arch/ia64/sn/kernel/sn2/sn2_smp.c

   1 /*
   2  * SN2 Platform specific SMP Support
   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) 2000-2006 Silicon Graphics, Inc. All rights reserved.
   9  */
  10
  11 #include <linux/init.h>
  12 #include <linux/kernel.h>
  13 #include <linux/spinlock.h>
  14 #include <linux/threads.h>
  15 #include <linux/sched.h>
  16 #include <linux/mm_types.h>
  17 #include <linux/smp.h>
  18 #include <linux/interrupt.h>
  19 #include <linux/irq.h>
  20 #include <linux/mmzone.h>
  21 #include <linux/module.h>
  22 #include <linux/bitops.h>
  23 #include <linux/nodemask.h>
  24 #include <linux/proc_fs.h>
  25 #include <linux/seq_file.h>
  26
  27 #include <asm/processor.h>
  28 #include <asm/irq.h>
  29 #include <asm/sal.h>
  30 #include <asm/delay.h>
  31 #include <asm/io.h>
  32 #include <asm/smp.h>
  33 #include <asm/tlb.h>
  34 #include <asm/numa.h>
  35 #include <asm/hw_irq.h>
  36 #include <asm/current.h>
  37 #include <asm/sn/sn_cpuid.h>
  38 #include <asm/sn/sn_sal.h>
  39 #include <asm/sn/addrs.h>
  40 #include <asm/sn/shub_mmr.h>
  41 #include <asm/sn/nodepda.h>
  42 #include <asm/sn/rw_mmr.h>
  43 #include <asm/sn/sn_feature_sets.h>
  44
  45 DEFINE_PER_CPU(struct ptc_stats, ptcstats);
  46 DECLARE_PER_CPU(struct ptc_stats, ptcstats);
  47
  48 static  __cacheline_aligned DEFINE_SPINLOCK(sn2_global_ptc_lock);
  49
  50 /* 0 = old algorithm (no IPI flushes), 1 = ipi deadlock flush, 2 = ipi instead of SHUB ptc, >2 = always ipi */
  51 static int sn2_flush_opt = 0;
  52
  53 extern unsigned long
  54 sn2_ptc_deadlock_recovery_core(volatile unsigned long *, unsigned long,
  55                                volatile unsigned long *, unsigned long,
  56                                volatile unsigned long *, unsigned long);
  57 void
  58 sn2_ptc_deadlock_recovery(nodemask_t, short, short, int,
  59                           volatile unsigned long *, unsigned long,
  60                           volatile unsigned long *, unsigned long);
  61
  62 /*
  63  * Note: some is the following is captured here to make degugging easier
  64  * (the macros make more sense if you see the debug patch - not posted)
  65  */
  66 #define sn2_ptctest     0
  67 #define local_node_uses_ptc_ga(sh1)     ((sh1) ? 1 : 0)
  68 #define max_active_pio(sh1)             ((sh1) ? 32 : 7)
  69 #define reset_max_active_on_deadlock()  1
  70 #define PTC_LOCK(sh1)                   ((sh1) ? &sn2_global_ptc_lock : &sn_nodepda->ptc_lock)
  71
  72 struct ptc_stats {
  73         unsigned long ptc_l;
  74         unsigned long change_rid;
  75         unsigned long shub_ptc_flushes;
  76         unsigned long nodes_flushed;
  77         unsigned long deadlocks;
  78         unsigned long deadlocks2;
  79         unsigned long lock_itc_clocks;
  80         unsigned long shub_itc_clocks;
  81         unsigned long shub_itc_clocks_max;
  82         unsigned long shub_ptc_flushes_not_my_mm;
  83         unsigned long shub_ipi_flushes;
  84         unsigned long shub_ipi_flushes_itc_clocks;
  85 };
  86
  87 #define sn2_ptctest     0
  88
  89 static inline unsigned long wait_piowc(void)
  90 {
  91         volatile unsigned long *piows;
  92         unsigned long zeroval, ws;
  93
  94         piows = pda->pio_write_status_addr;
  95         zeroval = pda->pio_write_status_val;
  96         do {
  97                 cpu_relax();
  98         } while (((ws = *piows) & SH_PIO_WRITE_STATUS_PENDING_WRITE_COUNT_MASK) != zeroval);
  99         return (ws & SH_PIO_WRITE_STATUS_WRITE_DEADLOCK_MASK) != 0;
 100 }
 101
 102 /**
 103  * sn_migrate - SN-specific task migration actions
 104  * @task: Task being migrated to new CPU
 105  *
 106  * SN2 PIO writes from separate CPUs are not guaranteed to arrive in order.
