1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
|
--- sys/amd64/amd64/cpu_switch.S.orig
+++ sys/amd64/amd64/cpu_switch.S
@@ -105,10 +105,10 @@
/* have we used fp, and need a save? */
cmpq %rdi,PCPU(FPCURTHREAD)
- jne 3f
+ jne 2f
movq PCB_SAVEFPU(%r8),%r8
clts
- cmpl $0,use_xsave
+ cmpl $0,use_xsave(%rip)
jne 1f
fxsave (%r8)
jmp 2f
@@ -120,12 +120,7 @@
/* This is patched to xsaveopt if supported, see fpuinit_bsp1() */
xsave (%r8)
movq %rcx,%rdx
-2: smsw %ax
- orb $CR0_TS,%al
- lmsw %ax
- xorl %eax,%eax
- movq %rax,PCPU(FPCURTHREAD)
-3:
+2:
/* Save is done. Now fire up new thread. Leave old vmspace. */
movq %rsi,%r12
movq %rdi,%r13
@@ -212,6 +207,8 @@
movq PCB_RBX(%r8),%rbx
movq PCB_RIP(%r8),%rax
movq %rax,(%rsp)
+ movq PCPU(CURTHREAD),%rdi
+ call fpu_activate_sw
ret
/*
--- sys/amd64/amd64/fpu.c.orig
+++ sys/amd64/amd64/fpu.c
@@ -139,6 +139,11 @@
SYSCTL_INT(_hw, HW_FLOATINGPT, floatingpoint, CTLFLAG_RD,
SYSCTL_NULL_INT_PTR, 1, "Floating point instructions executed in hardware");
+int lazy_fpu_switch = 0;
+SYSCTL_INT(_hw, OID_AUTO, lazy_fpu_switch, CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
+ &lazy_fpu_switch, 0,
+ "Lazily load FPU context after context switch");
+
int use_xsave; /* non-static for cpu_switch.S */
uint64_t xsave_mask; /* the same */
static uma_zone_t fpu_save_area_zone;
@@ -204,6 +209,7 @@
u_int cp[4];
uint64_t xsave_mask_user;
+ TUNABLE_INT_FETCH("hw.lazy_fpu_switch", &lazy_fpu_switch);
if ((cpu_feature2 & CPUID2_XSAVE) != 0) {
use_xsave = 1;
TUNABLE_INT_FETCH("hw.use_xsave", &use_xsave);
@@ -611,6 +617,45 @@
return (fpetable[(mxcsr & (~mxcsr >> 7)) & 0x3f]);
}
+static void
+restore_fpu_curthread(struct thread *td)
+{
+ struct pcb *pcb;
+
+ /*
+ * Record new context early in case frstor causes a trap.
+ */
+ PCPU_SET(fpcurthread, td);
+
+ stop_emulating();
+ fpu_clean_state();
+ pcb = td->td_pcb;
+
+ if ((pcb->pcb_flags & PCB_FPUINITDONE) == 0) {
+ /*
+ * This is the first time this thread has used the FPU or
+ * the PCB doesn't contain a clean FPU state. Explicitly
+ * load an initial state.
+ *
+ * We prefer to restore the state from the actual save
+ * area in PCB instead of directly loading from
+ * fpu_initialstate, to ignite the XSAVEOPT
+ * tracking engine.
+ */
+ bcopy(fpu_initialstate, pcb->pcb_save,
+ cpu_max_ext_state_size);
+ fpurestore(pcb->pcb_save);
+ if (pcb->pcb_initial_fpucw != __INITIAL_FPUCW__)
+ fldcw(pcb->pcb_initial_fpucw);
+ if (PCB_USER_FPU(pcb))
+ set_pcb_flags(pcb, PCB_FPUINITDONE |
+ PCB_USERFPUINITDONE);
+ else
+ set_pcb_flags(pcb, PCB_FPUINITDONE);
+ } else
+ fpurestore(pcb->pcb_save);
+}
+
/*
* Device Not Available (DNA, #NM) exception handler.
*
@@ -621,7 +666,9 @@
void
fpudna(void)
{
+ struct thread *td;
+ td = curthread;
/*
* This handler is entered with interrupts enabled, so context
* switches may occur before critical_enter() is executed. If
@@ -635,49 +682,38 @@
KASSERT((curpcb->pcb_flags & PCB_FPUNOSAVE) == 0,
("fpudna while in fpu_kern_enter(FPU_KERN_NOCTX)"));
- if (PCPU_GET(fpcurthread) == curthread) {
- printf("fpudna: fpcurthread == curthread\n");
+ if (__predict_false(PCPU_GET(fpcurthread) == td)) {
+ /*
+ * Some virtual machines seems to set %cr0.TS at
+ * arbitrary moments. Silently clear the TS bit
+ * regardless of the eager/lazy FPU context switch
+ * mode.
