arch/powerpc/platforms/powernv/pci-ioda.c

   1 /*
   2  * Support PCI/PCIe on PowerNV platforms
   3  *
   4  * Copyright 2011 Benjamin Herrenschmidt, IBM Corp.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the License, or (at your option) any later version.
  10  */
  11
  12 #undef DEBUG
  13
  14 #include <linux/kernel.h>
  15 #include <linux/pci.h>
  16 #include <linux/crash_dump.h>
  17 #include <linux/debugfs.h>
  18 #include <linux/delay.h>
  19 #include <linux/string.h>
  20 #include <linux/init.h>
  21 #include <linux/bootmem.h>
  22 #include <linux/irq.h>
  23 #include <linux/io.h>
  24 #include <linux/msi.h>
  25 #include <linux/memblock.h>
  26
  27 #include <asm/sections.h>
  28 #include <asm/io.h>
  29 #include <asm/prom.h>
  30 #include <asm/pci-bridge.h>
  31 #include <asm/machdep.h>
  32 #include <asm/msi_bitmap.h>
  33 #include <asm/ppc-pci.h>
  34 #include <asm/opal.h>
  35 #include <asm/iommu.h>
  36 #include <asm/tce.h>
  37 #include <asm/xics.h>
  38 #include <asm/debug.h>
  39 #include <asm/firmware.h>
  40 #include <asm/pnv-pci.h>
  41
  42 #include <misc/cxl-base.h>
  43
  44 #include "powernv.h"
  45 #include "pci.h"
  46
  47 /* 256M DMA window, 4K TCE pages, 8 bytes TCE */
  48 #define TCE32_TABLE_SIZE        ((0x10000000 / 0x1000) * 8)
  49
  50 static void pe_level_printk(const struct pnv_ioda_pe *pe, const char *level,
  51                             const char *fmt, ...)
  52 {
  53         struct va_format vaf;
  54         va_list args;
  55         char pfix[32];
  56
  57         va_start(args, fmt);
  58
  59         vaf.fmt = fmt;
  60         vaf.va = &args;
  61
  62         if (pe->flags & PNV_IODA_PE_DEV)
  63                 strlcpy(pfix, dev_name(&pe->pdev->dev), sizeof(pfix));
  64         else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
  65                 sprintf(pfix, "%04x:%02x     ",
  66                         pci_domain_nr(pe->pbus), pe->pbus->number);
  67 #ifdef CONFIG_PCI_IOV
  68         else if (pe->flags & PNV_IODA_PE_VF)
  69                 sprintf(pfix, "%04x:%02x:%2x.%d",
  70                         pci_domain_nr(pe->parent_dev->bus),
  71                         (pe->rid & 0xff00) >> 8,
  72                         PCI_SLOT(pe->rid), PCI_FUNC(pe->rid));
  73 #endif /* CONFIG_PCI_IOV*/
  74
  75         printk("%spci %s: [PE# %.3d] %pV",
  76                level, pfix, pe->pe_number, &vaf);
  77
  78         va_end(args);
  79 }
  80
  81 #define pe_err(pe, fmt, ...)                                    \
  82         pe_level_printk(pe, KERN_ERR, fmt, ##__VA_ARGS__)
  83 #define pe_warn(pe, fmt, ...)                                   \
  84         pe_level_printk(pe, KERN_WARNING, fmt, ##__VA_ARGS__)
  85 #define pe_info(pe, fmt, ...)                                   \
  86         pe_level_printk(pe, KERN_INFO, fmt, ##__VA_ARGS__)
  87
  88 static bool pnv_iommu_bypass_disabled __read_mostly;
  89
  90 static int __init iommu_setup(char *str)
  91 {
  92         if (!str)
  93                 return -EINVAL;
  94
  95         while (*str) {
  96                 if (!strncmp(str, "nobypass", 8)) {
  97                         pnv_iommu_bypass_disabled = true;
  98                         pr_info("PowerNV: IOMMU bypass window disabled.\n");
  99                         break;
 100                 }
 101                 str += strcspn(str, ",");
 102                 if (*str == ',')
 103                         str++;
 104         }
 105
 106         return 0;
 107 }
 108 early_param("iommu", iommu_setup);
 109
 110 /*
 111  * stdcix is only supposed to be used in hypervisor real mode as per
 112  * the architecture spec
 113  */
 114 static inline void __raw_rm_writeq(u64 val, volatile void __iomem *paddr)
 115 {
 116         __asm__ __volatile__("stdcix %0,0,%1"
 117                 : : "r" (val), "r" (paddr) : "memory");
 118 }
 119
 120 static inline bool pnv_pci_is_mem_pref_64(unsigned long flags)
 121 {
 122         return ((flags & (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH)) ==
 123                 (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH));
 124 }
 125
 126 static void pnv_ioda_reserve_pe(struct pnv_phb *phb, int pe_no)
 127 {
 128         if (!(pe_no >= 0 && pe_no < phb->ioda.total_pe)) {
 129                 pr_warn("%s: Invalid PE %d on PHB#%x\n",
 130                         __func__, pe_no, phb->hose->global_number);
 131                 return;
 132         }
 133
 134         if (test_and_set_bit(pe_no, phb->ioda.pe_alloc)) {
 135                 pr_warn("%s: PE %d was assigned on PHB#%x\n",
 136                         __func__, pe_no, phb->hose->global_number);
 137                 return;
 138         }
 139
 140         phb->ioda.pe_array[pe_no].phb = phb;
 141         phb->ioda.pe_array[pe_no].pe_number = pe_no;
 142 }
 143
 144 static int pnv_ioda_alloc_pe(struct pnv_phb *phb)
 145 {
 146         unsigned long pe;
 147
 148         do {
 149                 pe = find_next_zero_bit(phb->ioda.pe_alloc,
 150                                         phb->ioda.total_pe, 0);
 151                 if (pe >= phb->ioda.total_pe)
 152                         return IODA_INVALID_PE;
 153         } while(test_and_set_bit(pe, phb->ioda.pe_alloc));
 154
 155         phb->ioda.pe_array[pe].phb = phb;
 156         phb->ioda.pe_array[pe].pe_number = pe;
 157         return pe;
 158 }
 159
 160 static void pnv_ioda_free_pe(struct pnv_phb *phb, int pe)
 161 {
 162         WARN_ON(phb->ioda.pe_array[pe].pdev);
 163
 164         memset(&phb->ioda.pe_array[pe], 0, sizeof(struct pnv_ioda_pe));
 165         clear_bit(pe, phb->ioda.pe_alloc);
 166 }
 167
 168 /* The default M64 BAR is shared by all PEs */
 169 static int pnv_ioda2_init_m64(struct pnv_phb *phb)
 170 {
 171         const char *desc;
 172         struct resource *r;
 173         s64 rc;
 174
 175         /* Configure the default M64 BAR */
 176         rc = opal_pci_set_phb_mem_window(phb->opal_id,
 177                                          OPAL_M64_WINDOW_TYPE,
 178                                          phb->ioda.m64_bar_idx,
 179                                          phb->ioda.m64_base,
 180                                          0, /* unused */
 181                                          phb->ioda.m64_size);
 182         if (rc != OPAL_SUCCESS) {
 183                 desc = "configuring";
 184                 goto fail;
 185         }
 186
 187         /* Enable the default M64 BAR */
 188         rc = opal_pci_phb_mmio_enable(phb->opal_id,
 189                                       OPAL_M64_WINDOW_TYPE,
 190                                       phb->ioda.m64_bar_idx,
 191                                       OPAL_ENABLE_M64_SPLIT);
 192         if (rc != OPAL_SUCCESS) {
 193                 desc = "enabling";
 194                 goto fail;
 195         }
 196
 197         /* Mark the M64 BAR assigned */
 198         set_bit(phb->ioda.m64_bar_idx, &phb->ioda.m64_bar_alloc);
 199
 200         /*
 201          * Strip off the segment used by the reserved PE, which is
 202          * expected to be 0 or last one of PE capabicity.
 203          */
 204         r = &phb->hose->mem_resources[1];
 205         if (phb->ioda.reserved_pe == 0)
 206                 r->start += phb->ioda.m64_segsize;
 207         else if (phb->ioda.reserved_pe == (phb->ioda.total_pe - 1))
 208                 r->end -= phb->ioda.m64_segsize;
 209         else
 210                 pr_warn("  Cannot strip M64 segment for reserved PE#%d\n",
 211                         phb->ioda.reserved_pe);
 212
 213         return 0;
 214
 215 fail:
 216         pr_warn("  Failure %lld %s M64 BAR#%d\n",
 217                 rc, desc, phb->ioda.m64_bar_idx);
 218         opal_pci_phb_mmio_enable(phb->opal_id,
 219                                  OPAL_M64_WINDOW_TYPE,
 220                                  phb->ioda.m64_bar_idx,
 221                                  OPAL_DISABLE_M64);
 222         return -EIO;
 223 }
 224
 225 static void pnv_ioda2_reserve_m64_pe(struct pnv_phb *phb)
 226 {
 227         resource_size_t sgsz = phb->ioda.m64_segsize;
 228         struct pci_dev *pdev;
 229         struct resource *r;
 230         int base, step, i;
 231
 232         /*
 233          * Root bus always has full M64 range and root port has
 234          * M64 range used in reality. So we're checking root port
 235          * instead of root bus.
 236          */
 237         list_for_each_entry(pdev, &phb->hose->bus->devices, bus_list) {
 238                 for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++) {
 239                         r = &pdev->resource[PCI_BRIDGE_RESOURCES + i];
 240                         if (!r->parent ||
 241                             !pnv_pci_is_mem_pref_64(r->flags))
 242                                 continue;
 243
 244                         base = (r->start - phb->ioda.m64_base) / sgsz;
 245                         for (step = 0; step < resource_size(r) / sgsz; step++)
 246                                 pnv_ioda_reserve_pe(phb, base + step);
 247                 }
 248         }
 249 }
 250
 251 static int pnv_ioda2_pick_m64_pe(struct pnv_phb *phb,
 252                                  struct pci_bus *bus, int all)
 253 {
 254         resource_size_t segsz = phb->ioda.m64_segsize;
 255         struct pci_dev *pdev;
 256         struct resource *r;
 257         struct pnv_ioda_pe *master_pe, *pe;
 258         unsigned long size, *pe_alloc;
 259         bool found;
 260         int start, i, j;
 261
 262         /* Root bus shouldn't use M64 */
 263         if (pci_is_root_bus(bus))
 264                 return IODA_INVALID_PE;
 265
 266         /* We support only one M64 window on each bus */
 267         found = false;
 268         pci_bus_for_each_resource(bus, r, i) {
 269                 if (r && r->parent &&
 270                     pnv_pci_is_mem_pref_64(r->flags)) {
 271                         found = true;
 272                         break;
 273                 }
 274         }
 275
 276         /* No M64 window found ? */
 277         if (!found)
 278                 return IODA_INVALID_PE;
 279
 280         /* Allocate bitmap */
 281         size = _ALIGN_UP(phb->ioda.total_pe / 8, sizeof(unsigned long));
 282         pe_alloc = kzalloc(size, GFP_KERNEL);
 283         if (!pe_alloc) {
 284                 pr_warn("%s: Out of memory !\n",
 285                         __func__);
 286                 return IODA_INVALID_PE;
 287         }
 288
 289         /*
 290          * Figure out reserved PE numbers by the PE
 291          * the its child PEs.
 292          */
 293         start = (r->start - phb->ioda.m64_base) / segsz;
 294         for (i = 0; i < resource_size(r) / segsz; i++)
 295                 set_bit(start + i, pe_alloc);
 296
 297         if (all)
 298                 goto done;
 299
 300         /*
 301          * If the PE doesn't cover all subordinate buses,
 302          * we need subtract from reserved PEs for children.
 303          */
 304         list_for_each_entry(pdev, &bus->devices, bus_list) {
 305                 if (!pdev->subordinate)
 306                         continue;
 307
 308                 pci_bus_for_each_resource(pdev->subordinate, r, i) {
 309                         if (!r || !r->parent ||
 310                             !pnv_pci_is_mem_pref_64(r->flags))
 311                                 continue;
 312
 313                         start = (r->start - phb->ioda.m64_base) / segsz;
 314                         for (j = 0; j < resource_size(r) / segsz ; j++)
 315                                 clear_bit(start + j, pe_alloc);
 316                 }
 317         }
 318
 319         /*
 320          * the current bus might not own M64 window and that's all
 321          * contributed by its child buses. For the case, we needn't
 322          * pick M64 dependent PE#.
 323          */
 324         if (bitmap_empty(pe_alloc, phb->ioda.total_pe)) {
 325                 kfree(pe_alloc);
 326                 return IODA_INVALID_PE;
 327         }
 328
 329         /*
 330          * Figure out the master PE and put all slave PEs to master
 331          * PE's list to form compound PE.
