drivers/char/mmtimer.c

   1 /*
   2  * Timer device implementation for SGI SN platforms.
   3  *
   4  * This file is subject to the terms and conditions of the GNU General Public
   5  * License.  See the file "COPYING" in the main directory of this archive
   6  * for more details.
   7  *
   8  * Copyright (c) 2001-2006 Silicon Graphics, Inc.  All rights reserved.
   9  *
  10  * This driver exports an API that should be supportable by any HPET or IA-PC
  11  * multimedia timer.  The code below is currently specific to the SGI Altix
  12  * SHub RTC, however.
  13  *
  14  * 11/01/01 - jbarnes - initial revision
  15  * 9/10/04 - Christoph Lameter - remove interrupt support for kernel inclusion
  16  * 10/1/04 - Christoph Lameter - provide posix clock CLOCK_SGI_CYCLE
  17  * 10/13/04 - Christoph Lameter, Dimitri Sivanich - provide timer interrupt
  18  *              support via the posix timer interface
  19  */
  20
  21 #include <linux/types.h>
  22 #include <linux/kernel.h>
  23 #include <linux/ioctl.h>
  24 #include <linux/module.h>
  25 #include <linux/init.h>
  26 #include <linux/errno.h>
  27 #include <linux/mm.h>
  28 #include <linux/fs.h>
  29 #include <linux/mmtimer.h>
  30 #include <linux/miscdevice.h>
  31 #include <linux/posix-timers.h>
  32 #include <linux/interrupt.h>
  33 #include <linux/time.h>
  34 #include <linux/math64.h>
  35 #include <linux/mutex.h>
  36 #include <linux/slab.h>
  37
  38 #include <asm/uaccess.h>
  39 #include <asm/sn/addrs.h>
  40 #include <asm/sn/intr.h>
  41 #include <asm/sn/shub_mmr.h>
  42 #include <asm/sn/nodepda.h>
  43 #include <asm/sn/shubio.h>
  44
  45 MODULE_AUTHOR("Jesse Barnes <jbarnes@sgi.com>");
  46 MODULE_DESCRIPTION("SGI Altix RTC Timer");
  47 MODULE_LICENSE("GPL");
  48
  49 /* name of the device, usually in /dev */
  50 #define MMTIMER_NAME "mmtimer"
  51 #define MMTIMER_DESC "SGI Altix RTC Timer"
  52 #define MMTIMER_VERSION "2.1"
  53
  54 #define RTC_BITS 55 /* 55 bits for this implementation */
  55
  56 static struct k_clock sgi_clock;
  57
  58 extern unsigned long sn_rtc_cycles_per_second;
  59
  60 #define RTC_COUNTER_ADDR        ((long *)LOCAL_MMR_ADDR(SH_RTC))
  61
  62 #define rtc_time()              (*RTC_COUNTER_ADDR)
  63
  64 static DEFINE_MUTEX(mmtimer_mutex);
  65 static long mmtimer_ioctl(struct file *file, unsigned int cmd,
  66                                                 unsigned long arg);
  67 static int mmtimer_mmap(struct file *file, struct vm_area_struct *vma);
  68
  69 /*
  70  * Period in femtoseconds (10^-15 s)
  71  */
  72 static unsigned long mmtimer_femtoperiod = 0;
  73
  74 static const struct file_operations mmtimer_fops = {
  75         .owner = THIS_MODULE,
  76         .mmap = mmtimer_mmap,
  77         .unlocked_ioctl = mmtimer_ioctl,
  78         .llseek = noop_llseek,
  79 };
  80
  81 /*
  82  * We only have comparison registers RTC1-4 currently available per
  83  * node.  RTC0 is used by SAL.
