arch/ia64/mm/tlb.c

   1 /*
   2  * TLB support routines.
   3  *
   4  * Copyright (C) 1998-2001, 2003 Hewlett-Packard Co
   5  *      David Mosberger-Tang <davidm@hpl.hp.com>
   6  *
   7  * 08/02/00 A. Mallick <asit.k.mallick@intel.com>
   8  *              Modified RID allocation for SMP
   9  *          Goutham Rao <goutham.rao@intel.com>
  10  *              IPI based ptc implementation and A-step IPI implementation.
  11  * Rohit Seth <rohit.seth@intel.com>
  12  * Ken Chen <kenneth.w.chen@intel.com>
  13  * Christophe de Dinechin <ddd@hp.com>: Avoid ptc.e on memory allocation
  14  * Copyright (C) 2007 Intel Corp
  15  *      Fenghua Yu <fenghua.yu@intel.com>
  16  *      Add multiple ptc.g/ptc.ga instruction support in global tlb purge.
  17  */
  18 #include <linux/module.h>
  19 #include <linux/init.h>
  20 #include <linux/kernel.h>
  21 #include <linux/sched.h>
  22 #include <linux/smp.h>
  23 #include <linux/mm.h>
  24 #include <linux/bootmem.h>
  25
  26 #include <asm/delay.h>
  27 #include <asm/mmu_context.h>
  28 #include <asm/pgalloc.h>
  29 #include <asm/pal.h>
  30 #include <asm/tlbflush.h>
  31 #include <asm/dma.h>
  32 #include <asm/processor.h>
  33 #include <asm/sal.h>
  34 #include <asm/tlb.h>
  35
  36 static struct {
  37         u64 mask;               /* mask of supported purge page-sizes */
  38         unsigned long max_bits; /* log2 of largest supported purge page-size */
  39 } purge;
  40
  41 struct ia64_ctx ia64_ctx = {
  42         .lock = __SPIN_LOCK_UNLOCKED(ia64_ctx.lock),
  43         .next = 1,
  44         .max_ctx = ~0U
  45 };
  46
  47 DEFINE_PER_CPU(u8, ia64_need_tlb_flush);
  48 DEFINE_PER_CPU(u8, ia64_tr_num);  /*Number of TR slots in current processor*/
  49 DEFINE_PER_CPU(u8, ia64_tr_used); /*Max Slot number used by kernel*/
  50
  51 struct ia64_tr_entry *ia64_idtrs[NR_CPUS];
  52
  53 /*
  54  * Initializes the ia64_ctx.bitmap array based on max_ctx+1.
  55  * Called after cpu_init() has setup ia64_ctx.max_ctx based on
  56  * maximum RID that is supported by boot CPU.
  57  */
  58 void __init
  59 mmu_context_init (void)
  60 {
  61         ia64_ctx.bitmap = alloc_bootmem((ia64_ctx.max_ctx+1)>>3);
  62         ia64_ctx.flushmap = alloc_bootmem((ia64_ctx.max_ctx+1)>>3);
  63 }
  64
  65 /*
  66  * Acquire the ia64_ctx.lock before calling this function!
  67  */
  68 void
  69 wrap_mmu_context (struct mm_struct *mm)
  70 {
  71         int i, cpu;
  72         unsigned long flush_bit;
  73
  74         for (i=0; i <= ia64_ctx.max_ctx / BITS_PER_LONG; i++) {
  75                 flush_bit = xchg(&ia64_ctx.flushmap[i], 0);
  76                 ia64_ctx.bitmap[i] ^= flush_bit;
  77         }
  78
  79         /* use offset at 300 to skip daemons */
  80         ia64_ctx.next = find_next_zero_bit(ia64_ctx.bitmap,
  81                                 ia64_ctx.max_ctx, 300);
  82         ia64_ctx.limit = find_next_bit(ia64_ctx.bitmap,
  83                                 ia64_ctx.max_ctx, ia64_ctx.next);
  84
  85         /*
  86          * can't call flush_tlb_all() here because of race condition
  87          * with O(1) scheduler [EF]
  88          */
  89         cpu = get_cpu(); /* prevent preemption/migration */
  90         for_each_online_cpu(i)
  91                 if (i != cpu)
  92                         per_cpu(ia64_need_tlb_flush, i) = 1;
  93         put_cpu();
  94         local_flush_tlb_all();
  95 }
  96
  97 /*
  98  * Implement "spinaphores" ... like counting semaphores, but they
  99  * spin instead of sleeping.  If there are ever any other users for
 100  * this primitive it can be moved up to a spinaphore.h header.
