xref: /openbmc/linux/arch/arm64/kernel/fpsimd.c (revision 8e74a48d)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * FP/SIMD context switching and fault handling
4  *
5  * Copyright (C) 2012 ARM Ltd.
6  * Author: Catalin Marinas <catalin.marinas@arm.com>
7  */
8 
9 #include <linux/bitmap.h>
10 #include <linux/bitops.h>
11 #include <linux/bottom_half.h>
12 #include <linux/bug.h>
13 #include <linux/cache.h>
14 #include <linux/compat.h>
15 #include <linux/compiler.h>
16 #include <linux/cpu.h>
17 #include <linux/cpu_pm.h>
18 #include <linux/ctype.h>
19 #include <linux/kernel.h>
20 #include <linux/linkage.h>
21 #include <linux/irqflags.h>
22 #include <linux/init.h>
23 #include <linux/percpu.h>
24 #include <linux/prctl.h>
25 #include <linux/preempt.h>
26 #include <linux/ptrace.h>
27 #include <linux/sched/signal.h>
28 #include <linux/sched/task_stack.h>
29 #include <linux/signal.h>
30 #include <linux/slab.h>
31 #include <linux/stddef.h>
32 #include <linux/sysctl.h>
33 #include <linux/swab.h>
34 
35 #include <asm/esr.h>
36 #include <asm/exception.h>
37 #include <asm/fpsimd.h>
38 #include <asm/cpufeature.h>
39 #include <asm/cputype.h>
40 #include <asm/neon.h>
41 #include <asm/processor.h>
42 #include <asm/simd.h>
43 #include <asm/sigcontext.h>
44 #include <asm/sysreg.h>
45 #include <asm/traps.h>
46 #include <asm/virt.h>
47 
48 #define FPEXC_IOF	(1 << 0)
49 #define FPEXC_DZF	(1 << 1)
50 #define FPEXC_OFF	(1 << 2)
51 #define FPEXC_UFF	(1 << 3)
52 #define FPEXC_IXF	(1 << 4)
53 #define FPEXC_IDF	(1 << 7)
54 
55 /*
56  * (Note: in this discussion, statements about FPSIMD apply equally to SVE.)
57  *
58  * In order to reduce the number of times the FPSIMD state is needlessly saved
59  * and restored, we need to keep track of two things:
60  * (a) for each task, we need to remember which CPU was the last one to have
61  *     the task's FPSIMD state loaded into its FPSIMD registers;
62  * (b) for each CPU, we need to remember which task's userland FPSIMD state has
63  *     been loaded into its FPSIMD registers most recently, or whether it has
64  *     been used to perform kernel mode NEON in the meantime.
65  *
66  * For (a), we add a fpsimd_cpu field to thread_struct, which gets updated to
67  * the id of the current CPU every time the state is loaded onto a CPU. For (b),
68  * we add the per-cpu variable 'fpsimd_last_state' (below), which contains the
69  * address of the userland FPSIMD state of the task that was loaded onto the CPU
70  * the most recently, or NULL if kernel mode NEON has been performed after that.
71  *
72  * With this in place, we no longer have to restore the next FPSIMD state right
73  * when switching between tasks. Instead, we can defer this check to userland
74  * resume, at which time we verify whether the CPU's fpsimd_last_state and the
75  * task's fpsimd_cpu are still mutually in sync. If this is the case, we
76  * can omit the FPSIMD restore.
77  *
78  * As an optimization, we use the thread_info flag TIF_FOREIGN_FPSTATE to
79  * indicate whether or not the userland FPSIMD state of the current task is
80  * present in the registers. The flag is set unless the FPSIMD registers of this
81  * CPU currently contain the most recent userland FPSIMD state of the current
82  * task. If the task is behaving as a VMM, then this is will be managed by
83  * KVM which will clear it to indicate that the vcpu FPSIMD state is currently
84  * loaded on the CPU, allowing the state to be saved if a FPSIMD-aware
85  * softirq kicks in. Upon vcpu_put(), KVM will save the vcpu FP state and
86  * flag the register state as invalid.
87  *
88  * In order to allow softirq handlers to use FPSIMD, kernel_neon_begin() may
89  * save the task's FPSIMD context back to task_struct from softirq context.
90  * To prevent this from racing with the manipulation of the task's FPSIMD state
91  * from task context and thereby corrupting the state, it is necessary to
92  * protect any manipulation of a task's fpsimd_state or TIF_FOREIGN_FPSTATE
93  * flag with {, __}get_cpu_fpsimd_context(). This will still allow softirqs to
94  * run but prevent them to use FPSIMD.
95  *
96  * For a certain task, the sequence may look something like this:
97  * - the task gets scheduled in; if both the task's fpsimd_cpu field
98  *   contains the id of the current CPU, and the CPU's fpsimd_last_state per-cpu
99  *   variable points to the task's fpsimd_state, the TIF_FOREIGN_FPSTATE flag is
100  *   cleared, otherwise it is set;
101  *
102  * - the task returns to userland; if TIF_FOREIGN_FPSTATE is set, the task's
103  *   userland FPSIMD state is copied from memory to the registers, the task's
104  *   fpsimd_cpu field is set to the id of the current CPU, the current
105  *   CPU's fpsimd_last_state pointer is set to this task's fpsimd_state and the
106  *   TIF_FOREIGN_FPSTATE flag is cleared;
107  *
108  * - the task executes an ordinary syscall; upon return to userland, the
109  *   TIF_FOREIGN_FPSTATE flag will still be cleared, so no FPSIMD state is
110  *   restored;
111  *
112  * - the task executes a syscall which executes some NEON instructions; this is
113  *   preceded by a call to kernel_neon_begin(), which copies the task's FPSIMD
114  *   register contents to memory, clears the fpsimd_last_state per-cpu variable
115  *   and sets the TIF_FOREIGN_FPSTATE flag;
116  *
117  * - the task gets preempted after kernel_neon_end() is called; as we have not
118  *   returned from the 2nd syscall yet, TIF_FOREIGN_FPSTATE is still set so
119  *   whatever is in the FPSIMD registers is not saved to memory, but discarded.
120  */
121 struct fpsimd_last_state_struct {
122 	struct user_fpsimd_state *st;
123 	void *sve_state;
124 	unsigned int sve_vl;
125 };
126 
127 static DEFINE_PER_CPU(struct fpsimd_last_state_struct, fpsimd_last_state);
128 
129 __ro_after_init struct vl_info vl_info[ARM64_VEC_MAX] = {
130 #ifdef CONFIG_ARM64_SVE
131 	[ARM64_VEC_SVE] = {
132 		.type			= ARM64_VEC_SVE,
133 		.name			= "SVE",
134 		.min_vl			= SVE_VL_MIN,
135 		.max_vl			= SVE_VL_MIN,
136 		.max_virtualisable_vl	= SVE_VL_MIN,
137 	},
138 #endif
139 };
140 
141 static unsigned int vec_vl_inherit_flag(enum vec_type type)
142 {
143 	switch (type) {
144 	case ARM64_VEC_SVE:
145 		return TIF_SVE_VL_INHERIT;
146 	default:
147 		WARN_ON_ONCE(1);
148 		return 0;
149 	}
150 }
151 
152 struct vl_config {
153 	int __default_vl;		/* Default VL for tasks */
154 };
155 
156 static struct vl_config vl_config[ARM64_VEC_MAX];
157 
158 static inline int get_default_vl(enum vec_type type)
159 {
160 	return READ_ONCE(vl_config[type].__default_vl);
161 }
162 
163 #ifdef CONFIG_ARM64_SVE
164 
165 static inline int get_sve_default_vl(void)
166 {
167 	return get_default_vl(ARM64_VEC_SVE);
168 }
169 
170 static inline void set_default_vl(enum vec_type type, int val)
171 {
172 	WRITE_ONCE(vl_config[type].__default_vl, val);
173 }
174 
175 static inline void set_sve_default_vl(int val)
176 {
177 	set_default_vl(ARM64_VEC_SVE, val);
178 }
179 
180 static void __percpu *efi_sve_state;
181 
182 #else /* ! CONFIG_ARM64_SVE */
183 
184 /* Dummy declaration for code that will be optimised out: */
185 extern void __percpu *efi_sve_state;
186 
187 #endif /* ! CONFIG_ARM64_SVE */
188 
189 DEFINE_PER_CPU(bool, fpsimd_context_busy);
190 EXPORT_PER_CPU_SYMBOL(fpsimd_context_busy);
191 
192 static void fpsimd_bind_task_to_cpu(void);
193 
194 static void __get_cpu_fpsimd_context(void)
195 {
196 	bool busy = __this_cpu_xchg(fpsimd_context_busy, true);
197 
198 	WARN_ON(busy);
199 }
200 
201 /*
202  * Claim ownership of the CPU FPSIMD context for use by the calling context.
