1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * FP/SIMD context switching and fault handling 4 * 5 * Copyright (C) 2012 ARM Ltd. 6 * Author: Catalin Marinas <catalin.marinas@arm.com> 7 */ 8 9 #include <linux/bitmap.h> 10 #include <linux/bitops.h> 11 #include <linux/bottom_half.h> 12 #include <linux/bug.h> 13 #include <linux/cache.h> 14 #include <linux/compat.h> 15 #include <linux/compiler.h> 16 #include <linux/cpu.h> 17 #include <linux/cpu_pm.h> 18 #include <linux/ctype.h> 19 #include <linux/kernel.h> 20 #include <linux/linkage.h> 21 #include <linux/irqflags.h> 22 #include <linux/init.h> 23 #include <linux/percpu.h> 24 #include <linux/prctl.h> 25 #include <linux/preempt.h> 26 #include <linux/ptrace.h> 27 #include <linux/sched/signal.h> 28 #include <linux/sched/task_stack.h> 29 #include <linux/signal.h> 30 #include <linux/slab.h> 31 #include <linux/stddef.h> 32 #include <linux/sysctl.h> 33 #include <linux/swab.h> 34 35 #include <asm/esr.h> 36 #include <asm/exception.h> 37 #include <asm/fpsimd.h> 38 #include <asm/cpufeature.h> 39 #include <asm/cputype.h> 40 #include <asm/neon.h> 41 #include <asm/processor.h> 42 #include <asm/simd.h> 43 #include <asm/sigcontext.h> 44 #include <asm/sysreg.h> 45 #include <asm/traps.h> 46 #include <asm/virt.h> 47 48 #define FPEXC_IOF (1 << 0) 49 #define FPEXC_DZF (1 << 1) 50 #define FPEXC_OFF (1 << 2) 51 #define FPEXC_UFF (1 << 3) 52 #define FPEXC_IXF (1 << 4) 53 #define FPEXC_IDF (1 << 7) 54 55 /* 56 * (Note: in this discussion, statements about FPSIMD apply equally to SVE.) 57 * 58 * In order to reduce the number of times the FPSIMD state is needlessly saved 59 * and restored, we need to keep track of two things: 60 * (a) for each task, we need to remember which CPU was the last one to have 61 * the task's FPSIMD state loaded into its FPSIMD registers; 62 * (b) for each CPU, we need to remember which task's userland FPSIMD state has 63 * been loaded into its FPSIMD registers most recently, or whether it has 64 * been used to perform kernel mode NEON in the meantime. 65 * 66 * For (a), we add a fpsimd_cpu field to thread_struct, which gets updated to 67 * the id of the current CPU every time the state is loaded onto a CPU. For (b), 68 * we add the per-cpu variable 'fpsimd_last_state' (below), which contains the 69 * address of the userland FPSIMD state of the task that was loaded onto the CPU 70 * the most recently, or NULL if kernel mode NEON has been performed after that. 71 * 72 * With this in place, we no longer have to restore the next FPSIMD state right 73 * when switching between tasks. Instead, we can defer this check to userland 74 * resume, at which time we verify whether the CPU's fpsimd_last_state and the 75 * task's fpsimd_cpu are still mutually in sync. If this is the case, we 76 * can omit the FPSIMD restore. 77 * 78 * As an optimization, we use the thread_info flag TIF_FOREIGN_FPSTATE to 79 * indicate whether or not the userland FPSIMD state of the current task is 80 * present in the registers. The flag is set unless the FPSIMD registers of this 81 * CPU currently contain the most recent userland FPSIMD state of the current 82 * task. If the task is behaving as a VMM, then this is will be managed by 83 * KVM which will clear it to indicate that the vcpu FPSIMD state is currently 84 * loaded on the CPU, allowing the state to be saved if a FPSIMD-aware 85 * softirq kicks in. Upon vcpu_put(), KVM will save the vcpu FP state and 86 * flag the register state as invalid. 87 * 88 * In order to allow softirq handlers to use FPSIMD, kernel_neon_begin() may be 89 * called from softirq context, which will save the task's FPSIMD context back 90 * to task_struct. To prevent this from racing with the manipulation of the 91 * task's FPSIMD state from task context and thereby corrupting the state, it 92 * is necessary to protect any manipulation of a task's fpsimd_state or 93 * TIF_FOREIGN_FPSTATE flag with get_cpu_fpsimd_context(), which will suspend 94 * softirq servicing entirely until put_cpu_fpsimd_context() is called. 95 * 96 * For a certain task, the sequence may look something like this: 97 * - the task gets scheduled in; if both the task's fpsimd_cpu field 98 * contains the id of the current CPU, and the CPU's fpsimd_last_state per-cpu 99 * variable points to the task's fpsimd_state, the TIF_FOREIGN_FPSTATE flag is 100 * cleared, otherwise it is set; 101 * 102 * - the task returns to userland; if TIF_FOREIGN_FPSTATE is set, the task's 103 * userland FPSIMD state is copied from memory to the registers, the task's 104 * fpsimd_cpu field is set to the id of the current CPU, the current 105 * CPU's fpsimd_last_state pointer is set to this task's fpsimd_state and the 106 * TIF_FOREIGN_FPSTATE flag is cleared; 107 * 108 * - the task executes an ordinary syscall; upon return to userland, the 109 * TIF_FOREIGN_FPSTATE flag will still be cleared, so no FPSIMD state is 110 * restored; 111 * 112 * - the task executes a syscall which executes some NEON instructions; this is 113 * preceded by a call to kernel_neon_begin(), which copies the task's FPSIMD 114 * register contents to memory, clears the fpsimd_last_state per-cpu variable 115 * and sets the TIF_FOREIGN_FPSTATE flag; 116 * 117 * - the task gets preempted after kernel_neon_end() is called; as we have not 118 * returned from the 2nd syscall yet, TIF_FOREIGN_FPSTATE is still set so 119 * whatever is in the FPSIMD registers is not saved to memory, but discarded. 120 */ 121 122 static DEFINE_PER_CPU(struct cpu_fp_state, fpsimd_last_state); 123 124 __ro_after_init struct vl_info vl_info[ARM64_VEC_MAX] = { 125 #ifdef CONFIG_ARM64_SVE 126 [ARM64_VEC_SVE] = { 127 .type = ARM64_VEC_SVE, 128 .name = "SVE", 129 .min_vl = SVE_VL_MIN, 130 .max_vl = SVE_VL_MIN, 131 .max_virtualisable_vl = SVE_VL_MIN, 132 }, 133 #endif 134 #ifdef CONFIG_ARM64_SME 135 [ARM64_VEC_SME] = { 136 .type = ARM64_VEC_SME, 137 .name = "SME", 138 }, 139 #endif 140 }; 141 142 static unsigned int vec_vl_inherit_flag(enum vec_type type) 143 { 144 switch (type) { 145 case ARM64_VEC_SVE: 146 return TIF_SVE_VL_INHERIT; 147 case ARM64_VEC_SME: 148 return TIF_SME_VL_INHERIT; 149 default: 150 WARN_ON_ONCE(1); 151 return 0; 152 } 153 } 154 155 struct vl_config { 156 int __default_vl; /* Default VL for tasks */ 157 }; 158 159 static struct vl_config vl_config[ARM64_VEC_MAX]; 160 161 static inline int get_default_vl(enum vec_type type) 162 { 163 return READ_ONCE(vl_config[type].__default_vl); 164 } 165 166 #ifdef CONFIG_ARM64_SVE 167 168 static inline int get_sve_default_vl(void) 169 { 170 return get_default_vl(ARM64_VEC_SVE); 171 } 172 173 static inline void set_default_vl(enum vec_type type, int val) 174 { 175 WRITE_ONCE(vl_config[type].__default_vl, val); 176 } 177 178 static inline void set_sve_default_vl(int val) 179 { 180 set_default_vl(ARM64_VEC_SVE, val); 181 } 182 183 static void __percpu *efi_sve_state; 184 185 #else /* ! CONFIG_ARM64_SVE */ 186 187 /* Dummy declaration for code that will be optimised out: */ 188 extern void __percpu *efi_sve_state; 189 190 #endif /* ! CONFIG_ARM64_SVE */ 191 192 #ifdef CONFIG_ARM64_SME 193 194 static int get_sme_default_vl(void) 195 { 196 return get_default_vl(ARM64_VEC_SME); 197 } 198 199 static void set_sme_default_vl(int val) 200 { 201 set_default_vl(ARM64_VEC_SME, val); 202 } 203 204 static void sme_free(struct task_struct *); 205 206 #else 207 208 static inline void sme_free(struct task_struct *t) { } 209 210 #endif 211 212 static void fpsimd_bind_task_to_cpu(void); 213 214 /* 215 * Claim ownership of the CPU FPSIMD context for use by the calling context. 216 * 217 * The caller may freely manipulate the FPSIMD context metadata until 218 * put_cpu_fpsimd_context() is called. 219 * 220 * On RT kernels local_bh_disable() is not sufficient because it only 221 * serializes soft interrupt related sections via a local lock, but stays 222 * preemptible. Disabling preemption is the right choice here as bottom 223 * half processing is always in thread context on RT kernels so it 224 * implicitly prevents bottom half processing as well. 225 */ 226 static void get_cpu_fpsimd_context(void) 227 { 228 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 229 local_bh_disable(); 230 else 231 preempt_disable(); 232 } 233 234 /* 235 * Release the CPU FPSIMD context. 236 * 237 * Must be called from a context in which get_cpu_fpsimd_context() was 238 * previously called, with no call to put_cpu_fpsimd_context() in the 239 * meantime. 