author | jpk |
Fri, 24 Mar 2006 12:29:20 -0800 | |
changeset 1676 | 37f4a3e2bd99 |
parent 1058 | 0d3e7ce75bef |
child 3426 | e69c0764a03e |
permissions | -rw-r--r-- |
0 | 1 |
/* |
2 |
* CDDL HEADER START |
|
3 |
* |
|
4 |
* The contents of this file are subject to the terms of the |
|
5 |
* Common Development and Distribution License, Version 1.0 only |
|
6 |
* (the "License"). You may not use this file except in compliance |
|
7 |
* with the License. |
|
8 |
* |
|
9 |
* You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE |
|
10 |
* or http://www.opensolaris.org/os/licensing. |
|
11 |
* See the License for the specific language governing permissions |
|
12 |
* and limitations under the License. |
|
13 |
* |
|
14 |
* When distributing Covered Code, include this CDDL HEADER in each |
|
15 |
* file and include the License file at usr/src/OPENSOLARIS.LICENSE. |
|
16 |
* If applicable, add the following below this CDDL HEADER, with the |
|
17 |
* fields enclosed by brackets "[]" replaced with your own identifying |
|
18 |
* information: Portions Copyright [yyyy] [name of copyright owner] |
|
19 |
* |
|
20 |
* CDDL HEADER END |
|
21 |
*/ |
|
22 |
/* |
|
23 |
* Copyright 2005 Sun Microsystems, Inc. All rights reserved. |
|
24 |
* Use is subject to license terms. |
|
25 |
*/ |
|
26 |
||
27 |
#pragma ident "%Z%%M% %I% %E% SMI" |
|
28 |
||
29 |
#include <sys/types.h> |
|
30 |
#include <sys/param.h> |
|
31 |
#include <sys/systm.h> |
|
32 |
#include <sys/user.h> |
|
33 |
#include <sys/proc.h> |
|
34 |
#include <sys/cpuvar.h> |
|
35 |
#include <sys/thread.h> |
|
36 |
#include <sys/debug.h> |
|
37 |
#include <sys/msacct.h> |
|
38 |
#include <sys/time.h> |
|
39 |
||
40 |
/* |
|
41 |
* Mega-theory block comment: |
|
42 |
* |
|
43 |
* Microstate accounting uses finite states and the transitions between these |
|
44 |
* states to measure timing and accounting information. The state information |
|
45 |
* is presently tracked for threads (via microstate accounting) and cpus (via |
|
46 |
* cpu microstate accounting). In each case, these accounting mechanisms use |
|
47 |
* states and transitions to measure time spent in each state instead of |
|
48 |
* clock-based sampling methodologies. |
|
49 |
* |
|
50 |
* For microstate accounting: |
|
51 |
* state transitions are accomplished by calling new_mstate() to switch between |
|
52 |
* states. Transitions from a sleeping state (LMS_SLEEP and LMS_STOPPED) occur |
|
53 |
* by calling restore_mstate() which restores a thread to its previously running |
|
54 |
* state. This code is primarialy executed by the dispatcher in disp() before |
|
55 |
* running a process that was put to sleep. If the thread was not in a sleeping |
|
56 |
* state, this call has little effect other than to update the count of time the |
|
57 |
* thread has spent waiting on run-queues in its lifetime. |
|
58 |
* |
|
59 |
* For cpu microstate accounting: |
|
60 |
* Cpu microstate accounting is similar to the microstate accounting for threads |
|
61 |
* but it tracks user, system, and idle time for cpus. Cpu microstate |
|
62 |
* accounting does not track interrupt times as there is a pre-existing |
|
63 |
* interrupt accounting mechanism for this purpose. Cpu microstate accounting |
|
64 |
* tracks time that user threads have spent active, idle, or in the system on a |
|
65 |
* given cpu. Cpu microstate accounting has fewer states which allows it to |
|
66 |
* have better defined transitions. The states transition in the following |
