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

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Support PCI/PCIe on PowerNV platforms
   4  *
   5  * Copyright 2011 Benjamin Herrenschmidt, IBM Corp.
   6  */
   7
   8 #undef DEBUG
   9
  10 #include <linux/kernel.h>
  11 #include <linux/pci.h>
  12 #include <linux/crash_dump.h>
  13 #include <linux/delay.h>
  14 #include <linux/string.h>
  15 #include <linux/init.h>
  16 #include <linux/memblock.h>
  17 #include <linux/irq.h>
  18 #include <linux/io.h>
  19 #include <linux/msi.h>
  20 #include <linux/iommu.h>
  21 #include <linux/rculist.h>
  22 #include <linux/sizes.h>
  23 #include <linux/debugfs.h>
  24
  25 #include <asm/sections.h>
  26 #include <asm/io.h>
  27 #include <asm/prom.h>
  28 #include <asm/pci-bridge.h>
  29 #include <asm/machdep.h>
  30 #include <asm/msi_bitmap.h>
  31 #include <asm/ppc-pci.h>
  32 #include <asm/opal.h>
  33 #include <asm/iommu.h>
  34 #include <asm/tce.h>
  35 #include <asm/xics.h>
  36 #include <asm/firmware.h>
  37 #include <asm/pnv-pci.h>
  38 #include <asm/mmzone.h>
  39 #include <asm/xive.h>
  40
  41 #include <misc/cxl-base.h>
  42
  43 #include "powernv.h"
  44 #include "pci.h"
  45 #include "../../../../drivers/pci/pci.h"
  46
  47 #define PNV_IODA1_M64_NUM       16      /* Number of M64 BARs   */
  48 #define PNV_IODA1_M64_SEGS      8       /* Segments per M64 BAR */
  49 #define PNV_IODA1_DMA32_SEGSIZE 0x10000000
  50
  51 static const char * const pnv_phb_names[] = { "IODA1", "IODA2", "NPU_OCAPI" };
  52
  53 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable);
  54 static void pnv_pci_configure_bus(struct pci_bus *bus);
  55
  56 void pe_level_printk(const struct pnv_ioda_pe *pe, const char *level,
  57                             const char *fmt, ...)
  58 {
  59         struct va_format vaf;
  60         va_list args;
  61         char pfix[32];
  62
  63         va_start(args, fmt);
  64
  65         vaf.fmt = fmt;
  66         vaf.va = &args;
  67
  68         if (pe->flags & PNV_IODA_PE_DEV)
  69                 strlcpy(pfix, dev_name(&pe->pdev->dev), sizeof(pfix));
  70         else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
  71                 sprintf(pfix, "%04x:%02x     ",
  72                         pci_domain_nr(pe->pbus), pe->pbus->number);
  73 #ifdef CONFIG_PCI_IOV
  74         else if (pe->flags & PNV_IODA_PE_VF)
  75                 sprintf(pfix, "%04x:%02x:%2x.%d",
  76                         pci_domain_nr(pe->parent_dev->bus),
  77                         (pe->rid & 0xff00) >> 8,
  78                         PCI_SLOT(pe->rid), PCI_FUNC(pe->rid));
  79 #endif /* CONFIG_PCI_IOV*/
  80
  81         printk("%spci %s: [PE# %.2x] %pV",
  82                level, pfix, pe->pe_number, &vaf);
  83
  84         va_end(args);
  85 }
  86
  87 static bool pnv_iommu_bypass_disabled __read_mostly;
  88 static bool pci_reset_phbs __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 static int __init pci_reset_phbs_setup(char *str)
 111 {
 112         pci_reset_phbs = true;
 113         return 0;
 114 }
 115
 116 early_param("ppc_pci_reset_phbs", pci_reset_phbs_setup);
 117
 118 static struct pnv_ioda_pe *pnv_ioda_init_pe(struct pnv_phb *phb, int pe_no)
 119 {
 120         s64 rc;
 121
 122         phb->ioda.pe_array[pe_no].phb = phb;
 123         phb->ioda.pe_array[pe_no].pe_number = pe_no;
 124         phb->ioda.pe_array[pe_no].dma_setup_done = false;
 125
 126         /*
 127          * Clear the PE frozen state as it might be put into frozen state
 128          * in the last PCI remove path. It's not harmful to do so when the
 129          * PE is already in unfrozen state.
