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/cpu.h>
16 #include <linux/cpu_pm.h>
17 #include <linux/kernel.h>
18 #include <linux/linkage.h>
19 #include <linux/irqflags.h>
20 #include <linux/init.h>
21 #include <linux/percpu.h>
22 #include <linux/prctl.h>
23 #include <linux/preempt.h>
24 #include <linux/ptrace.h>
25 #include <linux/sched/signal.h>
26 #include <linux/sched/task_stack.h>
27 #include <linux/signal.h>
28 #include <linux/slab.h>
29 #include <linux/stddef.h>
30 #include <linux/sysctl.h>
31 #include <linux/swab.h>
34 #include <asm/fpsimd.h>
35 #include <asm/cpufeature.h>
36 #include <asm/cputype.h>
37 #include <asm/processor.h>
39 #include <asm/sigcontext.h>
40 #include <asm/sysreg.h>
41 #include <asm/traps.h>
44 #define FPEXC_IOF (1 << 0)
45 #define FPEXC_DZF (1 << 1)
46 #define FPEXC_OFF (1 << 2)
47 #define FPEXC_UFF (1 << 3)
48 #define FPEXC_IXF (1 << 4)
49 #define FPEXC_IDF (1 << 7)
52 * (Note: in this discussion, statements about FPSIMD apply equally to SVE.)
54 * In order to reduce the number of times the FPSIMD state is needlessly saved
55 * and restored, we need to keep track of two things:
56 * (a) for each task, we need to remember which CPU was the last one to have
57 * the task's FPSIMD state loaded into its FPSIMD registers;
58 * (b) for each CPU, we need to remember which task's userland FPSIMD state has
59 * been loaded into its FPSIMD registers most recently, or whether it has
60 * been used to perform kernel mode NEON in the meantime.
62 * For (a), we add a fpsimd_cpu field to thread_struct, which gets updated to
63 * the id of the current CPU every time the state is loaded onto a CPU. For (b),
64 * we add the per-cpu variable 'fpsimd_last_state' (below), which contains the
65 * address of the userland FPSIMD state of the task that was loaded onto the CPU
66 * the most recently, or NULL if kernel mode NEON has been performed after that.
68 * With this in place, we no longer have to restore the next FPSIMD state right
69 * when switching between tasks. Instead, we can defer this check to userland
70 * resume, at which time we verify whether the CPU's fpsimd_last_state and the
71 * task's fpsimd_cpu are still mutually in sync. If this is the case, we
72 * can omit the FPSIMD restore.
74 * As an optimization, we use the thread_info flag TIF_FOREIGN_FPSTATE to
75 * indicate whether or not the userland FPSIMD state of the current task is
76 * present in the registers. The flag is set unless the FPSIMD registers of this
77 * CPU currently contain the most recent userland FPSIMD state of the current
80 * In order to allow softirq handlers to use FPSIMD, kernel_neon_begin() may
81 * save the task's FPSIMD context back to task_struct from softirq context.
82 * To prevent this from racing with the manipulation of the task's FPSIMD state
83 * from task context and thereby corrupting the state, it is necessary to
84 * protect any manipulation of a task's fpsimd_state or TIF_FOREIGN_FPSTATE
85 * flag with {, __}get_cpu_fpsimd_context(). This will still allow softirqs to
86 * run but prevent them to use FPSIMD.
88 * For a certain task, the sequence may look something like this:
89 * - the task gets scheduled in; if both the task's fpsimd_cpu field
90 * contains the id of the current CPU, and the CPU's fpsimd_last_state per-cpu
91 * variable points to the task's fpsimd_state, the TIF_FOREIGN_FPSTATE flag is
92 * cleared, otherwise it is set;
94 * - the task returns to userland; if TIF_FOREIGN_FPSTATE is set, the task's
95 * userland FPSIMD state is copied from memory to the registers, the task's
96 * fpsimd_cpu field is set to the id of the current CPU, the current
97 * CPU's fpsimd_last_state pointer is set to this task's fpsimd_state and the
98 * TIF_FOREIGN_FPSTATE flag is cleared;
100 * - the task executes an ordinary syscall; upon return to userland, the
101 * TIF_FOREIGN_FPSTATE flag will still be cleared, so no FPSIMD state is
104 * - the task executes a syscall which executes some NEON instructions; this is
105 * preceded by a call to kernel_neon_begin(), which copies the task's FPSIMD
106 * register contents to memory, clears the fpsimd_last_state per-cpu variable
107 * and sets the TIF_FOREIGN_FPSTATE flag;
109 * - the task gets preempted after kernel_neon_end() is called; as we have not
110 * returned from the 2nd syscall yet, TIF_FOREIGN_FPSTATE is still set so
111 * whatever is in the FPSIMD registers is not saved to memory, but discarded.