203  *
204  * The caller may freely manipulate the FPSIMD context metadata until
205  * put_cpu_fpsimd_context() is called.
206  *
207  * The double-underscore version must only be called if you know the task
208  * can't be preempted.
209  */
210 static void get_cpu_fpsimd_context(void)
211 {
212 	local_bh_disable();
213 	__get_cpu_fpsimd_context();
214 }
215 
216 static void __put_cpu_fpsimd_context(void)
217 {
218 	bool busy = __this_cpu_xchg(fpsimd_context_busy, false);
219 
220 	WARN_ON(!busy); /* No matching get_cpu_fpsimd_context()? */
221 }
222 
223 /*
224  * Release the CPU FPSIMD context.
225  *
226  * Must be called from a context in which get_cpu_fpsimd_context() was
227  * previously called, with no call to put_cpu_fpsimd_context() in the
228  * meantime.
229  */
230 static void put_cpu_fpsimd_context(void)
231 {
232 	__put_cpu_fpsimd_context();
233 	local_bh_enable();
234 }
235 
236 static bool have_cpu_fpsimd_context(void)
237 {
238 	return !preemptible() && __this_cpu_read(fpsimd_context_busy);
239 }
240 
241 /*
242  * Call __sve_free() directly only if you know task can't be scheduled
243  * or preempted.
244  */
245 static void __sve_free(struct task_struct *task)
246 {
247 	kfree(task->thread.sve_state);
248 	task->thread.sve_state = NULL;
249 }
250 
251 static void sve_free(struct task_struct *task)
252 {
253 	WARN_ON(test_tsk_thread_flag(task, TIF_SVE));
254 
255 	__sve_free(task);
256 }
257 
258 unsigned int task_get_vl(const struct task_struct *task, enum vec_type type)
259 {
260 	return task->thread.vl[type];
261 }
262 
263 void task_set_vl(struct task_struct *task, enum vec_type type,
264 		 unsigned long vl)
265 {
266 	task->thread.vl[type] = vl;
267 }
268 
269 unsigned int task_get_vl_onexec(const struct task_struct *task,
270 				enum vec_type type)
271 {
272 	return task->thread.vl_onexec[type];
273 }
274 
275 void task_set_vl_onexec(struct task_struct *task, enum vec_type type,
276 			unsigned long vl)
277 {
278 	task->thread.vl_onexec[type] = vl;
279 }
280 
281 /*
282  * TIF_SVE controls whether a task can use SVE without trapping while
283  * in userspace, and also the way a task's FPSIMD/SVE state is stored
284  * in thread_struct.
285  *
286  * The kernel uses this flag to track whether a user task is actively
287  * using SVE, and therefore whether full SVE register state needs to
288  * be tracked.  If not, the cheaper FPSIMD context handling code can
289  * be used instead of the more costly SVE equivalents.
290  *
291  *  * TIF_SVE set:
292  *
293  *    The task can execute SVE instructions while in userspace without
294  *    trapping to the kernel.
295  *
296  *    When stored, Z0-Z31 (incorporating Vn in bits[127:0] or the
297  *    corresponding Zn), P0-P15 and FFR are encoded in in
298  *    task->thread.sve_state, formatted appropriately for vector
299  *    length task->thread.sve_vl.
300  *
301  *    task->thread.sve_state must point to a valid buffer at least
302  *    sve_state_size(task) bytes in size.
303  *
304  *    During any syscall, the kernel may optionally clear TIF_SVE and
305  *    discard the vector state except for the FPSIMD subset.
306  *
307  *  * TIF_SVE clear:
308  *
309  *    An attempt by the user task to execute an SVE instruction causes
310  *    do_sve_acc() to be called, which does some preparation and then
311  *    sets TIF_SVE.
312  *
313  *    When stored, FPSIMD registers V0-V31 are encoded in
314  *    task->thread.uw.fpsimd_state; bits [max : 128] for each of Z0-Z31 are
315  *    logically zero but not stored anywhere; P0-P15 and FFR are not
316  *    stored and have unspecified values from userspace's point of
317  *    view.  For hygiene purposes, the kernel zeroes them on next use,
318  *    but userspace is discouraged from relying on this.
319  *
320  *    task->thread.sve_state does not need to be non-NULL, valid or any
321  *    particular size: it must not be dereferenced.
322  *
323  *  * FPSR and FPCR are always stored in task->thread.uw.fpsimd_state
324  *    irrespective of whether TIF_SVE is clear or set, since these are
325  *    not vector length dependent.
326  */
327 
328 /*
329  * Update current's FPSIMD/SVE registers from thread_struct.
330  *
331  * This function should be called only when the FPSIMD/SVE state in
332  * thread_struct is known to be up to date, when preparing to enter
333  * userspace.
334  */
335 static void task_fpsimd_load(void)
336 {
337 	WARN_ON(!system_supports_fpsimd());
338 	WARN_ON(!have_cpu_fpsimd_context());
339 
340 	if (IS_ENABLED(CONFIG_ARM64_SVE) && test_thread_flag(TIF_SVE)) {
341 		sve_set_vq(sve_vq_from_vl(task_get_sve_vl(current)) - 1);
342 		sve_load_state(sve_pffr(&current->thread),
343 			       &current->thread.uw.fpsimd_state.fpsr, true);
344 	} else {
345 		fpsimd_load_state(&current->thread.uw.fpsimd_state);
346 	}
347 }
348 
349 /*
350  * Ensure FPSIMD/SVE storage in memory for the loaded context is up to
351  * date with respect to the CPU registers.
352  */
353 static void fpsimd_save(void)
354 {
355 	struct fpsimd_last_state_struct const *last =
356 		this_cpu_ptr(&fpsimd_last_state);
357 	/* set by fpsimd_bind_task_to_cpu() or fpsimd_bind_state_to_cpu() */
358 
359 	WARN_ON(!system_supports_fpsimd());
360 	WARN_ON(!have_cpu_fpsimd_context());
361 
362 	if (test_thread_flag(TIF_FOREIGN_FPSTATE))
363 		return;
364 
365 	if (IS_ENABLED(CONFIG_ARM64_SVE) &&
366 	    test_thread_flag(TIF_SVE)) {
367 		if (WARN_ON(sve_get_vl() != last->sve_vl)) {
368 			/*
369 			 * Can't save the user regs, so current would
370 			 * re-enter user with corrupt state.