 101  */
 102 struct spinaphore {
 103         unsigned long   ticket;
 104         unsigned long   serve;
 105 };
 106
 107 static inline void spinaphore_init(struct spinaphore *ss, int val)
 108 {
 109         ss->ticket = 0;
 110         ss->serve = val;
 111 }
 112
 113 static inline void down_spin(struct spinaphore *ss)
 114 {
 115         unsigned long t = ia64_fetchadd(1, &ss->ticket, acq), serve;
 116
 117         if (time_before(t, ss->serve))
 118                 return;
 119
 120         ia64_invala();
 121
 122         for (;;) {
 123                 asm volatile ("ld4.c.nc %0=[%1]" : "=r"(serve) : "r"(&ss->serve) : "memory");
 124                 if (time_before(t, serve))
 125                         return;
 126                 cpu_relax();
 127         }
 128 }
 129
 130 static inline void up_spin(struct spinaphore *ss)
 131 {
 132         ia64_fetchadd(1, &ss->serve, rel);
 133 }
 134
 135 static struct spinaphore ptcg_sem;
 136 static u16 nptcg = 1;
 137 static int need_ptcg_sem = 1;
 138 static int toolatetochangeptcgsem = 0;
 139
 140 /*
 141  * Kernel parameter "nptcg=" overrides max number of concurrent global TLB
 142  * purges which is reported from either PAL or SAL PALO.
 143  *
 144  * We don't have sanity checking for nptcg value. It's the user's responsibility
 145  * for valid nptcg value on the platform. Otherwise, kernel may hang in some
 146  * cases.
 147  */
 148 static int __init
 149 set_nptcg(char *str)
 150 {
 151         int value = 0;
 152
 153         get_option(&str, &value);
 154         setup_ptcg_sem(value, NPTCG_FROM_KERNEL_PARAMETER);
 155
 156         return 1;
 157 }
 158
 159 __setup("nptcg=", set_nptcg);
 160
 161 /*
 162  * Maximum number of simultaneous ptc.g purges in the system can
 163  * be defined by PAL_VM_SUMMARY (in which case we should take
 164  * the smallest value for any cpu in the system) or by the PAL
 165  * override table (in which case we should ignore the value from
 166  * PAL_VM_SUMMARY).
 167  *
 168  * Kernel parameter "nptcg=" overrides maximum number of simultanesous ptc.g
 169  * purges defined in either PAL_VM_SUMMARY or PAL override table. In this case,
 170  * we should ignore the value from either PAL_VM_SUMMARY or PAL override table.
 171  *
 172  * Complicating the logic here is the fact that num_possible_cpus()
 173  * isn't fully setup until we start bringing cpus online.
 174  */
 175 void
 176 setup_ptcg_sem(int max_purges, int nptcg_from)
 177 {
 178         static int kp_override;
 179         static int palo_override;
 180         static int firstcpu = 1;
 181
 182         if (toolatetochangeptcgsem) {
 183                 if (nptcg_from == NPTCG_FROM_PAL && max_purges == 0)
 184                         BUG_ON(1 < nptcg);
 185                 else
 186                         BUG_ON(max_purges < nptcg);
 187                 return;
 188         }
 189
 190         if (nptcg_from == NPTCG_FROM_KERNEL_PARAMETER) {
 191                 kp_override = 1;
 192                 nptcg = max_purges;
 193                 goto resetsema;
 194         }
 195         if (kp_override) {
 196                 need_ptcg_sem = num_possible_cpus() > nptcg;
 197                 return;
 198         }
 199
 200         if (nptcg_from == NPTCG_FROM_PALO) {
 201                 palo_override = 1;
 202
 203                 /* In PALO max_purges == 0 really means it! */
 204                 if (max_purges == 0)
 205                         panic("Whoa! Platform does not support global TLB purges.\n");
 206                 nptcg = max_purges;
 207                 if (nptcg == PALO_MAX_TLB_PURGES) {
 208                         need_ptcg_sem = 0;
 209                         return;
 210                 }
 211                 goto resetsema;
 212         }
 213         if (palo_override) {
 214                 if (nptcg != PALO_MAX_TLB_PURGES)
 215                         need_ptcg_sem = (num_possible_cpus() > nptcg);
 216                 return;
 217         }
 218
 219         /* In PAL_VM_SUMMARY max_purges == 0 actually means 1 */
 220         if (max_purges == 0) max_purges = 1;
 221
 222         if (firstcpu) {
 223                 nptcg = max_purges;
 224                 firstcpu = 0;
 225         }
 226         if (max_purges < nptcg)
 227                 nptcg = max_purges;
 228         if (nptcg == PAL_MAX_PURGES) {
 229                 need_ptcg_sem = 0;
 230                 return;
 231         } else
 232                 need_ptcg_sem = (num_possible_cpus() > nptcg);
 233
 234 resetsema:
 235         spinaphore_init(&ptcg_sem, max_purges);
 236 }
 237
 238 void
 239 ia64_global_tlb_purge (struct mm_struct *mm, unsigned long start,
 240                        unsigned long end, unsigned long nbits)
 241 {
 242         struct mm_struct *active_mm = current->active_mm;
 243
 244         toolatetochangeptcgsem = 1;
 245
 246         if (mm != active_mm) {
 247                 /* Restore region IDs for mm */
 248                 if (mm && active_mm) {
 249                         activate_context(mm);
 250                 } else {
 251                         flush_tlb_all();
 252                         return;
 253                 }
 254         }
 255
 256         if (need_ptcg_sem)
 257                 down_spin(&ptcg_sem);
 258
 259         do {
 260                 /*
 261                  * Flush ALAT entries also.
