2 * xsave/xrstor support.
4 * Author: Suresh Siddha <suresh.b.siddha@intel.com>
7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9 #include <linux/bootmem.h>
10 #include <linux/compat.h>
11 #include <linux/cpu.h>
13 #include <asm/fpu-internal.h>
14 #include <asm/sigframe.h>
18 * Supported feature mask by the CPU and the kernel.
23 * Represents init state for the supported extended state.
25 struct xsave_struct *init_xstate_buf;
27 static struct _fpx_sw_bytes fx_sw_reserved, fx_sw_reserved_ia32;
28 static unsigned int *xstate_offsets, *xstate_sizes;
29 static unsigned int xstate_comp_offsets[sizeof(pcntxt_mask)*8];
30 static unsigned int xstate_features;
33 * If a processor implementation discern that a processor state component is
34 * in its initialized state it may modify the corresponding bit in the
35 * xsave_hdr.xstate_bv as '0', with out modifying the corresponding memory
36 * layout in the case of xsaveopt. While presenting the xstate information to
37 * the user, we always ensure that the memory layout of a feature will be in
38 * the init state if the corresponding header bit is zero. This is to ensure
39 * that the user doesn't see some stale state in the memory layout during
40 * signal handling, debugging etc.
42 void __sanitize_i387_state(struct task_struct *tsk)
44 struct i387_fxsave_struct *fx = &tsk->thread.fpu.state->fxsave;
45 int feature_bit = 0x2;
51 xstate_bv = tsk->thread.fpu.state->xsave.xsave_hdr.xstate_bv;
54 * None of the feature bits are in init state. So nothing else
55 * to do for us, as the memory layout is up to date.
57 if ((xstate_bv & pcntxt_mask) == pcntxt_mask)
63 if (!(xstate_bv & XSTATE_FP)) {
70 memset(&fx->st_space[0], 0, 128);
74 * SSE is in init state
76 if (!(xstate_bv & XSTATE_SSE))
77 memset(&fx->xmm_space[0], 0, 256);
79 xstate_bv = (pcntxt_mask & ~xstate_bv) >> 2;
82 * Update all the other memory layouts for which the corresponding
83 * header bit is in the init state.
86 if (xstate_bv & 0x1) {
87 int offset = xstate_offsets[feature_bit];
88 int size = xstate_sizes[feature_bit];
90 memcpy(((void *) fx) + offset,
91 ((void *) init_xstate_buf) + offset,
101 * Check for the presence of extended state information in the
102 * user fpstate pointer in the sigcontext.
104 static inline int check_for_xstate(struct i387_fxsave_struct __user *buf,
105 void __user *fpstate,
106 struct _fpx_sw_bytes *fx_sw)
108 int min_xstate_size = sizeof(struct i387_fxsave_struct) +
109 sizeof(struct xsave_hdr_struct);
112 if (__copy_from_user(fx_sw, &buf->sw_reserved[0], sizeof(*fx_sw)))
115 /* Check for the first magic field and other error scenarios. */
116 if (fx_sw->magic1 != FP_XSTATE_MAGIC1 ||
117 fx_sw->xstate_size < min_xstate_size ||
118 fx_sw->xstate_size > xstate_size ||
119 fx_sw->xstate_size > fx_sw->extended_size)
123 * Check for the presence of second magic word at the end of memory
124 * layout. This detects the case where the user just copied the legacy
125 * fpstate layout with out copying the extended state information
126 * in the memory layout.
128 if (__get_user(magic2, (__u32 __user *)(fpstate + fx_sw->xstate_size))
129 || magic2 != FP_XSTATE_MAGIC2)
136 * Signal frame handlers.