 130          */
 131         rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no,
 132                                        OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 133         if (rc != OPAL_SUCCESS && rc != OPAL_UNSUPPORTED)
 134                 pr_warn("%s: Error %lld unfreezing PHB#%x-PE#%x\n",
 135                         __func__, rc, phb->hose->global_number, pe_no);
 136
 137         return &phb->ioda.pe_array[pe_no];
 138 }
 139
 140 static void pnv_ioda_reserve_pe(struct pnv_phb *phb, int pe_no)
 141 {
 142         if (!(pe_no >= 0 && pe_no < phb->ioda.total_pe_num)) {
 143                 pr_warn("%s: Invalid PE %x on PHB#%x\n",
 144                         __func__, pe_no, phb->hose->global_number);
 145                 return;
 146         }
 147
 148         mutex_lock(&phb->ioda.pe_alloc_mutex);
 149         if (test_and_set_bit(pe_no, phb->ioda.pe_alloc))
 150                 pr_debug("%s: PE %x was reserved on PHB#%x\n",
 151                          __func__, pe_no, phb->hose->global_number);
 152         mutex_unlock(&phb->ioda.pe_alloc_mutex);
 153
 154         pnv_ioda_init_pe(phb, pe_no);
 155 }
 156
 157 struct pnv_ioda_pe *pnv_ioda_alloc_pe(struct pnv_phb *phb, int count)
 158 {
 159         struct pnv_ioda_pe *ret = NULL;
 160         int run = 0, pe, i;
 161
 162         mutex_lock(&phb->ioda.pe_alloc_mutex);
 163
 164         /* scan backwards for a run of @count cleared bits */
 165         for (pe = phb->ioda.total_pe_num - 1; pe >= 0; pe--) {
 166                 if (test_bit(pe, phb->ioda.pe_alloc)) {
 167                         run = 0;
 168                         continue;
 169                 }
 170
 171                 run++;
 172                 if (run == count)
 173                         break;
 174         }
 175         if (run != count)
 176                 goto out;
 177
 178         for (i = pe; i < pe + count; i++) {
 179                 set_bit(i, phb->ioda.pe_alloc);
 180                 pnv_ioda_init_pe(phb, i);
 181         }
 182         ret = &phb->ioda.pe_array[pe];
 183
 184 out:
 185         mutex_unlock(&phb->ioda.pe_alloc_mutex);
 186         return ret;
 187 }
 188
 189 void pnv_ioda_free_pe(struct pnv_ioda_pe *pe)
 190 {
 191         struct pnv_phb *phb = pe->phb;
 192         unsigned int pe_num = pe->pe_number;
 193
 194         WARN_ON(pe->pdev);
 195         memset(pe, 0, sizeof(struct pnv_ioda_pe));
 196
 197         mutex_lock(&phb->ioda.pe_alloc_mutex);
 198         clear_bit(pe_num, phb->ioda.pe_alloc);
 199         mutex_unlock(&phb->ioda.pe_alloc_mutex);
 200 }
 201
 202 /* The default M64 BAR is shared by all PEs */
 203 static int pnv_ioda2_init_m64(struct pnv_phb *phb)
 204 {
 205         const char *desc;
 206         struct resource *r;
 207         s64 rc;
 208
 209         /* Configure the default M64 BAR */
 210         rc = opal_pci_set_phb_mem_window(phb->opal_id,
 211                                          OPAL_M64_WINDOW_TYPE,
 212                                          phb->ioda.m64_bar_idx,
 213                                          phb->ioda.m64_base,
 214                                          0, /* unused */
 215                                          phb->ioda.m64_size);
 216         if (rc != OPAL_SUCCESS) {
 217                 desc = "configuring";
 218                 goto fail;
 219         }
 220
 221         /* Enable the default M64 BAR */
 222         rc = opal_pci_phb_mmio_enable(phb->opal_id,
 223                                       OPAL_M64_WINDOW_TYPE,
 224                                       phb->ioda.m64_bar_idx,
 225                                       OPAL_ENABLE_M64_SPLIT);
 226         if (rc != OPAL_SUCCESS) {
 227                 desc = "enabling";
 228                 goto fail;
 229         }
 230
 231         /*
 232          * Exclude the segments for reserved and root bus PE, which
 233          * are first or last two PEs.
 234          */
 235         r = &phb->hose->mem_resources[1];
 236         if (phb->ioda.reserved_pe_idx == 0)
 237                 r->start += (2 * phb->ioda.m64_segsize);
 238         else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
 239                 r->end -= (2 * phb->ioda.m64_segsize);
 240         else
 241                 pr_warn("  Cannot strip M64 segment for reserved PE#%x\n",
 242                         phb->ioda.reserved_pe_idx);
 243
 244         return 0;
 245
 246 fail:
 247         pr_warn("  Failure %lld %s M64 BAR#%d\n",
 248                 rc, desc, phb->ioda.m64_bar_idx);
 249         opal_pci_phb_mmio_enable(phb->opal_id,
 250                                  OPAL_M64_WINDOW_TYPE,
 251                                  phb->ioda.m64_bar_idx,
 252                                  OPAL_DISABLE_M64);
 253         return -EIO;
 254 }
 255
 256 static void pnv_ioda_reserve_dev_m64_pe(struct pci_dev *pdev,
 257                                          unsigned long *pe_bitmap)
 258 {
 259         struct pnv_phb *phb = pci_bus_to_pnvhb(pdev->bus);
 260         struct resource *r;
 261         resource_size_t base, sgsz, start, end;
 262         int segno, i;
 263
 264         base = phb->ioda.m64_base;
 265         sgsz = phb->ioda.m64_segsize;
 266         for (i = 0; i <= PCI_ROM_RESOURCE; i++) {
 267                 r = &pdev->resource[i];
 268                 if (!r->parent || !pnv_pci_is_m64(phb, r))
 269                         continue;
 270
 271                 start = ALIGN_DOWN(r->start - base, sgsz);
 272                 end = ALIGN(r->end - base, sgsz);
 273                 for (segno = start / sgsz; segno < end / sgsz; segno++) {
 274                         if (pe_bitmap)
 275                                 set_bit(segno, pe_bitmap);
 276                         else
 277                                 pnv_ioda_reserve_pe(phb, segno);
 278                 }
 279         }
 280 }
 281
 282 static int pnv_ioda1_init_m64(struct pnv_phb *phb)
 283 {
 284         struct resource *r;
 285         int index;
 286
 287         /*
 288          * There are 16 M64 BARs, each of which has 8 segments. So
 289          * there are as many M64 segments as the maximum number of
 290          * PEs, which is 128.
