]>
Commit | Line | Data |
---|---|---|
53631b54 CM |
1 | /* |
2 | * FP/SIMD context switching and fault handling | |
3 | * | |
4 | * Copyright (C) 2012 ARM Ltd. | |
5 | * Author: Catalin Marinas <catalin.marinas@arm.com> | |
6 | * | |
7 | * This program is free software; you can redistribute it and/or modify | |
8 | * it under the terms of the GNU General Public License version 2 as | |
9 | * published by the Free Software Foundation. | |
10 | * | |
11 | * This program is distributed in the hope that it will be useful, | |
12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | * GNU General Public License for more details. | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
17 | * along with this program. If not, see <http://www.gnu.org/licenses/>. | |
18 | */ | |
19 | ||
7582e220 | 20 | #include <linux/bitmap.h> |
cb84d11e | 21 | #include <linux/bottom_half.h> |
bc0ee476 | 22 | #include <linux/bug.h> |
7582e220 | 23 | #include <linux/cache.h> |
bc0ee476 | 24 | #include <linux/compat.h> |
32365e64 | 25 | #include <linux/cpu.h> |
fb1ab1ab | 26 | #include <linux/cpu_pm.h> |
53631b54 | 27 | #include <linux/kernel.h> |
94ef7ecb | 28 | #include <linux/linkage.h> |
bc0ee476 | 29 | #include <linux/irqflags.h> |
53631b54 | 30 | #include <linux/init.h> |
cb84d11e | 31 | #include <linux/percpu.h> |
2d2123bc | 32 | #include <linux/prctl.h> |
4328825d | 33 | #include <linux/preempt.h> |
7582e220 | 34 | #include <linux/prctl.h> |
bc0ee476 | 35 | #include <linux/ptrace.h> |
3f07c014 | 36 | #include <linux/sched/signal.h> |
bc0ee476 | 37 | #include <linux/sched/task_stack.h> |
53631b54 | 38 | #include <linux/signal.h> |
bc0ee476 | 39 | #include <linux/slab.h> |
4ffa09a9 | 40 | #include <linux/sysctl.h> |
53631b54 CM |
41 | |
42 | #include <asm/fpsimd.h> | |
43 | #include <asm/cputype.h> | |
4328825d | 44 | #include <asm/simd.h> |
bc0ee476 DM |
45 | #include <asm/sigcontext.h> |
46 | #include <asm/sysreg.h> | |
47 | #include <asm/traps.h> | |
53631b54 CM |
48 | |
49 | #define FPEXC_IOF (1 << 0) | |
50 | #define FPEXC_DZF (1 << 1) | |
51 | #define FPEXC_OFF (1 << 2) | |
52 | #define FPEXC_UFF (1 << 3) | |
53 | #define FPEXC_IXF (1 << 4) | |
54 | #define FPEXC_IDF (1 << 7) | |
55 | ||
005f78cd | 56 | /* |
bc0ee476 DM |
57 | * (Note: in this discussion, statements about FPSIMD apply equally to SVE.) |
58 | * | |
005f78cd AB |
59 | * In order to reduce the number of times the FPSIMD state is needlessly saved |
60 | * and restored, we need to keep track of two things: | |
61 | * (a) for each task, we need to remember which CPU was the last one to have | |
62 | * the task's FPSIMD state loaded into its FPSIMD registers; | |
63 | * (b) for each CPU, we need to remember which task's userland FPSIMD state has | |
64 | * been loaded into its FPSIMD registers most recently, or whether it has | |
65 | * been used to perform kernel mode NEON in the meantime. | |
66 | * | |
67 | * For (a), we add a 'cpu' field to struct fpsimd_state, which gets updated to | |
ef769e32 | 68 | * the id of the current CPU every time the state is loaded onto a CPU. For (b), |
005f78cd AB |
69 | * we add the per-cpu variable 'fpsimd_last_state' (below), which contains the |
70 | * address of the userland FPSIMD state of the task that was loaded onto the CPU | |
71 | * the most recently, or NULL if kernel mode NEON has been performed after that. | |
72 | * | |
73 | * With this in place, we no longer have to restore the next FPSIMD state right | |
74 | * when switching between tasks. Instead, we can defer this check to userland | |
75 | * resume, at which time we verify whether the CPU's fpsimd_last_state and the | |
76 | * task's fpsimd_state.cpu are still mutually in sync. If this is the case, we | |
77 | * can omit the FPSIMD restore. | |
78 | * | |
79 | * As an optimization, we use the thread_info flag TIF_FOREIGN_FPSTATE to | |
80 | * indicate whether or not the userland FPSIMD state of the current task is | |
81 | * present in the registers. The flag is set unless the FPSIMD registers of this | |
82 | * CPU currently contain the most recent userland FPSIMD state of the current | |
83 | * task. | |
84 | * | |
cb84d11e DM |
85 | * In order to allow softirq handlers to use FPSIMD, kernel_neon_begin() may |
86 | * save the task's FPSIMD context back to task_struct from softirq context. | |
87 | * To prevent this from racing with the manipulation of the task's FPSIMD state | |
88 | * from task context and thereby corrupting the state, it is necessary to | |
89 | * protect any manipulation of a task's fpsimd_state or TIF_FOREIGN_FPSTATE | |
90 | * flag with local_bh_disable() unless softirqs are already masked. | |
91 | * | |
005f78cd AB |
92 | * For a certain task, the sequence may look something like this: |
93 | * - the task gets scheduled in; if both the task's fpsimd_state.cpu field | |
94 | * contains the id of the current CPU, and the CPU's fpsimd_last_state per-cpu | |
95 | * variable points to the task's fpsimd_state, the TIF_FOREIGN_FPSTATE flag is | |
96 | * cleared, otherwise it is set; | |
97 | * | |
98 | * - the task returns to userland; if TIF_FOREIGN_FPSTATE is set, the task's | |
99 | * userland FPSIMD state is copied from memory to the registers, the task's | |
100 | * fpsimd_state.cpu field is set to the id of the current CPU, the current | |
101 | * CPU's fpsimd_last_state pointer is set to this task's fpsimd_state and the | |
102 | * TIF_FOREIGN_FPSTATE flag is cleared; | |
103 | * | |
104 | * - the task executes an ordinary syscall; upon return to userland, the | |
105 | * TIF_FOREIGN_FPSTATE flag will still be cleared, so no FPSIMD state is | |
106 | * restored; | |
107 | * | |
108 | * - the task executes a syscall which executes some NEON instructions; this is | |
109 | * preceded by a call to kernel_neon_begin(), which copies the task's FPSIMD | |
110 | * register contents to memory, clears the fpsimd_last_state per-cpu variable | |
111 | * and sets the TIF_FOREIGN_FPSTATE flag; | |
112 | * | |
113 | * - the task gets preempted after kernel_neon_end() is called; as we have not | |
114 | * returned from the 2nd syscall yet, TIF_FOREIGN_FPSTATE is still set so | |