|
67 |
* order: |
|
68 |
* |
|
69 |
* CMS_USER <-> CMS_SYSTEM <-> CMS_IDLE |
|
70 |
* |
|
71 |
* In order to get to the idle state, the cpu microstate must first go through |
|
72 |
* the system state, and vice-versa for the user state from idle. The switching |
|
73 |
* of the microstates from user to system is done as part of the regular thread |
|
74 |
* microstate accounting code, except for the idle state which is switched by |
|
75 |
* the dispatcher before it runs the idle loop. |
|
76 |
* |
|
77 |
* Cpu percentages: |
|
78 |
* Cpu percentages are now handled by and based upon microstate accounting |
|
79 |
* information (the same is true for load averages). The routines which handle |
|
80 |
* the growing/shrinking and exponentiation of cpu percentages have been moved |
|
81 |
* here as it now makes more sense for them to be generated from the microstate |
|
82 |
* code. Cpu percentages are generated similarly to the way they were before; |
|
83 |
* however, now they are based upon high-resolution timestamps and the |
|
84 |
* timestamps are modified at various state changes instead of during a clock() |
|
85 |
* interrupt. This allows us to generate more accurate cpu percentages which |
|
86 |
* are also in-sync with microstate data. |
|
87 |
*/ |
|
88 |
||
89 |
/* |
|
90 |
* Initialize the microstate level and the |
|
91 |
* associated accounting information for an LWP. |
|
92 |
*/ |
|
93 |
void |
|
94 |
init_mstate( |
|
95 |
kthread_t *t, |
|
96 |
int init_state) |
|
97 |
{ |
|
98 |
struct mstate *ms; |
|
99 |
klwp_t *lwp; |
|
100 |
hrtime_t curtime; |
|
101 |
||
102 |
ASSERT(init_state != LMS_WAIT_CPU); |
|
103 |
ASSERT((unsigned)init_state < NMSTATES); |
|
104 |
||
105 |
if ((lwp = ttolwp(t)) != NULL) { |
|
106 |
ms = &lwp->lwp_mstate; |
|
107 |
curtime = gethrtime_unscaled(); |
|
108 |
ms->ms_prev = LMS_SYSTEM; |
|
109 |
ms->ms_start = curtime; |
|
110 |
ms->ms_term = 0; |
|
111 |
ms->ms_state_start = curtime; |
|
112 |
t->t_mstate = init_state; |
|
113 |
t->t_waitrq = 0; |
|
114 |
t->t_hrtime = curtime; |
|
115 |
if ((t->t_proc_flag & TP_MSACCT) == 0) |
|
116 |
t->t_proc_flag |= TP_MSACCT; |
|
117 |
bzero((caddr_t)&ms->ms_acct[0], sizeof (ms->ms_acct)); |
|
118 |
} |
|
119 |
} |
|
120 |
||
121 |
/* |
|
122 |
* Initialize the microstate level and associated accounting information |
|
123 |
* for the specified cpu |
|
124 |
*/ |
|
125 |
||
126 |
void |
|
127 |
init_cpu_mstate( |
|
128 |
cpu_t *cpu, |
|
129 |
int init_state) |
|
130 |
{ |
|
131 |
ASSERT(init_state != CMS_DISABLED); |
|
132 |
||
133 |
cpu->cpu_mstate = init_state; |
|
134 |
cpu->cpu_mstate_start = gethrtime_unscaled(); |
|
135 |
cpu->cpu_waitrq = 0; |
|
136 |
bzero((caddr_t)&cpu->cpu_acct[0], sizeof (cpu->cpu_acct)); |
|
137 |
} |
|
138 |
||
139 |
/* |
|
140 |
* sets cpu state to OFFLINE. We don't actually track this time, |
|
141 |
* but it serves as a useful placeholder state for when we're not |
|
142 |
* doing anything. |
|
143 |
*/ |
|
144 |
||
145 |
void |
|
146 |
term_cpu_mstate(struct cpu *cpu) |
|
147 |
{ |
|
148 |
ASSERT(cpu->cpu_mstate != CMS_DISABLED); |
|
149 |
cpu->cpu_mstate = CMS_DISABLED; |
|
150 |
cpu->cpu_mstate_start = 0; |
|
151 |
} |
|
152 |
||
1058 | 153 |
/* NEW_CPU_MSTATE comments inline in new_cpu_mstate below. */ |
154 |
||
155 |
#define NEW_CPU_MSTATE(state) \ |
|
156 |
gen = cpu->cpu_mstate_gen; \ |
|
157 |
cpu->cpu_mstate_gen = 0; \ |
|
158 |