 291          */
 292         for (index = 0; index < PNV_IODA1_M64_NUM; index++) {
 293                 unsigned long base, segsz = phb->ioda.m64_segsize;
 294                 int64_t rc;
 295
 296                 base = phb->ioda.m64_base +
 297                        index * PNV_IODA1_M64_SEGS * segsz;
 298                 rc = opal_pci_set_phb_mem_window(phb->opal_id,
 299                                 OPAL_M64_WINDOW_TYPE, index, base, 0,
 300                                 PNV_IODA1_M64_SEGS * segsz);
 301                 if (rc != OPAL_SUCCESS) {
 302                         pr_warn("  Error %lld setting M64 PHB#%x-BAR#%d\n",
 303                                 rc, phb->hose->global_number, index);
 304                         goto fail;
 305                 }
 306
 307                 rc = opal_pci_phb_mmio_enable(phb->opal_id,
 308                                 OPAL_M64_WINDOW_TYPE, index,
 309                                 OPAL_ENABLE_M64_SPLIT);
 310                 if (rc != OPAL_SUCCESS) {
 311                         pr_warn("  Error %lld enabling M64 PHB#%x-BAR#%d\n",
 312                                 rc, phb->hose->global_number, index);
 313                         goto fail;
 314                 }
 315         }
 316
 317         for (index = 0; index < phb->ioda.total_pe_num; index++) {
 318                 int64_t rc;
 319
 320                 /*
 321                  * P7IOC supports M64DT, which helps mapping M64 segment
 322                  * to one particular PE#. However, PHB3 has fixed mapping
 323                  * between M64 segment and PE#. In order to have same logic
 324                  * for P7IOC and PHB3, we enforce fixed mapping between M64
 325                  * segment and PE# on P7IOC.
 326                  */
 327                 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
 328                                 index, OPAL_M64_WINDOW_TYPE,
 329                                 index / PNV_IODA1_M64_SEGS,
 330                                 index % PNV_IODA1_M64_SEGS);
 331                 if (rc != OPAL_SUCCESS) {
 332                         pr_warn("%s: Error %lld mapping M64 for PHB#%x-PE#%x\n",
 333                                 __func__, rc, phb->hose->global_number,
 334                                 index);
 335                         goto fail;
 336                 }
 337         }
 338
 339         /*
 340          * Exclude the segments for reserved and root bus PE, which
 341          * are first or last two PEs.
 342          */
 343         r = &phb->hose->mem_resources[1];
 344         if (phb->ioda.reserved_pe_idx == 0)
 345                 r->start += (2 * phb->ioda.m64_segsize);
 346         else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
 347                 r->end -= (2 * phb->ioda.m64_segsize);
 348         else
 349                 WARN(1, "Wrong reserved PE#%x on PHB#%x\n",
 350                      phb->ioda.reserved_pe_idx, phb->hose->global_number);
 351
 352         return 0;
 353
 354 fail:
 355         for ( ; index >= 0; index--)
 356                 opal_pci_phb_mmio_enable(phb->opal_id,
 357                         OPAL_M64_WINDOW_TYPE, index, OPAL_DISABLE_M64);
 358
 359         return -EIO;
 360 }
 361
 362 static void pnv_ioda_reserve_m64_pe(struct pci_bus *bus,
 363                                     unsigned long *pe_bitmap,
 364                                     bool all)
 365 {
 366         struct pci_dev *pdev;
 367
 368         list_for_each_entry(pdev, &bus->devices, bus_list) {
 369                 pnv_ioda_reserve_dev_m64_pe(pdev, pe_bitmap);
 370
 371                 if (all && pdev->subordinate)
 372                         pnv_ioda_reserve_m64_pe(pdev->subordinate,
 373                                                 pe_bitmap, all);
 374         }
 375 }
 376
 377 static struct pnv_ioda_pe *pnv_ioda_pick_m64_pe(struct pci_bus *bus, bool all)
 378 {
 379         struct pnv_phb *phb = pci_bus_to_pnvhb(bus);
 380         struct pnv_ioda_pe *master_pe, *pe;
 381         unsigned long size, *pe_alloc;
 382         int i;
 383
 384         /* Root bus shouldn't use M64 */
 385         if (pci_is_root_bus(bus))
 386                 return NULL;
 387
 388         /* Allocate bitmap */
 389         size = ALIGN(phb->ioda.total_pe_num / 8, sizeof(unsigned long));
 390         pe_alloc = kzalloc(size, GFP_KERNEL);
 391         if (!pe_alloc) {
 392                 pr_warn("%s: Out of memory !\n",
 393                         __func__);
 394                 return NULL;
 395         }
 396
 397         /* Figure out reserved PE numbers by the PE */
 398         pnv_ioda_reserve_m64_pe(bus, pe_alloc, all);
 399
 400         /*
 401          * the current bus might not own M64 window and that's all
 402          * contributed by its child buses. For the case, we needn't
 403          * pick M64 dependent PE#.