240 */ 241 static void put_cpu_fpsimd_context(void) 242 { 243 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 244 local_bh_enable(); 245 else 246 preempt_enable(); 247 } 248 249 unsigned int task_get_vl(const struct task_struct *task, enum vec_type type) 250 { 251 return task->thread.vl[type]; 252 } 253 254 void task_set_vl(struct task_struct *task, enum vec_type type, 255 unsigned long vl) 256 { 257 task->thread.vl[type] = vl; 258 } 259 260 unsigned int task_get_vl_onexec(const struct task_struct *task, 261 enum vec_type type) 262 { 263 return task->thread.vl_onexec[type]; 264 } 265 266 void task_set_vl_onexec(struct task_struct *task, enum vec_type type, 267 unsigned long vl) 268 { 269 task->thread.vl_onexec[type] = vl; 270 } 271 272 /* 273 * TIF_SME controls whether a task can use SME without trapping while 274 * in userspace, when TIF_SME is set then we must have storage 275 * allocated in sve_state and sme_state to store the contents of both ZA 276 * and the SVE registers for both streaming and non-streaming modes. 277 * 278 * If both SVCR.ZA and SVCR.SM are disabled then at any point we 279 * may disable TIF_SME and reenable traps. 280 */ 281 282 283 /* 284 * TIF_SVE controls whether a task can use SVE without trapping while 285 * in userspace, and also (together with TIF_SME) the way a task's 286 * FPSIMD/SVE state is stored in thread_struct. 287 * 288 * The kernel uses this flag to track whether a user task is actively 289 * using SVE, and therefore whether full SVE register state needs to 290 * be tracked. If not, the cheaper FPSIMD context handling code can 291 * be used instead of the more costly SVE equivalents. 292 * 293 * * TIF_SVE or SVCR.SM set: 294 * 295 * The task can execute SVE instructions while in userspace without 296 * trapping to the kernel. 297 * 298 * During any syscall, the kernel may optionally clear TIF_SVE and 299 * discard the vector state except for the FPSIMD subset. 300 * 301 * * TIF_SVE clear: 302 * 303 * An attempt by the user task to execute an SVE instruction causes 304 * do_sve_acc() to be called, which does some preparation and then 305 * sets TIF_SVE. 306 * 307 * During any syscall, the kernel may optionally clear TIF_SVE and 308 * discard the vector state except for the FPSIMD subset. 309 * 310 * The data will be stored in one of two formats: 311 * 312 * * FPSIMD only - FP_STATE_FPSIMD: 313 * 314 * When the FPSIMD only state stored task->thread.fp_type is set to 315 * FP_STATE_FPSIMD, the FPSIMD registers V0-V31 are encoded in 316 * task->thread.uw.fpsimd_state; bits [max : 128] for each of Z0-Z31 are 317 * logically zero but not stored anywhere; P0-P15 and FFR are not 318 * stored and have unspecified values from userspace's point of 319 * view. For hygiene purposes, the kernel zeroes them on next use, 320 * but userspace is discouraged from relying on this. 321 * 322 * task->thread.sve_state does not need to be non-NULL, valid or any 323 * particular size: it must not be dereferenced and any data stored 324 * there should be considered stale and not referenced. 325 * 326 * * SVE state - FP_STATE_SVE: 327 * 328 * When the full SVE state is stored task->thread.fp_type is set to 329 * FP_STATE_SVE and Z0-Z31 (incorporating Vn in bits[127:0] or the 330 * corresponding Zn), P0-P15 and FFR are encoded in in 331 * task->thread.sve_state, formatted appropriately for vector 332 * length task->thread.sve_vl or, if SVCR.SM is set, 333 * task->thread.sme_vl. The storage for the vector registers in 334 * task->thread.uw.fpsimd_state should be ignored. 335 * 336 * task->thread.sve_state must point to a valid buffer at least 337 * sve_state_size(task) bytes in size. The data stored in 338 * task->thread.uw.fpsimd_state.vregs should be considered stale 339 * and not referenced. 340 * 341 * * FPSR and FPCR are always stored in task->thread.uw.fpsimd_state 342 * irrespective of whether TIF_SVE is clear or set, since these are 343 * not vector length dependent. 344 */ 345 346 /* 347 * Update current's FPSIMD/SVE registers from thread_struct. 348 * 349 * This function should be called only when the FPSIMD/SVE state in 350 * thread_struct is known to be up to date, when preparing to enter 351 * userspace. 352 */ 353 static void task_fpsimd_load(void) 354 { 355 bool restore_sve_regs = false; 356 bool restore_ffr; 357 358 WARN_ON(!system_supports_fpsimd()); 359 WARN_ON(preemptible()); 360 WARN_ON(test_thread_flag(TIF_KERNEL_FPSTATE)); 361 362 if (system_supports_fpmr()) 363 write_sysreg_s(current->thread.uw.fpmr, SYS_FPMR); 364 365 if (system_supports_sve() || system_supports_sme()) { 366 switch (current->thread.fp_type) { 367 case FP_STATE_FPSIMD: 368 /* Stop tracking SVE for this task until next use. */ 369 if (test_and_clear_thread_flag(TIF_SVE)) 370 sve_user_disable(); 371 break; 372 case FP_STATE_SVE: 373 if (!thread_sm_enabled(¤t->thread) && 374 !WARN_ON_ONCE(!test_and_set_thread_flag(TIF_SVE))) 375 sve_user_enable(); 376 377 if (test_thread_flag(TIF_SVE)) 378 sve_set_vq(sve_vq_from_vl(task_get_sve_vl(current)) - 1); 379 380 restore_sve_regs = true; 381 restore_ffr = true; 382 break; 383 default: 384 /* 385 * This indicates either a bug in 386 * fpsimd_save_user_state() or memory corruption, we 387 * should always record an explicit format 388 * when we save. We always at least have the 389 * memory allocated for FPSMID registers so 390 * try that and hope for the best. 391 */ 392 WARN_ON_ONCE(1); 393 clear_thread_flag(TIF_SVE); 394 break; 395 } 396 } 397 398 /* Restore SME, override SVE register configuration if needed */ 399 if (system_supports_sme()) { 400 unsigned long sme_vl = task_get_sme_vl(current); 401 402 /* Ensure VL is set up for restoring data */ 403 if (test_thread_flag(TIF_SME)) 404 sme_set_vq(sve_vq_from_vl(sme_vl) - 1); 405 406 write_sysreg_s(current->thread.svcr, SYS_SVCR); 407 408 if (thread_za_enabled(¤t->thread)) 409 sme_load_state(current->thread.sme_state, 410 system_supports_sme2()); 411 412 if (thread_sm_enabled(¤t->thread)) 413 restore_ffr = system_supports_fa64(); 414 } 415 416 if (restore_sve_regs) { 417 WARN_ON_ONCE(current->thread.fp_type != FP_STATE_SVE); 418 sve_load_state(sve_pffr(¤t->thread), 419 ¤t->thread.uw.fpsimd_state.fpsr, 420 restore_ffr); 421 } else { 422 WARN_ON_ONCE(current->thread.fp_type != FP_STATE_FPSIMD); 423 fpsimd_load_state(¤t->thread.uw.fpsimd_state); 424 } 425 } 426 427 /* 428 * Ensure FPSIMD/SVE storage in memory for the loaded context is up to 429 * date with respect to the CPU registers. Note carefully that the 430 * current context is the context last bound to the CPU stored in 431 * last, if KVM is involved this may be the guest VM context rather 432 * than the host thread for the VM pointed to by current. This means 433 * that we must always reference the state storage via last rather 434 * than via current, if we are saving KVM state then it will have 435 * ensured that the type of registers to save is set in last->to_save. 436 */ 437 static void fpsimd_save_user_state(void) 438 { 439 struct cpu_fp_state const *last = 440 this_cpu_ptr(&fpsimd_last_state); 441 /* set by fpsimd_bind_task_to_cpu() or fpsimd_bind_state_to_cpu() */ 442 bool save_sve_regs = false; 443 bool save_ffr; 444 unsigned int vl; 445 446 WARN_ON(!system_supports_fpsimd()); 447 WARN_ON(preemptible()); 448 449 if (test_thread_flag(TIF_FOREIGN_FPSTATE)) 450 return; 451 452 if (system_supports_fpmr()) 453 *(last->fpmr) = read_sysreg_s(SYS_FPMR); 454 455 /* 456 * If a task is in a syscall the ABI allows us to only 457 * preserve the state shared with FPSIMD so don't bother 458 * saving the full SVE state in that case. 459 */ 460 if ((last->to_save == FP_STATE_CURRENT && test_thread_flag(TIF_SVE) && 461 !in_syscall(current_pt_regs())) || 462 last->to_save == FP_STATE_SVE) { 463 save_sve_regs = true; 464 save_ffr = true; 465 vl = last->sve_vl; 466 } 467 468 if (system_supports_sme()) { 469 u64 *svcr = last->svcr; 470 471 *svcr = read_sysreg_s(SYS_SVCR); 472 473 if (*svcr & SVCR_ZA_MASK) 474 sme_save_state(last->sme_state, 475 system_supports_sme2()); 476 477 /* If we are in streaming mode override regular SVE. */ 478 if (*svcr & SVCR_SM_MASK) { 479 save_sve_regs = true; 480 save_ffr = system_supports_fa64(); 481 vl = last->sme_vl; 482 } 483 } 484 485 if (IS_ENABLED(CONFIG_ARM64_SVE) && save_sve_regs) { 486 /* Get the configured VL from RDVL, will account for SM */ 487 if (WARN_ON(sve_get_vl() != vl)) { 488 /* 489 * Can't save the user regs, so current would 490 * re-enter user with corrupt state. 