371 			 * There's no way to recover, so kill it:
372 			 */
373 			force_signal_inject(SIGKILL, SI_KERNEL, 0, 0);
374 			return;
375 		}
376 
377 		sve_save_state((char *)last->sve_state +
378 					sve_ffr_offset(last->sve_vl),
379 			       &last->st->fpsr, true);
380 	} else {
381 		fpsimd_save_state(last->st);
382 	}
383 }
384 
385 /*
386  * All vector length selection from userspace comes through here.
387  * We're on a slow path, so some sanity-checks are included.
388  * If things go wrong there's a bug somewhere, but try to fall back to a
389  * safe choice.
390  */
391 static unsigned int find_supported_vector_length(enum vec_type type,
392 						 unsigned int vl)
393 {
394 	struct vl_info *info = &vl_info[type];
395 	int bit;
396 	int max_vl = info->max_vl;
397 
398 	if (WARN_ON(!sve_vl_valid(vl)))
399 		vl = info->min_vl;
400 
401 	if (WARN_ON(!sve_vl_valid(max_vl)))
402 		max_vl = info->min_vl;
403 
404 	if (vl > max_vl)
405 		vl = max_vl;
406 
407 	bit = find_next_bit(info->vq_map, SVE_VQ_MAX,
408 			    __vq_to_bit(sve_vq_from_vl(vl)));
409 	return sve_vl_from_vq(__bit_to_vq(bit));
410 }
411 
412 #if defined(CONFIG_ARM64_SVE) && defined(CONFIG_SYSCTL)
413 
414 static int vec_proc_do_default_vl(struct ctl_table *table, int write,
415 				  void *buffer, size_t *lenp, loff_t *ppos)
416 {
417 	struct vl_info *info = table->extra1;
418 	enum vec_type type = info->type;
419 	int ret;
420 	int vl = get_default_vl(type);
421 	struct ctl_table tmp_table = {
422 		.data = &vl,
423 		.maxlen = sizeof(vl),
424 	};
425 
426 	ret = proc_dointvec(&tmp_table, write, buffer, lenp, ppos);
427 	if (ret || !write)
428 		return ret;
429 
430 	/* Writing -1 has the special meaning "set to max": */
431 	if (vl == -1)
432 		vl = info->max_vl;
433 
434 	if (!sve_vl_valid(vl))
435 		return -EINVAL;
436 
437 	set_default_vl(type, find_supported_vector_length(type, vl));
438 	return 0;
439 }
440 
441 static struct ctl_table sve_default_vl_table[] = {
442 	{
443 		.procname	= "sve_default_vector_length",
444 		.mode		= 0644,
445 		.proc_handler	= vec_proc_do_default_vl,
446 		.extra1		= &vl_info[ARM64_VEC_SVE],
447 	},
448 	{ }
449 };
450 
451 static int __init sve_sysctl_init(void)
452 {
453 	if (system_supports_sve())
454 		if (!register_sysctl("abi", sve_default_vl_table))
455 			return -EINVAL;
456 
457 	return 0;
458 }
459 
460 #else /* ! (CONFIG_ARM64_SVE && CONFIG_SYSCTL) */
461 static int __init sve_sysctl_init(void) { return 0; }
462 #endif /* ! (CONFIG_ARM64_SVE && CONFIG_SYSCTL) */
463 
464 #define ZREG(sve_state, vq, n) ((char *)(sve_state) +		\
465 	(SVE_SIG_ZREG_OFFSET(vq, n) - SVE_SIG_REGS_OFFSET))
466 
467 #ifdef CONFIG_CPU_BIG_ENDIAN
468 static __uint128_t arm64_cpu_to_le128(__uint128_t x)
469 {
470 	u64 a = swab64(x);
471 	u64 b = swab64(x >> 64);
472 
473 	return ((__uint128_t)a << 64) | b;
474 }
475 #else
476 static __uint128_t arm64_cpu_to_le128(__uint128_t x)
477 {
478 	return x;
479 }
480 #endif
481 
482 #define arm64_le128_to_cpu(x) arm64_cpu_to_le128(x)
483 
484 static void __fpsimd_to_sve(void *sst, struct user_fpsimd_state const *fst,
485 			    unsigned int vq)
486 {
487 	unsigned int i;
488 	__uint128_t *p;
489 
490 	for (i = 0; i < SVE_NUM_ZREGS; ++i) {
491 		p = (__uint128_t *)ZREG(sst, vq, i);
492 		*p = arm64_cpu_to_le128(fst->vregs[i]);
493 	}
494 }
495 
496 /*
497  * Transfer the FPSIMD state in task->thread.uw.fpsimd_state to
498  * task->thread.sve_state.
499  *
500  * Task can be a non-runnable task, or current.  In the latter case,
501  * the caller must have ownership of the cpu FPSIMD context before calling
502  * this function.
503  * task->thread.sve_state must point to at least sve_state_size(task)
504  * bytes of allocated kernel memory.
505  * task->thread.uw.fpsimd_state must be up to date before calling this
506  * function.
507  */
508 static void fpsimd_to_sve(struct task_struct *task)
509 {
510 	unsigned int vq;
511 	void *sst = task->thread.sve_state;
512 	struct user_fpsimd_state const *fst = &task->thread.uw.fpsimd_state;
513 
514 	if (!system_supports_sve())
515 		return;
516 
517 	vq = sve_vq_from_vl(task_get_sve_vl(task));
518 	__fpsimd_to_sve(sst, fst, vq);
519 }
520 
521 /*
522  * Transfer the SVE state in task->thread.sve_state to
523  * task->thread.uw.fpsimd_state.
524  *
525  * Task can be a non-runnable task, or current.  In the latter case,
526  * the caller must have ownership of the cpu FPSIMD context before calling
527  * this function.
528  * task->thread.sve_state must point to at least sve_state_size(task)
529  * bytes of allocated kernel memory.
530  * task->thread.sve_state must be up to date before calling this function.
531  */
532 static void sve_to_fpsimd(struct task_struct *task)
533 {
534 	unsigned int vq;
535 	void const *sst = task->thread.sve_state;
536 	struct user_fpsimd_state *fst = &task->thread.uw.fpsimd_state;
537 	unsigned int i;
538 	__uint128_t const *p;
539 
540 	if (!system_supports_sve())
541 		return;
542 
543 	vq = sve_vq_from_vl(task_get_sve_vl(task));
544 	for (i = 0; i < SVE_NUM_ZREGS; ++i) {
545 		p = (__uint128_t const *)ZREG(sst, vq, i);
546 		fst->vregs[i] = arm64_le128_to_cpu(*p);
547 	}
548 }
549 
550 #ifdef CONFIG_ARM64_SVE
551 
552 /*
553  * Return how many bytes of memory are required to store the full SVE
554  * state for task, given task's currently configured vector length.
555  */
556 static size_t sve_state_size(struct task_struct const *task)
557 {
558 	return SVE_SIG_REGS_SIZE(sve_vq_from_vl(task_get_sve_vl(task)));
559 }
560 
561 /*
562  * Ensure that task->thread.sve_state is allocated and sufficiently large.
563  *
564  * This function should be used only in preparation for replacing
565  * task->thread.sve_state with new data.  The memory is always zeroed
566  * here to prevent stale data from showing through: this is done in
567  * the interest of testability and predictability: except in the
568  * do_sve_acc() case, there is no ABI requirement to hide stale data
569  * written previously be task.