 107  * Context switching user threads which have memory-mapped MMIO may cause
 108  * PIOs to issue from separate CPUs, thus the PIO writes must be drained
 109  * from the previous CPU's Shub before execution resumes on the new CPU.
 110  */
 111 void sn_migrate(struct task_struct *task)
 112 {
 113         pda_t *last_pda = pdacpu(task_thread_info(task)->last_cpu);
 114         volatile unsigned long *adr = last_pda->pio_write_status_addr;
 115         unsigned long val = last_pda->pio_write_status_val;
 116
 117         /* Drain PIO writes from old CPU's Shub */
 118         while (unlikely((*adr & SH_PIO_WRITE_STATUS_PENDING_WRITE_COUNT_MASK)
 119                         != val))
 120                 cpu_relax();
 121 }
 122
 123 void sn_tlb_migrate_finish(struct mm_struct *mm)
 124 {
 125         /* flush_tlb_mm is inefficient if more than 1 users of mm */
 126         if (mm == current->mm && mm && atomic_read(&mm->mm_users) == 1)
 127                 flush_tlb_mm(mm);
 128 }
 129
 130 static void
 131 sn2_ipi_flush_all_tlb(struct mm_struct *mm)
 132 {
 133         unsigned long itc;
 134
 135         itc = ia64_get_itc();
 136         smp_flush_tlb_cpumask(*mm_cpumask(mm));
 137         itc = ia64_get_itc() - itc;
 138         __this_cpu_add(ptcstats.shub_ipi_flushes_itc_clocks, itc);
 139         __this_cpu_inc(ptcstats.shub_ipi_flushes);
 140 }
 141
 142 /**
 143  * sn2_global_tlb_purge - globally purge translation cache of virtual address range
 144  * @mm: mm_struct containing virtual address range
 145  * @start: start of virtual address range
 146  * @end: end of virtual address range
 147  * @nbits: specifies number of bytes to purge per instruction (num = 1<<(nbits & 0xfc))
 148  *
 149  * Purges the translation caches of all processors of the given virtual address
 150  * range.
 151  *
 152  * Note:
 153  *      - cpu_vm_mask is a bit mask that indicates which cpus have loaded the context.
 154  *      - cpu_vm_mask is converted into a nodemask of the nodes containing the
 155  *        cpus in cpu_vm_mask.
 156  *      - if only one bit is set in cpu_vm_mask & it is the current cpu & the
 157  *        process is purging its own virtual address range, then only the
 158  *        local TLB needs to be flushed. This flushing can be done using
 159  *        ptc.l. This is the common case & avoids the global spinlock.
 160  *      - if multiple cpus have loaded the context, then flushing has to be
 161  *        done with ptc.g/MMRs under protection of the global ptc_lock.