491 * There's no way to recover, so kill it: 492 */ 493 force_signal_inject(SIGKILL, SI_KERNEL, 0, 0); 494 return; 495 } 496 497 sve_save_state((char *)last->sve_state + 498 sve_ffr_offset(vl), 499 &last->st->fpsr, save_ffr); 500 *last->fp_type = FP_STATE_SVE; 501 } else { 502 fpsimd_save_state(last->st); 503 *last->fp_type = FP_STATE_FPSIMD; 504 } 505 } 506 507 /* 508 * All vector length selection from userspace comes through here. 509 * We're on a slow path, so some sanity-checks are included. 510 * If things go wrong there's a bug somewhere, but try to fall back to a 511 * safe choice. 512 */ 513 static unsigned int find_supported_vector_length(enum vec_type type, 514 unsigned int vl) 515 { 516 struct vl_info *info = &vl_info[type]; 517 int bit; 518 int max_vl = info->max_vl; 519 520 if (WARN_ON(!sve_vl_valid(vl))) 521 vl = info->min_vl; 522 523 if (WARN_ON(!sve_vl_valid(max_vl))) 524 max_vl = info->min_vl; 525 526 if (vl > max_vl) 527 vl = max_vl; 528 if (vl < info->min_vl) 529 vl = info->min_vl; 530 531 bit = find_next_bit(info->vq_map, SVE_VQ_MAX, 532 __vq_to_bit(sve_vq_from_vl(vl))); 533 return sve_vl_from_vq(__bit_to_vq(bit)); 534 } 535 536 #if defined(CONFIG_ARM64_SVE) && defined(CONFIG_SYSCTL) 537 538 static int vec_proc_do_default_vl(struct ctl_table *table, int write, 539 void *buffer, size_t *lenp, loff_t *ppos) 540 { 541 struct vl_info *info = table->extra1; 542 enum vec_type type = info->type; 543 int ret; 544 int vl = get_default_vl(type); 545 struct ctl_table tmp_table = { 546 .data = &vl, 547 .maxlen = sizeof(vl), 548 }; 549 550 ret = proc_dointvec(&tmp_table, write, buffer, lenp, ppos); 551 if (ret || !write) 552 return ret; 553 554 /* Writing -1 has the special meaning "set to max": */ 555 if (vl == -1) 556 vl = info->max_vl; 557 558 if (!sve_vl_valid(vl)) 559 return -EINVAL; 560 561 set_default_vl(type, find_supported_vector_length(type, vl)); 562 return 0; 563 } 564 565 static struct ctl_table sve_default_vl_table[] = { 566 { 567 .procname = "sve_default_vector_length", 568 .mode = 0644, 569 .proc_handler = vec_proc_do_default_vl, 570 .extra1 = &vl_info[ARM64_VEC_SVE], 571 }, 572 }; 573 574 static int __init sve_sysctl_init(void) 575 { 576 if (system_supports_sve()) 577 if (!register_sysctl("abi", sve_default_vl_table)) 578 return -EINVAL; 579 580 return 0; 581 } 582 583 #else /* ! (CONFIG_ARM64_SVE && CONFIG_SYSCTL) */ 584 static int __init sve_sysctl_init(void) { return 0; } 585 #endif /* ! (CONFIG_ARM64_SVE && CONFIG_SYSCTL) */ 586 587 #if defined(CONFIG_ARM64_SME) && defined(CONFIG_SYSCTL) 588 static struct ctl_table sme_default_vl_table[] = { 589 { 590 .procname = "sme_default_vector_length", 591 .mode = 0644, 592 .proc_handler = vec_proc_do_default_vl, 593 .extra1 = &vl_info[ARM64_VEC_SME], 594 }, 595 }; 596 597 static int __init sme_sysctl_init(void) 598 { 599 if (system_supports_sme()) 600 if (!register_sysctl("abi", sme_default_vl_table)) 601 return -EINVAL; 602 603 return 0; 604 } 605 606 #else /* ! (CONFIG_ARM64_SME && CONFIG_SYSCTL) */ 607 static int __init sme_sysctl_init(void) { return 0; } 608 #endif /* ! (CONFIG_ARM64_SME && CONFIG_SYSCTL) */ 609 610 #define ZREG(sve_state, vq, n) ((char *)(sve_state) + \ 611 (SVE_SIG_ZREG_OFFSET(vq, n) - SVE_SIG_REGS_OFFSET)) 612 613 #ifdef CONFIG_CPU_BIG_ENDIAN 614 static __uint128_t arm64_cpu_to_le128(__uint128_t x) 615 { 616 u64 a = swab64(x); 617 u64 b = swab64(x >> 64); 618 619 return ((__uint128_t)a << 64) | b; 620 } 621 #else 622 static __uint128_t arm64_cpu_to_le128(__uint128_t x) 623 { 624 return x; 625 } 626 #endif 627 628 #define arm64_le128_to_cpu(x) arm64_cpu_to_le128(x) 629 630 static void __fpsimd_to_sve(void *sst, struct user_fpsimd_state const *fst, 631 unsigned int vq) 632 { 633 unsigned int i; 634 __uint128_t *p; 635 636 for (i = 0; i < SVE_NUM_ZREGS; ++i) { 637 p = (__uint128_t *)ZREG(sst, vq, i); 638 *p = arm64_cpu_to_le128(fst->vregs[i]); 639 } 640 } 641 642 /* 643 * Transfer the FPSIMD state in task->thread.uw.fpsimd_state to 644 * task->thread.sve_state. 645 * 646 * Task can be a non-runnable task, or current. In the latter case, 647 * the caller must have ownership of the cpu FPSIMD context before calling 648 * this function. 649 * task->thread.sve_state must point to at least sve_state_size(task) 650 * bytes of allocated kernel memory. 651 * task->thread.uw.fpsimd_state must be up to date before calling this 652 * function. 653 */ 654 static void fpsimd_to_sve(struct task_struct *task) 655 { 656 unsigned int vq; 657 void *sst = task->thread.sve_state; 658 struct user_fpsimd_state const *fst = &task->thread.uw.fpsimd_state; 659 660 if (!system_supports_sve() && !system_supports_sme()) 661 return; 662 663 vq = sve_vq_from_vl(thread_get_cur_vl(&task->thread)); 664 __fpsimd_to_sve(sst, fst, vq); 665 } 666 667 /* 668 * Transfer the SVE state in task->thread.sve_state to 669 * task->thread.uw.fpsimd_state. 670 * 671 * Task can be a non-runnable task, or current. In the latter case, 672 * the caller must have ownership of the cpu FPSIMD context before calling 673 * this function. 674 * task->thread.sve_state must point to at least sve_state_size(task) 675 * bytes of allocated kernel memory. 676 * task->thread.sve_state must be up to date before calling this function. 677 */ 678 static void sve_to_fpsimd(struct task_struct *task) 679 { 680 unsigned int vq, vl; 681 void const *sst = task->thread.sve_state; 682 struct user_fpsimd_state *fst = &task->thread.uw.fpsimd_state; 683 unsigned int i; 684 __uint128_t const *p; 685 686 if (!system_supports_sve() && !system_supports_sme()) 687 return; 688 689 vl = thread_get_cur_vl(&task->thread); 690 vq = sve_vq_from_vl(vl); 691 for (i = 0; i < SVE_NUM_ZREGS; ++i) { 692 p = (__uint128_t const *)ZREG(sst, vq, i); 693 fst->vregs[i] = arm64_le128_to_cpu(*p); 694 } 695 } 696 697 void cpu_enable_fpmr(const struct arm64_cpu_capabilities *__always_unused p) 698 { 699 write_sysreg_s(read_sysreg_s(SYS_SCTLR_EL1) | SCTLR_EL1_EnFPM_MASK, 700 SYS_SCTLR_EL1); 701 } 702 703 #ifdef CONFIG_ARM64_SVE 704 /* 705 * Call __sve_free() directly only if you know task can't be scheduled 706 * or preempted. 707 */ 708 static void __sve_free(struct task_struct *task) 709 { 710 kfree(task->thread.sve_state); 711 task->thread.sve_state = NULL; 712 } 713 714 static void sve_free(struct task_struct *task) 715 { 716 WARN_ON(test_tsk_thread_flag(task, TIF_SVE)); 717 718 __sve_free(task); 719 } 720 721 /* 722 * Return how many bytes of memory are required to store the full SVE 723 * state for task, given task's currently configured vector length. 724 */ 725 size_t sve_state_size(struct task_struct const *task) 726 { 727 unsigned int vl = 0; 728 729 if (system_supports_sve()) 730 vl = task_get_sve_vl(task); 731 if (system_supports_sme()) 732 vl = max(vl, task_get_sme_vl(task)); 733 734 return SVE_SIG_REGS_SIZE(sve_vq_from_vl(vl)); 735 } 736 737 /* 738 * Ensure that task->thread.sve_state is allocated and sufficiently large. 739 * 740 * This function should be used only in preparation for replacing 741 * task->thread.sve_state with new data. The memory is always zeroed 742 * here to prevent stale data from showing through: this is done in 743 * the interest of testability and predictability: except in the 744 * do_sve_acc() case, there is no ABI requirement to hide stale data 745 * written previously be task. 746 */ 747 void sve_alloc(struct task_struct *task, bool flush) 748 { 749 if (task->thread.sve_state) { 750 if (flush) 751 memset(task->thread.sve_state, 0, 752 sve_state_size(task)); 753 return; 754 } 755 756 /* This is a small allocation (maximum ~8KB) and Should Not Fail. */ 757 task->thread.sve_state = 758 kzalloc(sve_state_size(task), GFP_KERNEL); 759 } 760 761 762 /* 763 * Force the FPSIMD state shared with SVE to be updated in the SVE state 764 * even if the SVE state is the current active state. 765 * 766 * This should only be called by ptrace. task must be non-runnable. 767 * task->thread.sve_state must point to at least sve_state_size(task) 768 * bytes of allocated kernel memory. 