570  */
571 void sve_alloc(struct task_struct *task)
572 {
573 	if (task->thread.sve_state) {
574 		memset(task->thread.sve_state, 0, sve_state_size(task));
575 		return;
576 	}
577 
578 	/* This is a small allocation (maximum ~8KB) and Should Not Fail. */
579 	task->thread.sve_state =
580 		kzalloc(sve_state_size(task), GFP_KERNEL);
581 }
582 
583 
584 /*
585  * Ensure that task->thread.sve_state is up to date with respect to
586  * the user task, irrespective of when SVE is in use or not.
587  *
588  * This should only be called by ptrace.  task must be non-runnable.
589  * task->thread.sve_state must point to at least sve_state_size(task)
590  * bytes of allocated kernel memory.
591  */
592 void fpsimd_sync_to_sve(struct task_struct *task)
593 {
594 	if (!test_tsk_thread_flag(task, TIF_SVE))
595 		fpsimd_to_sve(task);
596 }
597 
598 /*
599  * Ensure that task->thread.uw.fpsimd_state is up to date with respect to
600  * the user task, irrespective of whether SVE is in use or not.
601  *
602  * This should only be called by ptrace.  task must be non-runnable.
603  * task->thread.sve_state must point to at least sve_state_size(task)
604  * bytes of allocated kernel memory.
605  */
606 void sve_sync_to_fpsimd(struct task_struct *task)
607 {
608 	if (test_tsk_thread_flag(task, TIF_SVE))
609 		sve_to_fpsimd(task);
610 }
611 
612 /*
613  * Ensure that task->thread.sve_state is up to date with respect to
614  * the task->thread.uw.fpsimd_state.
615  *
616  * This should only be called by ptrace to merge new FPSIMD register
617  * values into a task for which SVE is currently active.
618  * task must be non-runnable.
619  * task->thread.sve_state must point to at least sve_state_size(task)
620  * bytes of allocated kernel memory.
621  * task->thread.uw.fpsimd_state must already have been initialised with
622  * the new FPSIMD register values to be merged in.
623  */
624 void sve_sync_from_fpsimd_zeropad(struct task_struct *task)
625 {
626 	unsigned int vq;
627 	void *sst = task->thread.sve_state;
628 	struct user_fpsimd_state const *fst = &task->thread.uw.fpsimd_state;
629 
630 	if (!test_tsk_thread_flag(task, TIF_SVE))
631 		return;
632 
633 	vq = sve_vq_from_vl(task_get_sve_vl(task));
634 
635 	memset(sst, 0, SVE_SIG_REGS_SIZE(vq));
636 	__fpsimd_to_sve(sst, fst, vq);
637 }
638 
639 int vec_set_vector_length(struct task_struct *task, enum vec_type type,
640 			  unsigned long vl, unsigned long flags)
641 {
642 	if (flags & ~(unsigned long)(PR_SVE_VL_INHERIT |
643 				     PR_SVE_SET_VL_ONEXEC))
644 		return -EINVAL;
645 
646 	if (!sve_vl_valid(vl))
647 		return -EINVAL;
648 
649 	/*
650 	 * Clamp to the maximum vector length that VL-agnostic code
651 	 * can work with.  A flag may be assigned in the future to
652 	 * allow setting of larger vector lengths without confusing
653 	 * older software.
654 	 */
655 	if (vl > VL_ARCH_MAX)
656 		vl = VL_ARCH_MAX;
657 
658 	vl = find_supported_vector_length(type, vl);
659 
660 	if (flags & (PR_SVE_VL_INHERIT |
661 		     PR_SVE_SET_VL_ONEXEC))
662 		task_set_vl_onexec(task, type, vl);
663 	else
664 		/* Reset VL to system default on next exec: */
665 		task_set_vl_onexec(task, type, 0);
666 
667 	/* Only actually set the VL if not deferred: */
668 	if (flags & PR_SVE_SET_VL_ONEXEC)
669 		goto out;
670 
671 	if (vl == task_get_vl(task, type))
672 		goto out;
673 
674 	/*
675 	 * To ensure the FPSIMD bits of the SVE vector registers are preserved,
676 	 * write any live register state back to task_struct, and convert to a
677 	 * regular FPSIMD thread.  Since the vector length can only be changed
678 	 * with a syscall we can't be in streaming mode while reconfiguring.
679 	 */
680 	if (task == current) {
681 		get_cpu_fpsimd_context();
682 
683 		fpsimd_save();
684 	}
685 
686 	fpsimd_flush_task_state(task);
687 	if (test_and_clear_tsk_thread_flag(task, TIF_SVE))
688 		sve_to_fpsimd(task);
689 
690 	if (task == current)
691 		put_cpu_fpsimd_context();
692 
693 	/*
694 	 * Force reallocation of task SVE state to the correct size
695 	 * on next use:
696 	 */
697 	sve_free(task);
698 
699 	task_set_vl(task, type, vl);
700 
701 out:
702 	update_tsk_thread_flag(task, vec_vl_inherit_flag(type),
703 			       flags & PR_SVE_VL_INHERIT);
704 
705 	return 0;
706 }
707 
708 /*
709  * Encode the current vector length and flags for return.
710  * This is only required for prctl(): ptrace has separate fields.
711  * SVE and SME use the same bits for _ONEXEC and _INHERIT.
712  *
713  * flags are as for vec_set_vector_length().
714  */
715 static int vec_prctl_status(enum vec_type type, unsigned long flags)
716 {
717 	int ret;
718 
719 	if (flags & PR_SVE_SET_VL_ONEXEC)
720 		ret = task_get_vl_onexec(current, type);
721 	else
722 		ret = task_get_vl(current, type);
723 
724 	if (test_thread_flag(vec_vl_inherit_flag(type)))
725 		ret |= PR_SVE_VL_INHERIT;
726 
727 	return ret;
728 }
729 
730 /* PR_SVE_SET_VL */
731 int sve_set_current_vl(unsigned long arg)
732 {
733 	unsigned long vl, flags;
734 	int ret;
735 
736 	vl = arg & PR_SVE_VL_LEN_MASK;
737 	flags = arg & ~vl;
738 
739 	if (!system_supports_sve() || is_compat_task())
740 		return -EINVAL;
741 
742 	ret = vec_set_vector_length(current, ARM64_VEC_SVE, vl, flags);
743 	if (ret)
744 		return ret;
745 
746 	return vec_prctl_status(ARM64_VEC_SVE, flags);
747 }
748 
749 /* PR_SVE_GET_VL */
750 int sve_get_current_vl(void)
751 {
752 	if (!system_supports_sve() || is_compat_task())
753 		return -EINVAL;
754 
755 	return vec_prctl_status(ARM64_VEC_SVE, 0);
756 }
757 
758 static void vec_probe_vqs(struct vl_info *info,
759 			  DECLARE_BITMAP(map, SVE_VQ_MAX))
760 {
761 	unsigned int vq, vl;
762 
763 	bitmap_zero(map, SVE_VQ_MAX);
764 
765 	for (vq = SVE_VQ_MAX; vq >= SVE_VQ_MIN; --vq) {
766 		write_vl(info->type, vq - 1); /* self-syncing */
767 		vl = sve_get_vl();
768 		vq = sve_vq_from_vl(vl); /* skip intervening lengths */
769 		set_bit(__vq_to_bit(vq), map);
770 	}
771 }
772 
773 /*
774  * Initialise the set of known supported VQs for the boot CPU.
775  * This is called during kernel boot, before secondary CPUs are brought up.
