fpu.h

  1/* SPDX-License-Identifier: GPL-2.0 */
  2/*
  3 * In-kernel FPU support functions
  4 *
  5 *
  6 * Consider these guidelines before using in-kernel FPU functions:
  7 *
  8 *  1. Use kernel_fpu_begin() and kernel_fpu_end() to enclose all in-kernel
  9 *     use of floating-point or vector registers and instructions.
 10 *
 11 *  2. For kernel_fpu_begin(), specify the vector register range you want to
 12 *     use with the KERNEL_VXR_* constants. Consider these usage guidelines:
 13 *
 14 *     a) If your function typically runs in process-context, use the lower
 15 *	  half of the vector registers, for example, specify KERNEL_VXR_LOW.
 16 *     b) If your function typically runs in soft-irq or hard-irq context,
 17 *	  prefer using the upper half of the vector registers, for example,
 18 *	  specify KERNEL_VXR_HIGH.
 19 *
 20 *     If you adhere to these guidelines, an interrupted process context
 21 *     does not require to save and restore vector registers because of
 22 *     disjoint register ranges.
 23 *
 24 *     Also note that the __kernel_fpu_begin()/__kernel_fpu_end() functions
 25 *     includes logic to save and restore up to 16 vector registers at once.
 26 *
 27 *  3. You can nest kernel_fpu_begin()/kernel_fpu_end() by using different
 28 *     struct kernel_fpu states.  Vector registers that are in use by outer
 29 *     levels are saved and restored.  You can minimize the save and restore
 30 *     effort by choosing disjoint vector register ranges.
 31 *
 32 *  5. To use vector floating-point instructions, specify the KERNEL_FPC
 33 *     flag to save and restore floating-point controls in addition to any
 34 *     vector register range.
 35 *
 36 *  6. To use floating-point registers and instructions only, specify the
 37 *     KERNEL_FPR flag.  This flag triggers a save and restore of vector
 38 *     registers V0 to V15 and floating-point controls.
 39 *
 40 * Copyright IBM Corp. 2015
 41 * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
 42 */
 43
 44#ifndef _ASM_S390_FPU_H
 45#define _ASM_S390_FPU_H
 46
 47#include <linux/cpufeature.h>
 48#include <linux/processor.h>
 49#include <linux/preempt.h>
 50#include <linux/string.h>
 51#include <linux/sched.h>
 52#include <asm/sigcontext.h>
 53#include <asm/fpu-types.h>
 54#include <asm/fpu-insn.h>
 55
 56enum {
 57	KERNEL_FPC_BIT = 0,
 58	KERNEL_VXR_V0V7_BIT,
 59	KERNEL_VXR_V8V15_BIT,
 60	KERNEL_VXR_V16V23_BIT,
 61	KERNEL_VXR_V24V31_BIT,
 62};
 63
 64#define KERNEL_FPC		BIT(KERNEL_FPC_BIT)
 65#define KERNEL_VXR_V0V7		BIT(KERNEL_VXR_V0V7_BIT)
 66#define KERNEL_VXR_V8V15	BIT(KERNEL_VXR_V8V15_BIT)
 67#define KERNEL_VXR_V16V23	BIT(KERNEL_VXR_V16V23_BIT)
 68#define KERNEL_VXR_V24V31	BIT(KERNEL_VXR_V24V31_BIT)
 69
 70#define KERNEL_VXR_LOW		(KERNEL_VXR_V0V7   | KERNEL_VXR_V8V15)
 71#define KERNEL_VXR_MID		(KERNEL_VXR_V8V15  | KERNEL_VXR_V16V23)
 72#define KERNEL_VXR_HIGH		(KERNEL_VXR_V16V23 | KERNEL_VXR_V24V31)
 73
 74#define KERNEL_VXR		(KERNEL_VXR_LOW	   | KERNEL_VXR_HIGH)
 75#define KERNEL_FPR		(KERNEL_FPC	   | KERNEL_VXR_LOW)
