2 * FP/SIMD context switching and fault handling
4 * Copyright (C) 2012 ARM Ltd.
5 * Author: Catalin Marinas <catalin.marinas@arm.com>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program. If not, see <http://www.gnu.org/licenses/>.
20 #include <linux/bitmap.h>
21 #include <linux/bottom_half.h>
22 #include <linux/bug.h>
23 #include <linux/cache.h>
24 #include <linux/compat.h>
25 #include <linux/cpu.h>
26 #include <linux/cpu_pm.h>
27 #include <linux/kernel.h>
28 #include <linux/linkage.h>
29 #include <linux/irqflags.h>
30 #include <linux/init.h>
31 #include <linux/percpu.h>
32 #include <linux/prctl.h>
33 #include <linux/preempt.h>
34 #include <linux/ptrace.h>
35 #include <linux/sched/signal.h>
36 #include <linux/sched/task_stack.h>
37 #include <linux/signal.h>
38 #include <linux/slab.h>
39 #include <linux/stddef.h>
40 #include <linux/sysctl.h>
43 #include <asm/fpsimd.h>
44 #include <asm/cpufeature.h>
45 #include <asm/cputype.h>
46 #include <asm/processor.h>
48 #include <asm/sigcontext.h>
49 #include <asm/sysreg.h>
50 #include <asm/traps.h>
52 #define FPEXC_IOF (1 << 0)
53 #define FPEXC_DZF (1 << 1)
54 #define FPEXC_OFF (1 << 2)
55 #define FPEXC_UFF (1 << 3)
56 #define FPEXC_IXF (1 << 4)
57 #define FPEXC_IDF (1 << 7)
60 * (Note: in this discussion, statements about FPSIMD apply equally to SVE.)
62 * In order to reduce the number of times the FPSIMD state is needlessly saved
63 * and restored, we need to keep track of two things:
64 * (a) for each task, we need to remember which CPU was the last one to have
65 * the task's FPSIMD state loaded into its FPSIMD registers;
66 * (b) for each CPU, we need to remember which task's userland FPSIMD state has
67 * been loaded into its FPSIMD registers most recently, or whether it has
68 * been used to perform kernel mode NEON in the meantime.
70 * For (a), we add a fpsimd_cpu field to thread_struct, which gets updated to
71 * the id of the current CPU every time the state is loaded onto a CPU. For (b),
72 * we add the per-cpu variable 'fpsimd_last_state' (below), which contains the
73 * address of the userland FPSIMD state of the task that was loaded onto the CPU
74 * the most recently, or NULL if kernel mode NEON has been performed after that.
76 * With this in place, we no longer have to restore the next FPSIMD state right
77 * when switching between tasks. Instead, we can defer this check to userland
78 * resume, at which time we verify whether the CPU's fpsimd_last_state and the
79 * task's fpsimd_cpu are still mutually in sync. If this is the case, we
80 * can omit the FPSIMD restore.
82 * As an optimization, we use the thread_info flag TIF_FOREIGN_FPSTATE to
83 * indicate whether or not the userland FPSIMD state of the current task is
84 * present in the registers. The flag is set unless the FPSIMD registers of this
85 * CPU currently contain the most recent userland FPSIMD state of the current
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 local_bh_disable() unless softirqs are already masked.
95 * For a certain task, the sequence may look something like this:
96 * - the task gets scheduled in; if both the task's fpsimd_cpu field
97 * contains the id of the current CPU, and the CPU's fpsimd_last_state per-cpu
98 * variable points to the task's fpsimd_state, the TIF_FOREIGN_FPSTATE flag is
99 * cleared, otherwise it is set;
101 * - the task returns to userland; if TIF_FOREIGN_FPSTATE is set, the task's
102 * userland FPSIMD state is copied from memory to the registers, the task's
103 * fpsimd_cpu field is set to the id of the current CPU, the current
104 * CPU's fpsimd_last_state pointer is set to this task's fpsimd_state and the
105 * TIF_FOREIGN_FPSTATE flag is cleared;
107 * - the task executes an ordinary syscall; upon return to userland, the
108 * TIF_FOREIGN_FPSTATE flag will still be cleared, so no FPSIMD state is
111 * - the task executes a syscall which executes some NEON instructions; this is
112 * preceded by a call to kernel_neon_begin(), which copies the task's FPSIMD
113 * register contents to memory, clears the fpsimd_last_state per-cpu variable
114 * and sets the TIF_FOREIGN_FPSTATE flag;
116 * - the task gets preempted after kernel_neon_end() is called; as we have not
117 * returned from the 2nd syscall yet, TIF_FOREIGN_FPSTATE is still set so
118 * whatever is in the FPSIMD registers is not saved to memory, but discarded.