  84  */
  85 /* Check for an RTC interrupt pending */
  86 static int mmtimer_int_pending(int comparator)
  87 {
  88         if (HUB_L((unsigned long *)LOCAL_MMR_ADDR(SH_EVENT_OCCURRED)) &
  89                         SH_EVENT_OCCURRED_RTC1_INT_MASK << comparator)
  90                 return 1;
  91         else
  92                 return 0;
  93 }
  94
  95 /* Clear the RTC interrupt pending bit */
  96 static void mmtimer_clr_int_pending(int comparator)
  97 {
  98         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_EVENT_OCCURRED_ALIAS),
  99                 SH_EVENT_OCCURRED_RTC1_INT_MASK << comparator);
 100 }
 101
 102 /* Setup timer on comparator RTC1 */
 103 static void mmtimer_setup_int_0(int cpu, u64 expires)
 104 {
 105         u64 val;
 106
 107         /* Disable interrupt */
 108         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_RTC1_INT_ENABLE), 0UL);
 109
 110         /* Initialize comparator value */
 111         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_INT_CMPB), -1L);
 112
 113         /* Clear pending bit */
 114         mmtimer_clr_int_pending(0);
 115
 116         val = ((u64)SGI_MMTIMER_VECTOR << SH_RTC1_INT_CONFIG_IDX_SHFT) |
 117                 ((u64)cpu_physical_id(cpu) <<
 118                         SH_RTC1_INT_CONFIG_PID_SHFT);
 119
 120         /* Set configuration */
 121         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_RTC1_INT_CONFIG), val);
 122
 123         /* Enable RTC interrupts */
 124         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_RTC1_INT_ENABLE), 1UL);
 125
 126         /* Initialize comparator value */
 127         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_INT_CMPB), expires);
 128
 129
 130 }
 131
 132 /* Setup timer on comparator RTC2 */
 133 static void mmtimer_setup_int_1(int cpu, u64 expires)
 134 {
 135         u64 val;
 136
 137         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_RTC2_INT_ENABLE), 0UL);
 138
 139         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_INT_CMPC), -1L);
 140
 141         mmtimer_clr_int_pending(1);
 142
 143         val = ((u64)SGI_MMTIMER_VECTOR << SH_RTC2_INT_CONFIG_IDX_SHFT) |
 144                 ((u64)cpu_physical_id(cpu) <<
 145                         SH_RTC2_INT_CONFIG_PID_SHFT);
 146
 147         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_RTC2_INT_CONFIG), val);
 148
 149         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_RTC2_INT_ENABLE), 1UL);
 150
 151         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_INT_CMPC), expires);
 152 }
 153
 154 /* Setup timer on comparator RTC3 */
 155 static void mmtimer_setup_int_2(int cpu, u64 expires)
 156 {
 157         u64 val;
 158
 159         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_RTC3_INT_ENABLE), 0UL);
 160
 161         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_INT_CMPD), -1L);
 162
 163         mmtimer_clr_int_pending(2);
 164
 165         val = ((u64)SGI_MMTIMER_VECTOR << SH_RTC3_INT_CONFIG_IDX_SHFT) |
 166                 ((u64)cpu_physical_id(cpu) <<
 167                         SH_RTC3_INT_CONFIG_PID_SHFT);
 168
 169         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_RTC3_INT_CONFIG), val);
 170
 171         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_RTC3_INT_ENABLE), 1UL);
 172
 173         HUB_S((u64 *)LOCAL_MMR_ADDR(SH_INT_CMPD), expires);
 174 }
 175
 176 /*
 177  * This function must be called with interrupts disabled and preemption off
 178  * in order to insure that the setup succeeds in a deterministic time frame.
 179  * It will check if the interrupt setup succeeded.