 404          */
 405         if (bitmap_empty(pe_alloc, phb->ioda.total_pe_num)) {
 406                 kfree(pe_alloc);
 407                 return NULL;
 408         }
 409
 410         /*
 411          * Figure out the master PE and put all slave PEs to master
 412          * PE's list to form compound PE.
 413          */
 414         master_pe = NULL;
 415         i = -1;
 416         while ((i = find_next_bit(pe_alloc, phb->ioda.total_pe_num, i + 1)) <
 417                 phb->ioda.total_pe_num) {
 418                 pe = &phb->ioda.pe_array[i];
 419
 420                 phb->ioda.m64_segmap[pe->pe_number] = pe->pe_number;
 421                 if (!master_pe) {
 422                         pe->flags |= PNV_IODA_PE_MASTER;
 423                         INIT_LIST_HEAD(&pe->slaves);
 424                         master_pe = pe;
 425                 } else {
 426                         pe->flags |= PNV_IODA_PE_SLAVE;
 427                         pe->master = master_pe;
 428                         list_add_tail(&pe->list, &master_pe->slaves);
 429                 }
 430         }
 431
 432         kfree(pe_alloc);
 433         return master_pe;
 434 }
 435
 436 static void __init pnv_ioda_parse_m64_window(struct pnv_phb *phb)
 437 {
 438         struct pci_controller *hose = phb->hose;
 439         struct device_node *dn = hose->dn;
 440         struct resource *res;
 441         u32 m64_range[2], i;
 442         const __be32 *r;
 443         u64 pci_addr;
 444
 445         if (phb->type != PNV_PHB_IODA1 && phb->type != PNV_PHB_IODA2) {
 446                 pr_info("  Not support M64 window\n");
 447                 return;
 448         }
 449
 450         if (!firmware_has_feature(FW_FEATURE_OPAL)) {
 451                 pr_info("  Firmware too old to support M64 window\n");
 452                 return;
 453         }
 454
 455         r = of_get_property(dn, "ibm,opal-m64-window", NULL);
 456         if (!r) {
 457                 pr_info("  No <ibm,opal-m64-window> on %pOF\n",
 458                         dn);
 459                 return;
 460         }
 461
 462         /*
 463          * Find the available M64 BAR range and pickup the last one for
 464          * covering the whole 64-bits space. We support only one range.
 465          */
 466         if (of_property_read_u32_array(dn, "ibm,opal-available-m64-ranges",
 467                                        m64_range, 2)) {
 468                 /* In absence of the property, assume 0..15 */
 469                 m64_range[0] = 0;
 470                 m64_range[1] = 16;
 471         }
 472         /* We only support 64 bits in our allocator */
 473         if (m64_range[1] > 63) {
 474                 pr_warn("%s: Limiting M64 range to 63 (from %d) on PHB#%x\n",
 475                         __func__, m64_range[1], phb->hose->global_number);
 476                 m64_range[1] = 63;
 477         }
 478         /* Empty range, no m64 */
 479         if (m64_range[1] <= m64_range[0]) {
 480                 pr_warn("%s: M64 empty, disabling M64 usage on PHB#%x\n",
 481                         __func__, phb->hose->global_number);
 482                 return;
 483         }
 484
 485         /* Configure M64 informations */
 486         res = &hose->mem_resources[1];
 487         res->name = dn->full_name;
 488         res->start = of_translate_address(dn, r + 2);
 489         res->end = res->start + of_read_number(r + 4, 2) - 1;
 490         res->flags = (IORESOURCE_MEM | IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
 491         pci_addr = of_read_number(r, 2);
 492         hose->mem_offset[1] = res->start - pci_addr;
 493
 494         phb->ioda.m64_size = resource_size(res);
 495         phb->ioda.m64_segsize = phb->ioda.m64_size / phb->ioda.total_pe_num;
 496         phb->ioda.m64_base = pci_addr;
 497
 498         /* This lines up nicely with the display from processing OF ranges */
 499         pr_info(" MEM 0x%016llx..0x%016llx -> 0x%016llx (M64 #%d..%d)\n",
 500                 res->start, res->end, pci_addr, m64_range[0],
 501                 m64_range[0] + m64_range[1] - 1);
 502
 503         /* Mark all M64 used up by default */
 504         phb->ioda.m64_bar_alloc = (unsigned long)-1;
 505
 506         /* Use last M64 BAR to cover M64 window */
 507         m64_range[1]--;
 508         phb->ioda.m64_bar_idx = m64_range[0] + m64_range[1];
 509
 510         pr_info(" Using M64 #%d as default window\n", phb->ioda.m64_bar_idx);
 511
 512         /* Mark remaining ones free */
 513         for (i = m64_range[0]; i < m64_range[1]; i++)
 514                 clear_bit(i, &phb->ioda.m64_bar_alloc);
 515
 516         /*
 517          * Setup init functions for M64 based on IODA version, IODA3 uses
 518          * the IODA2 code.
