xref: /titanic_41/usr/src/uts/sun4/sys/clock.h (revision 5bbb4db2c3f208d12bf0fd11769728f9e5ba66a2)
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 (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #ifndef _SYS_CLOCK_H
27 #define	_SYS_CLOCK_H
28 
29 #pragma ident	"%Z%%M%	%I%	%E% SMI"
30 
31 #ifdef	__cplusplus
32 extern "C" {
33 #endif
34 
35 #include <sys/spl.h>
36 #include <sys/time.h>
37 #include <sys/machclock.h>
38 
39 #ifndef _ASM
40 
41 #ifdef	_KERNEL
42 
43 extern void	setcpudelay(void);
44 
45 extern uint_t	nsec_scale;
46 extern uint_t	nsec_shift;
47 extern uint_t	nsec_per_sys_tick;
48 extern uint64_t	sys_tick_freq;
49 
50 extern int	traptrace_use_stick;
51 extern uint64_t	system_clock_freq;
52 extern uint_t	sys_clock_mhz;
53 
54 extern void mon_clock_init(void);
55 extern void mon_clock_start(void);
56 extern void mon_clock_stop(void);
57 extern void mon_clock_share(void);
58 extern void mon_clock_unshare(void);
59 
60 extern hrtime_t hrtime_base;
61 extern void hres_tick(void);
62 extern void	clkstart(void);
63 extern void cbe_level14();
64 extern hrtime_t tick2ns(hrtime_t, uint_t);
65 
66 typedef struct {
67 	uint64_t cbe_level1_inum;
68 	uint64_t cbe_level10_inum;
69 } cbe_data_t;
70 
71 #endif	/* _KERNEL */
72 
73 #endif	/* _ASM */
74 
75 
76 #define	CBE_LOW_PIL	1
77 #define	CBE_LOCK_PIL	LOCK_LEVEL
78 #define	CBE_HIGH_PIL	14
79 
80 #define	ADJ_SHIFT	4	/* used in get_hrestime and _level10 */
81 
82 /*
83  * Locking strategy for high-resolution timing services
84  *
85  * We generally construct timestamps from two or more components:
86  * a hardware time source and one or more software time sources.
87  * These components cannot all be loaded simultaneously, so we need
88  * some sort of locking strategy to generate consistent timestamps.
89  *
90  * To minimize lock contention and cache thrashing we employ the
91  * weakest possible synchronization model: writers (rare) serialize
92  * on an acquisition-counting mutex, described below; readers (common)
93  * execute in parallel with no synchronization at all -- they don't
94  * exclude other readers, and they don't even exclude writers.  Instead,
95  * readers just examine the writer lock's value before and after loading
96  * all the components of a timestamp to detect writer intervention.
97  * In the rare case when a writer does intervene, the reader will
98  * detect it, discard the timestamp and try again.
99  *
100  * The writer lock, hres_lock, is a 32-bit integer consisting of an
101  * 8-bit lock and a 24-bit acquisition count.  To acquire the lock we
102  * set the lock field with ldstub, which sets the low-order 8 bits to
103  * 0xff; to clear the lock, we increment it, which simultaneously clears
104  * the lock field (0xff --> 0x00) and increments the acquisition count
105  * (due to carry into bit 8).  Thus each acquisition transforms hres_lock
106  * from N:0 to N:ff, and each release transforms N:ff into (N+1):0.
107  *
108  * Readers can detect writer intervention by loading hres_lock before
109  * and after loading the time components they need; if either lock value
110  * contains 0xff in the low-order bits (lock held), or if the lock values
111  * are not equal (lock was acquired and released), a writer intervened
112  * and the reader must try again.  If the lock values are equal and the
113  * low-order 8 bits are clear, the timestamp must be valid.  We can check
114  * both of these conditions with a single compare instruction by checking
115  * whether old_hres_lock & ~1 == new_hres_lock, as illustrated by the
116  * following table of all possible lock states:
117  *
118  *	initial	& ~1	final		result of compare
119  *	------------	-----		-----------------
120  *	now:00		now:00		valid
121  *	now:00		now:ff		invalid
122  *	now:00		later:00	invalid
123  *	now:00		later:ff	invalid
124  *	now:fe		now:ff		invalid
125  *	now:fe		later:00	invalid
126  *	now:fe		later:ff	invalid
127  *
128  * Implementation considerations:
129  *
130  * (1) Load buffering.
131  *
132  * On a CPU that does load buffering we must ensure that the load of
133  * hres_lock completes before the load of any timestamp components.