 180  */
 181 static int mmtimer_setup(int cpu, int comparator, unsigned long expires,
 182         u64 *set_completion_time)
 183 {
 184         switch (comparator) {
 185         case 0:
 186                 mmtimer_setup_int_0(cpu, expires);
 187                 break;
 188         case 1:
 189                 mmtimer_setup_int_1(cpu, expires);
 190                 break;
 191         case 2:
 192                 mmtimer_setup_int_2(cpu, expires);
 193                 break;
 194         }
 195         /* We might've missed our expiration time */
 196         *set_completion_time = rtc_time();
 197         if (*set_completion_time <= expires)
 198                 return 1;
 199
 200         /*
 201          * If an interrupt is already pending then its okay
 202          * if not then we failed
 203          */
 204         return mmtimer_int_pending(comparator);
 205 }
 206
 207 static int mmtimer_disable_int(long nasid, int comparator)
 208 {
 209         switch (comparator) {
 210         case 0:
 211                 nasid == -1 ? HUB_S((u64 *)LOCAL_MMR_ADDR(SH_RTC1_INT_ENABLE),
 212                         0UL) : REMOTE_HUB_S(nasid, SH_RTC1_INT_ENABLE, 0UL);
 213                 break;
 214         case 1:
 215                 nasid == -1 ? HUB_S((u64 *)LOCAL_MMR_ADDR(SH_RTC2_INT_ENABLE),
 216                         0UL) : REMOTE_HUB_S(nasid, SH_RTC2_INT_ENABLE, 0UL);
 217                 break;
 218         case 2:
 219                 nasid == -1 ? HUB_S((u64 *)LOCAL_MMR_ADDR(SH_RTC3_INT_ENABLE),
 220                         0UL) : REMOTE_HUB_S(nasid, SH_RTC3_INT_ENABLE, 0UL);
 221                 break;
 222         default:
 223                 return -EFAULT;
 224         }
 225         return 0;
 226 }
 227
 228 #define COMPARATOR      1               /* The comparator to use */
 229
 230 #define TIMER_OFF       0xbadcabLL      /* Timer is not setup */
 231 #define TIMER_SET       0               /* Comparator is set for this timer */
 232
 233 #define MMTIMER_INTERVAL_RETRY_INCREMENT_DEFAULT 40
 234
 235 /* There is one of these for each timer */
 236 struct mmtimer {
 237         struct rb_node list;
 238         struct k_itimer *timer;
 239         int cpu;
 240 };
 241
 242 struct mmtimer_node {
 243         spinlock_t lock ____cacheline_aligned;
 244         struct rb_root timer_head;
 245         struct rb_node *next;
 246         struct tasklet_struct tasklet;
 247 };
 248 static struct mmtimer_node *timers;
 249
 250 static unsigned mmtimer_interval_retry_increment =
 251         MMTIMER_INTERVAL_RETRY_INCREMENT_DEFAULT;
 252 module_param(mmtimer_interval_retry_increment, uint, 0644);
 253 MODULE_PARM_DESC(mmtimer_interval_retry_increment,
 254         "RTC ticks to add to expiration on interval retry (default 40)");
 255
 256 /*
 257  * Add a new mmtimer struct to the node's mmtimer list.
 258  * This function assumes the struct mmtimer_node is locked.
 259  */
 260 static void mmtimer_add_list(struct mmtimer *n)
 261 {
 262         int nodeid = n->timer->it.mmtimer.node;
 263         unsigned long expires = n->timer->it.mmtimer.expires;
 264         struct rb_node **link = &timers[nodeid].timer_head.rb_node;
 265         struct rb_node *parent = NULL;
 266         struct mmtimer *x;
 267
 268         /*
 269          * Find the right place in the rbtree:
 270          */
 271         while (*link) {
 272                 parent = *link;
 273                 x = rb_entry(parent, struct mmtimer, list);
 274
 275                 if (expires < x->timer->it.mmtimer.expires)
 276                         link = &(*link)->rb_left;
 277                 else
 278                         link = &(*link)->rb_right;
 279         }
 280
 281         /*
 282          * Insert the timer to the rbtree and check whether it
 283          * replaces the first pending timer
 284          */
 285         rb_link_node(&n->list, parent, link);
 286         rb_insert_color(&n->list, &timers[nodeid].timer_head);
 287
 288         if (!timers[nodeid].next || expires < rb_entry(timers[nodeid].next,
 289                         struct mmtimer, list)->timer->it.mmtimer.expires)
 290                 timers[nodeid].next = &n->list;
 291 }
 292
 293 /*
 294  * Set the comparator for the next timer.
 295  * This function assumes the struct mmtimer_node is locked.
