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/prctl.h>
35 #include <linux/ptrace.h>
36 #include <linux/sched/signal.h>
37 #include <linux/sched/task_stack.h>
38 #include <linux/signal.h>
39 #include <linux/slab.h>
40 #include <linux/sysctl.h>
43 #include <asm/fpsimd.h>
44 #include <asm/cpufeature.h>
45 #include <asm/cputype.h>
47 #include <asm/sigcontext.h>
48 #include <asm/sysreg.h>
49 #include <asm/traps.h>
51 #define FPEXC_IOF (1 << 0)
52 #define FPEXC_DZF (1 << 1)
53 #define FPEXC_OFF (1 << 2)
54 #define FPEXC_UFF (1 << 3)
55 #define FPEXC_IXF (1 << 4)
56 #define FPEXC_IDF (1 << 7)
59 * (Note: in this discussion, statements about FPSIMD apply equally to SVE.)
61 * In order to reduce the number of times the FPSIMD state is needlessly saved
62 * and restored, we need to keep track of two things:
63 * (a) for each task, we need to remember which CPU was the last one to have
64 * the task's FPSIMD state loaded into its FPSIMD registers;
65 * (b) for each CPU, we need to remember which task's userland FPSIMD state has
66 * been loaded into its FPSIMD registers most recently, or whether it has
67 * been used to perform kernel mode NEON in the meantime.
69 * For (a), we add a fpsimd_cpu field to thread_struct, which gets updated to
70 * the id of the current CPU every time the state is loaded onto a CPU. For (b),
71 * we add the per-cpu variable 'fpsimd_last_state' (below), which contains the
72 * address of the userland FPSIMD state of the task that was loaded onto the CPU
73 * the most recently, or NULL if kernel mode NEON has been performed after that.
75 * With this in place, we no longer have to restore the next FPSIMD state right
76 * when switching between tasks. Instead, we can defer this check to userland
77 * resume, at which time we verify whether the CPU's fpsimd_last_state and the
78 * task's fpsimd_cpu are still mutually in sync. If this is the case, we
79 * can omit the FPSIMD restore.
81 * As an optimization, we use the thread_info flag TIF_FOREIGN_FPSTATE to
82 * indicate whether or not the userland FPSIMD state of the current task is
83 * present in the registers. The flag is set unless the FPSIMD registers of this
84 * CPU currently contain the most recent userland FPSIMD state of the current
87 * In order to allow softirq handlers to use FPSIMD, kernel_neon_begin() may
88 * save the task's FPSIMD context back to task_struct from softirq context.
89 * To prevent this from racing with the manipulation of the task's FPSIMD state
90 * from task context and thereby corrupting the state, it is necessary to
91 * protect any manipulation of a task's fpsimd_state or TIF_FOREIGN_FPSTATE
92 * flag with local_bh_disable() unless softirqs are already masked.
94 * For a certain task, the sequence may look something like this:
95 * - the task gets scheduled in; if both the task's fpsimd_cpu field
96 * contains the id of the current CPU, and the CPU's fpsimd_last_state per-cpu
97 * variable points to the task's fpsimd_state, the TIF_FOREIGN_FPSTATE flag is
98 * cleared, otherwise it is set;
100 * - the task returns to userland; if TIF_FOREIGN_FPSTATE is set, the task's
101 * userland FPSIMD state is copied from memory to the registers, the task's
102 * fpsimd_cpu field is set to the id of the current CPU, the current
103 * CPU's fpsimd_last_state pointer is set to this task's fpsimd_state and the
104 * TIF_FOREIGN_FPSTATE flag is cleared;
106 * - the task executes an ordinary syscall; upon return to userland, the
107 * TIF_FOREIGN_FPSTATE flag will still be cleared, so no FPSIMD state is
110 * - the task executes a syscall which executes some NEON instructions; this is
111 * preceded by a call to kernel_neon_begin(), which copies the task's FPSIMD
112 * register contents to memory, clears the fpsimd_last_state per-cpu variable
113 * and sets the TIF_FOREIGN_FPSTATE flag;
115 * - the task gets preempted after kernel_neon_end() is called; as we have not
116 * returned from the 2nd syscall yet, TIF_FOREIGN_FPSTATE is still set so
117 * whatever is in the FPSIMD registers is not saved to memory, but discarded.