138 static inline int save_fsave_header(struct task_struct *tsk, void __user *buf)
141 struct xsave_struct *xsave = &tsk->thread.fpu.state->xsave;
142 struct user_i387_ia32_struct env;
143 struct _fpstate_ia32 __user *fp = buf;
145 convert_from_fxsr(&env, tsk);
147 if (__copy_to_user(buf, &env, sizeof(env)) ||
148 __put_user(xsave->i387.swd, &fp->status) ||
149 __put_user(X86_FXSR_MAGIC, &fp->magic))
152 struct i387_fsave_struct __user *fp = buf;
154 if (__get_user(swd, &fp->swd) || __put_user(swd, &fp->status))
161 static inline int save_xstate_epilog(void __user *buf, int ia32_frame)
163 struct xsave_struct __user *x = buf;
164 struct _fpx_sw_bytes *sw_bytes;
168 /* Setup the bytes not touched by the [f]xsave and reserved for SW. */
169 sw_bytes = ia32_frame ? &fx_sw_reserved_ia32 : &fx_sw_reserved;
170 err = __copy_to_user(&x->i387.sw_reserved, sw_bytes, sizeof(*sw_bytes));
175 err |= __put_user(FP_XSTATE_MAGIC2, (__u32 *)(buf + xstate_size));
178 * Read the xstate_bv which we copied (directly from the cpu or
179 * from the state in task struct) to the user buffers.
181 err |= __get_user(xstate_bv, (__u32 *)&x->xsave_hdr.xstate_bv);
184 * For legacy compatible, we always set FP/SSE bits in the bit
185 * vector while saving the state to the user context. This will
186 * enable us capturing any changes(during sigreturn) to
187 * the FP/SSE bits by the legacy applications which don't touch
188 * xstate_bv in the xsave header.
190 * xsave aware apps can change the xstate_bv in the xsave
191 * header as well as change any contents in the memory layout.
192 * xrestore as part of sigreturn will capture all the changes.
194 xstate_bv |= XSTATE_FPSSE;
196 err |= __put_user(xstate_bv, (__u32 *)&x->xsave_hdr.xstate_bv);
201 static inline int save_user_xstate(struct xsave_struct __user *buf)
206 err = xsave_user(buf);
208 err = fxsave_user((struct i387_fxsave_struct __user *) buf);
210 err = fsave_user((struct i387_fsave_struct __user *) buf);
212 if (unlikely(err) && __clear_user(buf, xstate_size))
218 * Save the fpu, extended register state to the user signal frame.
220 * 'buf_fx' is the 64-byte aligned pointer at which the [f|fx|x]save
222 * 'buf' points to the 'buf_fx' or to the fsave header followed by 'buf_fx'.
224 * buf == buf_fx for 64-bit frames and 32-bit fsave frame.
225 * buf != buf_fx for 32-bit frames with fxstate.
227 * If the fpu, extended register state is live, save the state directly
228 * to the user frame pointed by the aligned pointer 'buf_fx'. Otherwise,
229 * copy the thread's fpu state to the user frame starting at 'buf_fx'.
231 * If this is a 32-bit frame with fxstate, put a fsave header before
232 * the aligned state at 'buf_fx'.
234 * For [f]xsave state, update the SW reserved fields in the [f]xsave frame
235 * indicating the absence/presence of the extended state to the user.
237 int save_xstate_sig(void __user *buf, void __user *buf_fx, int size)
239 struct xsave_struct *xsave = ¤t->thread.fpu.state->xsave;
240 struct task_struct *tsk = current;
241 int ia32_fxstate = (buf != buf_fx);
243 ia32_fxstate &= (config_enabled(CONFIG_X86_32) ||
244 config_enabled(CONFIG_IA32_EMULATION));
246 if (!access_ok(VERIFY_WRITE, buf, size))
249 if (!static_cpu_has(X86_FEATURE_FPU))
250 return fpregs_soft_get(current, NULL, 0,
251 sizeof(struct user_i387_ia32_struct), NULL,
252 (struct _fpstate_ia32 __user *) buf) ? -1 : 1;
254 if (user_has_fpu()) {
255 /* Save the live register state to the user directly. */
256 if (save_user_xstate(buf_fx))
258 /* Update the thread's fxstate to save the fsave header. */
260 fpu_fxsave(&tsk->thread.fpu);
262 sanitize_i387_state(tsk);
263 if (__copy_to_user(buf_fx, xsave, xstate_size))
267 /* Save the fsave header for the 32-bit frames. */
268 if ((ia32_fxstate || !use_fxsr()) && save_fsave_header(tsk, buf))
271 if (use_fxsr() && save_xstate_epilog(buf_fx, ia32_fxstate))
274 drop_init_fpu(tsk); /* trigger finit */
280 sanitize_restored_xstate(struct task_struct *tsk,
281 struct user_i387_ia32_struct *ia32_env,
282 u64 xstate_bv, int fx_only)
284 struct xsave_struct *xsave = &tsk->thread.fpu.state->xsave;
285 struct xsave_hdr_struct *xsave_hdr = &xsave->xsave_hdr;
288 /* These bits must be zero. */
289 memset(xsave_hdr->reserved, 0, 48);
292 * Init the state that is not present in the memory
293 * layout and not enabled by the OS.