115 | * whatever is in the FPSIMD registers is not saved to memory, but discarded. | |
116 | */ | |
cb968afc DM |
117 | struct fpsimd_last_state_struct { |
118 | struct fpsimd_state *st; | |
119 | bool sve_in_use; | |
120 | }; | |
121 | ||
122 | static DEFINE_PER_CPU(struct fpsimd_last_state_struct, fpsimd_last_state); | |
005f78cd | 123 | |
79ab047c | 124 | /* Default VL for tasks that don't set it explicitly: */ |
2e0f2478 | 125 | static int sve_default_vl = -1; |
79ab047c | 126 | |
7582e220 DM |
127 | #ifdef CONFIG_ARM64_SVE |
128 | ||
129 | /* Maximum supported vector length across all CPUs (initially poisoned) */ | |
130 | int __ro_after_init sve_max_vl = -1; | |
131 | /* Set of available vector lengths, as vq_to_bit(vq): */ | |
2e0f2478 | 132 | static __ro_after_init DECLARE_BITMAP(sve_vq_map, SVE_VQ_MAX); |
fdfa976c | 133 | static void __percpu *efi_sve_state; |
7582e220 DM |
134 | |
135 | #else /* ! CONFIG_ARM64_SVE */ | |
136 | ||
137 | /* Dummy declaration for code that will be optimised out: */ | |
2e0f2478 | 138 | extern __ro_after_init DECLARE_BITMAP(sve_vq_map, SVE_VQ_MAX); |
fdfa976c | 139 | extern void __percpu *efi_sve_state; |
7582e220 DM |
140 | |
141 | #endif /* ! CONFIG_ARM64_SVE */ | |
142 | ||
bc0ee476 DM |
143 | /* |
144 | * Call __sve_free() directly only if you know task can't be scheduled | |
145 | * or preempted. | |
146 | */ | |
147 | static void __sve_free(struct task_struct *task) | |
148 | { | |
149 | kfree(task->thread.sve_state); | |
150 | task->thread.sve_state = NULL; | |
151 | } | |
152 | ||
153 | static void sve_free(struct task_struct *task) | |
154 | { | |
155 | WARN_ON(test_tsk_thread_flag(task, TIF_SVE)); | |
156 | ||
157 | __sve_free(task); | |
158 | } | |
159 | ||
160 | ||
161 | /* Offset of FFR in the SVE register dump */ | |
162 | static size_t sve_ffr_offset(int vl) | |
163 | { | |
164 | return SVE_SIG_FFR_OFFSET(sve_vq_from_vl(vl)) - SVE_SIG_REGS_OFFSET; | |
165 | } | |
166 | ||
167 | static void *sve_pffr(struct task_struct *task) | |
168 | { | |
169 | return (char *)task->thread.sve_state + | |
170 | sve_ffr_offset(task->thread.sve_vl); | |
171 | } | |
172 | ||
173 | static void change_cpacr(u64 val, u64 mask) | |
174 | { | |
175 | u64 cpacr = read_sysreg(CPACR_EL1); | |
176 | u64 new = (cpacr & ~mask) | val; | |
177 | ||
178 | if (new != cpacr) | |
179 | write_sysreg(new, CPACR_EL1); | |
180 | } | |
181 | ||
182 | static void sve_user_disable(void) | |
183 | { | |
184 | change_cpacr(0, CPACR_EL1_ZEN_EL0EN); | |
185 | } | |
186 | ||
187 | static void sve_user_enable(void) | |
188 | { | |
189 | change_cpacr(CPACR_EL1_ZEN_EL0EN, CPACR_EL1_ZEN_EL0EN); | |
190 | } | |
191 | ||
192 | /* | |
193 | * TIF_SVE controls whether a task can use SVE without trapping while | |
194 | * in userspace, and also the way a task's FPSIMD/SVE state is stored | |
195 | * in thread_struct. | |
196 | * | |
197 | * The kernel uses this flag to track whether a user task is actively | |
198 | * using SVE, and therefore whether full SVE register state needs to | |
199 | * be tracked. If not, the cheaper FPSIMD context handling code can | |
200 | * be used instead of the more costly SVE equivalents. | |
201 | * | |
202 | * * TIF_SVE set: | |
203 | * | |
204 | * The task can execute SVE instructions while in userspace without | |
205 | * trapping to the kernel. | |
206 | * | |
207 | * When stored, Z0-Z31 (incorporating Vn in bits[127:0] or the | |
208 | * corresponding Zn), P0-P15 and FFR are encoded in in | |
209 | * task->thread.sve_state, formatted appropriately for vector | |
210 | * length task->thread.sve_vl. | |
211 | * | |
212 | * task->thread.sve_state must point to a valid buffer at least | |
213 | * sve_state_size(task) bytes in size. | |
214 | * | |
215 | * During any syscall, the kernel may optionally clear TIF_SVE and | |
216 | * discard the vector state except for the FPSIMD subset. | |
217 | * | |
218 | * * TIF_SVE clear: | |
219 | * | |
220 | * An attempt by the user task to execute an SVE instruction causes | |
221 | * do_sve_acc() to be called, which does some preparation and then | |
222 | * sets TIF_SVE. | |
223 | * | |
224 | * When stored, FPSIMD registers V0-V31 are encoded in | |
225 | * task->fpsimd_state; bits [max : 128] for each of Z0-Z31 are | |
226 | * logically zero but not stored anywhere; P0-P15 and FFR are not | |
227 | * stored and have unspecified values from userspace's point of | |
228 | * view. For hygiene purposes, the kernel zeroes them on next use, | |
229 | * but userspace is discouraged from relying on this. | |
230 | * | |
231 | * task->thread.sve_state does not need to be non-NULL, valid or any | |
232 | * particular size: it must not be dereferenced. | |
233 | * | |
234 | * * FPSR and FPCR are always stored in task->fpsimd_state irrespctive of | |
235 | * whether TIF_SVE is clear or set, since these are not vector length | |
236 | * dependent. | |
237 | */ | |
238 | ||
239 | /* | |
240 | * Update current's FPSIMD/SVE registers from thread_struct. | |
241 | * | |
242 | * This function should be called only when the FPSIMD/SVE state in | |
243 | * thread_struct is known to be up to date, when preparing to enter | |
244 | * userspace. | |
245 | * | |
246 | * Softirqs (and preemption) must be disabled. | |
247 | */ | |
248 | static void task_fpsimd_load(void) | |
249 | { | |
250 | WARN_ON(!in_softirq() && !irqs_disabled()); | |
251 | ||
252 | if (system_supports_sve() && test_thread_flag(TIF_SVE)) | |
253 | sve_load_state(sve_pffr(current), | |
254 | ¤t->thread.fpsimd_state.fpsr, | |
255 | sve_vq_from_vl(current->thread.sve_vl) - 1); | |
256 | else | |
257 | fpsimd_load_state(¤t->thread.fpsimd_state); | |
258 | ||
259 | if (system_supports_sve()) { | |
260 | /* Toggle SVE trapping for userspace if needed */ | |
261 | if (test_thread_flag(TIF_SVE)) | |
262 | sve_user_enable(); | |
263 | else | |
264 | sve_user_disable(); | |
265 | ||
266 | /* Serialised by exception return to user */ | |
267 | } | |
268 | } | |
269 | ||
270 | /* | |
271 | * Ensure current's FPSIMD/SVE storage in thread_struct is up to date | |