776  */
777 void __init vec_init_vq_map(enum vec_type type)
778 {
779 	struct vl_info *info = &vl_info[type];
780 	vec_probe_vqs(info, info->vq_map);
781 	bitmap_copy(info->vq_partial_map, info->vq_map, SVE_VQ_MAX);
782 }
783 
784 /*
785  * If we haven't committed to the set of supported VQs yet, filter out
786  * those not supported by the current CPU.
787  * This function is called during the bring-up of early secondary CPUs only.
788  */
789 void vec_update_vq_map(enum vec_type type)
790 {
791 	struct vl_info *info = &vl_info[type];
792 	DECLARE_BITMAP(tmp_map, SVE_VQ_MAX);
793 
794 	vec_probe_vqs(info, tmp_map);
795 	bitmap_and(info->vq_map, info->vq_map, tmp_map, SVE_VQ_MAX);
796 	bitmap_or(info->vq_partial_map, info->vq_partial_map, tmp_map,
797 		  SVE_VQ_MAX);
798 }
799 
800 /*
801  * Check whether the current CPU supports all VQs in the committed set.
802  * This function is called during the bring-up of late secondary CPUs only.
803  */
804 int vec_verify_vq_map(enum vec_type type)
805 {
806 	struct vl_info *info = &vl_info[type];
807 	DECLARE_BITMAP(tmp_map, SVE_VQ_MAX);
808 	unsigned long b;
809 
810 	vec_probe_vqs(info, tmp_map);
811 
812 	bitmap_complement(tmp_map, tmp_map, SVE_VQ_MAX);
813 	if (bitmap_intersects(tmp_map, info->vq_map, SVE_VQ_MAX)) {
814 		pr_warn("%s: cpu%d: Required vector length(s) missing\n",
815 			info->name, smp_processor_id());
816 		return -EINVAL;
817 	}
818 
819 	if (!IS_ENABLED(CONFIG_KVM) || !is_hyp_mode_available())
820 		return 0;
821 
822 	/*
823 	 * For KVM, it is necessary to ensure that this CPU doesn't
824 	 * support any vector length that guests may have probed as
825 	 * unsupported.
826 	 */
827 
828 	/* Recover the set of supported VQs: */
829 	bitmap_complement(tmp_map, tmp_map, SVE_VQ_MAX);
830 	/* Find VQs supported that are not globally supported: */
831 	bitmap_andnot(tmp_map, tmp_map, info->vq_map, SVE_VQ_MAX);
832 
833 	/* Find the lowest such VQ, if any: */
834 	b = find_last_bit(tmp_map, SVE_VQ_MAX);
835 	if (b >= SVE_VQ_MAX)
836 		return 0; /* no mismatches */
837 
838 	/*
839 	 * Mismatches above sve_max_virtualisable_vl are fine, since
840 	 * no guest is allowed to configure ZCR_EL2.LEN to exceed this:
841 	 */
842 	if (sve_vl_from_vq(__bit_to_vq(b)) <= info->max_virtualisable_vl) {
843 		pr_warn("%s: cpu%d: Unsupported vector length(s) present\n",
844 			info->name, smp_processor_id());
845 		return -EINVAL;
846 	}
847 
848 	return 0;
849 }
850 
851 static void __init sve_efi_setup(void)
852 {
853 	struct vl_info *info = &vl_info[ARM64_VEC_SVE];
854 
855 	if (!IS_ENABLED(CONFIG_EFI))
856 		return;
857 
858 	/*
859 	 * alloc_percpu() warns and prints a backtrace if this goes wrong.
860 	 * This is evidence of a crippled system and we are returning void,
861 	 * so no attempt is made to handle this situation here.
862 	 */
863 	if (!sve_vl_valid(info->max_vl))
864 		goto fail;
865 
866 	efi_sve_state = __alloc_percpu(
867 		SVE_SIG_REGS_SIZE(sve_vq_from_vl(info->max_vl)), SVE_VQ_BYTES);
868 	if (!efi_sve_state)
869 		goto fail;
870 
871 	return;
872 
873 fail:
874 	panic("Cannot allocate percpu memory for EFI SVE save/restore");
875 }
876 
877 /*
878  * Enable SVE for EL1.
879  * Intended for use by the cpufeatures code during CPU boot.
880  */
881 void sve_kernel_enable(const struct arm64_cpu_capabilities *__always_unused p)
882 {
883 	write_sysreg(read_sysreg(CPACR_EL1) | CPACR_EL1_ZEN_EL1EN, CPACR_EL1);
884 	isb();
885 }
886 
887 /*
888  * Read the pseudo-ZCR used by cpufeatures to identify the supported SVE
889  * vector length.
890  *
891  * Use only if SVE is present.
892  * This function clobbers the SVE vector length.
893  */
894 u64 read_zcr_features(void)
895 {
896 	u64 zcr;
897 	unsigned int vq_max;
898 
899 	/*
900 	 * Set the maximum possible VL, and write zeroes to all other
901 	 * bits to see if they stick.
902 	 */
903 	sve_kernel_enable(NULL);
904 	write_sysreg_s(ZCR_ELx_LEN_MASK, SYS_ZCR_EL1);
905 
906 	zcr = read_sysreg_s(SYS_ZCR_EL1);
907 	zcr &= ~(u64)ZCR_ELx_LEN_MASK; /* find sticky 1s outside LEN field */
908 	vq_max = sve_vq_from_vl(sve_get_vl());
909 	zcr |= vq_max - 1; /* set LEN field to maximum effective value */
910 
911 	return zcr;
912 }
913 
914 void __init sve_setup(void)
915 {
916 	struct vl_info *info = &vl_info[ARM64_VEC_SVE];
917 	u64 zcr;
918 	DECLARE_BITMAP(tmp_map, SVE_VQ_MAX);
919 	unsigned long b;
920 
921 	if (!system_supports_sve())
922 		return;
923 
924 	/*
925 	 * The SVE architecture mandates support for 128-bit vectors,
926 	 * so sve_vq_map must have at least SVE_VQ_MIN set.
927 	 * If something went wrong, at least try to patch it up:
928 	 */
929 	if (WARN_ON(!test_bit(__vq_to_bit(SVE_VQ_MIN), info->vq_map)))
930 		set_bit(__vq_to_bit(SVE_VQ_MIN), info->vq_map);
931 
932 	zcr = read_sanitised_ftr_reg(SYS_ZCR_EL1);
933 	info->max_vl = sve_vl_from_vq((zcr & ZCR_ELx_LEN_MASK) + 1);
934 
935 	/*
936 	 * Sanity-check that the max VL we determined through CPU features
937 	 * corresponds properly to sve_vq_map.  If not, do our best:
938 	 */
939 	if (WARN_ON(info->max_vl != find_supported_vector_length(ARM64_VEC_SVE,
940 								 info->max_vl)))
941 		info->max_vl = find_supported_vector_length(ARM64_VEC_SVE,
942 							    info->max_vl);
943 
944 	/*
945 	 * For the default VL, pick the maximum supported value <= 64.
946 	 * VL == 64 is guaranteed not to grow the signal frame.