119 struct fpsimd_last_state_struct {
120 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 = -1;
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 /* Offset of FFR in the SVE register dump */
164 static size_t sve_ffr_offset(int vl)
166 return SVE_SIG_FFR_OFFSET(sve_vq_from_vl(vl)) - SVE_SIG_REGS_OFFSET;
169 static void *sve_pffr(struct task_struct *task)
171 return (char *)task->thread.sve_state +
172 sve_ffr_offset(task->thread.sve_vl);
175 static void change_cpacr(u64 val, u64 mask)
177 u64 cpacr = read_sysreg(CPACR_EL1);
178 u64 new = (cpacr & ~mask) | val;
181 write_sysreg(new, CPACR_EL1);
184 static void sve_user_disable(void)
186 change_cpacr(0, CPACR_EL1_ZEN_EL0EN);
189 static void sve_user_enable(void)
191 change_cpacr(CPACR_EL1_ZEN_EL0EN, CPACR_EL1_ZEN_EL0EN);
195 * TIF_SVE controls whether a task can use SVE without trapping while
196 * in userspace, and also the way a task's FPSIMD/SVE state is stored
199 * The kernel uses this flag to track whether a user task is actively
200 * using SVE, and therefore whether full SVE register state needs to
201 * be tracked. If not, the cheaper FPSIMD context handling code can
202 * be used instead of the more costly SVE equivalents.
206 * The task can execute SVE instructions while in userspace without
207 * trapping to the kernel.
209 * When stored, Z0-Z31 (incorporating Vn in bits[127:0] or the
210 * corresponding Zn), P0-P15 and FFR are encoded in in
211 * task->thread.sve_state, formatted appropriately for vector
212 * length task->thread.sve_vl.
214 * task->thread.sve_state must point to a valid buffer at least
215 * sve_state_size(task) bytes in size.
217 * During any syscall, the kernel may optionally clear TIF_SVE and
218 * discard the vector state except for the FPSIMD subset.
222 * An attempt by the user task to execute an SVE instruction causes
223 * do_sve_acc() to be called, which does some preparation and then
226 * When stored, FPSIMD registers V0-V31 are encoded in
227 * task->thread.uw.fpsimd_state; bits [max : 128] for each of Z0-Z31 are
228 * logically zero but not stored anywhere; P0-P15 and FFR are not
229 * stored and have unspecified values from userspace's point of
230 * view. For hygiene purposes, the kernel zeroes them on next use,
231 * but userspace is discouraged from relying on this.
233 * task->thread.sve_state does not need to be non-NULL, valid or any
234 * particular size: it must not be dereferenced.
236 * * FPSR and FPCR are always stored in task->thread.uw.fpsimd_state
237 * irrespective of whether TIF_SVE is clear or set, since these are
238 * not vector length dependent.
242 * Update current's FPSIMD/SVE registers from thread_struct.
244 * This function should be called only when the FPSIMD/SVE state in
245 * thread_struct is known to be up to date, when preparing to enter
248 * Softirqs (and preemption) must be disabled.
250 static void task_fpsimd_load(void)
252 WARN_ON(!in_softirq() && !irqs_disabled());
254 if (system_supports_sve() && test_thread_flag(TIF_SVE))
255 sve_load_state(sve_pffr(current),
256 ¤t->thread.uw.fpsimd_state.fpsr,
257 sve_vq_from_vl(current->thread.sve_vl) - 1);
259 fpsimd_load_state(¤t->thread.uw.fpsimd_state);
261 if (system_supports_sve()) {
262 /* Toggle SVE trapping for userspace if needed */
263 if (test_thread_flag(TIF_SVE))
268 /* Serialised by exception return to user */
273 * Ensure current's FPSIMD/SVE storage in thread_struct is up to date
274 * with respect to the CPU registers.
276 * Softirqs (and preemption) must be disabled.
278 static void task_fpsimd_save(void)
280 WARN_ON(!in_softirq() && !irqs_disabled());
282 if (!test_thread_flag(TIF_FOREIGN_FPSTATE)) {
283 if (system_supports_sve() && test_thread_flag(TIF_SVE)) {
284 if (WARN_ON(sve_get_vl() != current->thread.sve_vl)) {
286 * Can't save the user regs, so current would
287 * re-enter user with corrupt state.
288 * There's no way to recover, so kill it:
290 force_signal_inject(SIGKILL, SI_KERNEL, 0);
294 sve_save_state(sve_pffr(current),
295 ¤t->thread.uw.fpsimd_state.fpsr);
297 fpsimd_save_state(¤t->thread.uw.fpsimd_state);
302 * Helpers to translate bit indices in sve_vq_map to VQ values (and
303 * vice versa). This allows find_next_bit() to be used to find the
304 * _maximum_ VQ not exceeding a certain value.