 296  */
 297 static void mmtimer_set_next_timer(int nodeid)
 298 {
 299         struct mmtimer_node *n = &timers[nodeid];
 300         struct mmtimer *x;
 301         struct k_itimer *t;
 302         u64 expires, exp, set_completion_time;
 303         int i;
 304
 305 restart:
 306         if (n->next == NULL)
 307                 return;
 308
 309         x = rb_entry(n->next, struct mmtimer, list);
 310         t = x->timer;
 311         if (!t->it.mmtimer.incr) {
 312                 /* Not an interval timer */
 313                 if (!mmtimer_setup(x->cpu, COMPARATOR,
 314                                         t->it.mmtimer.expires,
 315                                         &set_completion_time)) {
 316                         /* Late setup, fire now */
 317                         tasklet_schedule(&n->tasklet);
 318                 }
 319                 return;
 320         }
 321
 322         /* Interval timer */
 323         i = 0;
 324         expires = exp = t->it.mmtimer.expires;
 325         while (!mmtimer_setup(x->cpu, COMPARATOR, expires,
 326                                 &set_completion_time)) {
 327                 int to;
 328
 329                 i++;
 330                 expires = set_completion_time +
 331                                 mmtimer_interval_retry_increment + (1 << i);
 332                 /* Calculate overruns as we go. */
 333                 to = ((u64)(expires - exp) / t->it.mmtimer.incr);
 334                 if (to) {
 335                         t->it_overrun += to;
 336                         t->it.mmtimer.expires += t->it.mmtimer.incr * to;
 337                         exp = t->it.mmtimer.expires;
 338                 }
 339                 if (i > 20) {
 340                         printk(KERN_ALERT "mmtimer: cannot reschedule timer\n");
 341                         t->it.mmtimer.clock = TIMER_OFF;
 342                         n->next = rb_next(&x->list);
 343                         rb_erase(&x->list, &n->timer_head);
 344                         kfree(x);
 345                         goto restart;
 346                 }
 347         }
 348 }
 349
 350 /**
 351  * mmtimer_ioctl - ioctl interface for /dev/mmtimer
 352  * @file: file structure for the device
 353  * @cmd: command to execute
 354  * @arg: optional argument to command
 355  *
 356  * Executes the command specified by @cmd.  Returns 0 for success, < 0 for
 357  * failure.
 358  *
 359  * Valid commands:
 360  *
 361  * %MMTIMER_GETOFFSET - Should return the offset (relative to the start
 362  * of the page where the registers are mapped) for the counter in question.
 363  *
 364  * %MMTIMER_GETRES - Returns the resolution of the clock in femto (10^-15)
 365  * seconds
 366  *
 367  * %MMTIMER_GETFREQ - Copies the frequency of the clock in Hz to the address
 368  * specified by @arg
 369  *
 370  * %MMTIMER_GETBITS - Returns the number of bits in the clock's counter
 371  *
 372  * %MMTIMER_MMAPAVAIL - Returns 1 if the registers can be mmap'd into userspace
 373  *
 374  * %MMTIMER_GETCOUNTER - Gets the current value in the counter and places it
 375  * in the address specified by @arg.
 376  */
 377 static long mmtimer_ioctl(struct file *file, unsigned int cmd,
 378                                                 unsigned long arg)
 379 {
 380         int ret = 0;
 381
 382         mutex_lock(&mmtimer_mutex);
 383
 384         switch (cmd) {
 385         case MMTIMER_GETOFFSET: /* offset of the counter */
 386                 /*
 387                  * SN RTC registers are on their own 64k page
 388                  */
 389                 if(PAGE_SIZE <= (1 << 16))
 390                         ret = (((long)RTC_COUNTER_ADDR) & (PAGE_SIZE-1)) / 8;
 391                 else
 392                         ret = -ENOSYS;
 393                 break;
 394
 395         case MMTIMER_GETRES: /* resolution of the clock in 10^-15 s */
 396                 if(copy_to_user((unsigned long __user *)arg,
 397                                 &mmtimer_femtoperiod, sizeof(unsigned long)))
 398                         ret = -EFAULT;
 399                 break;
 400
 401         case MMTIMER_GETFREQ: /* frequency in Hz */
 402                 if(copy_to_user((unsigned long __user *)arg,
 403                                 &sn_rtc_cycles_per_second,
 404                                 sizeof(unsigned long)))
 405                         ret = -EFAULT;
 406                 break;
 407
 408         case MMTIMER_GETBITS: /* number of bits in the clock */
 409                 ret = RTC_BITS;
 410                 break;
 411
 412         case MMTIMER_MMAPAVAIL: /* can we mmap the clock into userspace? */
 413                 ret = (PAGE_SIZE <= (1 << 16)) ? 1 : 0;
 414                 break;
 415
 416         case MMTIMER_GETCOUNTER:
 417                 if(copy_to_user((unsigned long __user *)arg,
 418                                 RTC_COUNTER_ADDR, sizeof(unsigned long)))
 419                         ret = -EFAULT;
 420                 break;
 421         default:
 422                 ret = -ENOTTY;
 423                 break;
 424         }
 425         mutex_unlock(&mmtimer_mutex);
 426         return ret;
 427 }
 428
 429 /**
 430  * mmtimer_mmap - maps the clock's registers into userspace
 431  * @file: file structure for the device
 432  * @vma: VMA to map the registers into
 433  *
 434  * Calls remap_pfn_range() to map the clock's registers into
 435  * the calling process' address space.