 332          */
 333 done:
 334         master_pe = NULL;
 335         i = -1;
 336         while ((i = find_next_bit(pe_alloc, phb->ioda.total_pe, i + 1)) <
 337                 phb->ioda.total_pe) {
 338                 pe = &phb->ioda.pe_array[i];
 339
 340                 if (!master_pe) {
 341                         pe->flags |= PNV_IODA_PE_MASTER;
 342                         INIT_LIST_HEAD(&pe->slaves);
 343                         master_pe = pe;
 344                 } else {
 345                         pe->flags |= PNV_IODA_PE_SLAVE;
 346                         pe->master = master_pe;
 347                         list_add_tail(&pe->list, &master_pe->slaves);
 348                 }
 349         }
 350
 351         kfree(pe_alloc);
 352         return master_pe->pe_number;
 353 }
 354
 355 static void __init pnv_ioda_parse_m64_window(struct pnv_phb *phb)
 356 {
 357         struct pci_controller *hose = phb->hose;
 358         struct device_node *dn = hose->dn;
 359         struct resource *res;
 360         const u32 *r;
 361         u64 pci_addr;
 362
 363         /* FIXME: Support M64 for P7IOC */
 364         if (phb->type != PNV_PHB_IODA2) {
 365                 pr_info("  Not support M64 window\n");
 366                 return;
 367         }
 368
 369         if (!firmware_has_feature(FW_FEATURE_OPALv3)) {
 370                 pr_info("  Firmware too old to support M64 window\n");
 371                 return;
 372         }
 373
 374         r = of_get_property(dn, "ibm,opal-m64-window", NULL);
 375         if (!r) {
 376                 pr_info("  No <ibm,opal-m64-window> on %s\n",
 377                         dn->full_name);
 378                 return;
 379         }
 380
 381         res = &hose->mem_resources[1];
 382         res->start = of_translate_address(dn, r + 2);
 383         res->end = res->start + of_read_number(r + 4, 2) - 1;
 384         res->flags = (IORESOURCE_MEM | IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
 385         pci_addr = of_read_number(r, 2);
 386         hose->mem_offset[1] = res->start - pci_addr;
 387
 388         phb->ioda.m64_size = resource_size(res);
 389         phb->ioda.m64_segsize = phb->ioda.m64_size / phb->ioda.total_pe;
 390         phb->ioda.m64_base = pci_addr;
 391
 392         pr_info(" MEM64 0x%016llx..0x%016llx -> 0x%016llx\n",
 393                         res->start, res->end, pci_addr);
 394
 395         /* Use last M64 BAR to cover M64 window */
 396         phb->ioda.m64_bar_idx = 15;
 397         phb->init_m64 = pnv_ioda2_init_m64;
 398         phb->reserve_m64_pe = pnv_ioda2_reserve_m64_pe;
 399         phb->pick_m64_pe = pnv_ioda2_pick_m64_pe;
 400 }
 401
 402 static void pnv_ioda_freeze_pe(struct pnv_phb *phb, int pe_no)
 403 {
 404         struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_no];
 405         struct pnv_ioda_pe *slave;
 406         s64 rc;
 407
 408         /* Fetch master PE */
 409         if (pe->flags & PNV_IODA_PE_SLAVE) {
 410                 pe = pe->master;
 411                 if (WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)))
 412                         return;
 413
 414                 pe_no = pe->pe_number;
 415         }
 416
 417         /* Freeze master PE */
 418         rc = opal_pci_eeh_freeze_set(phb->opal_id,
 419                                      pe_no,
 420                                      OPAL_EEH_ACTION_SET_FREEZE_ALL);
 421         if (rc != OPAL_SUCCESS) {
 422                 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
 423                         __func__, rc, phb->hose->global_number, pe_no);
 424                 return;
 425         }
 426
 427         /* Freeze slave PEs */
 428         if (!(pe->flags & PNV_IODA_PE_MASTER))
 429                 return;
 430
 431         list_for_each_entry(slave, &pe->slaves, list) {
 432                 rc = opal_pci_eeh_freeze_set(phb->opal_id,
 433                                              slave->pe_number,
 434                                              OPAL_EEH_ACTION_SET_FREEZE_ALL);
 435                 if (rc != OPAL_SUCCESS)
 436                         pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
 437                                 __func__, rc, phb->hose->global_number,
 438                                 slave->pe_number);
 439         }
 440 }
 441
 442 static int pnv_ioda_unfreeze_pe(struct pnv_phb *phb, int pe_no, int opt)
 443 {
 444         struct pnv_ioda_pe *pe, *slave;
 445         s64 rc;
 446
 447         /* Find master PE */
 448         pe = &phb->ioda.pe_array[pe_no];
 449         if (pe->flags & PNV_IODA_PE_SLAVE) {
 450                 pe = pe->master;
 451                 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
 452                 pe_no = pe->pe_number;
 453         }
 454
 455         /* Clear frozen state for master PE */
 456         rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, opt);
 457         if (rc != OPAL_SUCCESS) {
 458                 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
 459                         __func__, rc, opt, phb->hose->global_number, pe_no);
 460                 return -EIO;
 461         }
 462
 463         if (!(pe->flags & PNV_IODA_PE_MASTER))
 464                 return 0;
 465
 466         /* Clear frozen state for slave PEs */
 467         list_for_each_entry(slave, &pe->slaves, list) {
 468                 rc = opal_pci_eeh_freeze_clear(phb->opal_id,
 469                                              slave->pe_number,
 470                                              opt);
 471                 if (rc != OPAL_SUCCESS) {
 472                         pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
 473                                 __func__, rc, opt, phb->hose->global_number,
 474                                 slave->pe_number);
 475                         return -EIO;
 476                 }
 477         }
 478
 479         return 0;
 480 }
 481
 482 static int pnv_ioda_get_pe_state(struct pnv_phb *phb, int pe_no)
 483 {
 484         struct pnv_ioda_pe *slave, *pe;
 485         u8 fstate, state;
 486         __be16 pcierr;
 487         s64 rc;
 488
 489         /* Sanity check on PE number */
 490         if (pe_no < 0 || pe_no >= phb->ioda.total_pe)
 491                 return OPAL_EEH_STOPPED_PERM_UNAVAIL;
 492
 493         /*
 494          * Fetch the master PE and the PE instance might be
 495          * not initialized yet.
 496          */
 497         pe = &phb->ioda.pe_array[pe_no];
 498         if (pe->flags & PNV_IODA_PE_SLAVE) {
 499                 pe = pe->master;
 500                 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
 501                 pe_no = pe->pe_number;
 502         }
 503
 504         /* Check the master PE */
 505         rc = opal_pci_eeh_freeze_status(phb->opal_id, pe_no,
 506                                         &state, &pcierr, NULL);
 507         if (rc != OPAL_SUCCESS) {
 508                 pr_warn("%s: Failure %lld getting "
 509                         "PHB#%x-PE#%x state\n",
 510                         __func__, rc,
 511                         phb->hose->global_number, pe_no);
 512                 return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
 513         }
 514
 515         /* Check the slave PE */
 516         if (!(pe->flags & PNV_IODA_PE_MASTER))
 517                 return state;
 518
 519         list_for_each_entry(slave, &pe->slaves, list) {
 520                 rc = opal_pci_eeh_freeze_status(phb->opal_id,
 521                                                 slave->pe_number,
 522                                                 &fstate,
 523                                                 &pcierr,
 524                                                 NULL);
 525                 if (rc != OPAL_SUCCESS) {
 526                         pr_warn("%s: Failure %lld getting "
 527                                 "PHB#%x-PE#%x state\n",
 528                                 __func__, rc,
 529                                 phb->hose->global_number, slave->pe_number);
 530                         return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
 531                 }
 532
 533                 /*
 534                  * Override the result based on the ascending
 535                  * priority.
 536                  */
 537                 if (fstate > state)
 538                         state = fstate;
 539         }
 540
 541         return state;
 542 }
 543
 544 /* Currently those 2 are only used when MSIs are enabled, this will change
 545  * but in the meantime, we need to protect them to avoid warnings
 546  */
 547 #ifdef CONFIG_PCI_MSI
 548 static struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev)
 549 {
 550         struct pci_controller *hose = pci_bus_to_host(dev->bus);
 551         struct pnv_phb *phb = hose->private_data;
 552         struct pci_dn *pdn = pci_get_pdn(dev);
 553
 554         if (!pdn)
 555                 return NULL;
 556         if (pdn->pe_number == IODA_INVALID_PE)
 557                 return NULL;
 558         return &phb->ioda.pe_array[pdn->pe_number];
 559 }
 560 #endif /* CONFIG_PCI_MSI */
 561
 562 static int pnv_ioda_set_one_peltv(struct pnv_phb *phb,
 563                                   struct pnv_ioda_pe *parent,
 564                                   struct pnv_ioda_pe *child,
 565                                   bool is_add)
 566 {
 567         const char *desc = is_add ? "adding" : "removing";
 568         uint8_t op = is_add ? OPAL_ADD_PE_TO_DOMAIN :
 569                               OPAL_REMOVE_PE_FROM_DOMAIN;
 570         struct pnv_ioda_pe *slave;
 571         long rc;
 572
 573         /* Parent PE affects child PE */
 574         rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
 575                                 child->pe_number, op);
 576         if (rc != OPAL_SUCCESS) {
 577                 pe_warn(child, "OPAL error %ld %s to parent PELTV\n",
 578                         rc, desc);
 579                 return -ENXIO;
 580         }
 581
 582         if (!(child->flags & PNV_IODA_PE_MASTER))
 583                 return 0;
 584
 585         /* Compound case: parent PE affects slave PEs */
 586         list_for_each_entry(slave, &child->slaves, list) {
 587                 rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
 588                                         slave->pe_number, op);
 589                 if (rc != OPAL_SUCCESS) {
 590                         pe_warn(slave, "OPAL error %ld %s to parent PELTV\n",
 591                                 rc, desc);
 592                         return -ENXIO;
 593                 }
 594         }
 595
 596         return 0;
 597 }
 598
 599 static int pnv_ioda_set_peltv(struct pnv_phb *phb,
 600                               struct pnv_ioda_pe *pe,
 601                               bool is_add)
 602 {
 603         struct pnv_ioda_pe *slave;
 604         struct pci_dev *pdev = NULL;
 605         int ret;
 606
 607         /*
 608          * Clear PE frozen state. If it's master PE, we need
 609          * clear slave PE frozen state as well.
 610          */
 611         if (is_add) {
 612                 opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
 613                                           OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 614                 if (pe->flags & PNV_IODA_PE_MASTER) {
 615                         list_for_each_entry(slave, &pe->slaves, list)
 616                                 opal_pci_eeh_freeze_clear(phb->opal_id,
 617                                                           slave->pe_number,
 618                                                           OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 619                 }
 620         }
 621
 622         /*
 623          * Associate PE in PELT. We need add the PE into the
 624          * corresponding PELT-V as well. Otherwise, the error
 625          * originated from the PE might contribute to other
 626          * PEs.
 627          */
 628         ret = pnv_ioda_set_one_peltv(phb, pe, pe, is_add);
 629         if (ret)
 630                 return ret;
 631
 632         /* For compound PEs, any one affects all of them */
 633         if (pe->flags & PNV_IODA_PE_MASTER) {
 634                 list_for_each_entry(slave, &pe->slaves, list) {
 635                         ret = pnv_ioda_set_one_peltv(phb, slave, pe, is_add);
 636                         if (ret)
 637                                 return ret;
 638                 }
 639         }
 640
 641         if (pe->flags & (PNV_IODA_PE_BUS_ALL | PNV_IODA_PE_BUS))
 642                 pdev = pe->pbus->self;
 643         else if (pe->flags & PNV_IODA_PE_DEV)
 644                 pdev = pe->pdev->bus->self;
 645 #ifdef CONFIG_PCI_IOV
 646         else if (pe->flags & PNV_IODA_PE_VF)
 647                 pdev = pe->parent_dev->bus->self;
 648 #endif /* CONFIG_PCI_IOV */
 649         while (pdev) {
 650                 struct pci_dn *pdn = pci_get_pdn(pdev);
 651                 struct pnv_ioda_pe *parent;
 652
 653                 if (pdn && pdn->pe_number != IODA_INVALID_PE) {
 654                         parent = &phb->ioda.pe_array[pdn->pe_number];
 655                         ret = pnv_ioda_set_one_peltv(phb, parent, pe, is_add);
 656                         if (ret)
 657                                 return ret;
 658                 }
 659
 660                 pdev = pdev->bus->self;
 661         }
 662
 663         return 0;
 664 }
 665
 666 #ifdef CONFIG_PCI_IOV
 667 static int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
 668 {
 669         struct pci_dev *parent;
 670         uint8_t bcomp, dcomp, fcomp;
 671         int64_t rc;
 672         long rid_end, rid;
 673
 674         /* Currently, we just deconfigure VF PE. Bus PE will always there.*/
 675         if (pe->pbus) {
 676                 int count;
 677
 678                 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
 679                 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
 680                 parent = pe->pbus->self;
 681                 if (pe->flags & PNV_IODA_PE_BUS_ALL)
 682                         count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
 683                 else
 684                         count = 1;
 685
 686                 switch(count) {
 687                 case  1: bcomp = OpalPciBusAll;         break;
 688                 case  2: bcomp = OpalPciBus7Bits;       break;
 689                 case  4: bcomp = OpalPciBus6Bits;       break;
 690                 case  8: bcomp = OpalPciBus5Bits;       break;
 691                 case 16: bcomp = OpalPciBus4Bits;       break;
 692                 case 32: bcomp = OpalPciBus3Bits;       break;
 693                 default:
 694                         dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
 695                                 count);
 696                         /* Do an exact match only */
 697                         bcomp = OpalPciBusAll;
 698                 }
 699                 rid_end = pe->rid + (count << 8);
 700         } else {
 701                 if (pe->flags & PNV_IODA_PE_VF)
 702                         parent = pe->parent_dev;
 703                 else
 704                         parent = pe->pdev->bus->self;
 705                 bcomp = OpalPciBusAll;
 706                 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
 707                 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
 708                 rid_end = pe->rid + 1;
 709         }
 710
 711         /* Clear the reverse map */
 712         for (rid = pe->rid; rid < rid_end; rid++)
 713                 phb->ioda.pe_rmap[rid] = 0;
 714
 715         /* Release from all parents PELT-V */
 716         while (parent) {
 717                 struct pci_dn *pdn = pci_get_pdn(parent);
 718                 if (pdn && pdn->pe_number != IODA_INVALID_PE) {
 719                         rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number,
 720                                                 pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
 721                         /* XXX What to do in case of error ? */
 722                 }
 723                 parent = parent->bus->self;
 724         }
 725
 726         opal_pci_eeh_freeze_set(phb->opal_id, pe->pe_number,
 727                                   OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 728
 729         /* Disassociate PE in PELT */
 730         rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number,
 731                                 pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
 732         if (rc)
 733                 pe_warn(pe, "OPAL error %ld remove self from PELTV\n", rc);
 734         rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
 735                              bcomp, dcomp, fcomp, OPAL_UNMAP_PE);
 736         if (rc)
 737                 pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
 738
 739         pe->pbus = NULL;
 740         pe->pdev = NULL;
 741         pe->parent_dev = NULL;
 742
 743         return 0;
 744 }
 745 #endif /* CONFIG_PCI_IOV */
 746
 747 static int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
 748 {
 749         struct pci_dev *parent;
 750         uint8_t bcomp, dcomp, fcomp;
 751         long rc, rid_end, rid;
 752
 753         /* Bus validation ? */
 754         if (pe->pbus) {
 755                 int count;
 756
 757                 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
 758                 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
 759                 parent = pe->pbus->self;
 760                 if (pe->flags & PNV_IODA_PE_BUS_ALL)
 761                         count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
 762                 else
 763                         count = 1;
 764
 765                 switch(count) {
 766                 case  1: bcomp = OpalPciBusAll;         break;
 767                 case  2: bcomp = OpalPciBus7Bits;       break;
 768                 case  4: bcomp = OpalPciBus6Bits;       break;
 769                 case  8: bcomp = OpalPciBus5Bits;       break;
 770                 case 16: bcomp = OpalPciBus4Bits;       break;
 771                 case 32: bcomp = OpalPciBus3Bits;       break;
 772                 default:
 773                         dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
 774                                 count);
 775                         /* Do an exact match only */
 776                         bcomp = OpalPciBusAll;
 777                 }
 778                 rid_end = pe->rid + (count << 8);
 779         } else {
 780 #ifdef CONFIG_PCI_IOV
 781                 if (pe->flags & PNV_IODA_PE_VF)
 782                         parent = pe->parent_dev;
 783                 else
 784 #endif /* CONFIG_PCI_IOV */
 785                         parent = pe->pdev->bus->self;
 786                 bcomp = OpalPciBusAll;
 787                 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
 788                 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
 789                 rid_end = pe->rid + 1;
 790         }
 791
 792         /*
 793          * Associate PE in PELT. We need add the PE into the
 794          * corresponding PELT-V as well. Otherwise, the error
 795          * originated from the PE might contribute to other
 796          * PEs.