307 static unsigned int vq_to_bit(unsigned int vq)
309 return SVE_VQ_MAX - vq;
312 static unsigned int bit_to_vq(unsigned int bit)
314 if (WARN_ON(bit >= SVE_VQ_MAX))
315 bit = SVE_VQ_MAX - 1;
317 return SVE_VQ_MAX - bit;
321 * All vector length selection from userspace comes through here.
322 * We're on a slow path, so some sanity-checks are included.
323 * If things go wrong there's a bug somewhere, but try to fall back to a
326 static unsigned int find_supported_vector_length(unsigned int vl)
329 int max_vl = sve_max_vl;
331 if (WARN_ON(!sve_vl_valid(vl)))
334 if (WARN_ON(!sve_vl_valid(max_vl)))
340 bit = find_next_bit(sve_vq_map, SVE_VQ_MAX,
341 vq_to_bit(sve_vq_from_vl(vl)));
342 return sve_vl_from_vq(bit_to_vq(bit));
347 static int sve_proc_do_default_vl(struct ctl_table *table, int write,
348 void __user *buffer, size_t *lenp,
352 int vl = sve_default_vl;
353 struct ctl_table tmp_table = {
355 .maxlen = sizeof(vl),
358 ret = proc_dointvec(&tmp_table, write, buffer, lenp, ppos);
362 /* Writing -1 has the special meaning "set to max": */
364 /* Fail safe if sve_max_vl wasn't initialised */
365 if (WARN_ON(!sve_vl_valid(sve_max_vl)))
373 if (!sve_vl_valid(vl))
376 vl = find_supported_vector_length(vl);
382 static struct ctl_table sve_default_vl_table[] = {
384 .procname = "sve_default_vector_length",
386 .proc_handler = sve_proc_do_default_vl,
391 static int __init sve_sysctl_init(void)
393 if (system_supports_sve())
394 if (!register_sysctl("abi", sve_default_vl_table))
400 #else /* ! CONFIG_SYSCTL */
401 static int __init sve_sysctl_init(void) { return 0; }
402 #endif /* ! CONFIG_SYSCTL */
404 #define ZREG(sve_state, vq, n) ((char *)(sve_state) + \
405 (SVE_SIG_ZREG_OFFSET(vq, n) - SVE_SIG_REGS_OFFSET))
408 * Transfer the FPSIMD state in task->thread.uw.fpsimd_state to
409 * task->thread.sve_state.
411 * Task can be a non-runnable task, or current. In the latter case,
412 * softirqs (and preemption) must be disabled.
413 * task->thread.sve_state must point to at least sve_state_size(task)
414 * bytes of allocated kernel memory.
415 * task->thread.uw.fpsimd_state must be up to date before calling this
418 static void fpsimd_to_sve(struct task_struct *task)
421 void *sst = task->thread.sve_state;
422 struct user_fpsimd_state const *fst = &task->thread.uw.fpsimd_state;
425 if (!system_supports_sve())
428 vq = sve_vq_from_vl(task->thread.sve_vl);
429 for (i = 0; i < 32; ++i)
430 memcpy(ZREG(sst, vq, i), &fst->vregs[i],
431 sizeof(fst->vregs[i]));
435 * Transfer the SVE state in task->thread.sve_state to
436 * task->thread.uw.fpsimd_state.
438 * Task can be a non-runnable task, or current. In the latter case,
439 * softirqs (and preemption) must be disabled.
440 * task->thread.sve_state must point to at least sve_state_size(task)
441 * bytes of allocated kernel memory.
442 * task->thread.sve_state must be up to date before calling this function.
444 static void sve_to_fpsimd(struct task_struct *task)
447 void const *sst = task->thread.sve_state;
448 struct user_fpsimd_state *fst = &task->thread.uw.fpsimd_state;
451 if (!system_supports_sve())
454 vq = sve_vq_from_vl(task->thread.sve_vl);
455 for (i = 0; i < 32; ++i)
456 memcpy(&fst->vregs[i], ZREG(sst, vq, i),
457 sizeof(fst->vregs[i]));
460 #ifdef CONFIG_ARM64_SVE
463 * Return how many bytes of memory are required to store the full SVE
464 * state for task, given task's currently configured vector length.
466 size_t sve_state_size(struct task_struct const *task)
468 return SVE_SIG_REGS_SIZE(sve_vq_from_vl(task->thread.sve_vl));
472 * Ensure that task->thread.sve_state is allocated and sufficiently large.
474 * This function should be used only in preparation for replacing
475 * task->thread.sve_state with new data. The memory is always zeroed
476 * here to prevent stale data from showing through: this is done in
477 * the interest of testability and predictability: except in the
478 * do_sve_acc() case, there is no ABI requirement to hide stale data
479 * written previously be task.
