• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  Copyright (C) 1994 Linus Torvalds
4  *
5  *  Pentium III FXSR, SSE support
6  *  General FPU state handling cleanups
7  *	Gareth Hughes <gareth@valinux.com>, May 2000
8  */
9 #include <asm/fpu/internal.h>
10 #include <asm/fpu/regset.h>
11 #include <asm/fpu/signal.h>
12 #include <asm/fpu/types.h>
13 #include <asm/traps.h>
14 #include <asm/irq_regs.h>
15 
16 #include <linux/hardirq.h>
17 #include <linux/pkeys.h>
18 
19 #define CREATE_TRACE_POINTS
20 #include <asm/trace/fpu.h>
21 
22 /*
23  * Represents the initial FPU state. It's mostly (but not completely) zeroes,
24  * depending on the FPU hardware format:
25  */
26 union fpregs_state init_fpstate __read_mostly;
27 
28 /* Track in-kernel FPU usage */
29 static DEFINE_PER_CPU(bool, in_kernel_fpu);
30 
31 /*
32  * Track which context is using the FPU on the CPU:
33  */
34 DEFINE_PER_CPU(struct fpu *, fpu_fpregs_owner_ctx);
35 
36 /*
37  * Can we use the FPU in kernel mode with the
38  * whole "kernel_fpu_begin/end()" sequence?
39  */
irq_fpu_usable(void)40 bool irq_fpu_usable(void)
41 {
42 	if (WARN_ON_ONCE(in_nmi()))
43 		return false;
44 
45 	/* In kernel FPU usage already active? */
46 	if (this_cpu_read(in_kernel_fpu))
47 		return false;
48 
49 	/*
50 	 * When not in NMI or hard interrupt context, FPU can be used in:
51 	 *
52 	 * - Task context except from within fpregs_lock()'ed critical
53 	 *   regions.
54 	 *
55 	 * - Soft interrupt processing context which cannot happen
56 	 *   while in a fpregs_lock()'ed critical region.
57 	 */
58 	if (!in_irq())
59 		return true;
60 
61 	/*
62 	 * In hard interrupt context it's safe when soft interrupts
63 	 * are enabled, which means the interrupt did not hit in
64 	 * a fpregs_lock()'ed critical region.
65 	 */
66 	return !softirq_count();
67 }
68 EXPORT_SYMBOL(irq_fpu_usable);
69 
70 /*
71  * These must be called with preempt disabled. Returns
72  * 'true' if the FPU state is still intact and we can
73  * keep registers active.
74  *
75  * The legacy FNSAVE instruction cleared all FPU state
76  * unconditionally, so registers are essentially destroyed.
77  * Modern FPU state can be kept in registers, if there are
78  * no pending FP exceptions.
79  */
copy_fpregs_to_fpstate(struct fpu * fpu)80 int copy_fpregs_to_fpstate(struct fpu *fpu)
81 {
82 	if (likely(use_xsave())) {
83 		copy_xregs_to_kernel(&fpu->state.xsave);
84 
85 		/*
86 		 * AVX512 state is tracked here because its use is
87 		 * known to slow the max clock speed of the core.
88 		 */
89 		if (fpu->state.xsave.header.xfeatures & XFEATURE_MASK_AVX512)
90 			fpu->avx512_timestamp = jiffies;
91 		return 1;
92 	}
93 
94 	if (likely(use_fxsr())) {
95 		copy_fxregs_to_kernel(fpu);
96 		return 1;
97 	}
98 
99 	/*
100 	 * Legacy FPU register saving, FNSAVE always clears FPU registers,
101 	 * so we have to mark them inactive:
102 	 */
103 	asm volatile("fnsave %[fp]; fwait" : [fp] "=m" (fpu->state.fsave));
104 
105 	return 0;
106 }
107 EXPORT_SYMBOL(copy_fpregs_to_fpstate);
108 
kernel_fpu_begin_mask(unsigned int kfpu_mask)109 void kernel_fpu_begin_mask(unsigned int kfpu_mask)
110 {
111 	preempt_disable();
112 
113 	WARN_ON_FPU(!irq_fpu_usable());
114 	WARN_ON_FPU(this_cpu_read(in_kernel_fpu));
115 
116 	this_cpu_write(in_kernel_fpu, true);
117 
118 	if (!(current->flags & PF_KTHREAD) &&
119 	    !test_thread_flag(TIF_NEED_FPU_LOAD)) {
120 		set_thread_flag(TIF_NEED_FPU_LOAD);
121 		/*
122 		 * Ignore return value -- we don't care if reg state
123 		 * is clobbered.