 797          */
 798         rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
 799                              bcomp, dcomp, fcomp, OPAL_MAP_PE);
 800         if (rc) {
 801                 pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
 802                 return -ENXIO;
 803         }
 804
 805         /* Configure PELTV */
 806         pnv_ioda_set_peltv(phb, pe, true);
 807
 808         /* Setup reverse map */
 809         for (rid = pe->rid; rid < rid_end; rid++)
 810                 phb->ioda.pe_rmap[rid] = pe->pe_number;
 811
 812         /* Setup one MVTs on IODA1 */
 813         if (phb->type != PNV_PHB_IODA1) {
 814                 pe->mve_number = 0;
 815                 goto out;
 816         }
 817
 818         pe->mve_number = pe->pe_number;
 819         rc = opal_pci_set_mve(phb->opal_id, pe->mve_number, pe->pe_number);
 820         if (rc != OPAL_SUCCESS) {
 821                 pe_err(pe, "OPAL error %ld setting up MVE %d\n",
 822                        rc, pe->mve_number);
 823                 pe->mve_number = -1;
 824         } else {
 825                 rc = opal_pci_set_mve_enable(phb->opal_id,
 826                                              pe->mve_number, OPAL_ENABLE_MVE);
 827                 if (rc) {
 828                         pe_err(pe, "OPAL error %ld enabling MVE %d\n",
 829                                rc, pe->mve_number);
 830                         pe->mve_number = -1;
 831                 }
 832         }
 833
 834 out:
 835         return 0;
 836 }
 837
 838 static void pnv_ioda_link_pe_by_weight(struct pnv_phb *phb,
 839                                        struct pnv_ioda_pe *pe)
 840 {
 841         struct pnv_ioda_pe *lpe;
 842
 843         list_for_each_entry(lpe, &phb->ioda.pe_dma_list, dma_link) {
 844                 if (lpe->dma_weight < pe->dma_weight) {
 845                         list_add_tail(&pe->dma_link, &lpe->dma_link);
 846                         return;
 847                 }
 848         }
 849         list_add_tail(&pe->dma_link, &phb->ioda.pe_dma_list);
 850 }
 851
 852 static unsigned int pnv_ioda_dma_weight(struct pci_dev *dev)
 853 {
 854         /* This is quite simplistic. The "base" weight of a device
 855          * is 10. 0 means no DMA is to be accounted for it.
 856          */
 857
 858         /* If it's a bridge, no DMA */
 859         if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
 860                 return 0;
 861
 862         /* Reduce the weight of slow USB controllers */
 863         if (dev->class == PCI_CLASS_SERIAL_USB_UHCI ||
 864             dev->class == PCI_CLASS_SERIAL_USB_OHCI ||
 865             dev->class == PCI_CLASS_SERIAL_USB_EHCI)
 866                 return 3;
 867
 868         /* Increase the weight of RAID (includes Obsidian) */
 869         if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID)
 870                 return 15;
 871
 872         /* Default */
 873         return 10;
 874 }
 875
 876 #ifdef CONFIG_PCI_IOV
 877 static int pnv_pci_vf_resource_shift(struct pci_dev *dev, int offset)
 878 {
 879         struct pci_dn *pdn = pci_get_pdn(dev);
 880         int i;
 881         struct resource *res, res2;
 882         resource_size_t size;
 883         u16 num_vfs;
 884
 885         if (!dev->is_physfn)
 886                 return -EINVAL;
 887
 888         /*
 889          * "offset" is in VFs.  The M64 windows are sized so that when they
 890          * are segmented, each segment is the same size as the IOV BAR.
 891          * Each segment is in a separate PE, and the high order bits of the
 892          * address are the PE number.  Therefore, each VF's BAR is in a
 893          * separate PE, and changing the IOV BAR start address changes the
 894          * range of PEs the VFs are in.
 895          */
 896         num_vfs = pdn->num_vfs;
 897         for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
 898                 res = &dev->resource[i + PCI_IOV_RESOURCES];
 899                 if (!res->flags || !res->parent)
 900                         continue;
 901
 902                 if (!pnv_pci_is_mem_pref_64(res->flags))
 903                         continue;
 904
 905                 /*
 906                  * The actual IOV BAR range is determined by the start address
 907                  * and the actual size for num_vfs VFs BAR.  This check is to
 908                  * make sure that after shifting, the range will not overlap
 909                  * with another device.
 910                  */
 911                 size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
 912                 res2.flags = res->flags;
 913                 res2.start = res->start + (size * offset);
 914                 res2.end = res2.start + (size * num_vfs) - 1;
 915
 916                 if (res2.end > res->end) {
 917                         dev_err(&dev->dev, "VF BAR%d: %pR would extend past %pR (trying to enable %d VFs shifted by %d)\n",
 918                                 i, &res2, res, num_vfs, offset);
 919                         return -EBUSY;
 920                 }
 921         }
 922
 923         /*
 924          * After doing so, there would be a "hole" in the /proc/iomem when
 925          * offset is a positive value. It looks like the device return some
 926          * mmio back to the system, which actually no one could use it.
 927          */
 928         for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
 929                 res = &dev->resource[i + PCI_IOV_RESOURCES];
 930                 if (!res->flags || !res->parent)
 931                         continue;
 932
 933                 if (!pnv_pci_is_mem_pref_64(res->flags))
 934                         continue;
 935
 936                 size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
 937                 res2 = *res;
 938                 res->start += size * offset;
 939
 940                 dev_info(&dev->dev, "VF BAR%d: %pR shifted to %pR (enabling %d VFs shifted by %d)\n",
 941                          i, &res2, res, num_vfs, offset);
 942                 pci_update_resource(dev, i + PCI_IOV_RESOURCES);
 943         }
 944         return 0;
 945 }
 946 #endif /* CONFIG_PCI_IOV */
 947
 948 #if 0
 949 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev)
 950 {
 951         struct pci_controller *hose = pci_bus_to_host(dev->bus);
 952         struct pnv_phb *phb = hose->private_data;
 953         struct pci_dn *pdn = pci_get_pdn(dev);
 954         struct pnv_ioda_pe *pe;
 955         int pe_num;
 956
 957         if (!pdn) {
 958                 pr_err("%s: Device tree node not associated properly\n",
 959                            pci_name(dev));
 960                 return NULL;
 961         }
 962         if (pdn->pe_number != IODA_INVALID_PE)
 963                 return NULL;
 964
 965         /* PE#0 has been pre-set */
 966         if (dev->bus->number == 0)
 967                 pe_num = 0;
 968         else
 969                 pe_num = pnv_ioda_alloc_pe(phb);
 970         if (pe_num == IODA_INVALID_PE) {
 971                 pr_warning("%s: Not enough PE# available, disabling device\n",
 972                            pci_name(dev));
 973                 return NULL;
 974         }
 975
 976         /* NOTE: We get only one ref to the pci_dev for the pdn, not for the
 977          * pointer in the PE data structure, both should be destroyed at the
 978          * same time. However, this needs to be looked at more closely again
 979          * once we actually start removing things (Hotplug, SR-IOV, ...)
 980          *
 981          * At some point we want to remove the PDN completely anyways
 982          */
 983         pe = &phb->ioda.pe_array[pe_num];
 984         pci_dev_get(dev);
 985         pdn->pcidev = dev;
 986         pdn->pe_number = pe_num;
 987         pe->pdev = dev;
 988         pe->pbus = NULL;
 989         pe->tce32_seg = -1;
 990         pe->mve_number = -1;
 991         pe->rid = dev->bus->number << 8 | pdn->devfn;
 992
 993         pe_info(pe, "Associated device to PE\n");
 994
 995         if (pnv_ioda_configure_pe(phb, pe)) {
 996                 /* XXX What do we do here ? */
 997                 if (pe_num)
 998                         pnv_ioda_free_pe(phb, pe_num);
 999                 pdn->pe_number = IODA_INVALID_PE;
1000                 pe->pdev = NULL;
1001                 pci_dev_put(dev);
1002                 return NULL;
1003         }
1004
1005         /* Assign a DMA weight to the device */
1006         pe->dma_weight = pnv_ioda_dma_weight(dev);
1007         if (pe->dma_weight != 0) {
1008                 phb->ioda.dma_weight += pe->dma_weight;
1009                 phb->ioda.dma_pe_count++;
1010         }
1011
1012         /* Link the PE */
1013         pnv_ioda_link_pe_by_weight(phb, pe);
1014
1015         return pe;
1016 }
1017 #endif /* Useful for SRIOV case */
1018
1019 static void pnv_ioda_setup_same_PE(struct pci_bus *bus, struct pnv_ioda_pe *pe)
1020 {
1021         struct pci_dev *dev;
1022
1023         list_for_each_entry(dev, &bus->devices, bus_list) {
1024                 struct pci_dn *pdn = pci_get_pdn(dev);
1025
1026                 if (pdn == NULL) {
1027                         pr_warn("%s: No device node associated with device !\n",
1028                                 pci_name(dev));
1029                         continue;
1030                 }
1031                 pdn->pe_number = pe->pe_number;
1032                 pe->dma_weight += pnv_ioda_dma_weight(dev);
1033                 if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1034                         pnv_ioda_setup_same_PE(dev->subordinate, pe);
1035         }
1036 }
1037
1038 /*
1039  * There're 2 types of PCI bus sensitive PEs: One that is compromised of
1040  * single PCI bus. Another one that contains the primary PCI bus and its
1041  * subordinate PCI devices and buses. The second type of PE is normally
1042  * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
1043  */
1044 static void pnv_ioda_setup_bus_PE(struct pci_bus *bus, int all)
1045 {
1046         struct pci_controller *hose = pci_bus_to_host(bus);
1047         struct pnv_phb *phb = hose->private_data;
1048         struct pnv_ioda_pe *pe;
1049         int pe_num = IODA_INVALID_PE;
1050
1051         /* Check if PE is determined by M64 */
1052         if (phb->pick_m64_pe)
1053                 pe_num = phb->pick_m64_pe(phb, bus, all);
1054
1055         /* The PE number isn't pinned by M64 */
1056         if (pe_num == IODA_INVALID_PE)
1057                 pe_num = pnv_ioda_alloc_pe(phb);
1058
1059         if (pe_num == IODA_INVALID_PE) {
1060                 pr_warning("%s: Not enough PE# available for PCI bus %04x:%02x\n",
1061                         __func__, pci_domain_nr(bus), bus->number);
1062                 return;
1063         }
1064
1065         pe = &phb->ioda.pe_array[pe_num];
1066         pe->flags |= (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS);
1067         pe->pbus = bus;
1068         pe->pdev = NULL;
1069         pe->tce32_seg = -1;
1070         pe->mve_number = -1;
1071         pe->rid = bus->busn_res.start << 8;
1072         pe->dma_weight = 0;
1073
1074         if (all)
1075                 pe_info(pe, "Secondary bus %d..%d associated with PE#%d\n",
1076                         bus->busn_res.start, bus->busn_res.end, pe_num);
1077         else
1078                 pe_info(pe, "Secondary bus %d associated with PE#%d\n",
1079                         bus->busn_res.start, pe_num);
1080
1081         if (pnv_ioda_configure_pe(phb, pe)) {
1082                 /* XXX What do we do here ? */
1083                 if (pe_num)
1084                         pnv_ioda_free_pe(phb, pe_num);
1085                 pe->pbus = NULL;
1086                 return;
1087         }
1088
1089         pe->tce32_table = kzalloc_node(sizeof(struct iommu_table),
1090                         GFP_KERNEL, hose->node);
1091         pe->tce32_table->data = pe;
1092
1093         /* Associate it with all child devices */
1094         pnv_ioda_setup_same_PE(bus, pe);
1095
1096         /* Put PE to the list */
1097         list_add_tail(&pe->list, &phb->ioda.pe_list);
1098
1099         /* Account for one DMA PE if at least one DMA capable device exist
1100          * below the bridge
1101          */
1102         if (pe->dma_weight != 0) {
1103                 phb->ioda.dma_weight += pe->dma_weight;
1104                 phb->ioda.dma_pe_count++;
1105         }
1106
1107         /* Link the PE */
1108         pnv_ioda_link_pe_by_weight(phb, pe);
1109 }
1110
1111 static void pnv_ioda_setup_PEs(struct pci_bus *bus)
1112 {
1113         struct pci_dev *dev;
1114
1115         pnv_ioda_setup_bus_PE(bus, 0);
1116
1117         list_for_each_entry(dev, &bus->devices, bus_list) {
1118                 if (dev->subordinate) {
1119                         if (pci_pcie_type(dev) == PCI_EXP_TYPE_PCI_BRIDGE)
1120                                 pnv_ioda_setup_bus_PE(dev->subordinate, 1);
1121                         else
1122                                 pnv_ioda_setup_PEs(dev->subordinate);
1123                 }
1124         }
1125 }
1126
1127 /*
1128  * Configure PEs so that the downstream PCI buses and devices
1129  * could have their associated PE#. Unfortunately, we didn't
1130  * figure out the way to identify the PLX bridge yet. So we
1131  * simply put the PCI bus and the subordinate behind the root
1132  * port to PE# here. The game rule here is expected to be changed
1133  * as soon as we can detected PLX bridge correctly.