134  * This is essential *even on a CPU that does in-order loads* because
135  * accessing the hardware time source may not involve a memory reference
136  * (e.g. rd %tick).  A convenient way to address this is to clear the
137  * lower bit (andn with 1) of the old lock value right away, since this
138  * generates a dependency on the load of hres_lock.  We have to do this
139  * anyway to perform the lock comparison described above.
140  *
141  * (2) Out-of-order loads.
142  *
143  * On a CPU that does out-of-order loads we must ensure that the loads
144  * of all timestamp components have completed before we load the final
145  * value of hres_lock.  This can be done either by generating load
146  * dependencies on the timestamp components or by membar #LoadLoad.
147  *
148  * (3) Interaction with the high level cyclic handler, hres_tick().
149  *
150  * One unusual property of hres_lock is that it's acquired in a high
151  * level cyclic handler, hres_tick().  Thus, hres_lock must be acquired at
152  * CBE_HIGH_PIL or higher to prevent single-CPU deadlock.
153  *
154  * (4) Cross-calls.
155  *
156  * If a cross-call happens while one CPU has hres_lock and another is
157  * trying to acquire it in the clock interrupt path, the system will
158  * deadlock: the first CPU will never release hres_lock since it's
159  * waiting to be released from the cross-call, and the cross-call can't
160  * complete because the second CPU is spinning on hres_lock with traps
161  * disabled.  Thus cross-calls must be blocked while holding hres_lock.
162  *
163  * Together, (3) and (4) imply that hres_lock should only be acquired
164  * at PIL >= max(XCALL_PIL, CBE_HIGH_PIL), or while traps are disabled.
165  */
166 #define	HRES_LOCK_OFFSET 3
167 
168 #define	CLOCK_LOCK(oldsplp)	\
169 	lock_set_spl((lock_t *)&hres_lock + HRES_LOCK_OFFSET, \
170 		ipltospl(CBE_HIGH_PIL), oldsplp)
171 
172 #define	CLOCK_UNLOCK(spl)	\
173 	membar_ldst_stst();	\
174 	hres_lock++;		\
175 	splx(spl);		\
176 	LOCKSTAT_RECORD0(LS_CLOCK_UNLOCK_RELEASE,	\
177 		(lock_t *)&hres_lock + HRES_LOCK_OFFSET);
178 
179 /*
180  * NATIVE_TIME_TO_NSEC_SCALE is called with NSEC_SHIFT to convert hi-res
181  * timestamps into nanoseconds. On systems that have a %stick register,
182  * hi-res timestamps are in %stick units. On systems that do not have a
183  * %stick register, hi-res timestamps are in %tick units.
184  *
185  * NATIVE_TIME_TO_NSEC_SCALE is called with TICK_NSEC_SHIFT to convert from
186  * %tick units to nanoseconds on all implementations whether %stick is
187  * available or not.
188  */
189 
190 /*
191  * At least 62.5 MHz CPU %tick frequency
192  */
193 
194 #define	TICK_NSEC_SHIFT	4
195 
196 /*
197  * Convert hi-res native time (V9's %tick in our case) into nanoseconds.
198  *
199  * The challenge is to multiply a %tick value by (NANOSEC / sys_tick_freq)
200  * without using floating point and without overflowing 64-bit integers.
201  * We assume that all sun4u systems will have a 16 nsec or better clock
202  * (i.e. faster than 62.5 MHz), which means that (ticks << 4) has units
203  * greater than one nanosecond, so converting from (ticks << 4) to nsec
204  * requires multiplication by a rational number, R, between 0 and 1.
205  * To avoid floating-point we precompute (R * 2^32) during boot and
206  * stash this away in nsec_scale.  Thus we can compute (tick * R) as
207  * (tick * nsec_scale) >> 32, which is accurate to about 1 part per billion.
208  *
209  * To avoid 64-bit overflow when multiplying (tick << 4) by nsec_scale,
210  * we split (tick << 4) into its high and low 32-bit pieces, H and L,
211  * multiply each piece separately, and add up the relevant bits of the
212  * partial products.  Putting it all together we have:
213  *
214  * nsec = (tick << 4) * R
215  *	= ((tick << 4) * nsec_scale) >> 32
216  *	= ((H << 32) + L) * nsec_scale) >> 32
217  *	= (H * nsec_scale) + ((L * nsec_scale) >> 32)
218  *
219  * The last line is the computation we actually perform: it requires no
220  * floating point and all intermediate results fit in 64-bit registers.
221  *
222  * Note that we require that tick is less than (1 << (64 - NSEC_SHIFT));
223  * greater values will result in overflow and misbehavior (not that this
224  * is a serious problem; (1 << (64 - NSEC_SHIFT)) nanoseconds is over
225  * thirty-six years).  Nonetheless, clients may wish to be aware of this
226  * limitation; NATIVE_TIME_MAX() returns this maximum native time.