124 		 */
125 		copy_fpregs_to_fpstate(&current->thread.fpu);
126 	}
127 	__cpu_invalidate_fpregs_state();
128 
129 	/* Put sane initial values into the control registers. */
130 	if (likely(kfpu_mask & KFPU_MXCSR) && boot_cpu_has(X86_FEATURE_XMM))
131 		ldmxcsr(MXCSR_DEFAULT);
132 
133 	if (unlikely(kfpu_mask & KFPU_387) && boot_cpu_has(X86_FEATURE_FPU))
134 		asm volatile ("fninit");
135 }
136 EXPORT_SYMBOL_GPL(kernel_fpu_begin_mask);
137 
kernel_fpu_end(void)138 void kernel_fpu_end(void)
139 {
140 	WARN_ON_FPU(!this_cpu_read(in_kernel_fpu));
141 
142 	this_cpu_write(in_kernel_fpu, false);
143 	preempt_enable();
144 }
145 EXPORT_SYMBOL_GPL(kernel_fpu_end);
146 
147 /*
148  * Save the FPU state (mark it for reload if necessary):
149  *
150  * This only ever gets called for the current task.
151  */
fpu__save(struct fpu * fpu)152 void fpu__save(struct fpu *fpu)
153 {
154 	WARN_ON_FPU(fpu != &current->thread.fpu);
155 
156 	fpregs_lock();
157 	trace_x86_fpu_before_save(fpu);
158 
159 	if (!test_thread_flag(TIF_NEED_FPU_LOAD)) {
160 		if (!copy_fpregs_to_fpstate(fpu)) {
161 			copy_kernel_to_fpregs(&fpu->state);
162 		}
163 	}
164 
165 	trace_x86_fpu_after_save(fpu);
166 	fpregs_unlock();
167 }
168 
169 /*
170  * Legacy x87 fpstate state init:
171  */
fpstate_init_fstate(struct fregs_state * fp)172 static inline void fpstate_init_fstate(struct fregs_state *fp)
173 {
174 	fp->cwd = 0xffff037fu;
175 	fp->swd = 0xffff0000u;
176 	fp->twd = 0xffffffffu;
177 	fp->fos = 0xffff0000u;
178 }
179 
fpstate_init(union fpregs_state * state)180 void fpstate_init(union fpregs_state *state)
181 {
182 	if (!static_cpu_has(X86_FEATURE_FPU)) {
183 		fpstate_init_soft(&state->soft);
184 		return;
185 	}
186 
187 	memset(state, 0, fpu_kernel_xstate_size);
188 
189 	if (static_cpu_has(X86_FEATURE_XSAVES))
190 		fpstate_init_xstate(&state->xsave);
191 	if (static_cpu_has(X86_FEATURE_FXSR))
192 		fpstate_init_fxstate(&state->fxsave);
193 	else
194 		fpstate_init_fstate(&state->fsave);
195 }
196 EXPORT_SYMBOL_GPL(fpstate_init);
197 
fpu__copy(struct task_struct * dst,struct task_struct * src)198 int fpu__copy(struct task_struct *dst, struct task_struct *src)
199 {
200 	struct fpu *dst_fpu = &dst->thread.fpu;
201 	struct fpu *src_fpu = &src->thread.fpu;
202 
203 	dst_fpu->last_cpu = -1;
204 
205 	if (!static_cpu_has(X86_FEATURE_FPU))
206 		return 0;
207 
208 	WARN_ON_FPU(src_fpu != &current->thread.fpu);
209 
210 	/*
211 	 * Don't let 'init optimized' areas of the XSAVE area
212 	 * leak into the child task:
213 	 */
214 	memset(&dst_fpu->state.xsave, 0, fpu_kernel_xstate_size);
215 
216 	/*
217 	 * If the FPU registers are not current just memcpy() the state.
218 	 * Otherwise save current FPU registers directly into the child's FPU
219 	 * context, without any memory-to-memory copying.
220 	 *
221 	 * ( The function 'fails' in the FNSAVE case, which destroys
222 	 *   register contents so we have to load them back. )
223 	 */
224 	fpregs_lock();
225 	if (test_thread_flag(TIF_NEED_FPU_LOAD))
226 		memcpy(&dst_fpu->state, &src_fpu->state, fpu_kernel_xstate_size);
227 
228 	else if (!copy_fpregs_to_fpstate(dst_fpu))
229 		copy_kernel_to_fpregs(&dst_fpu->state);
230 
231 	fpregs_unlock();
232 
233 	set_tsk_thread_flag(dst, TIF_NEED_FPU_LOAD);
234 
235 	trace_x86_fpu_copy_src(src_fpu);
236 	trace_x86_fpu_copy_dst(dst_fpu);
237 
238 	return 0;
239 }
240 
241 /*
242  * Activate the current task's in-memory FPU context,
243  * if it has not been used before:
244  */
fpu__initialize(struct fpu * fpu)245 static void fpu__initialize(struct fpu *fpu)
246 {
247 	WARN_ON_FPU(fpu != &current->thread.fpu);
248 
249 	set_thread_flag(TIF_NEED_FPU_LOAD);
250 	fpstate_init(&fpu->state);
251 	trace_x86_fpu_init_state(fpu);
252 }
253 
254 /*
255  * This function must be called before we read a task's fpstate.