1134  */
1135 static void pnv_pci_ioda_setup_PEs(void)
1136 {
1137         struct pci_controller *hose, *tmp;
1138         struct pnv_phb *phb;
1139
1140         list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1141                 phb = hose->private_data;
1142
1143                 /* M64 layout might affect PE allocation */
1144                 if (phb->reserve_m64_pe)
1145                         phb->reserve_m64_pe(phb);
1146
1147                 pnv_ioda_setup_PEs(hose->bus);
1148         }
1149 }
1150
1151 #ifdef CONFIG_PCI_IOV
1152 static int pnv_pci_vf_release_m64(struct pci_dev *pdev)
1153 {
1154         struct pci_bus        *bus;
1155         struct pci_controller *hose;
1156         struct pnv_phb        *phb;
1157         struct pci_dn         *pdn;
1158         int                    i, j;
1159
1160         bus = pdev->bus;
1161         hose = pci_bus_to_host(bus);
1162         phb = hose->private_data;
1163         pdn = pci_get_pdn(pdev);
1164
1165         for (i = 0; i < PCI_SRIOV_NUM_BARS; i++)
1166                 for (j = 0; j < M64_PER_IOV; j++) {
1167                         if (pdn->m64_wins[i][j] == IODA_INVALID_M64)
1168                                 continue;
1169                         opal_pci_phb_mmio_enable(phb->opal_id,
1170                                 OPAL_M64_WINDOW_TYPE, pdn->m64_wins[i][j], 0);
1171                         clear_bit(pdn->m64_wins[i][j], &phb->ioda.m64_bar_alloc);
1172                         pdn->m64_wins[i][j] = IODA_INVALID_M64;
1173                 }
1174
1175         return 0;
1176 }
1177
1178 static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)
1179 {
1180         struct pci_bus        *bus;
1181         struct pci_controller *hose;
1182         struct pnv_phb        *phb;
1183         struct pci_dn         *pdn;
1184         unsigned int           win;
1185         struct resource       *res;
1186         int                    i, j;
1187         int64_t                rc;
1188         int                    total_vfs;
1189         resource_size_t        size, start;
1190         int                    pe_num;
1191         int                    vf_groups;
1192         int                    vf_per_group;
1193
1194         bus = pdev->bus;
1195         hose = pci_bus_to_host(bus);
1196         phb = hose->private_data;
1197         pdn = pci_get_pdn(pdev);
1198         total_vfs = pci_sriov_get_totalvfs(pdev);
1199
1200         /* Initialize the m64_wins to IODA_INVALID_M64 */
1201         for (i = 0; i < PCI_SRIOV_NUM_BARS; i++)
1202                 for (j = 0; j < M64_PER_IOV; j++)
1203                         pdn->m64_wins[i][j] = IODA_INVALID_M64;
1204
1205         if (pdn->m64_per_iov == M64_PER_IOV) {
1206                 vf_groups = (num_vfs <= M64_PER_IOV) ? num_vfs: M64_PER_IOV;
1207                 vf_per_group = (num_vfs <= M64_PER_IOV)? 1:
1208                         roundup_pow_of_two(num_vfs) / pdn->m64_per_iov;
1209         } else {
1210                 vf_groups = 1;
1211                 vf_per_group = 1;
1212         }
1213
1214         for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
1215                 res = &pdev->resource[i + PCI_IOV_RESOURCES];
1216                 if (!res->flags || !res->parent)
1217                         continue;
1218
1219                 if (!pnv_pci_is_mem_pref_64(res->flags))
1220                         continue;
1221
1222                 for (j = 0; j < vf_groups; j++) {
1223                         do {
1224                                 win = find_next_zero_bit(&phb->ioda.m64_bar_alloc,
1225                                                 phb->ioda.m64_bar_idx + 1, 0);
1226
1227                                 if (win >= phb->ioda.m64_bar_idx + 1)
1228                                         goto m64_failed;
1229                         } while (test_and_set_bit(win, &phb->ioda.m64_bar_alloc));
1230
1231                         pdn->m64_wins[i][j] = win;
1232
1233                         if (pdn->m64_per_iov == M64_PER_IOV) {
1234                                 size = pci_iov_resource_size(pdev,
1235                                                         PCI_IOV_RESOURCES + i);
1236                                 size = size * vf_per_group;
1237                                 start = res->start + size * j;
1238                         } else {
1239                                 size = resource_size(res);
1240                                 start = res->start;
1241                         }
1242
1243                         /* Map the M64 here */
1244                         if (pdn->m64_per_iov == M64_PER_IOV) {
1245                                 pe_num = pdn->offset + j;
1246                                 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
1247                                                 pe_num, OPAL_M64_WINDOW_TYPE,
1248                                                 pdn->m64_wins[i][j], 0);
1249                         }
1250
1251                         rc = opal_pci_set_phb_mem_window(phb->opal_id,
1252                                                  OPAL_M64_WINDOW_TYPE,
1253                                                  pdn->m64_wins[i][j],
1254                                                  start,
1255                                                  0, /* unused */
1256                                                  size);
1257
1258
1259                         if (rc != OPAL_SUCCESS) {
1260                                 dev_err(&pdev->dev, "Failed to map M64 window #%d: %lld\n",
1261                                         win, rc);
1262                                 goto m64_failed;
1263                         }
1264
1265                         if (pdn->m64_per_iov == M64_PER_IOV)
1266                                 rc = opal_pci_phb_mmio_enable(phb->opal_id,
1267                                      OPAL_M64_WINDOW_TYPE, pdn->m64_wins[i][j], 2);
1268                         else
1269                                 rc = opal_pci_phb_mmio_enable(phb->opal_id,
1270                                      OPAL_M64_WINDOW_TYPE, pdn->m64_wins[i][j], 1);
1271
1272                         if (rc != OPAL_SUCCESS) {
1273                                 dev_err(&pdev->dev, "Failed to enable M64 window #%d: %llx\n",
1274                                         win, rc);
1275                                 goto m64_failed;
1276                         }
1277                 }
1278         }
1279         return 0;
1280
1281 m64_failed:
1282         pnv_pci_vf_release_m64(pdev);
1283         return -EBUSY;
1284 }
1285
1286 static void pnv_pci_ioda2_release_dma_pe(struct pci_dev *dev, struct pnv_ioda_pe *pe)
1287 {
1288         struct pci_bus        *bus;
1289         struct pci_controller *hose;
1290         struct pnv_phb        *phb;
1291         struct iommu_table    *tbl;
1292         unsigned long         addr;
1293         int64_t               rc;
1294
1295         bus = dev->bus;
1296         hose = pci_bus_to_host(bus);
1297         phb = hose->private_data;
1298         tbl = pe->tce32_table;
1299         addr = tbl->it_base;
1300
1301         opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
1302                                    pe->pe_number << 1, 1, __pa(addr),
1303                                    0, 0x1000);
1304
1305         rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
1306                                         pe->pe_number,
1307                                         (pe->pe_number << 1) + 1,
1308                                         pe->tce_bypass_base,
1309                                         0);
1310         if (rc)
1311                 pe_warn(pe, "OPAL error %ld release DMA window\n", rc);
1312
1313         iommu_free_table(tbl, of_node_full_name(dev->dev.of_node));
1314         free_pages(addr, get_order(TCE32_TABLE_SIZE));
1315         pe->tce32_table = NULL;
1316 }
1317
1318 static void pnv_ioda_release_vf_PE(struct pci_dev *pdev, u16 num_vfs)
1319 {
1320         struct pci_bus        *bus;
1321         struct pci_controller *hose;
1322         struct pnv_phb        *phb;
1323         struct pnv_ioda_pe    *pe, *pe_n;
1324         struct pci_dn         *pdn;
1325         u16                    vf_index;
1326         int64_t                rc;
1327
1328         bus = pdev->bus;
1329         hose = pci_bus_to_host(bus);
1330         phb = hose->private_data;
1331         pdn = pci_get_pdn(pdev);
1332
1333         if (!pdev->is_physfn)
1334                 return;
1335
1336         if (pdn->m64_per_iov == M64_PER_IOV && num_vfs > M64_PER_IOV) {
1337                 int   vf_group;
1338                 int   vf_per_group;
1339                 int   vf_index1;
1340
1341                 vf_per_group = roundup_pow_of_two(num_vfs) / pdn->m64_per_iov;
1342
1343                 for (vf_group = 0; vf_group < M64_PER_IOV; vf_group++)
1344                         for (vf_index = vf_group * vf_per_group;
1345                                 vf_index < (vf_group + 1) * vf_per_group &&
1346                                 vf_index < num_vfs;
1347                                 vf_index++)
1348                                 for (vf_index1 = vf_group * vf_per_group;
1349                                         vf_index1 < (vf_group + 1) * vf_per_group &&
1350                                         vf_index1 < num_vfs;
1351                                         vf_index1++){
1352
1353                                         rc = opal_pci_set_peltv(phb->opal_id,
1354                                                 pdn->offset + vf_index,
1355                                                 pdn->offset + vf_index1,
1356                                                 OPAL_REMOVE_PE_FROM_DOMAIN);
1357
1358                                         if (rc)
1359                                             dev_warn(&pdev->dev, "%s: Failed to unlink same group PE#%d(%lld)\n",
1360                                                 __func__,
1361                                                 pdn->offset + vf_index1, rc);
1362                                 }
1363         }
1364
1365         list_for_each_entry_safe(pe, pe_n, &phb->ioda.pe_list, list) {
1366                 if (pe->parent_dev != pdev)
1367                         continue;
1368
1369                 pnv_pci_ioda2_release_dma_pe(pdev, pe);
1370
1371                 /* Remove from list */
1372                 mutex_lock(&phb->ioda.pe_list_mutex);
1373                 list_del(&pe->list);
1374                 mutex_unlock(&phb->ioda.pe_list_mutex);
1375
1376                 pnv_ioda_deconfigure_pe(phb, pe);
1377
1378                 pnv_ioda_free_pe(phb, pe->pe_number);
1379         }
1380 }
1381
1382 void pnv_pci_sriov_disable(struct pci_dev *pdev)
1383 {
1384         struct pci_bus        *bus;
1385         struct pci_controller *hose;
1386         struct pnv_phb        *phb;
1387         struct pci_dn         *pdn;
1388         struct pci_sriov      *iov;
1389         u16 num_vfs;
1390
1391         bus = pdev->bus;
1392         hose = pci_bus_to_host(bus);
1393         phb = hose->private_data;
1394         pdn = pci_get_pdn(pdev);
1395         iov = pdev->sriov;
1396         num_vfs = pdn->num_vfs;
1397
1398         /* Release VF PEs */
1399         pnv_ioda_release_vf_PE(pdev, num_vfs);
1400
1401         if (phb->type == PNV_PHB_IODA2) {
1402                 if (pdn->m64_per_iov == 1)
1403                         pnv_pci_vf_resource_shift(pdev, -pdn->offset);
1404
1405                 /* Release M64 windows */
1406                 pnv_pci_vf_release_m64(pdev);
1407
1408                 /* Release PE numbers */
1409                 bitmap_clear(phb->ioda.pe_alloc, pdn->offset, num_vfs);
1410                 pdn->offset = 0;
1411         }
1412 }
1413
1414 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
1415                                        struct pnv_ioda_pe *pe);
1416 static void pnv_ioda_setup_vf_PE(struct pci_dev *pdev, u16 num_vfs)
1417 {
1418         struct pci_bus        *bus;
1419         struct pci_controller *hose;
1420         struct pnv_phb        *phb;
1421         struct pnv_ioda_pe    *pe;
1422         int                    pe_num;
1423         u16                    vf_index;
1424         struct pci_dn         *pdn;
1425         int64_t                rc;
1426
1427         bus = pdev->bus;
1428         hose = pci_bus_to_host(bus);
1429         phb = hose->private_data;
1430         pdn = pci_get_pdn(pdev);
1431
1432         if (!pdev->is_physfn)
1433                 return;
1434
1435         /* Reserve PE for each VF */
1436         for (vf_index = 0; vf_index < num_vfs; vf_index++) {
1437                 pe_num = pdn->offset + vf_index;
1438
1439                 pe = &phb->ioda.pe_array[pe_num];
1440                 pe->pe_number = pe_num;
1441                 pe->phb = phb;
1442                 pe->flags = PNV_IODA_PE_VF;
1443                 pe->pbus = NULL;
1444                 pe->parent_dev = pdev;
1445                 pe->tce32_seg = -1;
1446                 pe->mve_number = -1;
1447                 pe->rid = (pci_iov_virtfn_bus(pdev, vf_index) << 8) |
1448                            pci_iov_virtfn_devfn(pdev, vf_index);
1449
1450                 pe_info(pe, "VF %04d:%02d:%02d.%d associated with PE#%d\n",
1451                         hose->global_number, pdev->bus->number,
1452                         PCI_SLOT(pci_iov_virtfn_devfn(pdev, vf_index)),
1453                         PCI_FUNC(pci_iov_virtfn_devfn(pdev, vf_index)), pe_num);
1454
1455                 if (pnv_ioda_configure_pe(phb, pe)) {
1456                         /* XXX What do we do here ? */
1457                         if (pe_num)
1458                                 pnv_ioda_free_pe(phb, pe_num);
1459                         pe->pdev = NULL;
1460                         continue;
1461                 }
1462
1463                 pe->tce32_table = kzalloc_node(sizeof(struct iommu_table),
1464                                 GFP_KERNEL, hose->node);
1465                 pe->tce32_table->data = pe;
1466
1467                 /* Put PE to the list */
1468                 mutex_lock(&phb->ioda.pe_list_mutex);
1469                 list_add_tail(&pe->list, &phb->ioda.pe_list);
1470                 mutex_unlock(&phb->ioda.pe_list_mutex);
1471
1472                 pnv_pci_ioda2_setup_dma_pe(phb, pe);
1473         }
1474
1475         if (pdn->m64_per_iov == M64_PER_IOV && num_vfs > M64_PER_IOV) {
1476                 int   vf_group;
1477                 int   vf_per_group;
1478                 int   vf_index1;
1479
1480                 vf_per_group = roundup_pow_of_two(num_vfs) / pdn->m64_per_iov;
1481
1482                 for (vf_group = 0; vf_group < M64_PER_IOV; vf_group++) {
1483                         for (vf_index = vf_group * vf_per_group;
1484                              vf_index < (vf_group + 1) * vf_per_group &&
1485                              vf_index < num_vfs;
1486                              vf_index++) {
1487                                 for (vf_index1 = vf_group * vf_per_group;
1488                                      vf_index1 < (vf_group + 1) * vf_per_group &&
1489                                      vf_index1 < num_vfs;
1490                                      vf_index1++) {
1491
1492                                         rc = opal_pci_set_peltv(phb->opal_id,
1493                                                 pdn->offset + vf_index,
1494                                                 pdn->offset + vf_index1,
1495                                                 OPAL_ADD_PE_TO_DOMAIN);
1496
1497                                         if (rc)
1498                                             dev_warn(&pdev->dev, "%s: Failed to link same group PE#%d(%lld)\n",
1499                                                 __func__,
1500                                                 pdn->offset + vf_index1, rc);
1501                                 }
1502                         }
1503                 }
1504         }
1505 }
1506
1507 int pnv_pci_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1508 {
1509         struct pci_bus        *bus;
1510         struct pci_controller *hose;
1511         struct pnv_phb        *phb;
1512         struct pci_dn         *pdn;
1513         int                    ret;
1514
1515         bus = pdev->bus;
1516         hose = pci_bus_to_host(bus);
1517         phb = hose->private_data;
1518         pdn = pci_get_pdn(pdev);
1519
1520         if (phb->type == PNV_PHB_IODA2) {
1521                 /* Calculate available PE for required VFs */
1522                 mutex_lock(&phb->ioda.pe_alloc_mutex);
1523                 pdn->offset = bitmap_find_next_zero_area(
1524                         phb->ioda.pe_alloc, phb->ioda.total_pe,
1525                         0, num_vfs, 0);
1526                 if (pdn->offset >= phb->ioda.total_pe) {
1527                         mutex_unlock(&phb->ioda.pe_alloc_mutex);
1528                         dev_info(&pdev->dev, "Failed to enable VF%d\n", num_vfs);
1529                         pdn->offset = 0;
1530                         return -EBUSY;
1531                 }
1532                 bitmap_set(phb->ioda.pe_alloc, pdn->offset, num_vfs);
1533                 pdn->num_vfs = num_vfs;
1534                 mutex_unlock(&phb->ioda.pe_alloc_mutex);
1535
1536                 /* Assign M64 window accordingly */
1537                 ret = pnv_pci_vf_assign_m64(pdev, num_vfs);
1538                 if (ret) {
1539                         dev_info(&pdev->dev, "Not enough M64 window resources\n");
1540                         goto m64_failed;
1541                 }
1542
1543                 /*
1544                  * When using one M64 BAR to map one IOV BAR, we need to shift
1545                  * the IOV BAR according to the PE# allocated to the VFs.