296 xsave_hdr->xstate_bv = XSTATE_FPSSE;
298 xsave_hdr->xstate_bv &= (pcntxt_mask & xstate_bv);
303 * mscsr reserved bits must be masked to zero for security
306 xsave->i387.mxcsr &= mxcsr_feature_mask;
308 convert_to_fxsr(tsk, ia32_env);
313 * Restore the extended state if present. Otherwise, restore the FP/SSE state.
315 static inline int restore_user_xstate(void __user *buf, u64 xbv, int fx_only)
318 if ((unsigned long)buf % 64 || fx_only) {
319 u64 init_bv = pcntxt_mask & ~XSTATE_FPSSE;
320 xrstor_state(init_xstate_buf, init_bv);
321 return fxrstor_user(buf);
323 u64 init_bv = pcntxt_mask & ~xbv;
324 if (unlikely(init_bv))
325 xrstor_state(init_xstate_buf, init_bv);
326 return xrestore_user(buf, xbv);
328 } else if (use_fxsr()) {
329 return fxrstor_user(buf);
331 return frstor_user(buf);
334 int __restore_xstate_sig(void __user *buf, void __user *buf_fx, int size)
336 int ia32_fxstate = (buf != buf_fx);
337 struct task_struct *tsk = current;
338 int state_size = xstate_size;
342 ia32_fxstate &= (config_enabled(CONFIG_X86_32) ||
343 config_enabled(CONFIG_IA32_EMULATION));
350 if (!access_ok(VERIFY_READ, buf, size))
353 if (!used_math() && init_fpu(tsk))
356 if (!static_cpu_has(X86_FEATURE_FPU))
357 return fpregs_soft_set(current, NULL,
358 0, sizeof(struct user_i387_ia32_struct),
362 struct _fpx_sw_bytes fx_sw_user;
363 if (unlikely(check_for_xstate(buf_fx, buf_fx, &fx_sw_user))) {
365 * Couldn't find the extended state information in the
366 * memory layout. Restore just the FP/SSE and init all
367 * the other extended state.
369 state_size = sizeof(struct i387_fxsave_struct);
372 state_size = fx_sw_user.xstate_size;
373 xstate_bv = fx_sw_user.xstate_bv;
379 * For 32-bit frames with fxstate, copy the user state to the
380 * thread's fpu state, reconstruct fxstate from the fsave
381 * header. Sanitize the copied state etc.
383 struct xsave_struct *xsave = &tsk->thread.fpu.state->xsave;
384 struct user_i387_ia32_struct env;
388 * Drop the current fpu which clears used_math(). This ensures
389 * that any context-switch during the copy of the new state,
390 * avoids the intermediate state from getting restored/saved.
391 * Thus avoiding the new restored state from getting corrupted.
392 * We will be ready to restore/save the state only after
393 * set_used_math() is again set.
397 if (__copy_from_user(xsave, buf_fx, state_size) ||
398 __copy_from_user(&env, buf, sizeof(env))) {
401 sanitize_restored_xstate(tsk, &env, xstate_bv, fx_only);
406 math_state_restore();
411 * For 64-bit frames and 32-bit fsave frames, restore the user
412 * state to the registers directly (with exceptions handled).
415 if (restore_user_xstate(buf_fx, xstate_bv, fx_only)) {
425 * Prepare the SW reserved portion of the fxsave memory layout, indicating
426 * the presence of the extended state information in the memory layout
427 * pointed by the fpstate pointer in the sigcontext.
428 * This will be saved when ever the FP and extended state context is
429 * saved on the user stack during the signal handler delivery to the user.