272 | * with respect to the CPU registers. | |
273 | * | |
274 | * Softirqs (and preemption) must be disabled. | |
275 | */ | |
276 | static void task_fpsimd_save(void) | |
277 | { | |
278 | WARN_ON(!in_softirq() && !irqs_disabled()); | |
279 | ||
280 | if (!test_thread_flag(TIF_FOREIGN_FPSTATE)) { | |
281 | if (system_supports_sve() && test_thread_flag(TIF_SVE)) { | |
282 | if (WARN_ON(sve_get_vl() != current->thread.sve_vl)) { | |
283 | /* | |
284 | * Can't save the user regs, so current would | |
285 | * re-enter user with corrupt state. | |
286 | * There's no way to recover, so kill it: | |
287 | */ | |
288 | force_signal_inject( | |
289 | SIGKILL, 0, current_pt_regs(), 0); | |
290 | return; | |
291 | } | |
292 | ||
293 | sve_save_state(sve_pffr(current), | |
294 | ¤t->thread.fpsimd_state.fpsr); | |
295 | } else | |
296 | fpsimd_save_state(¤t->thread.fpsimd_state); | |
297 | } | |
298 | } | |
299 | ||
7582e220 DM |
300 | /* |
301 | * Helpers to translate bit indices in sve_vq_map to VQ values (and | |
302 | * vice versa). This allows find_next_bit() to be used to find the | |
303 | * _maximum_ VQ not exceeding a certain value. | |
304 | */ | |
305 | ||
306 | static unsigned int vq_to_bit(unsigned int vq) | |
307 | { | |
308 | return SVE_VQ_MAX - vq; | |
309 | } | |
310 | ||
311 | static unsigned int bit_to_vq(unsigned int bit) | |
312 | { | |
313 | if (WARN_ON(bit >= SVE_VQ_MAX)) | |
314 | bit = SVE_VQ_MAX - 1; | |
315 | ||
316 | return SVE_VQ_MAX - bit; | |
317 | } | |
318 | ||
319 | /* | |
320 | * All vector length selection from userspace comes through here. | |
321 | * We're on a slow path, so some sanity-checks are included. | |
322 | * If things go wrong there's a bug somewhere, but try to fall back to a | |
323 | * safe choice. | |
324 | */ | |
325 | static unsigned int find_supported_vector_length(unsigned int vl) | |
326 | { | |
327 | int bit; | |
328 | int max_vl = sve_max_vl; | |
329 | ||
330 | if (WARN_ON(!sve_vl_valid(vl))) | |
331 | vl = SVE_VL_MIN; | |
332 | ||
333 | if (WARN_ON(!sve_vl_valid(max_vl))) | |
334 | max_vl = SVE_VL_MIN; | |
335 | ||
336 | if (vl > max_vl) | |
337 | vl = max_vl; | |
338 | ||
339 | bit = find_next_bit(sve_vq_map, SVE_VQ_MAX, | |
340 | vq_to_bit(sve_vq_from_vl(vl))); | |
341 | return sve_vl_from_vq(bit_to_vq(bit)); | |
342 | } | |
343 | ||
4ffa09a9 DM |
344 | #ifdef CONFIG_SYSCTL |
345 | ||
346 | static int sve_proc_do_default_vl(struct ctl_table *table, int write, | |
347 | void __user *buffer, size_t *lenp, | |
348 | loff_t *ppos) | |
349 | { | |
350 | int ret; | |
351 | int vl = sve_default_vl; | |
352 | struct ctl_table tmp_table = { | |
353 | .data = &vl, | |
354 | .maxlen = sizeof(vl), | |
355 | }; | |
356 | ||
357 | ret = proc_dointvec(&tmp_table, write, buffer, lenp, ppos); | |
358 | if (ret || !write) | |
359 | return ret; | |
360 | ||
361 | /* Writing -1 has the special meaning "set to max": */ | |
362 | if (vl == -1) { | |
363 | /* Fail safe if sve_max_vl wasn't initialised */ | |
364 | if (WARN_ON(!sve_vl_valid(sve_max_vl))) | |
365 | vl = SVE_VL_MIN; | |
366 | else | |
367 | vl = sve_max_vl; | |
368 | ||
369 | goto chosen; | |
370 | } | |
371 | ||
372 | if (!sve_vl_valid(vl)) | |
373 | return -EINVAL; | |
374 | ||
375 | vl = find_supported_vector_length(vl); | |
376 | chosen: | |
377 | sve_default_vl = vl; | |
378 | return 0; | |
379 | } | |
380 | ||
381 | static struct ctl_table sve_default_vl_table[] = { | |
382 | { | |
383 | .procname = "sve_default_vector_length", | |
384 | .mode = 0644, | |
385 | .proc_handler = sve_proc_do_default_vl, | |
386 | }, | |
387 | { } | |
388 | }; | |
389 | ||
390 | static int __init sve_sysctl_init(void) | |
391 | { | |
392 | if (system_supports_sve()) | |
393 | if (!register_sysctl("abi", sve_default_vl_table)) | |
394 | return -EINVAL; | |
395 | ||
396 | return 0; | |
397 | } | |
398 | ||
399 | #else /* ! CONFIG_SYSCTL */ | |
400 | static int __init sve_sysctl_init(void) { return 0; } | |
401 | #endif /* ! CONFIG_SYSCTL */ | |
402 | ||
bc0ee476 DM |
403 | #define ZREG(sve_state, vq, n) ((char *)(sve_state) + \ |
404 | (SVE_SIG_ZREG_OFFSET(vq, n) - SVE_SIG_REGS_OFFSET)) | |
405 | ||
406 | /* | |
407 | * Transfer the FPSIMD state in task->thread.fpsimd_state to | |
408 | * task->thread.sve_state. | |
409 | * | |
410 | * Task can be a non-runnable task, or current. In the latter case, | |
411 | * softirqs (and preemption) must be disabled. | |
412 | * task->thread.sve_state must point to at least sve_state_size(task) | |
413 | * bytes of allocated kernel memory. | |
414 | * task->thread.fpsimd_state must be up to date before calling this function. | |
415 | */ | |
416 | static void fpsimd_to_sve(struct task_struct *task) | |
417 | { | |
418 | unsigned int vq; | |
419 | void *sst = task->thread.sve_state; | |
420 | struct fpsimd_state const *fst = &task->thread.fpsimd_state; | |
421 | unsigned int i; | |
422 | ||
423 | if (!system_supports_sve()) | |
424 | return; | |
425 | ||
426 | vq = sve_vq_from_vl(task->thread.sve_vl); | |
427 | for (i = 0; i < 32; ++i) | |
428 | memcpy(ZREG(sst, vq, i), &fst->vregs[i], | |
429 | sizeof(fst->vregs[i])); | |
430 | } | |
431 | ||
8cd969d2 DM |
432 | /* |
433 | * Transfer the SVE state in task->thread.sve_state to | |
434 | * task->thread.fpsimd_state. | |
435 | * | |
436 | * Task can be a non-runnable task, or current. In the latter case, | |
437 | * softirqs (and preemption) must be disabled. | |
438 | * task->thread.sve_state must point to at least sve_state_size(task) | |
439 | * bytes of allocated kernel memory. | |
440 | * task->thread.sve_state must be up to date before calling this function. | |
441 | */ | |
442 | static void sve_to_fpsimd(struct task_struct *task) | |
443 | { | |
444 | unsigned int vq; | |
445 | void const *sst = task->thread.sve_state; | |
446 | struct fpsimd_state *fst = &task->thread.fpsimd_state; | |
447 | unsigned int i; | |
448 | ||
449 | if (!system_supports_sve()) | |
450 | return; | |
451 | ||
452 | vq = sve_vq_from_vl(task->thread.sve_vl); | |
453 | for (i = 0; i < 32; ++i) | |
454 | memcpy(&fst->vregs[i], ZREG(sst, vq, i), | |
455 | sizeof(fst->vregs[i])); | |