1546                  * Otherwise, the PE# for the VF will conflict with others.
1547                  */
1548                 if (pdn->m64_per_iov == 1) {
1549                         ret = pnv_pci_vf_resource_shift(pdev, pdn->offset);
1550                         if (ret)
1551                                 goto m64_failed;
1552                 }
1553         }
1554
1555         /* Setup VF PEs */
1556         pnv_ioda_setup_vf_PE(pdev, num_vfs);
1557
1558         return 0;
1559
1560 m64_failed:
1561         bitmap_clear(phb->ioda.pe_alloc, pdn->offset, num_vfs);
1562         pdn->offset = 0;
1563
1564         return ret;
1565 }
1566
1567 int pcibios_sriov_disable(struct pci_dev *pdev)
1568 {
1569         pnv_pci_sriov_disable(pdev);
1570
1571         /* Release PCI data */
1572         remove_dev_pci_data(pdev);
1573         return 0;
1574 }
1575
1576 int pcibios_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1577 {
1578         /* Allocate PCI data */
1579         add_dev_pci_data(pdev);
1580
1581         pnv_pci_sriov_enable(pdev, num_vfs);
1582         return 0;
1583 }
1584 #endif /* CONFIG_PCI_IOV */
1585
1586 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb *phb, struct pci_dev *pdev)
1587 {
1588         struct pci_dn *pdn = pci_get_pdn(pdev);
1589         struct pnv_ioda_pe *pe;
1590
1591         /*
1592          * The function can be called while the PE#
1593          * hasn't been assigned. Do nothing for the
1594          * case.
1595          */
1596         if (!pdn || pdn->pe_number == IODA_INVALID_PE)
1597                 return;
1598
1599         pe = &phb->ioda.pe_array[pdn->pe_number];
1600         WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops);
1601         set_iommu_table_base(&pdev->dev, pe->tce32_table);
1602         /*
1603          * Note: iommu_add_device() will fail here as
1604          * for physical PE: the device is already added by now;
1605          * for virtual PE: sysfs entries are not ready yet and
1606          * tce_iommu_bus_notifier will add the device to a group later.
1607          */
1608 }
1609
1610 static int pnv_pci_ioda_dma_set_mask(struct pci_dev *pdev, u64 dma_mask)
1611 {
1612         struct pci_controller *hose = pci_bus_to_host(pdev->bus);
1613         struct pnv_phb *phb = hose->private_data;
1614         struct pci_dn *pdn = pci_get_pdn(pdev);
1615         struct pnv_ioda_pe *pe;
1616         uint64_t top;
1617         bool bypass = false;
1618
1619         if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1620                 return -ENODEV;;
1621
1622         pe = &phb->ioda.pe_array[pdn->pe_number];
1623         if (pe->tce_bypass_enabled) {
1624                 top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1;
1625                 bypass = (dma_mask >= top);
1626         }
1627
1628         if (bypass) {
1629                 dev_info(&pdev->dev, "Using 64-bit DMA iommu bypass\n");
1630                 set_dma_ops(&pdev->dev, &dma_direct_ops);
1631                 set_dma_offset(&pdev->dev, pe->tce_bypass_base);
1632         } else {
1633                 dev_info(&pdev->dev, "Using 32-bit DMA via iommu\n");
1634                 set_dma_ops(&pdev->dev, &dma_iommu_ops);
1635                 set_iommu_table_base(&pdev->dev, pe->tce32_table);
1636         }
1637         *pdev->dev.dma_mask = dma_mask;
1638         return 0;
1639 }
1640
1641 static u64 pnv_pci_ioda_dma_get_required_mask(struct pnv_phb *phb,
1642                                               struct pci_dev *pdev)
1643 {
1644         struct pci_dn *pdn = pci_get_pdn(pdev);
1645         struct pnv_ioda_pe *pe;
1646         u64 end, mask;
1647
1648         if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1649                 return 0;
1650
1651         pe = &phb->ioda.pe_array[pdn->pe_number];
1652         if (!pe->tce_bypass_enabled)
1653                 return __dma_get_required_mask(&pdev->dev);
1654
1655
1656         end = pe->tce_bypass_base + memblock_end_of_DRAM();
1657         mask = 1ULL << (fls64(end) - 1);
1658         mask += mask - 1;
1659
1660         return mask;
1661 }
1662
1663 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe,
1664                                    struct pci_bus *bus)
1665 {
1666         struct pci_dev *dev;
1667
1668         list_for_each_entry(dev, &bus->devices, bus_list) {
1669                 set_iommu_table_base(&dev->dev, pe->tce32_table);
1670                 iommu_add_device(&dev->dev);
1671
1672                 if (dev->subordinate)
1673                         pnv_ioda_setup_bus_dma(pe, dev->subordinate);
1674         }
1675 }
1676
1677 static void pnv_pci_ioda1_tce_invalidate(struct pnv_ioda_pe *pe,
1678                                          struct iommu_table *tbl,
1679                                          __be64 *startp, __be64 *endp, bool rm)
1680 {
1681         __be64 __iomem *invalidate = rm ?
1682                 (__be64 __iomem *)pe->tce_inval_reg_phys :
1683                 (__be64 __iomem *)tbl->it_index;
1684         unsigned long start, end, inc;
1685         const unsigned shift = tbl->it_page_shift;
1686
1687         start = __pa(startp);
1688         end = __pa(endp);
1689
1690         /* BML uses this case for p6/p7/galaxy2: Shift addr and put in node */
1691         if (tbl->it_busno) {
1692                 start <<= shift;
1693                 end <<= shift;
1694                 inc = 128ull << shift;
1695                 start |= tbl->it_busno;
1696                 end |= tbl->it_busno;
1697         } else if (tbl->it_type & TCE_PCI_SWINV_PAIR) {
1698                 /* p7ioc-style invalidation, 2 TCEs per write */
1699                 start |= (1ull << 63);
1700                 end |= (1ull << 63);
1701                 inc = 16;
1702         } else {
1703                 /* Default (older HW) */
1704                 inc = 128;
1705         }
1706
1707         end |= inc - 1; /* round up end to be different than start */
1708
1709         mb(); /* Ensure above stores are visible */
1710         while (start <= end) {
1711                 if (rm)
1712                         __raw_rm_writeq(cpu_to_be64(start), invalidate);
1713                 else
1714                         __raw_writeq(cpu_to_be64(start), invalidate);
1715                 start += inc;
1716         }
1717
1718         /*
1719          * The iommu layer will do another mb() for us on build()
1720          * and we don't care on free()
1721          */
1722 }
1723
1724 static void pnv_pci_ioda2_tce_invalidate(struct pnv_ioda_pe *pe,
1725                                          struct iommu_table *tbl,
1726                                          __be64 *startp, __be64 *endp, bool rm)
1727 {
1728         unsigned long start, end, inc;
1729         __be64 __iomem *invalidate = rm ?
1730                 (__be64 __iomem *)pe->tce_inval_reg_phys :
1731                 (__be64 __iomem *)tbl->it_index;
1732         const unsigned shift = tbl->it_page_shift;
1733
1734         /* We'll invalidate DMA address in PE scope */
1735         start = 0x2ull << 60;
1736         start |= (pe->pe_number & 0xFF);
1737         end = start;
1738
1739         /* Figure out the start, end and step */
1740         inc = tbl->it_offset + (((u64)startp - tbl->it_base) / sizeof(u64));
1741         start |= (inc << shift);
1742         inc = tbl->it_offset + (((u64)endp - tbl->it_base) / sizeof(u64));
1743         end |= (inc << shift);
1744         inc = (0x1ull << shift);
1745         mb();
1746
1747         while (start <= end) {
1748                 if (rm)
1749                         __raw_rm_writeq(cpu_to_be64(start), invalidate);
1750                 else
1751                         __raw_writeq(cpu_to_be64(start), invalidate);
1752                 start += inc;
1753         }
1754 }
1755
1756 void pnv_pci_ioda_tce_invalidate(struct iommu_table *tbl,
1757                                  __be64 *startp, __be64 *endp, bool rm)
1758 {
1759         struct pnv_ioda_pe *pe = tbl->data;
1760         struct pnv_phb *phb = pe->phb;
1761
1762         if (phb->type == PNV_PHB_IODA1)
1763                 pnv_pci_ioda1_tce_invalidate(pe, tbl, startp, endp, rm);
1764         else
1765                 pnv_pci_ioda2_tce_invalidate(pe, tbl, startp, endp, rm);
1766 }
1767
1768 static void pnv_pci_ioda_setup_dma_pe(struct pnv_phb *phb,
1769                                       struct pnv_ioda_pe *pe, unsigned int base,
1770                                       unsigned int segs)
1771 {
1772
1773         struct page *tce_mem = NULL;
1774         const __be64 *swinvp;
1775         struct iommu_table *tbl;
1776         unsigned int i;
1777         int64_t rc;
1778         void *addr;
1779
1780         /* XXX FIXME: Handle 64-bit only DMA devices */
1781         /* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
1782         /* XXX FIXME: Allocate multi-level tables on PHB3 */
1783
1784         /* We shouldn't already have a 32-bit DMA associated */
1785         if (WARN_ON(pe->tce32_seg >= 0))
1786                 return;
1787
1788         /* Grab a 32-bit TCE table */
1789         pe->tce32_seg = base;
1790         pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n",
1791                 (base << 28), ((base + segs) << 28) - 1);
1792
1793         /* XXX Currently, we allocate one big contiguous table for the
1794          * TCEs. We only really need one chunk per 256M of TCE space
1795          * (ie per segment) but that's an optimization for later, it
1796          * requires some added smarts with our get/put_tce implementation
1797          */
1798         tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
1799                                    get_order(TCE32_TABLE_SIZE * segs));
1800         if (!tce_mem) {
1801                 pe_err(pe, " Failed to allocate a 32-bit TCE memory\n");
1802                 goto fail;
1803         }
1804         addr = page_address(tce_mem);
1805         memset(addr, 0, TCE32_TABLE_SIZE * segs);
1806
1807         /* Configure HW */
1808         for (i = 0; i < segs; i++) {
1809                 rc = opal_pci_map_pe_dma_window(phb->opal_id,
1810                                               pe->pe_number,
1811                                               base + i, 1,
1812                                               __pa(addr) + TCE32_TABLE_SIZE * i,
1813                                               TCE32_TABLE_SIZE, 0x1000);
1814                 if (rc) {
1815                         pe_err(pe, " Failed to configure 32-bit TCE table,"
1816                                " err %ld\n", rc);
1817                         goto fail;
1818                 }
1819         }
1820
1821         /* Setup linux iommu table */
1822         tbl = pe->tce32_table;
1823         pnv_pci_setup_iommu_table(tbl, addr, TCE32_TABLE_SIZE * segs,
1824                                   base << 28, IOMMU_PAGE_SHIFT_4K);
1825
1826         /* OPAL variant of P7IOC SW invalidated TCEs */
1827         swinvp = of_get_property(phb->hose->dn, "ibm,opal-tce-kill", NULL);
1828         if (swinvp) {
1829                 /* We need a couple more fields -- an address and a data
1830                  * to or.  Since the bus is only printed out on table free
1831                  * errors, and on the first pass the data will be a relative
1832                  * bus number, print that out instead.
