1 // SPDX-License-Identifier: GPL-2.0-only
3 * FP/SIMD context switching and fault handling
5 * Copyright (C) 2012 ARM Ltd.
6 * Author: Catalin Marinas <catalin.marinas@arm.com>
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>
36 #include <asm/exception.h>
37 #include <asm/fpsimd.h>
38 #include <asm/cpufeature.h>
39 #include <asm/cputype.h>
41 #include <asm/processor.h>
43 #include <asm/sigcontext.h>
44 #include <asm/sysreg.h>
45 #include <asm/traps.h>
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)
56 * (Note: in this discussion, statements about FPSIMD apply equally to SVE.)
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.
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.
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.
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.
88 * In order to allow softirq handlers to use FPSIMD, kernel_neon_begin() may
89 * save the task's FPSIMD context back to task_struct from softirq context.
90 * To prevent this from racing with the manipulation of the task's FPSIMD state
91 * from task context and thereby corrupting the state, it is necessary to
92 * protect any manipulation of a task's fpsimd_state or TIF_FOREIGN_FPSTATE
93 * flag with {, __}get_cpu_fpsimd_context(). This will still allow softirqs to
94 * run but prevent them to use FPSIMD.
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;
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;
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
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;
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.
122 static DEFINE_PER_CPU(struct cpu_fp_state, fpsimd_last_state);
124 __ro_after_init struct vl_info vl_info[ARM64_VEC_MAX] = {
125 #ifdef CONFIG_ARM64_SVE
127 .type = ARM64_VEC_SVE,
129 .min_vl = SVE_VL_MIN,
130 .max_vl = SVE_VL_MIN,
131 .max_virtualisable_vl = SVE_VL_MIN,
134 #ifdef CONFIG_ARM64_SME
136 .type = ARM64_VEC_SME,
142 static unsigned int vec_vl_inherit_flag(enum vec_type type)
146 return TIF_SVE_VL_INHERIT;
148 return TIF_SME_VL_INHERIT;
156 int __default_vl; /* Default VL for tasks */
159 static struct vl_config vl_config[ARM64_VEC_MAX];
161 static inline int get_default_vl(enum vec_type type)
163 return READ_ONCE(vl_config[type].__default_vl);
166 #ifdef CONFIG_ARM64_SVE
168 static inline int get_sve_default_vl(void)
170 return get_default_vl(ARM64_VEC_SVE);
173 static inline void set_default_vl(enum vec_type type, int val)
175 WRITE_ONCE(vl_config[type].__default_vl, val);
178 static inline void set_sve_default_vl(int val)
180 set_default_vl(ARM64_VEC_SVE, val);
183 static void __percpu *efi_sve_state;
185 #else /* ! CONFIG_ARM64_SVE */
187 /* Dummy declaration for code that will be optimised out: */
188 extern void __percpu *efi_sve_state;
190 #endif /* ! CONFIG_ARM64_SVE */
192 #ifdef CONFIG_ARM64_SME
194 static int get_sme_default_vl(void)
196 return get_default_vl(ARM64_VEC_SME);
199 static void set_sme_default_vl(int val)
201 set_default_vl(ARM64_VEC_SME, val);
204 static void sme_free(struct task_struct *);
208 static inline void sme_free(struct task_struct *t) { }
212 DEFINE_PER_CPU(bool, fpsimd_context_busy);
213 EXPORT_PER_CPU_SYMBOL(fpsimd_context_busy);
215 static void fpsimd_bind_task_to_cpu(void);
217 static void __get_cpu_fpsimd_context(void)
219 bool busy = __this_cpu_xchg(fpsimd_context_busy, true);
225 * Claim ownership of the CPU FPSIMD context for use by the calling context.
227 * The caller may freely manipulate the FPSIMD context metadata until
228 * put_cpu_fpsimd_context() is called.
230 * The double-underscore version must only be called if you know the task
231 * can't be preempted.
233 * On RT kernels local_bh_disable() is not sufficient because it only
234 * serializes soft interrupt related sections via a local lock, but stays
235 * preemptible. Disabling preemption is the right choice here as bottom
236 * half processing is always in thread context on RT kernels so it
237 * implicitly prevents bottom half processing as well.
239 static void get_cpu_fpsimd_context(void)
241 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
245 __get_cpu_fpsimd_context();
248 static void __put_cpu_fpsimd_context(void)
250 bool busy = __this_cpu_xchg(fpsimd_context_busy, false);
252 WARN_ON(!busy); /* No matching get_cpu_fpsimd_context()? */
256 * Release the CPU FPSIMD context.
258 * Must be called from a context in which get_cpu_fpsimd_context() was
259 * previously called, with no call to put_cpu_fpsimd_context() in the
262 static void put_cpu_fpsimd_context(void)
264 __put_cpu_fpsimd_context();
265 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
271 static bool have_cpu_fpsimd_context(void)
273 return !preemptible() && __this_cpu_read(fpsimd_context_busy);
276 unsigned int task_get_vl(const struct task_struct *task, enum vec_type type)
278 return task->thread.vl[type];
281 void task_set_vl(struct task_struct *task, enum vec_type type,
284 task->thread.vl[type] = vl;
287 unsigned int task_get_vl_onexec(const struct task_struct *task,
290 return task->thread.vl_onexec[type];
293 void task_set_vl_onexec(struct task_struct *task, enum vec_type type,
296 task->thread.vl_onexec[type] = vl;
300 * TIF_SME controls whether a task can use SME without trapping while
301 * in userspace, when TIF_SME is set then we must have storage
302 * allocated in sve_state and sme_state to store the contents of both ZA
303 * and the SVE registers for both streaming and non-streaming modes.
305 * If both SVCR.ZA and SVCR.SM are disabled then at any point we
306 * may disable TIF_SME and reenable traps.
311 * TIF_SVE controls whether a task can use SVE without trapping while
312 * in userspace, and also (together with TIF_SME) the way a task's
313 * FPSIMD/SVE state is stored in thread_struct.
315 * The kernel uses this flag to track whether a user task is actively
316 * using SVE, and therefore whether full SVE register state needs to
317 * be tracked. If not, the cheaper FPSIMD context handling code can
318 * be used instead of the more costly SVE equivalents.
320 * * TIF_SVE or SVCR.SM set:
322 * The task can execute SVE instructions while in userspace without
323 * trapping to the kernel.
325 * During any syscall, the kernel may optionally clear TIF_SVE and
326 * discard the vector state except for the FPSIMD subset.
330 * An attempt by the user task to execute an SVE instruction causes
331 * do_sve_acc() to be called, which does some preparation and then
334 * During any syscall, the kernel may optionally clear TIF_SVE and
335 * discard the vector state except for the FPSIMD subset.
337 * The data will be stored in one of two formats:
339 * * FPSIMD only - FP_STATE_FPSIMD:
341 * When the FPSIMD only state stored task->thread.fp_type is set to
342 * FP_STATE_FPSIMD, the FPSIMD registers V0-V31 are encoded in
343 * task->thread.uw.fpsimd_state; bits [max : 128] for each of Z0-Z31 are
344 * logically zero but not stored anywhere; P0-P15 and FFR are not
345 * stored and have unspecified values from userspace's point of
346 * view. For hygiene purposes, the kernel zeroes them on next use,
347 * but userspace is discouraged from relying on this.
349 * task->thread.sve_state does not need to be non-NULL, valid or any
350 * particular size: it must not be dereferenced and any data stored
351 * there should be considered stale and not referenced.