256  *
257  * There's two cases where this gets called:
258  *
259  * - for the current task (when coredumping), in which case we have
260  *   to save the latest FPU registers into the fpstate,
261  *
262  * - or it's called for stopped tasks (ptrace), in which case the
263  *   registers were already saved by the context-switch code when
264  *   the task scheduled out.
265  *
266  * If the task has used the FPU before then save it.
267  */
fpu__prepare_read(struct fpu * fpu)268 void fpu__prepare_read(struct fpu *fpu)
269 {
270 	if (fpu == &current->thread.fpu)
271 		fpu__save(fpu);
272 }
273 
274 /*
275  * This function must be called before we write a task's fpstate.
276  *
277  * Invalidate any cached FPU registers.
278  *
279  * After this function call, after registers in the fpstate are
280  * modified and the child task has woken up, the child task will
281  * restore the modified FPU state from the modified context. If we
282  * didn't clear its cached status here then the cached in-registers
283  * state pending on its former CPU could be restored, corrupting
284  * the modifications.
285  */
fpu__prepare_write(struct fpu * fpu)286 void fpu__prepare_write(struct fpu *fpu)
287 {
288 	/*
289 	 * Only stopped child tasks can be used to modify the FPU
290 	 * state in the fpstate buffer:
291 	 */
292 	WARN_ON_FPU(fpu == &current->thread.fpu);
293 
294 	/* Invalidate any cached state: */
295 	__fpu_invalidate_fpregs_state(fpu);
296 }
297 
298 /*
299  * Drops current FPU state: deactivates the fpregs and
300  * the fpstate. NOTE: it still leaves previous contents
301  * in the fpregs in the eager-FPU case.
302  *
303  * This function can be used in cases where we know that
304  * a state-restore is coming: either an explicit one,
305  * or a reschedule.
306  */
fpu__drop(struct fpu * fpu)307 void fpu__drop(struct fpu *fpu)
308 {
309 	preempt_disable();
310 
311 	if (fpu == &current->thread.fpu) {
312 		/* Ignore delayed exceptions from user space */
313 		asm volatile("1: fwait\n"
314 			     "2:\n"
315 			     _ASM_EXTABLE(1b, 2b));
316 		fpregs_deactivate(fpu);
317 	}
318 
319 	trace_x86_fpu_dropped(fpu);
320 
321 	preempt_enable();
322 }
323 
324 /*
325  * Clear FPU registers by setting them up from the init fpstate.
326  * Caller must do fpregs_[un]lock() around it.
327  */
copy_init_fpstate_to_fpregs(u64 features_mask)328 static inline void copy_init_fpstate_to_fpregs(u64 features_mask)
329 {
330 	if (use_xsave())
331 		copy_kernel_to_xregs(&init_fpstate.xsave, features_mask);
332 	else if (static_cpu_has(X86_FEATURE_FXSR))
333 		copy_kernel_to_fxregs(&init_fpstate.fxsave);
334 	else
335 		copy_kernel_to_fregs(&init_fpstate.fsave);
336 
337 	if (boot_cpu_has(X86_FEATURE_OSPKE))
338 		copy_init_pkru_to_fpregs();
339 }
340 
341 /*
342  * Clear the FPU state back to init state.
343  *
344  * Called by sys_execve(), by the signal handler code and by various
345  * error paths.
346  */
fpu__clear(struct fpu * fpu,bool user_only)347 static void fpu__clear(struct fpu *fpu, bool user_only)
348 {
349 	WARN_ON_FPU(fpu != &current->thread.fpu);
350 
351 	if (!static_cpu_has(X86_FEATURE_FPU)) {
352 		fpu__drop(fpu);
353 		fpu__initialize(fpu);
354 		return;
355 	}
356 
357 	fpregs_lock();
358 
359 	if (user_only) {
360 		if (!fpregs_state_valid(fpu, smp_processor_id()) &&
361 		    xfeatures_mask_supervisor())
362 			copy_kernel_to_xregs(&fpu->state.xsave,
363 					     xfeatures_mask_supervisor());
364 		copy_init_fpstate_to_fpregs(xfeatures_mask_user());
365 	} else {
366 		copy_init_fpstate_to_fpregs(xfeatures_mask_all);
367 	}
368 
369 	fpregs_mark_activate();
370 	fpregs_unlock();
371 }
372 
fpu__clear_user_states(struct fpu * fpu)373 void fpu__clear_user_states(struct fpu *fpu)
374 {
375 	fpu__clear(fpu, true);
376 }
377 
fpu__clear_all(struct fpu * fpu)378 void fpu__clear_all(struct fpu *fpu)
379 {
380 	fpu__clear(fpu, false);
381 }
382 
383 /*
384  * Load FPU context before returning to userspace.