 519          */
 520         if (phb->type == PNV_PHB_IODA1)
 521                 phb->init_m64 = pnv_ioda1_init_m64;
 522         else
 523                 phb->init_m64 = pnv_ioda2_init_m64;
 524 }
 525
 526 static void pnv_ioda_freeze_pe(struct pnv_phb *phb, int pe_no)
 527 {
 528         struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_no];
 529         struct pnv_ioda_pe *slave;
 530         s64 rc;
 531
 532         /* Fetch master PE */
 533         if (pe->flags & PNV_IODA_PE_SLAVE) {
 534                 pe = pe->master;
 535                 if (WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)))
 536                         return;
 537
 538                 pe_no = pe->pe_number;
 539         }
 540
 541         /* Freeze master PE */
 542         rc = opal_pci_eeh_freeze_set(phb->opal_id,
 543                                      pe_no,
 544                                      OPAL_EEH_ACTION_SET_FREEZE_ALL);
 545         if (rc != OPAL_SUCCESS) {
 546                 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
 547                         __func__, rc, phb->hose->global_number, pe_no);
 548                 return;
 549         }
 550
 551         /* Freeze slave PEs */
 552         if (!(pe->flags & PNV_IODA_PE_MASTER))
 553                 return;
 554
 555         list_for_each_entry(slave, &pe->slaves, list) {
 556                 rc = opal_pci_eeh_freeze_set(phb->opal_id,
 557                                              slave->pe_number,
 558                                              OPAL_EEH_ACTION_SET_FREEZE_ALL);
 559                 if (rc != OPAL_SUCCESS)
 560                         pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
 561                                 __func__, rc, phb->hose->global_number,
 562                                 slave->pe_number);
 563         }
 564 }
 565
 566 static int pnv_ioda_unfreeze_pe(struct pnv_phb *phb, int pe_no, int opt)
 567 {
 568         struct pnv_ioda_pe *pe, *slave;
 569         s64 rc;
 570
 571         /* Find master PE */
 572         pe = &phb->ioda.pe_array[pe_no];
 573         if (pe->flags & PNV_IODA_PE_SLAVE) {
 574                 pe = pe->master;
 575                 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
 576                 pe_no = pe->pe_number;
 577         }
 578
 579         /* Clear frozen state for master PE */
 580         rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, opt);
 581         if (rc != OPAL_SUCCESS) {
 582                 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
 583                         __func__, rc, opt, phb->hose->global_number, pe_no);
 584                 return -EIO;
 585         }
 586
 587         if (!(pe->flags & PNV_IODA_PE_MASTER))
 588                 return 0;
 589
 590         /* Clear frozen state for slave PEs */
 591         list_for_each_entry(slave, &pe->slaves, list) {
 592                 rc = opal_pci_eeh_freeze_clear(phb->opal_id,
 593                                              slave->pe_number,
 594                                              opt);
 595                 if (rc != OPAL_SUCCESS) {
 596                         pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
 597                                 __func__, rc, opt, phb->hose->global_number,
 598                                 slave->pe_number);
 599                         return -EIO;
 600                 }
 601         }
 602
 603         return 0;
 604 }
 605
 606 static int pnv_ioda_get_pe_state(struct pnv_phb *phb, int pe_no)
 607 {
 608         struct pnv_ioda_pe *slave, *pe;
 609         u8 fstate = 0, state;
 610         __be16 pcierr = 0;
 611         s64 rc;
 612
 613         /* Sanity check on PE number */
 614         if (pe_no < 0 || pe_no >= phb->ioda.total_pe_num)
 615                 return OPAL_EEH_STOPPED_PERM_UNAVAIL;
 616
 617         /*
 618          * Fetch the master PE and the PE instance might be
 619          * not initialized yet.
 620          */
 621         pe = &phb->ioda.pe_array[pe_no];
 622         if (pe->flags & PNV_IODA_PE_SLAVE) {
 623                 pe = pe->master;
 624                 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
 625                 pe_no = pe->pe_number;
 626         }
 627
 628         /* Check the master PE */
 629         rc = opal_pci_eeh_freeze_status(phb->opal_id, pe_no,
 630                                         &state, &pcierr, NULL);
 631         if (rc != OPAL_SUCCESS) {
 632                 pr_warn("%s: Failure %lld getting "
 633                         "PHB#%x-PE#%x state\n",
 634                         __func__, rc,
 635                         phb->hose->global_number, pe_no);
 636                 return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
 637         }
 638
 639         /* Check the slave PE */
 640         if (!(pe->flags & PNV_IODA_PE_MASTER))
 641                 return state;
 642
 643         list_for_each_entry(slave, &pe->slaves, list) {
 644                 rc = opal_pci_eeh_freeze_status(phb->opal_id,
 645                                                 slave->pe_number,
 646                                                 &fstate,
 647                                                 &pcierr,
 648                                                 NULL);
 649                 if (rc != OPAL_SUCCESS) {
 650                         pr_warn("%s: Failure %lld getting "
 651                                 "PHB#%x-PE#%x state\n",
 652                                 __func__, rc,
 653                                 phb->hose->global_number, slave->pe_number);
 654                         return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
 655                 }
 656
 657                 /*
 658                  * Override the result based on the ascending
 659                  * priority.