120 struct fpsimd_last_state_struct {
121 struct user_fpsimd_state *st;
124 static DEFINE_PER_CPU(struct fpsimd_last_state_struct, fpsimd_last_state);
126 /* Default VL for tasks that don't set it explicitly: */
127 static int sve_default_vl = -1;
129 #ifdef CONFIG_ARM64_SVE
131 /* Maximum supported vector length across all CPUs (initially poisoned) */
132 int __ro_after_init sve_max_vl = SVE_VL_MIN;
133 /* Set of available vector lengths, as vq_to_bit(vq): */
134 static __ro_after_init DECLARE_BITMAP(sve_vq_map, SVE_VQ_MAX);
135 static void __percpu *efi_sve_state;
137 #else /* ! CONFIG_ARM64_SVE */
139 /* Dummy declaration for code that will be optimised out: */
140 extern __ro_after_init DECLARE_BITMAP(sve_vq_map, SVE_VQ_MAX);
141 extern void __percpu *efi_sve_state;
143 #endif /* ! CONFIG_ARM64_SVE */
146 * Call __sve_free() directly only if you know task can't be scheduled
149 static void __sve_free(struct task_struct *task)
151 kfree(task->thread.sve_state);
152 task->thread.sve_state = NULL;
155 static void sve_free(struct task_struct *task)
157 WARN_ON(test_tsk_thread_flag(task, TIF_SVE));
163 * TIF_SVE controls whether a task can use SVE without trapping while
164 * in userspace, and also the way a task's FPSIMD/SVE state is stored
167 * The kernel uses this flag to track whether a user task is actively
168 * using SVE, and therefore whether full SVE register state needs to
169 * be tracked. If not, the cheaper FPSIMD context handling code can
170 * be used instead of the more costly SVE equivalents.
174 * The task can execute SVE instructions while in userspace without
175 * trapping to the kernel.
177 * When stored, Z0-Z31 (incorporating Vn in bits[127:0] or the
178 * corresponding Zn), P0-P15 and FFR are encoded in in
179 * task->thread.sve_state, formatted appropriately for vector
180 * length task->thread.sve_vl.
182 * task->thread.sve_state must point to a valid buffer at least
183 * sve_state_size(task) bytes in size.
185 * During any syscall, the kernel may optionally clear TIF_SVE and
186 * discard the vector state except for the FPSIMD subset.
190 * An attempt by the user task to execute an SVE instruction causes
191 * do_sve_acc() to be called, which does some preparation and then
194 * When stored, FPSIMD registers V0-V31 are encoded in
195 * task->thread.uw.fpsimd_state; bits [max : 128] for each of Z0-Z31 are
196 * logically zero but not stored anywhere; P0-P15 and FFR are not
197 * stored and have unspecified values from userspace's point of
198 * view. For hygiene purposes, the kernel zeroes them on next use,
199 * but userspace is discouraged from relying on this.
201 * task->thread.sve_state does not need to be non-NULL, valid or any
202 * particular size: it must not be dereferenced.
204 * * FPSR and FPCR are always stored in task->thread.uw.fpsimd_state
205 * irrespective of whether TIF_SVE is clear or set, since these are
206 * not vector length dependent.
210 * Update current's FPSIMD/SVE registers from thread_struct.
212 * This function should be called only when the FPSIMD/SVE state in
213 * thread_struct is known to be up to date, when preparing to enter
216 * Softirqs (and preemption) must be disabled.
218 static void task_fpsimd_load(void)
220 WARN_ON(!in_softirq() && !irqs_disabled());
221 WARN_ON(!system_supports_fpsimd());
223 if (system_supports_sve() && test_thread_flag(TIF_SVE))
224 sve_load_state(sve_pffr(¤t->thread),
225 ¤t->thread.uw.fpsimd_state.fpsr,
226 sve_vq_from_vl(current->thread.sve_vl) - 1);
228 fpsimd_load_state(¤t->thread.uw.fpsimd_state);
232 * Ensure FPSIMD/SVE storage in memory for the loaded context is up to
233 * date with respect to the CPU registers.
235 * Softirqs (and preemption) must be disabled.
237 void fpsimd_save(void)
239 struct user_fpsimd_state *st = __this_cpu_read(fpsimd_last_state.st);
240 /* set by fpsimd_bind_task_to_cpu() or fpsimd_bind_state_to_cpu() */
242 WARN_ON(!system_supports_fpsimd());
243 WARN_ON(!in_softirq() && !irqs_disabled());
245 if (!test_thread_flag(TIF_FOREIGN_FPSTATE)) {
246 if (system_supports_sve() && test_thread_flag(TIF_SVE)) {
247 if (WARN_ON(sve_get_vl() != current->thread.sve_vl)) {
249 * Can't save the user regs, so current would
250 * re-enter user with corrupt state.
251 * There's no way to recover, so kill it:
253 force_signal_inject(SIGKILL, SI_KERNEL, 0);
257 sve_save_state(sve_pffr(¤t->thread), &st->fpsr);
259 fpsimd_save_state(st);
264 * Helpers to translate bit indices in sve_vq_map to VQ values (and
265 * vice versa). This allows find_next_bit() to be used to find the
266 * _maximum_ VQ not exceeding a certain value.