1833                  */
1834                 pe->tce_inval_reg_phys = be64_to_cpup(swinvp);
1835                 tbl->it_index = (unsigned long)ioremap(pe->tce_inval_reg_phys,
1836                                 8);
1837                 tbl->it_type |= (TCE_PCI_SWINV_CREATE |
1838                                  TCE_PCI_SWINV_FREE   |
1839                                  TCE_PCI_SWINV_PAIR);
1840         }
1841         iommu_init_table(tbl, phb->hose->node);
1842
1843         if (pe->flags & PNV_IODA_PE_DEV) {
1844                 iommu_register_group(tbl, phb->hose->global_number,
1845                                      pe->pe_number);
1846                 /*
1847                  * Setting table base here only for carrying iommu_group
1848                  * further down to let iommu_add_device() do the job.
1849                  * pnv_pci_ioda_dma_dev_setup will override it later anyway.
1850                  */
1851                 set_iommu_table_base(&pe->pdev->dev, tbl);
1852                 iommu_add_device(&pe->pdev->dev);
1853         } else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)) {
1854                 iommu_register_group(tbl, phb->hose->global_number,
1855                                      pe->pe_number);
1856                 pnv_ioda_setup_bus_dma(pe, pe->pbus);
1857         } else if (pe->flags & PNV_IODA_PE_VF) {
1858                 iommu_register_group(tbl, phb->hose->global_number,
1859                                      pe->pe_number);
1860         }
1861
1862         return;
1863  fail:
1864         /* XXX Failure: Try to fallback to 64-bit only ? */
1865         if (pe->tce32_seg >= 0)
1866                 pe->tce32_seg = -1;
1867         if (tce_mem)
1868                 __free_pages(tce_mem, get_order(TCE32_TABLE_SIZE * segs));
1869 }
1870
1871 static void pnv_pci_ioda2_set_bypass(struct iommu_table *tbl, bool enable)
1872 {
1873         struct pnv_ioda_pe *pe = tbl->data;
1874         uint16_t window_id = (pe->pe_number << 1 ) + 1;
1875         int64_t rc;
1876
1877         pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis");
1878         if (enable) {
1879                 phys_addr_t top = memblock_end_of_DRAM();
1880
1881                 top = roundup_pow_of_two(top);
1882                 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
1883                                                      pe->pe_number,
1884                                                      window_id,
1885                                                      pe->tce_bypass_base,
1886                                                      top);
1887         } else {
1888                 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
1889                                                      pe->pe_number,
1890                                                      window_id,
1891                                                      pe->tce_bypass_base,
1892                                                      0);
1893         }
1894         if (rc)
1895                 pe_err(pe, "OPAL error %lld configuring bypass window\n", rc);
1896         else
1897                 pe->tce_bypass_enabled = enable;
1898 }
1899
1900 static void pnv_pci_ioda2_setup_bypass_pe(struct pnv_phb *phb,
1901                                           struct pnv_ioda_pe *pe)
1902 {
1903         /* TVE #1 is selected by PCI address bit 59 */
1904         pe->tce_bypass_base = 1ull << 59;
1905
1906         /* Install set_bypass callback for VFIO */
1907         pe->tce32_table->set_bypass = pnv_pci_ioda2_set_bypass;
1908
1909         /* Enable bypass by default */
1910         pnv_pci_ioda2_set_bypass(pe->tce32_table, true);
1911 }
1912
1913 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
1914                                        struct pnv_ioda_pe *pe)
1915 {
1916         struct page *tce_mem = NULL;
1917         void *addr;
1918         const __be64 *swinvp;
1919         struct iommu_table *tbl;
1920         unsigned int tce_table_size, end;
1921         int64_t rc;
1922
1923         /* We shouldn't already have a 32-bit DMA associated */
1924         if (WARN_ON(pe->tce32_seg >= 0))
1925                 return;
1926
1927         /* The PE will reserve all possible 32-bits space */
1928         pe->tce32_seg = 0;
1929         end = (1 << ilog2(phb->ioda.m32_pci_base));
1930         tce_table_size = (end / 0x1000) * 8;
1931         pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n",
1932                 end);
1933
1934         /* Allocate TCE table */
1935         tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
1936                                    get_order(tce_table_size));
1937         if (!tce_mem) {
1938                 pe_err(pe, "Failed to allocate a 32-bit TCE memory\n");
1939                 goto fail;
1940         }
1941         addr = page_address(tce_mem);
1942         memset(addr, 0, tce_table_size);
1943
1944         /*
1945          * Map TCE table through TVT. The TVE index is the PE number
1946          * shifted by 1 bit for 32-bits DMA space.
1947          */
1948         rc = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
1949                                         pe->pe_number << 1, 1, __pa(addr),
1950                                         tce_table_size, 0x1000);
1951         if (rc) {
1952                 pe_err(pe, "Failed to configure 32-bit TCE table,"
1953                        " err %ld\n", rc);
1954                 goto fail;
1955         }
1956
1957         /* Setup linux iommu table */
1958         tbl = pe->tce32_table;
1959         pnv_pci_setup_iommu_table(tbl, addr, tce_table_size, 0,
1960                         IOMMU_PAGE_SHIFT_4K);
1961
1962         /* OPAL variant of PHB3 invalidated TCEs */
1963         swinvp = of_get_property(phb->hose->dn, "ibm,opal-tce-kill", NULL);
1964         if (swinvp) {
1965                 /* We need a couple more fields -- an address and a data
1966                  * to or.  Since the bus is only printed out on table free
1967                  * errors, and on the first pass the data will be a relative
1968                  * bus number, print that out instead.
1969                  */
1970                 pe->tce_inval_reg_phys = be64_to_cpup(swinvp);
1971                 tbl->it_index = (unsigned long)ioremap(pe->tce_inval_reg_phys,
1972                                 8);
1973                 tbl->it_type |= (TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE);
1974         }
1975         iommu_init_table(tbl, phb->hose->node);
1976
1977         if (pe->flags & PNV_IODA_PE_DEV) {
1978                 iommu_register_group(tbl, phb->hose->global_number,
1979                                      pe->pe_number);
1980                 /*
1981                  * Setting table base here only for carrying iommu_group
1982                  * further down to let iommu_add_device() do the job.
1983                  * pnv_pci_ioda_dma_dev_setup will override it later anyway.
1984                  */
1985                 set_iommu_table_base(&pe->pdev->dev, tbl);
1986                 iommu_add_device(&pe->pdev->dev);
1987         } else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)) {
1988                 iommu_register_group(tbl, phb->hose->global_number,
1989                                      pe->pe_number);
1990                 pnv_ioda_setup_bus_dma(pe, pe->pbus);
1991         } else if (pe->flags & PNV_IODA_PE_VF) {
1992                 iommu_register_group(tbl, phb->hose->global_number,
1993                                      pe->pe_number);
1994         }
1995
1996         /* Also create a bypass window */
1997         if (!pnv_iommu_bypass_disabled)
1998                 pnv_pci_ioda2_setup_bypass_pe(phb, pe);
1999
2000         return;
2001 fail:
2002         if (pe->tce32_seg >= 0)
2003                 pe->tce32_seg = -1;
2004         if (tce_mem)
2005                 __free_pages(tce_mem, get_order(tce_table_size));
2006 }
2007
2008 static void pnv_ioda_setup_dma(struct pnv_phb *phb)
2009 {
2010         struct pci_controller *hose = phb->hose;
2011         unsigned int residual, remaining, segs, tw, base;
2012         struct pnv_ioda_pe *pe;
2013
2014         /* If we have more PE# than segments available, hand out one
2015          * per PE until we run out and let the rest fail. If not,
2016          * then we assign at least one segment per PE, plus more based
2017          * on the amount of devices under that PE
2018          */
2019         if (phb->ioda.dma_pe_count > phb->ioda.tce32_count)
2020                 residual = 0;
2021         else
2022                 residual = phb->ioda.tce32_count -
2023                         phb->ioda.dma_pe_count;
2024
2025         pr_info("PCI: Domain %04x has %ld available 32-bit DMA segments\n",
2026                 hose->global_number, phb->ioda.tce32_count);
2027         pr_info("PCI: %d PE# for a total weight of %d\n",
2028                 phb->ioda.dma_pe_count, phb->ioda.dma_weight);
2029
2030         /* Walk our PE list and configure their DMA segments, hand them
2031          * out one base segment plus any residual segments based on
2032          * weight
2033          */
2034         remaining = phb->ioda.tce32_count;
2035         tw = phb->ioda.dma_weight;
2036         base = 0;
2037         list_for_each_entry(pe, &phb->ioda.pe_dma_list, dma_link) {
2038                 if (!pe->dma_weight)
2039                         continue;
2040                 if (!remaining) {
2041                         pe_warn(pe, "No DMA32 resources available\n");
2042                         continue;
2043                 }
2044                 segs = 1;
2045                 if (residual) {
2046                         segs += ((pe->dma_weight * residual)  + (tw / 2)) / tw;
2047                         if (segs > remaining)
2048                                 segs = remaining;
2049                 }
2050
2051                 /*
2052                  * For IODA2 compliant PHB3, we needn't care about the weight.
2053                  * The all available 32-bits DMA space will be assigned to
2054                  * the specific PE.
2055                  */
2056                 if (phb->type == PNV_PHB_IODA1) {
2057                         pe_info(pe, "DMA weight %d, assigned %d DMA32 segments\n",
2058                                 pe->dma_weight, segs);
2059                         pnv_pci_ioda_setup_dma_pe(phb, pe, base, segs);
2060                 } else {
2061                         pe_info(pe, "Assign DMA32 space\n");
2062                         segs = 0;
2063                         pnv_pci_ioda2_setup_dma_pe(phb, pe);
2064                 }
2065
2066                 remaining -= segs;
2067                 base += segs;
2068         }
2069 }
2070
2071 #ifdef CONFIG_PCI_MSI
2072 static void pnv_ioda2_msi_eoi(struct irq_data *d)
2073 {
2074         unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
2075         struct irq_chip *chip = irq_data_get_irq_chip(d);
2076         struct pnv_phb *phb = container_of(chip, struct pnv_phb,
2077                                            ioda.irq_chip);
2078         int64_t rc;
2079
2080         rc = opal_pci_msi_eoi(phb->opal_id, hw_irq);
2081         WARN_ON_ONCE(rc);
2082
2083         icp_native_eoi(d);
2084 }
2085
2086
2087 static void set_msi_irq_chip(struct pnv_phb *phb, unsigned int virq)
2088 {
2089         struct irq_data *idata;
2090         struct irq_chip *ichip;
2091
2092         if (phb->type != PNV_PHB_IODA2)
2093                 return;
2094
2095         if (!phb->ioda.irq_chip_init) {
2096                 /*
2097                  * First time we setup an MSI IRQ, we need to setup the
2098                  * corresponding IRQ chip to route correctly.
2099                  */
2100                 idata = irq_get_irq_data(virq);
2101                 ichip = irq_data_get_irq_chip(idata);
2102                 phb->ioda.irq_chip_init = 1;
2103                 phb->ioda.irq_chip = *ichip;
2104                 phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi;
2105         }
2106         irq_set_chip(virq, &phb->ioda.irq_chip);
2107 }
2108
2109 #ifdef CONFIG_CXL_BASE
2110
2111 struct device_node *pnv_pci_get_phb_node(struct pci_dev *dev)
2112 {
2113         struct pci_controller *hose = pci_bus_to_host(dev->bus);
2114
2115         return of_node_get(hose->dn);
2116 }
2117 EXPORT_SYMBOL(pnv_pci_get_phb_node);
2118
2119 int pnv_phb_to_cxl_mode(struct pci_dev *dev, uint64_t mode)
2120 {
2121         struct pci_controller *hose = pci_bus_to_host(dev->bus);
2122         struct pnv_phb *phb = hose->private_data;
2123         struct pnv_ioda_pe *pe;
2124         int rc;
2125
2126         pe = pnv_ioda_get_pe(dev);
2127         if (!pe)
2128                 return -ENODEV;
2129
2130         pe_info(pe, "Switching PHB to CXL\n");
2131
2132         rc = opal_pci_set_phb_cxl_mode(phb->opal_id, mode, pe->pe_number);
2133         if (rc)
2134                 dev_err(&dev->dev, "opal_pci_set_phb_cxl_mode failed: %i\n", rc);
2135
2136         return rc;
2137 }
2138 EXPORT_SYMBOL(pnv_phb_to_cxl_mode);
2139
2140 /* Find PHB for cxl dev and allocate MSI hwirqs?
