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.h>
  43
  44 #include "powernv.h"
  45 #include "pci.h"
  46
  47 static void pe_level_printk(const struct pnv_ioda_pe *pe, const char *level,
  48                             const char *fmt, ...)
  49 {
  50         struct va_format vaf;
  51         va_list args;
  52         char pfix[32];
  53
  54         va_start(args, fmt);
  55
  56         vaf.fmt = fmt;
  57         vaf.va = &args;
  58
  59         if (pe->pdev)
  60                 strlcpy(pfix, dev_name(&pe->pdev->dev), sizeof(pfix));
  61         else
  62                 sprintf(pfix, "%04x:%02x     ",
  63                         pci_domain_nr(pe->pbus), pe->pbus->number);
  64
  65         printk("%spci %s: [PE# %.3d] %pV",
  66                level, pfix, pe->pe_number, &vaf);
  67
  68         va_end(args);
  69 }
  70
  71 #define pe_err(pe, fmt, ...)                                    \
  72         pe_level_printk(pe, KERN_ERR, fmt, ##__VA_ARGS__)
  73 #define pe_warn(pe, fmt, ...)                                   \
  74         pe_level_printk(pe, KERN_WARNING, fmt, ##__VA_ARGS__)
  75 #define pe_info(pe, fmt, ...)                                   \
  76         pe_level_printk(pe, KERN_INFO, fmt, ##__VA_ARGS__)
  77
  78 /*
  79  * stdcix is only supposed to be used in hypervisor real mode as per
  80  * the architecture spec
  81  */
  82 static inline void __raw_rm_writeq(u64 val, volatile void __iomem *paddr)
  83 {
  84         __asm__ __volatile__("stdcix %0,0,%1"
  85                 : : "r" (val), "r" (paddr) : "memory");
  86 }
  87
  88 static inline bool pnv_pci_is_mem_pref_64(unsigned long flags)
  89 {
  90         return ((flags & (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH)) ==
  91                 (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH));
  92 }
  93
  94 static void pnv_ioda_reserve_pe(struct pnv_phb *phb, int pe_no)
  95 {
  96         if (!(pe_no >= 0 && pe_no < phb->ioda.total_pe)) {
  97                 pr_warn("%s: Invalid PE %d on PHB#%x\n",
  98                         __func__, pe_no, phb->hose->global_number);
  99                 return;
 100         }
 101
 102         if (test_and_set_bit(pe_no, phb->ioda.pe_alloc)) {
 103                 pr_warn("%s: PE %d was assigned on PHB#%x\n",
 104                         __func__, pe_no, phb->hose->global_number);
 105                 return;
 106         }
 107
 108         phb->ioda.pe_array[pe_no].phb = phb;
 109         phb->ioda.pe_array[pe_no].pe_number = pe_no;
 110 }
 111
 112 static int pnv_ioda_alloc_pe(struct pnv_phb *phb)
 113 {
 114         unsigned long pe;
 115
 116         do {
 117                 pe = find_next_zero_bit(phb->ioda.pe_alloc,
 118                                         phb->ioda.total_pe, 0);
 119                 if (pe >= phb->ioda.total_pe)
 120                         return IODA_INVALID_PE;
 121         } while(test_and_set_bit(pe, phb->ioda.pe_alloc));
 122
 123         phb->ioda.pe_array[pe].phb = phb;
 124         phb->ioda.pe_array[pe].pe_number = pe;
 125         return pe;
 126 }
 127
 128 static void pnv_ioda_free_pe(struct pnv_phb *phb, int pe)
 129 {
 130         WARN_ON(phb->ioda.pe_array[pe].pdev);
 131
 132         memset(&phb->ioda.pe_array[pe], 0, sizeof(struct pnv_ioda_pe));
 133         clear_bit(pe, phb->ioda.pe_alloc);
 134 }
 135
 136 /* The default M64 BAR is shared by all PEs */
 137 static int pnv_ioda2_init_m64(struct pnv_phb *phb)
 138 {
 139         const char *desc;
 140         struct resource *r;
 141         s64 rc;
 142
 143         /* Configure the default M64 BAR */
 144         rc = opal_pci_set_phb_mem_window(phb->opal_id,
 145                                          OPAL_M64_WINDOW_TYPE,
 146                                          phb->ioda.m64_bar_idx,
 147                                          phb->ioda.m64_base,
 148                                          0, /* unused */
 149                                          phb->ioda.m64_size);
 150         if (rc != OPAL_SUCCESS) {
 151                 desc = "configuring";
 152                 goto fail;
 153         }
 154
 155         /* Enable the default M64 BAR */
 156         rc = opal_pci_phb_mmio_enable(phb->opal_id,
 157                                       OPAL_M64_WINDOW_TYPE,
 158                                       phb->ioda.m64_bar_idx,
 159                                       OPAL_ENABLE_M64_SPLIT);
 160         if (rc != OPAL_SUCCESS) {
 161                 desc = "enabling";
 162                 goto fail;
 163         }
 164
 165         /* Mark the M64 BAR assigned */
 166         set_bit(phb->ioda.m64_bar_idx, &phb->ioda.m64_bar_alloc);
 167
 168         /*
 169          * Strip off the segment used by the reserved PE, which is
 170          * expected to be 0 or last one of PE capabicity.
 171          */
 172         r = &phb->hose->mem_resources[1];
 173         if (phb->ioda.reserved_pe == 0)
 174                 r->start += phb->ioda.m64_segsize;
 175         else if (phb->ioda.reserved_pe == (phb->ioda.total_pe - 1))
 176                 r->end -= phb->ioda.m64_segsize;
 177         else
 178                 pr_warn("  Cannot strip M64 segment for reserved PE#%d\n",
 179                         phb->ioda.reserved_pe);
 180
 181         return 0;
 182
 183 fail:
 184         pr_warn("  Failure %lld %s M64 BAR#%d\n",
 185                 rc, desc, phb->ioda.m64_bar_idx);
 186         opal_pci_phb_mmio_enable(phb->opal_id,
 187                                  OPAL_M64_WINDOW_TYPE,
 188                                  phb->ioda.m64_bar_idx,
 189                                  OPAL_DISABLE_M64);
 190         return -EIO;
 191 }
 192
 193 static void pnv_ioda2_reserve_m64_pe(struct pnv_phb *phb)
 194 {
 195         resource_size_t sgsz = phb->ioda.m64_segsize;
 196         struct pci_dev *pdev;
 197         struct resource *r;
 198         int base, step, i;
 199
 200         /*
 201          * Root bus always has full M64 range and root port has
 202          * M64 range used in reality. So we're checking root port
 203          * instead of root bus.
 204          */
 205         list_for_each_entry(pdev, &phb->hose->bus->devices, bus_list) {
 206                 for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++) {
 207                         r = &pdev->resource[PCI_BRIDGE_RESOURCES + i];
 208                         if (!r->parent ||
 209                             !pnv_pci_is_mem_pref_64(r->flags))
 210                                 continue;
 211
 212                         base = (r->start - phb->ioda.m64_base) / sgsz;
 213                         for (step = 0; step < resource_size(r) / sgsz; step++)
 214                                 pnv_ioda_reserve_pe(phb, base + step);
 215                 }
 216         }
 217 }
 218
 219 static int pnv_ioda2_pick_m64_pe(struct pnv_phb *phb,
 220                                  struct pci_bus *bus, int all)
 221 {
 222         resource_size_t segsz = phb->ioda.m64_segsize;
 223         struct pci_dev *pdev;
 224         struct resource *r;
 225         struct pnv_ioda_pe *master_pe, *pe;
 226         unsigned long size, *pe_alloc;
 227         bool found;
 228         int start, i, j;
 229
 230         /* Root bus shouldn't use M64 */
 231         if (pci_is_root_bus(bus))
 232                 return IODA_INVALID_PE;
 233
 234         /* We support only one M64 window on each bus */
 235         found = false;
 236         pci_bus_for_each_resource(bus, r, i) {
 237                 if (r && r->parent &&
 238                     pnv_pci_is_mem_pref_64(r->flags)) {
 239                         found = true;
 240                         break;
 241                 }
 242         }
 243
 244         /* No M64 window found ? */
 245         if (!found)
 246                 return IODA_INVALID_PE;
 247
 248         /* Allocate bitmap */
 249         size = _ALIGN_UP(phb->ioda.total_pe / 8, sizeof(unsigned long));
 250         pe_alloc = kzalloc(size, GFP_KERNEL);
 251         if (!pe_alloc) {
 252                 pr_warn("%s: Out of memory !\n",
 253                         __func__);
 254                 return IODA_INVALID_PE;
 255         }
 256
 257         /*
 258          * Figure out reserved PE numbers by the PE
 259          * the its child PEs.
 260          */
 261         start = (r->start - phb->ioda.m64_base) / segsz;
 262         for (i = 0; i < resource_size(r) / segsz; i++)
 263                 set_bit(start + i, pe_alloc);
 264
 265         if (all)
 266                 goto done;
 267
 268         /*
 269          * If the PE doesn't cover all subordinate buses,
 270          * we need subtract from reserved PEs for children.
 271          */
 272         list_for_each_entry(pdev, &bus->devices, bus_list) {
 273                 if (!pdev->subordinate)
 274                         continue;
 275
 276                 pci_bus_for_each_resource(pdev->subordinate, r, i) {
 277                         if (!r || !r->parent ||
 278                             !pnv_pci_is_mem_pref_64(r->flags))
 279                                 continue;
 280
 281                         start = (r->start - phb->ioda.m64_base) / segsz;
 282                         for (j = 0; j < resource_size(r) / segsz ; j++)
 283                                 clear_bit(start + j, pe_alloc);
 284                 }
 285         }
 286
 287         /*
 288          * the current bus might not own M64 window and that's all
 289          * contributed by its child buses. For the case, we needn't
 290          * pick M64 dependent PE#.
 291          */
 292         if (bitmap_empty(pe_alloc, phb->ioda.total_pe)) {
 293                 kfree(pe_alloc);
 294                 return IODA_INVALID_PE;
 295         }
 296
 297         /*
 298          * Figure out the master PE and put all slave PEs to master
 299          * PE's list to form compound PE.