431 static void prepare_fx_sw_frame(void)
433 int fsave_header_size = sizeof(struct i387_fsave_struct);
434 int size = xstate_size + FP_XSTATE_MAGIC2_SIZE;
436 if (config_enabled(CONFIG_X86_32))
437 size += fsave_header_size;
439 fx_sw_reserved.magic1 = FP_XSTATE_MAGIC1;
440 fx_sw_reserved.extended_size = size;
441 fx_sw_reserved.xstate_bv = pcntxt_mask;
442 fx_sw_reserved.xstate_size = xstate_size;
444 if (config_enabled(CONFIG_IA32_EMULATION)) {
445 fx_sw_reserved_ia32 = fx_sw_reserved;
446 fx_sw_reserved_ia32.extended_size += fsave_header_size;
451 * Enable the extended processor state save/restore feature
453 static inline void xstate_enable(void)
455 set_in_cr4(X86_CR4_OSXSAVE);
456 xsetbv(XCR_XFEATURE_ENABLED_MASK, pcntxt_mask);
460 * Record the offsets and sizes of different state managed by the xsave
463 static void __init setup_xstate_features(void)
465 int eax, ebx, ecx, edx, leaf = 0x2;
467 xstate_features = fls64(pcntxt_mask);
468 xstate_offsets = alloc_bootmem(xstate_features * sizeof(int));
469 xstate_sizes = alloc_bootmem(xstate_features * sizeof(int));
472 cpuid_count(XSTATE_CPUID, leaf, &eax, &ebx, &ecx, &edx);
477 xstate_offsets[leaf] = ebx;
478 xstate_sizes[leaf] = eax;
485 * This function sets up offsets and sizes of all extended states in
486 * xsave area. This supports both standard format and compacted format
487 * of the xsave aread.
492 void setup_xstate_comp(void)
494 unsigned int xstate_comp_sizes[sizeof(pcntxt_mask)*8];
498 * The FP xstates and SSE xstates are legacy states. They are always
499 * in the fixed offsets in the xsave area in either compacted form
502 xstate_comp_offsets[0] = 0;
503 xstate_comp_offsets[1] = offsetof(struct i387_fxsave_struct, xmm_space);
505 if (!cpu_has_xsaves) {
506 for (i = 2; i < xstate_features; i++) {
507 if (test_bit(i, (unsigned long *)&pcntxt_mask)) {
508 xstate_comp_offsets[i] = xstate_offsets[i];
509 xstate_comp_sizes[i] = xstate_sizes[i];
515 xstate_comp_offsets[2] = FXSAVE_SIZE + XSAVE_HDR_SIZE;
517 for (i = 2; i < xstate_features; i++) {
518 if (test_bit(i, (unsigned long *)&pcntxt_mask))
519 xstate_comp_sizes[i] = xstate_sizes[i];
521 xstate_comp_sizes[i] = 0;
524 xstate_comp_offsets[i] = xstate_comp_offsets[i-1]
525 + xstate_comp_sizes[i-1];
531 * setup the xstate image representing the init state
533 static void __init setup_init_fpu_buf(void)
536 * Setup init_xstate_buf to represent the init state of
537 * all the features managed by the xsave
539 init_xstate_buf = alloc_bootmem_align(xstate_size,
540 __alignof__(struct xsave_struct));
541 fx_finit(&init_xstate_buf->i387);
546 setup_xstate_features();
548 if (cpu_has_xsaves) {
549 init_xstate_buf->xsave_hdr.xcomp_bv =
550 (u64)1 << 63 | pcntxt_mask;
551 init_xstate_buf->xsave_hdr.xstate_bv = pcntxt_mask;
555 * Init all the features state with header_bv being 0x0
557 xrstor_state_booting(init_xstate_buf, -1);
559 * Dump the init state again. This is to identify the init state
560 * of any feature which is not represented by all zero's.
562 xsave_state_booting(init_xstate_buf, -1);
565 static enum { AUTO, ENABLE, DISABLE } eagerfpu = AUTO;
566 static int __init eager_fpu_setup(char *s)
568 if (!strcmp(s, "on"))
570 else if (!strcmp(s, "off"))
572 else if (!strcmp(s, "auto"))
576 __setup("eagerfpu=", eager_fpu_setup);
580 * Calculate total size of enabled xstates in XCR0/pcntxt_mask.