 262                  */
 263                 ia64_ptcga(start, (nbits << 2));
 264                 ia64_srlz_i();
 265                 start += (1UL << nbits);
 266         } while (start < end);
 267
 268         if (need_ptcg_sem)
 269                 up_spin(&ptcg_sem);
 270
 271         if (mm != active_mm) {
 272                 activate_context(active_mm);
 273         }
 274 }
 275
 276 void
 277 local_flush_tlb_all (void)
 278 {
 279         unsigned long i, j, flags, count0, count1, stride0, stride1, addr;
 280
 281         addr    = local_cpu_data->ptce_base;
 282         count0  = local_cpu_data->ptce_count[0];
 283         count1  = local_cpu_data->ptce_count[1];
 284         stride0 = local_cpu_data->ptce_stride[0];
 285         stride1 = local_cpu_data->ptce_stride[1];
 286
 287         local_irq_save(flags);
 288         for (i = 0; i < count0; ++i) {
 289                 for (j = 0; j < count1; ++j) {
 290                         ia64_ptce(addr);
 291                         addr += stride1;
 292                 }
 293                 addr += stride0;
 294         }
 295         local_irq_restore(flags);
 296         ia64_srlz_i();                  /* srlz.i implies srlz.d */
 297 }
 298
 299 void
 300 flush_tlb_range (struct vm_area_struct *vma, unsigned long start,
 301                  unsigned long end)
 302 {
 303         struct mm_struct *mm = vma->vm_mm;
 304         unsigned long size = end - start;
 305         unsigned long nbits;
 306
 307 #ifndef CONFIG_SMP
 308         if (mm != current->active_mm) {
 309                 mm->context = 0;
 310                 return;
 311         }
 312 #endif
 313
 314         nbits = ia64_fls(size + 0xfff);
 315         while (unlikely (((1UL << nbits) & purge.mask) == 0) &&
 316                         (nbits < purge.max_bits))
 317                 ++nbits;
 318         if (nbits > purge.max_bits)
 319                 nbits = purge.max_bits;
 320         start &= ~((1UL << nbits) - 1);
 321
 322         preempt_disable();
 323 #ifdef CONFIG_SMP
 324         if (mm != current->active_mm || cpumask_weight(mm_cpumask(mm)) != 1) {
 325                 platform_global_tlb_purge(mm, start, end, nbits);
 326                 preempt_enable();
 327                 return;
 328         }
 329 #endif
 330         do {
 331                 ia64_ptcl(start, (nbits<<2));
 332                 start += (1UL << nbits);
 333         } while (start < end);
 334         preempt_enable();
 335         ia64_srlz_i();                  /* srlz.i implies srlz.d */
 336 }
 337 EXPORT_SYMBOL(flush_tlb_range);
 338
 339 void __devinit
 340 ia64_tlb_init (void)
 341 {
 342         ia64_ptce_info_t uninitialized_var(ptce_info); /* GCC be quiet */
 343         u64 tr_pgbits;
 344         long status;
 345         pal_vm_info_1_u_t vm_info_1;
 346         pal_vm_info_2_u_t vm_info_2;
 347         int cpu = smp_processor_id();
 348
 349         if ((status = ia64_pal_vm_page_size(&tr_pgbits, &purge.mask)) != 0) {
 350                 printk(KERN_ERR "PAL_VM_PAGE_SIZE failed with status=%ld; "
 351                        "defaulting to architected purge page-sizes.\n", status);
 352                 purge.mask = 0x115557000UL;
 353         }
 354         purge.max_bits = ia64_fls(purge.mask);
 355
 356         ia64_get_ptce(&ptce_info);
 357         local_cpu_data->ptce_base = ptce_info.base;
 358         local_cpu_data->ptce_count[0] = ptce_info.count[0];
 359         local_cpu_data->ptce_count[1] = ptce_info.count[1];
 360         local_cpu_data->ptce_stride[0] = ptce_info.stride[0];
 361         local_cpu_data->ptce_stride[1] = ptce_info.stride[1];
 362
 363         local_flush_tlb_all();  /* nuke left overs from bootstrapping... */
 364         status = ia64_pal_vm_summary(&vm_info_1, &vm_info_2);
 365
 366         if (status) {