481 void sve_alloc(struct task_struct *task)
483 if (task->thread.sve_state) {
484 memset(task->thread.sve_state, 0, sve_state_size(current));
488 /* This is a small allocation (maximum ~8KB) and Should Not Fail. */
489 task->thread.sve_state =
490 kzalloc(sve_state_size(task), GFP_KERNEL);
493 * If future SVE revisions can have larger vectors though,
494 * this may cease to be true:
496 BUG_ON(!task->thread.sve_state);
501 * Ensure that task->thread.sve_state is up to date with respect to
502 * the user task, irrespective of when SVE is in use or not.
504 * This should only be called by ptrace. task must be non-runnable.
505 * task->thread.sve_state must point to at least sve_state_size(task)
506 * bytes of allocated kernel memory.
508 void fpsimd_sync_to_sve(struct task_struct *task)
510 if (!test_tsk_thread_flag(task, TIF_SVE))
515 * Ensure that task->thread.uw.fpsimd_state is up to date with respect to
516 * the user task, irrespective of whether SVE is in use or not.
518 * This should only be called by ptrace. task must be non-runnable.
519 * task->thread.sve_state must point to at least sve_state_size(task)
520 * bytes of allocated kernel memory.
522 void sve_sync_to_fpsimd(struct task_struct *task)
524 if (test_tsk_thread_flag(task, TIF_SVE))
529 * Ensure that task->thread.sve_state is up to date with respect to
530 * the task->thread.uw.fpsimd_state.
532 * This should only be called by ptrace to merge new FPSIMD register
533 * values into a task for which SVE is currently active.
534 * 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.
537 * task->thread.uw.fpsimd_state must already have been initialised with
538 * the new FPSIMD register values to be merged in.
540 void sve_sync_from_fpsimd_zeropad(struct task_struct *task)
543 void *sst = task->thread.sve_state;
544 struct user_fpsimd_state const *fst = &task->thread.uw.fpsimd_state;
547 if (!test_tsk_thread_flag(task, TIF_SVE))
550 vq = sve_vq_from_vl(task->thread.sve_vl);
552 memset(sst, 0, SVE_SIG_REGS_SIZE(vq));
554 for (i = 0; i < 32; ++i)
555 memcpy(ZREG(sst, vq, i), &fst->vregs[i],
556 sizeof(fst->vregs[i]));
559 int sve_set_vector_length(struct task_struct *task,
560 unsigned long vl, unsigned long flags)
562 if (flags & ~(unsigned long)(PR_SVE_VL_INHERIT |
563 PR_SVE_SET_VL_ONEXEC))
566 if (!sve_vl_valid(vl))
570 * Clamp to the maximum vector length that VL-agnostic SVE code can
571 * work with. A flag may be assigned in the future to allow setting
572 * of larger vector lengths without confusing older software.
574 if (vl > SVE_VL_ARCH_MAX)
575 vl = SVE_VL_ARCH_MAX;
577 vl = find_supported_vector_length(vl);
579 if (flags & (PR_SVE_VL_INHERIT |
580 PR_SVE_SET_VL_ONEXEC))
581 task->thread.sve_vl_onexec = vl;
583 /* Reset VL to system default on next exec: */
584 task->thread.sve_vl_onexec = 0;
586 /* Only actually set the VL if not deferred: */
587 if (flags & PR_SVE_SET_VL_ONEXEC)
590 if (vl == task->thread.sve_vl)
594 * To ensure the FPSIMD bits of the SVE vector registers are preserved,
595 * write any live register state back to task_struct, and convert to a
598 if (task == current) {
602 set_thread_flag(TIF_FOREIGN_FPSTATE);
605 fpsimd_flush_task_state(task);
606 if (test_and_clear_tsk_thread_flag(task, TIF_SVE))
613 * Force reallocation of task SVE state to the correct size
618 task->thread.sve_vl = vl;
621 if (flags & PR_SVE_VL_INHERIT)
622 set_tsk_thread_flag(task, TIF_SVE_VL_INHERIT);
624 clear_tsk_thread_flag(task, TIF_SVE_VL_INHERIT);
630 * Encode the current vector length and flags for return.
631 * This is only required for prctl(): ptrace has separate fields
633 * flags are as for sve_set_vector_length().
635 static int sve_prctl_status(unsigned long flags)
639 if (flags & PR_SVE_SET_VL_ONEXEC)
640 ret = current->thread.sve_vl_onexec;
642 ret = current->thread.sve_vl;
644 if (test_thread_flag(TIF_SVE_VL_INHERIT))
645 ret |= PR_SVE_VL_INHERIT;
651 int sve_set_current_vl(unsigned long arg)
653 unsigned long vl, flags;
656 vl = arg & PR_SVE_VL_LEN_MASK;
659 if (!system_supports_sve())
662 ret = sve_set_vector_length(current, vl, flags);
666 return sve_prctl_status(flags);
670 int sve_get_current_vl(void)
672 if (!system_supports_sve())
675 return sve_prctl_status(0);
679 * Bitmap for temporary storage of the per-CPU set of supported vector lengths
680 * during secondary boot.