 436  */
 437 static int mmtimer_mmap(struct file *file, struct vm_area_struct *vma)
 438 {
 439         unsigned long mmtimer_addr;
 440
 441         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
 442                 return -EINVAL;
 443
 444         if (vma->vm_flags & VM_WRITE)
 445                 return -EPERM;
 446
 447         if (PAGE_SIZE > (1 << 16))
 448                 return -ENOSYS;
 449
 450         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 451
 452         mmtimer_addr = __pa(RTC_COUNTER_ADDR);
 453         mmtimer_addr &= ~(PAGE_SIZE - 1);
 454         mmtimer_addr &= 0xfffffffffffffffUL;
 455
 456         if (remap_pfn_range(vma, vma->vm_start, mmtimer_addr >> PAGE_SHIFT,
 457                                         PAGE_SIZE, vma->vm_page_prot)) {
 458                 printk(KERN_ERR "remap_pfn_range failed in mmtimer.c\n");
 459                 return -EAGAIN;
 460         }
 461
 462         return 0;
 463 }
 464
 465 static struct miscdevice mmtimer_miscdev = {
 466         SGI_MMTIMER,
 467         MMTIMER_NAME,
 468         &mmtimer_fops
 469 };
 470
 471 static struct timespec sgi_clock_offset;
 472 static int sgi_clock_period;
 473
 474 /*
 475  * Posix Timer Interface
 476  */
 477
 478 static struct timespec sgi_clock_offset;
 479 static int sgi_clock_period;
 480
 481 static int sgi_clock_get(clockid_t clockid, struct timespec *tp)
 482 {
 483         u64 nsec;
 484
 485         nsec = rtc_time() * sgi_clock_period
 486                         + sgi_clock_offset.tv_nsec;
 487         *tp = ns_to_timespec(nsec);
 488         tp->tv_sec += sgi_clock_offset.tv_sec;
 489         return 0;
 490 };
 491
 492 static int sgi_clock_set(const clockid_t clockid, const struct timespec *tp)
 493 {
 494
 495         u64 nsec;
 496         u32 rem;
 497
 498         nsec = rtc_time() * sgi_clock_period;
 499
 500         sgi_clock_offset.tv_sec = tp->tv_sec - div_u64_rem(nsec, NSEC_PER_SEC, &rem);
 501
 502         if (rem <= tp->tv_nsec)
 503                 sgi_clock_offset.tv_nsec = tp->tv_sec - rem;
 504         else {
 505                 sgi_clock_offset.tv_nsec = tp->tv_sec + NSEC_PER_SEC - rem;
 506                 sgi_clock_offset.tv_sec--;
 507         }
 508         return 0;
 509 }
 510
 511 /**
 512  * mmtimer_interrupt - timer interrupt handler
 513  * @irq: irq received
 514  * @dev_id: device the irq came from
 515  *
 516  * Called when one of the comarators matches the counter, This
 517  * routine will send signals to processes that have requested
 518  * them.
 519  *
 520  * This interrupt is run in an interrupt context
 521  * by the SHUB. It is therefore safe to locally access SHub
 522  * registers.