456 | } | |
457 | ||
bc0ee476 DM |
458 | #ifdef CONFIG_ARM64_SVE |
459 | ||
460 | /* | |
461 | * Return how many bytes of memory are required to store the full SVE | |
462 | * state for task, given task's currently configured vector length. | |
463 | */ | |
464 | size_t sve_state_size(struct task_struct const *task) | |
465 | { | |
466 | return SVE_SIG_REGS_SIZE(sve_vq_from_vl(task->thread.sve_vl)); | |
467 | } | |
468 | ||
469 | /* | |
470 | * Ensure that task->thread.sve_state is allocated and sufficiently large. | |
471 | * | |
472 | * This function should be used only in preparation for replacing | |
473 | * task->thread.sve_state with new data. The memory is always zeroed | |
474 | * here to prevent stale data from showing through: this is done in | |
475 | * the interest of testability and predictability: except in the | |
476 | * do_sve_acc() case, there is no ABI requirement to hide stale data | |
477 | * written previously be task. | |
478 | */ | |
479 | void sve_alloc(struct task_struct *task) | |
480 | { | |
481 | if (task->thread.sve_state) { | |
482 | memset(task->thread.sve_state, 0, sve_state_size(current)); | |
483 | return; | |
484 | } | |
485 | ||
486 | /* This is a small allocation (maximum ~8KB) and Should Not Fail. */ | |
487 | task->thread.sve_state = | |
488 | kzalloc(sve_state_size(task), GFP_KERNEL); | |
489 | ||
490 | /* | |
491 | * If future SVE revisions can have larger vectors though, | |
492 | * this may cease to be true: | |
493 | */ | |
494 | BUG_ON(!task->thread.sve_state); | |
495 | } | |
496 | ||
43d4da2c DM |
497 | |
498 | /* | |
499 | * Ensure that task->thread.sve_state is up to date with respect to | |
500 | * the user task, irrespective of when SVE is in use or not. | |
501 | * | |
502 | * This should only be called by ptrace. task must be non-runnable. | |
503 | * task->thread.sve_state must point to at least sve_state_size(task) | |
504 | * bytes of allocated kernel memory. | |
505 | */ | |
506 | void fpsimd_sync_to_sve(struct task_struct *task) | |
507 | { | |
508 | if (!test_tsk_thread_flag(task, TIF_SVE)) | |
509 | fpsimd_to_sve(task); | |
510 | } | |
511 | ||
512 | /* | |
513 | * Ensure that task->thread.fpsimd_state is up to date with respect to | |
514 | * the user task, irrespective of whether SVE is in use or not. | |
515 | * | |
516 | * This should only be called by ptrace. task must be non-runnable. | |
517 | * task->thread.sve_state must point to at least sve_state_size(task) | |
518 | * bytes of allocated kernel memory. | |
519 | */ | |
520 | void sve_sync_to_fpsimd(struct task_struct *task) | |
521 | { | |
522 | if (test_tsk_thread_flag(task, TIF_SVE)) | |
523 | sve_to_fpsimd(task); | |
524 | } | |
525 | ||
526 | /* | |
527 | * Ensure that task->thread.sve_state is up to date with respect to | |
528 | * the task->thread.fpsimd_state. | |
529 | * | |
530 | * This should only be called by ptrace to merge new FPSIMD register | |
531 | * values into a task for which SVE is currently active. | |
532 | * task must be non-runnable. | |
533 | * task->thread.sve_state must point to at least sve_state_size(task) | |
534 | * bytes of allocated kernel memory. | |
535 | * task->thread.fpsimd_state must already have been initialised with | |
536 | * the new FPSIMD register values to be merged in. | |
537 | */ | |
538 | void sve_sync_from_fpsimd_zeropad(struct task_struct *task) | |
539 | { | |
540 | unsigned int vq; | |
541 | void *sst = task->thread.sve_state; | |
542 | struct fpsimd_state const *fst = &task->thread.fpsimd_state; | |
543 | unsigned int i; | |
544 | ||
545 | if (!test_tsk_thread_flag(task, TIF_SVE)) | |
546 | return; | |
547 | ||
548 | vq = sve_vq_from_vl(task->thread.sve_vl); | |
549 | ||
550 | memset(sst, 0, SVE_SIG_REGS_SIZE(vq)); | |
551 | ||
552 | for (i = 0; i < 32; ++i) | |
553 | memcpy(ZREG(sst, vq, i), &fst->vregs[i], | |
554 | sizeof(fst->vregs[i])); | |
555 | } | |
556 | ||
7582e220 DM |
557 | int sve_set_vector_length(struct task_struct *task, |
558 | unsigned long vl, unsigned long flags) | |
559 | { | |
560 | if (flags & ~(unsigned long)(PR_SVE_VL_INHERIT | | |
561 | PR_SVE_SET_VL_ONEXEC)) | |
562 | return -EINVAL; | |
563 | ||
564 | if (!sve_vl_valid(vl)) | |
565 | return -EINVAL; | |
566 | ||
567 | /* | |
568 | * Clamp to the maximum vector length that VL-agnostic SVE code can | |
569 | * work with. A flag may be assigned in the future to allow setting | |
570 | * of larger vector lengths without confusing older software. | |
571 | */ | |
572 | if (vl > SVE_VL_ARCH_MAX) | |
573 | vl = SVE_VL_ARCH_MAX; | |
574 | ||
575 | vl = find_supported_vector_length(vl); | |
576 | ||
577 | if (flags & (PR_SVE_VL_INHERIT | | |
578 | PR_SVE_SET_VL_ONEXEC)) | |
579 | task->thread.sve_vl_onexec = vl; | |
580 | else | |
581 | /* Reset VL to system default on next exec: */ | |
582 | task->thread.sve_vl_onexec = 0; | |
583 | ||
584 | /* Only actually set the VL if not deferred: */ | |
585 | if (flags & PR_SVE_SET_VL_ONEXEC) | |
586 | goto out; | |
587 | ||
588 | if (vl == task->thread.sve_vl) | |
589 | goto out; | |
590 | ||
591 | /* | |
592 | * To ensure the FPSIMD bits of the SVE vector registers are preserved, | |
593 | * write any live register state back to task_struct, and convert to a | |
594 | * non-SVE thread. | |
595 | */ | |
596 | if (task == current) { | |
597 | local_bh_disable(); | |
598 | ||
599 | task_fpsimd_save(); | |
600 | set_thread_flag(TIF_FOREIGN_FPSTATE); | |
601 | } | |
602 | ||
603 | fpsimd_flush_task_state(task); | |
604 | if (test_and_clear_tsk_thread_flag(task, TIF_SVE)) | |
605 | sve_to_fpsimd(task); | |
606 | ||
607 | if (task == current) | |
608 | local_bh_enable(); | |
609 | ||
610 | /* | |
611 | * Force reallocation of task SVE state to the correct size | |
612 | * on next use: | |
613 | */ | |
614 | sve_free(task); | |
615 | ||
616 | task->thread.sve_vl = vl; | |
617 | ||
618 | out: | |
619 | if (flags & PR_SVE_VL_INHERIT) | |
620 | set_tsk_thread_flag(task, TIF_SVE_VL_INHERIT); | |
621 | else | |
622 | clear_tsk_thread_flag(task, TIF_SVE_VL_INHERIT); | |
623 | ||
624 | return 0; | |
625 | } | |
626 | ||
2d2123bc DM |