 367                 printk(KERN_ERR "ia64_pal_vm_summary=%ld\n", status);
 368                 per_cpu(ia64_tr_num, cpu) = 8;
 369                 return;
 370         }
 371         per_cpu(ia64_tr_num, cpu) = vm_info_1.pal_vm_info_1_s.max_itr_entry+1;
 372         if (per_cpu(ia64_tr_num, cpu) >
 373                                 (vm_info_1.pal_vm_info_1_s.max_dtr_entry+1))
 374                 per_cpu(ia64_tr_num, cpu) =
 375                                 vm_info_1.pal_vm_info_1_s.max_dtr_entry+1;
 376         if (per_cpu(ia64_tr_num, cpu) > IA64_TR_ALLOC_MAX) {
 377                 static int justonce = 1;
 378                 per_cpu(ia64_tr_num, cpu) = IA64_TR_ALLOC_MAX;
 379                 if (justonce) {
 380                         justonce = 0;
 381                         printk(KERN_DEBUG "TR register number exceeds "
 382                                "IA64_TR_ALLOC_MAX!\n");
 383                 }
 384         }
 385 }
 386
 387 /*
 388  * is_tr_overlap
 389  *
 390  * Check overlap with inserted TRs.
 391  */
 392 static int is_tr_overlap(struct ia64_tr_entry *p, u64 va, u64 log_size)
 393 {
 394         u64 tr_log_size;
 395         u64 tr_end;
 396         u64 va_rr = ia64_get_rr(va);
 397         u64 va_rid = RR_TO_RID(va_rr);
 398         u64 va_end = va + (1<<log_size) - 1;
 399
 400         if (va_rid != RR_TO_RID(p->rr))
 401                 return 0;
 402         tr_log_size = (p->itir & 0xff) >> 2;
 403         tr_end = p->ifa + (1<<tr_log_size) - 1;
 404
 405         if (va > tr_end || p->ifa > va_end)
 406                 return 0;
 407         return 1;
 408
 409 }
 410
 411 /*
 412  * ia64_insert_tr in virtual mode. Allocate a TR slot
 413  *
 414  * target_mask : 0x1 : itr, 0x2 : dtr, 0x3 : idtr
 415  *
 416  * va   : virtual address.
 417  * pte  : pte entries inserted.
 418  * log_size: range to be covered.
 419  *
 420  * Return value:  <0 :  error No.
 421  *
 422  *                >=0 : slot number allocated for TR.
 423  * Must be called with preemption disabled.
 424  */
 425 int ia64_itr_entry(u64 target_mask, u64 va, u64 pte, u64 log_size)
 426 {
 427         int i, r;
 428         unsigned long psr;
 429         struct ia64_tr_entry *p;
 430         int cpu = smp_processor_id();
 431
 432         if (!ia64_idtrs[cpu]) {
 433                 ia64_idtrs[cpu] = kmalloc(2 * IA64_TR_ALLOC_MAX *
 434                                 sizeof (struct ia64_tr_entry), GFP_KERNEL);
 435                 if (!ia64_idtrs[cpu])
 436                         return -ENOMEM;
 437         }
 438         r = -EINVAL;
 439         /*Check overlap with existing TR entries*/
 440         if (target_mask & 0x1) {
 441                 p = ia64_idtrs[cpu];
 442                 for (i = IA64_TR_ALLOC_BASE; i <= per_cpu(ia64_tr_used, cpu);
 443                                                                 i++, p++) {
 444                         if (p->pte & 0x1)
 445                                 if (is_tr_overlap(p, va, log_size)) {
 446                                         printk(KERN_DEBUG "Overlapped Entry"
 447                                                 "Inserted for TR Reigster!!\n");
 448                                         goto out;
 449                         }
 450                 }
 451         }
 452         if (target_mask & 0x2) {
 453                 p = ia64_idtrs[cpu] + IA64_TR_ALLOC_MAX;
 454                 for (i = IA64_TR_ALLOC_BASE; i <= per_cpu(ia64_tr_used, cpu);
 455                                                                 i++, p++) {
 456                         if (p->pte & 0x1)
 457                                 if (is_tr_overlap(p, va, log_size)) {
 458                                         printk(KERN_DEBUG "Overlapped Entry"
 459                                                 "Inserted for TR Reigster!!\n");