682 static DECLARE_BITMAP(sve_secondary_vq_map, SVE_VQ_MAX);
684 static void sve_probe_vqs(DECLARE_BITMAP(map, SVE_VQ_MAX))
689 bitmap_zero(map, SVE_VQ_MAX);
691 zcr = ZCR_ELx_LEN_MASK;
692 zcr = read_sysreg_s(SYS_ZCR_EL1) & ~zcr;
694 for (vq = SVE_VQ_MAX; vq >= SVE_VQ_MIN; --vq) {
695 write_sysreg_s(zcr | (vq - 1), SYS_ZCR_EL1); /* self-syncing */
697 vq = sve_vq_from_vl(vl); /* skip intervening lengths */
698 set_bit(vq_to_bit(vq), map);
702 void __init sve_init_vq_map(void)
704 sve_probe_vqs(sve_vq_map);
708 * If we haven't committed to the set of supported VQs yet, filter out
709 * those not supported by the current CPU.
711 void sve_update_vq_map(void)
713 sve_probe_vqs(sve_secondary_vq_map);
714 bitmap_and(sve_vq_map, sve_vq_map, sve_secondary_vq_map, SVE_VQ_MAX);
717 /* Check whether the current CPU supports all VQs in the committed set */
718 int sve_verify_vq_map(void)
722 sve_probe_vqs(sve_secondary_vq_map);
723 bitmap_andnot(sve_secondary_vq_map, sve_vq_map, sve_secondary_vq_map,
725 if (!bitmap_empty(sve_secondary_vq_map, SVE_VQ_MAX)) {
726 pr_warn("SVE: cpu%d: Required vector length(s) missing\n",
734 static void __init sve_efi_setup(void)
736 if (!IS_ENABLED(CONFIG_EFI))
740 * alloc_percpu() warns and prints a backtrace if this goes wrong.
741 * This is evidence of a crippled system and we are returning void,
742 * so no attempt is made to handle this situation here.
744 if (!sve_vl_valid(sve_max_vl))
747 efi_sve_state = __alloc_percpu(
748 SVE_SIG_REGS_SIZE(sve_vq_from_vl(sve_max_vl)), SVE_VQ_BYTES);
755 panic("Cannot allocate percpu memory for EFI SVE save/restore");
759 * Enable SVE for EL1.
760 * Intended for use by the cpufeatures code during CPU boot.
762 void sve_kernel_enable(const struct arm64_cpu_capabilities *__always_unused p)
764 write_sysreg(read_sysreg(CPACR_EL1) | CPACR_EL1_ZEN_EL1EN, CPACR_EL1);
768 void __init sve_setup(void)
772 if (!system_supports_sve())
776 * The SVE architecture mandates support for 128-bit vectors,
777 * so sve_vq_map must have at least SVE_VQ_MIN set.
778 * If something went wrong, at least try to patch it up:
780 if (WARN_ON(!test_bit(vq_to_bit(SVE_VQ_MIN), sve_vq_map)))
781 set_bit(vq_to_bit(SVE_VQ_MIN), sve_vq_map);
783 zcr = read_sanitised_ftr_reg(SYS_ZCR_EL1);
784 sve_max_vl = sve_vl_from_vq((zcr & ZCR_ELx_LEN_MASK) + 1);
787 * Sanity-check that the max VL we determined through CPU features
788 * corresponds properly to sve_vq_map. If not, do our best:
790 if (WARN_ON(sve_max_vl != find_supported_vector_length(sve_max_vl)))
791 sve_max_vl = find_supported_vector_length(sve_max_vl);
794 * For the default VL, pick the maximum supported value <= 64.
795 * VL == 64 is guaranteed not to grow the signal frame.
797 sve_default_vl = find_supported_vector_length(64);
799 pr_info("SVE: maximum available vector length %u bytes per vector\n",
801 pr_info("SVE: default vector length %u bytes per vector\n",
808 * Called from the put_task_struct() path, which cannot get here
809 * unless dead_task is really dead and not schedulable.
811 void fpsimd_release_task(struct task_struct *dead_task)
813 __sve_free(dead_task);
816 #endif /* CONFIG_ARM64_SVE */
821 * Storage is allocated for the full SVE state, the current FPSIMD
822 * register contents are migrated across, and TIF_SVE is set so that
823 * the SVE access trap will be disabled the next time this task
824 * reaches ret_to_user.
