drivers/char/ftape/lowlevel/ftape-calibr.c

   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/errno.h>
  28 #include <linux/jiffies.h>
  29 #include <asm/system.h>
  30 #include <asm/io.h>
  31 #if defined(__alpha__)
  32 # include <asm/hwrpb.h>
  33 #elif defined(__x86_64__)
  34 # include <asm/msr.h>
  35 # include <asm/timex.h>
  36 #elif defined(__i386__)
  37 # include <linux/timex.h>
  38 #endif
  39 #include <linux/ftape.h>
  40 #include "../lowlevel/ftape-tracing.h"
  41 #include "../lowlevel/ftape-calibr.h"
  42 #include "../lowlevel/fdc-io.h"
  43
  44 #undef DEBUG
  45
  46 #if !defined(__alpha__) && !defined(__i386__) && !defined(__x86_64__)
  47 # error Ftape is not implemented for this architecture!
  48 #endif
  49
  50 #if defined(__alpha__) || defined(__x86_64__)
  51 static unsigned long ps_per_cycle = 0;
  52 #endif
  53
  54 static spinlock_t calibr_lock;
  55
  56 /*
  57  * Note: On Intel PCs, the clock ticks at 100 Hz (HZ==100) which is
  58  * too slow for certain timeouts (and that clock doesn't even tick
  59  * when interrupts are disabled).  For that reason, the 8254 timer is
  60  * used directly to implement fine-grained timeouts.  However, on
  61  * Alpha PCs, the 8254 is *not* used to implement the clock tick
  62  * (which is 1024 Hz, normally) and the 8254 timer runs at some
  63  * "random" frequency (it seems to run at 18Hz, but it's not safe to
  64  * rely on this value).  Instead, we use the Alpha's "rpcc"
  65  * instruction to read cycle counts.  As this is a 32 bit counter,
  66  * it will overflow only once per 30 seconds (on a 200MHz machine),
  67  * which is plenty.
  68  */
  69
  70 unsigned int ftape_timestamp(void)
  71 {
  72 #if defined(__alpha__)
  73         unsigned long r;
  74
  75         asm volatile ("rpcc %0" : "=r" (r));
  76         return r;
  77 #elif defined(__x86_64__)
  78         unsigned long r;
  79         rdtscl(r);
  80         return r;
  81 #elif defined(__i386__)
  82
  83 /*
  84  * Note that there is some time between counter underflowing and jiffies
  85  * increasing, so the code below won't always give correct output.
  86  * -Vojtech
  87  */
  88
  89         unsigned long flags;
  90         __u16 lo;
  91         __u16 hi;
  92
  93         spin_lock_irqsave(&calibr_lock, flags);
  94         outb_p(0x00, 0x43);     /* latch the count ASAP */
  95         lo = inb_p(0x40);       /* read the latched count */
  96         lo |= inb(0x40) << 8;
  97         hi = jiffies;
  98         spin_unlock_irqrestore(&calibr_lock, flags);
  99         return ((hi + 1) * (unsigned int) LATCH) - lo;  /* downcounter ! */
 100 #endif
 101 }
 102
 103 static unsigned int short_ftape_timestamp(void)
 104 {
 105 #if defined(__alpha__) || defined(__x86_64__)
 106         return ftape_timestamp();
 107 #elif defined(__i386__)
 108         unsigned int count;
 109         unsigned long flags;
 110
 111         spin_lock_irqsave(&calibr_lock, flags);
 112         outb_p(0x00, 0x43);     /* latch the count ASAP */
 113         count = inb_p(0x40);    /* read the latched count */
 114         count |= inb(0x40) << 8;
 115         spin_unlock_irqrestore(&calibr_lock, flags);
 116         return (LATCH - count); /* normal: downcounter */
 117 #endif
 118 }
 119
 120 static unsigned int diff(unsigned int t0, unsigned int t1)
 121 {
 122 #if defined(__alpha__) || defined(__x86_64__)
 123         return (t1 - t0);
 124 #elif defined(__i386__)
 125         /*
 126          * This is tricky: to work for both short and full ftape_timestamps
 127          * we'll have to discriminate between these.
 128          * If it _looks_ like short stamps with wrapping around we'll
 129          * asume it are. This will generate a small error if it really
 130          * was a (very large) delta from full ftape_timestamps.
 131          */
 132         return (t1 <= t0 && t0 <= LATCH) ? t1 + LATCH - t0 : t1 - t0;
 133 #endif
 134 }
 135
 136 static unsigned int usecs(unsigned int count)
 137 {
 138 #if defined(__alpha__) || defined(__x86_64__)
 139         return (ps_per_cycle * count) / 1000000UL;
 140 #elif defined(__i386__)
 141         return (10000 * count) / ((CLOCK_TICK_RATE + 50) / 100);
 142 #endif
 143 }
 144
 145 unsigned int ftape_timediff(unsigned int t0, unsigned int t1)
 146 {
 147         /*
 148          *  Calculate difference in usec for ftape_timestamp results t0 & t1.