269 static unsigned int vq_to_bit(unsigned int vq)
271 return SVE_VQ_MAX - vq;
274 static unsigned int bit_to_vq(unsigned int bit)
276 if (WARN_ON(bit >= SVE_VQ_MAX))
277 bit = SVE_VQ_MAX - 1;
279 return SVE_VQ_MAX - bit;
283 * All vector length selection from userspace comes through here.
284 * We're on a slow path, so some sanity-checks are included.
285 * If things go wrong there's a bug somewhere, but try to fall back to a
288 static unsigned int find_supported_vector_length(unsigned int vl)
291 int max_vl = sve_max_vl;
293 if (WARN_ON(!sve_vl_valid(vl)))
296 if (WARN_ON(!sve_vl_valid(max_vl)))
302 bit = find_next_bit(sve_vq_map, SVE_VQ_MAX,
303 vq_to_bit(sve_vq_from_vl(vl)));
304 return sve_vl_from_vq(bit_to_vq(bit));
307 #if defined(CONFIG_ARM64_SVE) && defined(CONFIG_SYSCTL)
309 static int sve_proc_do_default_vl(struct ctl_table *table, int write,
310 void __user *buffer, size_t *lenp,
314 int vl = sve_default_vl;
315 struct ctl_table tmp_table = {
317 .maxlen = sizeof(vl),
320 ret = proc_dointvec(&tmp_table, write, buffer, lenp, ppos);
324 /* Writing -1 has the special meaning "set to max": */
328 if (!sve_vl_valid(vl))
331 sve_default_vl = find_supported_vector_length(vl);
335 static struct ctl_table sve_default_vl_table[] = {
337 .procname = "sve_default_vector_length",
339 .proc_handler = sve_proc_do_default_vl,
344 static int __init sve_sysctl_init(void)
346 if (system_supports_sve())
347 if (!register_sysctl("abi", sve_default_vl_table))
353 #else /* ! (CONFIG_ARM64_SVE && CONFIG_SYSCTL) */
354 static int __init sve_sysctl_init(void) { return 0; }
355 #endif /* ! (CONFIG_ARM64_SVE && CONFIG_SYSCTL) */
357 #define ZREG(sve_state, vq, n) ((char *)(sve_state) + \
358 (SVE_SIG_ZREG_OFFSET(vq, n) - SVE_SIG_REGS_OFFSET))
361 * Transfer the FPSIMD state in task->thread.uw.fpsimd_state to
362 * task->thread.sve_state.
364 * Task can be a non-runnable task, or current. In the latter case,
365 * softirqs (and preemption) must be disabled.
366 * task->thread.sve_state must point to at least sve_state_size(task)
367 * bytes of allocated kernel memory.
368 * task->thread.uw.fpsimd_state must be up to date before calling this
371 static void fpsimd_to_sve(struct task_struct *task)
374 void *sst = task->thread.sve_state;
375 struct user_fpsimd_state const *fst = &task->thread.uw.fpsimd_state;
378 if (!system_supports_sve())
381 vq = sve_vq_from_vl(task->thread.sve_vl);
382 for (i = 0; i < 32; ++i)
383 memcpy(ZREG(sst, vq, i), &fst->vregs[i],
384 sizeof(fst->vregs[i]));
388 * Transfer the SVE state in task->thread.sve_state to
389 * task->thread.uw.fpsimd_state.
391 * Task can be a non-runnable task, or current. In the latter case,
392 * softirqs (and preemption) must be disabled.
393 * task->thread.sve_state must point to at least sve_state_size(task)
394 * bytes of allocated kernel memory.
395 * task->thread.sve_state must be up to date before calling this function.
397 static void sve_to_fpsimd(struct task_struct *task)
400 void const *sst = task->thread.sve_state;
401 struct user_fpsimd_state *fst = &task->thread.uw.fpsimd_state;
404 if (!system_supports_sve())
407 vq = sve_vq_from_vl(task->thread.sve_vl);
408 for (i = 0; i < 32; ++i)
409 memcpy(&fst->vregs[i], ZREG(sst, vq, i),
410 sizeof(fst->vregs[i]));
413 #ifdef CONFIG_ARM64_SVE
416 * Return how many bytes of memory are required to store the full SVE
417 * state for task, given task's currently configured vector length.
419 size_t sve_state_size(struct task_struct const *task)
421 return SVE_SIG_REGS_SIZE(sve_vq_from_vl(task->thread.sve_vl));
425 * Ensure that task->thread.sve_state is allocated and sufficiently large.
427 * This function should be used only in preparation for replacing
428 * task->thread.sve_state with new data. The memory is always zeroed
429 * here to prevent stale data from showing through: this is done in
430 * the interest of testability and predictability: except in the
431 * do_sve_acc() case, there is no ABI requirement to hide stale data
432 * written previously be task.