769 */ 770 void fpsimd_force_sync_to_sve(struct task_struct *task) 771 { 772 fpsimd_to_sve(task); 773 } 774 775 /* 776 * Ensure that task->thread.sve_state is up to date with respect to 777 * the user task, irrespective of when SVE is in use or not. 778 * 779 * This should only be called by ptrace. task must be non-runnable. 780 * task->thread.sve_state must point to at least sve_state_size(task) 781 * bytes of allocated kernel memory. 782 */ 783 void fpsimd_sync_to_sve(struct task_struct *task) 784 { 785 if (!test_tsk_thread_flag(task, TIF_SVE) && 786 !thread_sm_enabled(&task->thread)) 787 fpsimd_to_sve(task); 788 } 789 790 /* 791 * Ensure that task->thread.uw.fpsimd_state is up to date with respect to 792 * the user task, irrespective of whether SVE is in use or not. 793 * 794 * This should only be called by ptrace. task must be non-runnable. 795 * task->thread.sve_state must point to at least sve_state_size(task) 796 * bytes of allocated kernel memory. 797 */ 798 void sve_sync_to_fpsimd(struct task_struct *task) 799 { 800 if (task->thread.fp_type == FP_STATE_SVE) 801 sve_to_fpsimd(task); 802 } 803 804 /* 805 * Ensure that task->thread.sve_state is up to date with respect to 806 * the task->thread.uw.fpsimd_state. 807 * 808 * This should only be called by ptrace to merge new FPSIMD register 809 * values into a task for which SVE is currently active. 810 * task must be non-runnable. 811 * task->thread.sve_state must point to at least sve_state_size(task) 812 * bytes of allocated kernel memory. 813 * task->thread.uw.fpsimd_state must already have been initialised with 814 * the new FPSIMD register values to be merged in. 815 */ 816 void sve_sync_from_fpsimd_zeropad(struct task_struct *task) 817 { 818 unsigned int vq; 819 void *sst = task->thread.sve_state; 820 struct user_fpsimd_state const *fst = &task->thread.uw.fpsimd_state; 821 822 if (!test_tsk_thread_flag(task, TIF_SVE) && 823 !thread_sm_enabled(&task->thread)) 824 return; 825 826 vq = sve_vq_from_vl(thread_get_cur_vl(&task->thread)); 827 828 memset(sst, 0, SVE_SIG_REGS_SIZE(vq)); 829 __fpsimd_to_sve(sst, fst, vq); 830 } 831 832 int vec_set_vector_length(struct task_struct *task, enum vec_type type, 833 unsigned long vl, unsigned long flags) 834 { 835 bool free_sme = false; 836 837 if (flags & ~(unsigned long)(PR_SVE_VL_INHERIT | 838 PR_SVE_SET_VL_ONEXEC)) 839 return -EINVAL; 840 841 if (!sve_vl_valid(vl)) 842 return -EINVAL; 843 844 /* 845 * Clamp to the maximum vector length that VL-agnostic code 846 * can work with. A flag may be assigned in the future to 847 * allow setting of larger vector lengths without confusing 848 * older software. 849 */ 850 if (vl > VL_ARCH_MAX) 851 vl = VL_ARCH_MAX; 852 853 vl = find_supported_vector_length(type, vl); 854 855 if (flags & (PR_SVE_VL_INHERIT | 856 PR_SVE_SET_VL_ONEXEC)) 857 task_set_vl_onexec(task, type, vl); 858 else 859 /* Reset VL to system default on next exec: */ 860 task_set_vl_onexec(task, type, 0); 861 862 /* Only actually set the VL if not deferred: */ 863 if (flags & PR_SVE_SET_VL_ONEXEC) 864 goto out; 865 866 if (vl == task_get_vl(task, type)) 867 goto out; 868 869 /* 870 * To ensure the FPSIMD bits of the SVE vector registers are preserved, 871 * write any live register state back to task_struct, and convert to a 872 * regular FPSIMD thread. 873 */ 874 if (task == current) { 875 get_cpu_fpsimd_context(); 876 877 fpsimd_save_user_state(); 878 } 879 880 fpsimd_flush_task_state(task); 881 if (test_and_clear_tsk_thread_flag(task, TIF_SVE) || 882 thread_sm_enabled(&task->thread)) { 883 sve_to_fpsimd(task); 884 task->thread.fp_type = FP_STATE_FPSIMD; 885 } 886 887 if (system_supports_sme()) { 888 if (type == ARM64_VEC_SME || 889 !(task->thread.svcr & (SVCR_SM_MASK | SVCR_ZA_MASK))) { 890 /* 891 * We are changing the SME VL or weren't using 892 * SME anyway, discard the state and force a 893 * reallocation. 894 */ 895 task->thread.svcr &= ~(SVCR_SM_MASK | 896 SVCR_ZA_MASK); 897 clear_tsk_thread_flag(task, TIF_SME); 898 free_sme = true; 899 } 900 } 901 902 if (task == current) 903 put_cpu_fpsimd_context(); 904 905 task_set_vl(task, type, vl); 906 907 /* 908 * Free the changed states if they are not in use, SME will be 909 * reallocated to the correct size on next use and we just 910 * allocate SVE now in case it is needed for use in streaming 911 * mode. 912 */ 913 sve_free(task); 914 sve_alloc(task, true); 915 916 if (free_sme) 917 sme_free(task); 918 919 out: 920 update_tsk_thread_flag(task, vec_vl_inherit_flag(type), 921 flags & PR_SVE_VL_INHERIT); 922 923 return 0; 924 } 925 926 /* 927 * Encode the current vector length and flags for return. 928 * This is only required for prctl(): ptrace has separate fields. 929 * SVE and SME use the same bits for _ONEXEC and _INHERIT. 930 * 931 * flags are as for vec_set_vector_length(). 932 */ 933 static int vec_prctl_status(enum vec_type type, unsigned long flags) 934 { 935 int ret; 936 937 if (flags & PR_SVE_SET_VL_ONEXEC) 938 ret = task_get_vl_onexec(current, type); 939 else 940 ret = task_get_vl(current, type); 941 942 if (test_thread_flag(vec_vl_inherit_flag(type))) 943 ret |= PR_SVE_VL_INHERIT; 944 945 return ret; 946 } 947 948 /* PR_SVE_SET_VL */ 949 int sve_set_current_vl(unsigned long arg) 950 { 951 unsigned long vl, flags; 952 int ret; 953 954 vl = arg & PR_SVE_VL_LEN_MASK; 955 flags = arg & ~vl; 956 957 if (!system_supports_sve() || is_compat_task()) 958 return -EINVAL; 959 960 ret = vec_set_vector_length(current, ARM64_VEC_SVE, vl, flags); 961 if (ret) 962 return ret; 963 964 return vec_prctl_status(ARM64_VEC_SVE, flags); 965 } 966 967 /* PR_SVE_GET_VL */ 968 int sve_get_current_vl(void) 969 { 970 if (!system_supports_sve() || is_compat_task()) 971 return -EINVAL; 972 973 return vec_prctl_status(ARM64_VEC_SVE, 0); 974 } 975 976 #ifdef CONFIG_ARM64_SME 977 /* PR_SME_SET_VL */ 978 int sme_set_current_vl(unsigned long arg) 979 { 980 unsigned long vl, flags; 981 int ret; 982 983 vl = arg & PR_SME_VL_LEN_MASK; 984 flags = arg & ~vl; 985 986 if (!system_supports_sme() || is_compat_task()) 987 return -EINVAL; 988 989 ret = vec_set_vector_length(current, ARM64_VEC_SME, vl, flags); 990 if (ret) 991 return ret; 992 993 return vec_prctl_status(ARM64_VEC_SME, flags); 994 } 995 996 /* PR_SME_GET_VL */ 997 int sme_get_current_vl(void) 998 { 999 if (!system_supports_sme() || is_compat_task()) 1000 return -EINVAL; 1001 1002 return vec_prctl_status(ARM64_VEC_SME, 0); 1003 } 1004 #endif /* CONFIG_ARM64_SME */ 1005 1006 static void vec_probe_vqs(struct vl_info *info, 1007 DECLARE_BITMAP(map, SVE_VQ_MAX)) 1008 { 1009 unsigned int vq, vl; 1010 1011 bitmap_zero(map, SVE_VQ_MAX); 1012 1013 for (vq = SVE_VQ_MAX; vq >= SVE_VQ_MIN; --vq) { 1014 write_vl(info->type, vq - 1); /* self-syncing */ 1015 1016 switch (info->type) { 1017 case ARM64_VEC_SVE: 1018 vl = sve_get_vl(); 1019 break; 1020 case ARM64_VEC_SME: 1021 vl = sme_get_vl(); 1022 break; 1023 default: 1024 vl = 0; 1025 break; 1026 } 1027 1028 /* Minimum VL identified? */ 1029 if (sve_vq_from_vl(vl) > vq) 1030 break; 1031 1032 vq = sve_vq_from_vl(vl); /* skip intervening lengths */ 1033 set_bit(__vq_to_bit(vq), map); 1034 } 1035 } 1036 1037 /* 1038 * Initialise the set of known supported VQs for the boot CPU. 1039 * This is called during kernel boot, before secondary CPUs are brought up. 1040 */ 1041 void __init vec_init_vq_map(enum vec_type type) 1042 { 1043 struct vl_info *info = &vl_info[type]; 1044 vec_probe_vqs(info, info->vq_map); 1045 bitmap_copy(info->vq_partial_map, info->vq_map, SVE_VQ_MAX); 1046 } 1047 1048 /* 1049 * If we haven't committed to the set of supported VQs yet, filter out 1050 * those not supported by the current CPU. 1051 * This function is called during the bring-up of early secondary CPUs only. 1052 */ 1053 void vec_update_vq_map(enum vec_type type) 1054 { 1055 struct vl_info *info = &vl_info[type]; 1056 DECLARE_BITMAP(tmp_map, SVE_VQ_MAX); 1057 1058 vec_probe_vqs(info, tmp_map); 1059 bitmap_and(info->vq_map, info->vq_map, tmp_map, SVE_VQ_MAX); 1060 bitmap_or(info->vq_partial_map, info->vq_partial_map, tmp_map, 1061 SVE_VQ_MAX); 1062 } 1063 1064 /* 1065 * Check whether the current CPU supports all VQs in the committed set. 1066 * This function is called during the bring-up of late secondary CPUs only. 1067 */ 1068 int vec_verify_vq_map(enum vec_type type) 1069 { 1070 struct vl_info *info = &vl_info[type]; 1071 DECLARE_BITMAP(tmp_map, SVE_VQ_MAX); 1072 unsigned long b; 1073 1074 vec_probe_vqs(info, tmp_map); 1075 1076 bitmap_complement(tmp_map, tmp_map, SVE_VQ_MAX); 1077 if (bitmap_intersects(tmp_map, info->vq_map, SVE_VQ_MAX)) { 1078 pr_warn("%s: cpu%d: Required vector length(s) missing\n", 1079 info->name, smp_processor_id()); 1080 return -EINVAL; 1081 } 1082 1083 if (!IS_ENABLED(CONFIG_KVM) || !is_hyp_mode_available()) 1084 return 0; 1085 1086 /* 1087 * For KVM, it is necessary to ensure that this CPU doesn't 1088 * support any vector length that guests may have probed as 1089 * unsupported. 