarch64_user_vfp_state {
57 uint64_t vregs[64];
58 uint32_t fpsr;
59 uint32_t fpcr;
60 char pad[8];
61 } QEMU_PACKED;
63 QEMU_BUILD_BUG_ON(sizeof(struct aarch64_user_vfp_state) != 528);
65 /* struct user_sve_header from arch/arm64/include/uapi/asm/ptrace.h */
66 struct aarch64_user_sve_header {
67 uint32_t size;
68 uint32_t max_size;
69 uint16_t vl;
70 uint16_t max_vl;
71 uint16_t flags;
72 uint16_t reserved;
73 } QEMU_PACKED;
75 struct aarch64_note {
76 Elf64_Nhdr hdr;
77 char name[8]; /* align_up(sizeof("CORE"), 4) */
78 union {
79 struct aarch64_elf_prstatus prstatus;
80 struct aarch64_user_vfp_state vfp;
81 struct aarch64_user_sve_header sve;
83 } QEMU_PACKED;
85 #define AARCH64_NOTE_HEADER_SIZE offsetof(struct aarch64_note, prstatus)
86 #define AARCH64_PRSTATUS_NOTE_SIZE \
87 (AARCH64_NOTE_HEADER_SIZE + sizeof(struct aarch64_elf_prstatus))
88 #define AARCH64_PRFPREG_NOTE_SIZE \
89 (AARCH64_NOTE_HEADER_SIZE + sizeof(struct aarch64_user_vfp_state))
90 #define AARCH64_SVE_NOTE_SIZE(env) \
91 (AARCH64_NOTE_HEADER_SIZE + sve_size(env))
93 static void aarch64_note_init(struct aarch64_note *note, DumpState *s,
94 const char *name, Elf64_Word namesz,
95 Elf64_Word type, Elf64_Word descsz)
97 memset(note, 0, sizeof(*note));
99 note->hdr.n_namesz = cpu_to_dump32(s, namesz);
100 note->hdr.n_descsz = cpu_to_dump32(s, descsz);
101 note->hdr.n_type = cpu_to_dump32(s, type);
103 memcpy(note->name, name, namesz);
106 static int aarch64_write_elf64_prfpreg(WriteCoreDumpFunction f,
107 CPUARMState *env, int cpuid,
108 DumpState *s)
110 struct aarch64_note note;
111 int ret, i;
113 aarch64_note_init(&note, s, "CORE", 5, NT_PRFPREG, sizeof(note.vfp));
115 for (i = 0; i < 32; ++i) {
116 uint64_t *q = aa64_vfp_qreg(env, i);
117 note.vfp.vregs[2 * i + 0] = cpu_to_dump64(s, q[0]);
118 note.vfp.vregs[2 * i + 1] = cpu_to_dump64(s, q[1]);
121 if (s->dump_info.d_endian == ELFDATA2MSB) {
122 /* For AArch64 we must always swap the vfp.regs's 2n and 2n+1
123 * entries when generating BE notes, because even big endian
124 * hosts use 2n+1 for the high half.