353 * * SVE state - FP_STATE_SVE:
355 * When the full SVE state is stored task->thread.fp_type is set to
356 * FP_STATE_SVE and Z0-Z31 (incorporating Vn in bits[127:0] or the
357 * corresponding Zn), P0-P15 and FFR are encoded in in
358 * task->thread.sve_state, formatted appropriately for vector
359 * length task->thread.sve_vl or, if SVCR.SM is set,
360 * task->thread.sme_vl. The storage for the vector registers in
361 * task->thread.uw.fpsimd_state should be ignored.
363 * task->thread.sve_state must point to a valid buffer at least
364 * sve_state_size(task) bytes in size. The data stored in
365 * task->thread.uw.fpsimd_state.vregs should be considered stale
366 * and not referenced.
368 * * FPSR and FPCR are always stored in task->thread.uw.fpsimd_state
369 * irrespective of whether TIF_SVE is clear or set, since these are
370 * not vector length dependent.
374 * Update current's FPSIMD/SVE registers from thread_struct.
376 * This function should be called only when the FPSIMD/SVE state in
377 * thread_struct is known to be up to date, when preparing to enter
380 static void task_fpsimd_load(void)
382 bool restore_sve_regs = false;
385 WARN_ON(!system_supports_fpsimd());
386 WARN_ON(!have_cpu_fpsimd_context());
388 if (system_supports_sve() || system_supports_sme()) {
389 switch (current->thread.fp_type) {
390 case FP_STATE_FPSIMD:
391 /* Stop tracking SVE for this task until next use. */
392 if (test_and_clear_thread_flag(TIF_SVE))
396 if (!thread_sm_enabled(¤t->thread) &&
397 !WARN_ON_ONCE(!test_and_set_thread_flag(TIF_SVE)))
400 if (test_thread_flag(TIF_SVE))
401 sve_set_vq(sve_vq_from_vl(task_get_sve_vl(current)) - 1);
403 restore_sve_regs = true;
408 * This indicates either a bug in
409 * fpsimd_save() or memory corruption, we
410 * should always record an explicit format
411 * when we save. We always at least have the
412 * memory allocated for FPSMID registers so
413 * try that and hope for the best.
416 clear_thread_flag(TIF_SVE);
421 /* Restore SME, override SVE register configuration if needed */
422 if (system_supports_sme()) {
423 unsigned long sme_vl = task_get_sme_vl(current);
425 /* Ensure VL is set up for restoring data */
426 if (test_thread_flag(TIF_SME))
427 sme_set_vq(sve_vq_from_vl(sme_vl) - 1);
429 write_sysreg_s(current->thread.svcr, SYS_SVCR);
431 if (thread_za_enabled(¤t->thread))
432 sme_load_state(current->thread.sme_state,
433 system_supports_sme2());
435 if (thread_sm_enabled(¤t->thread))
436 restore_ffr = system_supports_fa64();
439 if (restore_sve_regs) {
440 WARN_ON_ONCE(current->thread.fp_type != FP_STATE_SVE);
441 sve_load_state(sve_pffr(¤t->thread),
442 ¤t->thread.uw.fpsimd_state.fpsr,
445 WARN_ON_ONCE(current->thread.fp_type != FP_STATE_FPSIMD);
446 fpsimd_load_state(¤t->thread.uw.fpsimd_state);
451 * Ensure FPSIMD/SVE storage in memory for the loaded context is up to
452 * date with respect to the CPU registers. Note carefully that the
453 * current context is the context last bound to the CPU stored in
454 * last, if KVM is involved this may be the guest VM context rather
455 * than the host thread for the VM pointed to by current. This means
456 * that we must always reference the state storage via last rather
457 * than via current, if we are saving KVM state then it will have
458 * ensured that the type of registers to save is set in last->to_save.
460 static void fpsimd_save(void)
462 struct cpu_fp_state const *last =
463 this_cpu_ptr(&fpsimd_last_state);
464 /* set by fpsimd_bind_task_to_cpu() or fpsimd_bind_state_to_cpu() */
465 bool save_sve_regs = false;
469 WARN_ON(!system_supports_fpsimd());
470 WARN_ON(!have_cpu_fpsimd_context());
472 if (test_thread_flag(TIF_FOREIGN_FPSTATE))
476 * If a task is in a syscall the ABI allows us to only
477 * preserve the state shared with FPSIMD so don't bother
478 * saving the full SVE state in that case.
480 if ((last->to_save == FP_STATE_CURRENT && test_thread_flag(TIF_SVE) &&
481 !in_syscall(current_pt_regs())) ||
482 last->to_save == FP_STATE_SVE) {
483 save_sve_regs = true;
488 if (system_supports_sme()) {
489 u64 *svcr = last->svcr;
491 *svcr = read_sysreg_s(SYS_SVCR);
493 if (*svcr & SVCR_ZA_MASK)
494 sme_save_state(last->sme_state,
495 system_supports_sme2());
497 /* If we are in streaming mode override regular SVE. */
498 if (*svcr & SVCR_SM_MASK) {
499 save_sve_regs = true;
500 save_ffr = system_supports_fa64();
505 if (IS_ENABLED(CONFIG_ARM64_SVE) && save_sve_regs) {
506 /* Get the configured VL from RDVL, will account for SM */
507 if (WARN_ON(sve_get_vl() != vl)) {
509 * Can't save the user regs, so current would
510 * re-enter user with corrupt state.
511 * There's no way to recover, so kill it:
513 force_signal_inject(SIGKILL, SI_KERNEL, 0, 0);
517 sve_save_state((char *)last->sve_state +
519 &last->st->fpsr, save_ffr);
520 *last->fp_type = FP_STATE_SVE;
522 fpsimd_save_state(last->st);
523 *last->fp_type = FP_STATE_FPSIMD;
528 * All vector length selection from userspace comes through here.
529 * We're on a slow path, so some sanity-checks are included.
530 * If things go wrong there's a bug somewhere, but try to fall back to a
533 static unsigned int find_supported_vector_length(enum vec_type type,
536 struct vl_info *info = &vl_info[type];
538 int max_vl = info->max_vl;
540 if (WARN_ON(!sve_vl_valid(vl)))
543 if (WARN_ON(!sve_vl_valid(max_vl)))
544 max_vl = info->min_vl;
548 if (vl < info->min_vl)
551 bit = find_next_bit(info->vq_map, SVE_VQ_MAX,
552 __vq_to_bit(sve_vq_from_vl(vl)));
553 return sve_vl_from_vq(__bit_to_vq(bit));
556 #if defined(CONFIG_ARM64_SVE) && defined(CONFIG_SYSCTL)
558 static int vec_proc_do_default_vl(struct ctl_table *table, int write,
559 void *buffer, size_t *lenp, loff_t *ppos)
561 struct vl_info *info = table->extra1;
562 enum vec_type type = info->type;
564 int vl = get_default_vl(type);
565 struct ctl_table tmp_table = {
567 .maxlen = sizeof(vl),
570 ret = proc_dointvec(&tmp_table, write, buffer, lenp, ppos);
574 /* Writing -1 has the special meaning "set to max": */
578 if (!sve_vl_valid(vl))
581 set_default_vl(type, find_supported_vector_length(type, vl));
585 static struct ctl_table sve_default_vl_table[] = {
587 .procname = "sve_default_vector_length",
589 .proc_handler = vec_proc_do_default_vl,
590 .extra1 = &vl_info[ARM64_VEC_SVE],
595 static int __init sve_sysctl_init(void)
597 if (system_supports_sve())
598 if (!register_sysctl("abi", sve_default_vl_table))
604 #else /* ! (CONFIG_ARM64_SVE && CONFIG_SYSCTL) */
605 static int __init sve_sysctl_init(void) { return 0; }
606 #endif /* ! (CONFIG_ARM64_SVE && CONFIG_SYSCTL) */
608 #if defined(CONFIG_ARM64_SME) && defined(CONFIG_SYSCTL)
609 static struct ctl_table sme_default_vl_table[] = {
611 .procname = "sme_default_vector_length",
613 .proc_handler = vec_proc_do_default_vl,
614 .extra1 = &vl_info[ARM64_VEC_SME],
619 static int __init sme_sysctl_init(void)
621 if (system_supports_sme())
622 if (!register_sysctl("abi", sme_default_vl_table))
628 #else /* ! (CONFIG_ARM64_SME && CONFIG_SYSCTL) */
629 static int __init sme_sysctl_init(void) { return 0; }
630 #endif /* ! (CONFIG_ARM64_SME && CONFIG_SYSCTL) */
632 #define ZREG(sve_state, vq, n) ((char *)(sve_state) + \
633 (SVE_SIG_ZREG_OFFSET(vq, n) - SVE_SIG_REGS_OFFSET))
635 #ifdef CONFIG_CPU_BIG_ENDIAN
636 static __uint128_t arm64_cpu_to_le128(__uint128_t x)
639 u64 b = swab64(x >> 64);
641 return ((__uint128_t)a << 64) | b;
644 static __uint128_t arm64_cpu_to_le128(__uint128_t x)
650 #define arm64_le128_to_cpu(x) arm64_cpu_to_le128(x)
652 static void __fpsimd_to_sve(void *sst, struct user_fpsimd_state const *fst,
658 for (i = 0; i < SVE_NUM_ZREGS; ++i) {
659 p = (__uint128_t *)ZREG(sst, vq, i);
660 *p = arm64_cpu_to_le128(fst->vregs[i]);
665 * Transfer the FPSIMD state in task->thread.uw.fpsimd_state to
666 * task->thread.sve_state.