1090 */ 1091 1092 /* Recover the set of supported VQs: */ 1093 bitmap_complement(tmp_map, tmp_map, SVE_VQ_MAX); 1094 /* Find VQs supported that are not globally supported: */ 1095 bitmap_andnot(tmp_map, tmp_map, info->vq_map, SVE_VQ_MAX); 1096 1097 /* Find the lowest such VQ, if any: */ 1098 b = find_last_bit(tmp_map, SVE_VQ_MAX); 1099 if (b >= SVE_VQ_MAX) 1100 return 0; /* no mismatches */ 1101 1102 /* 1103 * Mismatches above sve_max_virtualisable_vl are fine, since 1104 * no guest is allowed to configure ZCR_EL2.LEN to exceed this: 1105 */ 1106 if (sve_vl_from_vq(__bit_to_vq(b)) <= info->max_virtualisable_vl) { 1107 pr_warn("%s: cpu%d: Unsupported vector length(s) present\n", 1108 info->name, smp_processor_id()); 1109 return -EINVAL; 1110 } 1111 1112 return 0; 1113 } 1114 1115 static void __init sve_efi_setup(void) 1116 { 1117 int max_vl = 0; 1118 int i; 1119 1120 if (!IS_ENABLED(CONFIG_EFI)) 1121 return; 1122 1123 for (i = 0; i < ARRAY_SIZE(vl_info); i++) 1124 max_vl = max(vl_info[i].max_vl, max_vl); 1125 1126 /* 1127 * alloc_percpu() warns and prints a backtrace if this goes wrong. 1128 * This is evidence of a crippled system and we are returning void, 1129 * so no attempt is made to handle this situation here. 1130 */ 1131 if (!sve_vl_valid(max_vl)) 1132 goto fail; 1133 1134 efi_sve_state = __alloc_percpu( 1135 SVE_SIG_REGS_SIZE(sve_vq_from_vl(max_vl)), SVE_VQ_BYTES); 1136 if (!efi_sve_state) 1137 goto fail; 1138 1139 return; 1140 1141 fail: 1142 panic("Cannot allocate percpu memory for EFI SVE save/restore"); 1143 } 1144 1145 void cpu_enable_sve(const struct arm64_cpu_capabilities *__always_unused p) 1146 { 1147 write_sysreg(read_sysreg(CPACR_EL1) | CPACR_EL1_ZEN_EL1EN, CPACR_EL1); 1148 isb(); 1149 1150 write_sysreg_s(0, SYS_ZCR_EL1); 1151 } 1152 1153 void __init sve_setup(void) 1154 { 1155 struct vl_info *info = &vl_info[ARM64_VEC_SVE]; 1156 DECLARE_BITMAP(tmp_map, SVE_VQ_MAX); 1157 unsigned long b; 1158 int max_bit; 1159 1160 if (!system_supports_sve()) 1161 return; 1162 1163 /* 1164 * The SVE architecture mandates support for 128-bit vectors, 1165 * so sve_vq_map must have at least SVE_VQ_MIN set. 1166 * If something went wrong, at least try to patch it up: 1167 */ 1168 if (WARN_ON(!test_bit(__vq_to_bit(SVE_VQ_MIN), info->vq_map))) 1169 set_bit(__vq_to_bit(SVE_VQ_MIN), info->vq_map); 1170 1171 max_bit = find_first_bit(info->vq_map, SVE_VQ_MAX); 1172 info->max_vl = sve_vl_from_vq(__bit_to_vq(max_bit)); 1173 1174 /* 1175 * For the default VL, pick the maximum supported value <= 64. 1176 * VL == 64 is guaranteed not to grow the signal frame. 1177 */ 1178 set_sve_default_vl(find_supported_vector_length(ARM64_VEC_SVE, 64)); 1179 1180 bitmap_andnot(tmp_map, info->vq_partial_map, info->vq_map, 1181 SVE_VQ_MAX); 1182 1183 b = find_last_bit(tmp_map, SVE_VQ_MAX); 1184 if (b >= SVE_VQ_MAX) 1185 /* No non-virtualisable VLs found */ 1186 info->max_virtualisable_vl = SVE_VQ_MAX; 1187 else if (WARN_ON(b == SVE_VQ_MAX - 1)) 1188 /* No virtualisable VLs? This is architecturally forbidden. */ 1189 info->max_virtualisable_vl = SVE_VQ_MIN; 1190 else /* b + 1 < SVE_VQ_MAX */ 1191 info->max_virtualisable_vl = sve_vl_from_vq(__bit_to_vq(b + 1)); 1192 1193 if (info->max_virtualisable_vl > info->max_vl) 1194 info->max_virtualisable_vl = info->max_vl; 1195 1196 pr_info("%s: maximum available vector length %u bytes per vector\n", 1197 info->name, info->max_vl); 1198 pr_info("%s: default vector length %u bytes per vector\n", 1199 info->name, get_sve_default_vl()); 1200 1201 /* KVM decides whether to support mismatched systems. Just warn here: */ 1202 if (sve_max_virtualisable_vl() < sve_max_vl()) 1203 pr_warn("%s: unvirtualisable vector lengths present\n", 1204 info->name); 1205 1206 sve_efi_setup(); 1207 } 1208 1209 /* 1210 * Called from the put_task_struct() path, which cannot get here 1211 * unless dead_task is really dead and not schedulable. 1212 */ 1213 void fpsimd_release_task(struct task_struct *dead_task) 1214 { 1215 __sve_free(dead_task); 1216 sme_free(dead_task); 1217 } 1218 1219 #endif /* CONFIG_ARM64_SVE */ 1220 1221 #ifdef CONFIG_ARM64_SME 1222 1223 /* 1224 * Ensure that task->thread.sme_state is allocated and sufficiently large. 1225 * 1226 * This function should be used only in preparation for replacing 1227 * task->thread.sme_state with new data. The memory is always zeroed 1228 * here to prevent stale data from showing through: this is done in 1229 * the interest of testability and predictability, the architecture 1230 * guarantees that when ZA is enabled it will be zeroed. 1231 */ 1232 void sme_alloc(struct task_struct *task, bool flush) 1233 { 1234 if (task->thread.sme_state) { 1235 if (flush) 1236 memset(task->thread.sme_state, 0, 1237 sme_state_size(task)); 1238 return; 1239 } 1240 1241 /* This could potentially be up to 64K. */ 1242 task->thread.sme_state = 1243 kzalloc(sme_state_size(task), GFP_KERNEL); 1244 } 1245 1246 static void sme_free(struct task_struct *task) 1247 { 1248 kfree(task->thread.sme_state); 1249 task->thread.sme_state = NULL; 1250 } 1251 1252 void cpu_enable_sme(const struct arm64_cpu_capabilities *__always_unused p) 1253 { 1254 /* Set priority for all PEs to architecturally defined minimum */ 1255 write_sysreg_s(read_sysreg_s(SYS_SMPRI_EL1) & ~SMPRI_EL1_PRIORITY_MASK, 1256 SYS_SMPRI_EL1); 1257 1258 /* Allow SME in kernel */ 1259 write_sysreg(read_sysreg(CPACR_EL1) | CPACR_EL1_SMEN_EL1EN, CPACR_EL1); 1260 isb(); 1261 1262 /* Ensure all bits in SMCR are set to known values */ 1263 write_sysreg_s(0, SYS_SMCR_EL1); 1264 1265 /* Allow EL0 to access TPIDR2 */ 1266 write_sysreg(read_sysreg(SCTLR_EL1) | SCTLR_ELx_ENTP2, SCTLR_EL1); 1267 isb(); 1268 } 1269 1270 void cpu_enable_sme2(const struct arm64_cpu_capabilities *__always_unused p) 1271 { 1272 /* This must be enabled after SME */ 1273 BUILD_BUG_ON(ARM64_SME2 <= ARM64_SME); 1274 1275 /* Allow use of ZT0 */ 1276 write_sysreg_s(read_sysreg_s(SYS_SMCR_EL1) | SMCR_ELx_EZT0_MASK, 1277 SYS_SMCR_EL1); 1278 } 1279 1280 void cpu_enable_fa64(const struct arm64_cpu_capabilities *__always_unused p) 1281 { 1282 /* This must be enabled after SME */ 1283 BUILD_BUG_ON(ARM64_SME_FA64 <= ARM64_SME); 1284 1285 /* Allow use of FA64 */ 1286 write_sysreg_s(read_sysreg_s(SYS_SMCR_EL1) | SMCR_ELx_FA64_MASK, 1287 SYS_SMCR_EL1); 1288 } 1289 1290 void __init sme_setup(void) 1291 { 1292 struct vl_info *info = &vl_info[ARM64_VEC_SME]; 1293 int min_bit, max_bit; 1294 1295 if (!system_supports_sme()) 1296 return; 1297 1298 /* 1299 * SME doesn't require any particular vector length be 1300 * supported but it does require at least one. We should have 1301 * disabled the feature entirely while bringing up CPUs but 1302 * let's double check here. The bitmap is SVE_VQ_MAP sized for 1303 * sharing with SVE. 1304 */ 1305 WARN_ON(bitmap_empty(info->vq_map, SVE_VQ_MAX)); 1306 1307 min_bit = find_last_bit(info->vq_map, SVE_VQ_MAX); 1308 info->min_vl = sve_vl_from_vq(__bit_to_vq(min_bit)); 1309 1310 max_bit = find_first_bit(info->vq_map, SVE_VQ_MAX); 1311 info->max_vl = sve_vl_from_vq(__bit_to_vq(max_bit)); 1312 1313 WARN_ON(info->min_vl > info->max_vl); 1314 1315 /* 1316 * For the default VL, pick the maximum supported value <= 32 1317 * (256 bits) if there is one since this is guaranteed not to 1318 * grow the signal frame when in streaming mode, otherwise the 1319 * minimum available VL will be used. 1320 */ 1321 set_sme_default_vl(find_supported_vector_length(ARM64_VEC_SME, 32)); 1322 1323 pr_info("SME: minimum available vector length %u bytes per vector\n", 1324 info->min_vl); 1325 pr_info("SME: maximum available vector length %u bytes per vector\n", 1326 info->max_vl); 1327 pr_info("SME: default vector length %u bytes per vector\n", 1328 get_sme_default_vl()); 1329 } 1330 1331 void sme_suspend_exit(void) 1332 { 1333 u64 smcr = 0; 1334 1335 if (!system_supports_sme()) 1336 return; 1337 1338 if (system_supports_fa64()) 1339 