 76
 77void load_fpu_state(struct fpu *state, int flags);
 78void save_fpu_state(struct fpu *state, int flags);
 79void __kernel_fpu_begin(struct kernel_fpu *state, int flags);
 80void __kernel_fpu_end(struct kernel_fpu *state, int flags);
 81
 82static __always_inline void save_vx_regs(__vector128 *vxrs)
 83{
 84	fpu_vstm(0, 15, &vxrs[0]);
 85	fpu_vstm(16, 31, &vxrs[16]);
 86}
 87
 88static __always_inline void load_vx_regs(__vector128 *vxrs)
 89{
 90	fpu_vlm(0, 15, &vxrs[0]);
 91	fpu_vlm(16, 31, &vxrs[16]);
 92}
 93
 94static __always_inline void __save_fp_regs(freg_t *fprs, unsigned int offset)
 95{
 96	fpu_std(0, &fprs[0 * offset]);
 97	fpu_std(1, &fprs[1 * offset]);
 98	fpu_std(2, &fprs[2 * offset]);
 99	fpu_std(3, &fprs[3 * offset]);
100	fpu_std(4, &fprs[4 * offset]);
101	fpu_std(5, &fprs[5 * offset]);
102	fpu_std(6, &fprs[6 * offset]);
103	fpu_std(7, &fprs[7 * offset]);
104	fpu_std(8, &fprs[8 * offset]);
105	fpu_std(9, &fprs[9 * offset]);
106	fpu_std(10, &fprs[10 * offset]);
107	fpu_std(11, &fprs[11 * offset]);
108	fpu_std(12, &fprs[12 * offset]);
109	fpu_std(13, &fprs[13 * offset]);
110	fpu_std(14, &fprs[14 * offset]);
111	fpu_std(15, &fprs[15 * offset]);
112}
113
114static __always_inline void __load_fp_regs(freg_t *fprs, unsigned int offset)
115{
116	fpu_ld(0, &fprs[0 * offset]);
117	fpu_ld(1, &fprs[1 * offset]);
118	fpu_ld(2, &fprs[2 * offset]);
119	fpu_ld(3, &fprs[3 * offset]);
120	fpu_ld(4, &fprs[4 * offset]);
121	fpu_ld(5, &fprs[5 * offset]);
122	fpu_ld(6, &fprs[6 * offset]);
123	fpu_ld(7, &fprs[7 * offset]);
124	fpu_ld(8, &fprs[8 * offset]);
125	fpu_ld(9, &fprs[9 * offset]);
126	fpu_ld(10, &fprs[10 * offset]);
127	fpu_ld(11, &fprs[11 * offset]);
128	fpu_ld(12, &fprs[12 * offset]);
129	fpu_ld(13, &fprs[13 * offset]);
130	fpu_ld(14, &fprs[14 * offset]);
131	fpu_ld(15, &fprs[15 * offset]);
132}
133
134static __always_inline void save_fp_regs(freg_t *fprs)
135{
136	__save_fp_regs(fprs, sizeof(freg_t) / sizeof(freg_t));
137}
138
139static __always_inline void load_fp_regs(freg_t *fprs)
140{
141	__load_fp_regs(fprs, sizeof(freg_t) / sizeof(freg_t));
142}
143
144static __always_inline void save_fp_regs_vx(__vector128 *vxrs)
145{
146	freg_t *fprs = (freg_t *)&vxrs[0].high;
147
148	__save_fp_regs(fprs, sizeof(__vector128) / sizeof(freg_t));
149}
150
151static __always_inline void load_fp_regs_vx(__vector128 *vxrs)
152{
153	freg_t *fprs = (freg_t *)&vxrs[0].high;
154
155	__load_fp_regs(fprs, sizeof(__vector128) / sizeof(freg_t));
156}
157
158static inline void load_user_fpu_regs(void)
159{
160	struct thread_struct *thread = &current->thread;
161
162	if (!thread->ufpu_flags)
163		return;
164	load_fpu_state(&thread->ufpu, thread->ufpu_flags);
165	thread->ufpu_flags = 0;
166}
167
168static __always_inline void __save_user_fpu_regs(struct thread_struct *thread, int flags)
169{
170	save_fpu_state(&thread->ufpu, flags);
171	__atomic_or(flags, &thread->ufpu_flags);
172}
173
174static inline void save_user_fpu_regs(void)
175{
176	struct thread_struct *thread = &current->thread;
177	int mask, flags;