 660                  */
 661                 if (fstate > state)
 662                         state = fstate;
 663         }
 664
 665         return state;
 666 }
 667
 668 struct pnv_ioda_pe *pnv_pci_bdfn_to_pe(struct pnv_phb *phb, u16 bdfn)
 669 {
 670         int pe_number = phb->ioda.pe_rmap[bdfn];
 671
 672         if (pe_number == IODA_INVALID_PE)
 673                 return NULL;
 674
 675         return &phb->ioda.pe_array[pe_number];
 676 }
 677
 678 struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev)
 679 {
 680         struct pnv_phb *phb = pci_bus_to_pnvhb(dev->bus);
 681         struct pci_dn *pdn = pci_get_pdn(dev);
 682
 683         if (!pdn)
 684                 return NULL;
 685         if (pdn->pe_number == IODA_INVALID_PE)
 686                 return NULL;
 687         return &phb->ioda.pe_array[pdn->pe_number];
 688 }
 689
 690 static int pnv_ioda_set_one_peltv(struct pnv_phb *phb,
 691                                   struct pnv_ioda_pe *parent,
 692                                   struct pnv_ioda_pe *child,
 693                                   bool is_add)
 694 {
 695         const char *desc = is_add ? "adding" : "removing";
 696         uint8_t op = is_add ? OPAL_ADD_PE_TO_DOMAIN :
 697                               OPAL_REMOVE_PE_FROM_DOMAIN;
 698         struct pnv_ioda_pe *slave;
 699         long rc;
 700
 701         /* Parent PE affects child PE */
 702         rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
 703                                 child->pe_number, op);
 704         if (rc != OPAL_SUCCESS) {
 705                 pe_warn(child, "OPAL error %ld %s to parent PELTV\n",
 706                         rc, desc);
 707                 return -ENXIO;
 708         }
 709
 710         if (!(child->flags & PNV_IODA_PE_MASTER))
 711                 return 0;
 712
 713         /* Compound case: parent PE affects slave PEs */
 714         list_for_each_entry(slave, &child->slaves, list) {
 715                 rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
 716                                         slave->pe_number, op);
 717                 if (rc != OPAL_SUCCESS) {
 718                         pe_warn(slave, "OPAL error %ld %s to parent PELTV\n",
 719                                 rc, desc);
 720                         return -ENXIO;
 721                 }
 722         }
 723
 724         return 0;
 725 }
 726
 727 static int pnv_ioda_set_peltv(struct pnv_phb *phb,
 728                               struct pnv_ioda_pe *pe,
 729                               bool is_add)
 730 {
 731         struct pnv_ioda_pe *slave;
 732         struct pci_dev *pdev = NULL;
 733         int ret;
 734
 735         /*
 736          * Clear PE frozen state. If it's master PE, we need
 737          * clear slave PE frozen state as well.
 738          */
 739         if (is_add) {
 740                 opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
 741                                           OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 742                 if (pe->flags & PNV_IODA_PE_MASTER) {
 743                         list_for_each_entry(slave, &pe->slaves, list)
 744                                 opal_pci_eeh_freeze_clear(phb->opal_id,
 745                                                           slave->pe_number,
 746                                                           OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 747                 }
 748         }
 749
 750         /*
 751          * Associate PE in PELT. We need add the PE into the
 752          * corresponding PELT-V as well. Otherwise, the error
 753          * originated from the PE might contribute to other
 754          * PEs.