227  *
228  * We provide two versions of this macro: a "full-service" version that
229  * just converts ticks to nanoseconds and a higher-performance version that
230  * expects the scaling factor nsec_scale as its second argument (so that
231  * callers can distance the load of nsec_scale from its use).  Note that
232  * we take a fast path if we determine the ticks to be less than 32 bits
233  * (as it often is for the delta between %tick values for successive
234  * firings of the hres_tick() cyclic).
235  *
236  * Note that in the 32-bit path we don't even bother clearing NPT.
237  * We get away with this by making hardclk.c ensure than nsec_scale
238  * is even, so we can take advantage of the associativity of modular
239  * arithmetic: multiplying %tick by any even number, say 2*n, is
240  * equivalent to multiplying %tick by 2, then by n.  Multiplication
241  * by 2 is equivalent to shifting left by one, which clears NPT.
242  *
243  * Finally, note that the macros use the labels "6:" and "7:"; these
244  * labels must not be used across an invocation of either macro.
245  */
246 #define	NATIVE_TIME_TO_NSEC_SCALE(out, scr1, scr2, shift)		\
247 	srlx	out, 32, scr2;		/* check high 32 bits */	\
248 /* CSTYLED */ 								\
249 	brz,a,pt scr2, 6f;		/* if clear, 32-bit fast path */\
250 	mulx	out, scr1, out;		/* delay: 32-bit fast path */	\
251 	sllx	out, shift, out;	/* clear NPT and pre-scale */	\
252 	srlx	out, 32, scr2;		/* scr2 = hi32(tick<<4) = H */	\
253 	mulx	scr2, scr1, scr2;	/* scr2 = (H*F) */		\
254 	srl	out, 0, out;		/* out = lo32(tick<<4) = L */	\
255 	mulx	out, scr1, scr1;	/* scr1 = (L*F) */		\
256 	srlx	scr1, 32, scr1;		/* scr1 = (L*F) >> 32 */	\
257 	ba	7f;			/* branch over 32-bit path */	\
258 	add	scr1, scr2, out;	/* out = (H*F) + ((L*F) >> 32) */\
259 6:									\
260 	srlx	out, 32 - shift, out;					\
261 7:
262 
263 #define	NATIVE_TIME_TO_NSEC(out, scr1, scr2)				\
264 	sethi	%hi(nsec_scale), scr1;	/* load scaling factor */	\
265 	ld	[scr1 + %lo(nsec_scale)], scr1;				\
266 	NATIVE_TIME_TO_NSEC_SCALE(out, scr1, scr2, NSEC_SHIFT);
267 
268 #define	NATIVE_TIME_MAX(out)						\
269 	mov	-1, out;						\
270 	srlx	out, NSEC_SHIFT, out
271 
272 
273 /*
274  * The following macros are only for use in the cpu module.
275  */
276 #if defined(CPU_MODULE)
277 
278 /*
279  * NSEC_SHIFT and VTRACE_SHIFT constants are defined in
280  * <sys/machclock.h> file.
281  */
282 
283 
284 /*
285  * NOTE: the macros below assume that the various time-related variables
286  * (hrestime, hrestime_adj, hres_last_tick, timedelta, nsec_scale, etc)
287  * are all stored together on a 64-byte boundary.  The primary motivation
288  * is cache performance, but we also take advantage of a convenient side
289  * effect: these variables all have the same high 22 address bits, so only
290  * one sethi is needed to access them all.
291  */
292 
293 /*
294  * GET_HRESTIME() returns the value of hrestime, hrestime_adj and the
295  * number of nanoseconds since the last clock tick ('nslt').  It also
296  * sets 'nano' to the value NANOSEC (one billion).
297  *
298  * This macro assumes that all registers are globals or outs so they can
299  * safely contain 64-bit data, and that it's safe to use the label "5:".
300  * Further, this macro calls the NATIVE_TIME_TO_NSEC_SCALE which in turn
301  * uses the labels "6:" and "7:"; labels "5:", "6:" and "7:" must not
302  * be used across invocations of this macro.