smcr |= SMCR_ELx_FA64; 1340 if (system_supports_sme2()) 1341 smcr |= SMCR_ELx_EZT0; 1342 1343 write_sysreg_s(smcr, SYS_SMCR_EL1); 1344 write_sysreg_s(0, SYS_SMPRI_EL1); 1345 } 1346 1347 #endif /* CONFIG_ARM64_SME */ 1348 1349 static void sve_init_regs(void) 1350 { 1351 /* 1352 * Convert the FPSIMD state to SVE, zeroing all the state that 1353 * is not shared with FPSIMD. If (as is likely) the current 1354 * state is live in the registers then do this there and 1355 * update our metadata for the current task including 1356 * disabling the trap, otherwise update our in-memory copy. 1357 * We are guaranteed to not be in streaming mode, we can only 1358 * take a SVE trap when not in streaming mode and we can't be 1359 * in streaming mode when taking a SME trap. 1360 */ 1361 if (!test_thread_flag(TIF_FOREIGN_FPSTATE)) { 1362 unsigned long vq_minus_one = 1363 sve_vq_from_vl(task_get_sve_vl(current)) - 1; 1364 sve_set_vq(vq_minus_one); 1365 sve_flush_live(true, vq_minus_one); 1366 fpsimd_bind_task_to_cpu(); 1367 } else { 1368 fpsimd_to_sve(current); 1369 current->thread.fp_type = FP_STATE_SVE; 1370 } 1371 } 1372 1373 /* 1374 * Trapped SVE access 1375 * 1376 * Storage is allocated for the full SVE state, the current FPSIMD 1377 * register contents are migrated across, and the access trap is 1378 * disabled. 1379 * 1380 * TIF_SVE should be clear on entry: otherwise, fpsimd_restore_current_state() 1381 * would have disabled the SVE access trap for userspace during 1382 * ret_to_user, making an SVE access trap impossible in that case. 1383 */ 1384 void do_sve_acc(unsigned long esr, struct pt_regs *regs) 1385 { 1386 /* Even if we chose not to use SVE, the hardware could still trap: */ 1387 if (unlikely(!system_supports_sve()) || WARN_ON(is_compat_task())) { 1388 force_signal_inject(SIGILL, ILL_ILLOPC, regs->pc, 0); 1389 return; 1390 } 1391 1392 sve_alloc(current, true); 1393 if (!current->thread.sve_state) { 1394 force_sig(SIGKILL); 1395 return; 1396 } 1397 1398 get_cpu_fpsimd_context(); 1399 1400 if (test_and_set_thread_flag(TIF_SVE)) 1401 WARN_ON(1); /* SVE access shouldn't have trapped */ 1402 1403 /* 1404 * Even if the task can have used streaming mode we can only 1405 * generate SVE access traps in normal SVE mode and 1406 * transitioning out of streaming mode may discard any 1407 * streaming mode state. Always clear the high bits to avoid 1408 * any potential errors tracking what is properly initialised. 1409 */ 1410 sve_init_regs(); 1411 1412 put_cpu_fpsimd_context(); 1413 } 1414 1415 /* 1416 * Trapped SME access 1417 * 1418 * Storage is allocated for the full SVE and SME state, the current 1419 * FPSIMD register contents are migrated to SVE if SVE is not already 1420 * active, and the access trap is disabled. 1421 * 1422 * TIF_SME should be clear on entry: otherwise, fpsimd_restore_current_state() 1423 * would have disabled the SME access trap for userspace during 1424 * ret_to_user, making an SME access trap impossible in that case. 1425 */ 1426 void do_sme_acc(unsigned long esr, struct pt_regs *regs) 1427 { 1428 /* Even if we chose not to use SME, the hardware could still trap: */ 1429 if (unlikely(!system_supports_sme()) || WARN_ON(is_compat_task())) { 1430 force_signal_inject(SIGILL, ILL_ILLOPC, regs->pc, 0); 1431 return; 1432 } 1433 1434 /* 1435 * If this not a trap due to SME being disabled then something 1436 * is being used in the wrong mode, report as SIGILL. 1437 */ 1438 if (ESR_ELx_ISS(esr) != ESR_ELx_SME_ISS_SME_DISABLED) { 1439 force_signal_inject(SIGILL, ILL_ILLOPC, regs->pc, 0); 1440 return; 1441 } 1442 1443 sve_alloc(current, false); 1444 sme_alloc(current, true); 1445 if (!current->thread.sve_state || !current->thread.sme_state) { 1446 force_sig(SIGKILL); 1447 return; 1448 } 1449 1450 get_cpu_fpsimd_context(); 1451 1452 /* With TIF_SME userspace shouldn't generate any traps */ 1453 if (test_and_set_thread_flag(TIF_SME)) 1454 WARN_ON(1); 1455 1456 if (!test_thread_flag(TIF_FOREIGN_FPSTATE)) { 1457 unsigned long vq_minus_one = 1458 sve_vq_from_vl(task_get_sme_vl(current)) - 1; 1459 sme_set_vq(vq_minus_one); 1460 1461 fpsimd_bind_task_to_cpu(); 1462 } 1463 1464 put_cpu_fpsimd_context(); 1465 } 1466 1467 /* 1468 * Trapped FP/ASIMD access. 1469 */ 1470 void do_fpsimd_acc(unsigned long esr, struct pt_regs *regs) 1471 { 1472 /* Even if we chose not to use FPSIMD, the hardware could still trap: */ 1473 if (!system_supports_fpsimd()) { 1474 force_signal_inject(SIGILL, ILL_ILLOPC, regs->pc, 0); 1475 return; 1476 } 1477 1478 /* 1479 * When FPSIMD is enabled, we should never take a trap unless something 1480 * has gone very wrong. 1481 */ 1482 BUG(); 1483 } 1484 1485 /* 1486 * Raise a SIGFPE for the current process. 1487 */ 1488 void do_fpsimd_exc(unsigned long esr, struct pt_regs *regs) 1489 { 1490 unsigned int si_code = FPE_FLTUNK; 1491 1492 if (esr & ESR_ELx_FP_EXC_TFV) { 1493 if (esr & FPEXC_IOF) 1494 si_code = FPE_FLTINV; 1495 else if (esr & FPEXC_DZF) 1496 si_code = FPE_FLTDIV; 1497 else if (esr & FPEXC_OFF) 1498 si_code = FPE_FLTOVF; 1499 else if (esr & FPEXC_UFF) 1500 si_code = FPE_FLTUND; 1501 else if (esr & FPEXC_IXF) 1502 si_code = FPE_FLTRES; 1503 } 1504 1505 send_sig_fault(SIGFPE, si_code, 1506 (void __user *)instruction_pointer(regs), 1507 current); 1508 } 1509 1510 static void fpsimd_load_kernel_state(struct task_struct *task) 1511 { 1512 struct cpu_fp_state *last = this_cpu_ptr(&fpsimd_last_state); 1513 1514 /* 1515 * Elide the load if this CPU holds the most recent kernel mode 1516 * FPSIMD context of the current task. 1517 */ 1518 if (last->st == &task->thread.kernel_fpsimd_state && 1519 task->thread.kernel_fpsimd_cpu == smp_processor_id()) 1520 return; 1521 1522 fpsimd_load_state(&task->thread.kernel_fpsimd_state); 1523 } 1524 1525 static void fpsimd_save_kernel_state(struct task_struct *task) 1526 { 1527 struct cpu_fp_state cpu_fp_state = { 1528 .st = &task->thread.kernel_fpsimd_state, 1529 .to_save = FP_STATE_FPSIMD, 1530 }; 1531 1532 fpsimd_save_state(&task->thread.kernel_fpsimd_state); 1533 fpsimd_bind_state_to_cpu(&cpu_fp_state); 1534 1535 task->thread.kernel_fpsimd_cpu = smp_processor_id(); 1536 } 1537 1538 void fpsimd_thread_switch(struct task_struct *next) 1539 { 1540 bool wrong_task, wrong_cpu; 1541 1542 if (!system_supports_fpsimd()) 1543 return; 1544 1545 WARN_ON_ONCE(!irqs_disabled()); 1546 1547 /* Save unsaved fpsimd state, if any: */ 1548 if (test_thread_flag(TIF_KERNEL_FPSTATE)) 1549 fpsimd_save_kernel_state(current); 1550 else 1551 fpsimd_save_user_state(); 1552 1553 if (test_tsk_thread_flag(next, TIF_KERNEL_FPSTATE)) { 1554 fpsimd_load_kernel_state(next); 1555 set_tsk_thread_flag(next, TIF_FOREIGN_FPSTATE); 1556 } else { 1557 /* 1558 * Fix up TIF_FOREIGN_FPSTATE to correctly describe next's 1559 * state. For kernel threads, FPSIMD registers are never 1560 * loaded with user mode FPSIMD state and so wrong_task and 1561 * wrong_cpu will always be true. 1562 */ 1563 wrong_task = __this_cpu_read(fpsimd_last_state.st) != 1564 &next->thread.uw.fpsimd_state; 1565 wrong_cpu = next->thread.fpsimd_cpu != smp_processor_id(); 1566 1567 update_tsk_thread_flag(next, TIF_FOREIGN_FPSTATE, 1568 wrong_task || wrong_cpu); 1569 } 1570 } 1571 1572 static void fpsimd_flush_thread_vl(enum vec_type type) 1573 { 1574 int vl, supported_vl; 1575 1576 /* 1577 * Reset the task vector length as required. This is where we 1578 * ensure that all user tasks have a valid vector length 1579 * configured: no kernel task can become a user task without 1580 * an exec and hence a call to this function. By the time the 1581 * first call to this function is made, all early hardware 1582 * probing is complete, so __sve_default_vl should be valid. 