/* Need membar_producer() here if stores not ordered / TSO */ \ |
|
159 |
cpu->cpu_acct[cpu->cpu_mstate] += curtime - cpu->cpu_mstate_start; \ |
|
160 |
cpu->cpu_mstate = state; \ |
|
161 |
cpu->cpu_mstate_start = curtime; \ |
|
162 |
/* Need membar_producer() here if stores not ordered / TSO */ \ |
|
163 |
cpu->cpu_mstate_gen = (++gen == 0) ? 1 : gen; |
|
164 |
||
0 | 165 |
void |
590
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
166 |
new_cpu_mstate(int cmstate, hrtime_t curtime) |
0 | 167 |
{ |
590
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
168 |
cpu_t *cpu = CPU; |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
169 |
uint16_t gen; |
0 | 170 |
|
171 |
ASSERT(cpu->cpu_mstate != CMS_DISABLED); |
|
172 |
ASSERT(cmstate < NCMSTATES); |
|
173 |
ASSERT(cmstate != CMS_DISABLED); |
|
590
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
174 |
|
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
175 |
/* |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
176 |
* This function cannot be re-entrant on a given CPU. As such, |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
177 |
* we ASSERT and panic if we are called on behalf of an interrupt. |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
178 |
* The one exception is for an interrupt which has previously |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
179 |
* blocked. Such an interrupt is being scheduled by the dispatcher |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
180 |
* just like a normal thread, and as such cannot arrive here |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
181 |
* in a re-entrant manner. |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
182 |
*/ |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
183 |
|
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
184 |
ASSERT(!CPU_ON_INTR(cpu) && curthread->t_intr == NULL); |
0 | 185 |
ASSERT(curthread->t_preempt > 0 || curthread == cpu->cpu_idle_thread); |
186 |
||
590
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
187 |
/* |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
188 |
* LOCKING, or lack thereof: |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
189 |
* |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
190 |
* Updates to CPU mstate can only be made by the CPU |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
191 |
* itself, and the above check to ignore interrupts |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
192 |
* should prevent recursion into this function on a given |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
193 |
* processor. i.e. no possible write contention. |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
194 |
* |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
195 |
* However, reads of CPU mstate can occur at any time |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
196 |
* from any CPU. Any locking added to this code path |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
197 |
* would seriously impact syscall performance. So, |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
198 |
* instead we have a best-effort protection for readers. |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
199 |
* The reader will want to account for any time between |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
200 |
* cpu_mstate_start and the present time. This requires |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
201 |