 300          */
 301 done:
 302         master_pe = NULL;
 303         i = -1;
 304         while ((i = find_next_bit(pe_alloc, phb->ioda.total_pe, i + 1)) <
 305                 phb->ioda.total_pe) {
 306                 pe = &phb->ioda.pe_array[i];
 307
 308                 if (!master_pe) {
 309                         pe->flags |= PNV_IODA_PE_MASTER;
 310                         INIT_LIST_HEAD(&pe->slaves);
 311                         master_pe = pe;
 312                 } else {
 313                         pe->flags |= PNV_IODA_PE_SLAVE;
 314                         pe->master = master_pe;
 315                         list_add_tail(&pe->list, &master_pe->slaves);
 316                 }
 317         }
 318
 319         kfree(pe_alloc);
 320         return master_pe->pe_number;
 321 }
 322
 323 static void __init pnv_ioda_parse_m64_window(struct pnv_phb *phb)
 324 {
 325         struct pci_controller *hose = phb->hose;
 326         struct device_node *dn = hose->dn;
 327         struct resource *res;
 328         const u32 *r;
 329         u64 pci_addr;
 330
 331         /* FIXME: Support M64 for P7IOC */
 332         if (phb->type != PNV_PHB_IODA2) {
 333                 pr_info("  Not support M64 window\n");
 334                 return;
 335         }
 336
 337         if (!firmware_has_feature(FW_FEATURE_OPALv3)) {
 338                 pr_info("  Firmware too old to support M64 window\n");
 339                 return;
 340         }
 341
 342         r = of_get_property(dn, "ibm,opal-m64-window", NULL);
 343         if (!r) {
 344                 pr_info("  No <ibm,opal-m64-window> on %s\n",
 345                         dn->full_name);
 346                 return;
 347         }
 348
 349         res = &hose->mem_resources[1];
 350         res->start = of_translate_address(dn, r + 2);
 351         res->end = res->start + of_read_number(r + 4, 2) - 1;
 352         res->flags = (IORESOURCE_MEM | IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
 353         pci_addr = of_read_number(r, 2);
 354         hose->mem_offset[1] = res->start - pci_addr;
 355
 356         phb->ioda.m64_size = resource_size(res);
 357         phb->ioda.m64_segsize = phb->ioda.m64_size / phb->ioda.total_pe;
 358         phb->ioda.m64_base = pci_addr;
 359
 360         /* Use last M64 BAR to cover M64 window */
 361         phb->ioda.m64_bar_idx = 15;
 362         phb->init_m64 = pnv_ioda2_init_m64;
 363         phb->reserve_m64_pe = pnv_ioda2_reserve_m64_pe;
 364         phb->pick_m64_pe = pnv_ioda2_pick_m64_pe;
 365 }
 366
 367 static void pnv_ioda_freeze_pe(struct pnv_phb *phb, int pe_no)
 368 {
 369         struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_no];
 370         struct pnv_ioda_pe *slave;
 371         s64 rc;
 372
 373         /* Fetch master PE */
 374         if (pe->flags & PNV_IODA_PE_SLAVE) {
 375                 pe = pe->master;
 376                 if (WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)))
 377                         return;
 378
 379                 pe_no = pe->pe_number;
 380         }
 381
 382         /* Freeze master PE */
 383         rc = opal_pci_eeh_freeze_set(phb->opal_id,
 384                                      pe_no,
 385                                      OPAL_EEH_ACTION_SET_FREEZE_ALL);
 386         if (rc != OPAL_SUCCESS) {
 387                 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
 388                         __func__, rc, phb->hose->global_number, pe_no);
 389                 return;
 390         }
 391
 392         /* Freeze slave PEs */
 393         if (!(pe->flags & PNV_IODA_PE_MASTER))
 394                 return;
 395
 396         list_for_each_entry(slave, &pe->slaves, list) {
 397                 rc = opal_pci_eeh_freeze_set(phb->opal_id,
 398                                              slave->pe_number,
 399                                              OPAL_EEH_ACTION_SET_FREEZE_ALL);
 400                 if (rc != OPAL_SUCCESS)
 401                         pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
 402                                 __func__, rc, phb->hose->global_number,
 403                                 slave->pe_number);
 404         }
 405 }
 406
 407 static int pnv_ioda_unfreeze_pe(struct pnv_phb *phb, int pe_no, int opt)
 408 {
 409         struct pnv_ioda_pe *pe, *slave;
 410         s64 rc;
 411
 412         /* Find master PE */
 413         pe = &phb->ioda.pe_array[pe_no];
 414         if (pe->flags & PNV_IODA_PE_SLAVE) {
 415                 pe = pe->master;
 416                 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
 417                 pe_no = pe->pe_number;
 418         }
 419
 420         /* Clear frozen state for master PE */
 421         rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, opt);
 422         if (rc != OPAL_SUCCESS) {
 423                 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
 424                         __func__, rc, opt, phb->hose->global_number, pe_no);
 425                 return -EIO;
 426         }
 427
 428         if (!(pe->flags & PNV_IODA_PE_MASTER))
 429                 return 0;
 430
 431         /* Clear frozen state for slave PEs */
 432         list_for_each_entry(slave, &pe->slaves, list) {
 433                 rc = opal_pci_eeh_freeze_clear(phb->opal_id,
 434                                              slave->pe_number,
 435                                              opt);
 436                 if (rc != OPAL_SUCCESS) {
 437                         pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
 438                                 __func__, rc, opt, phb->hose->global_number,
 439                                 slave->pe_number);
 440                         return -EIO;
 441                 }
 442         }
 443
 444         return 0;
 445 }
 446
 447 static int pnv_ioda_get_pe_state(struct pnv_phb *phb, int pe_no)
 448 {
 449         struct pnv_ioda_pe *slave, *pe;
 450         u8 fstate, state;
 451         __be16 pcierr;
 452         s64 rc;
 453
 454         /* Sanity check on PE number */
 455         if (pe_no < 0 || pe_no >= phb->ioda.total_pe)
 456                 return OPAL_EEH_STOPPED_PERM_UNAVAIL;
 457
 458         /*
 459          * Fetch the master PE and the PE instance might be
 460          * not initialized yet.
 461          */
 462         pe = &phb->ioda.pe_array[pe_no];
 463         if (pe->flags & PNV_IODA_PE_SLAVE) {
 464                 pe = pe->master;
 465                 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
 466                 pe_no = pe->pe_number;
 467         }
 468
 469         /* Check the master PE */
 470         rc = opal_pci_eeh_freeze_status(phb->opal_id, pe_no,
 471                                         &state, &pcierr, NULL);
 472         if (rc != OPAL_SUCCESS) {
 473                 pr_warn("%s: Failure %lld getting "
 474                         "PHB#%x-PE#%x state\n",
 475                         __func__, rc,
 476                         phb->hose->global_number, pe_no);
 477                 return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
 478         }
 479
 480         /* Check the slave PE */
 481         if (!(pe->flags & PNV_IODA_PE_MASTER))
 482                 return state;
 483
 484         list_for_each_entry(slave, &pe->slaves, list) {
 485                 rc = opal_pci_eeh_freeze_status(phb->opal_id,
 486                                                 slave->pe_number,
 487                                                 &fstate,
 488                                                 &pcierr,
 489                                                 NULL);
 490                 if (rc != OPAL_SUCCESS) {
 491                         pr_warn("%s: Failure %lld getting "
 492                                 "PHB#%x-PE#%x state\n",
 493                                 __func__, rc,
 494                                 phb->hose->global_number, slave->pe_number);
 495                         return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
 496                 }
 497
 498                 /*
 499                  * Override the result based on the ascending
 500                  * priority.
 501                  */
 502                 if (fstate > state)
 503                         state = fstate;
 504         }
 505
 506         return state;
 507 }
 508
 509 /* Currently those 2 are only used when MSIs are enabled, this will change
 510  * but in the meantime, we need to protect them to avoid warnings
 511  */
 512 #ifdef CONFIG_PCI_MSI
 513 static struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev)
 514 {
 515         struct pci_controller *hose = pci_bus_to_host(dev->bus);
 516         struct pnv_phb *phb = hose->private_data;
 517         struct pci_dn *pdn = pci_get_pdn(dev);
 518
 519         if (!pdn)
 520                 return NULL;
 521         if (pdn->pe_number == IODA_INVALID_PE)
 522                 return NULL;
 523         return &phb->ioda.pe_array[pdn->pe_number];
 524 }
 525 #endif /* CONFIG_PCI_MSI */
 526
 527 static int pnv_ioda_set_one_peltv(struct pnv_phb *phb,
 528                                   struct pnv_ioda_pe *parent,
 529                                   struct pnv_ioda_pe *child,
 530                                   bool is_add)
 531 {
 532         const char *desc = is_add ? "adding" : "removing";
 533         uint8_t op = is_add ? OPAL_ADD_PE_TO_DOMAIN :
 534                               OPAL_REMOVE_PE_FROM_DOMAIN;
 535         struct pnv_ioda_pe *slave;
 536         long rc;
 537
 538         /* Parent PE affects child PE */
 539         rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
 540                                 child->pe_number, op);
 541         if (rc != OPAL_SUCCESS) {
 542                 pe_warn(child, "OPAL error %ld %s to parent PELTV\n",
 543                         rc, desc);
 544                 return -ENXIO;
 545         }
 546
 547         if (!(child->flags & PNV_IODA_PE_MASTER))
 548                 return 0;
 549
 550         /* Compound case: parent PE affects slave PEs */
 551         list_for_each_entry(slave, &child->slaves, list) {
 552                 rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
 553                                         slave->pe_number, op);
 554                 if (rc != OPAL_SUCCESS) {
 555                         pe_warn(slave, "OPAL error %ld %s to parent PELTV\n",
 556                                 rc, desc);
 557                         return -ENXIO;
 558                 }
 559         }
 560
 561         return 0;
 562 }
 563
 564 static int pnv_ioda_set_peltv(struct pnv_phb *phb,
 565                               struct pnv_ioda_pe *pe,
 566                               bool is_add)
 567 {
 568         struct pnv_ioda_pe *slave;
 569         struct pci_dev *pdev;
 570         int ret;
 571
 572         /*
 573          * Clear PE frozen state. If it's master PE, we need
 574          * clear slave PE frozen state as well.