434 void sve_alloc(struct task_struct *task)
436 if (task->thread.sve_state) {
437 memset(task->thread.sve_state, 0, sve_state_size(task));
441 /* This is a small allocation (maximum ~8KB) and Should Not Fail. */
442 task->thread.sve_state =
443 kzalloc(sve_state_size(task), GFP_KERNEL);
446 * If future SVE revisions can have larger vectors though,
447 * this may cease to be true:
449 BUG_ON(!task->thread.sve_state);
454 * Ensure that task->thread.sve_state is up to date with respect to
455 * the user task, irrespective of when SVE is in use or not.
457 * This should only be called by ptrace. task must be non-runnable.
458 * task->thread.sve_state must point to at least sve_state_size(task)
459 * bytes of allocated kernel memory.
461 void fpsimd_sync_to_sve(struct task_struct *task)
463 if (!test_tsk_thread_flag(task, TIF_SVE))
468 * Ensure that task->thread.uw.fpsimd_state is up to date with respect to
469 * the user task, irrespective of whether SVE is in use or not.
471 * This should only be called by ptrace. task must be non-runnable.
472 * task->thread.sve_state must point to at least sve_state_size(task)
473 * bytes of allocated kernel memory.
475 void sve_sync_to_fpsimd(struct task_struct *task)
477 if (test_tsk_thread_flag(task, TIF_SVE))
482 * Ensure that task->thread.sve_state is up to date with respect to
483 * the task->thread.uw.fpsimd_state.
485 * This should only be called by ptrace to merge new FPSIMD register
486 * values into a task for which SVE is currently active.
487 * task must be non-runnable.
488 * task->thread.sve_state must point to at least sve_state_size(task)
489 * bytes of allocated kernel memory.
490 * task->thread.uw.fpsimd_state must already have been initialised with
491 * the new FPSIMD register values to be merged in.
493 void sve_sync_from_fpsimd_zeropad(struct task_struct *task)
496 void *sst = task->thread.sve_state;
497 struct user_fpsimd_state const *fst = &task->thread.uw.fpsimd_state;
500 if (!test_tsk_thread_flag(task, TIF_SVE))
503 vq = sve_vq_from_vl(task->thread.sve_vl);
505 memset(sst, 0, SVE_SIG_REGS_SIZE(vq));
507 for (i = 0; i < 32; ++i)
508 memcpy(ZREG(sst, vq, i), &fst->vregs[i],
509 sizeof(fst->vregs[i]));
512 int sve_set_vector_length(struct task_struct *task,
513 unsigned long vl, unsigned long flags)
515 if (flags & ~(unsigned long)(PR_SVE_VL_INHERIT |
516 PR_SVE_SET_VL_ONEXEC))
519 if (!sve_vl_valid(vl))
523 * Clamp to the maximum vector length that VL-agnostic SVE code can
524 * work with. A flag may be assigned in the future to allow setting
525 * of larger vector lengths without confusing older software.
527 if (vl > SVE_VL_ARCH_MAX)
528 vl = SVE_VL_ARCH_MAX;
530 vl = find_supported_vector_length(vl);
532 if (flags & (PR_SVE_VL_INHERIT |
533 PR_SVE_SET_VL_ONEXEC))
534 task->thread.sve_vl_onexec = vl;
536 /* Reset VL to system default on next exec: */
537 task->thread.sve_vl_onexec = 0;
539 /* Only actually set the VL if not deferred: */
540 if (flags & PR_SVE_SET_VL_ONEXEC)
543 if (vl == task->thread.sve_vl)
547 * To ensure the FPSIMD bits of the SVE vector registers are preserved,
548 * write any live register state back to task_struct, and convert to a
551 if (task == current) {
555 set_thread_flag(TIF_FOREIGN_FPSTATE);
558 fpsimd_flush_task_state(task);
559 if (test_and_clear_tsk_thread_flag(task, TIF_SVE))
566 * Force reallocation of task SVE state to the correct size
571 task->thread.sve_vl = vl;
574 update_tsk_thread_flag(task, TIF_SVE_VL_INHERIT,
575 flags & PR_SVE_VL_INHERIT);
581 * Encode the current vector length and flags for return.
582 * This is only required for prctl(): ptrace has separate fields
584 * flags are as for sve_set_vector_length().
586 static int sve_prctl_status(unsigned long flags)
590 if (flags & PR_SVE_SET_VL_ONEXEC)
591 ret = current->thread.sve_vl_onexec;
593 ret = current->thread.sve_vl;
595 if (test_thread_flag(TIF_SVE_VL_INHERIT))
596 ret |= PR_SVE_VL_INHERIT;
602 int sve_set_current_vl(unsigned long arg)
604 unsigned long vl, flags;
607 vl = arg & PR_SVE_VL_LEN_MASK;
610 if (!system_supports_sve())
613 ret = sve_set_vector_length(current, vl, flags);
617 return sve_prctl_status(flags);
621 int sve_get_current_vl(void)
623 if (!system_supports_sve())
626 return sve_prctl_status(0);
630 * Bitmap for temporary storage of the per-CPU set of supported vector lengths
631 * during secondary boot.