178
179	mask = __atomic_or(KERNEL_FPC | KERNEL_VXR, &thread->kfpu_flags);
180	flags = ~READ_ONCE(thread->ufpu_flags) & (KERNEL_FPC | KERNEL_VXR);
181	if (flags)
182		__save_user_fpu_regs(thread, flags);
183	barrier();
184	WRITE_ONCE(thread->kfpu_flags, mask);
185}
186
187static __always_inline void _kernel_fpu_begin(struct kernel_fpu *state, int flags)
188{
189	struct thread_struct *thread = &current->thread;
190	int mask, uflags;
191
192	mask = __atomic_or(flags, &thread->kfpu_flags);
193	state->hdr.mask = mask;
194	uflags = READ_ONCE(thread->ufpu_flags);
195	if ((uflags & flags) != flags)
196		__save_user_fpu_regs(thread, ~uflags & flags);
197	if (mask & flags)
198		__kernel_fpu_begin(state, flags);
199}
200
201static __always_inline void _kernel_fpu_end(struct kernel_fpu *state, int flags)
202{
203	int mask = state->hdr.mask;
204
205	if (mask & flags)
206		__kernel_fpu_end(state, flags);
207	barrier();
208	WRITE_ONCE(current->thread.kfpu_flags, mask);
209}
210
211void __kernel_fpu_invalid_size(void);
212
213static __always_inline void kernel_fpu_check_size(int flags, unsigned int size)
214{
215	unsigned int cnt = 0;
216
217	if (flags & KERNEL_VXR_V0V7)
218		cnt += 8;
219	if (flags & KERNEL_VXR_V8V15)
220		cnt += 8;
221	if (flags & KERNEL_VXR_V16V23)
222		cnt += 8;
223	if (flags & KERNEL_VXR_V24V31)
224		cnt += 8;
225	if (cnt != size)
226		__kernel_fpu_invalid_size();
227}
228
229#define kernel_fpu_begin(state, flags)					\
230{									\
231	typeof(state) s = (state);					\
232	int _flags = (flags);						\
233									\
234	kernel_fpu_check_size(_flags, ARRAY_SIZE(s->vxrs));		\
235	_kernel_fpu_begin((struct kernel_fpu *)s, _flags);		\
236}
237
238#define kernel_fpu_end(state, flags)					\
239{									\
240	typeof(state) s = (state);					\
241	int _flags = (flags);						\
242									\
243	kernel_fpu_check_size(_flags, ARRAY_SIZE(s->vxrs));		\
244	_kernel_fpu_end((struct kernel_fpu *)s, _flags);		\
245}
246
247static inline void save_kernel_fpu_regs(struct thread_struct *thread)
248{
249	if (!thread->kfpu_flags)
250		return;
251	save_fpu_state(&thread->kfpu, thread->kfpu_flags);
252}
253
254static inline void restore_kernel_fpu_regs(struct thread_struct *thread)
255{
256	if (!thread->kfpu_flags)
257		return;
258	load_fpu_state(&thread->kfpu, thread->kfpu_flags);
259}
260
261static inline void convert_vx_to_fp(freg_t *fprs, __vector128 *vxrs)
262{
263	int i;
264
265	for (i = 0; i < __NUM_FPRS; i++)
266		fprs[i].ui = vxrs[i].high;
267}
268
269static inline void convert_fp_to_vx(__vector128 *vxrs, freg_t *fprs)
270{
271	int i;
272
273	for (i = 0; i < __NUM_FPRS; i++)
274		vxrs[i].high = fprs[i].ui;
275}
276
277static inline void fpregs_store(_s390_fp_regs *fpregs, struct fpu *fpu)
278{
279	fpregs->pad = 0;
280	fpregs->fpc = fpu->fpc;
281	convert_vx_to_fp((freg_t *)&fpregs->fprs, fpu->vxrs);
282}
283
284static inline void fpregs_load(_s390_fp_regs *fpregs, struct fpu *fpu)
285{
286	fpu->fpc = fpregs->fpc;
287	convert_fp_to_vx(fpu->vxrs, (freg_t *)&fpregs->fprs);
288}
289
290#endif /* _ASM_S390_FPU_H */