* some guarantees that the reader is getting coherent |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
202 |
* information. |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
203 |
* |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
204 |
* We use a generation counter, which is set to 0 before |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
205 |
* we start making changes, and is set to a new value |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
206 |
* after we're done. Someone reading the CPU mstate |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
207 |
* should check for the same non-zero value of this |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
208 |
* counter both before and after reading all state. The |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
209 |
* important point is that the reader is not a |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
210 |
* performance-critical path, but this function is. |
1058 | 211 |
* |
212 |
* The ordering of writes is critical. cpu_mstate_gen must |
|
213 |
* be visibly zero on all CPUs before we change cpu_mstate |
|
214 |
* and cpu_mstate_start. Additionally, cpu_mstate_gen must |
|
215 |
* not be restored to oldgen+1 until after all of the other |
|
216 |
* writes have become visible. |
|
217 |
* |
|
218 |
* Normally one puts membar_producer() calls to accomplish |
|
219 |
* this. Unfortunately this routine is extremely performance |
|
220 |
* critical (esp. in syscall_mstate below) and we cannot |
|
221 |
* afford the additional time, particularly on some x86 |
|
222 |
* architectures with extremely slow sfence calls. On a |
|
223 |
* CPU which guarantees write ordering (including sparc, x86, |
|
224 |
* and amd64) this is not a problem. The compiler could still |
|
225 |
* reorder the writes, so we make the four cpu fields |
|
226 |
* volatile to prevent this. |
|
227 |
* |
|
228 |
* TSO warning: should we port to a non-TSO (or equivalent) |
|
229 |
* CPU, this will break. |
|
230 |
* |
|
231 |
* The reader stills needs the membar_consumer() calls because, |
|
232 |
* although the volatiles prevent the compiler from reordering |
|
233 |
* loads, the CPU can still do so. |
|
590
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
234 |
*/ |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
235 |
|
1058 | 236 |
NEW_CPU_MSTATE(cmstate); |
0 | 237 |
} |
238 |
||
239 |
/* |
|
240 |
* Return an aggregation of microstate times in scaled nanoseconds (high-res |
|
241 |
* time). This keeps in mind that p_acct is already scaled, and ms_acct is |
|
242 |
* not. |
|
243 |
*/ |
|
244 |
hrtime_t |
|
245 |
mstate_aggr_state(proc_t *p, int a_state) |
|
246 |
{ |
|
247 |
struct mstate *ms; |
|
248 |
kthread_t *t; |
|
249 |
klwp_t *lwp; |
|
250 |
hrtime_t aggr_time; |
|
251 |
hrtime_t scaledtime; |
|
252 |
||
253 |
ASSERT(MUTEX_HELD(&p->p_lock)); |
|
254 |
ASSERT((unsigned)a_state < NMSTATES); |
|
255 |
||
256 |
aggr_time = p->p_acct[a_state]; |
|
257 |
if (a_state == LMS_SYSTEM) |
|
258 |
aggr_time += p->p_acct[LMS_TRAP]; |
|
259 |
||
260 |
t = p->p_tlist; |
|
261 |
if (t == NULL) |
|
262 |
return (aggr_time); |
|
263 |
||
264 |
do { |
|
265 |
if (t->t_proc_flag & TP_LWPEXIT) |
|
266 |
continue; |
|
267 |
||
268 |
lwp = ttolwp(t); |
|
269 |
ms = &lwp->lwp_mstate; |
|
270 |
scaledtime = ms->ms_acct[a_state]; |
|
271 |
scalehrtime(&scaledtime); |
|
272 |
aggr_time += scaledtime; |
|
273 |
if (a_state == LMS_SYSTEM) { |