 162  */
 163
 164 void
 165 sn2_global_tlb_purge(struct mm_struct *mm, unsigned long start,
 166                      unsigned long end, unsigned long nbits)
 167 {
 168         int i, ibegin, shub1, cnode, mynasid, cpu, lcpu = 0, nasid;
 169         int mymm = (mm == current->active_mm && mm == current->mm);
 170         int use_cpu_ptcga;
 171         volatile unsigned long *ptc0, *ptc1;
 172         unsigned long itc, itc2, flags, data0 = 0, data1 = 0, rr_value, old_rr = 0;
 173         short nix;
 174         nodemask_t nodes_flushed;
 175         int active, max_active, deadlock, flush_opt = sn2_flush_opt;
 176
 177         if (flush_opt > 2) {
 178                 sn2_ipi_flush_all_tlb(mm);
 179                 return;
 180         }
 181
 182         nodes_clear(nodes_flushed);
 183         i = 0;
 184
 185         for_each_cpu(cpu, mm_cpumask(mm)) {
 186                 cnode = cpu_to_node(cpu);
 187                 node_set(cnode, nodes_flushed);
 188                 lcpu = cpu;
 189                 i++;
 190         }
 191
 192         if (i == 0)
 193                 return;
 194
 195         preempt_disable();
 196
 197         if (likely(i == 1 && lcpu == smp_processor_id() && mymm)) {
 198                 do {
 199                         ia64_ptcl(start, nbits << 2);
 200                         start += (1UL << nbits);
 201                 } while (start < end);
 202                 ia64_srlz_i();
 203                 __this_cpu_inc(ptcstats.ptc_l);
 204                 preempt_enable();
 205                 return;
 206         }
 207
 208         if (atomic_read(&mm->mm_users) == 1 && mymm) {
 209                 flush_tlb_mm(mm);
 210                 __this_cpu_inc(ptcstats.change_rid);
 211                 preempt_enable();
 212                 return;
 213         }
 214
 215         if (flush_opt == 2) {
 216                 sn2_ipi_flush_all_tlb(mm);
 217                 preempt_enable();
 218                 return;
 219         }
 220
 221         itc = ia64_get_itc();
 222         nix = nodes_weight(nodes_flushed);
 223
 224         rr_value = (mm->context << 3) | REGION_NUMBER(start);
 225
 226         shub1 = is_shub1();
 227         if (shub1) {
 228                 data0 = (1UL << SH1_PTC_0_A_SHFT) |
 229                         (nbits << SH1_PTC_0_PS_SHFT) |
 230                         (rr_value << SH1_PTC_0_RID_SHFT) |
 231                         (1UL << SH1_PTC_0_START_SHFT);
 232                 ptc0 = (long *)GLOBAL_MMR_PHYS_ADDR(0, SH1_PTC_0);
 233                 ptc1 = (long *)GLOBAL_MMR_PHYS_ADDR(0, SH1_PTC_1);
 234         } else {
 235                 data0 = (1UL << SH2_PTC_A_SHFT) |
 236                         (nbits << SH2_PTC_PS_SHFT) |
 237                         (1UL << SH2_PTC_START_SHFT);
 238                 ptc0 = (long *)GLOBAL_MMR_PHYS_ADDR(0, SH2_PTC +
 239                         (rr_value << SH2_PTC_RID_SHFT));
 240                 ptc1 = NULL;
 241         }
 242
 243
 244         mynasid = get_nasid();
 245         use_cpu_ptcga = local_node_uses_ptc_ga(shub1);
 246         max_active = max_active_pio(shub1);
 247
 248         itc = ia64_get_itc();
 249         spin_lock_irqsave(PTC_LOCK(shub1), flags);
 250         itc2 = ia64_get_itc();
 251
 252         __this_cpu_add(ptcstats.lock_itc_clocks, itc2 - itc);
 253         __this_cpu_inc(ptcstats.shub_ptc_flushes);
 254         __this_cpu_add(ptcstats.nodes_flushed, nix);
 255         if (!mymm)
 256                  __this_cpu_inc(ptcstats.shub_ptc_flushes_not_my_mm);
 257
 258         if (use_cpu_ptcga && !mymm) {
 259                 old_rr = ia64_get_rr(start);
 260                 ia64_set_rr(start, (old_rr & 0xff) | (rr_value << 8));
 261                 ia64_srlz_d();
 262         }
 263
 264         wait_piowc();
 265         do {
 266                 if (shub1)
 267                         data1 = start | (1UL << SH1_PTC_1_START_SHFT);
 268                 else
 269                         data0 = (data0 & ~SH2_PTC_ADDR_MASK) | (start & SH2_PTC_ADDR_MASK);
 270                 deadlock = 0;
 271                 active = 0;