113 struct fpsimd_last_state_struct {
114 struct user_fpsimd_state *st;
119 static DEFINE_PER_CPU(struct fpsimd_last_state_struct, fpsimd_last_state);
121 /* Default VL for tasks that don't set it explicitly: */
122 static int sve_default_vl = -1;
124 #ifdef CONFIG_ARM64_SVE
126 /* Maximum supported vector length across all CPUs (initially poisoned) */
127 int __ro_after_init sve_max_vl = SVE_VL_MIN;
128 int __ro_after_init sve_max_virtualisable_vl = SVE_VL_MIN;
131 * Set of available vector lengths,
132 * where length vq encoded as bit __vq_to_bit(vq):
134 __ro_after_init DECLARE_BITMAP(sve_vq_map, SVE_VQ_MAX);
135 /* Set of vector lengths present on at least one cpu: */
136 static __ro_after_init DECLARE_BITMAP(sve_vq_partial_map, SVE_VQ_MAX);
138 static void __percpu *efi_sve_state;
140 #else /* ! CONFIG_ARM64_SVE */
142 /* Dummy declaration for code that will be optimised out: */
143 extern __ro_after_init DECLARE_BITMAP(sve_vq_map, SVE_VQ_MAX);
144 extern __ro_after_init DECLARE_BITMAP(sve_vq_partial_map, SVE_VQ_MAX);
145 extern void __percpu *efi_sve_state;
147 #endif /* ! CONFIG_ARM64_SVE */
149 DEFINE_PER_CPU(bool, fpsimd_context_busy);
150 EXPORT_PER_CPU_SYMBOL(fpsimd_context_busy);
152 static void __get_cpu_fpsimd_context(void)
154 bool busy = __this_cpu_xchg(fpsimd_context_busy, true);
160 * Claim ownership of the CPU FPSIMD context for use by the calling context.
162 * The caller may freely manipulate the FPSIMD context metadata until
163 * put_cpu_fpsimd_context() is called.
165 * The double-underscore version must only be called if you know the task
166 * can't be preempted.
168 static void get_cpu_fpsimd_context(void)
171 __get_cpu_fpsimd_context();
174 static void __put_cpu_fpsimd_context(void)
176 bool busy = __this_cpu_xchg(fpsimd_context_busy, false);
178 WARN_ON(!busy); /* No matching get_cpu_fpsimd_context()? */
182 * Release the CPU FPSIMD context.
184 * Must be called from a context in which get_cpu_fpsimd_context() was
185 * previously called, with no call to put_cpu_fpsimd_context() in the
188 static void put_cpu_fpsimd_context(void)
190 __put_cpu_fpsimd_context();
194 static bool have_cpu_fpsimd_context(void)
196 return !preemptible() && __this_cpu_read(fpsimd_context_busy);
200 * Call __sve_free() directly only if you know task can't be scheduled
203 static void __sve_free(struct task_struct *task)
205 kfree(task->thread.sve_state);
206 task->thread.sve_state = NULL;
209 static void sve_free(struct task_struct *task)
211 WARN_ON(test_tsk_thread_flag(task, TIF_SVE));
217 * TIF_SVE controls whether a task can use SVE without trapping while
218 * in userspace, and also the way a task's FPSIMD/SVE state is stored
221 * The kernel uses this flag to track whether a user task is actively
222 * using SVE, and therefore whether full SVE register state needs to
223 * be tracked. If not, the cheaper FPSIMD context handling code can
224 * be used instead of the more costly SVE equivalents.
228 * The task can execute SVE instructions while in userspace without
229 * trapping to the kernel.
231 * When stored, Z0-Z31 (incorporating Vn in bits[127:0] or the
232 * corresponding Zn), P0-P15 and FFR are encoded in in
233 * task->thread.sve_state, formatted appropriately for vector
234 * length task->thread.sve_vl.
236 * task->thread.sve_state must point to a valid buffer at least
237 * sve_state_size(task) bytes in size.
239 * During any syscall, the kernel may optionally clear TIF_SVE and
240 * discard the vector state except for the FPSIMD subset.
244 * An attempt by the user task to execute an SVE instruction causes
245 * do_sve_acc() to be called, which does some preparation and then
248 * When stored, FPSIMD registers V0-V31 are encoded in
249 * task->thread.uw.fpsimd_state; bits [max : 128] for each of Z0-Z31 are
250 * logically zero but not stored anywhere; P0-P15 and FFR are not
251 * stored and have unspecified values from userspace's point of
252 * view. For hygiene purposes, the kernel zeroes them on next use,
253 * but userspace is discouraged from relying on this.
255 * task->thread.sve_state does not need to be non-NULL, valid or any
256 * particular size: it must not be dereferenced.
258 * * FPSR and FPCR are always stored in task->thread.uw.fpsimd_state
259 * irrespective of whether TIF_SVE is clear or set, since these are
260 * not vector length dependent.
264 * Update current's FPSIMD/SVE registers from thread_struct.
266 * This function should be called only when the FPSIMD/SVE state in
267 * thread_struct is known to be up to date, when preparing to enter
270 static void task_fpsimd_load(void)
272 WARN_ON(!system_supports_fpsimd());
273 WARN_ON(!have_cpu_fpsimd_context());
275 if (system_supports_sve() && test_thread_flag(TIF_SVE))
276 sve_load_state(sve_pffr(¤t->thread),
277 ¤t->thread.uw.fpsimd_state.fpsr,
278 sve_vq_from_vl(current->thread.sve_vl) - 1);
280 fpsimd_load_state(¤t->thread.uw.fpsimd_state);
284 * Ensure FPSIMD/SVE storage in memory for the loaded context is up to
285 * date with respect to the CPU registers.
287 static void fpsimd_save(void)
289 struct fpsimd_last_state_struct const *last =
290 this_cpu_ptr(&fpsimd_last_state);
291 /* set by fpsimd_bind_task_to_cpu() or fpsimd_bind_state_to_cpu() */
293 WARN_ON(!system_supports_fpsimd());
294 WARN_ON(!have_cpu_fpsimd_context());
296 if (!test_thread_flag(TIF_FOREIGN_FPSTATE)) {
297 if (system_supports_sve() && test_thread_flag(TIF_SVE)) {
298 if (WARN_ON(sve_get_vl() != last->sve_vl)) {
300 * Can't save the user regs, so current would
301 * re-enter user with corrupt state.