|
274 |
scaledtime = ms->ms_acct[LMS_TRAP]; |
|
275 |
scalehrtime(&scaledtime); |
|
276 |
aggr_time += scaledtime; |
|
277 |
} |
|
278 |
} while ((t = t->t_forw) != p->p_tlist); |
|
279 |
||
280 |
return (aggr_time); |
|
281 |
} |
|
282 |
||
1058 | 283 |
|
0 | 284 |
void |
285 |
syscall_mstate(int fromms, int toms) |
|
286 |
{ |
|
287 |
kthread_t *t = curthread; |
|
288 |
struct mstate *ms; |
|
289 |
hrtime_t *mstimep; |
|
290 |
hrtime_t curtime; |
|
291 |
klwp_t *lwp; |
|
292 |
hrtime_t newtime; |
|
1058 | 293 |
cpu_t *cpu; |
294 |
uint16_t gen; |
|
0 | 295 |
|
296 |
if ((lwp = ttolwp(t)) == NULL) |
|
297 |
return; |
|
298 |
||
299 |
ASSERT(fromms < NMSTATES); |
|
300 |
ASSERT(toms < NMSTATES); |
|
301 |
||
302 |
ms = &lwp->lwp_mstate; |
|
303 |
mstimep = &ms->ms_acct[fromms]; |
|
304 |
curtime = gethrtime_unscaled(); |
|
305 |
newtime = curtime - ms->ms_state_start; |
|
306 |
while (newtime < 0) { |
|
307 |
curtime = gethrtime_unscaled(); |
|
308 |
newtime = curtime - ms->ms_state_start; |
|
309 |
} |
|
310 |
*mstimep += newtime; |
|
311 |
t->t_mstate = toms; |
|
312 |
ms->ms_state_start = curtime; |
|
313 |
ms->ms_prev = fromms; |
|
590
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
314 |
kpreempt_disable(); /* don't change CPU while changing CPU's state */ |
1058 | 315 |
cpu = CPU; |
316 |
ASSERT(cpu == t->t_cpu); |
|
317 |
if ((toms != LMS_USER) && (cpu->cpu_mstate != CMS_SYSTEM)) { |
|
318 |
NEW_CPU_MSTATE(CMS_SYSTEM); |
|
319 |
} else if ((toms == LMS_USER) && (cpu->cpu_mstate != CMS_USER)) { |
|
320 |
NEW_CPU_MSTATE(CMS_USER); |
|
321 |
} |
|
0 | 322 |
kpreempt_enable(); |
323 |
} |
|
324 |
||
1058 | 325 |
#undef NEW_CPU_MSTATE |
326 |
||
0 | 327 |
/* |
328 |
* The following is for computing the percentage of cpu time used recently |
|
329 |
* by an lwp. The function cpu_decay() is also called from /proc code. |
|
330 |
* |
|
331 |
* exp_x(x): |
|
332 |
* Given x as a 64-bit non-negative scaled integer of arbitrary magnitude, |
|
333 |
* Return exp(-x) as a 64-bit scaled integer in the range [0 .. 1]. |
|
334 |
* |
|
335 |
* Scaling for 64-bit scaled integer: |
|
336 |
* The binary point is to the right of the high-order bit |
|
337 |
* of the low-order 32-bit word. |
|
338 |
*/ |
|
339 |
||
340 |
#define LSHIFT 31 |
|
341 |
#define LSI_ONE ((uint32_t)1 << LSHIFT) /* 32-bit scaled integer 1 */ |
|
342 |
||
343 |
#ifdef DEBUG |
|
344 |
uint_t expx_cnt = 0; /* number of calls to exp_x() */ |
|
345 |
uint_t expx_mul = 0; /* number of long multiplies in exp_x() */ |
|
346 |
#endif |
|
347 |
||
348 |
static uint64_t |
|
349 |
exp_x(uint64_t x) |
|
350 |
{ |
|
351 |
int i; |
|
352 |
uint64_t ull; |
|
353 |
uint32_t ui; |
|
354 |
||
355 |
#ifdef DEBUG |
|
356 |
expx_cnt++; |
|
357 |
#endif |
|
358 |
/* |
|
359 |
* By the formula: |
|
360 |
* exp(-x) = exp(-x/2) * exp(-x/2) |
|
361 |
* we keep halving x until it becomes small enough for |
|
362 |
* the following approximation to be accurate enough: |
|
363 |
* exp(-x) = 1 - x |
|
364 |
* We reduce x until it is less than 1/4 (the 2 in LSHIFT-2 below). |
|
365 |
* Our final error will be smaller than 4% . |
|
366 |
*/ |
|
367 |
||
368 |
/* |
|
369 |
* Use a uint64_t for the initial shift calculation. |
|
370 |
*/ |
|
371 |
ull = x >> (LSHIFT-2); |
|
372 |
||
373 |
/* |
|
374 |
* Short circuit: |
|
375 |
* A number this large produces effectively 0 (actually .005). |
|
376 |
* This way, we will never do more than 5 multiplies. |
|
377 |