 523  */
 524 static irqreturn_t
 525 mmtimer_interrupt(int irq, void *dev_id)
 526 {
 527         unsigned long expires = 0;
 528         int result = IRQ_NONE;
 529         unsigned indx = cpu_to_node(smp_processor_id());
 530         struct mmtimer *base;
 531
 532         spin_lock(&timers[indx].lock);
 533         base = rb_entry(timers[indx].next, struct mmtimer, list);
 534         if (base == NULL) {
 535                 spin_unlock(&timers[indx].lock);
 536                 return result;
 537         }
 538
 539         if (base->cpu == smp_processor_id()) {
 540                 if (base->timer)
 541                         expires = base->timer->it.mmtimer.expires;
 542                 /* expires test won't work with shared irqs */
 543                 if ((mmtimer_int_pending(COMPARATOR) > 0) ||
 544                         (expires && (expires <= rtc_time()))) {
 545                         mmtimer_clr_int_pending(COMPARATOR);
 546                         tasklet_schedule(&timers[indx].tasklet);
 547                         result = IRQ_HANDLED;
 548                 }
 549         }
 550         spin_unlock(&timers[indx].lock);
 551         return result;
 552 }
 553
 554 static void mmtimer_tasklet(unsigned long data)
 555 {
 556         int nodeid = data;
 557         struct mmtimer_node *mn = &timers[nodeid];
 558         struct mmtimer *x;
 559         struct k_itimer *t;
 560         unsigned long flags;
 561
 562         /* Send signal and deal with periodic signals */
 563         spin_lock_irqsave(&mn->lock, flags);
 564         if (!mn->next)
 565                 goto out;
 566
 567         x = rb_entry(mn->next, struct mmtimer, list);
 568         t = x->timer;
 569
 570         if (t->it.mmtimer.clock == TIMER_OFF)
 571                 goto out;
 572
 573         t->it_overrun = 0;
 574
 575         mn->next = rb_next(&x->list);
 576         rb_erase(&x->list, &mn->timer_head);
 577
 578         if (posix_timer_event(t, 0) != 0)
 579                 t->it_overrun++;
 580
 581         if(t->it.mmtimer.incr) {
 582                 t->it.mmtimer.expires += t->it.mmtimer.incr;
 583                 mmtimer_add_list(x);
 584         } else {
 585                 /* Ensure we don't false trigger in mmtimer_interrupt */
 586                 t->it.mmtimer.clock = TIMER_OFF;
 587                 t->it.mmtimer.expires = 0;
 588                 kfree(x);
 589         }
 590         /* Set comparator for next timer, if there is one */
 591         mmtimer_set_next_timer(nodeid);
 592
 593         t->it_overrun_last = t->it_overrun;
 594 out:
 595         spin_unlock_irqrestore(&mn->lock, flags);
 596 }
 597
 598 static int sgi_timer_create(struct k_itimer *timer)
 599 {
 600         /* Insure that a newly created timer is off */
 601         timer->it.mmtimer.clock = TIMER_OFF;
 602         return 0;
 603 }
 604
 605 /* This does not really delete a timer. It just insures
 606  * that the timer is not active
 607  *
 608  * Assumption: it_lock is already held with irq's disabled
 609  */
 610 static int sgi_timer_del(struct k_itimer *timr)
 611 {
 612         cnodeid_t nodeid = timr->it.mmtimer.node;
 613         unsigned long irqflags;
 614
 615         spin_lock_irqsave(&timers[nodeid].lock, irqflags);
 616         if (timr->it.mmtimer.clock != TIMER_OFF) {
 617                 unsigned long expires = timr->it.mmtimer.expires;
 618                 struct rb_node *n = timers[nodeid].timer_head.rb_node;
 619                 struct mmtimer *uninitialized_var(t);
 620                 int r = 0;
 621
 622                 timr->it.mmtimer.clock = TIMER_OFF;
 623                 timr->it.mmtimer.expires = 0;
 624
 625                 while (n) {
 626                         t = rb_entry(n, struct mmtimer, list);