633 static DECLARE_BITMAP(sve_secondary_vq_map, SVE_VQ_MAX);
635 static void sve_probe_vqs(DECLARE_BITMAP(map, SVE_VQ_MAX))
640 bitmap_zero(map, SVE_VQ_MAX);
642 zcr = ZCR_ELx_LEN_MASK;
643 zcr = read_sysreg_s(SYS_ZCR_EL1) & ~zcr;
645 for (vq = SVE_VQ_MAX; vq >= SVE_VQ_MIN; --vq) {
646 write_sysreg_s(zcr | (vq - 1), SYS_ZCR_EL1); /* self-syncing */
648 vq = sve_vq_from_vl(vl); /* skip intervening lengths */
649 set_bit(vq_to_bit(vq), map);
653 void __init sve_init_vq_map(void)
655 sve_probe_vqs(sve_vq_map);
659 * If we haven't committed to the set of supported VQs yet, filter out
660 * those not supported by the current CPU.
662 void sve_update_vq_map(void)
664 sve_probe_vqs(sve_secondary_vq_map);
665 bitmap_and(sve_vq_map, sve_vq_map, sve_secondary_vq_map, SVE_VQ_MAX);
668 /* Check whether the current CPU supports all VQs in the committed set */
669 int sve_verify_vq_map(void)
673 sve_probe_vqs(sve_secondary_vq_map);
674 bitmap_andnot(sve_secondary_vq_map, sve_vq_map, sve_secondary_vq_map,
676 if (!bitmap_empty(sve_secondary_vq_map, SVE_VQ_MAX)) {
677 pr_warn("SVE: cpu%d: Required vector length(s) missing\n",
685 static void __init sve_efi_setup(void)
687 if (!IS_ENABLED(CONFIG_EFI))
691 * alloc_percpu() warns and prints a backtrace if this goes wrong.
692 * This is evidence of a crippled system and we are returning void,
693 * so no attempt is made to handle this situation here.
695 if (!sve_vl_valid(sve_max_vl))
698 efi_sve_state = __alloc_percpu(
699 SVE_SIG_REGS_SIZE(sve_vq_from_vl(sve_max_vl)), SVE_VQ_BYTES);
706 panic("Cannot allocate percpu memory for EFI SVE save/restore");
710 * Enable SVE for EL1.
711 * Intended for use by the cpufeatures code during CPU boot.
713 void sve_kernel_enable(const struct arm64_cpu_capabilities *__always_unused p)
715 write_sysreg(read_sysreg(CPACR_EL1) | CPACR_EL1_ZEN_EL1EN, CPACR_EL1);
720 * Read the pseudo-ZCR used by cpufeatures to identify the supported SVE
723 * Use only if SVE is present.
724 * This function clobbers the SVE vector length.
726 u64 read_zcr_features(void)
732 * Set the maximum possible VL, and write zeroes to all other
733 * bits to see if they stick.
735 sve_kernel_enable(NULL);
736 write_sysreg_s(ZCR_ELx_LEN_MASK, SYS_ZCR_EL1);
738 zcr = read_sysreg_s(SYS_ZCR_EL1);
739 zcr &= ~(u64)ZCR_ELx_LEN_MASK; /* find sticky 1s outside LEN field */
740 vq_max = sve_vq_from_vl(sve_get_vl());
741 zcr |= vq_max - 1; /* set LEN field to maximum effective value */
746 void __init sve_setup(void)
750 if (!system_supports_sve())
754 * The SVE architecture mandates support for 128-bit vectors,
755 * so sve_vq_map must have at least SVE_VQ_MIN set.
756 * If something went wrong, at least try to patch it up:
758 if (WARN_ON(!test_bit(vq_to_bit(SVE_VQ_MIN), sve_vq_map)))
759 set_bit(vq_to_bit(SVE_VQ_MIN), sve_vq_map);
761 zcr = read_sanitised_ftr_reg(SYS_ZCR_EL1);
762 sve_max_vl = sve_vl_from_vq((zcr & ZCR_ELx_LEN_MASK) + 1);
765 * Sanity-check that the max VL we determined through CPU features
766 * corresponds properly to sve_vq_map. If not, do our best:
768 if (WARN_ON(sve_max_vl != find_supported_vector_length(sve_max_vl)))
769 sve_max_vl = find_supported_vector_length(sve_max_vl);
772 * For the default VL, pick the maximum supported value <= 64.
773 * VL == 64 is guaranteed not to grow the signal frame.