*/ |
|
378 |
if (ull >= (1 << 5)) |
|
379 |
return (0); |
|
380 |
||
381 |
ui = ull; /* OK. Now we can use a uint_t. */ |
|
382 |
for (i = 0; ui != 0; i++) |
|
383 |
ui >>= 1; |
|
384 |
||
385 |
if (i != 0) { |
|
386 |
#ifdef DEBUG |
|
387 |
expx_mul += i; /* seldom happens */ |
|
388 |
#endif |
|
389 |
x >>= i; |
|
390 |
} |
|
391 |
||
392 |
/* |
|
393 |
* Now we compute 1 - x and square it the number of times |
|
394 |
* that we halved x above to produce the final result: |
|
395 |
*/ |
|
396 |
x = LSI_ONE - x; |
|
397 |
while (i--) |
|
398 |
x = (x * x) >> LSHIFT; |
|
399 |
||
400 |
return (x); |
|
401 |
} |
|
402 |
||
403 |
/* |
|
404 |
* Given the old percent cpu and a time delta in nanoseconds, |
|
405 |
* return the new decayed percent cpu: pct * exp(-tau), |
|
406 |
* where 'tau' is the time delta multiplied by a decay factor. |
|
407 |
* We have chosen the decay factor (cpu_decay_factor in param.c) |
|
408 |
* to make the decay over five seconds be approximately 20%. |
|
409 |
* |
|
410 |
* 'pct' is a 32-bit scaled integer <= 1 |
|
411 |
* The binary point is to the right of the high-order bit |
|
412 |
* of the 32-bit word. |
|
413 |
*/ |
|
414 |
static uint32_t |
|
415 |
cpu_decay(uint32_t pct, hrtime_t nsec) |
|
416 |
{ |
|
417 |
uint64_t delta = (uint64_t)nsec; |
|
418 |
||
419 |
delta /= cpu_decay_factor; |
|
420 |
return ((pct * exp_x(delta)) >> LSHIFT); |
|
421 |
} |
|
422 |
||
423 |
/* |
|
424 |
* Given the old percent cpu and a time delta in nanoseconds, |
|
425 |
* return the new grown percent cpu: 1 - ( 1 - pct ) * exp(-tau) |
|
426 |
*/ |
|
427 |
static uint32_t |
|
428 |
cpu_grow(uint32_t pct, hrtime_t nsec) |
|
429 |
{ |
|
430 |
return (LSI_ONE - cpu_decay(LSI_ONE - pct, nsec)); |
|
431 |
} |
|
432 |
||
433 |
||
434 |
/* |
|
435 |
* Defined to determine whether a lwp is still on a processor. |
|
436 |
*/ |
|
437 |
||
438 |
#define T_ONPROC(kt) \ |
|
439 |
((kt)->t_mstate < LMS_SLEEP) |
|
440 |
#define T_OFFPROC(kt) \ |
|
441 |
((kt)->t_mstate >= LMS_SLEEP) |
|
442 |
||
443 |
uint_t |
|
444 |
cpu_update_pct(kthread_t *t, hrtime_t newtime) |
|
445 |
{ |
|
446 |
hrtime_t delta; |
|
447 |
hrtime_t hrlb; |
|
448 |
uint_t pctcpu; |
|
449 |
uint_t npctcpu; |
|
450 |
||
451 |
/* |
|
452 |
* This routine can get called at PIL > 0, this *has* to be |
|
453 |
* done atomically. Holding locks here causes bad things to happen. |
|
454 |
* (read: deadlock). |
|
455 |
*/ |
|
456 |
||
457 |
do { |
|
458 |
if (T_ONPROC(t) && t->t_waitrq == 0) { |
|
459 |
hrlb = t->t_hrtime; |
|
460 |
delta = newtime - hrlb; |
|
461 |
if (delta < 0) { |
|
462 |
newtime = gethrtime_unscaled(); |
|
463 |
delta = newtime - hrlb; |
|
464 |
} |
|
465 |
t->t_hrtime = newtime; |
|
466 |
scalehrtime(&delta); |
|
467 |
pctcpu = t->t_pctcpu; |
|
468 |
npctcpu = cpu_grow(pctcpu, delta); |
|
469 |
} else { |
|
470 |
hrlb = t->t_hrtime; |
|
471 |
delta = newtime - hrlb; |
|
472 |
if (delta < 0) { |
|
473 |
newtime = gethrtime_unscaled(); |
|
474 |
delta = newtime - hrlb; |
|
475 |
} |
|
476 |
t->t_hrtime = newtime; |
|
477 |
scalehrtime(&delta); |
|
478 |
pctcpu = t->t_pctcpu; |
|
479 |
npctcpu = cpu_decay(pctcpu, delta); |
|
480 |
} |
|
481 |
} while (cas32(&t->t_pctcpu, pctcpu, npctcpu) != pctcpu); |
|
482 |
||
483 |
return (npctcpu); |
|
484 |
} |
|
485 |
||
486 |
/* |
|
487 |
* Change the microstate level for the LWP and update the |
|
488 |
* associated accounting information. Return the previous |
|
489 |
* LWP state. |