302 * There's no way to recover, so kill it:
304 force_signal_inject(SIGKILL, SI_KERNEL, 0);
308 sve_save_state((char *)last->sve_state +
309 sve_ffr_offset(last->sve_vl),
312 fpsimd_save_state(last->st);
317 * All vector length selection from userspace comes through here.
318 * We're on a slow path, so some sanity-checks are included.
319 * If things go wrong there's a bug somewhere, but try to fall back to a
322 static unsigned int find_supported_vector_length(unsigned int vl)
325 int max_vl = sve_max_vl;
327 if (WARN_ON(!sve_vl_valid(vl)))
330 if (WARN_ON(!sve_vl_valid(max_vl)))
336 bit = find_next_bit(sve_vq_map, SVE_VQ_MAX,
337 __vq_to_bit(sve_vq_from_vl(vl)));
338 return sve_vl_from_vq(__bit_to_vq(bit));
343 static int sve_proc_do_default_vl(struct ctl_table *table, int write,
344 void *buffer, size_t *lenp, loff_t *ppos)
347 int vl = sve_default_vl;
348 struct ctl_table tmp_table = {
350 .maxlen = sizeof(vl),
353 ret = proc_dointvec(&tmp_table, write, buffer, lenp, ppos);
357 /* Writing -1 has the special meaning "set to max": */
361 if (!sve_vl_valid(vl))
364 sve_default_vl = find_supported_vector_length(vl);
368 static struct ctl_table sve_default_vl_table[] = {
370 .procname = "sve_default_vector_length",
372 .proc_handler = sve_proc_do_default_vl,
377 static int __init sve_sysctl_init(void)
379 if (system_supports_sve())
380 if (!register_sysctl("abi", sve_default_vl_table))
386 #else /* ! CONFIG_SYSCTL */
387 static int __init sve_sysctl_init(void) { return 0; }
388 #endif /* ! CONFIG_SYSCTL */
390 #define ZREG(sve_state, vq, n) ((char *)(sve_state) + \
391 (SVE_SIG_ZREG_OFFSET(vq, n) - SVE_SIG_REGS_OFFSET))
393 #ifdef CONFIG_CPU_BIG_ENDIAN
394 static __uint128_t arm64_cpu_to_le128(__uint128_t x)
397 u64 b = swab64(x >> 64);
399 return ((__uint128_t)a << 64) | b;
402 static __uint128_t arm64_cpu_to_le128(__uint128_t x)
408 #define arm64_le128_to_cpu(x) arm64_cpu_to_le128(x)
410 static void __fpsimd_to_sve(void *sst, struct user_fpsimd_state const *fst,
416 for (i = 0; i < SVE_NUM_ZREGS; ++i) {
417 p = (__uint128_t *)ZREG(sst, vq, i);
418 *p = arm64_cpu_to_le128(fst->vregs[i]);
423 * Transfer the FPSIMD state in task->thread.uw.fpsimd_state to
424 * task->thread.sve_state.
426 * Task can be a non-runnable task, or current. In the latter case,
427 * the caller must have ownership of the cpu FPSIMD context before calling
429 * task->thread.sve_state must point to at least sve_state_size(task)
430 * bytes of allocated kernel memory.
431 * task->thread.uw.fpsimd_state must be up to date before calling this
434 static void fpsimd_to_sve(struct task_struct *task)
437 void *sst = task->thread.sve_state;
438 struct user_fpsimd_state const *fst = &task->thread.uw.fpsimd_state;
440 if (!system_supports_sve())
443 vq = sve_vq_from_vl(task->thread.sve_vl);
444 __fpsimd_to_sve(sst, fst, vq);
448 * Transfer the SVE state in task->thread.sve_state to
449 * task->thread.uw.fpsimd_state.
451 * Task can be a non-runnable task, or current. In the latter case,
452 * the caller must have ownership of the cpu FPSIMD context before calling
454 * task->thread.sve_state must point to at least sve_state_size(task)
455 * bytes of allocated kernel memory.
456 * task->thread.sve_state must be up to date before calling this function.
458 static void sve_to_fpsimd(struct task_struct *task)
461 void const *sst = task->thread.sve_state;
462 struct user_fpsimd_state *fst = &task->thread.uw.fpsimd_state;
464 __uint128_t const *p;
466 if (!system_supports_sve())
469 vq = sve_vq_from_vl(task->thread.sve_vl);
470 for (i = 0; i < SVE_NUM_ZREGS; ++i) {
471 p = (__uint128_t const *)ZREG(sst, vq, i);
472 fst->vregs[i] = arm64_le128_to_cpu(*p);
476 #ifdef CONFIG_ARM64_SVE
479 * Return how many bytes of memory are required to store the full SVE
480 * state for task, given task's currently configured vector length.
482 size_t sve_state_size(struct task_struct const *task)
484 return SVE_SIG_REGS_SIZE(sve_vq_from_vl(task->thread.sve_vl));
488 * Ensure that task->thread.sve_state is allocated and sufficiently large.
490 * This function should be used only in preparation for replacing
491 * task->thread.sve_state with new data. The memory is always zeroed
492 * here to prevent stale data from showing through: this is done in
493 * the interest of testability and predictability: except in the
494 * do_sve_acc() case, there is no ABI requirement to hide stale data
495 * written previously be task.