1583 * If a bug causes this to go wrong, we make some noise and 1584 * try to fudge thread.sve_vl to a safe value here. 1585 */ 1586 vl = task_get_vl_onexec(current, type); 1587 if (!vl) 1588 vl = get_default_vl(type); 1589 1590 if (WARN_ON(!sve_vl_valid(vl))) 1591 vl = vl_info[type].min_vl; 1592 1593 supported_vl = find_supported_vector_length(type, vl); 1594 if (WARN_ON(supported_vl != vl)) 1595 vl = supported_vl; 1596 1597 task_set_vl(current, type, vl); 1598 1599 /* 1600 * If the task is not set to inherit, ensure that the vector 1601 * length will be reset by a subsequent exec: 1602 */ 1603 if (!test_thread_flag(vec_vl_inherit_flag(type))) 1604 task_set_vl_onexec(current, type, 0); 1605 } 1606 1607 void fpsimd_flush_thread(void) 1608 { 1609 void *sve_state = NULL; 1610 void *sme_state = NULL; 1611 1612 if (!system_supports_fpsimd()) 1613 return; 1614 1615 get_cpu_fpsimd_context(); 1616 1617 fpsimd_flush_task_state(current); 1618 memset(¤t->thread.uw.fpsimd_state, 0, 1619 sizeof(current->thread.uw.fpsimd_state)); 1620 1621 if (system_supports_sve()) { 1622 clear_thread_flag(TIF_SVE); 1623 1624 /* Defer kfree() while in atomic context */ 1625 sve_state = current->thread.sve_state; 1626 current->thread.sve_state = NULL; 1627 1628 fpsimd_flush_thread_vl(ARM64_VEC_SVE); 1629 } 1630 1631 if (system_supports_sme()) { 1632 clear_thread_flag(TIF_SME); 1633 1634 /* Defer kfree() while in atomic context */ 1635 sme_state = current->thread.sme_state; 1636 current->thread.sme_state = NULL; 1637 1638 fpsimd_flush_thread_vl(ARM64_VEC_SME); 1639 current->thread.svcr = 0; 1640 } 1641 1642 current->thread.fp_type = FP_STATE_FPSIMD; 1643 1644 put_cpu_fpsimd_context(); 1645 kfree(sve_state); 1646 kfree(sme_state); 1647 } 1648 1649 /* 1650 * Save the userland FPSIMD state of 'current' to memory, but only if the state 1651 * currently held in the registers does in fact belong to 'current' 1652 */ 1653 void fpsimd_preserve_current_state(void) 1654 { 1655 if (!system_supports_fpsimd()) 1656 return; 1657 1658 get_cpu_fpsimd_context(); 1659 fpsimd_save_user_state(); 1660 put_cpu_fpsimd_context(); 1661 } 1662 1663 /* 1664 * Like fpsimd_preserve_current_state(), but ensure that 1665 * current->thread.uw.fpsimd_state is updated so that it can be copied to 1666 * the signal frame. 1667 */ 1668 void fpsimd_signal_preserve_current_state(void) 1669 { 1670 fpsimd_preserve_current_state(); 1671 if (current->thread.fp_type == FP_STATE_SVE) 1672 sve_to_fpsimd(current); 1673 } 1674 1675 /* 1676 * Called by KVM when entering the guest. 1677 */ 1678 void fpsimd_kvm_prepare(void) 1679 { 1680 if (!system_supports_sve()) 1681 return; 1682 1683 /* 1684 * KVM does not save host SVE state since we can only enter 1685 * the guest from a syscall so the ABI means that only the 1686 * non-saved SVE state needs to be saved. If we have left 1687 * SVE enabled for performance reasons then update the task 1688 * state to be FPSIMD only. 1689 */ 1690 get_cpu_fpsimd_context(); 1691 1692 if (test_and_clear_thread_flag(TIF_SVE)) { 1693 sve_to_fpsimd(current); 1694 current->thread.fp_type = FP_STATE_FPSIMD; 1695 } 1696 1697 put_cpu_fpsimd_context(); 1698 } 1699 1700 /* 1701 * Associate current's FPSIMD context with this cpu 1702 * The caller must have ownership of the cpu FPSIMD context before calling 1703 * this function. 1704 */ 1705 static void fpsimd_bind_task_to_cpu(void) 1706 { 1707 struct cpu_fp_state *last = this_cpu_ptr(&fpsimd_last_state); 1708 1709 WARN_ON(!system_supports_fpsimd()); 1710 last->st = ¤t->thread.uw.fpsimd_state; 1711 last->sve_state = current->thread.sve_state; 1712 last->sme_state = current->thread.sme_state; 1713 last->sve_vl = task_get_sve_vl(current); 1714 last->sme_vl = task_get_sme_vl(current); 1715 last->svcr = ¤t->thread.svcr; 1716 last->fpmr = ¤t->thread.uw.fpmr; 1717 last->fp_type = ¤t->thread.fp_type; 1718 last->to_save = FP_STATE_CURRENT; 1719 current->thread.fpsimd_cpu = smp_processor_id(); 1720 1721 /* 1722 * Toggle SVE and SME trapping for userspace if needed, these 1723 * are serialsied by ret_to_user(). 1724 */ 1725 if (system_supports_sme()) { 1726 if (test_thread_flag(TIF_SME)) 1727 sme_user_enable(); 1728 else 1729 sme_user_disable(); 1730 } 1731 1732 if (system_supports_sve()) { 1733 if (test_thread_flag(TIF_SVE)) 1734 sve_user_enable(); 1735 else 1736 sve_user_disable(); 1737 } 1738 } 1739 1740 void fpsimd_bind_state_to_cpu(struct cpu_fp_state *state) 1741 { 1742 struct cpu_fp_state *last = this_cpu_ptr(&fpsimd_last_state); 1743 1744 WARN_ON(!system_supports_fpsimd()); 1745 WARN_ON(!in_softirq() && !irqs_disabled()); 1746 1747 *last = *state; 1748 } 1749 1750 /* 1751 * Load the userland FPSIMD state of 'current' from memory, but only if the 1752 * FPSIMD state already held in the registers is /not/ the most recent FPSIMD 1753 * state of 'current'. This is called when we are preparing to return to 1754 * userspace to ensure that userspace sees a good register state. 1755 */ 1756 void fpsimd_restore_current_state(void) 1757 { 1758 /* 1759 * TIF_FOREIGN_FPSTATE is set on the init task and copied by 1760 * arch_dup_task_struct() regardless of whether FP/SIMD is detected. 1761 * Thus user threads can have this set even when FP/SIMD hasn't been 1762 * detected. 1763 * 1764 * When FP/SIMD is detected, begin_new_exec() will set 1765 * TIF_FOREIGN_FPSTATE via flush_thread() -> fpsimd_flush_thread(), 1766 * and fpsimd_thread_switch() will set TIF_FOREIGN_FPSTATE when 1767 * switching tasks. We detect FP/SIMD before we exec the first user 1768 * process, ensuring this has TIF_FOREIGN_FPSTATE set and 1769 * do_notify_resume() will call fpsimd_restore_current_state() to 1770 * install the user FP/SIMD context. 1771 * 1772 * When FP/SIMD is not detected, nothing else will clear or set 1773 * TIF_FOREIGN_FPSTATE prior to the first return to userspace, and 1774 * we must clear TIF_FOREIGN_FPSTATE to avoid do_notify_resume() 1775 * looping forever calling fpsimd_restore_current_state(). 1776 */ 1777 if (!system_supports_fpsimd()) { 1778 clear_thread_flag(TIF_FOREIGN_FPSTATE); 1779 return; 1780 } 1781 1782 get_cpu_fpsimd_context(); 1783 1784 if (test_and_clear_thread_flag(TIF_FOREIGN_FPSTATE)) { 1785 task_fpsimd_load(); 1786 fpsimd_bind_task_to_cpu(); 1787 } 1788 1789 put_cpu_fpsimd_context(); 1790 } 1791 1792 /* 1793 * Load an updated userland FPSIMD state for 'current' from memory and set the 1794 * flag that indicates that the FPSIMD register contents are the most recent 1795 * FPSIMD state of 'current'. This is used by the signal code to restore the 1796 * register state when returning from a signal handler in FPSIMD only cases, 1797 * any SVE context will be discarded. 1798 */ 1799 void fpsimd_update_current_state(struct user_fpsimd_state const *state) 1800 { 1801 if (WARN_ON(!system_supports_fpsimd())) 1802 return; 1803 1804 get_cpu_fpsimd_context(); 1805 1806 current->thread.uw.fpsimd_state = *state; 1807 if (test_thread_flag(TIF_SVE)) 1808 fpsimd_to_sve(current); 1809 1810 task_fpsimd_load(); 1811 fpsimd_bind_task_to_cpu(); 1812 1813 clear_thread_flag(TIF_FOREIGN_FPSTATE); 1814 1815 put_cpu_fpsimd_context(); 1816 } 1817 1818 /* 1819 * Invalidate live CPU copies of task t's FPSIMD state 1820 * 1821 * This function may be called with preemption enabled. The barrier() 1822 * ensures that the assignment to fpsimd_cpu is visible to any 1823 * preemption/softirq that could race with set_tsk_thread_flag(), so 1824 * that TIF_FOREIGN_FPSTATE cannot be spuriously re-cleared. 1825 * 1826 * The final barrier ensures that TIF_FOREIGN_FPSTATE is seen set by any 1827 * subsequent code. 1828 */ 1829 void fpsimd_flush_task_state(struct task_struct *t) 1830 { 1831 t->thread.fpsimd_cpu = NR_CPUS; 1832 /* 1833 * If we don't support fpsimd, bail out after we have 1834 * reset the fpsimd_cpu for this task and clear the 1835 * FPSTATE. 1836 */ 1837 if (!system_supports_fpsimd()) 1838 return; 1839 barrier(); 1840 set_tsk_thread_flag(t, TIF_FOREIGN_FPSTATE); 1841 1842 barrier(); 1843 } 1844 1845 /* 1846 * Invalidate any task's FPSIMD state that is present on this cpu. 1847 * The FPSIMD context should be acquired with get_cpu_fpsimd_context() 1848 * before calling this function. 1849 */ 1850 static void fpsimd_flush_cpu_state(void) 1851 { 1852 WARN_ON(!system_supports_fpsimd()); 1853 __this_cpu_write(fpsimd_last_state.st, NULL); 1854 1855 /* 1856 * Leaving streaming mode enabled will cause issues for any kernel 1857 * NEON and leaving streaming mode or ZA enabled may increase power 1858 * consumption. 1859 */ 1860 if (system_supports_sme()) 1861 sme_smstop(); 1862 1863 set_thread_flag(TIF_FOREIGN_FPSTATE); 1864 } 1865 1866 /* 1867 * Save the FPSIMD state to memory and invalidate cpu view. 1868 * This function must be called with preemption disabled. 