627 | /* |
628 | * Encode the current vector length and flags for return. | |
629 | * This is only required for prctl(): ptrace has separate fields | |
630 | * | |
631 | * flags are as for sve_set_vector_length(). | |
632 | */ | |
633 | static int sve_prctl_status(unsigned long flags) | |
634 | { | |
635 | int ret; | |
636 | ||
637 | if (flags & PR_SVE_SET_VL_ONEXEC) | |
638 | ret = current->thread.sve_vl_onexec; | |
639 | else | |
640 | ret = current->thread.sve_vl; | |
641 | ||
642 | if (test_thread_flag(TIF_SVE_VL_INHERIT)) | |
643 | ret |= PR_SVE_VL_INHERIT; | |
644 | ||
645 | return ret; | |
646 | } | |
647 | ||
648 | /* PR_SVE_SET_VL */ | |
649 | int sve_set_current_vl(unsigned long arg) | |
650 | { | |
651 | unsigned long vl, flags; | |
652 | int ret; | |
653 | ||
654 | vl = arg & PR_SVE_VL_LEN_MASK; | |
655 | flags = arg & ~vl; | |
656 | ||
657 | if (!system_supports_sve()) | |
658 | return -EINVAL; | |
659 | ||
660 | ret = sve_set_vector_length(current, vl, flags); | |
661 | if (ret) | |
662 | return ret; | |
663 | ||
664 | return sve_prctl_status(flags); | |
665 | } | |
666 | ||
667 | /* PR_SVE_GET_VL */ | |
668 | int sve_get_current_vl(void) | |
669 | { | |
670 | if (!system_supports_sve()) | |
671 | return -EINVAL; | |
672 | ||
673 | return sve_prctl_status(0); | |
674 | } | |
675 | ||
2e0f2478 DM |
676 | /* |
677 | * Bitmap for temporary storage of the per-CPU set of supported vector lengths | |
678 | * during secondary boot. | |
679 | */ | |
680 | static DECLARE_BITMAP(sve_secondary_vq_map, SVE_VQ_MAX); | |
681 | ||
682 | static void sve_probe_vqs(DECLARE_BITMAP(map, SVE_VQ_MAX)) | |
683 | { | |
684 | unsigned int vq, vl; | |
685 | unsigned long zcr; | |
686 | ||
687 | bitmap_zero(map, SVE_VQ_MAX); | |
688 | ||
689 | zcr = ZCR_ELx_LEN_MASK; | |
690 | zcr = read_sysreg_s(SYS_ZCR_EL1) & ~zcr; | |
691 | ||
692 | for (vq = SVE_VQ_MAX; vq >= SVE_VQ_MIN; --vq) { | |
693 | write_sysreg_s(zcr | (vq - 1), SYS_ZCR_EL1); /* self-syncing */ | |
694 | vl = sve_get_vl(); | |
695 | vq = sve_vq_from_vl(vl); /* skip intervening lengths */ | |
696 | set_bit(vq_to_bit(vq), map); | |
697 | } | |
698 | } | |
699 | ||
700 | void __init sve_init_vq_map(void) | |
701 | { | |
702 | sve_probe_vqs(sve_vq_map); | |
703 | } | |
704 | ||
705 | /* | |
706 | * If we haven't committed to the set of supported VQs yet, filter out | |
707 | * those not supported by the current CPU. | |
708 | */ | |
709 | void sve_update_vq_map(void) | |
710 | { | |
711 | sve_probe_vqs(sve_secondary_vq_map); | |
712 | bitmap_and(sve_vq_map, sve_vq_map, sve_secondary_vq_map, SVE_VQ_MAX); | |
713 | } | |
714 | ||
715 | /* Check whether the current CPU supports all VQs in the committed set */ | |
716 | int sve_verify_vq_map(void) | |
717 | { | |
718 | int ret = 0; | |
719 | ||
720 | sve_probe_vqs(sve_secondary_vq_map); | |
721 | bitmap_andnot(sve_secondary_vq_map, sve_vq_map, sve_secondary_vq_map, | |
722 | SVE_VQ_MAX); | |
723 | if (!bitmap_empty(sve_secondary_vq_map, SVE_VQ_MAX)) { | |
724 | pr_warn("SVE: cpu%d: Required vector length(s) missing\n", | |
725 | smp_processor_id()); | |
726 | ret = -EINVAL; | |
727 | } | |
728 | ||
729 | return ret; | |
730 | } | |
731 | ||
fdfa976c DM |
732 | static void __init sve_efi_setup(void) |
733 | { | |
734 | if (!IS_ENABLED(CONFIG_EFI)) | |
735 | return; | |
736 | ||
737 | /* | |
738 | * alloc_percpu() warns and prints a backtrace if this goes wrong. | |
739 | * This is evidence of a crippled system and we are returning void, | |
740 | * so no attempt is made to handle this situation here. | |
741 | */ | |
742 | if (!sve_vl_valid(sve_max_vl)) | |
743 | goto fail; | |
744 | ||
745 | efi_sve_state = __alloc_percpu( | |
746 | SVE_SIG_REGS_SIZE(sve_vq_from_vl(sve_max_vl)), SVE_VQ_BYTES); | |
747 | if (!efi_sve_state) | |
748 | goto fail; | |
749 | ||
750 | return; | |
751 | ||
752 | fail: | |
753 | panic("Cannot allocate percpu memory for EFI SVE save/restore"); | |
754 | } | |
755 | ||
2e0f2478 DM |
756 | /* |
757 | * Enable SVE for EL1. | |
758 | * Intended for use by the cpufeatures code during CPU boot. | |
759 | */ | |
760 | int sve_kernel_enable(void *__always_unused p) | |
761 | { | |
762 | write_sysreg(read_sysreg(CPACR_EL1) | CPACR_EL1_ZEN_EL1EN, CPACR_EL1); | |
763 | isb(); | |
764 | ||
765 | return 0; | |
766 | } | |
767 | ||
768 | void __init sve_setup(void) | |
769 | { | |
770 | u64 zcr; | |
771 | ||
772 | if (!system_supports_sve()) | |
773 | return; | |
774 | ||
775 | /* | |
776 | * The SVE architecture mandates support for 128-bit vectors, | |
777 | * so sve_vq_map must have at least SVE_VQ_MIN set. | |
778 | * If something went wrong, at least try to patch it up: | |
779 | */ | |
780 | if (WARN_ON(!test_bit(vq_to_bit(SVE_VQ_MIN), sve_vq_map))) | |
781 | set_bit(vq_to_bit(SVE_VQ_MIN), sve_vq_map); | |
782 | ||
783 | zcr = read_sanitised_ftr_reg(SYS_ZCR_EL1); | |
784 | sve_max_vl = sve_vl_from_vq((zcr & ZCR_ELx_LEN_MASK) + 1); | |
785 | ||
786 | /* | |
787 | * Sanity-check that the max VL we determined through CPU features | |
788 | * corresponds properly to sve_vq_map. If not, do our best: | |
789 | */ | |
790 | if (WARN_ON(sve_max_vl != find_supported_vector_length(sve_max_vl))) | |
791 | sve_max_vl = find_supported_vector_length(sve_max_vl); | |
792 | ||
793 | /* | |
794 | * For the default VL, pick the maximum supported value <= 64. | |
795 | * VL == 64 is guaranteed not to grow the signal frame. | |
796 | */ | |
797 | sve_default_vl = find_supported_vector_length(64); | |
798 | ||
799 | pr_info("SVE: maximum available vector length %u bytes per vector\n", | |
800 | sve_max_vl); | |
801 | pr_info("SVE: default vector length %u bytes per vector\n", | |
802 | sve_default_vl); | |
fdfa976c DM |
803 | |
804 | sve_efi_setup(); | |
2e0f2478 DM |
805 | } |
806 | ||
bc0ee476 DM |
807 | /* |
808 | * Called from the put_task_struct() path, which cannot get here | |
809 | * unless dead_task is really dead and not schedulable. | |
810 | */ | |
811 | void fpsimd_release_task(struct task_struct *dead_task) | |
812 | { | |
813 | __sve_free(dead_task); | |
814 | } | |
815 | ||
816 | #endif /* CONFIG_ARM64_SVE */ | |