775 sve_default_vl = find_supported_vector_length(64);
777 pr_info("SVE: maximum available vector length %u bytes per vector\n",
779 pr_info("SVE: default vector length %u bytes per vector\n",
786 * Called from the put_task_struct() path, which cannot get here
787 * unless dead_task is really dead and not schedulable.
789 void fpsimd_release_task(struct task_struct *dead_task)
791 __sve_free(dead_task);
794 #endif /* CONFIG_ARM64_SVE */
799 * Storage is allocated for the full SVE state, the current FPSIMD
800 * register contents are migrated across, and TIF_SVE is set so that
801 * the SVE access trap will be disabled the next time this task
802 * reaches ret_to_user.
804 * TIF_SVE should be clear on entry: otherwise, task_fpsimd_load()
805 * would have disabled the SVE access trap for userspace during
806 * ret_to_user, making an SVE access trap impossible in that case.
808 asmlinkage void do_sve_acc(unsigned int esr, struct pt_regs *regs)
810 /* Even if we chose not to use SVE, the hardware could still trap: */
811 if (unlikely(!system_supports_sve()) || WARN_ON(is_compat_task())) {
812 force_signal_inject(SIGILL, ILL_ILLOPC, regs->pc);
821 fpsimd_to_sve(current);
823 /* Force ret_to_user to reload the registers: */
824 fpsimd_flush_task_state(current);
825 set_thread_flag(TIF_FOREIGN_FPSTATE);
827 if (test_and_set_thread_flag(TIF_SVE))
828 WARN_ON(1); /* SVE access shouldn't have trapped */
834 * Trapped FP/ASIMD access.
836 asmlinkage void do_fpsimd_acc(unsigned int esr, struct pt_regs *regs)
838 /* TODO: implement lazy context saving/restoring */
843 * Raise a SIGFPE for the current process.
845 asmlinkage void do_fpsimd_exc(unsigned int esr, struct pt_regs *regs)
848 unsigned int si_code = FPE_FLTUNK;
850 if (esr & ESR_ELx_FP_EXC_TFV) {
852 si_code = FPE_FLTINV;
853 else if (esr & FPEXC_DZF)
854 si_code = FPE_FLTDIV;
855 else if (esr & FPEXC_OFF)
856 si_code = FPE_FLTOVF;
857 else if (esr & FPEXC_UFF)
858 si_code = FPE_FLTUND;
859 else if (esr & FPEXC_IXF)
860 si_code = FPE_FLTRES;
863 clear_siginfo(&info);
864 info.si_signo = SIGFPE;
865 info.si_code = si_code;
866 info.si_addr = (void __user *)instruction_pointer(regs);
868 send_sig_info(SIGFPE, &info, current);
871 void fpsimd_thread_switch(struct task_struct *next)
873 bool wrong_task, wrong_cpu;
875 if (!system_supports_fpsimd())
878 /* Save unsaved fpsimd state, if any: */
882 * Fix up TIF_FOREIGN_FPSTATE to correctly describe next's
883 * state. For kernel threads, FPSIMD registers are never loaded
884 * and wrong_task and wrong_cpu will always be true.
886 wrong_task = __this_cpu_read(fpsimd_last_state.st) !=
887 &next->thread.uw.fpsimd_state;
888 wrong_cpu = next->thread.fpsimd_cpu != smp_processor_id();
890 update_tsk_thread_flag(next, TIF_FOREIGN_FPSTATE,
891 wrong_task || wrong_cpu);
894 void fpsimd_flush_thread(void)
896 int vl, supported_vl;
898 if (!system_supports_fpsimd())
903 memset(¤t->thread.uw.fpsimd_state, 0,
904 sizeof(current->thread.uw.fpsimd_state));
905 fpsimd_flush_task_state(current);
907 if (system_supports_sve()) {
908 clear_thread_flag(TIF_SVE);
912 * Reset the task vector length as required.
913 * This is where we ensure that all user tasks have a valid
914 * vector length configured: no kernel task can become a user
915 * task without an exec and hence a call to this function.
916 * By the time the first call to this function is made, all
917 * early hardware probing is complete, so sve_default_vl
919 * If a bug causes this to go wrong, we make some noise and
920 * try to fudge thread.sve_vl to a safe value here.
922 vl = current->thread.sve_vl_onexec ?