 755          */
 756         ret = pnv_ioda_set_one_peltv(phb, pe, pe, is_add);
 757         if (ret)
 758                 return ret;
 759
 760         /* For compound PEs, any one affects all of them */
 761         if (pe->flags & PNV_IODA_PE_MASTER) {
 762                 list_for_each_entry(slave, &pe->slaves, list) {
 763                         ret = pnv_ioda_set_one_peltv(phb, slave, pe, is_add);
 764                         if (ret)
 765                                 return ret;
 766                 }
 767         }
 768
 769         if (pe->flags & (PNV_IODA_PE_BUS_ALL | PNV_IODA_PE_BUS))
 770                 pdev = pe->pbus->self;
 771         else if (pe->flags & PNV_IODA_PE_DEV)
 772                 pdev = pe->pdev->bus->self;
 773 #ifdef CONFIG_PCI_IOV
 774         else if (pe->flags & PNV_IODA_PE_VF)
 775                 pdev = pe->parent_dev;
 776 #endif /* CONFIG_PCI_IOV */
 777         while (pdev) {
 778                 struct pci_dn *pdn = pci_get_pdn(pdev);
 779                 struct pnv_ioda_pe *parent;
 780
 781                 if (pdn && pdn->pe_number != IODA_INVALID_PE) {
 782                         parent = &phb->ioda.pe_array[pdn->pe_number];
 783                         ret = pnv_ioda_set_one_peltv(phb, parent, pe, is_add);
 784                         if (ret)
 785                                 return ret;
 786                 }
 787
 788                 pdev = pdev->bus->self;
 789         }
 790
 791         return 0;
 792 }
 793
 794 static void pnv_ioda_unset_peltv(struct pnv_phb *phb,
 795                                  struct pnv_ioda_pe *pe,
 796                                  struct pci_dev *parent)
 797 {
 798         int64_t rc;
 799
 800         while (parent) {
 801                 struct pci_dn *pdn = pci_get_pdn(parent);
 802
 803                 if (pdn && pdn->pe_number != IODA_INVALID_PE) {
 804                         rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number,
 805                                                 pe->pe_number,
 806                                                 OPAL_REMOVE_PE_FROM_DOMAIN);
 807                         /* XXX What to do in case of error ? */
 808                 }
 809                 parent = parent->bus->self;
 810         }
 811
 812         opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
 813                                   OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 814
 815         /* Disassociate PE in PELT */
 816         rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number,
 817                                 pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
 818         if (rc)
 819                 pe_warn(pe, "OPAL error %lld remove self from PELTV\n", rc);
 820 }
 821
 822 int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
 823 {
 824         struct pci_dev *parent;
 825         uint8_t bcomp, dcomp, fcomp;
 826         int64_t rc;
 827         long rid_end, rid;
 828
 829         /* Currently, we just deconfigure VF PE. Bus PE will always there.*/
 830         if (pe->pbus) {
 831                 int count;
 832
 833                 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
 834                 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
 835                 parent = pe->pbus->self;
 836                 if (pe->flags & PNV_IODA_PE_BUS_ALL)
 837                         count = resource_size(&pe->pbus->busn_res);
 838                 else
 839                         count = 1;
 840
 841                 switch(count) {
 842                 case  1: bcomp = OpalPciBusAll;         break;
 843                 case  2: bcomp = OpalPciBus7Bits;       break;
 844                 case  4: bcomp = OpalPciBus6Bits;       break;
 845                 case  8: bcomp = OpalPciBus5Bits;       break;
 846                 case 16: bcomp = OpalPciBus4Bits;       break;
 847                 case 32: bcomp = OpalPciBus3Bits;       break;
 848                 default:
 849                         dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
 850                                 count);
 851                         /* Do an exact match only */
 852                         bcomp = OpalPciBusAll;
 853                 }
 854                 rid_end = pe->rid + (count << 8);
 855         } else {
 856 #ifdef CONFIG_PCI_IOV
 857                 if (pe->flags & PNV_IODA_PE_VF)
 858                         parent = pe->parent_dev;
 859                 else
 860 #endif
 861                         parent = pe->pdev->bus->self;
 862                 bcomp = OpalPciBusAll;
 863                 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
 864                 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
 865                 rid_end = pe->rid + 1;
 866         }
 867
 868         /* Clear the reverse map */
 869         for (rid = pe->rid; rid < rid_end; rid++)
 870                 phb->ioda.pe_rmap[rid] = IODA_INVALID_PE;
 871
 872         /*
 873          * Release from all parents PELT-V. NPUs don't have a PELTV
 874          * table
 875          */
 876         if (phb->type != PNV_PHB_NPU_OCAPI)
 877                 pnv_ioda_unset_peltv(phb, pe, parent);
 878
 879         rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
 880                              bcomp, dcomp, fcomp, OPAL_UNMAP_PE);
 881         if (rc)
 882                 pe_err(pe, "OPAL error %lld trying to setup PELT table\n", rc);
 883
 884         pe->pbus = NULL;
 885         pe->pdev = NULL;
 886 #ifdef CONFIG_PCI_IOV
 887         pe->parent_dev = NULL;
 888 #endif
 889
 890         return 0;
 891 }
 892
 893 int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
 894 {
 895         uint8_t bcomp, dcomp, fcomp;
 896         long rc, rid_end, rid;
 897
 898         /* Bus validation ? */
 899         if (pe->pbus) {
 900                 int count;
 901
 902                 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
 903                 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
 904                 if (pe->flags & PNV_IODA_PE_BUS_ALL)
 905                         count = resource_size(&pe->pbus->busn_res);
 906                 else
 907                         count = 1;
 908
 909                 switch(count) {
 910                 case  1: bcomp = OpalPciBusAll;         break;
 911                 case  2: bcomp = OpalPciBus7Bits;       break;
 912                 case  4: bcomp = OpalPciBus6Bits;       break;
 913                 case  8: bcomp = OpalPciBus5Bits;       break;
 914                 case 16: bcomp = OpalPciBus4Bits;       break;
 915                 case 32: bcomp = OpalPciBus3Bits;       break;
 916                 default:
 917                         dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
 918                                 count);
 919                         /* Do an exact match only */
 920                         bcomp = OpalPciBusAll;
 921                 }
 922                 rid_end = pe->rid + (count << 8);
 923         } else {
 924                 bcomp = OpalPciBusAll;
 925                 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
 926                 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
 927                 rid_end = pe->rid + 1;
 928         }
 929
 930         /*
 931          * Associate PE in PELT. We need add the PE into the
 932          * corresponding PELT-V as well. Otherwise, the error
 933          * originated from the PE might contribute to other
 934          * PEs.