668 * Task can be a non-runnable task, or current. In the latter case,
669 * the caller must have ownership of the cpu FPSIMD context before calling
671 * task->thread.sve_state must point to at least sve_state_size(task)
672 * bytes of allocated kernel memory.
673 * task->thread.uw.fpsimd_state must be up to date before calling this
676 static void fpsimd_to_sve(struct task_struct *task)
679 void *sst = task->thread.sve_state;
680 struct user_fpsimd_state const *fst = &task->thread.uw.fpsimd_state;
682 if (!system_supports_sve())
685 vq = sve_vq_from_vl(thread_get_cur_vl(&task->thread));
686 __fpsimd_to_sve(sst, fst, vq);
690 * Transfer the SVE state in task->thread.sve_state to
691 * task->thread.uw.fpsimd_state.
693 * Task can be a non-runnable task, or current. In the latter case,
694 * the caller must have ownership of the cpu FPSIMD context before calling
696 * task->thread.sve_state must point to at least sve_state_size(task)
697 * bytes of allocated kernel memory.
698 * task->thread.sve_state must be up to date before calling this function.
700 static void sve_to_fpsimd(struct task_struct *task)
703 void const *sst = task->thread.sve_state;
704 struct user_fpsimd_state *fst = &task->thread.uw.fpsimd_state;
706 __uint128_t const *p;
708 if (!system_supports_sve())
711 vl = thread_get_cur_vl(&task->thread);
712 vq = sve_vq_from_vl(vl);
713 for (i = 0; i < SVE_NUM_ZREGS; ++i) {
714 p = (__uint128_t const *)ZREG(sst, vq, i);
715 fst->vregs[i] = arm64_le128_to_cpu(*p);
719 #ifdef CONFIG_ARM64_SVE
721 * Call __sve_free() directly only if you know task can't be scheduled
724 static void __sve_free(struct task_struct *task)
726 kfree(task->thread.sve_state);
727 task->thread.sve_state = NULL;
730 static void sve_free(struct task_struct *task)
732 WARN_ON(test_tsk_thread_flag(task, TIF_SVE));
738 * Return how many bytes of memory are required to store the full SVE
739 * state for task, given task's currently configured vector length.
741 size_t sve_state_size(struct task_struct const *task)
745 if (system_supports_sve())
746 vl = task_get_sve_vl(task);
747 if (system_supports_sme())
748 vl = max(vl, task_get_sme_vl(task));
750 return SVE_SIG_REGS_SIZE(sve_vq_from_vl(vl));
754 * Ensure that task->thread.sve_state is allocated and sufficiently large.
756 * This function should be used only in preparation for replacing
757 * task->thread.sve_state with new data. The memory is always zeroed
758 * here to prevent stale data from showing through: this is done in
759 * the interest of testability and predictability: except in the
760 * do_sve_acc() case, there is no ABI requirement to hide stale data
761 * written previously be task.
763 void sve_alloc(struct task_struct *task, bool flush)
765 if (task->thread.sve_state) {
767 memset(task->thread.sve_state, 0,
768 sve_state_size(task));
772 /* This is a small allocation (maximum ~8KB) and Should Not Fail. */
773 task->thread.sve_state =
774 kzalloc(sve_state_size(task), GFP_KERNEL);
779 * Force the FPSIMD state shared with SVE to be updated in the SVE state
780 * even if the SVE state is the current active state.
782 * This should only be called by ptrace. task must be non-runnable.
783 * task->thread.sve_state must point to at least sve_state_size(task)
784 * bytes of allocated kernel memory.
786 void fpsimd_force_sync_to_sve(struct task_struct *task)
792 * Ensure that task->thread.sve_state is up to date with respect to
793 * the user task, irrespective of when SVE is in use or not.
795 * This should only be called by ptrace. task must be non-runnable.
796 * task->thread.sve_state must point to at least sve_state_size(task)
797 * bytes of allocated kernel memory.
799 void fpsimd_sync_to_sve(struct task_struct *task)
801 if (!test_tsk_thread_flag(task, TIF_SVE) &&
802 !thread_sm_enabled(&task->thread))
807 * Ensure that task->thread.uw.fpsimd_state is up to date with respect to
808 * the user task, irrespective of whether SVE is in use or not.
810 * This should only be called by ptrace. 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.
814 void sve_sync_to_fpsimd(struct task_struct *task)
816 if (task->thread.fp_type == FP_STATE_SVE)
821 * Ensure that task->thread.sve_state is up to date with respect to
822 * the task->thread.uw.fpsimd_state.
824 * This should only be called by ptrace to merge new FPSIMD register
825 * values into a task for which SVE is currently active.
826 * task must be non-runnable.
827 * task->thread.sve_state must point to at least sve_state_size(task)
828 * bytes of allocated kernel memory.
829 * task->thread.uw.fpsimd_state must already have been initialised with
830 * the new FPSIMD register values to be merged in.
832 void sve_sync_from_fpsimd_zeropad(struct task_struct *task)
835 void *sst = task->thread.sve_state;
836 struct user_fpsimd_state const *fst = &task->thread.uw.fpsimd_state;
838 if (!test_tsk_thread_flag(task, TIF_SVE))
841 vq = sve_vq_from_vl(thread_get_cur_vl(&task->thread));
843 memset(sst, 0, SVE_SIG_REGS_SIZE(vq));
844 __fpsimd_to_sve(sst, fst, vq);
847 int vec_set_vector_length(struct task_struct *task, enum vec_type type,
848 unsigned long vl, unsigned long flags)
850 bool free_sme = false;
852 if (flags & ~(unsigned long)(PR_SVE_VL_INHERIT |
853 PR_SVE_SET_VL_ONEXEC))
856 if (!sve_vl_valid(vl))
860 * Clamp to the maximum vector length that VL-agnostic code
861 * can work with. A flag may be assigned in the future to
862 * allow setting of larger vector lengths without confusing
865 if (vl > VL_ARCH_MAX)
868 vl = find_supported_vector_length(type, vl);
870 if (flags & (PR_SVE_VL_INHERIT |
871 PR_SVE_SET_VL_ONEXEC))
872 task_set_vl_onexec(task, type, vl);
874 /* Reset VL to system default on next exec: */
875 task_set_vl_onexec(task, type, 0);
877 /* Only actually set the VL if not deferred: */
878 if (flags & PR_SVE_SET_VL_ONEXEC)
881 if (vl == task_get_vl(task, type))
885 * To ensure the FPSIMD bits of the SVE vector registers are preserved,
886 * write any live register state back to task_struct, and convert to a
887 * regular FPSIMD thread.