 272                 ibegin = 0;
 273                 i = 0;
 274                 for_each_node_mask(cnode, nodes_flushed) {
 275                         nasid = cnodeid_to_nasid(cnode);
 276                         if (use_cpu_ptcga && unlikely(nasid == mynasid)) {
 277                                 ia64_ptcga(start, nbits << 2);
 278                                 ia64_srlz_i();
 279                         } else {
 280                                 ptc0 = CHANGE_NASID(nasid, ptc0);
 281                                 if (ptc1)
 282                                         ptc1 = CHANGE_NASID(nasid, ptc1);
 283                                 pio_atomic_phys_write_mmrs(ptc0, data0, ptc1, data1);
 284                                 active++;
 285                         }
 286                         if (active >= max_active || i == (nix - 1)) {
 287                                 if ((deadlock = wait_piowc())) {
 288                                         if (flush_opt == 1)
 289                                                 goto done;
 290                                         sn2_ptc_deadlock_recovery(nodes_flushed, ibegin, i, mynasid, ptc0, data0, ptc1, data1);
 291                                         if (reset_max_active_on_deadlock())
 292                                                 max_active = 1;
 293                                 }
 294                                 active = 0;
 295                                 ibegin = i + 1;
 296                         }
 297                         i++;
 298                 }
 299                 start += (1UL << nbits);
 300         } while (start < end);
 301
 302 done:
 303         itc2 = ia64_get_itc() - itc2;
 304         __this_cpu_add(ptcstats.shub_itc_clocks, itc2);
 305         if (itc2 > __this_cpu_read(ptcstats.shub_itc_clocks_max))
 306                 __this_cpu_write(ptcstats.shub_itc_clocks_max, itc2);
 307
 308         if (old_rr) {
 309                 ia64_set_rr(start, old_rr);
 310                 ia64_srlz_d();
 311         }
 312
 313         spin_unlock_irqrestore(PTC_LOCK(shub1), flags);
 314
 315         if (flush_opt == 1 && deadlock) {
 316                 __this_cpu_inc(ptcstats.deadlocks);
 317                 sn2_ipi_flush_all_tlb(mm);
 318         }
 319
 320         preempt_enable();
 321 }
 322
 323 /*
 324  * sn2_ptc_deadlock_recovery
 325  *
 326  * Recover from PTC deadlocks conditions. Recovery requires stepping thru each
 327  * TLB flush transaction.  The recovery sequence is somewhat tricky & is
 328  * coded in assembly language.
 329  */
 330
 331 void
 332 sn2_ptc_deadlock_recovery(nodemask_t nodes, short ib, short ie, int mynasid,
 333                           volatile unsigned long *ptc0, unsigned long data0,
 334                           volatile unsigned long *ptc1, unsigned long data1)
 335 {
 336         short nasid, i;
 337         int cnode;
 338         unsigned long *piows, zeroval, n;
 339
 340         __this_cpu_inc(ptcstats.deadlocks);
 341
 342         piows = (unsigned long *) pda->pio_write_status_addr;
 343         zeroval = pda->pio_write_status_val;
 344
 345         i = 0;
 346         for_each_node_mask(cnode, nodes) {
 347                 if (i < ib)
 348                         goto next;
 349
 350                 if (i > ie)
 351                         break;
 352
 353                 nasid = cnodeid_to_nasid(cnode);
 354                 if (local_node_uses_ptc_ga(is_shub1()) && nasid == mynasid)
 355                         goto next;
 356
 357                 ptc0 = CHANGE_NASID(nasid, ptc0);
 358                 if (ptc1)
 359                         ptc1 = CHANGE_NASID(nasid, ptc1);
 360
 361                 n = sn2_ptc_deadlock_recovery_core(ptc0, data0, ptc1, data1, piows, zeroval);
 362                 __this_cpu_add(ptcstats.deadlocks2, n);
 363 next:
 364                 i++;
 365         }
 366
 367 }
 368
 369 /**
 370  * sn_send_IPI_phys - send an IPI to a Nasid and slice
 371  * @nasid: nasid to receive the interrupt (may be outside partition)
 372  * @physid: physical cpuid to receive the interrupt.