497 void sve_alloc(struct task_struct *task)
499 if (task->thread.sve_state) {
500 memset(task->thread.sve_state, 0, sve_state_size(current));
504 /* This is a small allocation (maximum ~8KB) and Should Not Fail. */
505 task->thread.sve_state =
506 kzalloc(sve_state_size(task), GFP_KERNEL);
509 * If future SVE revisions can have larger vectors though,
510 * this may cease to be true:
512 BUG_ON(!task->thread.sve_state);
517 * Ensure that task->thread.sve_state is up to date with respect to
518 * the user task, irrespective of when SVE is in use or not.
520 * This should only be called by ptrace. task must be non-runnable.
521 * task->thread.sve_state must point to at least sve_state_size(task)
522 * bytes of allocated kernel memory.
524 void fpsimd_sync_to_sve(struct task_struct *task)
526 if (!test_tsk_thread_flag(task, TIF_SVE))
531 * Ensure that task->thread.uw.fpsimd_state is up to date with respect to
532 * the user task, irrespective of whether SVE is in use or not.
534 * This should only be called by ptrace. task must be non-runnable.
535 * task->thread.sve_state must point to at least sve_state_size(task)
536 * bytes of allocated kernel memory.
538 void sve_sync_to_fpsimd(struct task_struct *task)
540 if (test_tsk_thread_flag(task, TIF_SVE))
545 * Ensure that task->thread.sve_state is up to date with respect to
546 * the task->thread.uw.fpsimd_state.
548 * This should only be called by ptrace to merge new FPSIMD register
549 * values into a task for which SVE is currently active.
550 * task must be non-runnable.
551 * task->thread.sve_state must point to at least sve_state_size(task)
552 * bytes of allocated kernel memory.
553 * task->thread.uw.fpsimd_state must already have been initialised with
554 * the new FPSIMD register values to be merged in.
556 void sve_sync_from_fpsimd_zeropad(struct task_struct *task)
559 void *sst = task->thread.sve_state;
560 struct user_fpsimd_state const *fst = &task->thread.uw.fpsimd_state;
562 if (!test_tsk_thread_flag(task, TIF_SVE))
565 vq = sve_vq_from_vl(task->thread.sve_vl);
567 memset(sst, 0, SVE_SIG_REGS_SIZE(vq));
568 __fpsimd_to_sve(sst, fst, vq);
571 int sve_set_vector_length(struct task_struct *task,
572 unsigned long vl, unsigned long flags)
574 if (flags & ~(unsigned long)(PR_SVE_VL_INHERIT |
575 PR_SVE_SET_VL_ONEXEC))
578 if (!sve_vl_valid(vl))
582 * Clamp to the maximum vector length that VL-agnostic SVE code can
583 * work with. A flag may be assigned in the future to allow setting
584 * of larger vector lengths without confusing older software.
586 if (vl > SVE_VL_ARCH_MAX)
587 vl = SVE_VL_ARCH_MAX;
589 vl = find_supported_vector_length(vl);
591 if (flags & (PR_SVE_VL_INHERIT |
592 PR_SVE_SET_VL_ONEXEC))
593 task->thread.sve_vl_onexec = vl;
595 /* Reset VL to system default on next exec: */
596 task->thread.sve_vl_onexec = 0;
598 /* Only actually set the VL if not deferred: */
599 if (flags & PR_SVE_SET_VL_ONEXEC)
602 if (vl == task->thread.sve_vl)
606 * To ensure the FPSIMD bits of the SVE vector registers are preserved,
607 * write any live register state back to task_struct, and convert to a
610 if (task == current) {
611 get_cpu_fpsimd_context();
616 fpsimd_flush_task_state(task);
617 if (test_and_clear_tsk_thread_flag(task, TIF_SVE))
621 put_cpu_fpsimd_context();
624 * Force reallocation of task SVE state to the correct size
629 task->thread.sve_vl = vl;
632 update_tsk_thread_flag(task, TIF_SVE_VL_INHERIT,
633 flags & PR_SVE_VL_INHERIT);
639 * Encode the current vector length and flags for return.
640 * This is only required for prctl(): ptrace has separate fields
642 * flags are as for sve_set_vector_length().
644 static int sve_prctl_status(unsigned long flags)
648 if (flags & PR_SVE_SET_VL_ONEXEC)
649 ret = current->thread.sve_vl_onexec;
651 ret = current->thread.sve_vl;
653 if (test_thread_flag(TIF_SVE_VL_INHERIT))
654 ret |= PR_SVE_VL_INHERIT;
660 int sve_set_current_vl(unsigned long arg)
662 unsigned long vl, flags;
665 vl = arg & PR_SVE_VL_LEN_MASK;
668 if (!system_supports_sve())
671 ret = sve_set_vector_length(current, vl, flags);
675 return sve_prctl_status(flags);
679 int sve_get_current_vl(void)
681 if (!system_supports_sve())
684 return sve_prctl_status(0);
687 static void sve_probe_vqs(DECLARE_BITMAP(map, SVE_VQ_MAX))
692 bitmap_zero(map, SVE_VQ_MAX);
694 zcr = ZCR_ELx_LEN_MASK;
695 zcr = read_sysreg_s(SYS_ZCR_EL1) & ~zcr;
697 for (vq = SVE_VQ_MAX; vq >= SVE_VQ_MIN; --vq) {
698 write_sysreg_s(zcr | (vq - 1), SYS_ZCR_EL1); /* self-syncing */
700 vq = sve_vq_from_vl(vl); /* skip intervening lengths */
701 set_bit(__vq_to_bit(vq), map);
706 * Initialise the set of known supported VQs for the boot CPU.
707 * This is called during kernel boot, before secondary CPUs are brought up.