889 if (task == current) {
890 get_cpu_fpsimd_context();
895 fpsimd_flush_task_state(task);
896 if (test_and_clear_tsk_thread_flag(task, TIF_SVE) ||
897 thread_sm_enabled(&task->thread)) {
899 task->thread.fp_type = FP_STATE_FPSIMD;
902 if (system_supports_sme()) {
903 if (type == ARM64_VEC_SME ||
904 !(task->thread.svcr & (SVCR_SM_MASK | SVCR_ZA_MASK))) {
906 * We are changing the SME VL or weren't using
907 * SME anyway, discard the state and force a
910 task->thread.svcr &= ~(SVCR_SM_MASK |
912 clear_thread_flag(TIF_SME);
918 put_cpu_fpsimd_context();
920 task_set_vl(task, type, vl);
923 * Free the changed states if they are not in use, SME will be
924 * reallocated to the correct size on next use and we just
925 * allocate SVE now in case it is needed for use in streaming
928 if (system_supports_sve()) {
930 sve_alloc(task, true);
937 update_tsk_thread_flag(task, vec_vl_inherit_flag(type),
938 flags & PR_SVE_VL_INHERIT);
944 * Encode the current vector length and flags for return.
945 * This is only required for prctl(): ptrace has separate fields.
946 * SVE and SME use the same bits for _ONEXEC and _INHERIT.
948 * flags are as for vec_set_vector_length().
950 static int vec_prctl_status(enum vec_type type, unsigned long flags)
954 if (flags & PR_SVE_SET_VL_ONEXEC)
955 ret = task_get_vl_onexec(current, type);
957 ret = task_get_vl(current, type);
959 if (test_thread_flag(vec_vl_inherit_flag(type)))
960 ret |= PR_SVE_VL_INHERIT;
966 int sve_set_current_vl(unsigned long arg)
968 unsigned long vl, flags;
971 vl = arg & PR_SVE_VL_LEN_MASK;
974 if (!system_supports_sve() || is_compat_task())
977 ret = vec_set_vector_length(current, ARM64_VEC_SVE, vl, flags);
981 return vec_prctl_status(ARM64_VEC_SVE, flags);
985 int sve_get_current_vl(void)
987 if (!system_supports_sve() || is_compat_task())
990 return vec_prctl_status(ARM64_VEC_SVE, 0);
993 #ifdef CONFIG_ARM64_SME
995 int sme_set_current_vl(unsigned long arg)
997 unsigned long vl, flags;
1000 vl = arg & PR_SME_VL_LEN_MASK;
1003 if (!system_supports_sme() || is_compat_task())
1006 ret = vec_set_vector_length(current, ARM64_VEC_SME, vl, flags);
1010 return vec_prctl_status(ARM64_VEC_SME, flags);
1014 int sme_get_current_vl(void)
1016 if (!system_supports_sme() || is_compat_task())
1019 return vec_prctl_status(ARM64_VEC_SME, 0);
1021 #endif /* CONFIG_ARM64_SME */
1023 static void vec_probe_vqs(struct vl_info *info,
1024 DECLARE_BITMAP(map, SVE_VQ_MAX))
1026 unsigned int vq, vl;
1028 bitmap_zero(map, SVE_VQ_MAX);
1030 for (vq = SVE_VQ_MAX; vq >= SVE_VQ_MIN; --vq) {
1031 write_vl(info->type, vq - 1); /* self-syncing */
1033 switch (info->type) {
1045 /* Minimum VL identified? */
1046 if (sve_vq_from_vl(vl) > vq)
1049 vq = sve_vq_from_vl(vl); /* skip intervening lengths */
1050 set_bit(__vq_to_bit(vq), map);
1055 * Initialise the set of known supported VQs for the boot CPU.
1056 * This is called during kernel boot, before secondary CPUs are brought up.
1058 void __init vec_init_vq_map(enum vec_type type)
1060 struct vl_info *info = &vl_info[type];
1061 vec_probe_vqs(info, info->vq_map);
1062 bitmap_copy(info->vq_partial_map, info->vq_map, SVE_VQ_MAX);
1066 * If we haven't committed to the set of supported VQs yet, filter out
1067 * those not supported by the current CPU.
1068 * This function is called during the bring-up of early secondary CPUs only.
1070 void vec_update_vq_map(enum vec_type type)
1072 struct vl_info *info = &vl_info[type];
1073 DECLARE_BITMAP(tmp_map, SVE_VQ_MAX);
1075 vec_probe_vqs(info, tmp_map);
1076 bitmap_and(info->vq_map, info->vq_map, tmp_map, SVE_VQ_MAX);
1077 bitmap_or(info->vq_partial_map, info->vq_partial_map, tmp_map,
1082 * Check whether the current CPU supports all VQs in the committed set.
1083 * This function is called during the bring-up of late secondary CPUs only.
1085 int vec_verify_vq_map(enum vec_type type)
1087 struct vl_info *info = &vl_info[type];
1088 DECLARE_BITMAP(tmp_map, SVE_VQ_MAX);
1091 vec_probe_vqs(info, tmp_map);
1093 bitmap_complement(tmp_map, tmp_map, SVE_VQ_MAX);
1094 if (bitmap_intersects(tmp_map, info->vq_map, SVE_VQ_MAX)) {
1095 pr_warn("%s: cpu%d: Required vector length(s) missing\n",
1096 info->name, smp_processor_id());
1100 if (!IS_ENABLED(CONFIG_KVM) || !is_hyp_mode_available())
1104 * For KVM, it is necessary to ensure that this CPU doesn't
1105 * support any vector length that guests may have probed as
1109 /* Recover the set of supported VQs: */
1110 bitmap_complement(tmp_map, tmp_map, SVE_VQ_MAX);
1111 /* Find VQs supported that are not globally supported: */
1112 bitmap_andnot(tmp_map, tmp_map, info->vq_map, SVE_VQ_MAX);
1114 /* Find the lowest such VQ, if any: */
1115 b = find_last_bit(tmp_map, SVE_VQ_MAX);
1116 if (b >= SVE_VQ_MAX)
1117 return 0; /* no mismatches */
1120 * Mismatches above sve_max_virtualisable_vl are fine, since
1121 * no guest is allowed to configure ZCR_EL2.LEN to exceed this:
1123 if (sve_vl_from_vq(__bit_to_vq(b)) <= info->max_virtualisable_vl) {
1124 pr_warn("%s: cpu%d: Unsupported vector length(s) present\n",
1125 info->name, smp_processor_id());
1132 static void __init sve_efi_setup(void)
1137 if (!IS_ENABLED(CONFIG_EFI))
1140 for (i = 0; i < ARRAY_SIZE(vl_info); i++)
1141 max_vl = max(vl_info[i].max_vl, max_vl);
1144 * alloc_percpu() warns and prints a backtrace if this goes wrong.
1145 * This is evidence of a crippled system and we are returning void,
1146 * so no attempt is made to handle this situation here.
