1 /*
2  *      Copyright (C) 1993-1996 Bas Laarhoven.
3 
4  This program is free software; you can redistribute it and/or modify
5  it under the terms of the GNU General Public License as published by
6  the Free Software Foundation; either version 2, or (at your option)
7  any later version.
8 
9  This program is distributed in the hope that it will be useful,
10  but WITHOUT ANY WARRANTY; without even the implied warranty of
11  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  GNU General Public License for more details.
13 
14  You should have received a copy of the GNU General Public License
15  along with this program; see the file COPYING.  If not, write to
16  the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
17 
18  *
19  * $Source: /homes/cvs/ftape-stacked/ftape/lowlevel/ftape-calibr.c,v $
20  * $Revision: 1.2 $
21  * $Date: 1997/10/05 19:18:08 $
22  *
23  *      GP calibration routine for processor speed dependent
24  *      functions.
25  */
26 
27 #include <linux/config.h>
28 #include <linux/errno.h>
29 #include <linux/sched.h>
30 #include <asm/system.h>
31 #include <asm/io.h>
32 #if defined(__alpha__)
33 # include <asm/hwrpb.h>
34 #elif defined(__x86_64__)
35 # include <asm/msr.h>
36 # include <asm/timex.h>
37 #elif defined(__i386__)
38 # include <linux/timex.h>
39 #endif
40 #include <linux/ftape.h>
41 #include "../lowlevel/ftape-tracing.h"
42 #include "../lowlevel/ftape-calibr.h"
43 #include "../lowlevel/fdc-io.h"
44 
45 #undef DEBUG
46 
47 #if !defined(__alpha__) && !defined(__i386__) && !defined(__x86_64__)
48 # error Ftape is not implemented for this architecture!
49 #endif
50 
51 #if defined(__alpha__) || defined(__x86_64__)
52 static unsigned long ps_per_cycle = 0;
53 #endif
54 
55 #if defined(__i386__)
56 extern spinlock_t i8253_lock;
57 #endif
58 
59 /*
60  * Note: On Intel PCs, the clock ticks at 100 Hz (HZ==100) which is
61  * too slow for certain timeouts (and that clock doesn't even tick
62  * when interrupts are disabled).  For that reason, the 8254 timer is
63  * used directly to implement fine-grained timeouts.  However, on
64  * Alpha PCs, the 8254 is *not* used to implement the clock tick
65  * (which is 1024 Hz, normally) and the 8254 timer runs at some
66  * "random" frequency (it seems to run at 18Hz, but its not safe to
67  * rely on this value).  Instead, we use the Alpha's "rpcc"
68  * instruction to read cycle counts.  As this is a 32 bit counter,
69  * it will overflow only once per 30 seconds (on a 200MHz machine),
70  * which is plenty.
71  */
72 
ftape_timestamp(void)73 unsigned int ftape_timestamp(void)
74 {
75 #if defined(__alpha__)
76 	unsigned long r;
77 	asm volatile ("rpcc %0" : "=r" (r));
78 	return r;
79 #elif defined(__x86_64__)
80 	unsigned long r;
81 	rdtscl(r);
82 	return r;
83 #elif defined(__i386__)
84 
85 /*
86  * Note that there is some time between counter underflowing and jiffies
87  * increasing, so the code below won't always give correct output.
88  * -Vojtech
89  */
90 
91 	unsigned long flags;
92 	__u16 lo;
93 	__u16 hi;
94 
95 	spin_lock_irqsave(&i8253_lock, flags);
96 	outb_p(0x00, 0x43);	/* latch the count ASAP */
97 	lo = inb_p(0x40);	/* read the latched count */
98 	lo |= inb(0x40) << 8;
99 	hi = jiffies;
100 	spin_unlock_irqrestore(&i8253_lock, flags);
101 
102 	return ((hi + 1) * (unsigned int) LATCH) - lo;  /* downcounter ! */
103 #endif
104 }
105 
short_ftape_timestamp(void)106 static unsigned int short_ftape_timestamp(void)
107 {
108 #if defined(__alpha__) || defined(__x86_64__)
109 	return ftape_timestamp();
110 #elif defined(__i386__)
111 	unsigned int count;
112  	unsigned long flags;
113 
114 	spin_lock_irqsave(&i8253_lock, flags);
115  	outb_p(0x00, 0x43);	/* latch the count ASAP */
116 	count = inb_p(0x40);	/* read the latched count */
117 	count |= inb(0x40) << 8;
118 	spin_unlock_irqrestore(&i8253_lock, flags);
119 
120 	return (LATCH - count);	/* normal: downcounter */
121 #endif
122 }
123 
diff(unsigned int t0,unsigned int t1)124 static unsigned int diff(unsigned int t0, unsigned int t1)
125 {
126 #if defined(__alpha__) || defined(__x86_64__)
127 	return (t1 - t0);
128 #elif defined(__i386__)
129 	/*
130 	 * This is tricky: to work for both short and full ftape_timestamps
131 	 * we'll have to discriminate between these.
132 	 * If it _looks_ like short stamps with wrapping around we'll
133 	 * asume it are. This will generate a small error if it really
134 	 * was a (very large) delta from full ftape_timestamps.