709 void __init sve_init_vq_map(void)
711 sve_probe_vqs(sve_vq_map);
712 bitmap_copy(sve_vq_partial_map, sve_vq_map, SVE_VQ_MAX);
716 * If we haven't committed to the set of supported VQs yet, filter out
717 * those not supported by the current CPU.
718 * This function is called during the bring-up of early secondary CPUs only.
720 void sve_update_vq_map(void)
722 DECLARE_BITMAP(tmp_map, SVE_VQ_MAX);
724 sve_probe_vqs(tmp_map);
725 bitmap_and(sve_vq_map, sve_vq_map, tmp_map, SVE_VQ_MAX);
726 bitmap_or(sve_vq_partial_map, sve_vq_partial_map, tmp_map, SVE_VQ_MAX);
730 * Check whether the current CPU supports all VQs in the committed set.
731 * This function is called during the bring-up of late secondary CPUs only.
733 int sve_verify_vq_map(void)
735 DECLARE_BITMAP(tmp_map, SVE_VQ_MAX);
738 sve_probe_vqs(tmp_map);
740 bitmap_complement(tmp_map, tmp_map, SVE_VQ_MAX);
741 if (bitmap_intersects(tmp_map, sve_vq_map, SVE_VQ_MAX)) {
742 pr_warn("SVE: cpu%d: Required vector length(s) missing\n",
747 if (!IS_ENABLED(CONFIG_KVM) || !is_hyp_mode_available())
751 * For KVM, it is necessary to ensure that this CPU doesn't
752 * support any vector length that guests may have probed as
756 /* Recover the set of supported VQs: */
757 bitmap_complement(tmp_map, tmp_map, SVE_VQ_MAX);
758 /* Find VQs supported that are not globally supported: */
759 bitmap_andnot(tmp_map, tmp_map, sve_vq_map, SVE_VQ_MAX);
761 /* Find the lowest such VQ, if any: */
762 b = find_last_bit(tmp_map, SVE_VQ_MAX);
764 return 0; /* no mismatches */
767 * Mismatches above sve_max_virtualisable_vl are fine, since
768 * no guest is allowed to configure ZCR_EL2.LEN to exceed this:
770 if (sve_vl_from_vq(__bit_to_vq(b)) <= sve_max_virtualisable_vl) {
771 pr_warn("SVE: cpu%d: Unsupported vector length(s) present\n",
779 static void __init sve_efi_setup(void)
781 if (!IS_ENABLED(CONFIG_EFI))
785 * alloc_percpu() warns and prints a backtrace if this goes wrong.
786 * This is evidence of a crippled system and we are returning void,
787 * so no attempt is made to handle this situation here.
789 if (!sve_vl_valid(sve_max_vl))
792 efi_sve_state = __alloc_percpu(
793 SVE_SIG_REGS_SIZE(sve_vq_from_vl(sve_max_vl)), SVE_VQ_BYTES);
800 panic("Cannot allocate percpu memory for EFI SVE save/restore");
804 * Enable SVE for EL1.
805 * Intended for use by the cpufeatures code during CPU boot.
807 void sve_kernel_enable(const struct arm64_cpu_capabilities *__always_unused p)
809 write_sysreg(read_sysreg(CPACR_EL1) | CPACR_EL1_ZEN_EL1EN, CPACR_EL1);
814 * Read the pseudo-ZCR used by cpufeatures to identify the supported SVE
817 * Use only if SVE is present.
818 * This function clobbers the SVE vector length.
820 u64 read_zcr_features(void)
826 * Set the maximum possible VL, and write zeroes to all other
827 * bits to see if they stick.
829 sve_kernel_enable(NULL);
830 write_sysreg_s(ZCR_ELx_LEN_MASK, SYS_ZCR_EL1);
832 zcr = read_sysreg_s(SYS_ZCR_EL1);
833 zcr &= ~(u64)ZCR_ELx_LEN_MASK; /* find sticky 1s outside LEN field */
834 vq_max = sve_vq_from_vl(sve_get_vl());
835 zcr |= vq_max - 1; /* set LEN field to maximum effective value */
840 void __init sve_setup(void)
843 DECLARE_BITMAP(tmp_map, SVE_VQ_MAX);
846 if (!system_supports_sve())
850 * The SVE architecture mandates support for 128-bit vectors,
851 * so sve_vq_map must have at least SVE_VQ_MIN set.
852 * If something went wrong, at least try to patch it up:
854 if (WARN_ON(!test_bit(__vq_to_bit(SVE_VQ_MIN), sve_vq_map)))
855 set_bit(__vq_to_bit(SVE_VQ_MIN), sve_vq_map);
857 zcr = read_sanitised_ftr_reg(SYS_ZCR_EL1);
858 sve_max_vl = sve_vl_from_vq((zcr & ZCR_ELx_LEN_MASK) + 1);
861 * Sanity-check that the max VL we determined through CPU features
862 * corresponds properly to sve_vq_map. If not, do our best:
864 if (WARN_ON(sve_max_vl != find_supported_vector_length(sve_max_vl)))
865 sve_max_vl = find_supported_vector_length(sve_max_vl);
868 * For the default VL, pick the maximum supported value <= 64.
869 * VL == 64 is guaranteed not to grow the signal frame.