|
490 |
*/ |
|
491 |
int |
|
492 |
new_mstate(kthread_t *t, int new_state) |
|
493 |
{ |
|
494 |
struct mstate *ms; |
|
495 |
unsigned state; |
|
496 |
hrtime_t *mstimep; |
|
497 |
hrtime_t curtime; |
|
498 |
hrtime_t newtime; |
|
499 |
hrtime_t oldtime; |
|
500 |
klwp_t *lwp; |
|
501 |
||
502 |
ASSERT(new_state != LMS_WAIT_CPU); |
|
503 |
ASSERT((unsigned)new_state < NMSTATES); |
|
504 |
ASSERT(t == curthread || THREAD_LOCK_HELD(t)); |
|
505 |
||
506 |
if ((lwp = ttolwp(t)) == NULL) |
|
507 |
return (LMS_SYSTEM); |
|
508 |
||
509 |
curtime = gethrtime_unscaled(); |
|
510 |
||
511 |
/* adjust cpu percentages before we go any further */ |
|
512 |
(void) cpu_update_pct(t, curtime); |
|
513 |
||
514 |
ms = &lwp->lwp_mstate; |
|
515 |
state = t->t_mstate; |
|
516 |
do { |
|
517 |
switch (state) { |
|
518 |
case LMS_TFAULT: |
|
519 |
case LMS_DFAULT: |
|
520 |
case LMS_KFAULT: |
|
521 |
case LMS_USER_LOCK: |
|
522 |
mstimep = &ms->ms_acct[LMS_SYSTEM]; |
|
523 |
break; |
|
524 |
default: |
|
525 |
mstimep = &ms->ms_acct[state]; |
|
526 |
break; |
|
527 |
} |
|
528 |
newtime = curtime - ms->ms_state_start; |
|
529 |
if (newtime < 0) { |
|
530 |
curtime = gethrtime_unscaled(); |
|
531 |
oldtime = *mstimep - 1; /* force CAS to fail */ |
|
532 |
continue; |
|
533 |
} |
|
534 |
oldtime = *mstimep; |
|
535 |
newtime += oldtime; |
|
536 |
t->t_mstate = new_state; |
|
537 |
ms->ms_state_start = curtime; |
|
538 |
} while (cas64((uint64_t *)mstimep, oldtime, newtime) != oldtime); |
|
539 |
/* |
|
540 |
* Remember the previous running microstate. |
|
541 |
*/ |
|
542 |
if (state != LMS_SLEEP && state != LMS_STOPPED) |
|
543 |
ms->ms_prev = state; |
|
544 |
||
545 |
/* |
|
546 |
* Switch CPU microstate if appropriate |
|
547 |
*/ |
|
590
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
548 |
|
0 | 549 |
kpreempt_disable(); /* MUST disable kpreempt before touching t->cpu */ |
590
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
550 |
ASSERT(t->t_cpu == CPU); |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
551 |
if (!CPU_ON_INTR(t->t_cpu) && curthread->t_intr == NULL) { |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
552 |
if (new_state == LMS_USER && t->t_cpu->cpu_mstate != CMS_USER) |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
553 |
new_cpu_mstate(CMS_USER, curtime); |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
554 |
else if (new_state != LMS_USER && |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
555 |
t->t_cpu->cpu_mstate != CMS_SYSTEM) |
94ad906873b6
5062435 cpu_X_system kstats should include interrupt time
esolom
parents:
477
diff
changeset
|
556 |
new_cpu_mstate(CMS_SYSTEM, curtime); |
0 | 557 |
} |
558 |
kpreempt_enable(); |
|
559 |
||
560 |
return (ms->ms_prev); |
|
561 |
} |
|
562 |
||
563 |
static long waitrqis0 = 0; |
|
564 |
||
565 |
/* |
|
566 |
* Restore the LWP microstate to the previous runnable state. |
|
567 |
* Called from disp() with the newly selected lwp. |
|
568 |
*/ |
|
569 |
void |
|
570 |
restore_mstate(kthread_t *t) |
|
571 |
{ |
|
572 |
struct mstate *ms; |
|
573 |
hrtime_t *mstimep; |
|
574 |
klwp_t *lwp; |
|
575 |
hrtime_t curtime; |
|
576 |
hrtime_t waitrq; |
|
577 |
hrtime_t newtime; |
|
578 |
hrtime_t oldtime; |
|
579 |
||
580 |
if ((lwp = ttolwp(t)) == NULL) |
|
581 |
return; |
|
582 |
||
583 |
curtime = gethrtime_unscaled(); |
|
584 |
(void) cpu_update_pct(t, curtime); |
|
585 |
ms = &lwp->lwp_mstate; |
|
586 |