1148 if (!sve_vl_valid(max_vl))
1151 efi_sve_state = __alloc_percpu(
1152 SVE_SIG_REGS_SIZE(sve_vq_from_vl(max_vl)), SVE_VQ_BYTES);
1159 panic("Cannot allocate percpu memory for EFI SVE save/restore");
1163 * Enable SVE for EL1.
1164 * Intended for use by the cpufeatures code during CPU boot.
1166 void sve_kernel_enable(const struct arm64_cpu_capabilities *__always_unused p)
1168 write_sysreg(read_sysreg(CPACR_EL1) | CPACR_EL1_ZEN_EL1EN, CPACR_EL1);
1173 * Read the pseudo-ZCR used by cpufeatures to identify the supported SVE
1176 * Use only if SVE is present.
1177 * This function clobbers the SVE vector length.
1179 u64 read_zcr_features(void)
1182 unsigned int vq_max;
1185 * Set the maximum possible VL, and write zeroes to all other
1186 * bits to see if they stick.
1188 sve_kernel_enable(NULL);
1189 write_sysreg_s(ZCR_ELx_LEN_MASK, SYS_ZCR_EL1);
1191 zcr = read_sysreg_s(SYS_ZCR_EL1);
1192 zcr &= ~(u64)ZCR_ELx_LEN_MASK; /* find sticky 1s outside LEN field */
1193 vq_max = sve_vq_from_vl(sve_get_vl());
1194 zcr |= vq_max - 1; /* set LEN field to maximum effective value */
1199 void __init sve_setup(void)
1201 struct vl_info *info = &vl_info[ARM64_VEC_SVE];
1203 DECLARE_BITMAP(tmp_map, SVE_VQ_MAX);
1206 if (!system_supports_sve())
1210 * The SVE architecture mandates support for 128-bit vectors,
1211 * so sve_vq_map must have at least SVE_VQ_MIN set.
1212 * If something went wrong, at least try to patch it up:
1214 if (WARN_ON(!test_bit(__vq_to_bit(SVE_VQ_MIN), info->vq_map)))
1215 set_bit(__vq_to_bit(SVE_VQ_MIN), info->vq_map);
1217 zcr = read_sanitised_ftr_reg(SYS_ZCR_EL1);
1218 info->max_vl = sve_vl_from_vq((zcr & ZCR_ELx_LEN_MASK) + 1);
1221 * Sanity-check that the max VL we determined through CPU features
1222 * corresponds properly to sve_vq_map. If not, do our best:
1224 if (WARN_ON(info->max_vl != find_supported_vector_length(ARM64_VEC_SVE,
1226 info->max_vl = find_supported_vector_length(ARM64_VEC_SVE,
1230 * For the default VL, pick the maximum supported value <= 64.
1231 * VL == 64 is guaranteed not to grow the signal frame.
1233 set_sve_default_vl(find_supported_vector_length(ARM64_VEC_SVE, 64));
1235 bitmap_andnot(tmp_map, info->vq_partial_map, info->vq_map,
1238 b = find_last_bit(tmp_map, SVE_VQ_MAX);
1239 if (b >= SVE_VQ_MAX)
1240 /* No non-virtualisable VLs found */
1241 info->max_virtualisable_vl = SVE_VQ_MAX;
1242 else if (WARN_ON(b == SVE_VQ_MAX - 1))
1243 /* No virtualisable VLs? This is architecturally forbidden. */
1244 info->max_virtualisable_vl = SVE_VQ_MIN;
1245 else /* b + 1 < SVE_VQ_MAX */
1246 info->max_virtualisable_vl = sve_vl_from_vq(__bit_to_vq(b + 1));
1248 if (info->max_virtualisable_vl > info->max_vl)
1249 info->max_virtualisable_vl = info->max_vl;
1251 pr_info("%s: maximum available vector length %u bytes per vector\n",
1252 info->name, info->max_vl);
1253 pr_info("%s: default vector length %u bytes per vector\n",
1254 info->name, get_sve_default_vl());
1256 /* KVM decides whether to support mismatched systems. Just warn here: */
1257 if (sve_max_virtualisable_vl() < sve_max_vl())
1258 pr_warn("%s: unvirtualisable vector lengths present\n",
1265 * Called from the put_task_struct() path, which cannot get here
1266 * unless dead_task is really dead and not schedulable.
1268 void fpsimd_release_task(struct task_struct *dead_task)
1270 __sve_free(dead_task);
1271 sme_free(dead_task);
1274 #endif /* CONFIG_ARM64_SVE */
1276 #ifdef CONFIG_ARM64_SME
1279 * Ensure that task->thread.sme_state is allocated and sufficiently large.
1281 * This function should be used only in preparation for replacing
1282 * task->thread.sme_state with new data. The memory is always zeroed
1283 * here to prevent stale data from showing through: this is done in
1284 * the interest of testability and predictability, the architecture
1285 * guarantees that when ZA is enabled it will be zeroed.
1287 void sme_alloc(struct task_struct *task)
1289 if (task->thread.sme_state) {
1290 memset(task->thread.sme_state, 0, sme_state_size(task));
1294 /* This could potentially be up to 64K. */
1295 task->thread.sme_state =
1296 kzalloc(sme_state_size(task), GFP_KERNEL);
1299 static void sme_free(struct task_struct *task)
1301 kfree(task->thread.sme_state);
1302 task->thread.sme_state = NULL;
1305 void sme_kernel_enable(const struct arm64_cpu_capabilities *__always_unused p)
1307 /* Set priority for all PEs to architecturally defined minimum */
1308 write_sysreg_s(read_sysreg_s(SYS_SMPRI_EL1) & ~SMPRI_EL1_PRIORITY_MASK,
1311 /* Allow SME in kernel */
1312 write_sysreg(read_sysreg(CPACR_EL1) | CPACR_EL1_SMEN_EL1EN, CPACR_EL1);
1315 /* Allow EL0 to access TPIDR2 */
1316 write_sysreg(read_sysreg(SCTLR_EL1) | SCTLR_ELx_ENTP2, SCTLR_EL1);
1321 * This must be called after sme_kernel_enable(), we rely on the
1322 * feature table being sorted to ensure this.
1324 void sme2_kernel_enable(const struct arm64_cpu_capabilities *__always_unused p)
1326 /* Allow use of ZT0 */
1327 write_sysreg_s(read_sysreg_s(SYS_SMCR_EL1) | SMCR_ELx_EZT0_MASK,
1332 * This must be called after sme_kernel_enable(), we rely on the
1333 * feature table being sorted to ensure this.
1335 void fa64_kernel_enable(const struct arm64_cpu_capabilities *__always_unused p)
1337 /* Allow use of FA64 */
1338 write_sysreg_s(read_sysreg_s(SYS_SMCR_EL1) | SMCR_ELx_FA64_MASK,
1343 * Read the pseudo-SMCR used by cpufeatures to identify the supported
1346 * Use only if SME is present.
1347 * This function clobbers the SME vector length.
1349 u64 read_smcr_features(void)
1352 unsigned int vq_max;
1354 sme_kernel_enable(NULL);
1357 * Set the maximum possible VL.
1359 write_sysreg_s(read_sysreg_s(SYS_SMCR_EL1) | SMCR_ELx_LEN_MASK,
1362 smcr = read_sysreg_s(SYS_SMCR_EL1);
1363 smcr &= ~(u64)SMCR_ELx_LEN_MASK; /* Only the LEN field */
1364 vq_max = sve_vq_from_vl(sme_get_vl());
1365 smcr |= vq_max - 1; /* set LEN field to maximum effective value */
1370 void __init sme_setup(void)
1372 struct vl_info *info = &vl_info[ARM64_VEC_SME];
1376 if (!system_supports_sme())
1380 * SME doesn't require any particular vector length be
1381 * supported but it does require at least one. We should have
1382 * disabled the feature entirely while bringing up CPUs but
1383 * let's double check here.