871 sve_default_vl = find_supported_vector_length(64);
873 bitmap_andnot(tmp_map, sve_vq_partial_map, sve_vq_map,
876 b = find_last_bit(tmp_map, SVE_VQ_MAX);
878 /* No non-virtualisable VLs found */
879 sve_max_virtualisable_vl = SVE_VQ_MAX;
880 else if (WARN_ON(b == SVE_VQ_MAX - 1))
881 /* No virtualisable VLs? This is architecturally forbidden. */
882 sve_max_virtualisable_vl = SVE_VQ_MIN;
883 else /* b + 1 < SVE_VQ_MAX */
884 sve_max_virtualisable_vl = sve_vl_from_vq(__bit_to_vq(b + 1));
886 if (sve_max_virtualisable_vl > sve_max_vl)
887 sve_max_virtualisable_vl = sve_max_vl;
889 pr_info("SVE: maximum available vector length %u bytes per vector\n",
891 pr_info("SVE: default vector length %u bytes per vector\n",
894 /* KVM decides whether to support mismatched systems. Just warn here: */
895 if (sve_max_virtualisable_vl < sve_max_vl)
896 pr_warn("SVE: unvirtualisable vector lengths present\n");
902 * Called from the put_task_struct() path, which cannot get here
903 * unless dead_task is really dead and not schedulable.
905 void fpsimd_release_task(struct task_struct *dead_task)
907 __sve_free(dead_task);
910 #endif /* CONFIG_ARM64_SVE */
915 * Storage is allocated for the full SVE state, the current FPSIMD
916 * register contents are migrated across, and TIF_SVE is set so that
917 * the SVE access trap will be disabled the next time this task
918 * reaches ret_to_user.
920 * TIF_SVE should be clear on entry: otherwise, task_fpsimd_load()
921 * would have disabled the SVE access trap for userspace during
922 * ret_to_user, making an SVE access trap impossible in that case.
924 void do_sve_acc(unsigned int esr, struct pt_regs *regs)
926 /* Even if we chose not to use SVE, the hardware could still trap: */
927 if (unlikely(!system_supports_sve()) || WARN_ON(is_compat_task())) {
928 force_signal_inject(SIGILL, ILL_ILLOPC, regs->pc);
934 get_cpu_fpsimd_context();
938 /* Force ret_to_user to reload the registers: */
939 fpsimd_flush_task_state(current);
941 fpsimd_to_sve(current);
942 if (test_and_set_thread_flag(TIF_SVE))
943 WARN_ON(1); /* SVE access shouldn't have trapped */
945 put_cpu_fpsimd_context();
949 * Trapped FP/ASIMD access.
951 void do_fpsimd_acc(unsigned int esr, struct pt_regs *regs)
953 /* TODO: implement lazy context saving/restoring */
958 * Raise a SIGFPE for the current process.
960 void do_fpsimd_exc(unsigned int esr, struct pt_regs *regs)
962 unsigned int si_code = FPE_FLTUNK;
964 if (esr & ESR_ELx_FP_EXC_TFV) {
966 si_code = FPE_FLTINV;
967 else if (esr & FPEXC_DZF)
968 si_code = FPE_FLTDIV;
969 else if (esr & FPEXC_OFF)
970 si_code = FPE_FLTOVF;
971 else if (esr & FPEXC_UFF)
972 si_code = FPE_FLTUND;
973 else if (esr & FPEXC_IXF)
974 si_code = FPE_FLTRES;
977 send_sig_fault(SIGFPE, si_code,
978 (void __user *)instruction_pointer(regs),
982 void fpsimd_thread_switch(struct task_struct *next)
984 bool wrong_task, wrong_cpu;
986 if (!system_supports_fpsimd())
989 __get_cpu_fpsimd_context();
991 /* Save unsaved fpsimd state, if any: */
995 * Fix up TIF_FOREIGN_FPSTATE to correctly describe next's
996 * state. For kernel threads, FPSIMD registers are never loaded
997 * and wrong_task and wrong_cpu will always be true.
999 wrong_task = __this_cpu_read(fpsimd_last_state.st) !=
1000 &next->thread.uw.fpsimd_state;
1001 wrong_cpu = next->thread.fpsimd_cpu != smp_processor_id();
1003 update_tsk_thread_flag(next, TIF_FOREIGN_FPSTATE,
1004 wrong_task || wrong_cpu);
1006 __put_cpu_fpsimd_context();
1009 void fpsimd_flush_thread(void)
1011 int vl, supported_vl;
1013 if (!system_supports_fpsimd())
1016 get_cpu_fpsimd_context();
1018 fpsimd_flush_task_state(current);
1019 memset(¤t->thread.uw.fpsimd_state, 0,
1020 sizeof(current->thread.uw.fpsimd_state));
1022 if (system_supports_sve()) {
1023 clear_thread_flag(TIF_SVE);
1027 * Reset the task vector length as required.
1028 * This is where we ensure that all user tasks have a valid
1029 * vector length configured: no kernel task can become a user
1030 * task without an exec and hence a call to this function.
1031 * By the time the first call to this function is made, all
1032 * early hardware probing is complete, so sve_default_vl
1034 * If a bug causes this to go wrong, we make some noise and
1035 * try to fudge thread.sve_vl to a safe value here.
1037 vl = current->thread.sve_vl_onexec ?