 373  * @vector: command to send
 374  * @delivery_mode: delivery mechanism
 375  *
 376  * Sends an IPI (interprocessor interrupt) to the processor specified by
 377  * @physid
 378  *
 379  * @delivery_mode can be one of the following
 380  *
 381  * %IA64_IPI_DM_INT - pend an interrupt
 382  * %IA64_IPI_DM_PMI - pend a PMI
 383  * %IA64_IPI_DM_NMI - pend an NMI
 384  * %IA64_IPI_DM_INIT - pend an INIT interrupt
 385  */
 386 void sn_send_IPI_phys(int nasid, long physid, int vector, int delivery_mode)
 387 {
 388         long val;
 389         unsigned long flags = 0;
 390         volatile long *p;
 391
 392         p = (long *)GLOBAL_MMR_PHYS_ADDR(nasid, SH_IPI_INT);
 393         val = (1UL << SH_IPI_INT_SEND_SHFT) |
 394             (physid << SH_IPI_INT_PID_SHFT) |
 395             ((long)delivery_mode << SH_IPI_INT_TYPE_SHFT) |
 396             ((long)vector << SH_IPI_INT_IDX_SHFT) |
 397             (0x000feeUL << SH_IPI_INT_BASE_SHFT);
 398
 399         mb();
 400         if (enable_shub_wars_1_1()) {
 401                 spin_lock_irqsave(&sn2_global_ptc_lock, flags);
 402         }
 403         pio_phys_write_mmr(p, val);
 404         if (enable_shub_wars_1_1()) {
 405                 wait_piowc();
 406                 spin_unlock_irqrestore(&sn2_global_ptc_lock, flags);
 407         }
 408
 409 }
 410
 411 EXPORT_SYMBOL(sn_send_IPI_phys);
 412
 413 /**
 414  * sn2_send_IPI - send an IPI to a processor
 415  * @cpuid: target of the IPI
 416  * @vector: command to send
 417  * @delivery_mode: delivery mechanism
 418  * @redirect: redirect the IPI?
 419  *
 420  * Sends an IPI (InterProcessor Interrupt) to the processor specified by
 421  * @cpuid.  @vector specifies the command to send, while @delivery_mode can
 422  * be one of the following
 423  *
 424  * %IA64_IPI_DM_INT - pend an interrupt
 425  * %IA64_IPI_DM_PMI - pend a PMI
 426  * %IA64_IPI_DM_NMI - pend an NMI
 427  * %IA64_IPI_DM_INIT - pend an INIT interrupt
 428  */
 429 void sn2_send_IPI(int cpuid, int vector, int delivery_mode, int redirect)
 430 {
 431         long physid;
 432         int nasid;
 433
 434         physid = cpu_physical_id(cpuid);
 435         nasid = cpuid_to_nasid(cpuid);
 436
 437         /* the following is used only when starting cpus at boot time */
 438         if (unlikely(nasid == -1))
 439                 ia64_sn_get_sapic_info(physid, &nasid, NULL, NULL);
 440
 441         sn_send_IPI_phys(nasid, physid, vector, delivery_mode);
 442 }
 443
 444 #ifdef CONFIG_HOTPLUG_CPU
 445 /**
 446  * sn_cpu_disable_allowed - Determine if a CPU can be disabled.
 447  * @cpu - CPU that is requested to be disabled.
 448  *
 449  * CPU disable is only allowed on SHub2 systems running with a PROM
 450  * that supports CPU disable. It is not permitted to disable the boot processor.
 451  */
 452 bool sn_cpu_disable_allowed(int cpu)
 453 {
 454         if (is_shub2() && sn_prom_feature_available(PRF_CPU_DISABLE_SUPPORT)) {
 455                 if (cpu != 0)
 456                         return true;
 457                 else
 458                         printk(KERN_WARNING
 459                               "Disabling the boot processor is not allowed.\n");
 460
 461         } else
 462                 printk(KERN_WARNING
 463                        "CPU disable is not supported on this system.\n");
 464
 465         return false;
 466 }
 467 #endif /* CONFIG_HOTPLUG_CPU */
 468
 469 #ifdef CONFIG_PROC_FS
 470
 471 #define PTC_BASENAME    "sgi_sn/ptc_statistics"
 472
 473 static void *sn2_ptc_seq_start(struct seq_file *file, loff_t * offset)
 474 {
 475         if (*offset < nr_cpu_ids)
 476                 return offset;
 477         return NULL;
 478 }
 479
 480 static void *sn2_ptc_seq_next(struct seq_file *file, void *data, loff_t * offset)
 481 {
 482         (*offset)++;
 483         if (*offset < nr_cpu_ids)
 484                 return offset;
 485         return NULL;
 486 }
 487
 488 static void sn2_ptc_seq_stop(struct seq_file *file, void *data)
 489 {
 490 }
 491
 492 static int sn2_ptc_seq_show(struct seq_file *file, void *data)
 493 {
 494         struct ptc_stats *stat;
 495         int cpu;
 496
 497         cpu = *(loff_t *) data;
 498
 499         if (!cpu) {
 500                 seq_printf(file,
 501                            "# cpu ptc_l newrid ptc_flushes nodes_flushed deadlocks lock_nsec shub_nsec shub_nsec_max not_my_mm deadlock2 ipi_fluches ipi_nsec\n");
 502                 seq_printf(file, "# ptctest %d, flushopt %d\n", sn2_ptctest, sn2_flush_opt);
 503         }
 504
 505         if (cpu < nr_cpu_ids && cpu_online(cpu)) {
 506                 stat = &per_cpu(ptcstats, cpu);
 507                 seq_printf(file, "cpu %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld\n", cpu, stat->ptc_l,
 508                                 stat->change_rid, stat->shub_ptc_flushes, stat->nodes_flushed,
 509                                 stat->deadlocks,
 510                                 1000 * stat->lock_itc_clocks / per_cpu(ia64_cpu_info, cpu).cyc_per_usec,
 511                                 1000 * stat->shub_itc_clocks / per_cpu(ia64_cpu_info, cpu).cyc_per_usec,
 512                                 1000 * stat->shub_itc_clocks_max / per_cpu(ia64_cpu_info, cpu).cyc_per_usec,
 513                                 stat->shub_ptc_flushes_not_my_mm,
 514                                 stat->deadlocks2,
 515                                 stat->shub_ipi_flushes,
 516                                 1000 * stat->shub_ipi_flushes_itc_clocks / per_cpu(ia64_cpu_info, cpu).cyc_per_usec);
 517         }
 518         return 0;
 519 }
 520
 521 static ssize_t sn2_ptc_proc_write(struct file *file, const char __user *user, size_t count, loff_t *data)
 522 {
 523         int cpu;
 524         char optstr[64];
 525
 526         if (count == 0 || count > sizeof(optstr))
 527                 return -EINVAL;
 528         if (copy_from_user(optstr, user, count))
 529                 return -EFAULT;
 530         optstr[count - 1] = '\0';
 531         sn2_flush_opt = simple_strtoul(optstr, NULL, 0);
 532
 533         for_each_online_cpu(cpu)
 534                 memset(&per_cpu(ptcstats, cpu), 0, sizeof(struct ptc_stats));
 535
 536         return count;
 537 }
 538
 539 static const struct seq_operations sn2_ptc_seq_ops = {
 540         .start = sn2_ptc_seq_start,
 541         .next = sn2_ptc_seq_next,
 542         .stop = sn2_ptc_seq_stop,
 543         .show = sn2_ptc_seq_show
 544 };
 545
 546 static int sn2_ptc_proc_open(struct inode *inode, struct file *file)
 547 {
 548         return seq_open(file, &sn2_ptc_seq_ops);
 549 }
 550
 551 static const struct file_operations proc_sn2_ptc_operations = {
 552         .open = sn2_ptc_proc_open,
 553         .read = seq_read,
 554         .write = sn2_ptc_proc_write,
 555         .llseek = seq_lseek,
 556         .release = seq_release,
 557 };
 558
 559 static struct proc_dir_entry *proc_sn2_ptc;
 560
 561 static int __init sn2_ptc_init(void)
 562 {
 563         if (!ia64_platform_is("sn2"))
 564                 return 0;
 565
 566         proc_sn2_ptc = proc_create(PTC_BASENAME, 0444,
 567                                    NULL, &proc_sn2_ptc_operations);
 568         if (!proc_sn2_ptc) {
 569                 printk(KERN_ERR "unable to create %s proc entry", PTC_BASENAME);
 570                 return -EINVAL;
 571         }
 572         spin_lock_init(&sn2_global_ptc_lock);
 573         return 0;
 574 }
 575
 576 static void __exit sn2_ptc_exit(void)
 577 {
 578         remove_proc_entry(PTC_BASENAME, NULL);
 579 }
 580
 581 module_init(sn2_ptc_init);
 582 module_exit(sn2_ptc_exit);
 583 #endif /* CONFIG_PROC_FS */
 584