 575          */
 576         if (is_add) {
 577                 opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
 578                                           OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 579                 if (pe->flags & PNV_IODA_PE_MASTER) {
 580                         list_for_each_entry(slave, &pe->slaves, list)
 581                                 opal_pci_eeh_freeze_clear(phb->opal_id,
 582                                                           slave->pe_number,
 583                                                           OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 584                 }
 585         }
 586
 587         /*
 588          * Associate PE in PELT. We need add the PE into the
 589          * corresponding PELT-V as well. Otherwise, the error
 590          * originated from the PE might contribute to other
 591          * PEs.
 592          */
 593         ret = pnv_ioda_set_one_peltv(phb, pe, pe, is_add);
 594         if (ret)
 595                 return ret;
 596
 597         /* For compound PEs, any one affects all of them */
 598         if (pe->flags & PNV_IODA_PE_MASTER) {
 599                 list_for_each_entry(slave, &pe->slaves, list) {
 600                         ret = pnv_ioda_set_one_peltv(phb, slave, pe, is_add);
 601                         if (ret)
 602                                 return ret;
 603                 }
 604         }
 605
 606         if (pe->flags & (PNV_IODA_PE_BUS_ALL | PNV_IODA_PE_BUS))
 607                 pdev = pe->pbus->self;
 608         else
 609                 pdev = pe->pdev->bus->self;
 610         while (pdev) {
 611                 struct pci_dn *pdn = pci_get_pdn(pdev);
 612                 struct pnv_ioda_pe *parent;
 613
 614                 if (pdn && pdn->pe_number != IODA_INVALID_PE) {
 615                         parent = &phb->ioda.pe_array[pdn->pe_number];
 616                         ret = pnv_ioda_set_one_peltv(phb, parent, pe, is_add);
 617                         if (ret)
 618                                 return ret;
 619                 }
 620
 621                 pdev = pdev->bus->self;
 622         }
 623
 624         return 0;
 625 }
 626
 627 static int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
 628 {
 629         struct pci_dev *parent;
 630         uint8_t bcomp, dcomp, fcomp;
 631         long rc, rid_end, rid;
 632
 633         /* Bus validation ? */
 634         if (pe->pbus) {
 635                 int count;
 636
 637                 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
 638                 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
 639                 parent = pe->pbus->self;
 640                 if (pe->flags & PNV_IODA_PE_BUS_ALL)
 641                         count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
 642                 else
 643                         count = 1;
 644
 645                 switch(count) {
 646                 case  1: bcomp = OpalPciBusAll;         break;
 647                 case  2: bcomp = OpalPciBus7Bits;       break;
 648                 case  4: bcomp = OpalPciBus6Bits;       break;
 649                 case  8: bcomp = OpalPciBus5Bits;       break;
 650                 case 16: bcomp = OpalPciBus4Bits;       break;
 651                 case 32: bcomp = OpalPciBus3Bits;       break;
 652                 default:
 653                         pr_err("%s: Number of subordinate busses %d"
 654                                " unsupported\n",
 655                                pci_name(pe->pbus->self), count);
 656                         /* Do an exact match only */
 657                         bcomp = OpalPciBusAll;
 658                 }
 659                 rid_end = pe->rid + (count << 8);
 660         } else {
 661                 parent = pe->pdev->bus->self;
 662                 bcomp = OpalPciBusAll;
 663                 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
 664                 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
 665                 rid_end = pe->rid + 1;
 666         }
 667
 668         /*
 669          * Associate PE in PELT. We need add the PE into the
 670          * corresponding PELT-V as well. Otherwise, the error
 671          * originated from the PE might contribute to other
 672          * PEs.
 673          */
 674         rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
 675                              bcomp, dcomp, fcomp, OPAL_MAP_PE);
 676         if (rc) {
 677                 pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
 678                 return -ENXIO;
 679         }
 680
 681         /* Configure PELTV */
 682         pnv_ioda_set_peltv(phb, pe, true);
 683
 684         /* Setup reverse map */
 685         for (rid = pe->rid; rid < rid_end; rid++)
 686                 phb->ioda.pe_rmap[rid] = pe->pe_number;
 687
 688         /* Setup one MVTs on IODA1 */
 689         if (phb->type != PNV_PHB_IODA1) {
 690                 pe->mve_number = 0;
 691                 goto out;
 692         }
 693
 694         pe->mve_number = pe->pe_number;
 695         rc = opal_pci_set_mve(phb->opal_id, pe->mve_number, pe->pe_number);
 696         if (rc != OPAL_SUCCESS) {
 697                 pe_err(pe, "OPAL error %ld setting up MVE %d\n",
 698                        rc, pe->mve_number);
 699                 pe->mve_number = -1;
 700         } else {
 701                 rc = opal_pci_set_mve_enable(phb->opal_id,
 702                                              pe->mve_number, OPAL_ENABLE_MVE);
 703                 if (rc) {
 704                         pe_err(pe, "OPAL error %ld enabling MVE %d\n",
 705                                rc, pe->mve_number);
 706                         pe->mve_number = -1;
 707                 }
 708         }
 709
 710 out:
 711         return 0;
 712 }
 713
 714 static void pnv_ioda_link_pe_by_weight(struct pnv_phb *phb,
 715                                        struct pnv_ioda_pe *pe)
 716 {
 717         struct pnv_ioda_pe *lpe;
 718
 719         list_for_each_entry(lpe, &phb->ioda.pe_dma_list, dma_link) {
 720                 if (lpe->dma_weight < pe->dma_weight) {
 721                         list_add_tail(&pe->dma_link, &lpe->dma_link);
 722                         return;
 723                 }
 724         }
 725         list_add_tail(&pe->dma_link, &phb->ioda.pe_dma_list);
 726 }
 727
 728 static unsigned int pnv_ioda_dma_weight(struct pci_dev *dev)
 729 {
 730         /* This is quite simplistic. The "base" weight of a device
 731          * is 10. 0 means no DMA is to be accounted for it.
 732          */
 733
 734         /* If it's a bridge, no DMA */
 735         if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
 736                 return 0;
 737
 738         /* Reduce the weight of slow USB controllers */
 739         if (dev->class == PCI_CLASS_SERIAL_USB_UHCI ||
 740             dev->class == PCI_CLASS_SERIAL_USB_OHCI ||
 741             dev->class == PCI_CLASS_SERIAL_USB_EHCI)
 742                 return 3;
 743
 744         /* Increase the weight of RAID (includes Obsidian) */
 745         if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID)
 746                 return 15;
 747
 748         /* Default */
 749         return 10;
 750 }
 751
 752 #if 0
 753 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev)
 754 {
 755         struct pci_controller *hose = pci_bus_to_host(dev->bus);
 756         struct pnv_phb *phb = hose->private_data;
 757         struct pci_dn *pdn = pci_get_pdn(dev);
 758         struct pnv_ioda_pe *pe;
 759         int pe_num;
 760
 761         if (!pdn) {
 762                 pr_err("%s: Device tree node not associated properly\n",
 763                            pci_name(dev));
 764                 return NULL;
 765         }
 766         if (pdn->pe_number != IODA_INVALID_PE)
 767                 return NULL;
 768
 769         /* PE#0 has been pre-set */
 770         if (dev->bus->number == 0)
 771                 pe_num = 0;
 772         else
 773                 pe_num = pnv_ioda_alloc_pe(phb);
 774         if (pe_num == IODA_INVALID_PE) {
 775                 pr_warning("%s: Not enough PE# available, disabling device\n",
 776                            pci_name(dev));
 777                 return NULL;
 778         }
 779
 780         /* NOTE: We get only one ref to the pci_dev for the pdn, not for the
 781          * pointer in the PE data structure, both should be destroyed at the
 782          * same time. However, this needs to be looked at more closely again
 783          * once we actually start removing things (Hotplug, SR-IOV, ...)