1038 current->thread.sve_vl_onexec : sve_default_vl;
1040 if (WARN_ON(!sve_vl_valid(vl)))
1043 supported_vl = find_supported_vector_length(vl);
1044 if (WARN_ON(supported_vl != vl))
1047 current->thread.sve_vl = vl;
1050 * If the task is not set to inherit, ensure that the vector
1051 * length will be reset by a subsequent exec:
1053 if (!test_thread_flag(TIF_SVE_VL_INHERIT))
1054 current->thread.sve_vl_onexec = 0;
1057 put_cpu_fpsimd_context();
1061 * Save the userland FPSIMD state of 'current' to memory, but only if the state
1062 * currently held in the registers does in fact belong to 'current'
1064 void fpsimd_preserve_current_state(void)
1066 if (!system_supports_fpsimd())
1069 get_cpu_fpsimd_context();
1071 put_cpu_fpsimd_context();
1075 * Like fpsimd_preserve_current_state(), but ensure that
1076 * current->thread.uw.fpsimd_state is updated so that it can be copied to
1079 void fpsimd_signal_preserve_current_state(void)
1081 fpsimd_preserve_current_state();
1082 if (system_supports_sve() && test_thread_flag(TIF_SVE))
1083 sve_to_fpsimd(current);
1087 * Associate current's FPSIMD context with this cpu
1088 * The caller must have ownership of the cpu FPSIMD context before calling
1091 void fpsimd_bind_task_to_cpu(void)
1093 struct fpsimd_last_state_struct *last =
1094 this_cpu_ptr(&fpsimd_last_state);
1096 WARN_ON(!system_supports_fpsimd());
1097 last->st = ¤t->thread.uw.fpsimd_state;
1098 last->sve_state = current->thread.sve_state;
1099 last->sve_vl = current->thread.sve_vl;
1100 current->thread.fpsimd_cpu = smp_processor_id();
1102 if (system_supports_sve()) {
1103 /* Toggle SVE trapping for userspace if needed */
1104 if (test_thread_flag(TIF_SVE))
1109 /* Serialised by exception return to user */
1113 void fpsimd_bind_state_to_cpu(struct user_fpsimd_state *st, void *sve_state,
1114 unsigned int sve_vl)
1116 struct fpsimd_last_state_struct *last =
1117 this_cpu_ptr(&fpsimd_last_state);
1119 WARN_ON(!system_supports_fpsimd());
1120 WARN_ON(!in_softirq() && !irqs_disabled());
1123 last->sve_state = sve_state;
1124 last->sve_vl = sve_vl;
1128 * Load the userland FPSIMD state of 'current' from memory, but only if the
1129 * FPSIMD state already held in the registers is /not/ the most recent FPSIMD
1130 * state of 'current'
1132 void fpsimd_restore_current_state(void)
1135 * For the tasks that were created before we detected the absence of
1136 * FP/SIMD, the TIF_FOREIGN_FPSTATE could be set via fpsimd_thread_switch(),
1137 * e.g, init. This could be then inherited by the children processes.
1138 * If we later detect that the system doesn't support FP/SIMD,
1139 * we must clear the flag for all the tasks to indicate that the
1140 * FPSTATE is clean (as we can't have one) to avoid looping for ever in
1141 * do_notify_resume().
1143 if (!system_supports_fpsimd()) {
1144 clear_thread_flag(TIF_FOREIGN_FPSTATE);
1148 get_cpu_fpsimd_context();
1150 if (test_and_clear_thread_flag(TIF_FOREIGN_FPSTATE)) {
1152 fpsimd_bind_task_to_cpu();
1155 put_cpu_fpsimd_context();
1159 * Load an updated userland FPSIMD state for 'current' from memory and set the
1160 * flag that indicates that the FPSIMD register contents are the most recent
1161 * FPSIMD state of 'current'
1163 void fpsimd_update_current_state(struct user_fpsimd_state const *state)
1165 if (WARN_ON(!system_supports_fpsimd()))
1168 get_cpu_fpsimd_context();
1170 current->thread.uw.fpsimd_state = *state;
1171 if (system_supports_sve() && test_thread_flag(TIF_SVE))
1172 fpsimd_to_sve(current);
1175 fpsimd_bind_task_to_cpu();
1177 clear_thread_flag(TIF_FOREIGN_FPSTATE);
1179 put_cpu_fpsimd_context();
1183 * Invalidate live CPU copies of task t's FPSIMD state
1185 * This function may be called with preemption enabled. The barrier()
1186 * ensures that the assignment to fpsimd_cpu is visible to any
1187 * preemption/softirq that could race with set_tsk_thread_flag(), so
1188 * that TIF_FOREIGN_FPSTATE cannot be spuriously re-cleared.
1190 * The final barrier ensures that TIF_FOREIGN_FPSTATE is seen set by any
1193 void fpsimd_flush_task_state(struct task_struct *t)
1195 t->thread.fpsimd_cpu = NR_CPUS;
1197 * If we don't support fpsimd, bail out after we have
1198 * reset the fpsimd_cpu for this task and clear the
1201 if (!system_supports_fpsimd())
1204 set_tsk_thread_flag(t, TIF_FOREIGN_FPSTATE);
1210 * Invalidate any task's FPSIMD state that is present on this cpu.
1211 * The FPSIMD context should be acquired with get_cpu_fpsimd_context()
1212 * before calling this function.
1214 static void fpsimd_flush_cpu_state(void)
1216 WARN_ON(!system_supports_fpsimd());
1217 __this_cpu_write(fpsimd_last_state.st, NULL);
1218 set_thread_flag(TIF_FOREIGN_FPSTATE);
1222 * Save the FPSIMD state to memory and invalidate cpu view.
1223 * This function must be called with preemption disabled.
1225 void fpsimd_save_and_flush_cpu_state(void)
1227 if (!system_supports_fpsimd())
1229 WARN_ON(preemptible());
1230 __get_cpu_fpsimd_context();
1232 fpsimd_flush_cpu_state();
1233 __put_cpu_fpsimd_context();
1236 #ifdef CONFIG_KERNEL_MODE_NEON
1239 * Kernel-side NEON support functions
1243 * kernel_neon_begin(): obtain the CPU FPSIMD registers for use by the calling
1246 * Must not be called unless may_use_simd() returns true.