826 * TIF_SVE should be clear on entry: otherwise, task_fpsimd_load()
827 * would have disabled the SVE access trap for userspace during
828 * ret_to_user, making an SVE access trap impossible in that case.
830 asmlinkage void do_sve_acc(unsigned int esr, struct pt_regs *regs)
832 /* Even if we chose not to use SVE, the hardware could still trap: */
833 if (unlikely(!system_supports_sve()) || WARN_ON(is_compat_task())) {
834 force_signal_inject(SIGILL, ILL_ILLOPC, regs->pc);
843 fpsimd_to_sve(current);
845 /* Force ret_to_user to reload the registers: */
846 fpsimd_flush_task_state(current);
847 set_thread_flag(TIF_FOREIGN_FPSTATE);
849 if (test_and_set_thread_flag(TIF_SVE))
850 WARN_ON(1); /* SVE access shouldn't have trapped */
856 * Trapped FP/ASIMD access.
858 asmlinkage void do_fpsimd_acc(unsigned int esr, struct pt_regs *regs)
860 /* TODO: implement lazy context saving/restoring */
865 * Raise a SIGFPE for the current process.
867 asmlinkage void do_fpsimd_exc(unsigned int esr, struct pt_regs *regs)
870 unsigned int si_code = FPE_FLTUNK;
872 if (esr & ESR_ELx_FP_EXC_TFV) {
874 si_code = FPE_FLTINV;
875 else if (esr & FPEXC_DZF)
876 si_code = FPE_FLTDIV;
877 else if (esr & FPEXC_OFF)
878 si_code = FPE_FLTOVF;
879 else if (esr & FPEXC_UFF)
880 si_code = FPE_FLTUND;
881 else if (esr & FPEXC_IXF)
882 si_code = FPE_FLTRES;
885 clear_siginfo(&info);
886 info.si_signo = SIGFPE;
887 info.si_code = si_code;
888 info.si_addr = (void __user *)instruction_pointer(regs);
890 send_sig_info(SIGFPE, &info, current);
893 void fpsimd_thread_switch(struct task_struct *next)
895 if (!system_supports_fpsimd())
898 * Save the current FPSIMD state to memory, but only if whatever is in
899 * the registers is in fact the most recent userland FPSIMD state of
907 * If we are switching to a task whose most recent userland
908 * FPSIMD state is already in the registers of *this* cpu,
909 * we can skip loading the state from memory. Otherwise, set
910 * the TIF_FOREIGN_FPSTATE flag so the state will be loaded
911 * upon the next return to userland.
913 if (__this_cpu_read(fpsimd_last_state.st) ==
914 &next->thread.uw.fpsimd_state
915 && next->thread.fpsimd_cpu == smp_processor_id())
916 clear_tsk_thread_flag(next, TIF_FOREIGN_FPSTATE);
918 set_tsk_thread_flag(next, TIF_FOREIGN_FPSTATE);
922 void fpsimd_flush_thread(void)
924 int vl, supported_vl;
926 if (!system_supports_fpsimd())
931 memset(¤t->thread.uw.fpsimd_state, 0,
932 sizeof(current->thread.uw.fpsimd_state));
933 fpsimd_flush_task_state(current);
935 if (system_supports_sve()) {
936 clear_thread_flag(TIF_SVE);
940 * Reset the task vector length as required.
941 * This is where we ensure that all user tasks have a valid
942 * vector length configured: no kernel task can become a user
943 * task without an exec and hence a call to this function.
944 * By the time the first call to this function is made, all
945 * early hardware probing is complete, so sve_default_vl
947 * If a bug causes this to go wrong, we make some noise and
948 * try to fudge thread.sve_vl to a safe value here.
950 vl = current->thread.sve_vl_onexec ?
951 current->thread.sve_vl_onexec : sve_default_vl;
953 if (WARN_ON(!sve_vl_valid(vl)))
956 supported_vl = find_supported_vector_length(vl);
957 if (WARN_ON(supported_vl != vl))
960 current->thread.sve_vl = vl;
963 * If the task is not set to inherit, ensure that the vector
964 * length will be reset by a subsequent exec:
966 if (!test_thread_flag(TIF_SVE_VL_INHERIT))
967 current->thread.sve_vl_onexec = 0;
970 set_thread_flag(TIF_FOREIGN_FPSTATE);
976 * Save the userland FPSIMD state of 'current' to memory, but only if the state
977 * currently held in the registers does in fact belong to 'current'
979 void fpsimd_preserve_current_state(void)
981 if (!system_supports_fpsimd())
990 * Like fpsimd_preserve_current_state(), but ensure that
991 * current->thread.uw.fpsimd_state is updated so that it can be copied to
994 void fpsimd_signal_preserve_current_state(void)
996 fpsimd_preserve_current_state();
997 if (system_supports_sve() && test_thread_flag(TIF_SVE))
998 sve_to_fpsimd(current);
1002 * Associate current's FPSIMD context with this cpu
1003 * Preemption must be disabled when calling this function.