2141  * Returns the absolute hardware IRQ number
2142  */
2143 int pnv_cxl_alloc_hwirqs(struct pci_dev *dev, int num)
2144 {
2145         struct pci_controller *hose = pci_bus_to_host(dev->bus);
2146         struct pnv_phb *phb = hose->private_data;
2147         int hwirq = msi_bitmap_alloc_hwirqs(&phb->msi_bmp, num);
2148
2149         if (hwirq < 0) {
2150                 dev_warn(&dev->dev, "Failed to find a free MSI\n");
2151                 return -ENOSPC;
2152         }
2153
2154         return phb->msi_base + hwirq;
2155 }
2156 EXPORT_SYMBOL(pnv_cxl_alloc_hwirqs);
2157
2158 void pnv_cxl_release_hwirqs(struct pci_dev *dev, int hwirq, int num)
2159 {
2160         struct pci_controller *hose = pci_bus_to_host(dev->bus);
2161         struct pnv_phb *phb = hose->private_data;
2162
2163         msi_bitmap_free_hwirqs(&phb->msi_bmp, hwirq - phb->msi_base, num);
2164 }
2165 EXPORT_SYMBOL(pnv_cxl_release_hwirqs);
2166
2167 void pnv_cxl_release_hwirq_ranges(struct cxl_irq_ranges *irqs,
2168                                   struct pci_dev *dev)
2169 {
2170         struct pci_controller *hose = pci_bus_to_host(dev->bus);
2171         struct pnv_phb *phb = hose->private_data;
2172         int i, hwirq;
2173
2174         for (i = 1; i < CXL_IRQ_RANGES; i++) {
2175                 if (!irqs->range[i])
2176                         continue;
2177                 pr_devel("cxl release irq range 0x%x: offset: 0x%lx  limit: %ld\n",
2178                          i, irqs->offset[i],
2179                          irqs->range[i]);
2180                 hwirq = irqs->offset[i] - phb->msi_base;
2181                 msi_bitmap_free_hwirqs(&phb->msi_bmp, hwirq,
2182                                        irqs->range[i]);
2183         }
2184 }
2185 EXPORT_SYMBOL(pnv_cxl_release_hwirq_ranges);
2186
2187 int pnv_cxl_alloc_hwirq_ranges(struct cxl_irq_ranges *irqs,
2188                                struct pci_dev *dev, int num)
2189 {
2190         struct pci_controller *hose = pci_bus_to_host(dev->bus);
2191         struct pnv_phb *phb = hose->private_data;
2192         int i, hwirq, try;
2193
2194         memset(irqs, 0, sizeof(struct cxl_irq_ranges));
2195
2196         /* 0 is reserved for the multiplexed PSL DSI interrupt */
2197         for (i = 1; i < CXL_IRQ_RANGES && num; i++) {
2198                 try = num;
2199                 while (try) {
2200                         hwirq = msi_bitmap_alloc_hwirqs(&phb->msi_bmp, try);
2201                         if (hwirq >= 0)
2202                                 break;
2203                         try /= 2;
2204                 }
2205                 if (!try)
2206                         goto fail;
2207
2208                 irqs->offset[i] = phb->msi_base + hwirq;
2209                 irqs->range[i] = try;
2210                 pr_devel("cxl alloc irq range 0x%x: offset: 0x%lx  limit: %li\n",
2211                          i, irqs->offset[i], irqs->range[i]);
2212                 num -= try;
2213         }
2214         if (num)
2215                 goto fail;
2216
2217         return 0;
2218 fail:
2219         pnv_cxl_release_hwirq_ranges(irqs, dev);
2220         return -ENOSPC;
2221 }
2222 EXPORT_SYMBOL(pnv_cxl_alloc_hwirq_ranges);
2223
2224 int pnv_cxl_get_irq_count(struct pci_dev *dev)
2225 {
2226         struct pci_controller *hose = pci_bus_to_host(dev->bus);
2227         struct pnv_phb *phb = hose->private_data;
2228
2229         return phb->msi_bmp.irq_count;
2230 }
2231 EXPORT_SYMBOL(pnv_cxl_get_irq_count);
2232
2233 int pnv_cxl_ioda_msi_setup(struct pci_dev *dev, unsigned int hwirq,
2234                            unsigned int virq)
2235 {
2236         struct pci_controller *hose = pci_bus_to_host(dev->bus);
2237         struct pnv_phb *phb = hose->private_data;
2238         unsigned int xive_num = hwirq - phb->msi_base;
2239         struct pnv_ioda_pe *pe;
2240         int rc;
2241
2242         if (!(pe = pnv_ioda_get_pe(dev)))
2243                 return -ENODEV;
2244
2245         /* Assign XIVE to PE */
2246         rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
2247         if (rc) {
2248                 pe_warn(pe, "%s: OPAL error %d setting msi_base 0x%x "
2249                         "hwirq 0x%x XIVE 0x%x PE\n",
2250                         pci_name(dev), rc, phb->msi_base, hwirq, xive_num);
2251                 return -EIO;
2252         }
2253         set_msi_irq_chip(phb, virq);
2254
2255         return 0;
2256 }
2257 EXPORT_SYMBOL(pnv_cxl_ioda_msi_setup);
2258 #endif
2259
2260 static int pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev,
2261                                   unsigned int hwirq, unsigned int virq,
2262                                   unsigned int is_64, struct msi_msg *msg)
2263 {
2264         struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev);
2265         unsigned int xive_num = hwirq - phb->msi_base;
2266         __be32 data;
2267         int rc;
2268
2269         /* No PE assigned ? bail out ... no MSI for you ! */
2270         if (pe == NULL)
2271                 return -ENXIO;
2272
2273         /* Check if we have an MVE */
2274         if (pe->mve_number < 0)
2275                 return -ENXIO;
2276
2277         /* Force 32-bit MSI on some broken devices */
2278         if (dev->no_64bit_msi)
2279                 is_64 = 0;
2280
2281         /* Assign XIVE to PE */
2282         rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
2283         if (rc) {
2284                 pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
2285                         pci_name(dev), rc, xive_num);
2286                 return -EIO;
2287         }
2288
2289         if (is_64) {
2290                 __be64 addr64;
2291
2292                 rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1,
2293                                      &addr64, &data);
2294                 if (rc) {
2295                         pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
2296                                 pci_name(dev), rc);
2297                         return -EIO;
2298                 }
2299                 msg->address_hi = be64_to_cpu(addr64) >> 32;
2300                 msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful;
2301         } else {
2302                 __be32 addr32;
2303
2304                 rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1,
2305                                      &addr32, &data);
2306                 if (rc) {
2307                         pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
2308                                 pci_name(dev), rc);
2309                         return -EIO;
2310                 }
2311                 msg->address_hi = 0;
2312                 msg->address_lo = be32_to_cpu(addr32);
2313         }
2314         msg->data = be32_to_cpu(data);
2315
2316         set_msi_irq_chip(phb, virq);
2317
2318         pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d),"
2319                  " address=%x_%08x data=%x PE# %d\n",
2320                  pci_name(dev), is_64 ? "64" : "32", hwirq, xive_num,
2321                  msg->address_hi, msg->address_lo, data, pe->pe_number);
2322
2323         return 0;
2324 }
2325
2326 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb)
2327 {
2328         unsigned int count;
2329         const __be32 *prop = of_get_property(phb->hose->dn,
2330                                              "ibm,opal-msi-ranges", NULL);
2331         if (!prop) {
2332                 /* BML Fallback */
2333                 prop = of_get_property(phb->hose->dn, "msi-ranges", NULL);
2334         }
2335         if (!prop)
2336                 return;
2337
2338         phb->msi_base = be32_to_cpup(prop);
2339         count = be32_to_cpup(prop + 1);
2340         if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) {
2341                 pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
2342                        phb->hose->global_number);
2343                 return;
2344         }
2345
2346         phb->msi_setup = pnv_pci_ioda_msi_setup;
2347         phb->msi32_support = 1;
2348         pr_info("  Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
2349                 count, phb->msi_base);
2350 }
2351 #else
2352 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb) { }
2353 #endif /* CONFIG_PCI_MSI */
2354
2355 #ifdef CONFIG_PCI_IOV
2356 static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev *pdev)
2357 {
2358         struct pci_controller *hose;
2359         struct pnv_phb *phb;
2360         struct resource *res;
2361         int i;
2362         resource_size_t size;
2363         struct pci_dn *pdn;
2364         int mul, total_vfs;
2365
2366         if (!pdev->is_physfn || pdev->is_added)
2367                 return;
2368
2369         hose = pci_bus_to_host(pdev->bus);
2370         phb = hose->private_data;
2371
2372         pdn = pci_get_pdn(pdev);
2373         pdn->vfs_expanded = 0;
2374
2375         total_vfs = pci_sriov_get_totalvfs(pdev);
2376         pdn->m64_per_iov = 1;
2377         mul = phb->ioda.total_pe;
2378
2379         for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2380                 res = &pdev->resource[i + PCI_IOV_RESOURCES];
2381                 if (!res->flags || res->parent)
2382                         continue;
2383                 if (!pnv_pci_is_mem_pref_64(res->flags)) {
2384                         dev_warn(&pdev->dev, " non M64 VF BAR%d: %pR\n",
2385                                  i, res);
2386                         continue;
2387                 }
2388
2389                 size = pci_iov_resource_size(pdev, i + PCI_IOV_RESOURCES);
2390
2391                 /* bigger than 64M */
2392                 if (size > (1 << 26)) {
2393                         dev_info(&pdev->dev, "PowerNV: VF BAR%d: %pR IOV size is bigger than 64M, roundup power2\n",
2394                                  i, res);
2395                         pdn->m64_per_iov = M64_PER_IOV;
2396                         mul = roundup_pow_of_two(total_vfs);
2397                         break;
2398                 }
2399         }
2400
2401         for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2402                 res = &pdev->resource[i + PCI_IOV_RESOURCES];
2403                 if (!res->flags || res->parent)
2404                         continue;
2405                 if (!pnv_pci_is_mem_pref_64(res->flags)) {
2406                         dev_warn(&pdev->dev, "Skipping expanding VF BAR%d: %pR\n",
2407                                  i, res);
2408                         continue;
2409                 }
2410
2411                 dev_dbg(&pdev->dev, " Fixing VF BAR%d: %pR to\n", i, res);
2412                 size = pci_iov_resource_size(pdev, i + PCI_IOV_RESOURCES);
2413                 res->end = res->start + size * mul - 1;
2414                 dev_dbg(&pdev->dev, "                       %pR\n", res);
2415                 dev_info(&pdev->dev, "VF BAR%d: %pR (expanded to %d VFs for PE alignment)",
2416                          i, res, mul);
2417         }
2418         pdn->vfs_expanded = mul;
2419 }
2420 #endif /* CONFIG_PCI_IOV */
2421
2422 /*
2423  * This function is supposed to be called on basis of PE from top
2424  * to bottom style. So the the I/O or MMIO segment assigned to
2425  * parent PE could be overrided by its child PEs if necessary.
2426  */
2427 static void pnv_ioda_setup_pe_seg(struct pci_controller *hose,
2428                                   struct pnv_ioda_pe *pe)
2429 {
2430         struct pnv_phb *phb = hose->private_data;
2431         struct pci_bus_region region;
2432         struct resource *res;
2433         int i, index;
2434         int rc;
2435
2436         /*
2437          * NOTE: We only care PCI bus based PE for now. For PCI
2438          * device based PE, for example SRIOV sensitive VF should
2439          * be figured out later.
2440          */
2441         BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)));
2442
2443         pci_bus_for_each_resource(pe->pbus, res, i) {
2444                 if (!res || !res->flags ||
2445                     res->start > res->end)
2446                         continue;
2447
2448                 if (res->flags & IORESOURCE_IO) {
2449                         region.start = res->start - phb->ioda.io_pci_base;
2450                         region.end   = res->end - phb->ioda.io_pci_base;
2451                         index = region.start / phb->ioda.io_segsize;
2452
2453                         while (index < phb->ioda.total_pe &&
2454                                region.start <= region.end) {
2455                                 phb->ioda.io_segmap[index] = pe->pe_number;
2456                                 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2457                                         pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index);
2458                                 if (rc != OPAL_SUCCESS) {
2459                                         pr_err("%s: OPAL error %d when mapping IO "
2460                                                "segment #%d to PE#%d\n",
2461                                                __func__, rc, index, pe->pe_number);
2462                                         break;
2463                                 }
2464
2465                                 region.start += phb->ioda.io_segsize;
2466                                 index++;
2467                         }
2468                 } else if ((res->flags & IORESOURCE_MEM) &&
2469                            !pnv_pci_is_mem_pref_64(res->flags)) {
2470                         region.start = res->start -
2471                                        hose->mem_offset[0] -
2472                                        phb->ioda.m32_pci_base;
2473                         region.end   = res->end -
2474                                        hose->mem_offset[0] -
2475                                        phb->ioda.m32_pci_base;
2476                         index = region.start / phb->ioda.m32_segsize;
2477
2478                         while (index < phb->ioda.total_pe &&
2479                                region.start <= region.end) {
2480                                 phb->ioda.m32_segmap[index] = pe->pe_number;
2481                                 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2482                                         pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index);
2483                                 if (rc != OPAL_SUCCESS) {
2484                                         pr_err("%s: OPAL error %d when mapping M32 "
2485                                                "segment#%d to PE#%d",
2486                                                __func__, rc, index, pe->pe_number);
2487                                         break;
2488                                 }
2489
2490                                 region.start += phb->ioda.m32_segsize;
2491                                 index++;
2492                         }
2493                 }
2494         }
2495 }
2496
2497 static void pnv_pci_ioda_setup_seg(void)
2498 {
2499         struct pci_controller *tmp, *hose;
2500         struct pnv_phb *phb;
2501         struct pnv_ioda_pe *pe;
2502
2503         list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
2504                 phb = hose->private_data;
2505                 list_for_each_entry(pe, &phb->ioda.pe_list, list) {
2506                         pnv_ioda_setup_pe_seg(hose, pe);
2507                 }
2508         }
2509 }
2510
2511 static void pnv_pci_ioda_setup_DMA(void)
2512 {
2513         struct pci_controller *hose, *tmp;
2514         struct pnv_phb *phb;
2515
2516         list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
2517                 pnv_ioda_setup_dma(hose->private_data);
2518
2519                 /* Mark the PHB initialization done */
2520                 phb = hose->private_data;
2521                 phb->initialized = 1;
2522         }
2523 }
2524
2525 static void pnv_pci_ioda_create_dbgfs(void)
2526 {
2527 #ifdef CONFIG_DEBUG_FS
2528         struct pci_controller *hose, *tmp;
2529         struct pnv_phb *phb;
2530         char name[16];
2531
2532         list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
2533                 phb = hose->private_data;
2534
2535                 sprintf(name, "PCI%04x", hose->global_number);
2536                 phb->dbgfs = debugfs_create_dir(name, powerpc_debugfs_root);
2537                 if (!phb->dbgfs)
2538                         pr_warning("%s: Error on creating debugfs on PHB#%x\n",
2539                                 __func__, hose->global_number);
2540         }
2541 #endif /* CONFIG_DEBUG_FS */
2542 }
2543
2544 static void pnv_pci_ioda_fixup(void)
2545 {
2546         pnv_pci_ioda_setup_PEs();
2547         pnv_pci_ioda_setup_seg();
2548         pnv_pci_ioda_setup_DMA();
2549
2550         pnv_pci_ioda_create_dbgfs();
2551
2552 #ifdef CONFIG_EEH
2553         eeh_init();
2554         eeh_addr_cache_build();
2555 #endif
2556 }
2557
2558 /*
2559  * Returns the alignment for I/O or memory windows for P2P
2560  * bridges. That actually depends on how PEs are segmented.