126 for (i = 0; i < 32; ++i) {
127 uint64_t tmp = note.vfp.vregs[2*i];
128 note.vfp.vregs[2 * i] = note.vfp.vregs[2 * i + 1];
129 note.vfp.vregs[2 * i + 1] = tmp;
133 note.vfp.fpsr = cpu_to_dump32(s, vfp_get_fpsr(env));
134 note.vfp.fpcr = cpu_to_dump32(s, vfp_get_fpcr(env));
136 ret = f(&note, AARCH64_PRFPREG_NOTE_SIZE, s);
137 if (ret < 0) {
138 return -1;
141 return 0;
144 #ifdef TARGET_AARCH64
145 static off_t sve_zreg_offset(uint32_t vq, int n)
147 off_t off = sizeof(struct aarch64_user_sve_header);
148 return ROUND_UP(off, 16) + vq * 16 * n;
151 static off_t sve_preg_offset(uint32_t vq, int n)
153 return sve_zreg_offset(vq, 32) + vq * 16 / 8 * n;
156 static off_t sve_fpsr_offset(uint32_t vq)
158 off_t off = sve_preg_offset(vq, 17);
159 return ROUND_UP(off, 16);
162 static off_t sve_fpcr_offset(uint32_t vq)
164 return sve_fpsr_offset(vq) + sizeof(uint32_t);
167 static uint32_t sve_current_vq(CPUARMState *env)
169 return sve_zcr_len_for_el(env, arm_current_el(env)) + 1;
172 static size_t sve_size_vq(uint32_t vq)
174 off_t off = sve_fpcr_offset(vq) + sizeof(uint32_t);
175 return ROUND_UP(off, 16);
178 static size_t sve_size(CPUARMState *env)
180 return sve_size_vq(sve_current_vq(env));
183 static int aarch64_write_elf64_sve(WriteCoreDumpFunction f,
184 CPUARMState *env, int cpuid,
185 DumpState *s)
187 struct aarch64_note *note;
188 ARMCPU *cpu = env_archcpu(env);
189 uint32_t vq = sve_current_vq(env);
190 uint64_t tmp[ARM_MAX_VQ * 2], *r;
191 uint32_t fpr;
192 uint8_t *buf;
193 int ret, i;
195 note = g_malloc0(AARCH64_SVE_NOTE_SIZE(env));
196 buf = (uint8_t *)&note->sve;
198 aarch64_note_init(note, s, "LINUX", 6, NT_ARM_SVE, sve_size_vq(vq));
200 note->sve.size = cpu_to_dump32(s, sve_size_vq(vq));
201 note->sve.max_size = cpu_to_dump32(s, sve_size_vq(cpu->sve_max_vq));
202 note->sve.vl = cpu_to_dump16(s, vq * 16);
203 note->sve.max_vl = cpu_to_dump16(s, cpu->sve_max_vq * 16);
204 note->sve.flags = cpu_to_dump16(s, 1);
206 for (i = 0; i < 32; ++i) {
207 r = sve_bswap64(tmp, &env->vfp.zregs[i].d[0], vq * 2);
208 memcpy(&buf[sve_zreg_offset(vq, i)], r, vq * 16);
211 for (i = 0; i < 17; ++i) {
212 r = sve_bswap64(tmp, r = &env->vfp.pregs[i].p[0],
213 DIV_ROUND_UP(vq * 2, 8));
214 memcpy(&buf[sve_preg_offset(vq, i)], r, vq * 16 / 8);
217 fpr = cpu_to_dump32(s, vfp_get_fpsr(env));
218 memcpy(&buf[sve_fpsr_offset(vq)], &fpr, sizeof(uint32_t));
220 fpr = cpu_to_dump32(s, vfp_get_fpcr(env));
221 memcpy(&buf[sve_fpcr_offset(vq)], &fpr, sizeof(uint32_t));
223 ret = f(note, AARCH64_SVE_NOTE_SIZE(env), s);
224 g_free(note);
226 if (ret < 0) {
227 return -1;
230 return 0;
232 #endif
234 int arm_cpu_write_elf64_note(WriteCoreDumpFunction f, CPUState *cs,
235 int cpuid, void *opaque)
237 struct aarch64_note note;
238 ARMCPU *cpu = ARM_CPU(cs);
239 CPUARMState *env = &cpu->env;
240 DumpState *s = opaque;
241 uint64_t pstate, sp;
242 int ret, i;
244 aarch64_note_init(&note, s, "CORE", 5, NT_PRSTATUS, sizeof(note.prstatus));