1005 static void fpsimd_bind_to_cpu(void)
1007 struct fpsimd_last_state_struct *last =
1008 this_cpu_ptr(&fpsimd_last_state);
1010 last->st = ¤t->thread.uw.fpsimd_state;
1011 last->sve_in_use = test_thread_flag(TIF_SVE);
1012 current->thread.fpsimd_cpu = smp_processor_id();
1016 * Load the userland FPSIMD state of 'current' from memory, but only if the
1017 * FPSIMD state already held in the registers is /not/ the most recent FPSIMD
1018 * state of 'current'
1020 void fpsimd_restore_current_state(void)
1022 if (!system_supports_fpsimd())
1027 if (test_and_clear_thread_flag(TIF_FOREIGN_FPSTATE)) {
1029 fpsimd_bind_to_cpu();
1036 * Load an updated userland FPSIMD state for 'current' from memory and set the
1037 * flag that indicates that the FPSIMD register contents are the most recent
1038 * FPSIMD state of 'current'
1040 void fpsimd_update_current_state(struct user_fpsimd_state const *state)
1042 if (!system_supports_fpsimd())
1047 current->thread.uw.fpsimd_state = *state;
1048 if (system_supports_sve() && test_thread_flag(TIF_SVE))
1049 fpsimd_to_sve(current);
1053 if (test_and_clear_thread_flag(TIF_FOREIGN_FPSTATE))
1054 fpsimd_bind_to_cpu();
1060 * Invalidate live CPU copies of task t's FPSIMD state
1062 void fpsimd_flush_task_state(struct task_struct *t)
1064 t->thread.fpsimd_cpu = NR_CPUS;
1067 static inline void fpsimd_flush_cpu_state(void)
1069 __this_cpu_write(fpsimd_last_state.st, NULL);
1073 * Invalidate any task SVE state currently held in this CPU's regs.
1075 * This is used to prevent the kernel from trying to reuse SVE register data
1076 * that is detroyed by KVM guest enter/exit. This function should go away when
1077 * KVM SVE support is implemented. Don't use it for anything else.
1079 #ifdef CONFIG_ARM64_SVE
1080 void sve_flush_cpu_state(void)
1082 struct fpsimd_last_state_struct const *last =
1083 this_cpu_ptr(&fpsimd_last_state);
1085 if (last->st && last->sve_in_use)
1086 fpsimd_flush_cpu_state();
1088 #endif /* CONFIG_ARM64_SVE */
1090 #ifdef CONFIG_KERNEL_MODE_NEON
1092 DEFINE_PER_CPU(bool, kernel_neon_busy);
1093 EXPORT_PER_CPU_SYMBOL(kernel_neon_busy);
1096 * Kernel-side NEON support functions
1100 * kernel_neon_begin(): obtain the CPU FPSIMD registers for use by the calling
1103 * Must not be called unless may_use_simd() returns true.
1104 * Task context in the FPSIMD registers is saved back to memory as necessary.
1106 * A matching call to kernel_neon_end() must be made before returning from the
1109 * The caller may freely use the FPSIMD registers until kernel_neon_end() is
1112 void kernel_neon_begin(void)
1114 if (WARN_ON(!system_supports_fpsimd()))
1117 BUG_ON(!may_use_simd());
1121 __this_cpu_write(kernel_neon_busy, true);
1123 /* Save unsaved task fpsimd state, if any: */
1126 set_thread_flag(TIF_FOREIGN_FPSTATE);
1129 /* Invalidate any task state remaining in the fpsimd regs: */
1130 fpsimd_flush_cpu_state();
1136 EXPORT_SYMBOL(kernel_neon_begin);
1139 * kernel_neon_end(): give the CPU FPSIMD registers back to the current task
1141 * Must be called from a context in which kernel_neon_begin() was previously
1142 * called, with no call to kernel_neon_end() in the meantime.
1144 * The caller must not use the FPSIMD registers after this function is called,
1145 * unless kernel_neon_begin() is called again in the meantime.