+ */
stop_emulating();
- critical_exit();
- return;
+ } else {
+ if (__predict_false(PCPU_GET(fpcurthread) != NULL)) {
+ panic(
+ "fpudna: fpcurthread = %p (%d), curthread = %p (%d)\n",
+ PCPU_GET(fpcurthread),
+ PCPU_GET(fpcurthread)->td_tid, td, td->td_tid);
+ }
+ restore_fpu_curthread(td);
}
- if (PCPU_GET(fpcurthread) != NULL) {
- panic("fpudna: fpcurthread = %p (%d), curthread = %p (%d)\n",
- PCPU_GET(fpcurthread), PCPU_GET(fpcurthread)->td_tid,
- curthread, curthread->td_tid);
- }
- stop_emulating();
- /*
- * Record new context early in case frstor causes a trap.
- */
- PCPU_SET(fpcurthread, curthread);
+ critical_exit();
+}
- fpu_clean_state();
+void fpu_activate_sw(struct thread *td); /* Called from the context switch */
+void
+fpu_activate_sw(struct thread *td)
+{
- if ((curpcb->pcb_flags & PCB_FPUINITDONE) == 0) {
- /*
- * This is the first time this thread has used the FPU or
- * the PCB doesn't contain a clean FPU state. Explicitly
- * load an initial state.
- *
- * We prefer to restore the state from the actual save
- * area in PCB instead of directly loading from
- * fpu_initialstate, to ignite the XSAVEOPT
- * tracking engine.
- */
- bcopy(fpu_initialstate, curpcb->pcb_save,
- cpu_max_ext_state_size);
- fpurestore(curpcb->pcb_save);
- if (curpcb->pcb_initial_fpucw != __INITIAL_FPUCW__)
- fldcw(curpcb->pcb_initial_fpucw);
- if (PCB_USER_FPU(curpcb))
- set_pcb_flags(curpcb,
- PCB_FPUINITDONE | PCB_USERFPUINITDONE);
- else
- set_pcb_flags(curpcb, PCB_FPUINITDONE);
- } else
- fpurestore(curpcb->pcb_save);
- critical_exit();
+ if (lazy_fpu_switch || (td->td_pflags & TDP_KTHREAD) != 0 ||
+ !PCB_USER_FPU(td->td_pcb)) {
+ PCPU_SET(fpcurthread, NULL);
+ start_emulating();
+ } else if (PCPU_GET(fpcurthread) != td) {
+ restore_fpu_curthread(td);
+ }
}
void
--- sys/i386/i386/swtch.s.orig
+++ sys/i386/i386/swtch.s
@@ -293,6 +293,12 @@
cpu_switch_load_gs:
mov PCB_GS(%edx),%gs
+ pushl %edx
+ pushl PCPU(CURTHREAD)
+ call npxswitch
+ popl %edx
+ popl %edx
+
/* Test if debug registers should be restored. */
testl $PCB_DBREGS,PCB_FLAGS(%edx)
jz 1f
--- sys/i386/isa/npx.c.orig
+++ sys/i386/isa/npx.c
@@ -191,6 +191,11 @@
SYSCTL_INT(_hw, HW_FLOATINGPT, floatingpoint, CTLFLAG_RD,
&hw_float, 0, "Floating point instructions executed in hardware");
+int lazy_fpu_switch = 0;
+SYSCTL_INT(_hw, OID_AUTO, lazy_fpu_switch, CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
+ &lazy_fpu_switch, 0,
+ "Lazily load FPU context after context switch");
+
int use_xsave;
uint64_t xsave_mask;
static uma_zone_t fpu_save_area_zone;
@@ -327,6 +332,7 @@
u_int cp[4];
uint64_t xsave_mask_user;
+ TUNABLE_INT_FETCH("hw.lazy_fpu_switch", &lazy_fpu_switch);
if (cpu_fxsr && (cpu_feature2 & CPUID2_XSAVE) != 0) {
use_xsave = 1;
TUNABLE_INT_FETCH("hw.use_xsave", &use_xsave);
@@ -785,47 +791,20 @@
return (fpetable[(mxcsr & (~mxcsr >> 7)) & 0x3f]);
}
-/*
- * Implement device not available (DNA) exception
- *
- * It would be better to switch FP context here (if curthread != fpcurthread)
- * and not necessarily for every context switch, but it is too hard to
- * access foreign pcb's.