1247 * Task context in the FPSIMD registers is saved back to memory as necessary.
1249 * A matching call to kernel_neon_end() must be made before returning from the
1252 * The caller may freely use the FPSIMD registers until kernel_neon_end() is
1255 void kernel_neon_begin(void)
1257 if (WARN_ON(!system_supports_fpsimd()))
1260 BUG_ON(!may_use_simd());
1262 get_cpu_fpsimd_context();
1264 /* Save unsaved fpsimd state, if any: */
1267 /* Invalidate any task state remaining in the fpsimd regs: */
1268 fpsimd_flush_cpu_state();
1270 EXPORT_SYMBOL(kernel_neon_begin);
1273 * kernel_neon_end(): give the CPU FPSIMD registers back to the current task
1275 * Must be called from a context in which kernel_neon_begin() was previously
1276 * called, with no call to kernel_neon_end() in the meantime.
1278 * The caller must not use the FPSIMD registers after this function is called,
1279 * unless kernel_neon_begin() is called again in the meantime.
1281 void kernel_neon_end(void)
1283 if (!system_supports_fpsimd())
1286 put_cpu_fpsimd_context();
1288 EXPORT_SYMBOL(kernel_neon_end);
1292 static DEFINE_PER_CPU(struct user_fpsimd_state, efi_fpsimd_state);
1293 static DEFINE_PER_CPU(bool, efi_fpsimd_state_used);
1294 static DEFINE_PER_CPU(bool, efi_sve_state_used);
1297 * EFI runtime services support functions
1299 * The ABI for EFI runtime services allows EFI to use FPSIMD during the call.
1300 * This means that for EFI (and only for EFI), we have to assume that FPSIMD
1301 * is always used rather than being an optional accelerator.
1303 * These functions provide the necessary support for ensuring FPSIMD
1304 * save/restore in the contexts from which EFI is used.
1306 * Do not use them for any other purpose -- if tempted to do so, you are
1307 * either doing something wrong or you need to propose some refactoring.
1311 * __efi_fpsimd_begin(): prepare FPSIMD for making an EFI runtime services call
1313 void __efi_fpsimd_begin(void)
1315 if (!system_supports_fpsimd())
1318 WARN_ON(preemptible());
1320 if (may_use_simd()) {
1321 kernel_neon_begin();
1324 * If !efi_sve_state, SVE can't be in use yet and doesn't need
1327 if (system_supports_sve() && likely(efi_sve_state)) {
1328 char *sve_state = this_cpu_ptr(efi_sve_state);
1330 __this_cpu_write(efi_sve_state_used, true);
1332 sve_save_state(sve_state + sve_ffr_offset(sve_max_vl),
1333 &this_cpu_ptr(&efi_fpsimd_state)->fpsr);
1335 fpsimd_save_state(this_cpu_ptr(&efi_fpsimd_state));
1338 __this_cpu_write(efi_fpsimd_state_used, true);
1343 * __efi_fpsimd_end(): clean up FPSIMD after an EFI runtime services call
1345 void __efi_fpsimd_end(void)
1347 if (!system_supports_fpsimd())
1350 if (!__this_cpu_xchg(efi_fpsimd_state_used, false)) {
1353 if (system_supports_sve() &&
1354 likely(__this_cpu_read(efi_sve_state_used))) {
1355 char const *sve_state = this_cpu_ptr(efi_sve_state);
1357 sve_load_state(sve_state + sve_ffr_offset(sve_max_vl),
1358 &this_cpu_ptr(&efi_fpsimd_state)->fpsr,
1359 sve_vq_from_vl(sve_get_vl()) - 1);
1361 __this_cpu_write(efi_sve_state_used, false);
1363 fpsimd_load_state(this_cpu_ptr(&efi_fpsimd_state));
1368 #endif /* CONFIG_EFI */
1370 #endif /* CONFIG_KERNEL_MODE_NEON */
1372 #ifdef CONFIG_CPU_PM
1373 static int fpsimd_cpu_pm_notifier(struct notifier_block *self,
1374 unsigned long cmd, void *v)
1378 fpsimd_save_and_flush_cpu_state();
1382 case CPU_PM_ENTER_FAILED:
1389 static struct notifier_block fpsimd_cpu_pm_notifier_block = {
1390 .notifier_call = fpsimd_cpu_pm_notifier,
1393 static void __init fpsimd_pm_init(void)
1395 cpu_pm_register_notifier(&fpsimd_cpu_pm_notifier_block);
1399 static inline void fpsimd_pm_init(void) { }
1400 #endif /* CONFIG_CPU_PM */
1402 #ifdef CONFIG_HOTPLUG_CPU
1403 static int fpsimd_cpu_dead(unsigned int cpu)
1405 per_cpu(fpsimd_last_state.st, cpu) = NULL;
1409 static inline void fpsimd_hotplug_init(void)
1411 cpuhp_setup_state_nocalls(CPUHP_ARM64_FPSIMD_DEAD, "arm64/fpsimd:dead",
1412 NULL, fpsimd_cpu_dead);
1416 static inline void fpsimd_hotplug_init(void) { }
1420 * FP/SIMD support code initialisation.
1422 static int __init fpsimd_init(void)
1424 if (cpu_have_named_feature(FP)) {
1426 fpsimd_hotplug_init();
1428 pr_notice("Floating-point is not implemented\n");
1431 if (!cpu_have_named_feature(ASIMD))
1432 pr_notice("Advanced SIMD is not implemented\n");
1434 return sve_sysctl_init();
1436 core_initcall(fpsimd_init);