923 current->thread.sve_vl_onexec : sve_default_vl;
925 if (WARN_ON(!sve_vl_valid(vl)))
928 supported_vl = find_supported_vector_length(vl);
929 if (WARN_ON(supported_vl != vl))
932 current->thread.sve_vl = vl;
935 * If the task is not set to inherit, ensure that the vector
936 * length will be reset by a subsequent exec:
938 if (!test_thread_flag(TIF_SVE_VL_INHERIT))
939 current->thread.sve_vl_onexec = 0;
942 set_thread_flag(TIF_FOREIGN_FPSTATE);
948 * Save the userland FPSIMD state of 'current' to memory, but only if the state
949 * currently held in the registers does in fact belong to 'current'
951 void fpsimd_preserve_current_state(void)
953 if (!system_supports_fpsimd())
962 * Like fpsimd_preserve_current_state(), but ensure that
963 * current->thread.uw.fpsimd_state is updated so that it can be copied to
966 void fpsimd_signal_preserve_current_state(void)
968 fpsimd_preserve_current_state();
969 if (system_supports_sve() && test_thread_flag(TIF_SVE))
970 sve_to_fpsimd(current);
974 * Associate current's FPSIMD context with this cpu
975 * Preemption must be disabled when calling this function.
977 void fpsimd_bind_task_to_cpu(void)
979 struct fpsimd_last_state_struct *last =
980 this_cpu_ptr(&fpsimd_last_state);
982 WARN_ON(!system_supports_fpsimd());
983 last->st = ¤t->thread.uw.fpsimd_state;
984 current->thread.fpsimd_cpu = smp_processor_id();
986 if (system_supports_sve()) {
987 /* Toggle SVE trapping for userspace if needed */
988 if (test_thread_flag(TIF_SVE))
993 /* Serialised by exception return to user */
997 void fpsimd_bind_state_to_cpu(struct user_fpsimd_state *st)
999 struct fpsimd_last_state_struct *last =
1000 this_cpu_ptr(&fpsimd_last_state);
1002 WARN_ON(!system_supports_fpsimd());
1003 WARN_ON(!in_softirq() && !irqs_disabled());
1009 * Load the userland FPSIMD state of 'current' from memory, but only if the
1010 * FPSIMD state already held in the registers is /not/ the most recent FPSIMD
1011 * state of 'current'
1013 void fpsimd_restore_current_state(void)
1016 * For the tasks that were created before we detected the absence of
1017 * FP/SIMD, the TIF_FOREIGN_FPSTATE could be set via fpsimd_thread_switch(),
1018 * e.g, init. This could be then inherited by the children processes.
1019 * If we later detect that the system doesn't support FP/SIMD,
1020 * we must clear the flag for all the tasks to indicate that the
1021 * FPSTATE is clean (as we can't have one) to avoid looping for ever in
1022 * do_notify_resume().
1024 if (!system_supports_fpsimd()) {
1025 clear_thread_flag(TIF_FOREIGN_FPSTATE);
1031 if (test_and_clear_thread_flag(TIF_FOREIGN_FPSTATE)) {
1033 fpsimd_bind_task_to_cpu();
1040 * Load an updated userland FPSIMD state for 'current' from memory and set the
1041 * flag that indicates that the FPSIMD register contents are the most recent
1042 * FPSIMD state of 'current'
1044 void fpsimd_update_current_state(struct user_fpsimd_state const *state)
1046 if (WARN_ON(!system_supports_fpsimd()))
1051 current->thread.uw.fpsimd_state = *state;
1052 if (system_supports_sve() && test_thread_flag(TIF_SVE))
1053 fpsimd_to_sve(current);
1056 fpsimd_bind_task_to_cpu();
1058 clear_thread_flag(TIF_FOREIGN_FPSTATE);
1064 * Invalidate live CPU copies of task t's FPSIMD state
1066 void fpsimd_flush_task_state(struct task_struct *t)
1068 t->thread.fpsimd_cpu = NR_CPUS;
1071 void fpsimd_flush_cpu_state(void)
1073 WARN_ON(!system_supports_fpsimd());
1074 __this_cpu_write(fpsimd_last_state.st, NULL);
1075 set_thread_flag(TIF_FOREIGN_FPSTATE);
1078 #ifdef CONFIG_KERNEL_MODE_NEON
1080 DEFINE_PER_CPU(bool, kernel_neon_busy);
1081 EXPORT_PER_CPU_SYMBOL(kernel_neon_busy);
1084 * Kernel-side NEON support functions
1088 * kernel_neon_begin(): obtain the CPU FPSIMD registers for use by the calling
1091 * Must not be called unless may_use_simd() returns true.
1092 * Task context in the FPSIMD registers is saved back to memory as necessary.
1094 * A matching call to kernel_neon_end() must be made before returning from the
1097 * The caller may freely use the FPSIMD registers until kernel_neon_end() is
1100 void kernel_neon_begin(void)
1102 if (WARN_ON(!system_supports_fpsimd()))
1105 BUG_ON(!may_use_simd());
1109 __this_cpu_write(kernel_neon_busy, true);
1111 /* Save unsaved fpsimd state, if any: */
1114 /* Invalidate any task state remaining in the fpsimd regs: */
1115 fpsimd_flush_cpu_state();
1121 EXPORT_SYMBOL(kernel_neon_begin);
1124 * kernel_neon_end(): give the CPU FPSIMD registers back to the current task
1126 * Must be called from a context in which kernel_neon_begin() was previously
1127 * called, with no call to kernel_neon_end() in the meantime.