947 	 */
948 	set_sve_default_vl(find_supported_vector_length(ARM64_VEC_SVE, 64));
949 
950 	bitmap_andnot(tmp_map, info->vq_partial_map, info->vq_map,
951 		      SVE_VQ_MAX);
952 
953 	b = find_last_bit(tmp_map, SVE_VQ_MAX);
954 	if (b >= SVE_VQ_MAX)
955 		/* No non-virtualisable VLs found */
956 		info->max_virtualisable_vl = SVE_VQ_MAX;
957 	else if (WARN_ON(b == SVE_VQ_MAX - 1))
958 		/* No virtualisable VLs?  This is architecturally forbidden. */
959 		info->max_virtualisable_vl = SVE_VQ_MIN;
960 	else /* b + 1 < SVE_VQ_MAX */
961 		info->max_virtualisable_vl = sve_vl_from_vq(__bit_to_vq(b + 1));
962 
963 	if (info->max_virtualisable_vl > info->max_vl)
964 		info->max_virtualisable_vl = info->max_vl;
965 
966 	pr_info("%s: maximum available vector length %u bytes per vector\n",
967 		info->name, info->max_vl);
968 	pr_info("%s: default vector length %u bytes per vector\n",
969 		info->name, get_sve_default_vl());
970 
971 	/* KVM decides whether to support mismatched systems. Just warn here: */
972 	if (sve_max_virtualisable_vl() < sve_max_vl())
973 		pr_warn("%s: unvirtualisable vector lengths present\n",
974 			info->name);
975 
976 	sve_efi_setup();
977 }
978 
979 /*
980  * Called from the put_task_struct() path, which cannot get here
981  * unless dead_task is really dead and not schedulable.
982  */
983 void fpsimd_release_task(struct task_struct *dead_task)
984 {
985 	__sve_free(dead_task);
986 }
987 
988 #endif /* CONFIG_ARM64_SVE */
989 
990 /*
991  * Trapped SVE access
992  *
993  * Storage is allocated for the full SVE state, the current FPSIMD
994  * register contents are migrated across, and the access trap is
995  * disabled.
996  *
997  * TIF_SVE should be clear on entry: otherwise, fpsimd_restore_current_state()
998  * would have disabled the SVE access trap for userspace during
999  * ret_to_user, making an SVE access trap impossible in that case.
1000  */
1001 void do_sve_acc(unsigned int esr, struct pt_regs *regs)
1002 {
1003 	/* Even if we chose not to use SVE, the hardware could still trap: */
1004 	if (unlikely(!system_supports_sve()) || WARN_ON(is_compat_task())) {
1005 		force_signal_inject(SIGILL, ILL_ILLOPC, regs->pc, 0);
1006 		return;
1007 	}
1008 
1009 	sve_alloc(current);
1010 	if (!current->thread.sve_state) {
1011 		force_sig(SIGKILL);
1012 		return;
1013 	}
1014 
1015 	get_cpu_fpsimd_context();
1016 
1017 	if (test_and_set_thread_flag(TIF_SVE))
1018 		WARN_ON(1); /* SVE access shouldn't have trapped */
1019 
1020 	/*
1021 	 * Convert the FPSIMD state to SVE, zeroing all the state that
1022 	 * is not shared with FPSIMD. If (as is likely) the current
1023 	 * state is live in the registers then do this there and
1024 	 * update our metadata for the current task including
1025 	 * disabling the trap, otherwise update our in-memory copy.
1026 	 */
1027 	if (!test_thread_flag(TIF_FOREIGN_FPSTATE)) {
1028 		unsigned long vq_minus_one =
1029 			sve_vq_from_vl(task_get_sve_vl(current)) - 1;
1030 		sve_set_vq(vq_minus_one);
1031 		sve_flush_live(true, vq_minus_one);
1032 		fpsimd_bind_task_to_cpu();
1033 	} else {
1034 		fpsimd_to_sve(current);
1035 	}
1036 
1037 	put_cpu_fpsimd_context();
1038 }
1039 
1040 /*
1041  * Trapped FP/ASIMD access.
1042  */
1043 void do_fpsimd_acc(unsigned int esr, struct pt_regs *regs)
1044 {
1045 	/* TODO: implement lazy context saving/restoring */
1046 	WARN_ON(1);
1047 }
1048 
1049 /*
1050  * Raise a SIGFPE for the current process.
1051  */
1052 void do_fpsimd_exc(unsigned int esr, struct pt_regs *regs)
1053 {
1054 	unsigned int si_code = FPE_FLTUNK;
1055 
1056 	if (esr & ESR_ELx_FP_EXC_TFV) {
1057 		if (esr & FPEXC_IOF)
1058 			si_code = FPE_FLTINV;
1059 		else if (esr & FPEXC_DZF)
1060 			si_code = FPE_FLTDIV;
1061 		else if (esr & FPEXC_OFF)
1062 			si_code = FPE_FLTOVF;
1063 		else if (esr & FPEXC_UFF)
1064 			si_code = FPE_FLTUND;
1065 		else if (esr & FPEXC_IXF)
1066 			si_code = FPE_FLTRES;
1067 	}
1068 
1069 	send_sig_fault(SIGFPE, si_code,
1070 		       (void __user *)instruction_pointer(regs),
1071 		       current);
1072 }
1073 
1074 void fpsimd_thread_switch(struct task_struct *next)
1075 {
1076 	bool wrong_task, wrong_cpu;
1077 
1078 	if (!system_supports_fpsimd())
1079 		return;
1080 
1081 	__get_cpu_fpsimd_context();
1082 
1083 	/* Save unsaved fpsimd state, if any: */
1084 	fpsimd_save();
1085 
1086 	/*
1087 	 * Fix up TIF_FOREIGN_FPSTATE to correctly describe next's
1088 	 * state.  For kernel threads, FPSIMD registers are never loaded
1089 	 * and wrong_task and wrong_cpu will always be true.
1090 	 */
1091 	wrong_task = __this_cpu_read(fpsimd_last_state.st) !=
1092 					&next->thread.uw.fpsimd_state;
1093 	wrong_cpu = next->thread.fpsimd_cpu != smp_processor_id();
1094 
1095 	update_tsk_thread_flag(next, TIF_FOREIGN_FPSTATE,
1096 			       wrong_task || wrong_cpu);
1097 
1098 	__put_cpu_fpsimd_context();
1099 }
1100 
1101 static void fpsimd_flush_thread_vl(enum vec_type type)
1102 {
1103 	int vl, supported_vl;
1104 
1105 	/*
1106 	 * Reset the task vector length as required.  This is where we
1107 	 * ensure that all user tasks have a valid vector length
1108 	 * configured: no kernel task can become a user task without
1109 	 * an exec and hence a call to this function.  By the time the
1110 	 * first call to this function is made, all early hardware
1111 	 * probing is complete, so __sve_default_vl should be valid.
1112 	 * If a bug causes this to go wrong, we make some noise and
1113 	 * try to fudge thread.sve_vl to a safe value here.
1114 	 */
1115 	vl = task_get_vl_onexec(current, type);
1116 	if (!vl)
1117 		vl = get_default_vl(type);
1118 
1119 	if (WARN_ON(!sve_vl_valid(vl)))
1120 		vl = vl_info[type].min_vl;
1121 
1122 	supported_vl = find_supported_vector_length(type, vl);
1123 	if (WARN_ON(supported_vl != vl))
1124 		vl = supported_vl;
1125 
1126 	task_set_vl(current, type, vl);
1127 
1128 	/*
1129 	 * If the task is not set to inherit, ensure that the vector
1130 	 * length will be reset by a subsequent exec:
1131 	 */
1132 	if (!test_thread_flag(vec_vl_inherit_flag(type)))
1133 		task_set_vl_onexec(current, type, 0);
1134 }
1135 
1136 void fpsimd_flush_thread(void)
1137 {
1138 	if (!system_supports_fpsimd())
1139 		return;
1140 
1141 	get_cpu_fpsimd_context();
1142 
1143 	fpsimd_flush_task_state(current);
1144 	memset(&current->thread.uw.fpsimd_state, 0,
1145 	       sizeof(current->thread.uw.fpsimd_state));
1146 
1147 	if (system_supports_sve()) {
1148 		clear_thread_flag(TIF_SVE);
1149 		sve_free(current);
1150 		fpsimd_flush_thread_vl(ARM64_VEC_SVE);
1151 	}
1152 
1153 	put_cpu_fpsimd_context();
1154 }
1155 
1156 /*
1157  * Save the userland FPSIMD state of 'current' to memory, but only if the state
1158  * currently held in the registers does in fact belong to 'current'
1159  */
1160 void fpsimd_preserve_current_state(void)
1161 {
1162 	if (!system_supports_fpsimd())
1163 		return;
1164 
1165 	get_cpu_fpsimd_context();
1166 	fpsimd_save();
1167 	put_cpu_fpsimd_context();
1168 }
1169 
1170 /*
1171  * Like fpsimd_preserve_current_state(), but ensure that
1172  * current->thread.uw.fpsimd_state is updated so that it can be copied to
1173  * the signal frame.