ASSERT((unsigned)t->t_mstate < NMSTATES); |
|
587 |
do { |
|
588 |
switch (t->t_mstate) { |
|
589 |
case LMS_SLEEP: |
|
590 |
/* |
|
591 |
* Update the timer for the current sleep state. |
|
592 |
*/ |
|
593 |
ASSERT((unsigned)ms->ms_prev < NMSTATES); |
|
594 |
switch (ms->ms_prev) { |
|
595 |
case LMS_TFAULT: |
|
596 |
case LMS_DFAULT: |
|
597 |
case LMS_KFAULT: |
|
598 |
case LMS_USER_LOCK: |
|
599 |
mstimep = &ms->ms_acct[ms->ms_prev]; |
|
600 |
break; |
|
601 |
default: |
|
602 |
mstimep = &ms->ms_acct[LMS_SLEEP]; |
|
603 |
break; |
|
604 |
} |
|
605 |
/* |
|
606 |
* Return to the previous run state. |
|
607 |
*/ |
|
608 |
t->t_mstate = ms->ms_prev; |
|
609 |
break; |
|
610 |
case LMS_STOPPED: |
|
611 |
mstimep = &ms->ms_acct[LMS_STOPPED]; |
|
612 |
/* |
|
613 |
* Return to the previous run state. |
|
614 |
*/ |
|
615 |
t->t_mstate = ms->ms_prev; |
|
616 |
break; |
|
617 |
case LMS_TFAULT: |
|
618 |
case LMS_DFAULT: |
|
619 |
case LMS_KFAULT: |
|
620 |
case LMS_USER_LOCK: |
|
621 |
mstimep = &ms->ms_acct[LMS_SYSTEM]; |
|
622 |
break; |
|
623 |
default: |
|
624 |
mstimep = &ms->ms_acct[t->t_mstate]; |
|
625 |
break; |
|
626 |
} |
|
627 |
waitrq = t->t_waitrq; /* hopefully atomic */ |
|
628 |
t->t_waitrq = 0; |
|
629 |
if (waitrq == 0) { /* should only happen during boot */ |
|
630 |
waitrq = curtime; |
|
631 |
waitrqis0++; |
|
632 |
} |
|
633 |
newtime = waitrq - ms->ms_state_start; |
|
634 |
if (newtime < 0) { |
|
635 |
curtime = gethrtime_unscaled(); |
|
636 |
oldtime = *mstimep - 1; /* force CAS to fail */ |
|
637 |
continue; |
|
638 |
} |
|
639 |
oldtime = *mstimep; |
|
640 |
newtime += oldtime; |
|
641 |
} while (cas64((uint64_t *)mstimep, oldtime, newtime) != oldtime); |
|
642 |
/* |
|
643 |
* Update the WAIT_CPU timer and per-cpu waitrq total. |
|
644 |
*/ |
|
645 |
ms->ms_acct[LMS_WAIT_CPU] += (curtime - waitrq); |
|
477
8043c92eb425
6217375 CPU unconfigure, t_disp_queue, and restore_mstate() duke it out
mishra
parents:
0
diff
changeset
|
646 |
CPU->cpu_waitrq += (curtime - waitrq); |
0 | 647 |
ms->ms_state_start = curtime; |
648 |
} |
|
649 |
||
650 |
/* |
|
651 |
* Copy lwp microstate accounting and resource usage information |
|
652 |
* to the process. (lwp is terminating) |
|
653 |
*/ |
|
654 |
void |
|
655 |
term_mstate(kthread_t *t) |
|
656 |
{ |
|
657 |
struct mstate *ms; |
|
658 |
proc_t *p = ttoproc(t); |
|
659 |
klwp_t *lwp = ttolwp(t); |
|
660 |
int i; |
|
661 |
hrtime_t tmp; |
|
662 |
||
663 |
ASSERT(MUTEX_HELD(&p->p_lock)); |
|
664 |
||
665 |
ms = &lwp->lwp_mstate; |
|
666 |
(void) new_mstate(t, LMS_STOPPED); |
|
667 |
ms->ms_term = ms->ms_state_start; |
|
668 |
tmp = ms->ms_term - ms->ms_start; |
|
669 |
scalehrtime(&tmp); |
|
670 |
p->p_mlreal += tmp; |
|
671 |
for (i = 0; i < NMSTATES; i++) { |
|
672 |
tmp = ms->ms_acct[i]; |
|
673 |
scalehrtime(&tmp); |
|
674 |
p->p_acct[i] += tmp; |
|
675 |
} |
|
676 |
p->p_ru.minflt += lwp->lwp_ru.minflt; |
|
677 |
p->p_ru.majflt += lwp->lwp_ru.majflt; |
|
678 |
p->p_ru.nswap += lwp->lwp_ru.nswap; |
|
679 |
p->p_ru.inblock += lwp->lwp_ru.inblock; |
|
680 |
p->p_ru.oublock += lwp->lwp_ru.oublock; |
|
681 |
p->p_ru.msgsnd += lwp->lwp_ru.msgsnd; |
|
682 |
p->p_ru.msgrcv += lwp->lwp_ru.msgrcv; |
|
683 |
p->p_ru.nsignals += lwp->lwp_ru.nsignals; |
|
684 |
p->p_ru.nvcsw += lwp->lwp_ru.nvcsw; |
|
685 |
p->p_ru.nivcsw += lwp->lwp_ru.nivcsw; |
|
686 |
p->p_ru.sysc += lwp->lwp_ru.sysc; |
|
687 |
p->p_ru.ioch += lwp->lwp_ru.ioch; |
|
688 |
p->p_defunct++; |
|
689 |
} |