 460                                         goto out;
 461                                 }
 462                 }
 463         }
 464
 465         for (i = IA64_TR_ALLOC_BASE; i < per_cpu(ia64_tr_num, cpu); i++) {
 466                 switch (target_mask & 0x3) {
 467                 case 1:
 468                         if (!((ia64_idtrs[cpu] + i)->pte & 0x1))
 469                                 goto found;
 470                         continue;
 471                 case 2:
 472                         if (!((ia64_idtrs[cpu] + IA64_TR_ALLOC_MAX + i)->pte & 0x1))
 473                                 goto found;
 474                         continue;
 475                 case 3:
 476                         if (!((ia64_idtrs[cpu] + i)->pte & 0x1) &&
 477                             !((ia64_idtrs[cpu] + IA64_TR_ALLOC_MAX + i)->pte & 0x1))
 478                                 goto found;
 479                         continue;
 480                 default:
 481                         r = -EINVAL;
 482                         goto out;
 483                 }
 484         }
 485 found:
 486         if (i >= per_cpu(ia64_tr_num, cpu))
 487                 return -EBUSY;
 488
 489         /*Record tr info for mca hander use!*/
 490         if (i > per_cpu(ia64_tr_used, cpu))
 491                 per_cpu(ia64_tr_used, cpu) = i;
 492
 493         psr = ia64_clear_ic();
 494         if (target_mask & 0x1) {
 495                 ia64_itr(0x1, i, va, pte, log_size);
 496                 ia64_srlz_i();
 497                 p = ia64_idtrs[cpu] + i;
 498                 p->ifa = va;
 499                 p->pte = pte;
 500                 p->itir = log_size << 2;
 501                 p->rr = ia64_get_rr(va);
 502         }
 503         if (target_mask & 0x2) {
 504                 ia64_itr(0x2, i, va, pte, log_size);
 505                 ia64_srlz_i();
 506                 p = ia64_idtrs[cpu] + IA64_TR_ALLOC_MAX + i;
 507                 p->ifa = va;
 508                 p->pte = pte;
 509                 p->itir = log_size << 2;
 510                 p->rr = ia64_get_rr(va);
 511         }
 512         ia64_set_psr(psr);
 513         r = i;
 514 out:
 515         return r;
 516 }
 517 EXPORT_SYMBOL_GPL(ia64_itr_entry);
 518
 519 /*
 520  * ia64_purge_tr
 521  *
 522  * target_mask: 0x1: purge itr, 0x2 : purge dtr, 0x3 purge idtr.
 523  * slot: slot number to be freed.
 524  *
 525  * Must be called with preemption disabled.
 526  */
 527 void ia64_ptr_entry(u64 target_mask, int slot)
 528 {
 529         int cpu = smp_processor_id();
 530         int i;
 531         struct ia64_tr_entry *p;
 532
 533         if (slot < IA64_TR_ALLOC_BASE || slot >= per_cpu(ia64_tr_num, cpu))
 534                 return;
 535
 536         if (target_mask & 0x1) {
 537                 p = ia64_idtrs[cpu] + slot;
 538                 if ((p->pte&0x1) && is_tr_overlap(p, p->ifa, p->itir>>2)) {
 539                         p->pte = 0;
 540                         ia64_ptr(0x1, p->ifa, p->itir>>2);
 541                         ia64_srlz_i();
 542                 }
 543         }
 544
 545         if (target_mask & 0x2) {
 546                 p = ia64_idtrs[cpu] + IA64_TR_ALLOC_MAX + slot;
 547                 if ((p->pte & 0x1) && is_tr_overlap(p, p->ifa, p->itir>>2)) {
 548                         p->pte = 0;
 549                         ia64_ptr(0x2, p->ifa, p->itir>>2);
 550                         ia64_srlz_i();
 551                 }
 552         }
 553
 554         for (i = per_cpu(ia64_tr_used, cpu); i >= IA64_TR_ALLOC_BASE; i--) {
 555                 if (((ia64_idtrs[cpu] + i)->pte & 0x1) ||
 556                     ((ia64_idtrs[cpu] + IA64_TR_ALLOC_MAX + i)->pte & 0x1))
 557                         break;
 558         }
 559         per_cpu(ia64_tr_used, cpu) = i;
 560 }
 561 EXPORT_SYMBOL_GPL(ia64_ptr_entry);