303  */
304 #define	GET_HRESTIME(hrestsec, hrestnsec, adj, nslt, nano, scr, hrlock, \
305     gnt1, gnt2) \
306 5:	sethi	%hi(hres_lock), scr;					\
307 	lduw	[scr + %lo(hres_lock)], hrlock;	/* load clock lock */	\
308 	lduw	[scr + %lo(nsec_scale)], nano;	/* tick-to-ns factor */	\
309 	andn	hrlock, 1, hrlock;  	/* see comments above! */	\
310 	ldx	[scr + %lo(hres_last_tick)], nslt;			\
311 	ldn	[scr + %lo(hrestime)], hrestsec; /* load hrestime.sec */\
312 	add	scr, %lo(hrestime), hrestnsec;				\
313 	ldn	[hrestnsec + CLONGSIZE], hrestnsec;			\
314 	GET_NATIVE_TIME(adj, gnt1, gnt2);	/* get current %tick */	\
315 	subcc	adj, nslt, nslt; /* nslt = ticks since last clockint */	\
316 	movneg	%xcc, %g0, nslt; /* ignore neg delta from tick skew */	\
317 	ldx	[scr + %lo(hrestime_adj)], adj; /* load hrestime_adj */	\
318 	/* membar #LoadLoad; (see comment (2) above) */			\
319 	lduw	[scr + %lo(hres_lock)], scr; /* load clock lock */	\
320 	NATIVE_TIME_TO_NSEC_SCALE(nslt, nano, gnt1, NSEC_SHIFT);	\
321 	sethi	%hi(NANOSEC), nano;					\
322 	xor	hrlock, scr, scr;					\
323 /* CSTYLED */ 								\
324 	brnz,pn	scr, 5b;						\
325 	or	nano, %lo(NANOSEC), nano;
326 
327 /*
328  * Similar to above, but returns current gethrtime() value in 'base'.
329  */
330 #define	GET_HRTIME(base, now, nslt, scale, scr, hrlock, gnt1, gnt2)	\
331 5:	sethi	%hi(hres_lock), scr;					\
332 	lduw	[scr + %lo(hres_lock)], hrlock;	/* load clock lock */	\
333 	lduw	[scr + %lo(nsec_scale)], scale;	/* tick-to-ns factor */	\
334 	andn	hrlock, 1, hrlock;  	/* see comments above! */	\
335 	ldx	[scr + %lo(hres_last_tick)], nslt;			\
336 	ldx	[scr + %lo(hrtime_base)], base;	/* load hrtime_base */	\
337 	GET_NATIVE_TIME(now, gnt1, gnt2);	/* get current %tick */	\
338 	subcc	now, nslt, nslt; /* nslt = ticks since last clockint */	\
339 	movneg	%xcc, %g0, nslt; /* ignore neg delta from tick skew */	\
340 	/* membar #LoadLoad; (see comment (2) above) */			\
341 	ld	[scr + %lo(hres_lock)], scr; /* load clock lock */	\
342 	NATIVE_TIME_TO_NSEC_SCALE(nslt, scale, gnt1, NSEC_SHIFT);	\
343 	xor	hrlock, scr, scr;					\
344 /* CSTYLED */ 								\
345 	brnz,pn	scr, 5b;						\
346 	add	base, nslt, base;
347 
348 /*
349  * Maximum-performance timestamp for kernel tracing.  We don't bother
350  * clearing NPT because vtrace expresses everything in 32-bit deltas,
351  * so only the low-order 32 bits matter.  We do shift down a few bits,
352  * however, so that the trace framework doesn't emit a ridiculous number
353  * of 32_bit_elapsed_time records (trace points are more expensive when
354  * the time since the last trace point doesn't fit in a 16-bit delta).
355  * We currently shift by 4 (divide by 16) on the grounds that (1) there's
356  * no point making the timing finer-grained than the trace point latency,
357  * which exceeds 16 cycles; and (2) the cost and probe effect of many
358  * 32-bit time records far exceeds the cost of the 'srlx' instruction.
359  */
360 #define	GET_VTRACE_TIME(out, scr1, scr2)				\
361 	GET_NATIVE_TIME(out, scr1, scr2);	/* get current %tick */	\
362 	srlx	out, VTRACE_SHIFT, out;
363 
364 /*
365  * Full 64-bit version for those truly rare occasions when you need it.
366  * Currently this is only needed to generate the TR_START_TIME record.
367  */
368 #define	GET_VTRACE_TIME_64(out, scr1, scr2)				\
369 	GET_NATIVE_TIME(out, scr1, scr2);	/* get current %tick */	\
370 	add	out, out, out;						\
371 	srlx	out, VTRACE_SHIFT + 1, out;
372 
373 /*
374  * Return the rate at which the vtrace clock runs.
375  */
376 #define	GET_VTRACE_FREQUENCY(out, scr1, scr2)				\
377 	sethi	%hi(sys_tick_freq), out;				\
378 	ldx	[out + %lo(sys_tick_freq)], out;			\
379 	srlx	out, VTRACE_SHIFT, out;
380 
381 #endif /* CPU_MODULE */
382 
383 #ifdef	__cplusplus
384 }
385 #endif
386 
387 #endif	/* !_SYS_CLOCK_H */
388