817 | ||
818 | /* | |
819 | * Trapped SVE access | |
820 | * | |
821 | * Storage is allocated for the full SVE state, the current FPSIMD | |
822 | * register contents are migrated across, and TIF_SVE is set so that | |
823 | * the SVE access trap will be disabled the next time this task | |
824 | * reaches ret_to_user. | |
825 | * | |
826 | * TIF_SVE should be clear on entry: otherwise, task_fpsimd_load() | |
827 | * would have disabled the SVE access trap for userspace during | |
828 | * ret_to_user, making an SVE access trap impossible in that case. | |
829 | */ | |
830 | asmlinkage void do_sve_acc(unsigned int esr, struct pt_regs *regs) | |
831 | { | |
832 | /* Even if we chose not to use SVE, the hardware could still trap: */ | |
833 | if (unlikely(!system_supports_sve()) || WARN_ON(is_compat_task())) { | |
834 | force_signal_inject(SIGILL, ILL_ILLOPC, regs, 0); | |
835 | return; | |
836 | } | |
837 | ||
838 | sve_alloc(current); | |
839 | ||
840 | local_bh_disable(); | |
841 | ||
842 | task_fpsimd_save(); | |
843 | fpsimd_to_sve(current); | |
844 | ||
845 | /* Force ret_to_user to reload the registers: */ | |
846 | fpsimd_flush_task_state(current); | |
847 | set_thread_flag(TIF_FOREIGN_FPSTATE); | |
848 | ||
849 | if (test_and_set_thread_flag(TIF_SVE)) | |
850 | WARN_ON(1); /* SVE access shouldn't have trapped */ | |
851 | ||
852 | local_bh_enable(); | |
853 | } | |
854 | ||
53631b54 CM |
855 | /* |
856 | * Trapped FP/ASIMD access. | |
857 | */ | |
94ef7ecb | 858 | asmlinkage void do_fpsimd_acc(unsigned int esr, struct pt_regs *regs) |
53631b54 CM |
859 | { |
860 | /* TODO: implement lazy context saving/restoring */ | |
861 | WARN_ON(1); | |
862 | } | |
863 | ||
864 | /* | |
865 | * Raise a SIGFPE for the current process. | |
866 | */ | |
94ef7ecb | 867 | asmlinkage void do_fpsimd_exc(unsigned int esr, struct pt_regs *regs) |
53631b54 CM |
868 | { |
869 | siginfo_t info; | |
526c3ddb | 870 | unsigned int si_code = FPE_FIXME; |
53631b54 CM |
871 | |
872 | if (esr & FPEXC_IOF) | |
873 | si_code = FPE_FLTINV; | |
874 | else if (esr & FPEXC_DZF) | |
875 | si_code = FPE_FLTDIV; | |
876 | else if (esr & FPEXC_OFF) | |
877 | si_code = FPE_FLTOVF; | |
878 | else if (esr & FPEXC_UFF) | |
879 | si_code = FPE_FLTUND; | |
880 | else if (esr & FPEXC_IXF) | |
881 | si_code = FPE_FLTRES; | |
882 | ||
883 | memset(&info, 0, sizeof(info)); | |
884 | info.si_signo = SIGFPE; | |
885 | info.si_code = si_code; | |
886 | info.si_addr = (void __user *)instruction_pointer(regs); | |
887 | ||
888 | send_sig_info(SIGFPE, &info, current); | |
889 | } | |
890 | ||
891 | void fpsimd_thread_switch(struct task_struct *next) | |
892 | { | |
82e0191a SP |
893 | if (!system_supports_fpsimd()) |
894 | return; | |
005f78cd AB |
895 | /* |
896 | * Save the current FPSIMD state to memory, but only if whatever is in | |
897 | * the registers is in fact the most recent userland FPSIMD state of | |
898 | * 'current'. | |
899 | */ | |
bc0ee476 DM |
900 | if (current->mm) |
901 | task_fpsimd_save(); | |
005f78cd AB |
902 | |
903 | if (next->mm) { | |
904 | /* | |
905 | * If we are switching to a task whose most recent userland | |
906 | * FPSIMD state is already in the registers of *this* cpu, | |
907 | * we can skip loading the state from memory. Otherwise, set | |
908 | * the TIF_FOREIGN_FPSTATE flag so the state will be loaded | |
909 | * upon the next return to userland. | |
910 | */ | |
911 | struct fpsimd_state *st = &next->thread.fpsimd_state; | |
912 | ||
cb968afc | 913 | if (__this_cpu_read(fpsimd_last_state.st) == st |
005f78cd | 914 | && st->cpu == smp_processor_id()) |
9cf5b54f | 915 | clear_tsk_thread_flag(next, TIF_FOREIGN_FPSTATE); |
005f78cd | 916 | else |
9cf5b54f | 917 | set_tsk_thread_flag(next, TIF_FOREIGN_FPSTATE); |
005f78cd | 918 | } |
53631b54 CM |
919 | } |
920 | ||
921 | void fpsimd_flush_thread(void) | |
922 | { | |
7582e220 | 923 | int vl, supported_vl; |
bc0ee476 | 924 | |
82e0191a SP |
925 | if (!system_supports_fpsimd()) |
926 | return; | |
cb84d11e DM |
927 | |
928 | local_bh_disable(); | |
929 | ||
53631b54 | 930 | memset(¤t->thread.fpsimd_state, 0, sizeof(struct fpsimd_state)); |
674c242c | 931 | fpsimd_flush_task_state(current); |
bc0ee476 DM |
932 | |
933 | if (system_supports_sve()) { | |
934 | clear_thread_flag(TIF_SVE); | |
935 | sve_free(current); | |
936 | ||
937 | /* | |
938 | * Reset the task vector length as required. | |
939 | * This is where we ensure that all user tasks have a valid | |
940 | * vector length configured: no kernel task can become a user | |
941 | * task without an exec and hence a call to this function. | |
2e0f2478 DM |
942 | * By the time the first call to this function is made, all |
943 | * early hardware probing is complete, so sve_default_vl | |
944 | * should be valid. | |
bc0ee476 DM |
945 | * If a bug causes this to go wrong, we make some noise and |
946 | * try to fudge thread.sve_vl to a safe value here. | |
947 | */ | |
79ab047c DM |
948 | vl = current->thread.sve_vl_onexec ? |
949 | current->thread.sve_vl_onexec : sve_default_vl; | |
bc0ee476 DM |
950 | |
951 | if (WARN_ON(!sve_vl_valid(vl))) | |
952 | vl = SVE_VL_MIN; | |
953 | ||
7582e220 DM |
954 | supported_vl = find_supported_vector_length(vl); |
955 | if (WARN_ON(supported_vl != vl)) | |
956 | vl = supported_vl; | |
957 | ||
bc0ee476 | 958 | current->thread.sve_vl = vl; |
79ab047c DM |
959 | |
960 | /* | |
961 | * If the task is not set to inherit, ensure that the vector | |
962 | * length will be reset by a subsequent exec: | |
963 | */ | |
964 | if (!test_thread_flag(TIF_SVE_VL_INHERIT)) | |
965 | current->thread.sve_vl_onexec = 0; | |
bc0ee476 DM |
966 | } |
967 | ||
005f78cd | 968 | set_thread_flag(TIF_FOREIGN_FPSTATE); |
cb84d11e DM |
969 | |
970 | local_bh_enable(); | |
53631b54 CM |
971 | } |
972 | ||
c51f9269 | 973 | /* |
005f78cd AB |
974 | * Save the userland FPSIMD state of 'current' to memory, but only if the state |
975 | * currently held in the registers does in fact belong to 'current' | |
c51f9269 AB |
976 | */ |
977 | void fpsimd_preserve_current_state(void) | |
978 | { | |
82e0191a SP |