 784          *
 785          * At some point we want to remove the PDN completely anyways
 786          */
 787         pe = &phb->ioda.pe_array[pe_num];
 788         pci_dev_get(dev);
 789         pdn->pcidev = dev;
 790         pdn->pe_number = pe_num;
 791         pe->pdev = dev;
 792         pe->pbus = NULL;
 793         pe->tce32_seg = -1;
 794         pe->mve_number = -1;
 795         pe->rid = dev->bus->number << 8 | pdn->devfn;
 796
 797         pe_info(pe, "Associated device to PE\n");
 798
 799         if (pnv_ioda_configure_pe(phb, pe)) {
 800                 /* XXX What do we do here ? */
 801                 if (pe_num)
 802                         pnv_ioda_free_pe(phb, pe_num);
 803                 pdn->pe_number = IODA_INVALID_PE;
 804                 pe->pdev = NULL;
 805                 pci_dev_put(dev);
 806                 return NULL;
 807         }
 808
 809         /* Assign a DMA weight to the device */
 810         pe->dma_weight = pnv_ioda_dma_weight(dev);
 811         if (pe->dma_weight != 0) {
 812                 phb->ioda.dma_weight += pe->dma_weight;
 813                 phb->ioda.dma_pe_count++;
 814         }
 815
 816         /* Link the PE */
 817         pnv_ioda_link_pe_by_weight(phb, pe);
 818
 819         return pe;
 820 }
 821 #endif /* Useful for SRIOV case */
 822
 823 static void pnv_ioda_setup_same_PE(struct pci_bus *bus, struct pnv_ioda_pe *pe)
 824 {
 825         struct pci_dev *dev;
 826
 827         list_for_each_entry(dev, &bus->devices, bus_list) {
 828                 struct pci_dn *pdn = pci_get_pdn(dev);
 829
 830                 if (pdn == NULL) {
 831                         pr_warn("%s: No device node associated with device !\n",
 832                                 pci_name(dev));
 833                         continue;
 834                 }
 835                 pdn->pcidev = dev;
 836                 pdn->pe_number = pe->pe_number;
 837                 pe->dma_weight += pnv_ioda_dma_weight(dev);
 838                 if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
 839                         pnv_ioda_setup_same_PE(dev->subordinate, pe);
 840         }
 841 }
 842
 843 /*
 844  * There're 2 types of PCI bus sensitive PEs: One that is compromised of
 845  * single PCI bus. Another one that contains the primary PCI bus and its
 846  * subordinate PCI devices and buses. The second type of PE is normally
 847  * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
 848  */
 849 static void pnv_ioda_setup_bus_PE(struct pci_bus *bus, int all)
 850 {
 851         struct pci_controller *hose = pci_bus_to_host(bus);
 852         struct pnv_phb *phb = hose->private_data;
 853         struct pnv_ioda_pe *pe;
 854         int pe_num = IODA_INVALID_PE;
 855
 856         /* Check if PE is determined by M64 */
 857         if (phb->pick_m64_pe)
 858                 pe_num = phb->pick_m64_pe(phb, bus, all);
 859
 860         /* The PE number isn't pinned by M64 */
 861         if (pe_num == IODA_INVALID_PE)
 862                 pe_num = pnv_ioda_alloc_pe(phb);
 863
 864         if (pe_num == IODA_INVALID_PE) {
 865                 pr_warning("%s: Not enough PE# available for PCI bus %04x:%02x\n",
 866                         __func__, pci_domain_nr(bus), bus->number);
 867                 return;
 868         }
 869
 870         pe = &phb->ioda.pe_array[pe_num];
 871         pe->flags |= (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS);
 872         pe->pbus = bus;
 873         pe->pdev = NULL;
 874         pe->tce32_seg = -1;
 875         pe->mve_number = -1;
 876         pe->rid = bus->busn_res.start << 8;
 877         pe->dma_weight = 0;
 878
 879         if (all)
 880                 pe_info(pe, "Secondary bus %d..%d associated with PE#%d\n",
 881                         bus->busn_res.start, bus->busn_res.end, pe_num);
 882         else
 883                 pe_info(pe, "Secondary bus %d associated with PE#%d\n",
 884                         bus->busn_res.start, pe_num);
 885
 886         if (pnv_ioda_configure_pe(phb, pe)) {
 887                 /* XXX What do we do here ? */
 888                 if (pe_num)
 889                         pnv_ioda_free_pe(phb, pe_num);
 890                 pe->pbus = NULL;
 891                 return;
 892         }
 893
 894         /* Associate it with all child devices */
 895         pnv_ioda_setup_same_PE(bus, pe);
 896
 897         /* Put PE to the list */
 898         list_add_tail(&pe->list, &phb->ioda.pe_list);
 899
 900         /* Account for one DMA PE if at least one DMA capable device exist
 901          * below the bridge
 902          */
 903         if (pe->dma_weight != 0) {
 904                 phb->ioda.dma_weight += pe->dma_weight;
 905                 phb->ioda.dma_pe_count++;
 906         }
 907
 908         /* Link the PE */
 909         pnv_ioda_link_pe_by_weight(phb, pe);
 910 }
 911
 912 static void pnv_ioda_setup_PEs(struct pci_bus *bus)
 913 {
 914         struct pci_dev *dev;
 915
 916         pnv_ioda_setup_bus_PE(bus, 0);
 917
 918         list_for_each_entry(dev, &bus->devices, bus_list) {
 919                 if (dev->subordinate) {
 920                         if (pci_pcie_type(dev) == PCI_EXP_TYPE_PCI_BRIDGE)
 921                                 pnv_ioda_setup_bus_PE(dev->subordinate, 1);
 922                         else
 923                                 pnv_ioda_setup_PEs(dev->subordinate);
 924                 }
 925         }
 926 }
 927
 928 /*
 929  * Configure PEs so that the downstream PCI buses and devices
 930  * could have their associated PE#. Unfortunately, we didn't
 931  * figure out the way to identify the PLX bridge yet. So we
 932  * simply put the PCI bus and the subordinate behind the root
 933  * port to PE# here. The game rule here is expected to be changed
 934  * as soon as we can detected PLX bridge correctly.
 935  */
 936 static void pnv_pci_ioda_setup_PEs(void)
 937 {
 938         struct pci_controller *hose, *tmp;
 939         struct pnv_phb *phb;
 940
 941         list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
 942                 phb = hose->private_data;
 943
 944                 /* M64 layout might affect PE allocation */
 945                 if (phb->reserve_m64_pe)
 946                         phb->reserve_m64_pe(phb);
 947
 948                 pnv_ioda_setup_PEs(hose->bus);
 949         }
 950 }
 951
 952 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb *phb, struct pci_dev *pdev)
 953 {
 954         struct pci_dn *pdn = pci_get_pdn(pdev);
 955         struct pnv_ioda_pe *pe;
 956
 957         /*
 958          * The function can be called while the PE#
 959          * hasn't been assigned. Do nothing for the
 960          * case.
 961          */
 962         if (!pdn || pdn->pe_number == IODA_INVALID_PE)
 963                 return;
 964
 965         pe = &phb->ioda.pe_array[pdn->pe_number];
 966         WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops);
 967         set_iommu_table_base_and_group(&pdev->dev, &pe->tce32_table);
 968 }
 969
 970 static int pnv_pci_ioda_dma_set_mask(struct pnv_phb *phb,
 971                                      struct pci_dev *pdev, u64 dma_mask)
 972 {
 973         struct pci_dn *pdn = pci_get_pdn(pdev);
 974         struct pnv_ioda_pe *pe;
 975         uint64_t top;
 976         bool bypass = false;
 977
 978         if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
 979                 return -ENODEV;;
 980
 981         pe = &phb->ioda.pe_array[pdn->pe_number];
 982         if (pe->tce_bypass_enabled) {
 983                 top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1;
 984                 bypass = (dma_mask >= top);
 985         }
 986
 987         if (bypass) {
 988                 dev_info(&pdev->dev, "Using 64-bit DMA iommu bypass\n");
 989                 set_dma_ops(&pdev->dev, &dma_direct_ops);
 990                 set_dma_offset(&pdev->dev, pe->tce_bypass_base);
 991         } else {
 992                 dev_info(&pdev->dev, "Using 32-bit DMA via iommu\n");
 993                 set_dma_ops(&pdev->dev, &dma_iommu_ops);
 994                 set_iommu_table_base(&pdev->dev, &pe->tce32_table);
 995         }
 996         *pdev->dev.dma_mask = dma_mask;
 997         return 0;
 998 }
 999
1000 static u64 pnv_pci_ioda_dma_get_required_mask(struct pnv_phb *phb,
1001                                               struct pci_dev *pdev)
1002 {
1003         struct pci_dn *pdn = pci_get_pdn(pdev);
1004         struct pnv_ioda_pe *pe;
1005         u64 end, mask;
1006
1007         if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1008                 return 0;
1009
1010         pe = &phb->ioda.pe_array[pdn->pe_number];
1011         if (!pe->tce_bypass_enabled)
1012                 return __dma_get_required_mask(&pdev->dev);
1013
1014
1015         end = pe->tce_bypass_base + memblock_end_of_DRAM();
1016         mask = 1ULL << (fls64(end) - 1);
1017         mask += mask - 1;
1018
1019         return mask;
1020 }
1021
1022 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe,
1023                                    struct pci_bus *bus,
1024                                    bool add_to_iommu_group)
1025 {
1026         struct pci_dev *dev;
1027
1028         list_for_each_entry(dev, &bus->devices, bus_list) {
1029                 if (add_to_iommu_group)
1030                         set_iommu_table_base_and_group(&dev->dev,
1031                                                        &pe->tce32_table);
1032                 else
1033                         set_iommu_table_base(&dev->dev, &pe->tce32_table);
1034
1035                 if (dev->subordinate)
1036                         pnv_ioda_setup_bus_dma(pe, dev->subordinate,
1037                                                add_to_iommu_group);
1038         }
1039 }
1040
1041 static void pnv_pci_ioda1_tce_invalidate(struct pnv_ioda_pe *pe,
1042                                          struct iommu_table *tbl,
1043                                          __be64 *startp, __be64 *endp, bool rm)
1044 {
1045         __be64 __iomem *invalidate = rm ?
1046                 (__be64 __iomem *)pe->tce_inval_reg_phys :
1047                 (__be64 __iomem *)tbl->it_index;
1048         unsigned long start, end, inc;
1049         const unsigned shift = tbl->it_page_shift;
1050
1051         start = __pa(startp);
1052         end = __pa(endp);
1053
1054         /* BML uses this case for p6/p7/galaxy2: Shift addr and put in node */
1055         if (tbl->it_busno) {
1056                 start <<= shift;
1057                 end <<= shift;
1058                 inc = 128ull << shift;
1059                 start |= tbl->it_busno;
1060                 end |= tbl->it_busno;
1061         } else if (tbl->it_type & TCE_PCI_SWINV_PAIR) {
1062                 /* p7ioc-style invalidation, 2 TCEs per write */
1063                 start |= (1ull << 63);
1064                 end |= (1ull << 63);
1065                 inc = 16;
1066         } else {
1067                 /* Default (older HW) */
1068                 inc = 128;
1069         }
1070
1071         end |= inc - 1; /* round up end to be different than start */
1072
1073         mb(); /* Ensure above stores are visible */
1074         while (start <= end) {
1075                 if (rm)
1076                         __raw_rm_writeq(cpu_to_be64(start), invalidate);
1077                 else
1078                         __raw_writeq(cpu_to_be64(start), invalidate);
1079                 start += inc;
1080         }
1081
1082         /*
1083          * The iommu layer will do another mb() for us on build()
1084          * and we don't care on free()
1085          */
1086 }
1087
1088 static void pnv_pci_ioda2_tce_invalidate(struct pnv_ioda_pe *pe,
1089                                          struct iommu_table *tbl,
1090                                          __be64 *startp, __be64 *endp, bool rm)
1091 {
1092         unsigned long start, end, inc;
1093         __be64 __iomem *invalidate = rm ?
1094                 (__be64 __iomem *)pe->tce_inval_reg_phys :
1095                 (__be64 __iomem *)tbl->it_index;
1096         const unsigned shift = tbl->it_page_shift;
1097
1098         /* We'll invalidate DMA address in PE scope */
1099         start = 0x2ull << 60;
1100         start |= (pe->pe_number & 0xFF);
1101         end = start;
1102
1103         /* Figure out the start, end and step */
1104         inc = tbl->it_offset + (((u64)startp - tbl->it_base) / sizeof(u64));
1105         start |= (inc << shift);
1106         inc = tbl->it_offset + (((u64)endp - tbl->it_base) / sizeof(u64));
1107         end |= (inc << shift);
1108         inc = (0x1ull << shift);
1109         mb();
1110
1111         while (start <= end) {
1112                 if (rm)
1113                         __raw_rm_writeq(cpu_to_be64(start), invalidate);
1114                 else
1115                         __raw_writeq(cpu_to_be64(start), invalidate);
1116                 start += inc;
1117         }
1118 }
1119
1120 void pnv_pci_ioda_tce_invalidate(struct iommu_table *tbl,
1121                                  __be64 *startp, __be64 *endp, bool rm)
1122 {
1123         struct pnv_ioda_pe *pe = container_of(tbl, struct pnv_ioda_pe,
1124                                               tce32_table);
1125         struct pnv_phb *phb = pe->phb;
1126
1127         if (phb->type == PNV_PHB_IODA1)
1128                 pnv_pci_ioda1_tce_invalidate(pe, tbl, startp, endp, rm);
1129         else
1130                 pnv_pci_ioda2_tce_invalidate(pe, tbl, startp, endp, rm);
1131 }
1132
1133 static void pnv_pci_ioda_setup_dma_pe(struct pnv_phb *phb,
1134                                       struct pnv_ioda_pe *pe, unsigned int base,
1135                                       unsigned int segs)
1136 {
1137
1138         struct page *tce_mem = NULL;
1139         const __be64 *swinvp;
1140         struct iommu_table *tbl;
1141         unsigned int i;
1142         int64_t rc;
1143         void *addr;
1144
1145         /* 256M DMA window, 4K TCE pages, 8 bytes TCE */
1146 #define TCE32_TABLE_SIZE        ((0x10000000 / 0x1000) * 8)
1147
1148         /* XXX FIXME: Handle 64-bit only DMA devices */
1149         /* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
1150         /* XXX FIXME: Allocate multi-level tables on PHB3 */
1151
1152         /* We shouldn't already have a 32-bit DMA associated */
1153         if (WARN_ON(pe->tce32_seg >= 0))
1154                 return;
1155
1156         /* Grab a 32-bit TCE table */
1157         pe->tce32_seg = base;
1158         pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n",
1159                 (base << 28), ((base + segs) << 28) - 1);
1160
1161         /* XXX Currently, we allocate one big contiguous table for the
1162          * TCEs. We only really need one chunk per 256M of TCE space
1163          * (ie per segment) but that's an optimization for later, it
1164          * requires some added smarts with our get/put_tce implementation
1165          */
1166         tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
1167                                    get_order(TCE32_TABLE_SIZE * segs));
1168         if (!tce_mem) {
1169                 pe_err(pe, " Failed to allocate a 32-bit TCE memory\n");
1170                 goto fail;
1171         }
1172         addr = page_address(tce_mem);
1173         memset(addr, 0, TCE32_TABLE_SIZE * segs);
1174
1175         /* Configure HW */
1176         for (i = 0; i < segs; i++) {
1177                 rc = opal_pci_map_pe_dma_window(phb->opal_id,
1178                                               pe->pe_number,
1179                                               base + i, 1,
1180                                               __pa(addr) + TCE32_TABLE_SIZE * i,
1181                                               TCE32_TABLE_SIZE, 0x1000);
1182                 if (rc) {
1183                         pe_err(pe, " Failed to configure 32-bit TCE table,"
1184                                " err %ld\n", rc);
1185                         goto fail;
1186                 }
1187         }
1188
1189         /* Setup linux iommu table */
1190         tbl = &pe->tce32_table;
1191         pnv_pci_setup_iommu_table(tbl, addr, TCE32_TABLE_SIZE * segs,
1192                                   base << 28, IOMMU_PAGE_SHIFT_4K);
1193
1194         /* OPAL variant of P7IOC SW invalidated TCEs */
1195         swinvp = of_get_property(phb->hose->dn, "ibm,opal-tce-kill", NULL);
1196         if (swinvp) {
1197                 /* We need a couple more fields -- an address and a data
1198                  * to or.  Since the bus is only printed out on table free
1199                  * errors, and on the first pass the data will be a relative
1200                  * bus number, print that out instead.
1201                  */
1202                 pe->tce_inval_reg_phys = be64_to_cpup(swinvp);
1203                 tbl->it_index = (unsigned long)ioremap(pe->tce_inval_reg_phys,
1204                                 8);
1205                 tbl->it_type |= (TCE_PCI_SWINV_CREATE |
1206                                  TCE_PCI_SWINV_FREE   |
1207                                  TCE_PCI_SWINV_PAIR);
1208         }
1209         iommu_init_table(tbl, phb->hose->node);
1210         iommu_register_group(tbl, phb->hose->global_number, pe->pe_number);
1211
1212         if (pe->pdev)
1213                 set_iommu_table_base_and_group(&pe->pdev->dev, tbl);
1214         else
1215                 pnv_ioda_setup_bus_dma(pe, pe->pbus, true);
1216
1217         return;
1218  fail:
1219         /* XXX Failure: Try to fallback to 64-bit only ? */
1220         if (pe->tce32_seg >= 0)
1221                 pe->tce32_seg = -1;
1222         if (tce_mem)
1223                 __free_pages(tce_mem, get_order(TCE32_TABLE_SIZE * segs));
1224 }
1225
1226 static void pnv_pci_ioda2_set_bypass(struct iommu_table *tbl, bool enable)
1227 {
1228         struct pnv_ioda_pe *pe = container_of(tbl, struct pnv_ioda_pe,
1229                                               tce32_table);
1230         uint16_t window_id = (pe->pe_number << 1 ) + 1;
1231         int64_t rc;
1232
1233         pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis");
1234         if (enable) {
1235                 phys_addr_t top = memblock_end_of_DRAM();
1236
1237                 top = roundup_pow_of_two(top);
1238                 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
1239                                                      pe->pe_number,
1240                                                      window_id,
1241                                                      pe->tce_bypass_base,
1242                                                      top);
1243         } else {
1244                 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
1245                                                      pe->pe_number,
1246                                                      window_id,
1247                                                      pe->tce_bypass_base,
1248                                                      0);
1249
1250                 /*
1251                  * EEH needs the mapping between IOMMU table and group
1252                  * of those VFIO/KVM pass-through devices. We can postpone
1253                  * resetting DMA ops until the DMA mask is configured in
1254                  * host side.
1255                  */
1256                 if (pe->pdev)
1257                         set_iommu_table_base(&pe->pdev->dev, tbl);
1258                 else
1259                         pnv_ioda_setup_bus_dma(pe, pe->pbus, false);
1260         }
1261         if (rc)
1262                 pe_err(pe, "OPAL error %lld configuring bypass window\n", rc);
1263         else
1264                 pe->tce_bypass_enabled = enable;
1265 }
1266
1267 static void pnv_pci_ioda2_setup_bypass_pe(struct pnv_phb *phb,
1268                                           struct pnv_ioda_pe *pe)
1269 {
1270         /* TVE #1 is selected by PCI address bit 59 */
1271         pe->tce_bypass_base = 1ull << 59;
1272
1273         /* Install set_bypass callback for VFIO */
1274         pe->tce32_table.set_bypass = pnv_pci_ioda2_set_bypass;
1275
1276         /* Enable bypass by default */
1277         pnv_pci_ioda2_set_bypass(&pe->tce32_table, true);
1278 }
1279
1280 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
1281                                        struct pnv_ioda_pe *pe)
1282 {
1283         struct page *tce_mem = NULL;
1284         void *addr;
1285         const __be64 *swinvp;
1286         struct iommu_table *tbl;
1287         unsigned int tce_table_size, end;
1288         int64_t rc;
1289
1290         /* We shouldn't already have a 32-bit DMA associated */
1291         if (WARN_ON(pe->tce32_seg >= 0))
1292                 return;
1293
1294         /* The PE will reserve all possible 32-bits space */
1295         pe->tce32_seg = 0;
1296         end = (1 << ilog2(phb->ioda.m32_pci_base));
1297         tce_table_size = (end / 0x1000) * 8;
1298         pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n",
1299                 end);
1300
1301         /* Allocate TCE table */
1302         tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
1303                                    get_order(tce_table_size));
1304         if (!tce_mem) {
1305                 pe_err(pe, "Failed to allocate a 32-bit TCE memory\n");
1306                 goto fail;
1307         }
1308         addr = page_address(tce_mem);
1309         memset(addr, 0, tce_table_size);
1310
1311         /*
1312          * Map TCE table through TVT. The TVE index is the PE number
1313          * shifted by 1 bit for 32-bits DMA space.
1314          */
1315         rc = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
1316                                         pe->pe_number << 1, 1, __pa(addr),
1317                                         tce_table_size, 0x1000);
1318         if (rc) {
1319                 pe_err(pe, "Failed to configure 32-bit TCE table,"
1320                        " err %ld\n", rc);
1321                 goto fail;
1322         }
1323
1324         /* Setup linux iommu table */
1325         tbl = &pe->tce32_table;
1326         pnv_pci_setup_iommu_table(tbl, addr, tce_table_size, 0,
1327                         IOMMU_PAGE_SHIFT_4K);
1328
1329         /* OPAL variant of PHB3 invalidated TCEs */
1330         swinvp = of_get_property(phb->hose->dn, "ibm,opal-tce-kill", NULL);
1331         if (swinvp) {
1332                 /* We need a couple more fields -- an address and a data
1333                  * to or.  Since the bus is only printed out on table free
1334                  * errors, and on the first pass the data will be a relative
1335                  * bus number, print that out instead.
1336                  */
1337                 pe->tce_inval_reg_phys = be64_to_cpup(swinvp);
1338                 tbl->it_index = (unsigned long)ioremap(pe->tce_inval_reg_phys,
1339                                 8);
1340                 tbl->it_type |= (TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE);
1341         }
1342         iommu_init_table(tbl, phb->hose->node);
1343         iommu_register_group(tbl, phb->hose->global_number, pe->pe_number);
1344
1345         if (pe->pdev)
1346                 set_iommu_table_base_and_group(&pe->pdev->dev, tbl);
1347         else
1348                 pnv_ioda_setup_bus_dma(pe, pe->pbus, true);
1349
1350         /* Also create a bypass window */
1351         pnv_pci_ioda2_setup_bypass_pe(phb, pe);
1352         return;
1353 fail:
1354         if (pe->tce32_seg >= 0)
1355                 pe->tce32_seg = -1;
1356         if (tce_mem)
1357                 __free_pages(tce_mem, get_order(tce_table_size));
1358 }
1359
1360 static void pnv_ioda_setup_dma(struct pnv_phb *phb)
1361 {
1362         struct pci_controller *hose = phb->hose;
1363         unsigned int residual, remaining, segs, tw, base;
1364         struct pnv_ioda_pe *pe;
1365
1366         /* If we have more PE# than segments available, hand out one
1367          * per PE until we run out and let the rest fail. If not,
1368          * then we assign at least one segment per PE, plus more based
1369          * on the amount of devices under that PE
1370          */
1371         if (phb->ioda.dma_pe_count > phb->ioda.tce32_count)
1372                 residual = 0;
1373         else
1374                 residual = phb->ioda.tce32_count -
1375                         phb->ioda.dma_pe_count;
1376
1377         pr_info("PCI: Domain %04x has %ld available 32-bit DMA segments\n",
1378                 hose->global_number, phb->ioda.tce32_count);
1379         pr_info("PCI: %d PE# for a total weight of %d\n",
1380                 phb->ioda.dma_pe_count, phb->ioda.dma_weight);
1381
1382         /* Walk our PE list and configure their DMA segments, hand them
1383          * out one base segment plus any residual segments based on
1384          * weight
1385          */
1386         remaining = phb->ioda.tce32_count;
1387         tw = phb->ioda.dma_weight;
1388         base = 0;
1389         list_for_each_entry(pe, &phb->ioda.pe_dma_list, dma_link) {
1390                 if (!pe->dma_weight)
1391                         continue;
1392                 if (!remaining) {
1393                         pe_warn(pe, "No DMA32 resources available\n");
1394                         continue;
1395                 }
1396                 segs = 1;
1397                 if (residual) {
1398                         segs += ((pe->dma_weight * residual)  + (tw / 2)) / tw;
1399                         if (segs > remaining)
1400                                 segs = remaining;
1401                 }
1402
1403                 /*
1404                  * For IODA2 compliant PHB3, we needn't care about the weight.
1405                  * The all available 32-bits DMA space will be assigned to
1406                  * the specific PE.
1407                  */
1408                 if (phb->type == PNV_PHB_IODA1) {
1409                         pe_info(pe, "DMA weight %d, assigned %d DMA32 segments\n",
1410                                 pe->dma_weight, segs);
1411                         pnv_pci_ioda_setup_dma_pe(phb, pe, base, segs);
1412                 } else {
1413                         pe_info(pe, "Assign DMA32 space\n");
1414                         segs = 0;
1415                         pnv_pci_ioda2_setup_dma_pe(phb, pe);
1416                 }
1417
1418                 remaining -= segs;
1419                 base += segs;
1420         }
1421 }
1422
1423 #ifdef CONFIG_PCI_MSI
1424 static void pnv_ioda2_msi_eoi(struct irq_data *d)
1425 {
1426         unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
1427         struct irq_chip *chip = irq_data_get_irq_chip(d);
1428         struct pnv_phb *phb = container_of(chip, struct pnv_phb,
1429                                            ioda.irq_chip);
1430         int64_t rc;
1431
1432         rc = opal_pci_msi_eoi(phb->opal_id, hw_irq);
1433         WARN_ON_ONCE(rc);
1434
1435         icp_native_eoi(d);
1436 }
1437
1438
1439 static void set_msi_irq_chip(struct pnv_phb *phb, unsigned int virq)
1440 {
1441         struct irq_data *idata;
1442         struct irq_chip *ichip;
1443
1444         if (phb->type != PNV_PHB_IODA2)
1445                 return;
1446
1447         if (!phb->ioda.irq_chip_init) {
1448                 /*
1449                  * First time we setup an MSI IRQ, we need to setup the
1450                  * corresponding IRQ chip to route correctly.
1451                  */
1452                 idata = irq_get_irq_data(virq);
1453                 ichip = irq_data_get_irq_chip(idata);
1454                 phb->ioda.irq_chip_init = 1;
1455                 phb->ioda.irq_chip = *ichip;
1456                 phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi;
1457         }
1458         irq_set_chip(virq, &phb->ioda.irq_chip);
1459 }
1460
1461 #ifdef CONFIG_CXL_BASE
1462
1463 struct device_node *pnv_pci_to_phb_node(struct pci_dev *dev)
1464 {
1465         struct pci_controller *hose = pci_bus_to_host(dev->bus);
1466
1467         return hose->dn;
1468 }
1469 EXPORT_SYMBOL(pnv_pci_to_phb_node);
1470
1471 int pnv_phb_to_cxl(struct pci_dev *dev)
1472 {
1473         struct pci_controller *hose = pci_bus_to_host(dev->bus);
1474         struct pnv_phb *phb = hose->private_data;
1475         struct pnv_ioda_pe *pe;
1476         int rc;
1477
1478         pe = pnv_ioda_get_pe(dev);
1479         if (!pe)
1480                 return -ENODEV;
1481
1482         pe_info(pe, "Switching PHB to CXL\n");
1483
1484         rc = opal_pci_set_phb_cxl_mode(phb->opal_id, 1, pe->pe_number);
1485         if (rc)
1486                 dev_err(&dev->dev, "opal_pci_set_phb_cxl_mode failed: %i\n", rc);
1487
1488         return rc;
1489 }
1490 EXPORT_SYMBOL(pnv_phb_to_cxl);
1491
1492 /* Find PHB for cxl dev and allocate MSI hwirqs?
1493  * Returns the absolute hardware IRQ number
1494  */
1495 int pnv_cxl_alloc_hwirqs(struct pci_dev *dev, int num)
1496 {
1497         struct pci_controller *hose = pci_bus_to_host(dev->bus);
1498         struct pnv_phb *phb = hose->private_data;
1499         int hwirq = msi_bitmap_alloc_hwirqs(&phb->msi_bmp, num);
1500
1501         if (hwirq < 0) {
1502                 dev_warn(&dev->dev, "Failed to find a free MSI\n");
1503                 return -ENOSPC;
1504         }
1505
1506         return phb->msi_base + hwirq;
1507 }
1508 EXPORT_SYMBOL(pnv_cxl_alloc_hwirqs);
1509
1510 void pnv_cxl_release_hwirqs(struct pci_dev *dev, int hwirq, int num)
1511 {
1512         struct pci_controller *hose = pci_bus_to_host(dev->bus);
1513         struct pnv_phb *phb = hose->private_data;
1514
1515         msi_bitmap_free_hwirqs(&phb->msi_bmp, hwirq - phb->msi_base, num);
1516 }
1517 EXPORT_SYMBOL(pnv_cxl_release_hwirqs);
1518
1519 void pnv_cxl_release_hwirq_ranges(struct cxl_irq_ranges *irqs,
1520                                   struct pci_dev *dev)
1521 {
1522         struct pci_controller *hose = pci_bus_to_host(dev->bus);
1523         struct pnv_phb *phb = hose->private_data;
1524         int i, hwirq;
1525
1526         for (i = 1; i < CXL_IRQ_RANGES; i++) {
1527                 if (!irqs->range[i])
1528                         continue;
1529                 pr_devel("cxl release irq range 0x%x: offset: 0x%lx  limit: %ld\n",
1530                          i, irqs->offset[i],
1531                          irqs->range[i]);
1532                 hwirq = irqs->offset[i] - phb->msi_base;
1533                 msi_bitmap_free_hwirqs(&phb->msi_bmp, hwirq,
1534                                        irqs->range[i]);
1535         }
1536 }
1537 EXPORT_SYMBOL(pnv_cxl_release_hwirq_ranges);
1538
1539 int pnv_cxl_alloc_hwirq_ranges(struct cxl_irq_ranges *irqs,
1540                                struct pci_dev *dev, int num)
1541 {
1542         struct pci_controller *hose = pci_bus_to_host(dev->bus);
1543         struct pnv_phb *phb = hose->private_data;
1544         int i, hwirq, try;
1545
1546         memset(irqs, 0, sizeof(struct cxl_irq_ranges));
1547
1548         /* 0 is reserved for the multiplexed PSL DSI interrupt */
1549         for (i = 1; i < CXL_IRQ_RANGES && num; i++) {
1550                 try = num;
1551                 while (try) {
1552                         hwirq = msi_bitmap_alloc_hwirqs(&phb->msi_bmp, try);
1553                         if (hwirq >= 0)
1554                                 break;
1555                         try /= 2;
1556                 }
1557                 if (!try)
1558                         goto fail;
1559
1560                 irqs->offset[i] = phb->msi_base + hwirq;
1561                 irqs->range[i] = try;
1562                 pr_devel("cxl alloc irq range 0x%x: offset: 0x%lx  limit: %li\n",
1563                          i, irqs->offset[i], irqs->range[i]);
1564                 num -= try;
1565         }
1566         if (num)
1567                 goto fail;
1568
1569         return 0;
1570 fail:
1571         pnv_cxl_release_hwirq_ranges(irqs, dev);
1572         return -ENOSPC;
1573 }
1574 EXPORT_SYMBOL(pnv_cxl_alloc_hwirq_ranges);
1575
1576 int pnv_cxl_get_irq_count(struct pci_dev *dev)
1577 {
1578         struct pci_controller *hose = pci_bus_to_host(dev->bus);
1579         struct pnv_phb *phb = hose->private_data;
1580
1581         return phb->msi_bmp.irq_count;
1582 }
1583 EXPORT_SYMBOL(pnv_cxl_get_irq_count);
1584
1585 int pnv_cxl_ioda_msi_setup(struct pci_dev *dev, unsigned int hwirq,
1586                            unsigned int virq)
1587 {
1588         struct pci_controller *hose = pci_bus_to_host(dev->bus);
1589         struct pnv_phb *phb = hose->private_data;
1590         unsigned int xive_num = hwirq - phb->msi_base;
1591         struct pnv_ioda_pe *pe;
1592         int rc;
1593
1594         if (!(pe = pnv_ioda_get_pe(dev)))
1595                 return -ENODEV;
1596
1597         /* Assign XIVE to PE */
1598         rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
1599         if (rc) {
1600                 pe_warn(pe, "%s: OPAL error %d setting msi_base 0x%x "
1601                         "hwirq 0x%x XIVE 0x%x PE\n",
1602                         pci_name(dev), rc, phb->msi_base, hwirq, xive_num);
1603                 return -EIO;
1604         }
1605         set_msi_irq_chip(phb, virq);
1606
1607         return 0;
1608 }
1609 EXPORT_SYMBOL(pnv_cxl_ioda_msi_setup);
1610 #endif
1611
1612 static int pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev,
1613                                   unsigned int hwirq, unsigned int virq,
1614                                   unsigned int is_64, struct msi_msg *msg)
1615 {
1616         struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev);
1617         struct pci_dn *pdn = pci_get_pdn(dev);
1618         unsigned int xive_num = hwirq - phb->msi_base;
1619         __be32 data;
1620         int rc;
1621
1622         /* No PE assigned ? bail out ... no MSI for you ! */
1623         if (pe == NULL)
1624                 return -ENXIO;
1625
1626         /* Check if we have an MVE */
1627         if (pe->mve_number < 0)
1628                 return -ENXIO;
1629
1630         /* Force 32-bit MSI on some broken devices */
1631         if (pdn && pdn->force_32bit_msi)
1632                 is_64 = 0;
1633
1634         /* Assign XIVE to PE */
1635         rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
1636         if (rc) {
1637                 pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
1638                         pci_name(dev), rc, xive_num);
1639                 return -EIO;
1640         }
1641
1642         if (is_64) {
1643                 __be64 addr64;
1644
1645                 rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1,
1646                                      &addr64, &data);
1647                 if (rc) {
1648                         pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
1649                                 pci_name(dev), rc);
1650                         return -EIO;
1651                 }
1652                 msg->address_hi = be64_to_cpu(addr64) >> 32;
1653                 msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful;
1654         } else {
1655                 __be32 addr32;
1656
1657                 rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1,
1658                                      &addr32, &data);
1659                 if (rc) {
1660                         pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
1661                                 pci_name(dev), rc);
1662                         return -EIO;
1663                 }
1664                 msg->address_hi = 0;
1665                 msg->address_lo = be32_to_cpu(addr32);
1666         }
1667         msg->data = be32_to_cpu(data);
1668
1669         set_msi_irq_chip(phb, virq);
1670
1671         pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d),"
1672                  " address=%x_%08x data=%x PE# %d\n",
1673                  pci_name(dev), is_64 ? "64" : "32", hwirq, xive_num,
1674                  msg->address_hi, msg->address_lo, data, pe->pe_number);
1675
1676         return 0;
1677 }
1678
1679 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb)
1680 {
1681         unsigned int count;
1682         const __be32 *prop = of_get_property(phb->hose->dn,
1683                                              "ibm,opal-msi-ranges", NULL);
1684         if (!prop) {
1685                 /* BML Fallback */
1686                 prop = of_get_property(phb->hose->dn, "msi-ranges", NULL);
1687         }
1688         if (!prop)
1689                 return;
1690
1691         phb->msi_base = be32_to_cpup(prop);
1692         count = be32_to_cpup(prop + 1);
1693         if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) {
1694                 pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
1695                        phb->hose->global_number);
1696                 return;
1697         }
1698
1699         phb->msi_setup = pnv_pci_ioda_msi_setup;
1700         phb->msi32_support = 1;
1701         pr_info("  Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
1702                 count, phb->msi_base);
1703 }
1704 #else
1705 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb) { }
1706 #endif /* CONFIG_PCI_MSI */
1707
1708 /*
1709  * This function is supposed to be called on basis of PE from top
1710  * to bottom style. So the the I/O or MMIO segment assigned to
1711  * parent PE could be overrided by its child PEs if necessary.
1712  */
1713 static void pnv_ioda_setup_pe_seg(struct pci_controller *hose,
1714                                   struct pnv_ioda_pe *pe)
1715 {
1716         struct pnv_phb *phb = hose->private_data;
1717         struct pci_bus_region region;
1718         struct resource *res;
1719         int i, index;
1720         int rc;
1721
1722         /*
1723          * NOTE: We only care PCI bus based PE for now. For PCI
1724          * device based PE, for example SRIOV sensitive VF should
1725          * be figured out later.
1726          */
1727         BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)));
1728
1729         pci_bus_for_each_resource(pe->pbus, res, i) {
1730                 if (!res || !res->flags ||
1731                     res->start > res->end)
1732                         continue;
1733
1734                 if (res->flags & IORESOURCE_IO) {
1735                         region.start = res->start - phb->ioda.io_pci_base;
1736                         region.end   = res->end - phb->ioda.io_pci_base;
1737                         index = region.start / phb->ioda.io_segsize;
1738
1739                         while (index < phb->ioda.total_pe &&
1740                                region.start <= region.end) {
1741                                 phb->ioda.io_segmap[index] = pe->pe_number;
1742                                 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
1743                                         pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index);
1744                                 if (rc != OPAL_SUCCESS) {
1745                                         pr_err("%s: OPAL error %d when mapping IO "
1746                                                "segment #%d to PE#%d\n",
1747                                                __func__, rc, index, pe->pe_number);
1748                                         break;
1749                                 }
1750
1751                                 region.start += phb->ioda.io_segsize;
1752                                 index++;
1753                         }
1754                 } else if (res->flags & IORESOURCE_MEM) {
1755                         region.start = res->start -
1756                                        hose->mem_offset[0] -
1757                                        phb->ioda.m32_pci_base;
1758                         region.end   = res->end -
1759                                        hose->mem_offset[0] -
1760                                        phb->ioda.m32_pci_base;
1761                         index = region.start / phb->ioda.m32_segsize;
1762
1763                         while (index < phb->ioda.total_pe &&
1764                                region.start <= region.end) {
1765                                 phb->ioda.m32_segmap[index] = pe->pe_number;
1766                                 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
1767                                         pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index);
1768                                 if (rc != OPAL_SUCCESS) {
1769                                         pr_err("%s: OPAL error %d when mapping M32 "
1770                                                "segment#%d to PE#%d",
1771                                                __func__, rc, index, pe->pe_number);
1772                                         break;
1773                                 }
1774
1775                                 region.start += phb->ioda.m32_segsize;
1776                                 index++;
1777                         }
1778                 }
1779         }
1780 }
1781
1782 static void pnv_pci_ioda_setup_seg(void)
1783 {
1784         struct pci_controller *tmp, *hose;
1785         struct pnv_phb *phb;
1786         struct pnv_ioda_pe *pe;
1787
1788         list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1789                 phb = hose->private_data;
1790                 list_for_each_entry(pe, &phb->ioda.pe_list, list) {
1791                         pnv_ioda_setup_pe_seg(hose, pe);
1792                 }
1793         }
1794 }
1795
1796 static void pnv_pci_ioda_setup_DMA(void)
1797 {
1798         struct pci_controller *hose, *tmp;
1799         struct pnv_phb *phb;
1800
1801         list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1802                 pnv_ioda_setup_dma(hose->private_data);
1803
1804                 /* Mark the PHB initialization done */
1805                 phb = hose->private_data;
1806                 phb->initialized = 1;
1807         }
1808 }
1809
1810 static void pnv_pci_ioda_create_dbgfs(void)
1811 {
1812 #ifdef CONFIG_DEBUG_FS
1813         struct pci_controller *hose, *tmp;
1814         struct pnv_phb *phb;
1815         char name[16];
1816
1817         list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1818                 phb = hose->private_data;
1819
1820                 sprintf(name, "PCI%04x", hose->global_number);
1821                 phb->dbgfs = debugfs_create_dir(name, powerpc_debugfs_root);
1822                 if (!phb->dbgfs)
1823                         pr_warning("%s: Error on creating debugfs on PHB#%x\n",
1824                                 __func__, hose->global_number);
1825         }
1826 #endif /* CONFIG_DEBUG_FS */
1827 }
1828
1829 static void pnv_pci_ioda_fixup(void)
1830 {
1831         pnv_pci_ioda_setup_PEs();
1832         pnv_pci_ioda_setup_seg();
1833         pnv_pci_ioda_setup_DMA();
1834
1835         pnv_pci_ioda_create_dbgfs();
1836
1837 #ifdef CONFIG_EEH
1838         eeh_init();
1839         eeh_addr_cache_build();
1840 #endif
1841 }
1842
1843 /*
1844  * Returns the alignment for I/O or memory windows for P2P
1845  * bridges. That actually depends on how PEs are segmented.
1846  * For now, we return I/O or M32 segment size for PE sensitive
1847  * P2P bridges. Otherwise, the default values (4KiB for I/O,
1848  * 1MiB for memory) will be returned.
1849  *
1850  * The current PCI bus might be put into one PE, which was
1851  * create against the parent PCI bridge. For that case, we
1852  * needn't enlarge the alignment so that we can save some
1853  * resources.
1854  */
1855 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus,
1856                                                 unsigned long type)
1857 {
1858         struct pci_dev *bridge;
1859         struct pci_controller *hose = pci_bus_to_host(bus);
1860         struct pnv_phb *phb = hose->private_data;
1861         int num_pci_bridges = 0;
1862
1863         bridge = bus->self;
1864         while (bridge) {
1865                 if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) {
1866                         num_pci_bridges++;
1867                         if (num_pci_bridges >= 2)
1868                                 return 1;
1869                 }
1870
1871                 bridge = bridge->bus->self;
1872         }
1873
1874         /* We fail back to M32 if M64 isn't supported */
1875         if (phb->ioda.m64_segsize &&
1876             pnv_pci_is_mem_pref_64(type))
1877                 return phb->ioda.m64_segsize;
1878         if (type & IORESOURCE_MEM)
1879                 return phb->ioda.m32_segsize;
1880
1881         return phb->ioda.io_segsize;
1882 }
1883
1884 /* Prevent enabling devices for which we couldn't properly
1885  * assign a PE
1886  */
1887 static int pnv_pci_enable_device_hook(struct pci_dev *dev)
1888 {
1889         struct pci_controller *hose = pci_bus_to_host(dev->bus);
1890         struct pnv_phb *phb = hose->private_data;
1891         struct pci_dn *pdn;
1892
1893         /* The function is probably called while the PEs have
1894          * not be created yet. For example, resource reassignment
1895          * during PCI probe period. We just skip the check if
1896          * PEs isn't ready.
1897          */
1898         if (!phb->initialized)
1899                 return 0;
1900
1901         pdn = pci_get_pdn(dev);
1902         if (!pdn || pdn->pe_number == IODA_INVALID_PE)
1903                 return -EINVAL;
1904
1905         return 0;
1906 }
1907
1908 static u32 pnv_ioda_bdfn_to_pe(struct pnv_phb *phb, struct pci_bus *bus,
1909                                u32 devfn)
1910 {
1911         return phb->ioda.pe_rmap[(bus->number << 8) | devfn];
1912 }
1913
1914 static void pnv_pci_ioda_shutdown(struct pnv_phb *phb)
1915 {
1916         opal_pci_reset(phb->opal_id, OPAL_RESET_PCI_IODA_TABLE,
1917                        OPAL_ASSERT_RESET);
1918 }
1919
1920 static void __init pnv_pci_init_ioda_phb(struct device_node *np,
1921                                          u64 hub_id, int ioda_type)
1922 {
1923         struct pci_controller *hose;
1924         struct pnv_phb *phb;
1925         unsigned long size, m32map_off, pemap_off, iomap_off = 0;
1926         const __be64 *prop64;
1927         const __be32 *prop32;
1928         int len;
1929         u64 phb_id;
1930         void *aux;
1931         long rc;
1932
1933         pr_info("Initializing IODA%d OPAL PHB %s\n", ioda_type, np->full_name);
1934
1935         prop64 = of_get_property(np, "ibm,opal-phbid", NULL);
1936         if (!prop64) {
1937                 pr_err("  Missing \"ibm,opal-phbid\" property !\n");
1938                 return;
1939         }
1940         phb_id = be64_to_cpup(prop64);
1941         pr_debug("  PHB-ID  : 0x%016llx\n", phb_id);
1942
1943         phb = memblock_virt_alloc(sizeof(struct pnv_phb), 0);
1944
1945         /* Allocate PCI controller */
1946         phb->hose = hose = pcibios_alloc_controller(np);
1947         if (!phb->hose) {
1948                 pr_err("  Can't allocate PCI controller for %s\n",
1949                        np->full_name);
1950                 memblock_free(__pa(phb), sizeof(struct pnv_phb));
1951                 return;
1952         }
1953
1954         spin_lock_init(&phb->lock);
1955         prop32 = of_get_property(np, "bus-range", &len);
1956         if (prop32 && len == 8) {
1957                 hose->first_busno = be32_to_cpu(prop32[0]);
1958                 hose->last_busno = be32_to_cpu(prop32[1]);
1959         } else {
1960                 pr_warn("  Broken <bus-range> on %s\n", np->full_name);
1961                 hose->first_busno = 0;
1962                 hose->last_busno = 0xff;
1963         }
1964         hose->private_data = phb;
1965         phb->hub_id = hub_id;
1966         phb->opal_id = phb_id;
1967         phb->type = ioda_type;
1968
1969         /* Detect specific models for error handling */
1970         if (of_device_is_compatible(np, "ibm,p7ioc-pciex"))
1971                 phb->model = PNV_PHB_MODEL_P7IOC;
1972         else if (of_device_is_compatible(np, "ibm,power8-pciex"))
1973                 phb->model = PNV_PHB_MODEL_PHB3;
1974         else
1975                 phb->model = PNV_PHB_MODEL_UNKNOWN;
1976
1977         /* Parse 32-bit and IO ranges (if any) */
1978         pci_process_bridge_OF_ranges(hose, np, !hose->global_number);
1979
1980         /* Get registers */
1981         phb->regs = of_iomap(np, 0);
1982         if (phb->regs == NULL)
1983                 pr_err("  Failed to map registers !\n");
1984
1985         /* Initialize more IODA stuff */
1986         phb->ioda.total_pe = 1;
1987         prop32 = of_get_property(np, "ibm,opal-num-pes", NULL);
1988         if (prop32)
1989                 phb->ioda.total_pe = be32_to_cpup(prop32);
1990         prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL);
1991         if (prop32)
1992                 phb->ioda.reserved_pe = be32_to_cpup(prop32);
1993
1994         /* Parse 64-bit MMIO range */
1995         pnv_ioda_parse_m64_window(phb);
1996
1997         phb->ioda.m32_size = resource_size(&hose->mem_resources[0]);
1998         /* FW Has already off top 64k of M32 space (MSI space) */
1999         phb->ioda.m32_size += 0x10000;
2000
2001         phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe;
2002         phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0];
2003         phb->ioda.io_size = hose->pci_io_size;
2004         phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe;
2005         phb->ioda.io_pci_base = 0; /* XXX calculate this ? */
2006
2007         /* Allocate aux data & arrays. We don't have IO ports on PHB3 */
2008         size = _ALIGN_UP(phb->ioda.total_pe / 8, sizeof(unsigned long));
2009         m32map_off = size;
2010         size += phb->ioda.total_pe * sizeof(phb->ioda.m32_segmap[0]);
2011         if (phb->type == PNV_PHB_IODA1) {
2012                 iomap_off = size;
2013                 size += phb->ioda.total_pe * sizeof(phb->ioda.io_segmap[0]);
2014         }
2015         pemap_off = size;
2016         size += phb->ioda.total_pe * sizeof(struct pnv_ioda_pe);
2017         aux = memblock_virt_alloc(size, 0);
2018         phb->ioda.pe_alloc = aux;
2019         phb->ioda.m32_segmap = aux + m32map_off;
2020         if (phb->type == PNV_PHB_IODA1)
2021                 phb->ioda.io_segmap = aux + iomap_off;
2022         phb->ioda.pe_array = aux + pemap_off;
2023         set_bit(phb->ioda.reserved_pe, phb->ioda.pe_alloc);
2024
2025         INIT_LIST_HEAD(&phb->ioda.pe_dma_list);
2026         INIT_LIST_HEAD(&phb->ioda.pe_list);
2027
2028         /* Calculate how many 32-bit TCE segments we have */
2029         phb->ioda.tce32_count = phb->ioda.m32_pci_base >> 28;
2030
2031 #if 0 /* We should really do that ... */
2032         rc = opal_pci_set_phb_mem_window(opal->phb_id,
2033                                          window_type,
2034                                          window_num,
2035                                          starting_real_address,
2036                                          starting_pci_address,
2037                                          segment_size);
2038 #endif
2039
2040         pr_info("  %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
2041                 phb->ioda.total_pe, phb->ioda.reserved_pe,
2042                 phb->ioda.m32_size, phb->ioda.m32_segsize);
2043         if (phb->ioda.m64_size)
2044                 pr_info("                 M64: 0x%lx [segment=0x%lx]\n",
2045                         phb->ioda.m64_size, phb->ioda.m64_segsize);
2046         if (phb->ioda.io_size)
2047                 pr_info("                  IO: 0x%x [segment=0x%x]\n",
2048                         phb->ioda.io_size, phb->ioda.io_segsize);
2049
2050
2051         phb->hose->ops = &pnv_pci_ops;
2052         phb->get_pe_state = pnv_ioda_get_pe_state;
2053         phb->freeze_pe = pnv_ioda_freeze_pe;
2054         phb->unfreeze_pe = pnv_ioda_unfreeze_pe;
2055 #ifdef CONFIG_EEH
2056         phb->eeh_ops = &ioda_eeh_ops;
2057 #endif
2058
2059         /* Setup RID -> PE mapping function */
2060         phb->bdfn_to_pe = pnv_ioda_bdfn_to_pe;
2061
2062         /* Setup TCEs */
2063         phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup;
2064         phb->dma_set_mask = pnv_pci_ioda_dma_set_mask;
2065         phb->dma_get_required_mask = pnv_pci_ioda_dma_get_required_mask;
2066
2067         /* Setup shutdown function for kexec */
2068         phb->shutdown = pnv_pci_ioda_shutdown;
2069
2070         /* Setup MSI support */
2071         pnv_pci_init_ioda_msis(phb);
2072
2073         /*
2074          * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
2075          * to let the PCI core do resource assignment. It's supposed
2076          * that the PCI core will do correct I/O and MMIO alignment
2077          * for the P2P bridge bars so that each PCI bus (excluding
2078          * the child P2P bridges) can form individual PE.
2079          */
2080         ppc_md.pcibios_fixup = pnv_pci_ioda_fixup;
2081         ppc_md.pcibios_enable_device_hook = pnv_pci_enable_device_hook;
2082         ppc_md.pcibios_window_alignment = pnv_pci_window_alignment;
2083         ppc_md.pcibios_reset_secondary_bus = pnv_pci_reset_secondary_bus;
2084         pci_add_flags(PCI_REASSIGN_ALL_RSRC);
2085
2086         /* Reset IODA tables to a clean state */
2087         rc = opal_pci_reset(phb_id, OPAL_RESET_PCI_IODA_TABLE, OPAL_ASSERT_RESET);
2088         if (rc)
2089                 pr_warning("  OPAL Error %ld performing IODA table reset !\n", rc);
2090
2091         /* If we're running in kdump kerenl, the previous kerenl never
2092          * shutdown PCI devices correctly. We already got IODA table
2093          * cleaned out. So we have to issue PHB reset to stop all PCI
2094          * transactions from previous kerenl.
2095          */
2096         if (is_kdump_kernel()) {
2097                 pr_info("  Issue PHB reset ...\n");
2098                 ioda_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL);
2099                 ioda_eeh_phb_reset(hose, OPAL_DEASSERT_RESET);
2100         }
2101
2102         /* Remove M64 resource if we can't configure it successfully */
2103         if (!phb->init_m64 || phb->init_m64(phb))
2104                 hose->mem_resources[1].flags = 0;
2105 }
2106
2107 void __init pnv_pci_init_ioda2_phb(struct device_node *np)
2108 {
2109         pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2);
2110 }
2111
2112 void __init pnv_pci_init_ioda_hub(struct device_node *np)
2113 {
2114         struct device_node *phbn;
2115         const __be64 *prop64;
2116         u64 hub_id;
2117
2118         pr_info("Probing IODA IO-Hub %s\n", np->full_name);
2119
2120         prop64 = of_get_property(np, "ibm,opal-hubid", NULL);
2121         if (!prop64) {
2122                 pr_err(" Missing \"ibm,opal-hubid\" property !\n");
2123                 return;
2124         }
2125         hub_id = be64_to_cpup(prop64);
2126         pr_devel(" HUB-ID : 0x%016llx\n", hub_id);
2127
2128         /* Count child PHBs */
2129         for_each_child_of_node(np, phbn) {
2130                 /* Look for IODA1 PHBs */
2131                 if (of_device_is_compatible(phbn, "ibm,ioda-phb"))
2132                         pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1);
2133         }
2134 }