1174  */
1175 void fpsimd_signal_preserve_current_state(void)
1176 {
1177 	fpsimd_preserve_current_state();
1178 	if (test_thread_flag(TIF_SVE))
1179 		sve_to_fpsimd(current);
1180 }
1181 
1182 /*
1183  * Associate current's FPSIMD context with this cpu
1184  * The caller must have ownership of the cpu FPSIMD context before calling
1185  * this function.
1186  */
1187 static void fpsimd_bind_task_to_cpu(void)
1188 {
1189 	struct fpsimd_last_state_struct *last =
1190 		this_cpu_ptr(&fpsimd_last_state);
1191 
1192 	WARN_ON(!system_supports_fpsimd());
1193 	last->st = &current->thread.uw.fpsimd_state;
1194 	last->sve_state = current->thread.sve_state;
1195 	last->sve_vl = task_get_sve_vl(current);
1196 	current->thread.fpsimd_cpu = smp_processor_id();
1197 
1198 	if (system_supports_sve()) {
1199 		/* Toggle SVE trapping for userspace if needed */
1200 		if (test_thread_flag(TIF_SVE))
1201 			sve_user_enable();
1202 		else
1203 			sve_user_disable();
1204 
1205 		/* Serialised by exception return to user */
1206 	}
1207 }
1208 
1209 void fpsimd_bind_state_to_cpu(struct user_fpsimd_state *st, void *sve_state,
1210 			      unsigned int sve_vl)
1211 {
1212 	struct fpsimd_last_state_struct *last =
1213 		this_cpu_ptr(&fpsimd_last_state);
1214 
1215 	WARN_ON(!system_supports_fpsimd());
1216 	WARN_ON(!in_softirq() && !irqs_disabled());
1217 
1218 	last->st = st;
1219 	last->sve_state = sve_state;
1220 	last->sve_vl = sve_vl;
1221 }
1222 
1223 /*
1224  * Load the userland FPSIMD state of 'current' from memory, but only if the
1225  * FPSIMD state already held in the registers is /not/ the most recent FPSIMD
1226  * state of 'current'.  This is called when we are preparing to return to
1227  * userspace to ensure that userspace sees a good register state.
1228  */
1229 void fpsimd_restore_current_state(void)
1230 {
1231 	/*
1232 	 * For the tasks that were created before we detected the absence of
1233 	 * FP/SIMD, the TIF_FOREIGN_FPSTATE could be set via fpsimd_thread_switch(),
1234 	 * e.g, init. This could be then inherited by the children processes.
1235 	 * If we later detect that the system doesn't support FP/SIMD,
1236 	 * we must clear the flag for  all the tasks to indicate that the
1237 	 * FPSTATE is clean (as we can't have one) to avoid looping for ever in
1238 	 * do_notify_resume().
1239 	 */
1240 	if (!system_supports_fpsimd()) {
1241 		clear_thread_flag(TIF_FOREIGN_FPSTATE);
1242 		return;
1243 	}
1244 
1245 	get_cpu_fpsimd_context();
1246 
1247 	if (test_and_clear_thread_flag(TIF_FOREIGN_FPSTATE)) {
1248 		task_fpsimd_load();
1249 		fpsimd_bind_task_to_cpu();
1250 	}
1251 
1252 	put_cpu_fpsimd_context();
1253 }
1254 
1255 /*
1256  * Load an updated userland FPSIMD state for 'current' from memory and set the
1257  * flag that indicates that the FPSIMD register contents are the most recent
1258  * FPSIMD state of 'current'. This is used by the signal code to restore the
1259  * register state when returning from a signal handler in FPSIMD only cases,
1260  * any SVE context will be discarded.
1261  */
1262 void fpsimd_update_current_state(struct user_fpsimd_state const *state)
1263 {
1264 	if (WARN_ON(!system_supports_fpsimd()))
1265 		return;
1266 
1267 	get_cpu_fpsimd_context();
1268 
1269 	current->thread.uw.fpsimd_state = *state;
1270 	if (test_thread_flag(TIF_SVE))
1271 		fpsimd_to_sve(current);
1272 
1273 	task_fpsimd_load();
1274 	fpsimd_bind_task_to_cpu();
1275 
1276 	clear_thread_flag(TIF_FOREIGN_FPSTATE);
1277 
1278 	put_cpu_fpsimd_context();
1279 }
1280 
1281 /*
1282  * Invalidate live CPU copies of task t's FPSIMD state
1283  *
1284  * This function may be called with preemption enabled.  The barrier()
1285  * ensures that the assignment to fpsimd_cpu is visible to any
1286  * preemption/softirq that could race with set_tsk_thread_flag(), so
1287  * that TIF_FOREIGN_FPSTATE cannot be spuriously re-cleared.
1288  *
1289  * The final barrier ensures that TIF_FOREIGN_FPSTATE is seen set by any
1290  * subsequent code.
1291  */
1292 void fpsimd_flush_task_state(struct task_struct *t)
1293 {
1294 	t->thread.fpsimd_cpu = NR_CPUS;
1295 	/*
1296 	 * If we don't support fpsimd, bail out after we have
1297 	 * reset the fpsimd_cpu for this task and clear the
1298 	 * FPSTATE.
1299 	 */
1300 	if (!system_supports_fpsimd())
1301 		return;
1302 	barrier();
1303 	set_tsk_thread_flag(t, TIF_FOREIGN_FPSTATE);
1304 
1305 	barrier();
1306 }
1307 
1308 /*
1309  * Invalidate any task's FPSIMD state that is present on this cpu.
1310  * The FPSIMD context should be acquired with get_cpu_fpsimd_context()
1311  * before calling this function.
1312  */
1313 static void fpsimd_flush_cpu_state(void)
1314 {
1315 	WARN_ON(!system_supports_fpsimd());
1316 	__this_cpu_write(fpsimd_last_state.st, NULL);
1317 	set_thread_flag(TIF_FOREIGN_FPSTATE);
1318 }
1319 
1320 /*
1321  * Save the FPSIMD state to memory and invalidate cpu view.
1322  * This function must be called with preemption disabled.
1323  */
1324 void fpsimd_save_and_flush_cpu_state(void)
1325 {
1326 	if (!system_supports_fpsimd())
1327 		return;
1328 	WARN_ON(preemptible());
1329 	__get_cpu_fpsimd_context();
1330 	fpsimd_save();
1331 	fpsimd_flush_cpu_state();
1332 	__put_cpu_fpsimd_context();
1333 }
1334 
1335 #ifdef CONFIG_KERNEL_MODE_NEON
1336 
1337 /*
1338  * Kernel-side NEON support functions
1339  */
1340 
1341 /*
1342  * kernel_neon_begin(): obtain the CPU FPSIMD registers for use by the calling
1343  * context
1344  *
1345  * Must not be called unless may_use_simd() returns true.