 627                         if (t->timer == timr)
 628                                 break;
 629
 630                         if (expires < t->timer->it.mmtimer.expires)
 631                                 n = n->rb_left;
 632                         else
 633                                 n = n->rb_right;
 634                 }
 635
 636                 if (!n) {
 637                         spin_unlock_irqrestore(&timers[nodeid].lock, irqflags);
 638                         return 0;
 639                 }
 640
 641                 if (timers[nodeid].next == n) {
 642                         timers[nodeid].next = rb_next(n);
 643                         r = 1;
 644                 }
 645
 646                 rb_erase(n, &timers[nodeid].timer_head);
 647                 kfree(t);
 648
 649                 if (r) {
 650                         mmtimer_disable_int(cnodeid_to_nasid(nodeid),
 651                                 COMPARATOR);
 652                         mmtimer_set_next_timer(nodeid);
 653                 }
 654         }
 655         spin_unlock_irqrestore(&timers[nodeid].lock, irqflags);
 656         return 0;
 657 }
 658
 659 /* Assumption: it_lock is already held with irq's disabled */
 660 static void sgi_timer_get(struct k_itimer *timr, struct itimerspec *cur_setting)
 661 {
 662
 663         if (timr->it.mmtimer.clock == TIMER_OFF) {
 664                 cur_setting->it_interval.tv_nsec = 0;
 665                 cur_setting->it_interval.tv_sec = 0;
 666                 cur_setting->it_value.tv_nsec = 0;
 667                 cur_setting->it_value.tv_sec =0;
 668                 return;
 669         }
 670
 671         cur_setting->it_interval = ns_to_timespec(timr->it.mmtimer.incr * sgi_clock_period);
 672         cur_setting->it_value = ns_to_timespec((timr->it.mmtimer.expires - rtc_time()) * sgi_clock_period);
 673 }
 674
 675
 676 static int sgi_timer_set(struct k_itimer *timr, int flags,
 677         struct itimerspec * new_setting,
 678         struct itimerspec * old_setting)
 679 {
 680         unsigned long when, period, irqflags;
 681         int err = 0;
 682         cnodeid_t nodeid;
 683         struct mmtimer *base;
 684         struct rb_node *n;
 685
 686         if (old_setting)
 687                 sgi_timer_get(timr, old_setting);
 688
 689         sgi_timer_del(timr);
 690         when = timespec_to_ns(&new_setting->it_value);
 691         period = timespec_to_ns(&new_setting->it_interval);
 692
 693         if (when == 0)
 694                 /* Clear timer */
 695                 return 0;
 696
 697         base = kmalloc(sizeof(struct mmtimer), GFP_KERNEL);
 698         if (base == NULL)
 699                 return -ENOMEM;
 700
 701         if (flags & TIMER_ABSTIME) {
 702                 struct timespec n;
 703                 unsigned long now;
 704
 705                 getnstimeofday(&n);
 706                 now = timespec_to_ns(&n);
 707                 if (when > now)
 708                         when -= now;
 709                 else
 710                         /* Fire the timer immediately */
 711                         when = 0;
 712         }
 713
 714         /*
 715          * Convert to sgi clock period. Need to keep rtc_time() as near as possible
 716          * to getnstimeofday() in order to be as faithful as possible to the time
 717          * specified.
 718          */
 719         when = (when + sgi_clock_period - 1) / sgi_clock_period + rtc_time();
 720         period = (period + sgi_clock_period - 1)  / sgi_clock_period;
 721
 722         /*
 723          * We are allocating a local SHub comparator. If we would be moved to another
 724          * cpu then another SHub may be local to us. Prohibit that by switching off
 725          * preemption.