246 note.prstatus.pr_pid = cpu_to_dump32(s, cpuid);
247 note.prstatus.pr_fpvalid = cpu_to_dump32(s, 1);
249 if (!is_a64(env)) {
250 aarch64_sync_32_to_64(env);
251 pstate = cpsr_read(env);
252 sp = 0;
253 } else {
254 pstate = pstate_read(env);
255 sp = env->xregs[31];
258 for (i = 0; i < 31; ++i) {
259 note.prstatus.pr_reg.regs[i] = cpu_to_dump64(s, env->xregs[i]);
261 note.prstatus.pr_reg.sp = cpu_to_dump64(s, sp);
262 note.prstatus.pr_reg.pc = cpu_to_dump64(s, env->pc);
263 note.prstatus.pr_reg.pstate = cpu_to_dump64(s, pstate);
265 ret = f(&note, AARCH64_PRSTATUS_NOTE_SIZE, s);
266 if (ret < 0) {
267 return -1;
270 ret = aarch64_write_elf64_prfpreg(f, env, cpuid, s);
271 if (ret) {
272 return ret;
275 #ifdef TARGET_AARCH64
276 if (cpu_isar_feature(aa64_sve, cpu)) {
277 ret = aarch64_write_elf64_sve(f, env, cpuid, s);
279 #endif
281 return ret;
284 /* struct pt_regs from arch/arm/include/asm/ptrace.h */
285 struct arm_user_regs {
286 uint32_t regs[17];
287 char pad[4];
288 } QEMU_PACKED;
290 QEMU_BUILD_BUG_ON(sizeof(struct arm_user_regs) != 72);
292 /* struct elf_prstatus from include/uapi/linux/elfcore.h */
293 struct arm_elf_prstatus {
294 char pad1[24]; /* 24 == offsetof(struct elf_prstatus, pr_pid) */
295 uint32_t pr_pid;
296 char pad2[44]; /* 44 == offsetof(struct elf_prstatus, pr_reg) -
297 offsetof(struct elf_prstatus, pr_ppid) */
298 struct arm_user_regs pr_reg;
299 uint32_t pr_fpvalid;
300 } QEMU_PACKED arm_elf_prstatus;
302 QEMU_BUILD_BUG_ON(sizeof(struct arm_elf_prstatus) != 148);
304 /* struct user_vfp from arch/arm/include/asm/user.h */
305 struct arm_user_vfp_state {
306 uint64_t vregs[32];
307 uint32_t fpscr;
308 } QEMU_PACKED;
310 QEMU_BUILD_BUG_ON(sizeof(struct arm_user_vfp_state) != 260);
312 struct arm_note {
313 Elf32_Nhdr hdr;
314 char name[8]; /* align_up(sizeof("LINUX"), 4) */
315 union {
316 struct arm_elf_prstatus prstatus;
317 struct arm_user_vfp_state vfp;
319 } QEMU_PACKED;
321 #define ARM_NOTE_HEADER_SIZE offsetof(struct arm_note, prstatus)
322 #define ARM_PRSTATUS_NOTE_SIZE \
323 (ARM_NOTE_HEADER_SIZE + sizeof(struct arm_elf_prstatus))
324 #define ARM_VFP_NOTE_SIZE \
325 (ARM_NOTE_HEADER_SIZE + sizeof(struct arm_user_vfp_state))
327 static void arm_note_init(struct arm_note *note, DumpState *s,
328 const char *name, Elf32_Word namesz,
329 Elf32_Word type, Elf32_Word descsz)
331 memset(note, 0, sizeof(*note));
333 note->hdr.n_namesz = cpu_to_dump32(s, namesz);
334 note->hdr.n_descsz = cpu_to_dump32(s, descsz);
335 note->hdr.n_type = cpu_to_dump32(s, type);
337 memcpy(note->name, name, namesz);
340 static int arm_write_elf32_vfp(WriteCoreDumpFunction f, CPUARMState *env,
341 int cpuid, DumpState *s)
343 struct arm_note note;
344 int ret, i;
346 arm_note_init(&note, s, "LINUX", 6, NT_ARM_VFP, sizeof(note.vfp));
348 for (i = 0; i < 32; ++i) {
349 note.vfp.vregs[i] = cpu_to_dump64(s, *aa32_vfp_dreg(env, i));
352 note.vfp.fpscr = cpu_to_dump32(s, vfp_get_fpscr(env));
354 ret = f(&note, ARM_VFP_NOTE_SIZE, s);