135 	 */
136 	return (t1 <= t0 && t0 <= LATCH) ? t1 + LATCH - t0 : t1 - t0;
137 #endif
138 }
139 
usecs(unsigned int count)140 static unsigned int usecs(unsigned int count)
141 {
142 #if defined(__alpha__) || defined(__x86_64__)
143 	return (ps_per_cycle * count) / 1000000UL;
144 #elif defined(__i386__)
145 	return (10000 * count) / ((CLOCK_TICK_RATE + 50) / 100);
146 #endif
147 }
148 
ftape_timediff(unsigned int t0,unsigned int t1)149 unsigned int ftape_timediff(unsigned int t0, unsigned int t1)
150 {
151 	/*
152 	 *  Calculate difference in usec for ftape_timestamp results t0 & t1.
153 	 *  Note that on the i386 platform with short time-stamps, the
154 	 *  maximum allowed timespan is 1/HZ or we'll lose ticks!
155 	 */
156 	return usecs(diff(t0, t1));
157 }
158 
159 /*      To get an indication of the I/O performance,
160  *      measure the duration of the inb() function.
161  */
time_inb(void)162 static void time_inb(void)
163 {
164 	int i;
165 	int t0, t1;
166 	unsigned long flags;
167 	int status;
168 	TRACE_FUN(ft_t_any);
169 
170 	save_flags(flags);
171 	cli();
172 	t0 = short_ftape_timestamp();
173 	for (i = 0; i < 1000; ++i)
174 		status = inb(fdc.msr);
175 	t1 = short_ftape_timestamp();
176 	restore_flags(flags);
177 
178 	TRACE(ft_t_info, "inb() duration: %d nsec", ftape_timediff(t0, t1));
179 	TRACE_EXIT;
180 }
181 
init_clock(void)182 static void init_clock(void)
183 {
184 	TRACE_FUN(ft_t_any);
185 
186 #if defined(__x86_64__)
187 	ps_per_cycle = 1000000000UL / cpu_khz;
188 #elif defined(__alpha__)
189 	extern struct hwrpb_struct *hwrpb;
190 		ps_per_cycle = (1000*1000*1000*1000UL) / hwrpb->cycle_freq;
191 #endif
192 	TRACE_EXIT;
193 }
194 
195 /*
196  *      Input:  function taking int count as parameter.
197  *              pointers to calculated calibration variables.
198  */
ftape_calibrate(char * name,void (* fun)(unsigned int),unsigned int * calibr_count,unsigned int * calibr_time)199 void ftape_calibrate(char *name,
200 		    void (*fun) (unsigned int),
201 		    unsigned int *calibr_count,
202 		    unsigned int *calibr_time)
203 {
204 	static int first_time = 1;
205 	int i;
206 	unsigned int tc = 0;
207 	unsigned int count;
208 	unsigned int time;
209 #if defined(__i386__)
210 	unsigned int old_tc = 0;
211 	unsigned int old_count = 1;
212 	unsigned int old_time = 1;
213 #endif
214 	TRACE_FUN(ft_t_flow);
215 
216 	if (first_time) {             /* get idea of I/O performance */
217 		init_clock();
218 		time_inb();
219 		first_time = 0;
220 	}
221 	/*    value of timeout must be set so that on very slow systems
222 	 *    it will give a time less than one jiffy, and on
223 	 *    very fast systems it'll give reasonable precision.
224 	 */
225 
226 	count = 40;
227 	for (i = 0; i < 15; ++i) {
228 		unsigned int t0;
229 		unsigned int t1;
230 		unsigned int once;
231 		unsigned int multiple;
232 		unsigned long flags;
233 
234 		*calibr_count =
235 		*calibr_time = count;	/* set TC to 1 */
236 		save_flags(flags);
237 		cli();
238 		fun(0);		/* dummy, get code into cache */
239 		t0 = short_ftape_timestamp();
240 		fun(0);		/* overhead + one test */
241 		t1 = short_ftape_timestamp();
242 		once = diff(t0, t1);
243 		t0 = short_ftape_timestamp();
244 		fun(count);		/* overhead + count tests */
245 		t1 = short_ftape_timestamp();
246 		multiple = diff(t0, t1);
247 		restore_flags(flags);
248 		time = ftape_timediff(0, multiple - once);
249 		tc = (1000 * time) / (count - 1);
250 		TRACE(ft_t_any, "once:%3d us,%6d times:%6d us, TC:%5d ns",
251 			usecs(once), count - 1, usecs(multiple), tc);
252 #if defined(__alpha__) || defined(__x86_64__)
253 		/*
254 		 * Increase the calibration count exponentially until the
255 		 * calibration time exceeds 100 ms.
256 		 */
257 		if (time >= 100*1000)
258 			break;
259 #elif defined(__i386__)
260 		/*
261 		 * increase the count until the resulting time nears 2/HZ,
262 		 * then the tc will drop sharply because we lose LATCH counts.
263 		 */
264 		if (tc <= old_tc / 2) {
265 			time = old_time;
266 			count = old_count;
267 			break;
268 		}
269 		old_tc = tc;
270 		old_count = count;
271 		old_time = time;
272 #endif
273 		count *= 2;
274 	}
275 	*calibr_count = count - 1;
276 	*calibr_time  = time;
277 	TRACE(ft_t_info, "TC for `%s()' = %d nsec (at %d counts)",
278 	     name, (1000 * *calibr_time) / *calibr_count, *calibr_count);
279 	TRACE_EXIT;
280 }
281