1385 WARN_ON(bitmap_empty(info->vq_map, SVE_VQ_MAX));
1387 min_bit = find_last_bit(info->vq_map, SVE_VQ_MAX);
1388 info->min_vl = sve_vl_from_vq(__bit_to_vq(min_bit));
1390 smcr = read_sanitised_ftr_reg(SYS_SMCR_EL1);
1391 info->max_vl = sve_vl_from_vq((smcr & SMCR_ELx_LEN_MASK) + 1);
1394 * Sanity-check that the max VL we determined through CPU features
1395 * corresponds properly to sme_vq_map. If not, do our best:
1397 if (WARN_ON(info->max_vl != find_supported_vector_length(ARM64_VEC_SME,
1399 info->max_vl = find_supported_vector_length(ARM64_VEC_SME,
1402 WARN_ON(info->min_vl > info->max_vl);
1405 * For the default VL, pick the maximum supported value <= 32
1406 * (256 bits) if there is one since this is guaranteed not to
1407 * grow the signal frame when in streaming mode, otherwise the
1408 * minimum available VL will be used.
1410 set_sme_default_vl(find_supported_vector_length(ARM64_VEC_SME, 32));
1412 pr_info("SME: minimum available vector length %u bytes per vector\n",
1414 pr_info("SME: maximum available vector length %u bytes per vector\n",
1416 pr_info("SME: default vector length %u bytes per vector\n",
1417 get_sme_default_vl());
1420 #endif /* CONFIG_ARM64_SME */
1422 static void sve_init_regs(void)
1425 * Convert the FPSIMD state to SVE, zeroing all the state that
1426 * is not shared with FPSIMD. If (as is likely) the current
1427 * state is live in the registers then do this there and
1428 * update our metadata for the current task including
1429 * disabling the trap, otherwise update our in-memory copy.
1430 * We are guaranteed to not be in streaming mode, we can only
1431 * take a SVE trap when not in streaming mode and we can't be
1432 * in streaming mode when taking a SME trap.
1434 if (!test_thread_flag(TIF_FOREIGN_FPSTATE)) {
1435 unsigned long vq_minus_one =
1436 sve_vq_from_vl(task_get_sve_vl(current)) - 1;
1437 sve_set_vq(vq_minus_one);
1438 sve_flush_live(true, vq_minus_one);
1439 fpsimd_bind_task_to_cpu();
1441 fpsimd_to_sve(current);
1442 current->thread.fp_type = FP_STATE_SVE;
1447 * Trapped SVE access
1449 * Storage is allocated for the full SVE state, the current FPSIMD
1450 * register contents are migrated across, and the access trap is
1453 * TIF_SVE should be clear on entry: otherwise, fpsimd_restore_current_state()
1454 * would have disabled the SVE access trap for userspace during
1455 * ret_to_user, making an SVE access trap impossible in that case.
1457 void do_sve_acc(unsigned long esr, struct pt_regs *regs)
1459 /* Even if we chose not to use SVE, the hardware could still trap: */
1460 if (unlikely(!system_supports_sve()) || WARN_ON(is_compat_task())) {
1461 force_signal_inject(SIGILL, ILL_ILLOPC, regs->pc, 0);
1465 sve_alloc(current, true);
1466 if (!current->thread.sve_state) {
1471 get_cpu_fpsimd_context();
1473 if (test_and_set_thread_flag(TIF_SVE))
1474 WARN_ON(1); /* SVE access shouldn't have trapped */
1477 * Even if the task can have used streaming mode we can only
1478 * generate SVE access traps in normal SVE mode and
1479 * transitioning out of streaming mode may discard any
1480 * streaming mode state. Always clear the high bits to avoid
1481 * any potential errors tracking what is properly initialised.
1485 put_cpu_fpsimd_context();
1489 * Trapped SME access
1491 * Storage is allocated for the full SVE and SME state, the current
1492 * FPSIMD register contents are migrated to SVE if SVE is not already
1493 * active, and the access trap is disabled.
1495 * TIF_SME should be clear on entry: otherwise, fpsimd_restore_current_state()
1496 * would have disabled the SME access trap for userspace during
1497 * ret_to_user, making an SME access trap impossible in that case.
1499 void do_sme_acc(unsigned long esr, struct pt_regs *regs)
1501 /* Even if we chose not to use SME, the hardware could still trap: */
1502 if (unlikely(!system_supports_sme()) || WARN_ON(is_compat_task())) {
1503 force_signal_inject(SIGILL, ILL_ILLOPC, regs->pc, 0);
1508 * If this not a trap due to SME being disabled then something
1509 * is being used in the wrong mode, report as SIGILL.
1511 if (ESR_ELx_ISS(esr) != ESR_ELx_SME_ISS_SME_DISABLED) {
1512 force_signal_inject(SIGILL, ILL_ILLOPC, regs->pc, 0);
1516 sve_alloc(current, false);
1518 if (!current->thread.sve_state || !current->thread.sme_state) {
1523 get_cpu_fpsimd_context();
1525 /* With TIF_SME userspace shouldn't generate any traps */
1526 if (test_and_set_thread_flag(TIF_SME))
1529 if (!test_thread_flag(TIF_FOREIGN_FPSTATE)) {
1530 unsigned long vq_minus_one =
1531 sve_vq_from_vl(task_get_sme_vl(current)) - 1;
1532 sme_set_vq(vq_minus_one);
1534 fpsimd_bind_task_to_cpu();
1537 put_cpu_fpsimd_context();
1541 * Trapped FP/ASIMD access.
1543 void do_fpsimd_acc(unsigned long esr, struct pt_regs *regs)
1545 /* TODO: implement lazy context saving/restoring */
1550 * Raise a SIGFPE for the current process.
1552 void do_fpsimd_exc(unsigned long esr, struct pt_regs *regs)
1554 unsigned int si_code = FPE_FLTUNK;
1556 if (esr & ESR_ELx_FP_EXC_TFV) {
1557 if (esr & FPEXC_IOF)
1558 si_code = FPE_FLTINV;
1559 else if (esr & FPEXC_DZF)
1560 si_code = FPE_FLTDIV;
1561 else if (esr & FPEXC_OFF)
1562 si_code = FPE_FLTOVF;
1563 else if (esr & FPEXC_UFF)
1564 si_code = FPE_FLTUND;
1565 else if (esr & FPEXC_IXF)
1566 si_code = FPE_FLTRES;
1569 send_sig_fault(SIGFPE, si_code,
1570 (void __user *)instruction_pointer(regs),
1574 void fpsimd_thread_switch(struct task_struct *next)
1576 bool wrong_task, wrong_cpu;
1578 if (!system_supports_fpsimd())
1581 __get_cpu_fpsimd_context();
1583 /* Save unsaved fpsimd state, if any: */
1587 * Fix up TIF_FOREIGN_FPSTATE to correctly describe next's
1588 * state. For kernel threads, FPSIMD registers are never loaded
1589 * and wrong_task and wrong_cpu will always be true.
1591 wrong_task = __this_cpu_read(fpsimd_last_state.st) !=
1592 &next->thread.uw.fpsimd_state;
1593 wrong_cpu = next->thread.fpsimd_cpu != smp_processor_id();
1595 update_tsk_thread_flag(next, TIF_FOREIGN_FPSTATE,
1596 wrong_task || wrong_cpu);
1598 __put_cpu_fpsimd_context();
1601 static void fpsimd_flush_thread_vl(enum vec_type type)
1603 int vl, supported_vl;
1606 * Reset the task vector length as required. This is where we
1607 * ensure that all user tasks have a valid vector length
1608 * configured: no kernel task can become a user task without
1609 * an exec and hence a call to this function. By the time the
1610 * first call to this function is made, all early hardware
1611 * probing is complete, so __sve_default_vl should be valid.