1147 void kernel_neon_end(void)
1151 if (!system_supports_fpsimd())
1154 busy = __this_cpu_xchg(kernel_neon_busy, false);
1155 WARN_ON(!busy); /* No matching kernel_neon_begin()? */
1159 EXPORT_SYMBOL(kernel_neon_end);
1163 static DEFINE_PER_CPU(struct user_fpsimd_state, efi_fpsimd_state);
1164 static DEFINE_PER_CPU(bool, efi_fpsimd_state_used);
1165 static DEFINE_PER_CPU(bool, efi_sve_state_used);
1168 * EFI runtime services support functions
1170 * The ABI for EFI runtime services allows EFI to use FPSIMD during the call.
1171 * This means that for EFI (and only for EFI), we have to assume that FPSIMD
1172 * is always used rather than being an optional accelerator.
1174 * These functions provide the necessary support for ensuring FPSIMD
1175 * save/restore in the contexts from which EFI is used.
1177 * Do not use them for any other purpose -- if tempted to do so, you are
1178 * either doing something wrong or you need to propose some refactoring.
1182 * __efi_fpsimd_begin(): prepare FPSIMD for making an EFI runtime services call
1184 void __efi_fpsimd_begin(void)
1186 if (!system_supports_fpsimd())
1189 WARN_ON(preemptible());
1191 if (may_use_simd()) {
1192 kernel_neon_begin();
1195 * If !efi_sve_state, SVE can't be in use yet and doesn't need
1198 if (system_supports_sve() && likely(efi_sve_state)) {
1199 char *sve_state = this_cpu_ptr(efi_sve_state);
1201 __this_cpu_write(efi_sve_state_used, true);
1203 sve_save_state(sve_state + sve_ffr_offset(sve_max_vl),
1204 &this_cpu_ptr(&efi_fpsimd_state)->fpsr);
1206 fpsimd_save_state(this_cpu_ptr(&efi_fpsimd_state));
1209 __this_cpu_write(efi_fpsimd_state_used, true);
1214 * __efi_fpsimd_end(): clean up FPSIMD after an EFI runtime services call
1216 void __efi_fpsimd_end(void)
1218 if (!system_supports_fpsimd())
1221 if (!__this_cpu_xchg(efi_fpsimd_state_used, false)) {
1224 if (system_supports_sve() &&
1225 likely(__this_cpu_read(efi_sve_state_used))) {
1226 char const *sve_state = this_cpu_ptr(efi_sve_state);
1228 sve_load_state(sve_state + sve_ffr_offset(sve_max_vl),
1229 &this_cpu_ptr(&efi_fpsimd_state)->fpsr,
1230 sve_vq_from_vl(sve_get_vl()) - 1);
1232 __this_cpu_write(efi_sve_state_used, false);
1234 fpsimd_load_state(this_cpu_ptr(&efi_fpsimd_state));
1239 #endif /* CONFIG_EFI */
1241 #endif /* CONFIG_KERNEL_MODE_NEON */
1243 #ifdef CONFIG_CPU_PM
1244 static int fpsimd_cpu_pm_notifier(struct notifier_block *self,
1245 unsigned long cmd, void *v)
1251 fpsimd_flush_cpu_state();
1255 set_thread_flag(TIF_FOREIGN_FPSTATE);
1257 case CPU_PM_ENTER_FAILED:
1264 static struct notifier_block fpsimd_cpu_pm_notifier_block = {
1265 .notifier_call = fpsimd_cpu_pm_notifier,
1268 static void __init fpsimd_pm_init(void)
1270 cpu_pm_register_notifier(&fpsimd_cpu_pm_notifier_block);
1274 static inline void fpsimd_pm_init(void) { }
1275 #endif /* CONFIG_CPU_PM */
1277 #ifdef CONFIG_HOTPLUG_CPU
1278 static int fpsimd_cpu_dead(unsigned int cpu)
1280 per_cpu(fpsimd_last_state.st, cpu) = NULL;
1284 static inline void fpsimd_hotplug_init(void)
1286 cpuhp_setup_state_nocalls(CPUHP_ARM64_FPSIMD_DEAD, "arm64/fpsimd:dead",
1287 NULL, fpsimd_cpu_dead);
1291 static inline void fpsimd_hotplug_init(void) { }
1295 * FP/SIMD support code initialisation.
1297 static int __init fpsimd_init(void)
1299 if (elf_hwcap & HWCAP_FP) {
1301 fpsimd_hotplug_init();
1303 pr_notice("Floating-point is not implemented\n");
1306 if (!(elf_hwcap & HWCAP_ASIMD))
1307 pr_notice("Advanced SIMD is not implemented\n");
1309 return sve_sysctl_init();
1311 core_initcall(fpsimd_init);