2561  * For now, we return I/O or M32 segment size for PE sensitive
2562  * P2P bridges. Otherwise, the default values (4KiB for I/O,
2563  * 1MiB for memory) will be returned.
2564  *
2565  * The current PCI bus might be put into one PE, which was
2566  * create against the parent PCI bridge. For that case, we
2567  * needn't enlarge the alignment so that we can save some
2568  * resources.
2569  */
2570 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus,
2571                                                 unsigned long type)
2572 {
2573         struct pci_dev *bridge;
2574         struct pci_controller *hose = pci_bus_to_host(bus);
2575         struct pnv_phb *phb = hose->private_data;
2576         int num_pci_bridges = 0;
2577
2578         bridge = bus->self;
2579         while (bridge) {
2580                 if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) {
2581                         num_pci_bridges++;
2582                         if (num_pci_bridges >= 2)
2583                                 return 1;
2584                 }
2585
2586                 bridge = bridge->bus->self;
2587         }
2588
2589         /* We fail back to M32 if M64 isn't supported */
2590         if (phb->ioda.m64_segsize &&
2591             pnv_pci_is_mem_pref_64(type))
2592                 return phb->ioda.m64_segsize;
2593         if (type & IORESOURCE_MEM)
2594                 return phb->ioda.m32_segsize;
2595
2596         return phb->ioda.io_segsize;
2597 }
2598
2599 #ifdef CONFIG_PCI_IOV
2600 static resource_size_t pnv_pci_iov_resource_alignment(struct pci_dev *pdev,
2601                                                       int resno)
2602 {
2603         struct pci_dn *pdn = pci_get_pdn(pdev);
2604         resource_size_t align, iov_align;
2605
2606         iov_align = resource_size(&pdev->resource[resno]);
2607         if (iov_align)
2608                 return iov_align;
2609
2610         align = pci_iov_resource_size(pdev, resno);
2611         if (pdn->vfs_expanded)
2612                 return pdn->vfs_expanded * align;
2613
2614         return align;
2615 }
2616 #endif /* CONFIG_PCI_IOV */
2617
2618 /* Prevent enabling devices for which we couldn't properly
2619  * assign a PE
2620  */
2621 static bool pnv_pci_enable_device_hook(struct pci_dev *dev)
2622 {
2623         struct pci_controller *hose = pci_bus_to_host(dev->bus);
2624         struct pnv_phb *phb = hose->private_data;
2625         struct pci_dn *pdn;
2626
2627         /* The function is probably called while the PEs have
2628          * not be created yet. For example, resource reassignment
2629          * during PCI probe period. We just skip the check if
2630          * PEs isn't ready.
2631          */
2632         if (!phb->initialized)
2633                 return true;
2634
2635         pdn = pci_get_pdn(dev);
2636         if (!pdn || pdn->pe_number == IODA_INVALID_PE)
2637                 return false;
2638
2639         return true;
2640 }
2641
2642 static u32 pnv_ioda_bdfn_to_pe(struct pnv_phb *phb, struct pci_bus *bus,
2643                                u32 devfn)
2644 {
2645         return phb->ioda.pe_rmap[(bus->number << 8) | devfn];
2646 }
2647
2648 static void pnv_pci_ioda_shutdown(struct pci_controller *hose)
2649 {
2650         struct pnv_phb *phb = hose->private_data;
2651
2652         opal_pci_reset(phb->opal_id, OPAL_RESET_PCI_IODA_TABLE,
2653                        OPAL_ASSERT_RESET);
2654 }
2655
2656 static const struct pci_controller_ops pnv_pci_ioda_controller_ops = {
2657        .dma_dev_setup = pnv_pci_dma_dev_setup,
2658 #ifdef CONFIG_PCI_MSI
2659        .setup_msi_irqs = pnv_setup_msi_irqs,
2660        .teardown_msi_irqs = pnv_teardown_msi_irqs,
2661 #endif
2662        .enable_device_hook = pnv_pci_enable_device_hook,
2663        .window_alignment = pnv_pci_window_alignment,
2664        .reset_secondary_bus = pnv_pci_reset_secondary_bus,
2665        .dma_set_mask = pnv_pci_ioda_dma_set_mask,
2666        .shutdown = pnv_pci_ioda_shutdown,
2667 };
2668
2669 static void __init pnv_pci_init_ioda_phb(struct device_node *np,
2670                                          u64 hub_id, int ioda_type)
2671 {
2672         struct pci_controller *hose;
2673         struct pnv_phb *phb;
2674         unsigned long size, m32map_off, pemap_off, iomap_off = 0;
2675         const __be64 *prop64;
2676         const __be32 *prop32;
2677         int len;
2678         u64 phb_id;
2679         void *aux;
2680         long rc;
2681
2682         pr_info("Initializing IODA%d OPAL PHB %s\n", ioda_type, np->full_name);
2683
2684         prop64 = of_get_property(np, "ibm,opal-phbid", NULL);
2685         if (!prop64) {
2686                 pr_err("  Missing \"ibm,opal-phbid\" property !\n");
2687                 return;
2688         }
2689         phb_id = be64_to_cpup(prop64);
2690         pr_debug("  PHB-ID  : 0x%016llx\n", phb_id);
2691
2692         phb = memblock_virt_alloc(sizeof(struct pnv_phb), 0);
2693
2694         /* Allocate PCI controller */
2695         phb->hose = hose = pcibios_alloc_controller(np);
2696         if (!phb->hose) {
2697                 pr_err("  Can't allocate PCI controller for %s\n",
2698                        np->full_name);
2699                 memblock_free(__pa(phb), sizeof(struct pnv_phb));
2700                 return;
2701         }
2702
2703         spin_lock_init(&phb->lock);
2704         prop32 = of_get_property(np, "bus-range", &len);
2705         if (prop32 && len == 8) {
2706                 hose->first_busno = be32_to_cpu(prop32[0]);
2707                 hose->last_busno = be32_to_cpu(prop32[1]);
2708         } else {
2709                 pr_warn("  Broken <bus-range> on %s\n", np->full_name);
2710                 hose->first_busno = 0;
2711                 hose->last_busno = 0xff;
2712         }
2713         hose->private_data = phb;
2714         phb->hub_id = hub_id;
2715         phb->opal_id = phb_id;
2716         phb->type = ioda_type;
2717         mutex_init(&phb->ioda.pe_alloc_mutex);
2718
2719         /* Detect specific models for error handling */
2720         if (of_device_is_compatible(np, "ibm,p7ioc-pciex"))
2721                 phb->model = PNV_PHB_MODEL_P7IOC;
2722         else if (of_device_is_compatible(np, "ibm,power8-pciex"))
2723                 phb->model = PNV_PHB_MODEL_PHB3;
2724         else
2725                 phb->model = PNV_PHB_MODEL_UNKNOWN;
2726
2727         /* Parse 32-bit and IO ranges (if any) */
2728         pci_process_bridge_OF_ranges(hose, np, !hose->global_number);
2729
2730         /* Get registers */
2731         phb->regs = of_iomap(np, 0);
2732         if (phb->regs == NULL)
2733                 pr_err("  Failed to map registers !\n");
2734
2735         /* Initialize more IODA stuff */
2736         phb->ioda.total_pe = 1;
2737         prop32 = of_get_property(np, "ibm,opal-num-pes", NULL);
2738         if (prop32)
2739                 phb->ioda.total_pe = be32_to_cpup(prop32);
2740         prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL);
2741         if (prop32)
2742                 phb->ioda.reserved_pe = be32_to_cpup(prop32);
2743
2744         /* Parse 64-bit MMIO range */
2745         pnv_ioda_parse_m64_window(phb);
2746
2747         phb->ioda.m32_size = resource_size(&hose->mem_resources[0]);
2748         /* FW Has already off top 64k of M32 space (MSI space) */
2749         phb->ioda.m32_size += 0x10000;
2750
2751         phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe;
2752         phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0];
2753         phb->ioda.io_size = hose->pci_io_size;
2754         phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe;
2755         phb->ioda.io_pci_base = 0; /* XXX calculate this ? */
2756
2757         /* Allocate aux data & arrays. We don't have IO ports on PHB3 */
2758         size = _ALIGN_UP(phb->ioda.total_pe / 8, sizeof(unsigned long));
2759         m32map_off = size;
2760         size += phb->ioda.total_pe * sizeof(phb->ioda.m32_segmap[0]);
2761         if (phb->type == PNV_PHB_IODA1) {
2762                 iomap_off = size;
2763                 size += phb->ioda.total_pe * sizeof(phb->ioda.io_segmap[0]);
2764         }
2765         pemap_off = size;
2766         size += phb->ioda.total_pe * sizeof(struct pnv_ioda_pe);
2767         aux = memblock_virt_alloc(size, 0);
2768         phb->ioda.pe_alloc = aux;
2769         phb->ioda.m32_segmap = aux + m32map_off;
2770         if (phb->type == PNV_PHB_IODA1)
2771                 phb->ioda.io_segmap = aux + iomap_off;
2772         phb->ioda.pe_array = aux + pemap_off;
2773         set_bit(phb->ioda.reserved_pe, phb->ioda.pe_alloc);
2774
2775         INIT_LIST_HEAD(&phb->ioda.pe_dma_list);
2776         INIT_LIST_HEAD(&phb->ioda.pe_list);
2777         mutex_init(&phb->ioda.pe_list_mutex);
2778
2779         /* Calculate how many 32-bit TCE segments we have */
2780         phb->ioda.tce32_count = phb->ioda.m32_pci_base >> 28;
2781
2782 #if 0 /* We should really do that ... */
2783         rc = opal_pci_set_phb_mem_window(opal->phb_id,
2784                                          window_type,
2785                                          window_num,
2786                                          starting_real_address,
2787                                          starting_pci_address,
2788                                          segment_size);
2789 #endif
2790
2791         pr_info("  %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
2792                 phb->ioda.total_pe, phb->ioda.reserved_pe,
2793                 phb->ioda.m32_size, phb->ioda.m32_segsize);
2794         if (phb->ioda.m64_size)
2795                 pr_info("                 M64: 0x%lx [segment=0x%lx]\n",
2796                         phb->ioda.m64_size, phb->ioda.m64_segsize);
2797         if (phb->ioda.io_size)
2798                 pr_info("                  IO: 0x%x [segment=0x%x]\n",
2799                         phb->ioda.io_size, phb->ioda.io_segsize);
2800
2801
2802         phb->hose->ops = &pnv_pci_ops;
2803         phb->get_pe_state = pnv_ioda_get_pe_state;
2804         phb->freeze_pe = pnv_ioda_freeze_pe;
2805         phb->unfreeze_pe = pnv_ioda_unfreeze_pe;
2806
2807         /* Setup RID -> PE mapping function */
2808         phb->bdfn_to_pe = pnv_ioda_bdfn_to_pe;
2809
2810         /* Setup TCEs */
2811         phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup;
2812         phb->dma_get_required_mask = pnv_pci_ioda_dma_get_required_mask;
2813
2814         /* Setup MSI support */
2815         pnv_pci_init_ioda_msis(phb);
2816
2817         /*
2818          * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
2819          * to let the PCI core do resource assignment. It's supposed
2820          * that the PCI core will do correct I/O and MMIO alignment
2821          * for the P2P bridge bars so that each PCI bus (excluding
2822          * the child P2P bridges) can form individual PE.
2823          */
2824         ppc_md.pcibios_fixup = pnv_pci_ioda_fixup;
2825         hose->controller_ops = pnv_pci_ioda_controller_ops;
2826
2827 #ifdef CONFIG_PCI_IOV
2828         ppc_md.pcibios_fixup_sriov = pnv_pci_ioda_fixup_iov_resources;
2829         ppc_md.pcibios_iov_resource_alignment = pnv_pci_iov_resource_alignment;
2830 #endif
2831
2832         pci_add_flags(PCI_REASSIGN_ALL_RSRC);
2833
2834         /* Reset IODA tables to a clean state */
2835         rc = opal_pci_reset(phb_id, OPAL_RESET_PCI_IODA_TABLE, OPAL_ASSERT_RESET);
2836         if (rc)
2837                 pr_warning("  OPAL Error %ld performing IODA table reset !\n", rc);
2838
2839         /* If we're running in kdump kerenl, the previous kerenl never
2840          * shutdown PCI devices correctly. We already got IODA table
2841          * cleaned out. So we have to issue PHB reset to stop all PCI
2842          * transactions from previous kerenl.
2843          */
2844         if (is_kdump_kernel()) {
2845                 pr_info("  Issue PHB reset ...\n");
2846                 pnv_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL);
2847                 pnv_eeh_phb_reset(hose, EEH_RESET_DEACTIVATE);
2848         }
2849
2850         /* Remove M64 resource if we can't configure it successfully */
2851         if (!phb->init_m64 || phb->init_m64(phb))
2852                 hose->mem_resources[1].flags = 0;
2853 }
2854
2855 void __init pnv_pci_init_ioda2_phb(struct device_node *np)
2856 {
2857         pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2);
2858 }
2859
2860 void __init pnv_pci_init_ioda_hub(struct device_node *np)
2861 {
2862         struct device_node *phbn;
2863         const __be64 *prop64;
2864         u64 hub_id;
2865
2866         pr_info("Probing IODA IO-Hub %s\n", np->full_name);
2867
2868         prop64 = of_get_property(np, "ibm,opal-hubid", NULL);
2869         if (!prop64) {
2870                 pr_err(" Missing \"ibm,opal-hubid\" property !\n");
2871                 return;
2872         }
2873         hub_id = be64_to_cpup(prop64);
2874         pr_devel(" HUB-ID : 0x%016llx\n", hub_id);
2875
2876         /* Count child PHBs */
2877         for_each_child_of_node(np, phbn) {
2878                 /* Look for IODA1 PHBs */
2879                 if (of_device_is_compatible(phbn, "ibm,ioda-phb"))
2880                         pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1);
2881         }
2882 }