1869 */ 1870 void fpsimd_save_and_flush_cpu_state(void) 1871 { 1872 unsigned long flags; 1873 1874 if (!system_supports_fpsimd()) 1875 return; 1876 WARN_ON(preemptible()); 1877 local_irq_save(flags); 1878 fpsimd_save_user_state(); 1879 fpsimd_flush_cpu_state(); 1880 local_irq_restore(flags); 1881 } 1882 1883 #ifdef CONFIG_KERNEL_MODE_NEON 1884 1885 /* 1886 * Kernel-side NEON support functions 1887 */ 1888 1889 /* 1890 * kernel_neon_begin(): obtain the CPU FPSIMD registers for use by the calling 1891 * context 1892 * 1893 * Must not be called unless may_use_simd() returns true. 1894 * Task context in the FPSIMD registers is saved back to memory as necessary. 1895 * 1896 * A matching call to kernel_neon_end() must be made before returning from the 1897 * calling context. 1898 * 1899 * The caller may freely use the FPSIMD registers until kernel_neon_end() is 1900 * called. 1901 */ 1902 void kernel_neon_begin(void) 1903 { 1904 if (WARN_ON(!system_supports_fpsimd())) 1905 return; 1906 1907 BUG_ON(!may_use_simd()); 1908 1909 get_cpu_fpsimd_context(); 1910 1911 /* Save unsaved fpsimd state, if any: */ 1912 if (test_thread_flag(TIF_KERNEL_FPSTATE)) { 1913 BUG_ON(IS_ENABLED(CONFIG_PREEMPT_RT) || !in_serving_softirq()); 1914 fpsimd_save_kernel_state(current); 1915 } else { 1916 fpsimd_save_user_state(); 1917 1918 /* 1919 * Set the thread flag so that the kernel mode FPSIMD state 1920 * will be context switched along with the rest of the task 1921 * state. 1922 * 1923 * On non-PREEMPT_RT, softirqs may interrupt task level kernel 1924 * mode FPSIMD, but the task will not be preemptible so setting 1925 * TIF_KERNEL_FPSTATE for those would be both wrong (as it 1926 * would mark the task context FPSIMD state as requiring a 1927 * context switch) and unnecessary. 1928 * 1929 * On PREEMPT_RT, softirqs are serviced from a separate thread, 1930 * which is scheduled as usual, and this guarantees that these 1931 * softirqs are not interrupting use of the FPSIMD in kernel 1932 * mode in task context. So in this case, setting the flag here 1933 * is always appropriate. 1934 */ 1935 if (IS_ENABLED(CONFIG_PREEMPT_RT) || !in_serving_softirq()) 1936 set_thread_flag(TIF_KERNEL_FPSTATE); 1937 } 1938 1939 /* Invalidate any task state remaining in the fpsimd regs: */ 1940 fpsimd_flush_cpu_state(); 1941 1942 put_cpu_fpsimd_context(); 1943 } 1944 EXPORT_SYMBOL_GPL(kernel_neon_begin); 1945 1946 /* 1947 * kernel_neon_end(): give the CPU FPSIMD registers back to the current task 1948 * 1949 * Must be called from a context in which kernel_neon_begin() was previously 1950 * called, with no call to kernel_neon_end() in the meantime. 1951 * 1952 * The caller must not use the FPSIMD registers after this function is called, 1953 * unless kernel_neon_begin() is called again in the meantime. 1954 */ 1955 void kernel_neon_end(void) 1956 { 1957 if (!system_supports_fpsimd()) 1958 return; 1959 1960 /* 1961 * If we are returning from a nested use of kernel mode FPSIMD, restore 1962 * the task context kernel mode FPSIMD state. This can only happen when 1963 * running in softirq context on non-PREEMPT_RT. 1964 */ 1965 if (!IS_ENABLED(CONFIG_PREEMPT_RT) && in_serving_softirq() && 1966 test_thread_flag(TIF_KERNEL_FPSTATE)) 1967 fpsimd_load_kernel_state(current); 1968 else 1969 clear_thread_flag(TIF_KERNEL_FPSTATE); 1970 } 1971 EXPORT_SYMBOL_GPL(kernel_neon_end); 1972 1973 #ifdef CONFIG_EFI 1974 1975 static DEFINE_PER_CPU(struct user_fpsimd_state, efi_fpsimd_state); 1976 static DEFINE_PER_CPU(bool, efi_fpsimd_state_used); 1977 static DEFINE_PER_CPU(bool, efi_sve_state_used); 1978 static DEFINE_PER_CPU(bool, efi_sm_state); 1979 1980 /* 1981 * EFI runtime services support functions 1982 * 1983 * The ABI for EFI runtime services allows EFI to use FPSIMD during the call. 1984 * This means that for EFI (and only for EFI), we have to assume that FPSIMD 1985 * is always used rather than being an optional accelerator. 1986 * 1987 * These functions provide the necessary support for ensuring FPSIMD 1988 * save/restore in the contexts from which EFI is used. 1989 * 1990 * Do not use them for any other purpose -- if tempted to do so, you are 1991 * either doing something wrong or you need to propose some refactoring. 1992 */ 1993 1994 /* 1995 * __efi_fpsimd_begin(): prepare FPSIMD for making an EFI runtime services call 1996 */ 1997 void __efi_fpsimd_begin(void) 1998 { 1999 if (!system_supports_fpsimd()) 2000 return; 2001 2002 WARN_ON(preemptible()); 2003 2004 if (may_use_simd()) { 2005 kernel_neon_begin(); 2006 } else { 2007 /* 2008 * If !efi_sve_state, SVE can't be in use yet and doesn't need 2009 * preserving: 2010 */ 2011 if (system_supports_sve() && likely(efi_sve_state)) { 2012 char *sve_state = this_cpu_ptr(efi_sve_state); 2013 bool ffr = true; 2014 u64 svcr; 2015 2016 __this_cpu_write(efi_sve_state_used, true); 2017 2018 if (system_supports_sme()) { 2019 svcr = read_sysreg_s(SYS_SVCR); 2020 2021 __this_cpu_write(efi_sm_state, 2022 svcr & SVCR_SM_MASK); 2023 2024 /* 2025 * Unless we have FA64 FFR does not 2026 * exist in streaming mode. 2027 */ 2028 if (!system_supports_fa64()) 2029 ffr = !(svcr & SVCR_SM_MASK); 2030 } 2031 2032 sve_save_state(sve_state + sve_ffr_offset(sve_max_vl()), 2033 &this_cpu_ptr(&efi_fpsimd_state)->fpsr, 2034 ffr); 2035 2036 if (system_supports_sme()) 2037 sysreg_clear_set_s(SYS_SVCR, 2038 SVCR_SM_MASK, 0); 2039 2040 } else { 2041 fpsimd_save_state(this_cpu_ptr(&efi_fpsimd_state)); 2042 } 2043 2044 __this_cpu_write(efi_fpsimd_state_used, true); 2045 } 2046 } 2047 2048 /* 2049 * __efi_fpsimd_end(): clean up FPSIMD after an EFI runtime services call 2050 */ 2051 void __efi_fpsimd_end(void) 2052 { 2053 if (!system_supports_fpsimd()) 2054 return; 2055 2056 if (!__this_cpu_xchg(efi_fpsimd_state_used, false)) { 2057 kernel_neon_end(); 2058 } else { 2059 if (system_supports_sve() && 2060 likely(__this_cpu_read(efi_sve_state_used))) { 2061 char const *sve_state = this_cpu_ptr(efi_sve_state); 2062 bool ffr = true; 2063 2064 /* 2065 * Restore streaming mode; EFI calls are 2066 * normal function calls so should not return in 2067 * streaming mode. 2068 */ 2069 if (system_supports_sme()) { 2070 if (__this_cpu_read(efi_sm_state)) { 2071 sysreg_clear_set_s(SYS_SVCR, 2072 0, 2073 SVCR_SM_MASK); 2074 2075 /* 2076 * Unless we have FA64 FFR does not 2077 * exist in streaming mode. 2078 */ 2079 if (!system_supports_fa64()) 2080 ffr = false; 2081 } 2082 } 2083 2084 sve_load_state(sve_state + sve_ffr_offset(sve_max_vl()), 2085 &this_cpu_ptr(&efi_fpsimd_state)->fpsr, 2086 ffr); 2087 2088 __this_cpu_write(efi_sve_state_used, false); 2089 } else { 2090 fpsimd_load_state(this_cpu_ptr(&efi_fpsimd_state)); 2091 } 2092 } 2093 } 2094 2095 #endif /* CONFIG_EFI */ 2096 2097 #endif /* CONFIG_KERNEL_MODE_NEON */ 2098 2099 #ifdef CONFIG_CPU_PM 2100 static int fpsimd_cpu_pm_notifier(struct notifier_block *self, 2101 unsigned long cmd, void *v) 2102 { 2103 switch (cmd) { 2104 case CPU_PM_ENTER: 2105 fpsimd_save_and_flush_cpu_state(); 2106 break; 2107 case CPU_PM_EXIT: 2108 break; 2109 case CPU_PM_ENTER_FAILED: 2110 default: 2111 return NOTIFY_DONE; 2112 } 2113 return NOTIFY_OK; 2114 } 2115 2116 static struct notifier_block fpsimd_cpu_pm_notifier_block = { 2117 .notifier_call = fpsimd_cpu_pm_notifier, 2118 }; 2119 2120 static void __init fpsimd_pm_init(void) 2121 { 2122 cpu_pm_register_notifier(&fpsimd_cpu_pm_notifier_block); 2123 } 2124 2125 #else 2126 static inline void fpsimd_pm_init(void) { } 2127 #endif /* CONFIG_CPU_PM */ 2128 2129 #ifdef CONFIG_HOTPLUG_CPU 2130 static int fpsimd_cpu_dead(unsigned int cpu) 2131 { 2132 per_cpu(fpsimd_last_state.st, cpu) = NULL; 2133 return 0; 2134 } 2135 2136 static inline void fpsimd_hotplug_init(void) 2137 { 2138 cpuhp_setup_state_nocalls(CPUHP_ARM64_FPSIMD_DEAD, "arm64/fpsimd:dead", 2139 NULL, fpsimd_cpu_dead); 2140 } 2141 2142 #else 2143 static inline void fpsimd_hotplug_init(void) { } 2144 #endif 2145 2146 void cpu_enable_fpsimd(const struct arm64_cpu_capabilities *__always_unused p) 2147 { 2148 unsigned long enable = CPACR_EL1_FPEN_EL1EN | CPACR_EL1_FPEN_EL0EN; 2149 write_sysreg(read_sysreg(CPACR_EL1) | enable, CPACR_EL1); 2150 isb(); 2151 } 2152 2153 /* 2154 * FP/SIMD support code initialisation. 2155 */ 2156 static int __init fpsimd_init(void) 2157 { 2158 if (cpu_have_named_feature(FP)) { 2159 fpsimd_pm_init(); 2160 fpsimd_hotplug_init(); 2161 } else { 2162 pr_notice("Floating-point is not implemented\n"); 2163 } 2164 2165 if (!cpu_have_named_feature(ASIMD)) 2166 pr_notice("Advanced SIMD is not implemented\n"); 2167 2168 2169 sve_sysctl_init(); 2170 sme_sysctl_init(); 2171 2172 return 0; 2173 } 2174 core_initcall(fpsimd_init); 2175