1129 * The caller must not use the FPSIMD registers after this function is called,
1130 * unless kernel_neon_begin() is called again in the meantime.
1132 void kernel_neon_end(void)
1136 if (!system_supports_fpsimd())
1139 busy = __this_cpu_xchg(kernel_neon_busy, false);
1140 WARN_ON(!busy); /* No matching kernel_neon_begin()? */
1144 EXPORT_SYMBOL(kernel_neon_end);
1148 static DEFINE_PER_CPU(struct user_fpsimd_state, efi_fpsimd_state);
1149 static DEFINE_PER_CPU(bool, efi_fpsimd_state_used);
1150 static DEFINE_PER_CPU(bool, efi_sve_state_used);
1153 * EFI runtime services support functions
1155 * The ABI for EFI runtime services allows EFI to use FPSIMD during the call.
1156 * This means that for EFI (and only for EFI), we have to assume that FPSIMD
1157 * is always used rather than being an optional accelerator.
1159 * These functions provide the necessary support for ensuring FPSIMD
1160 * save/restore in the contexts from which EFI is used.
1162 * Do not use them for any other purpose -- if tempted to do so, you are
1163 * either doing something wrong or you need to propose some refactoring.
1167 * __efi_fpsimd_begin(): prepare FPSIMD for making an EFI runtime services call
1169 void __efi_fpsimd_begin(void)
1171 if (!system_supports_fpsimd())
1174 WARN_ON(preemptible());
1176 if (may_use_simd()) {
1177 kernel_neon_begin();
1180 * If !efi_sve_state, SVE can't be in use yet and doesn't need
1183 if (system_supports_sve() && likely(efi_sve_state)) {
1184 char *sve_state = this_cpu_ptr(efi_sve_state);
1186 __this_cpu_write(efi_sve_state_used, true);
1188 sve_save_state(sve_state + sve_ffr_offset(sve_max_vl),
1189 &this_cpu_ptr(&efi_fpsimd_state)->fpsr);
1191 fpsimd_save_state(this_cpu_ptr(&efi_fpsimd_state));
1194 __this_cpu_write(efi_fpsimd_state_used, true);
1199 * __efi_fpsimd_end(): clean up FPSIMD after an EFI runtime services call
1201 void __efi_fpsimd_end(void)
1203 if (!system_supports_fpsimd())
1206 if (!__this_cpu_xchg(efi_fpsimd_state_used, false)) {
1209 if (system_supports_sve() &&
1210 likely(__this_cpu_read(efi_sve_state_used))) {
1211 char const *sve_state = this_cpu_ptr(efi_sve_state);
1213 sve_load_state(sve_state + sve_ffr_offset(sve_max_vl),
1214 &this_cpu_ptr(&efi_fpsimd_state)->fpsr,
1215 sve_vq_from_vl(sve_get_vl()) - 1);
1217 __this_cpu_write(efi_sve_state_used, false);
1219 fpsimd_load_state(this_cpu_ptr(&efi_fpsimd_state));
1224 #endif /* CONFIG_EFI */
1226 #endif /* CONFIG_KERNEL_MODE_NEON */
1228 #ifdef CONFIG_CPU_PM
1229 static int fpsimd_cpu_pm_notifier(struct notifier_block *self,
1230 unsigned long cmd, void *v)
1235 fpsimd_flush_cpu_state();
1239 case CPU_PM_ENTER_FAILED:
1246 static struct notifier_block fpsimd_cpu_pm_notifier_block = {
1247 .notifier_call = fpsimd_cpu_pm_notifier,
1250 static void __init fpsimd_pm_init(void)
1252 cpu_pm_register_notifier(&fpsimd_cpu_pm_notifier_block);
1256 static inline void fpsimd_pm_init(void) { }
1257 #endif /* CONFIG_CPU_PM */
1259 #ifdef CONFIG_HOTPLUG_CPU
1260 static int fpsimd_cpu_dead(unsigned int cpu)
1262 per_cpu(fpsimd_last_state.st, cpu) = NULL;
1266 static inline void fpsimd_hotplug_init(void)
1268 cpuhp_setup_state_nocalls(CPUHP_ARM64_FPSIMD_DEAD, "arm64/fpsimd:dead",
1269 NULL, fpsimd_cpu_dead);
1273 static inline void fpsimd_hotplug_init(void) { }
1277 * FP/SIMD support code initialisation.
1279 static int __init fpsimd_init(void)
1281 if (elf_hwcap & HWCAP_FP) {
1283 fpsimd_hotplug_init();
1285 pr_notice("Floating-point is not implemented\n");
1288 if (!(elf_hwcap & HWCAP_ASIMD))
1289 pr_notice("Advanced SIMD is not implemented\n");
1291 return sve_sysctl_init();
1293 core_initcall(fpsimd_init);