385  */
switch_fpu_return(void)386 void switch_fpu_return(void)
387 {
388 	if (!static_cpu_has(X86_FEATURE_FPU))
389 		return;
390 
391 	__fpregs_load_activate();
392 }
393 EXPORT_SYMBOL_GPL(switch_fpu_return);
394 
395 #ifdef CONFIG_X86_DEBUG_FPU
396 /*
397  * If current FPU state according to its tracking (loaded FPU context on this
398  * CPU) is not valid then we must have TIF_NEED_FPU_LOAD set so the context is
399  * loaded on return to userland.
400  */
fpregs_assert_state_consistent(void)401 void fpregs_assert_state_consistent(void)
402 {
403 	struct fpu *fpu = &current->thread.fpu;
404 
405 	if (test_thread_flag(TIF_NEED_FPU_LOAD))
406 		return;
407 
408 	WARN_ON_FPU(!fpregs_state_valid(fpu, smp_processor_id()));
409 }
410 EXPORT_SYMBOL_GPL(fpregs_assert_state_consistent);
411 #endif
412 
fpregs_mark_activate(void)413 void fpregs_mark_activate(void)
414 {
415 	struct fpu *fpu = &current->thread.fpu;
416 
417 	fpregs_activate(fpu);
418 	fpu->last_cpu = smp_processor_id();
419 	clear_thread_flag(TIF_NEED_FPU_LOAD);
420 }
421 EXPORT_SYMBOL_GPL(fpregs_mark_activate);
422 
423 /*
424  * x87 math exception handling:
425  */
426 
fpu__exception_code(struct fpu * fpu,int trap_nr)427 int fpu__exception_code(struct fpu *fpu, int trap_nr)
428 {
429 	int err;
430 
431 	if (trap_nr == X86_TRAP_MF) {
432 		unsigned short cwd, swd;
433 		/*
434 		 * (~cwd & swd) will mask out exceptions that are not set to unmasked
435 		 * status.  0x3f is the exception bits in these regs, 0x200 is the
436 		 * C1 reg you need in case of a stack fault, 0x040 is the stack
437 		 * fault bit.  We should only be taking one exception at a time,
438 		 * so if this combination doesn't produce any single exception,
439 		 * then we have a bad program that isn't synchronizing its FPU usage
440 		 * and it will suffer the consequences since we won't be able to
441 		 * fully reproduce the context of the exception.
442 		 */
443 		if (boot_cpu_has(X86_FEATURE_FXSR)) {
444 			cwd = fpu->state.fxsave.cwd;
445 			swd = fpu->state.fxsave.swd;
446 		} else {
447 			cwd = (unsigned short)fpu->state.fsave.cwd;
448 			swd = (unsigned short)fpu->state.fsave.swd;
449 		}
450 
451 		err = swd & ~cwd;
452 	} else {
453 		/*
454 		 * The SIMD FPU exceptions are handled a little differently, as there
455 		 * is only a single status/control register.  Thus, to determine which
456 		 * unmasked exception was caught we must mask the exception mask bits
457 		 * at 0x1f80, and then use these to mask the exception bits at 0x3f.
458 		 */
459 		unsigned short mxcsr = MXCSR_DEFAULT;
460 
461 		if (boot_cpu_has(X86_FEATURE_XMM))
462 			mxcsr = fpu->state.fxsave.mxcsr;
463 
464 		err = ~(mxcsr >> 7) & mxcsr;
465 	}
466 
467 	if (err & 0x001) {	/* Invalid op */
468 		/*
469 		 * swd & 0x240 == 0x040: Stack Underflow
470 		 * swd & 0x240 == 0x240: Stack Overflow
471 		 * User must clear the SF bit (0x40) if set
472 		 */
473 		return FPE_FLTINV;
474 	} else if (err & 0x004) { /* Divide by Zero */
475 		return FPE_FLTDIV;
476 	} else if (err & 0x008) { /* Overflow */
477 		return FPE_FLTOVF;
478 	} else if (err & 0x012) { /* Denormal, Underflow */
479 		return FPE_FLTUND;
480 	} else if (err & 0x020) { /* Precision */
481 		return FPE_FLTRES;
482 	}
483 
484 	/*
485 	 * If we're using IRQ 13, or supposedly even some trap
486 	 * X86_TRAP_MF implementations, it's possible
487 	 * we get a spurious trap, which is not an error.
488 	 */
489 	return 0;
490 }
491