1612 * If a bug causes this to go wrong, we make some noise and
1613 * try to fudge thread.sve_vl to a safe value here.
1615 vl = task_get_vl_onexec(current, type);
1617 vl = get_default_vl(type);
1619 if (WARN_ON(!sve_vl_valid(vl)))
1620 vl = vl_info[type].min_vl;
1622 supported_vl = find_supported_vector_length(type, vl);
1623 if (WARN_ON(supported_vl != vl))
1626 task_set_vl(current, type, vl);
1629 * If the task is not set to inherit, ensure that the vector
1630 * length will be reset by a subsequent exec:
1632 if (!test_thread_flag(vec_vl_inherit_flag(type)))
1633 task_set_vl_onexec(current, type, 0);
1636 void fpsimd_flush_thread(void)
1638 void *sve_state = NULL;
1639 void *sme_state = NULL;
1641 if (!system_supports_fpsimd())
1644 get_cpu_fpsimd_context();
1646 fpsimd_flush_task_state(current);
1647 memset(¤t->thread.uw.fpsimd_state, 0,
1648 sizeof(current->thread.uw.fpsimd_state));
1650 if (system_supports_sve()) {
1651 clear_thread_flag(TIF_SVE);
1653 /* Defer kfree() while in atomic context */
1654 sve_state = current->thread.sve_state;
1655 current->thread.sve_state = NULL;
1657 fpsimd_flush_thread_vl(ARM64_VEC_SVE);
1660 if (system_supports_sme()) {
1661 clear_thread_flag(TIF_SME);
1663 /* Defer kfree() while in atomic context */
1664 sme_state = current->thread.sme_state;
1665 current->thread.sme_state = NULL;
1667 fpsimd_flush_thread_vl(ARM64_VEC_SME);
1668 current->thread.svcr = 0;
1671 current->thread.fp_type = FP_STATE_FPSIMD;
1673 put_cpu_fpsimd_context();
1679 * Save the userland FPSIMD state of 'current' to memory, but only if the state
1680 * currently held in the registers does in fact belong to 'current'
1682 void fpsimd_preserve_current_state(void)
1684 if (!system_supports_fpsimd())
1687 get_cpu_fpsimd_context();
1689 put_cpu_fpsimd_context();
1693 * Like fpsimd_preserve_current_state(), but ensure that
1694 * current->thread.uw.fpsimd_state is updated so that it can be copied to
1697 void fpsimd_signal_preserve_current_state(void)
1699 fpsimd_preserve_current_state();
1700 if (test_thread_flag(TIF_SVE))
1701 sve_to_fpsimd(current);
1705 * Called by KVM when entering the guest.
1707 void fpsimd_kvm_prepare(void)
1709 if (!system_supports_sve())
1713 * KVM does not save host SVE state since we can only enter
1714 * the guest from a syscall so the ABI means that only the
1715 * non-saved SVE state needs to be saved. If we have left
1716 * SVE enabled for performance reasons then update the task
1717 * state to be FPSIMD only.
1719 get_cpu_fpsimd_context();
1721 if (test_and_clear_thread_flag(TIF_SVE)) {
1722 sve_to_fpsimd(current);
1723 current->thread.fp_type = FP_STATE_FPSIMD;
1726 put_cpu_fpsimd_context();
1730 * Associate current's FPSIMD context with this cpu
1731 * The caller must have ownership of the cpu FPSIMD context before calling
1734 static void fpsimd_bind_task_to_cpu(void)
1736 struct cpu_fp_state *last = this_cpu_ptr(&fpsimd_last_state);
1738 WARN_ON(!system_supports_fpsimd());
1739 last->st = ¤t->thread.uw.fpsimd_state;
1740 last->sve_state = current->thread.sve_state;
1741 last->sme_state = current->thread.sme_state;
1742 last->sve_vl = task_get_sve_vl(current);
1743 last->sme_vl = task_get_sme_vl(current);
1744 last->svcr = ¤t->thread.svcr;
1745 last->fp_type = ¤t->thread.fp_type;
1746 last->to_save = FP_STATE_CURRENT;
1747 current->thread.fpsimd_cpu = smp_processor_id();
1750 * Toggle SVE and SME trapping for userspace if needed, these
1751 * are serialsied by ret_to_user().
1753 if (system_supports_sme()) {
1754 if (test_thread_flag(TIF_SME))
1760 if (system_supports_sve()) {
1761 if (test_thread_flag(TIF_SVE))
1768 void fpsimd_bind_state_to_cpu(struct cpu_fp_state *state)
1770 struct cpu_fp_state *last = this_cpu_ptr(&fpsimd_last_state);
1772 WARN_ON(!system_supports_fpsimd());
1773 WARN_ON(!in_softirq() && !irqs_disabled());
1779 * Load the userland FPSIMD state of 'current' from memory, but only if the
1780 * FPSIMD state already held in the registers is /not/ the most recent FPSIMD
1781 * state of 'current'. This is called when we are preparing to return to
1782 * userspace to ensure that userspace sees a good register state.
1784 void fpsimd_restore_current_state(void)
1787 * For the tasks that were created before we detected the absence of
1788 * FP/SIMD, the TIF_FOREIGN_FPSTATE could be set via fpsimd_thread_switch(),
1789 * e.g, init. This could be then inherited by the children processes.
1790 * If we later detect that the system doesn't support FP/SIMD,
1791 * we must clear the flag for all the tasks to indicate that the
1792 * FPSTATE is clean (as we can't have one) to avoid looping for ever in
1793 * do_notify_resume().
1795 if (!system_supports_fpsimd()) {
1796 clear_thread_flag(TIF_FOREIGN_FPSTATE);
1800 get_cpu_fpsimd_context();
1802 if (test_and_clear_thread_flag(TIF_FOREIGN_FPSTATE)) {
1804 fpsimd_bind_task_to_cpu();
1807 put_cpu_fpsimd_context();
1811 * Load an updated userland FPSIMD state for 'current' from memory and set the
1812 * flag that indicates that the FPSIMD register contents are the most recent
1813 * FPSIMD state of 'current'. This is used by the signal code to restore the
1814 * register state when returning from a signal handler in FPSIMD only cases,
1815 * any SVE context will be discarded.
1817 void fpsimd_update_current_state(struct user_fpsimd_state const *state)
1819 if (WARN_ON(!system_supports_fpsimd()))
1822 get_cpu_fpsimd_context();
1824 current->thread.uw.fpsimd_state = *state;
1825 if (test_thread_flag(TIF_SVE))
1826 fpsimd_to_sve(current);
1829 fpsimd_bind_task_to_cpu();
1831 clear_thread_flag(TIF_FOREIGN_FPSTATE);
1833 put_cpu_fpsimd_context();
1837 * Invalidate live CPU copies of task t's FPSIMD state
1839 * This function may be called with preemption enabled. The barrier()
1840 * ensures that the assignment to fpsimd_cpu is visible to any
1841 * preemption/softirq that could race with set_tsk_thread_flag(), so
1842 * that TIF_FOREIGN_FPSTATE cannot be spuriously re-cleared.
1844 * The final barrier ensures that TIF_FOREIGN_FPSTATE is seen set by any
1847 void fpsimd_flush_task_state(struct task_struct *t)
1849 t->thread.fpsimd_cpu = NR_CPUS;
1851 * If we don't support fpsimd, bail out after we have
1852 * reset the fpsimd_cpu for this task and clear the
1855 if (!system_supports_fpsimd())
1858 set_tsk_thread_flag(t, TIF_FOREIGN_FPSTATE);
1864 * Invalidate any task's FPSIMD state that is present on this cpu.