979 | if (!system_supports_fpsimd()) |
980 | return; | |
cb84d11e DM |
981 | |
982 | local_bh_disable(); | |
8cd969d2 | 983 | task_fpsimd_save(); |
cb84d11e | 984 | local_bh_enable(); |
c51f9269 AB |
985 | } |
986 | ||
8cd969d2 DM |
987 | /* |
988 | * Like fpsimd_preserve_current_state(), but ensure that | |
989 | * current->thread.fpsimd_state is updated so that it can be copied to | |
990 | * the signal frame. | |
991 | */ | |
992 | void fpsimd_signal_preserve_current_state(void) | |
993 | { | |
994 | fpsimd_preserve_current_state(); | |
995 | if (system_supports_sve() && test_thread_flag(TIF_SVE)) | |
996 | sve_to_fpsimd(current); | |
997 | } | |
998 | ||
8884b7bd DM |
999 | /* |
1000 | * Associate current's FPSIMD context with this cpu | |
1001 | * Preemption must be disabled when calling this function. | |
1002 | */ | |
1003 | static void fpsimd_bind_to_cpu(void) | |
1004 | { | |
cb968afc DM |
1005 | struct fpsimd_last_state_struct *last = |
1006 | this_cpu_ptr(&fpsimd_last_state); | |
8884b7bd DM |
1007 | struct fpsimd_state *st = ¤t->thread.fpsimd_state; |
1008 | ||
cb968afc DM |
1009 | last->st = st; |
1010 | last->sve_in_use = test_thread_flag(TIF_SVE); | |
8884b7bd DM |
1011 | st->cpu = smp_processor_id(); |
1012 | } | |
1013 | ||
c51f9269 | 1014 | /* |
005f78cd AB |
1015 | * Load the userland FPSIMD state of 'current' from memory, but only if the |
1016 | * FPSIMD state already held in the registers is /not/ the most recent FPSIMD | |
1017 | * state of 'current' | |
1018 | */ | |
1019 | void fpsimd_restore_current_state(void) | |
1020 | { | |
82e0191a SP |
1021 | if (!system_supports_fpsimd()) |
1022 | return; | |
cb84d11e DM |
1023 | |
1024 | local_bh_disable(); | |
1025 | ||
005f78cd | 1026 | if (test_and_clear_thread_flag(TIF_FOREIGN_FPSTATE)) { |
bc0ee476 | 1027 | task_fpsimd_load(); |
8884b7bd | 1028 | fpsimd_bind_to_cpu(); |
005f78cd | 1029 | } |
cb84d11e DM |
1030 | |
1031 | local_bh_enable(); | |
005f78cd AB |
1032 | } |
1033 | ||
1034 | /* | |
1035 | * Load an updated userland FPSIMD state for 'current' from memory and set the | |
1036 | * flag that indicates that the FPSIMD register contents are the most recent | |
1037 | * FPSIMD state of 'current' | |
c51f9269 | 1038 | */ |
0abdeff5 | 1039 | void fpsimd_update_current_state(struct user_fpsimd_state const *state) |
c51f9269 | 1040 | { |
82e0191a SP |
1041 | if (!system_supports_fpsimd()) |
1042 | return; | |
cb84d11e DM |
1043 | |
1044 | local_bh_disable(); | |
1045 | ||
0abdeff5 | 1046 | current->thread.fpsimd_state.user_fpsimd = *state; |
9de52a75 | 1047 | if (system_supports_sve() && test_thread_flag(TIF_SVE)) |
8cd969d2 | 1048 | fpsimd_to_sve(current); |
9de52a75 | 1049 | |
8cd969d2 DM |
1050 | task_fpsimd_load(); |
1051 | ||
8884b7bd DM |
1052 | if (test_and_clear_thread_flag(TIF_FOREIGN_FPSTATE)) |
1053 | fpsimd_bind_to_cpu(); | |
cb84d11e DM |
1054 | |
1055 | local_bh_enable(); | |
c51f9269 AB |
1056 | } |
1057 | ||
005f78cd AB |
1058 | /* |
1059 | * Invalidate live CPU copies of task t's FPSIMD state | |
1060 | */ | |
1061 | void fpsimd_flush_task_state(struct task_struct *t) | |
1062 | { | |
1063 | t->thread.fpsimd_state.cpu = NR_CPUS; | |
1064 | } | |
1065 | ||
17eed27b DM |
1066 | static inline void fpsimd_flush_cpu_state(void) |
1067 | { | |
cb968afc | 1068 | __this_cpu_write(fpsimd_last_state.st, NULL); |
17eed27b DM |
1069 | } |
1070 | ||
1071 | /* | |
1072 | * Invalidate any task SVE state currently held in this CPU's regs. | |
1073 | * | |
1074 | * This is used to prevent the kernel from trying to reuse SVE register data | |
1075 | * that is detroyed by KVM guest enter/exit. This function should go away when | |
1076 | * KVM SVE support is implemented. Don't use it for anything else. | |
1077 | */ | |
1078 | #ifdef CONFIG_ARM64_SVE | |
1079 | void sve_flush_cpu_state(void) | |
1080 | { | |
cb968afc DM |
1081 | struct fpsimd_last_state_struct const *last = |
1082 | this_cpu_ptr(&fpsimd_last_state); | |
17eed27b | 1083 | |
cb968afc | 1084 | if (last->st && last->sve_in_use) |
17eed27b DM |
1085 | fpsimd_flush_cpu_state(); |
1086 | } | |
1087 | #endif /* CONFIG_ARM64_SVE */ | |
1088 | ||
4cfb3613 AB |
1089 | #ifdef CONFIG_KERNEL_MODE_NEON |
1090 | ||
cb84d11e | 1091 | DEFINE_PER_CPU(bool, kernel_neon_busy); |
11cefd5a | 1092 | EXPORT_PER_CPU_SYMBOL(kernel_neon_busy); |
190f1ca8 | 1093 | |
4cfb3613 AB |
1094 | /* |
1095 | * Kernel-side NEON support functions | |
1096 | */ | |
cb84d11e DM |
1097 | |
1098 | /* | |
1099 | * kernel_neon_begin(): obtain the CPU FPSIMD registers for use by the calling | |
1100 | * context | |
1101 | * | |
1102 | * Must not be called unless may_use_simd() returns true. | |
1103 | * Task context in the FPSIMD registers is saved back to memory as necessary. | |
1104 | * | |
1105 | * A matching call to kernel_neon_end() must be made before returning from the | |
1106 | * calling context. | |
1107 | * | |
1108 | * The caller may freely use the FPSIMD registers until kernel_neon_end() is | |
1109 | * called. | |
1110 | */ | |
1111 | void kernel_neon_begin(void) | |
4cfb3613 | 1112 | { |
82e0191a SP |
1113 | if (WARN_ON(!system_supports_fpsimd())) |
1114 | return; | |
4cfb3613 | 1115 | |
cb84d11e DM |
1116 | BUG_ON(!may_use_simd()); |
1117 | ||
1118 | local_bh_disable(); | |
1119 | ||
1120 | __this_cpu_write(kernel_neon_busy, true); | |
1121 | ||
1122 | /* Save unsaved task fpsimd state, if any: */ | |
1bd3f936 DM |
1123 | if (current->mm) { |
1124 | task_fpsimd_save(); | |
1125 | set_thread_flag(TIF_FOREIGN_FPSTATE); | |
1126 | } | |
cb84d11e DM |
1127 | |
1128 | /* Invalidate any task state remaining in the fpsimd regs: */ | |
17eed27b | 1129 | fpsimd_flush_cpu_state(); |
cb84d11e DM |
1130 | |
1131 | preempt_disable(); | |
1132 | ||
1133 | local_bh_enable(); | |
4cfb3613 | 1134 | } |
cb84d11e | 1135 | EXPORT_SYMBOL(kernel_neon_begin); |
4cfb3613 | 1136 | |
cb84d11e DM |
1137 | /* |
1138 | * kernel_neon_end(): give the CPU FPSIMD registers back to the current task | |
1139 | * | |
1140 | * Must be called from a context in which kernel_neon_begin() was previously | |
1141 | * called, with no call to kernel_neon_end() in the meantime. | |