 726          */
 727         preempt_disable();
 728
 729         nodeid =  cpu_to_node(smp_processor_id());
 730
 731         /* Lock the node timer structure */
 732         spin_lock_irqsave(&timers[nodeid].lock, irqflags);
 733
 734         base->timer = timr;
 735         base->cpu = smp_processor_id();
 736
 737         timr->it.mmtimer.clock = TIMER_SET;
 738         timr->it.mmtimer.node = nodeid;
 739         timr->it.mmtimer.incr = period;
 740         timr->it.mmtimer.expires = when;
 741
 742         n = timers[nodeid].next;
 743
 744         /* Add the new struct mmtimer to node's timer list */
 745         mmtimer_add_list(base);
 746
 747         if (timers[nodeid].next == n) {
 748                 /* No need to reprogram comparator for now */
 749                 spin_unlock_irqrestore(&timers[nodeid].lock, irqflags);
 750                 preempt_enable();
 751                 return err;
 752         }
 753
 754         /* We need to reprogram the comparator */
 755         if (n)
 756                 mmtimer_disable_int(cnodeid_to_nasid(nodeid), COMPARATOR);
 757
 758         mmtimer_set_next_timer(nodeid);
 759
 760         /* Unlock the node timer structure */
 761         spin_unlock_irqrestore(&timers[nodeid].lock, irqflags);
 762
 763         preempt_enable();
 764
 765         return err;
 766 }
 767
 768 static int sgi_clock_getres(const clockid_t which_clock, struct timespec *tp)
 769 {
 770         tp->tv_sec = 0;
 771         tp->tv_nsec = sgi_clock_period;
 772         return 0;
 773 }
 774
 775 static struct k_clock sgi_clock = {
 776         .clock_set      = sgi_clock_set,
 777         .clock_get      = sgi_clock_get,
 778         .clock_getres   = sgi_clock_getres,
 779         .timer_create   = sgi_timer_create,
 780         .timer_set      = sgi_timer_set,
 781         .timer_del      = sgi_timer_del,
 782         .timer_get      = sgi_timer_get
 783 };
 784
 785 /**
 786  * mmtimer_init - device initialization routine
 787  *
 788  * Does initial setup for the mmtimer device.
 789  */
 790 static int __init mmtimer_init(void)
 791 {
 792         cnodeid_t node, maxn = -1;
 793
 794         if (!ia64_platform_is("sn2"))
 795                 return 0;
 796
 797         /*
 798          * Sanity check the cycles/sec variable
 799          */
 800         if (sn_rtc_cycles_per_second < 100000) {
 801                 printk(KERN_ERR "%s: unable to determine clock frequency\n",
 802                        MMTIMER_NAME);
 803                 goto out1;
 804         }
 805
 806         mmtimer_femtoperiod = ((unsigned long)1E15 + sn_rtc_cycles_per_second /
 807                                2) / sn_rtc_cycles_per_second;
 808
 809         if (request_irq(SGI_MMTIMER_VECTOR, mmtimer_interrupt, IRQF_PERCPU, MMTIMER_NAME, NULL)) {
 810                 printk(KERN_WARNING "%s: unable to allocate interrupt.",
 811                         MMTIMER_NAME);
 812                 goto out1;
 813         }
 814
 815         if (misc_register(&mmtimer_miscdev)) {
 816                 printk(KERN_ERR "%s: failed to register device\n",
 817                        MMTIMER_NAME);
 818                 goto out2;
 819         }
 820
 821         /* Get max numbered node, calculate slots needed */
 822         for_each_online_node(node) {
 823                 maxn = node;
 824         }
 825         maxn++;
 826
 827         /* Allocate list of node ptrs to mmtimer_t's */
 828         timers = kzalloc(sizeof(struct mmtimer_node)*maxn, GFP_KERNEL);
 829         if (!timers) {
 830                 printk(KERN_ERR "%s: failed to allocate memory for device\n",
 831                                 MMTIMER_NAME);
 832                 goto out3;
 833         }
 834
 835         /* Initialize struct mmtimer's for each online node */
 836         for_each_online_node(node) {
 837                 spin_lock_init(&timers[node].lock);
 838                 tasklet_init(&timers[node].tasklet, mmtimer_tasklet,
 839                         (unsigned long) node);
 840         }
 841
 842         sgi_clock_period = NSEC_PER_SEC / sn_rtc_cycles_per_second;
 843         posix_timers_register_clock(CLOCK_SGI_CYCLE, &sgi_clock);
 844
 845         printk(KERN_INFO "%s: v%s, %ld MHz\n", MMTIMER_DESC, MMTIMER_VERSION,
 846                sn_rtc_cycles_per_second/(unsigned long)1E6);
 847
 848         return 0;
 849
 850 out3:
 851         misc_deregister(&mmtimer_miscdev);
 852 out2:
 853         free_irq(SGI_MMTIMER_VECTOR, NULL);
 854 out1:
 855         return -1;
 856 }
 857
 858 module_init(mmtimer_init);