1865 * The FPSIMD context should be acquired with get_cpu_fpsimd_context()
1866 * before calling this function.
1868 static void fpsimd_flush_cpu_state(void)
1870 WARN_ON(!system_supports_fpsimd());
1871 __this_cpu_write(fpsimd_last_state.st, NULL);
1874 * Leaving streaming mode enabled will cause issues for any kernel
1875 * NEON and leaving streaming mode or ZA enabled may increase power
1878 if (system_supports_sme())
1881 set_thread_flag(TIF_FOREIGN_FPSTATE);
1885 * Save the FPSIMD state to memory and invalidate cpu view.
1886 * This function must be called with preemption disabled.
1888 void fpsimd_save_and_flush_cpu_state(void)
1890 if (!system_supports_fpsimd())
1892 WARN_ON(preemptible());
1893 __get_cpu_fpsimd_context();
1895 fpsimd_flush_cpu_state();
1896 __put_cpu_fpsimd_context();
1899 #ifdef CONFIG_KERNEL_MODE_NEON
1902 * Kernel-side NEON support functions
1906 * kernel_neon_begin(): obtain the CPU FPSIMD registers for use by the calling
1909 * Must not be called unless may_use_simd() returns true.
1910 * Task context in the FPSIMD registers is saved back to memory as necessary.
1912 * A matching call to kernel_neon_end() must be made before returning from the
1915 * The caller may freely use the FPSIMD registers until kernel_neon_end() is
1918 void kernel_neon_begin(void)
1920 if (WARN_ON(!system_supports_fpsimd()))
1923 BUG_ON(!may_use_simd());
1925 get_cpu_fpsimd_context();
1927 /* Save unsaved fpsimd state, if any: */
1930 /* Invalidate any task state remaining in the fpsimd regs: */
1931 fpsimd_flush_cpu_state();
1933 EXPORT_SYMBOL_GPL(kernel_neon_begin);
1936 * kernel_neon_end(): give the CPU FPSIMD registers back to the current task
1938 * Must be called from a context in which kernel_neon_begin() was previously
1939 * called, with no call to kernel_neon_end() in the meantime.
1941 * The caller must not use the FPSIMD registers after this function is called,
1942 * unless kernel_neon_begin() is called again in the meantime.
1944 void kernel_neon_end(void)
1946 if (!system_supports_fpsimd())
1949 put_cpu_fpsimd_context();
1951 EXPORT_SYMBOL_GPL(kernel_neon_end);
1955 static DEFINE_PER_CPU(struct user_fpsimd_state, efi_fpsimd_state);
1956 static DEFINE_PER_CPU(bool, efi_fpsimd_state_used);
1957 static DEFINE_PER_CPU(bool, efi_sve_state_used);
1958 static DEFINE_PER_CPU(bool, efi_sm_state);
1961 * EFI runtime services support functions
1963 * The ABI for EFI runtime services allows EFI to use FPSIMD during the call.
1964 * This means that for EFI (and only for EFI), we have to assume that FPSIMD
1965 * is always used rather than being an optional accelerator.
1967 * These functions provide the necessary support for ensuring FPSIMD
1968 * save/restore in the contexts from which EFI is used.
1970 * Do not use them for any other purpose -- if tempted to do so, you are
1971 * either doing something wrong or you need to propose some refactoring.
1975 * __efi_fpsimd_begin(): prepare FPSIMD for making an EFI runtime services call
1977 void __efi_fpsimd_begin(void)
1979 if (!system_supports_fpsimd())
1982 WARN_ON(preemptible());
1984 if (may_use_simd()) {
1985 kernel_neon_begin();
1988 * If !efi_sve_state, SVE can't be in use yet and doesn't need
1991 if (system_supports_sve() && likely(efi_sve_state)) {
1992 char *sve_state = this_cpu_ptr(efi_sve_state);
1996 __this_cpu_write(efi_sve_state_used, true);
1998 if (system_supports_sme()) {
1999 svcr = read_sysreg_s(SYS_SVCR);
2001 __this_cpu_write(efi_sm_state,
2002 svcr & SVCR_SM_MASK);
2005 * Unless we have FA64 FFR does not
2006 * exist in streaming mode.
2008 if (!system_supports_fa64())
2009 ffr = !(svcr & SVCR_SM_MASK);
2012 sve_save_state(sve_state + sve_ffr_offset(sve_max_vl()),
2013 &this_cpu_ptr(&efi_fpsimd_state)->fpsr,
2016 if (system_supports_sme())
2017 sysreg_clear_set_s(SYS_SVCR,
2021 fpsimd_save_state(this_cpu_ptr(&efi_fpsimd_state));
2024 __this_cpu_write(efi_fpsimd_state_used, true);
2029 * __efi_fpsimd_end(): clean up FPSIMD after an EFI runtime services call
2031 void __efi_fpsimd_end(void)
2033 if (!system_supports_fpsimd())
2036 if (!__this_cpu_xchg(efi_fpsimd_state_used, false)) {
2039 if (system_supports_sve() &&
2040 likely(__this_cpu_read(efi_sve_state_used))) {
2041 char const *sve_state = this_cpu_ptr(efi_sve_state);
2045 * Restore streaming mode; EFI calls are
2046 * normal function calls so should not return in
2049 if (system_supports_sme()) {
2050 if (__this_cpu_read(efi_sm_state)) {
2051 sysreg_clear_set_s(SYS_SVCR,
2056 * Unless we have FA64 FFR does not
2057 * exist in streaming mode.
2059 if (!system_supports_fa64())
2064 sve_load_state(sve_state + sve_ffr_offset(sve_max_vl()),
2065 &this_cpu_ptr(&efi_fpsimd_state)->fpsr,
2068 __this_cpu_write(efi_sve_state_used, false);
2070 fpsimd_load_state(this_cpu_ptr(&efi_fpsimd_state));
2075 #endif /* CONFIG_EFI */
2077 #endif /* CONFIG_KERNEL_MODE_NEON */
2079 #ifdef CONFIG_CPU_PM
2080 static int fpsimd_cpu_pm_notifier(struct notifier_block *self,
2081 unsigned long cmd, void *v)
2085 fpsimd_save_and_flush_cpu_state();
2089 case CPU_PM_ENTER_FAILED:
2096 static struct notifier_block fpsimd_cpu_pm_notifier_block = {
2097 .notifier_call = fpsimd_cpu_pm_notifier,
2100 static void __init fpsimd_pm_init(void)
2102 cpu_pm_register_notifier(&fpsimd_cpu_pm_notifier_block);
2106 static inline void fpsimd_pm_init(void) { }
2107 #endif /* CONFIG_CPU_PM */
2109 #ifdef CONFIG_HOTPLUG_CPU
2110 static int fpsimd_cpu_dead(unsigned int cpu)
2112 per_cpu(fpsimd_last_state.st, cpu) = NULL;
2116 static inline void fpsimd_hotplug_init(void)
2118 cpuhp_setup_state_nocalls(CPUHP_ARM64_FPSIMD_DEAD, "arm64/fpsimd:dead",
2119 NULL, fpsimd_cpu_dead);
2123 static inline void fpsimd_hotplug_init(void) { }
2127 * FP/SIMD support code initialisation.
2129 static int __init fpsimd_init(void)
2131 if (cpu_have_named_feature(FP)) {
2133 fpsimd_hotplug_init();
2135 pr_notice("Floating-point is not implemented\n");
2138 if (!cpu_have_named_feature(ASIMD))
2139 pr_notice("Advanced SIMD is not implemented\n");
2147 core_initcall(fpsimd_init);