- */
-
-static int err_count = 0;
-
-int
-npxdna(void)
+static void
+restore_npx_curthread(struct thread *td, struct pcb *pcb)
{
- if (!hw_float)
- return (0);
- critical_enter();
- if (PCPU_GET(fpcurthread) == curthread) {
- printf("npxdna: fpcurthread == curthread %d times\n",
- ++err_count);
- stop_emulating();
- critical_exit();
- return (1);
- }
- if (PCPU_GET(fpcurthread) != NULL) {
- printf("npxdna: fpcurthread = %p (%d), curthread = %p (%d)\n",
- PCPU_GET(fpcurthread),
- PCPU_GET(fpcurthread)->td_proc->p_pid,
- curthread, curthread->td_proc->p_pid);
- panic("npxdna");
- }
- stop_emulating();
/*
* Record new context early in case frstor causes a trap.
*/
- PCPU_SET(fpcurthread, curthread);
+ PCPU_SET(fpcurthread, td);
+ stop_emulating();
if (cpu_fxsr)
fpu_clean_state();
- if ((curpcb->pcb_flags & PCB_NPXINITDONE) == 0) {
+ if ((pcb->pcb_flags & PCB_NPXINITDONE) == 0) {
/*
* This is the first time this thread has used the FPU or
* the PCB doesn't contain a clean FPU state. Explicitly
@@ -836,18 +815,54 @@
* npx_initialstate, to ignite the XSAVEOPT
* tracking engine.
*/
- bcopy(npx_initialstate, curpcb->pcb_save, cpu_max_ext_state_size);
- fpurstor(curpcb->pcb_save);
- if (curpcb->pcb_initial_npxcw != __INITIAL_NPXCW__)
- fldcw(curpcb->pcb_initial_npxcw);
- curpcb->pcb_flags |= PCB_NPXINITDONE;
- if (PCB_USER_FPU(curpcb))
- curpcb->pcb_flags |= PCB_NPXUSERINITDONE;
+ bcopy(npx_initialstate, pcb->pcb_save, cpu_max_ext_state_size);
+ fpurstor(pcb->pcb_save);
+ if (pcb->pcb_initial_npxcw != __INITIAL_NPXCW__)
+ fldcw(pcb->pcb_initial_npxcw);
+ pcb->pcb_flags |= PCB_NPXINITDONE;
+ if (PCB_USER_FPU(pcb))
+ pcb->pcb_flags |= PCB_NPXUSERINITDONE;
} else {
- fpurstor(curpcb->pcb_save);
+ fpurstor(pcb->pcb_save);
}
+}
+
+/*
+ * Implement device not available (DNA) exception
+ *
+ * It would be better to switch FP context here (if curthread != fpcurthread)
+ * and not necessarily for every context switch, but it is too hard to
+ * access foreign pcb's.
+ */
+int
+npxdna(void)
+{
+ struct thread *td;
+
+ if (!hw_float)
+ return (0);
+ td = curthread;
+ critical_enter();
+ if (__predict_false(PCPU_GET(fpcurthread) == td)) {
+ /*
+ * Some virtual machines seems to set %cr0.TS at
+ * arbitrary moments. Silently clear the TS bit
+ * regardless of the eager/lazy FPU context switch
+ * mode.
+ */
+ stop_emulating();
+ } else {
+ if (__predict_false(PCPU_GET(fpcurthread) != NULL)) {
+ printf(
+ "npxdna: fpcurthread = %p (%d), curthread = %p (%d)\n",
+ PCPU_GET(fpcurthread),
+ PCPU_GET(fpcurthread)->td_proc->p_pid,
+ td, td->td_proc->p_pid);
+ panic("npxdna");
+ }
+ restore_npx_curthread(td, td->td_pcb);
+ }
critical_exit();
-
return (1);
}
@@ -869,10 +884,22 @@
xsaveopt((char *)addr, xsave_mask);
else
fpusave(addr);
- start_emulating();
- PCPU_SET(fpcurthread, NULL);
}
+void npxswitch(struct thread *td, struct pcb *pcb);
+void
+npxswitch(struct thread *td, struct pcb *pcb)
+{
+
+ if (lazy_fpu_switch || (td->td_pflags & TDP_KTHREAD) != 0 ||
+ !PCB_USER_FPU(pcb)) {
+ start_emulating();
+ PCPU_SET(fpcurthread, NULL);
+ } else if (PCPU_GET(fpcurthread) != td) {
+ restore_npx_curthread(td, pcb);
+ }
+}
+
/*
* Unconditionally save the current co-processor state across suspend and
* resume.
|