1142 | * | |
1143 | * The caller must not use the FPSIMD registers after this function is called, | |
1144 | * unless kernel_neon_begin() is called again in the meantime. | |
1145 | */ | |
4cfb3613 AB |
1146 | void kernel_neon_end(void) |
1147 | { | |
cb84d11e DM |
1148 | bool busy; |
1149 | ||
82e0191a SP |
1150 | if (!system_supports_fpsimd()) |
1151 | return; | |
cb84d11e DM |
1152 | |
1153 | busy = __this_cpu_xchg(kernel_neon_busy, false); | |
1154 | WARN_ON(!busy); /* No matching kernel_neon_begin()? */ | |
1155 | ||
1156 | preempt_enable(); | |
4cfb3613 AB |
1157 | } |
1158 | EXPORT_SYMBOL(kernel_neon_end); | |
1159 | ||
e580b8bc DM |
1160 | #ifdef CONFIG_EFI |
1161 | ||
3b66023d DM |
1162 | static DEFINE_PER_CPU(struct fpsimd_state, efi_fpsimd_state); |
1163 | static DEFINE_PER_CPU(bool, efi_fpsimd_state_used); | |
fdfa976c | 1164 | static DEFINE_PER_CPU(bool, efi_sve_state_used); |
4328825d DM |
1165 | |
1166 | /* | |
1167 | * EFI runtime services support functions | |
1168 | * | |
1169 | * The ABI for EFI runtime services allows EFI to use FPSIMD during the call. | |
1170 | * This means that for EFI (and only for EFI), we have to assume that FPSIMD | |
1171 | * is always used rather than being an optional accelerator. | |
1172 | * | |
1173 | * These functions provide the necessary support for ensuring FPSIMD | |
1174 | * save/restore in the contexts from which EFI is used. | |
1175 | * | |
1176 | * Do not use them for any other purpose -- if tempted to do so, you are | |
1177 | * either doing something wrong or you need to propose some refactoring. | |
1178 | */ | |
1179 | ||
1180 | /* | |
1181 | * __efi_fpsimd_begin(): prepare FPSIMD for making an EFI runtime services call | |
1182 | */ | |
1183 | void __efi_fpsimd_begin(void) | |
1184 | { | |
1185 | if (!system_supports_fpsimd()) | |
1186 | return; | |
1187 | ||
1188 | WARN_ON(preemptible()); | |
1189 | ||
fdfa976c | 1190 | if (may_use_simd()) { |
4328825d | 1191 | kernel_neon_begin(); |
fdfa976c DM |
1192 | } else { |
1193 | /* | |
1194 | * If !efi_sve_state, SVE can't be in use yet and doesn't need | |
1195 | * preserving: | |
1196 | */ | |
1197 | if (system_supports_sve() && likely(efi_sve_state)) { | |
1198 | char *sve_state = this_cpu_ptr(efi_sve_state); | |
1199 | ||
1200 | __this_cpu_write(efi_sve_state_used, true); | |
1201 | ||
1202 | sve_save_state(sve_state + sve_ffr_offset(sve_max_vl), | |
1203 | &this_cpu_ptr(&efi_fpsimd_state)->fpsr); | |
1204 | } else { | |
1205 | fpsimd_save_state(this_cpu_ptr(&efi_fpsimd_state)); | |
1206 | } | |
1207 | ||
4328825d DM |
1208 | __this_cpu_write(efi_fpsimd_state_used, true); |
1209 | } | |
1210 | } | |
1211 | ||
1212 | /* | |
1213 | * __efi_fpsimd_end(): clean up FPSIMD after an EFI runtime services call | |
1214 | */ | |
1215 | void __efi_fpsimd_end(void) | |
1216 | { | |
1217 | if (!system_supports_fpsimd()) | |
1218 | return; | |
1219 | ||
fdfa976c | 1220 | if (!__this_cpu_xchg(efi_fpsimd_state_used, false)) { |
4328825d | 1221 | kernel_neon_end(); |
fdfa976c DM |
1222 | } else { |
1223 | if (system_supports_sve() && | |
1224 | likely(__this_cpu_read(efi_sve_state_used))) { | |
1225 | char const *sve_state = this_cpu_ptr(efi_sve_state); | |
1226 | ||
1227 | sve_load_state(sve_state + sve_ffr_offset(sve_max_vl), | |
1228 | &this_cpu_ptr(&efi_fpsimd_state)->fpsr, | |
1229 | sve_vq_from_vl(sve_get_vl()) - 1); | |
1230 | ||
1231 | __this_cpu_write(efi_sve_state_used, false); | |
1232 | } else { | |
1233 | fpsimd_load_state(this_cpu_ptr(&efi_fpsimd_state)); | |
1234 | } | |
1235 | } | |
4328825d DM |
1236 | } |
1237 | ||
e580b8bc DM |
1238 | #endif /* CONFIG_EFI */ |
1239 | ||
4cfb3613 AB |
1240 | #endif /* CONFIG_KERNEL_MODE_NEON */ |
1241 | ||
fb1ab1ab LP |
1242 | #ifdef CONFIG_CPU_PM |
1243 | static int fpsimd_cpu_pm_notifier(struct notifier_block *self, | |
1244 | unsigned long cmd, void *v) | |
1245 | { | |
1246 | switch (cmd) { | |
1247 | case CPU_PM_ENTER: | |
bc0ee476 DM |
1248 | if (current->mm) |
1249 | task_fpsimd_save(); | |
17eed27b | 1250 | fpsimd_flush_cpu_state(); |
fb1ab1ab LP |
1251 | break; |
1252 | case CPU_PM_EXIT: | |
1253 | if (current->mm) | |
005f78cd | 1254 | set_thread_flag(TIF_FOREIGN_FPSTATE); |
fb1ab1ab LP |
1255 | break; |
1256 | case CPU_PM_ENTER_FAILED: | |
1257 | default: | |
1258 | return NOTIFY_DONE; | |
1259 | } | |
1260 | return NOTIFY_OK; | |
1261 | } | |
1262 | ||
1263 | static struct notifier_block fpsimd_cpu_pm_notifier_block = { | |
1264 | .notifier_call = fpsimd_cpu_pm_notifier, | |
1265 | }; | |
1266 | ||
a7c61a34 | 1267 | static void __init fpsimd_pm_init(void) |
fb1ab1ab LP |
1268 | { |
1269 | cpu_pm_register_notifier(&fpsimd_cpu_pm_notifier_block); | |
1270 | } | |
1271 | ||
1272 | #else | |
1273 | static inline void fpsimd_pm_init(void) { } | |
1274 | #endif /* CONFIG_CPU_PM */ | |
1275 | ||
32365e64 | 1276 | #ifdef CONFIG_HOTPLUG_CPU |
c23a7266 | 1277 | static int fpsimd_cpu_dead(unsigned int cpu) |
32365e64 | 1278 | { |
cb968afc | 1279 | per_cpu(fpsimd_last_state.st, cpu) = NULL; |
c23a7266 | 1280 | return 0; |
32365e64 JL |
1281 | } |
1282 | ||
32365e64 JL |
1283 | static inline void fpsimd_hotplug_init(void) |
1284 | { | |
c23a7266 SAS |
1285 | cpuhp_setup_state_nocalls(CPUHP_ARM64_FPSIMD_DEAD, "arm64/fpsimd:dead", |
1286 | NULL, fpsimd_cpu_dead); | |
32365e64 JL |
1287 | } |
1288 | ||
1289 | #else | |
1290 | static inline void fpsimd_hotplug_init(void) { } | |
1291 | #endif | |
1292 | ||
53631b54 CM |
1293 | /* |
1294 | * FP/SIMD support code initialisation. | |
1295 | */ | |
1296 | static int __init fpsimd_init(void) | |
1297 | { | |
fe80f9f2 SP |
1298 | if (elf_hwcap & HWCAP_FP) { |
1299 | fpsimd_pm_init(); | |
1300 | fpsimd_hotplug_init(); | |
1301 | } else { | |
53631b54 | 1302 | pr_notice("Floating-point is not implemented\n"); |
53631b54 | 1303 | } |
53631b54 | 1304 | |
fe80f9f2 | 1305 | if (!(elf_hwcap & HWCAP_ASIMD)) |
53631b54 | 1306 | pr_notice("Advanced SIMD is not implemented\n"); |
fb1ab1ab | 1307 | |
4ffa09a9 | 1308 | return sve_sysctl_init(); |
53631b54 | 1309 | } |
ae2e972d | 1310 | core_initcall(fpsimd_init); |