582 static void __init init_xstate_size(void)
584 unsigned int eax, ebx, ecx, edx;
587 if (!cpu_has_xsaves) {
588 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
593 xstate_size = FXSAVE_SIZE + XSAVE_HDR_SIZE;
594 for (i = 2; i < 64; i++) {
595 if (test_bit(i, (unsigned long *)&pcntxt_mask)) {
596 cpuid_count(XSTATE_CPUID, i, &eax, &ebx, &ecx, &edx);
603 * Enable and initialize the xsave feature.
605 static void __init xstate_enable_boot_cpu(void)
607 unsigned int eax, ebx, ecx, edx;
609 if (boot_cpu_data.cpuid_level < XSTATE_CPUID) {
610 WARN(1, KERN_ERR "XSTATE_CPUID missing\n");
614 cpuid_count(XSTATE_CPUID, 0, &eax, &ebx, &ecx, &edx);
615 pcntxt_mask = eax + ((u64)edx << 32);
617 if ((pcntxt_mask & XSTATE_FPSSE) != XSTATE_FPSSE) {
618 pr_err("FP/SSE not shown under xsave features 0x%llx\n",
624 * Support only the state known to OS.
626 pcntxt_mask = pcntxt_mask & XCNTXT_MASK;
631 * Recompute the context size for enabled features
635 update_regset_xstate_info(xstate_size, pcntxt_mask);
636 prepare_fx_sw_frame();
637 setup_init_fpu_buf();
639 /* Auto enable eagerfpu for xsaveopt */
640 if (cpu_has_xsaveopt && eagerfpu != DISABLE)
643 if (pcntxt_mask & XSTATE_EAGER) {
644 if (eagerfpu == DISABLE) {
645 pr_err("eagerfpu not present, disabling some xstate features: 0x%llx\n",
646 pcntxt_mask & XSTATE_EAGER);
647 pcntxt_mask &= ~XSTATE_EAGER;
653 pr_info("enabled xstate_bv 0x%llx, cntxt size 0x%x using %s\n",
654 pcntxt_mask, xstate_size,
655 cpu_has_xsaves ? "compacted form" : "standard form");
659 * For the very first instance, this calls xstate_enable_boot_cpu();
660 * for all subsequent instances, this calls xstate_enable().
662 * This is somewhat obfuscated due to the lack of powerful enough
663 * overrides for the section checks.
665 void xsave_init(void)
667 static __refdata void (*next_func)(void) = xstate_enable_boot_cpu;
668 void (*this_func)(void);
673 this_func = next_func;
674 next_func = xstate_enable;
678 static inline void __init eager_fpu_init_bp(void)
680 current->thread.fpu.state =
681 alloc_bootmem_align(xstate_size, __alignof__(struct xsave_struct));
682 if (!init_xstate_buf)
683 setup_init_fpu_buf();
686 void eager_fpu_init(void)
688 static __refdata void (*boot_func)(void) = eager_fpu_init_bp;
691 current_thread_info()->status = 0;
693 if (eagerfpu == ENABLE)
694 setup_force_cpu_cap(X86_FEATURE_EAGER_FPU);
696 if (!cpu_has_eager_fpu) {
707 * This is same as math_state_restore(). But use_xsave() is
708 * not yet patched to use math_state_restore().
711 __thread_fpu_begin(current);
713 xrstor_state(init_xstate_buf, -1);
715 fxrstor_checking(&init_xstate_buf->i387);
719 * Given the xsave area and a state inside, this function returns the
720 * address of the state.
722 * This is the API that is called to get xstate address in either
723 * standard format or compacted format of xsave area.
726 * xsave: base address of the xsave area;
727 * xstate: state which is defined in xsave.h (e.g. XSTATE_FP, XSTATE_SSE,
730 * address of the state in the xsave area.
732 void *get_xsave_addr(struct xsave_struct *xsave, int xstate)
734 int feature = fls64(xstate) - 1;
735 if (!test_bit(feature, (unsigned long *)&pcntxt_mask))
738 return (void *)xsave + xstate_comp_offsets[feature];