1346  * Task context in the FPSIMD registers is saved back to memory as necessary.
1347  *
1348  * A matching call to kernel_neon_end() must be made before returning from the
1349  * calling context.
1350  *
1351  * The caller may freely use the FPSIMD registers until kernel_neon_end() is
1352  * called.
1353  */
1354 void kernel_neon_begin(void)
1355 {
1356 	if (WARN_ON(!system_supports_fpsimd()))
1357 		return;
1358 
1359 	BUG_ON(!may_use_simd());
1360 
1361 	get_cpu_fpsimd_context();
1362 
1363 	/* Save unsaved fpsimd state, if any: */
1364 	fpsimd_save();
1365 
1366 	/* Invalidate any task state remaining in the fpsimd regs: */
1367 	fpsimd_flush_cpu_state();
1368 }
1369 EXPORT_SYMBOL(kernel_neon_begin);
1370 
1371 /*
1372  * kernel_neon_end(): give the CPU FPSIMD registers back to the current task
1373  *
1374  * Must be called from a context in which kernel_neon_begin() was previously
1375  * called, with no call to kernel_neon_end() in the meantime.
1376  *
1377  * The caller must not use the FPSIMD registers after this function is called,
1378  * unless kernel_neon_begin() is called again in the meantime.
1379  */
1380 void kernel_neon_end(void)
1381 {
1382 	if (!system_supports_fpsimd())
1383 		return;
1384 
1385 	put_cpu_fpsimd_context();
1386 }
1387 EXPORT_SYMBOL(kernel_neon_end);
1388 
1389 #ifdef CONFIG_EFI
1390 
1391 static DEFINE_PER_CPU(struct user_fpsimd_state, efi_fpsimd_state);
1392 static DEFINE_PER_CPU(bool, efi_fpsimd_state_used);
1393 static DEFINE_PER_CPU(bool, efi_sve_state_used);
1394 
1395 /*
1396  * EFI runtime services support functions
1397  *
1398  * The ABI for EFI runtime services allows EFI to use FPSIMD during the call.
1399  * This means that for EFI (and only for EFI), we have to assume that FPSIMD
1400  * is always used rather than being an optional accelerator.
1401  *
1402  * These functions provide the necessary support for ensuring FPSIMD
1403  * save/restore in the contexts from which EFI is used.
1404  *
1405  * Do not use them for any other purpose -- if tempted to do so, you are
1406  * either doing something wrong or you need to propose some refactoring.
1407  */
1408 
1409 /*
1410  * __efi_fpsimd_begin(): prepare FPSIMD for making an EFI runtime services call
1411  */
1412 void __efi_fpsimd_begin(void)
1413 {
1414 	if (!system_supports_fpsimd())
1415 		return;
1416 
1417 	WARN_ON(preemptible());
1418 
1419 	if (may_use_simd()) {
1420 		kernel_neon_begin();
1421 	} else {
1422 		/*
1423 		 * If !efi_sve_state, SVE can't be in use yet and doesn't need
1424 		 * preserving:
1425 		 */
1426 		if (system_supports_sve() && likely(efi_sve_state)) {
1427 			char *sve_state = this_cpu_ptr(efi_sve_state);
1428 
1429 			__this_cpu_write(efi_sve_state_used, true);
1430 
1431 			sve_save_state(sve_state + sve_ffr_offset(sve_max_vl()),
1432 				       &this_cpu_ptr(&efi_fpsimd_state)->fpsr,
1433 				       true);
1434 		} else {
1435 			fpsimd_save_state(this_cpu_ptr(&efi_fpsimd_state));
1436 		}
1437 
1438 		__this_cpu_write(efi_fpsimd_state_used, true);
1439 	}
1440 }
1441 
1442 /*
1443  * __efi_fpsimd_end(): clean up FPSIMD after an EFI runtime services call
1444  */
1445 void __efi_fpsimd_end(void)
1446 {
1447 	if (!system_supports_fpsimd())
1448 		return;
1449 
1450 	if (!__this_cpu_xchg(efi_fpsimd_state_used, false)) {
1451 		kernel_neon_end();
1452 	} else {
1453 		if (system_supports_sve() &&
1454 		    likely(__this_cpu_read(efi_sve_state_used))) {
1455 			char const *sve_state = this_cpu_ptr(efi_sve_state);
1456 
1457 			sve_set_vq(sve_vq_from_vl(sve_get_vl()) - 1);
1458 			sve_load_state(sve_state + sve_ffr_offset(sve_max_vl()),
1459 				       &this_cpu_ptr(&efi_fpsimd_state)->fpsr,
1460 				       true);
1461 
1462 			__this_cpu_write(efi_sve_state_used, false);
1463 		} else {
1464 			fpsimd_load_state(this_cpu_ptr(&efi_fpsimd_state));
1465 		}
1466 	}
1467 }
1468 
1469 #endif /* CONFIG_EFI */
1470 
1471 #endif /* CONFIG_KERNEL_MODE_NEON */
1472 
1473 #ifdef CONFIG_CPU_PM
1474 static int fpsimd_cpu_pm_notifier(struct notifier_block *self,
1475 				  unsigned long cmd, void *v)
1476 {
1477 	switch (cmd) {
1478 	case CPU_PM_ENTER:
1479 		fpsimd_save_and_flush_cpu_state();
1480 		break;
1481 	case CPU_PM_EXIT:
1482 		break;
1483 	case CPU_PM_ENTER_FAILED:
1484 	default:
1485 		return NOTIFY_DONE;
1486 	}
1487 	return NOTIFY_OK;
1488 }
1489 
1490 static struct notifier_block fpsimd_cpu_pm_notifier_block = {
1491 	.notifier_call = fpsimd_cpu_pm_notifier,
1492 };
1493 
1494 static void __init fpsimd_pm_init(void)
1495 {
1496 	cpu_pm_register_notifier(&fpsimd_cpu_pm_notifier_block);
1497 }
1498 
1499 #else
1500 static inline void fpsimd_pm_init(void) { }
1501 #endif /* CONFIG_CPU_PM */
1502 
1503 #ifdef CONFIG_HOTPLUG_CPU
1504 static int fpsimd_cpu_dead(unsigned int cpu)
1505 {
1506 	per_cpu(fpsimd_last_state.st, cpu) = NULL;
1507 	return 0;
1508 }
1509 
1510 static inline void fpsimd_hotplug_init(void)
1511 {
1512 	cpuhp_setup_state_nocalls(CPUHP_ARM64_FPSIMD_DEAD, "arm64/fpsimd:dead",
1513 				  NULL, fpsimd_cpu_dead);
1514 }
1515 
1516 #else
1517 static inline void fpsimd_hotplug_init(void) { }
1518 #endif
1519 
1520 /*
1521  * FP/SIMD support code initialisation.
1522  */
1523 static int __init fpsimd_init(void)
1524 {
1525 	if (cpu_have_named_feature(FP)) {
1526 		fpsimd_pm_init();
1527 		fpsimd_hotplug_init();
1528 	} else {
1529 		pr_notice("Floating-point is not implemented\n");
1530 	}
1531 
1532 	if (!cpu_have_named_feature(ASIMD))
1533 		pr_notice("Advanced SIMD is not implemented\n");
1534 
1535 	return sve_sysctl_init();
1536 }
1537 core_initcall(fpsimd_init);
1538