 149          *  Note that on the i386 platform with short time-stamps, the
 150          *  maximum allowed timespan is 1/HZ or we'll lose ticks!
 151          */
 152         return usecs(diff(t0, t1));
 153 }
 154
 155 /*      To get an indication of the I/O performance,
 156  *      measure the duration of the inb() function.
 157  */
 158 static void time_inb(void)
 159 {
 160         int i;
 161         int t0, t1;
 162         unsigned long flags;
 163         int status;
 164         TRACE_FUN(ft_t_any);
 165
 166         spin_lock_irqsave(&calibr_lock, flags);
 167         t0 = short_ftape_timestamp();
 168         for (i = 0; i < 1000; ++i) {
 169                 status = inb(fdc.msr);
 170         }
 171         t1 = short_ftape_timestamp();
 172         spin_unlock_irqrestore(&calibr_lock, flags);
 173         TRACE(ft_t_info, "inb() duration: %d nsec", ftape_timediff(t0, t1));
 174         TRACE_EXIT;
 175 }
 176
 177 static void init_clock(void)
 178 {
 179         TRACE_FUN(ft_t_any);
 180
 181 #if defined(__x86_64__)
 182         ps_per_cycle = 1000000000UL / cpu_khz;
 183 #elif defined(__alpha__)
 184         extern struct hwrpb_struct *hwrpb;
 185         ps_per_cycle = (1000*1000*1000*1000UL) / hwrpb->cycle_freq;
 186 #endif
 187         TRACE_EXIT;
 188 }
 189
 190 /*
 191  *      Input:  function taking int count as parameter.
 192  *              pointers to calculated calibration variables.
 193  */
 194 void ftape_calibrate(char *name,
 195                     void (*fun) (unsigned int),
 196                     unsigned int *calibr_count,
 197                     unsigned int *calibr_time)
 198 {
 199         static int first_time = 1;
 200         int i;
 201         unsigned int tc = 0;
 202         unsigned int count;
 203         unsigned int time;
 204 #if defined(__i386__)
 205         unsigned int old_tc = 0;
 206         unsigned int old_count = 1;
 207         unsigned int old_time = 1;
 208 #endif
 209         TRACE_FUN(ft_t_flow);
 210
 211         if (first_time) {             /* get idea of I/O performance */
 212                 init_clock();
 213                 time_inb();
 214                 first_time = 0;
 215         }
 216         /*    value of timeout must be set so that on very slow systems
 217          *    it will give a time less than one jiffy, and on
 218          *    very fast systems it'll give reasonable precision.
 219          */
 220
 221         count = 40;
 222         for (i = 0; i < 15; ++i) {
 223                 unsigned int t0;
 224                 unsigned int t1;
 225                 unsigned int once;
 226                 unsigned int multiple;
 227                 unsigned long flags;
 228
 229                 *calibr_count =
 230                 *calibr_time = count;   /* set TC to 1 */
 231                 spin_lock_irqsave(&calibr_lock, flags);
 232                 fun(0);         /* dummy, get code into cache */
 233                 t0 = short_ftape_timestamp();
 234                 fun(0);         /* overhead + one test */
 235                 t1 = short_ftape_timestamp();
 236                 once = diff(t0, t1);
 237                 t0 = short_ftape_timestamp();
 238                 fun(count);             /* overhead + count tests */
 239                 t1 = short_ftape_timestamp();
 240                 multiple = diff(t0, t1);
 241                 spin_unlock_irqrestore(&calibr_lock, flags);
 242                 time = ftape_timediff(0, multiple - once);
 243                 tc = (1000 * time) / (count - 1);
 244                 TRACE(ft_t_any, "once:%3d us,%6d times:%6d us, TC:%5d ns",
 245                         usecs(once), count - 1, usecs(multiple), tc);
 246 #if defined(__alpha__) || defined(__x86_64__)
 247                 /*
 248                  * Increase the calibration count exponentially until the
 249                  * calibration time exceeds 100 ms.
 250                  */
 251                 if (time >= 100*1000) {
 252                         break;
 253                 }
 254 #elif defined(__i386__)
 255                 /*
 256                  * increase the count until the resulting time nears 2/HZ,
 257                  * then the tc will drop sharply because we lose LATCH counts.
 258                  */
 259                 if (tc <= old_tc / 2) {
 260                         time = old_time;
 261                         count = old_count;
 262                         break;
 263                 }
 264                 old_tc = tc;
 265                 old_count = count;
 266                 old_time = time;
 267 #endif
 268                 count *= 2;
 269         }
 270         *calibr_count = count - 1;
 271         *calibr_time  = time;
 272         TRACE(ft_t_info, "TC for `%s()' = %d nsec (at %d counts)",
 273              name, (1000 * *calibr_time) / *calibr_count, *calibr_count);
 274         TRACE_EXIT;
 275 }