355 if (ret < 0) {
356 return -1;
359 return 0;
362 int arm_cpu_write_elf32_note(WriteCoreDumpFunction f, CPUState *cs,
363 int cpuid, void *opaque)
365 struct arm_note note;
366 ARMCPU *cpu = ARM_CPU(cs);
367 CPUARMState *env = &cpu->env;
368 DumpState *s = opaque;
369 int ret, i;
370 bool fpvalid = cpu_isar_feature(aa32_vfp_simd, cpu);
372 arm_note_init(&note, s, "CORE", 5, NT_PRSTATUS, sizeof(note.prstatus));
374 note.prstatus.pr_pid = cpu_to_dump32(s, cpuid);
375 note.prstatus.pr_fpvalid = cpu_to_dump32(s, fpvalid);
377 for (i = 0; i < 16; ++i) {
378 note.prstatus.pr_reg.regs[i] = cpu_to_dump32(s, env->regs[i]);
380 note.prstatus.pr_reg.regs[16] = cpu_to_dump32(s, cpsr_read(env));
382 ret = f(&note, ARM_PRSTATUS_NOTE_SIZE, s);
383 if (ret < 0) {
384 return -1;
385 } else if (fpvalid) {
386 return arm_write_elf32_vfp(f, env, cpuid, s);
389 return 0;
392 int cpu_get_dump_info(ArchDumpInfo *info,
393 const GuestPhysBlockList *guest_phys_blocks)
395 ARMCPU *cpu;
396 CPUARMState *env;
397 GuestPhysBlock *block;
398 hwaddr lowest_addr = ULLONG_MAX;
400 if (first_cpu == NULL) {
401 return -1;
404 cpu = ARM_CPU(first_cpu);
405 env = &cpu->env;
407 /* Take a best guess at the phys_base. If we get it wrong then crash
408 * will need '--machdep phys_offset=<phys-offset>' added to its command
409 * line, which isn't any worse than assuming we can use zero, but being
410 * wrong. This is the same algorithm the crash utility uses when
411 * attempting to guess as it loads non-dumpfile formatted files.
413 QTAILQ_FOREACH(block, &guest_phys_blocks->head, next) {
414 if (block->target_start < lowest_addr) {
415 lowest_addr = block->target_start;
419 if (arm_feature(env, ARM_FEATURE_AARCH64)) {
420 info->d_machine = EM_AARCH64;
421 info->d_class = ELFCLASS64;
422 info->page_size = (1 << 16); /* aarch64 max pagesize */
423 if (lowest_addr != ULLONG_MAX) {
424 info->phys_base = lowest_addr;
426 } else {
427 info->d_machine = EM_ARM;
428 info->d_class = ELFCLASS32;
429 info->page_size = (1 << 12);
430 if (lowest_addr < UINT_MAX) {
431 info->phys_base = lowest_addr;
435 /* We assume the relevant endianness is that of EL1; this is right
436 * for kernels, but might give the wrong answer if you're trying to
437 * dump a hypervisor that happens to be running an opposite-endian
438 * kernel.
440 info->d_endian = (env->cp15.sctlr_el[1] & SCTLR_EE) != 0
441 ? ELFDATA2MSB : ELFDATA2LSB;
443 return 0;
446 ssize_t cpu_get_note_size(int class, int machine, int nr_cpus)
448 ARMCPU *cpu = ARM_CPU(first_cpu);
449 size_t note_size;
451 if (class == ELFCLASS64) {
452 note_size = AARCH64_PRSTATUS_NOTE_SIZE;
453 note_size += AARCH64_PRFPREG_NOTE_SIZE;
454 #ifdef TARGET_AARCH64
455 if (cpu_isar_feature(aa64_sve, cpu)) {
456 note_size += AARCH64_SVE_NOTE_SIZE(&cpu->env);
458 #endif
459 } else {
460 note_size = ARM_PRSTATUS_NOTE_SIZE;
461 if (cpu_isar_feature(aa32_vfp_simd, cpu)) {
462 note_size += ARM_VFP_NOTE_SIZE;
466 return note_size * nr_cpus;