~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/arch/arm64/kernel/fpsimd.c

Version: ~ [ linux-5.12-rc7 ] ~ [ linux-5.11.13 ] ~ [ linux-5.10.29 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.111 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.186 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.230 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.266 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.266 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

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

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp