• 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 /*
29  * Track whether the kernel is using the FPU state
30  * currently.
31  *
32  * This flag is used:
33  *
34  *   - by IRQ context code to potentially use the FPU
35  *     if it's unused.
36  *
37  *   - to debug kernel_fpu_begin()/end() correctness
38  */
39 static DEFINE_PER_CPU(bool, in_kernel_fpu);
40 
41 /*
42  * Track which context is using the FPU on the CPU:
43  */
44 DEFINE_PER_CPU(struct fpu *, fpu_fpregs_owner_ctx);
45 
kernel_fpu_disabled(void)46 static bool kernel_fpu_disabled(void)
47 {
48 	return this_cpu_read(in_kernel_fpu);
49 }
50 
interrupted_kernel_fpu_idle(void)51 static bool interrupted_kernel_fpu_idle(void)
52 {
53 	return !kernel_fpu_disabled();
54 }
55 
56 /*
57  * Were we in user mode (or vm86 mode) when we were
58  * interrupted?
59  *
60  * Doing kernel_fpu_begin/end() is ok if we are running
61  * in an interrupt context from user mode - we'll just
62  * save the FPU state as required.
63  */
interrupted_user_mode(void)64 static bool interrupted_user_mode(void)
65 {
66 	struct pt_regs *regs = get_irq_regs();
67 	return regs && user_mode(regs);
68 }
69 
70 /*
71  * Can we use the FPU in kernel mode with the
72  * whole "kernel_fpu_begin/end()" sequence?
73  *
74  * It's always ok in process context (ie "not interrupt")
75  * but it is sometimes ok even from an irq.
76  */
irq_fpu_usable(void)77 bool irq_fpu_usable(void)
78 {
79 	return !in_interrupt() ||
80 		interrupted_user_mode() ||
81 		interrupted_kernel_fpu_idle();
82 }
83 EXPORT_SYMBOL(irq_fpu_usable);
84 
kernel_fpu_begin_mask(unsigned int kfpu_mask)85 void kernel_fpu_begin_mask(unsigned int kfpu_mask)
86 {
87 	preempt_disable();
88 
89 	WARN_ON_FPU(!irq_fpu_usable());
90 	WARN_ON_FPU(this_cpu_read(in_kernel_fpu));
91 
92 	this_cpu_write(in_kernel_fpu, true);
93 
94 	if (!(current->flags & PF_KTHREAD) &&
95 	    !test_thread_flag(TIF_NEED_FPU_LOAD)) {
96 		set_thread_flag(TIF_NEED_FPU_LOAD);
97 		/*
98 		 * Ignore return value -- we don't care if reg state
99 		 * is clobbered.
100 		 */
101 		copy_fpregs_to_fpstate(&current->thread.fpu);
102 	}
103 	__cpu_invalidate_fpregs_state();
104 
105 	/* Put sane initial values into the control registers. */
106 	if (likely(kfpu_mask & KFPU_MXCSR) && boot_cpu_has(X86_FEATURE_XMM))
107 		ldmxcsr(MXCSR_DEFAULT);
108 
109 	if (unlikely(kfpu_mask & KFPU_387) && boot_cpu_has(X86_FEATURE_FPU))
110 		asm volatile ("fninit");
111 }
112 EXPORT_SYMBOL_GPL(kernel_fpu_begin_mask);
113 
kernel_fpu_end(void)114 void kernel_fpu_end(void)
115 {
116 	WARN_ON_FPU(!this_cpu_read(in_kernel_fpu));
117 
118 	this_cpu_write(in_kernel_fpu, false);
119 	preempt_enable();
120 }
121 EXPORT_SYMBOL_GPL(kernel_fpu_end);
122 
123 /*
124  * Save the FPU state (mark it for reload if necessary):
125  *
126  * This only ever gets called for the current task.
127  */
fpu__save(struct fpu * fpu)128 void fpu__save(struct fpu *fpu)
129 {
130 	WARN_ON_FPU(fpu != &current->thread.fpu);
131 
132 	fpregs_lock();
133 	trace_x86_fpu_before_save(fpu);
134 
135 	if (!test_thread_flag(TIF_NEED_FPU_LOAD)) {
136 		if (!copy_fpregs_to_fpstate(fpu)) {
137 			copy_kernel_to_fpregs(&fpu->state);
138 		}
139 	}
140 
141 	trace_x86_fpu_after_save(fpu);
142 	fpregs_unlock();
143 }
144 
145 /*
146  * Legacy x87 fpstate state init:
147  */
fpstate_init_fstate(struct fregs_state * fp)148 static inline void fpstate_init_fstate(struct fregs_state *fp)
149 {
150 	fp->cwd = 0xffff037fu;
151 	fp->swd = 0xffff0000u;
152 	fp->twd = 0xffffffffu;
153 	fp->fos = 0xffff0000u;
154 }
155 
fpstate_init(union fpregs_state * state)156 void fpstate_init(union fpregs_state *state)
157 {
158 	if (!static_cpu_has(X86_FEATURE_FPU)) {
159 		fpstate_init_soft(&state->soft);
160 		return;
161 	}
162 
163 	memset(state, 0, fpu_kernel_xstate_size);
164 
165 	if (static_cpu_has(X86_FEATURE_XSAVES))
166 		fpstate_init_xstate(&state->xsave);
167 	if (static_cpu_has(X86_FEATURE_FXSR))
168 		fpstate_init_fxstate(&state->fxsave);
169 	else
170 		fpstate_init_fstate(&state->fsave);
171 }
172 EXPORT_SYMBOL_GPL(fpstate_init);
173 
fpu__copy(struct task_struct * dst,struct task_struct * src)174 int fpu__copy(struct task_struct *dst, struct task_struct *src)
175 {
176 	struct fpu *dst_fpu = &dst->thread.fpu;
177 	struct fpu *src_fpu = &src->thread.fpu;
178 
179 	dst_fpu->last_cpu = -1;
180 
181 	if (!static_cpu_has(X86_FEATURE_FPU))
182 		return 0;
183 
184 	WARN_ON_FPU(src_fpu != &current->thread.fpu);
185 
186 	/*
187 	 * Don't let 'init optimized' areas of the XSAVE area
188 	 * leak into the child task:
189 	 */
190 	memset(&dst_fpu->state.xsave, 0, fpu_kernel_xstate_size);
191 
192 	/*
193 	 * If the FPU registers are not current just memcpy() the state.
194 	 * Otherwise save current FPU registers directly into the child's FPU
195 	 * context, without any memory-to-memory copying.
196 	 *
197 	 * ( The function 'fails' in the FNSAVE case, which destroys
198 	 *   register contents so we have to load them back. )
199 	 */
200 	fpregs_lock();
201 	if (test_thread_flag(TIF_NEED_FPU_LOAD))
202 		memcpy(&dst_fpu->state, &src_fpu->state, fpu_kernel_xstate_size);
203 
204 	else if (!copy_fpregs_to_fpstate(dst_fpu))
205 		copy_kernel_to_fpregs(&dst_fpu->state);
206 
207 	fpregs_unlock();
208 
209 	set_tsk_thread_flag(dst, TIF_NEED_FPU_LOAD);
210 
211 	trace_x86_fpu_copy_src(src_fpu);
212 	trace_x86_fpu_copy_dst(dst_fpu);
213 
214 	return 0;
215 }
216 
217 /*
218  * Activate the current task's in-memory FPU context,
219  * if it has not been used before:
220  */
fpu__initialize(struct fpu * fpu)221 static void fpu__initialize(struct fpu *fpu)
222 {
223 	WARN_ON_FPU(fpu != &current->thread.fpu);
224 
225 	set_thread_flag(TIF_NEED_FPU_LOAD);
226 	fpstate_init(&fpu->state);
227 	trace_x86_fpu_init_state(fpu);
228 }
229 
230 /*
231  * This function must be called before we read a task's fpstate.
232  *
233  * There's two cases where this gets called:
234  *
235  * - for the current task (when coredumping), in which case we have
236  *   to save the latest FPU registers into the fpstate,
237  *
238  * - or it's called for stopped tasks (ptrace), in which case the
239  *   registers were already saved by the context-switch code when
240  *   the task scheduled out.
241  *
242  * If the task has used the FPU before then save it.
243  */
fpu__prepare_read(struct fpu * fpu)244 void fpu__prepare_read(struct fpu *fpu)
245 {
246 	if (fpu == &current->thread.fpu)
247 		fpu__save(fpu);
248 }
249 
250 /*
251  * This function must be called before we write a task's fpstate.
252  *
253  * Invalidate any cached FPU registers.
254  *
255  * After this function call, after registers in the fpstate are
256  * modified and the child task has woken up, the child task will
257  * restore the modified FPU state from the modified context. If we
258  * didn't clear its cached status here then the cached in-registers
259  * state pending on its former CPU could be restored, corrupting
260  * the modifications.
261  */
fpu__prepare_write(struct fpu * fpu)262 void fpu__prepare_write(struct fpu *fpu)
263 {
264 	/*
265 	 * Only stopped child tasks can be used to modify the FPU
266 	 * state in the fpstate buffer:
267 	 */
268 	WARN_ON_FPU(fpu == &current->thread.fpu);
269 
270 	/* Invalidate any cached state: */
271 	__fpu_invalidate_fpregs_state(fpu);
272 }
273 
274 /*
275  * Drops current FPU state: deactivates the fpregs and
276  * the fpstate. NOTE: it still leaves previous contents
277  * in the fpregs in the eager-FPU case.
278  *
279  * This function can be used in cases where we know that
280  * a state-restore is coming: either an explicit one,
281  * or a reschedule.
282  */
fpu__drop(struct fpu * fpu)283 void fpu__drop(struct fpu *fpu)
284 {
285 	preempt_disable();
286 
287 	if (fpu == &current->thread.fpu) {
288 		/* Ignore delayed exceptions from user space */
289 		asm volatile("1: fwait\n"
290 			     "2:\n"
291 			     _ASM_EXTABLE(1b, 2b));
292 		fpregs_deactivate(fpu);
293 	}
294 
295 	trace_x86_fpu_dropped(fpu);
296 
297 	preempt_enable();
298 }
299 
300 /*
301  * Clear FPU registers by setting them up from
302  * the init fpstate:
303  */
copy_init_fpstate_to_fpregs(void)304 static inline void copy_init_fpstate_to_fpregs(void)
305 {
306 	fpregs_lock();
307 
308 	if (use_xsave())
309 		copy_kernel_to_xregs(&init_fpstate.xsave, -1);
310 	else if (static_cpu_has(X86_FEATURE_FXSR))
311 		copy_kernel_to_fxregs(&init_fpstate.fxsave);
312 	else
313 		copy_kernel_to_fregs(&init_fpstate.fsave);
314 
315 	if (boot_cpu_has(X86_FEATURE_OSPKE))
316 		copy_init_pkru_to_fpregs();
317 
318 	fpregs_mark_activate();
319 	fpregs_unlock();
320 }
321 
322 /*
323  * Clear the FPU state back to init state.
324  *
325  * Called by sys_execve(), by the signal handler code and by various
326  * error paths.
327  */
fpu__clear(struct fpu * fpu)328 void fpu__clear(struct fpu *fpu)
329 {
330 	WARN_ON_FPU(fpu != &current->thread.fpu); /* Almost certainly an anomaly */
331 
332 	fpu__drop(fpu);
333 
334 	/*
335 	 * Make sure fpstate is cleared and initialized.
336 	 */
337 	fpu__initialize(fpu);
338 	if (static_cpu_has(X86_FEATURE_FPU))
339 		copy_init_fpstate_to_fpregs();
340 }
341 
342 /*
343  * Load FPU context before returning to userspace.
344  */
switch_fpu_return(void)345 void switch_fpu_return(void)
346 {
347 	if (!static_cpu_has(X86_FEATURE_FPU))
348 		return;
349 
350 	__fpregs_load_activate();
351 }
352 EXPORT_SYMBOL_GPL(switch_fpu_return);
353 
354 #ifdef CONFIG_X86_DEBUG_FPU
355 /*
356  * If current FPU state according to its tracking (loaded FPU context on this
357  * CPU) is not valid then we must have TIF_NEED_FPU_LOAD set so the context is
358  * loaded on return to userland.
359  */
fpregs_assert_state_consistent(void)360 void fpregs_assert_state_consistent(void)
361 {
362 	struct fpu *fpu = &current->thread.fpu;
363 
364 	if (test_thread_flag(TIF_NEED_FPU_LOAD))
365 		return;
366 
367 	WARN_ON_FPU(!fpregs_state_valid(fpu, smp_processor_id()));
368 }
369 EXPORT_SYMBOL_GPL(fpregs_assert_state_consistent);
370 #endif
371 
fpregs_mark_activate(void)372 void fpregs_mark_activate(void)
373 {
374 	struct fpu *fpu = &current->thread.fpu;
375 
376 	fpregs_activate(fpu);
377 	fpu->last_cpu = smp_processor_id();
378 	clear_thread_flag(TIF_NEED_FPU_LOAD);
379 }
380 EXPORT_SYMBOL_GPL(fpregs_mark_activate);
381 
382 /*
383  * x87 math exception handling:
384  */
385 
fpu__exception_code(struct fpu * fpu,int trap_nr)386 int fpu__exception_code(struct fpu *fpu, int trap_nr)
387 {
388 	int err;
389 
390 	if (trap_nr == X86_TRAP_MF) {
391 		unsigned short cwd, swd;
392 		/*
393 		 * (~cwd & swd) will mask out exceptions that are not set to unmasked
394 		 * status.  0x3f is the exception bits in these regs, 0x200 is the
395 		 * C1 reg you need in case of a stack fault, 0x040 is the stack
396 		 * fault bit.  We should only be taking one exception at a time,
397 		 * so if this combination doesn't produce any single exception,
398 		 * then we have a bad program that isn't synchronizing its FPU usage
399 		 * and it will suffer the consequences since we won't be able to
400 		 * fully reproduce the context of the exception.
401 		 */
402 		if (boot_cpu_has(X86_FEATURE_FXSR)) {
403 			cwd = fpu->state.fxsave.cwd;
404 			swd = fpu->state.fxsave.swd;
405 		} else {
406 			cwd = (unsigned short)fpu->state.fsave.cwd;
407 			swd = (unsigned short)fpu->state.fsave.swd;
408 		}
409 
410 		err = swd & ~cwd;
411 	} else {
412 		/*
413 		 * The SIMD FPU exceptions are handled a little differently, as there
414 		 * is only a single status/control register.  Thus, to determine which
415 		 * unmasked exception was caught we must mask the exception mask bits
416 		 * at 0x1f80, and then use these to mask the exception bits at 0x3f.
417 		 */
418 		unsigned short mxcsr = MXCSR_DEFAULT;
419 
420 		if (boot_cpu_has(X86_FEATURE_XMM))
421 			mxcsr = fpu->state.fxsave.mxcsr;
422 
423 		err = ~(mxcsr >> 7) & mxcsr;
424 	}
425 
426 	if (err & 0x001) {	/* Invalid op */
427 		/*
428 		 * swd & 0x240 == 0x040: Stack Underflow
429 		 * swd & 0x240 == 0x240: Stack Overflow
430 		 * User must clear the SF bit (0x40) if set
431 		 */
432 		return FPE_FLTINV;
433 	} else if (err & 0x004) { /* Divide by Zero */
434 		return FPE_FLTDIV;
435 	} else if (err & 0x008) { /* Overflow */
436 		return FPE_FLTOVF;
437 	} else if (err & 0x012) { /* Denormal, Underflow */
438 		return FPE_FLTUND;
439 	} else if (err & 0x020) { /* Precision */
440 		return FPE_FLTRES;
441 	}
442 
443 	/*
444 	 * If we're using IRQ 13, or supposedly even some trap
445 	 * X86_TRAP_MF implementations, it's possible
446 	 * we get a spurious trap, which is not an error.
447 	 */
448 	return 0;
449 }
450