 935          */
 936         rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
 937                              bcomp, dcomp, fcomp, OPAL_MAP_PE);
 938         if (rc) {
 939                 pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
 940                 return -ENXIO;
 941         }
 942
 943         /*
 944          * Configure PELTV. NPUs don't have a PELTV table so skip
 945          * configuration on them.
 946          */
 947         if (phb->type != PNV_PHB_NPU_OCAPI)
 948                 pnv_ioda_set_peltv(phb, pe, true);
 949
 950         /* Setup reverse map */
 951         for (rid = pe->rid; rid < rid_end; rid++)
 952                 phb->ioda.pe_rmap[rid] = pe->pe_number;
 953
 954         /* Setup one MVTs on IODA1 */
 955         if (phb->type != PNV_PHB_IODA1) {
 956                 pe->mve_number = 0;
 957                 goto out;
 958         }
 959
 960         pe->mve_number = pe->pe_number;
 961         rc = opal_pci_set_mve(phb->opal_id, pe->mve_number, pe->pe_number);
 962         if (rc != OPAL_SUCCESS) {
 963                 pe_err(pe, "OPAL error %ld setting up MVE %x\n",
 964                        rc, pe->mve_number);
 965                 pe->mve_number = -1;
 966         } else {
 967                 rc = opal_pci_set_mve_enable(phb->opal_id,
 968                                              pe->mve_number, OPAL_ENABLE_MVE);
 969                 if (rc) {
 970                         pe_err(pe, "OPAL error %ld enabling MVE %x\n",
 971                                rc, pe->mve_number);
 972                         pe->mve_number = -1;
 973                 }
 974         }
 975
 976 out:
 977         return 0;
 978 }
 979
 980 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev)
 981 {
 982         struct pnv_phb *phb = pci_bus_to_pnvhb(dev->bus);
 983         struct pci_dn *pdn = pci_get_pdn(dev);
 984         struct pnv_ioda_pe *pe;
 985
 986         if (!pdn) {
 987                 pr_err("%s: Device tree node not associated properly\n",
 988                            pci_name(dev));
 989                 return NULL;
 990         }
 991         if (pdn->pe_number != IODA_INVALID_PE)
 992                 return NULL;
 993
 994         pe = pnv_ioda_alloc_pe(phb, 1);
 995         if (!pe) {
 996                 pr_warn("%s: Not enough PE# available, disabling device\n",
 997                         pci_name(dev));
 998                 return NULL;
 999         }
1000
1001         /* NOTE: We don't get a reference for the pointer in the PE
1002          * data structure, both the device and PE structures should be
1003          * destroyed at the same time.
1004          *
1005          * At some point we want to remove the PDN completely anyways
1006          */
1007         pdn->pe_number = pe->pe_number;
1008         pe->flags = PNV_IODA_PE_DEV;
1009         pe->pdev = dev;
1010         pe->pbus = NULL;
1011         pe->mve_number = -1;
1012         pe->rid = dev->bus->number << 8 | pdn->devfn;
1013         pe->device_count++;
1014
1015         pe_info(pe, "Associated device to PE\n");
1016
1017         if (pnv_ioda_configure_pe(phb, pe)) {
1018                 /* XXX What do we do here ? */
1019                 pnv_ioda_free_pe(pe);
1020                 pdn->pe_number = IODA_INVALID_PE;
1021                 pe->pdev = NULL;
1022                 return NULL;
1023         }
1024
1025         /* Put PE to the list */
1026         mutex_lock(&phb->ioda.pe_list_mutex);
1027         list_add_tail(&pe->list, &phb->ioda.pe_list);
1028         mutex_unlock(&phb->ioda.pe_list_mutex);
1029         return pe;
1030 }
1031
1032 /*
1033  * There're 2 types of PCI bus sensitive PEs: One that is compromised of
1034  * single PCI bus. Another one that contains the primary PCI bus and its
1035  * subordinate PCI devices and buses. The second type of PE is normally
1036  * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
1037  */
1038 static struct pnv_ioda_pe *pnv_ioda_setup_bus_PE(struct pci_bus *bus, bool all)
1039 {
1040         struct pnv_phb *phb = pci_bus_to_pnvhb(bus);
1041         struct pnv_ioda_pe *pe = NULL;
1042         unsigned int pe_num;
1043
1044         /*
1045          * In partial hotplug case, the PE instance might be still alive.
1046          * We should reuse it instead of allocating a new one.
1047          */
1048         pe_num = phb->ioda.pe_rmap[bus->number << 8];
1049         if (WARN_ON(pe_num != IODA_INVALID_PE)) {
1050                 pe = &phb->ioda.pe_array[pe_num];
1051                 return NULL;
1052         }
1053
1054         /* PE number for root bus should have been reserved */
1055         if (pci_is_root_bus(bus))
1056                 pe = &phb->ioda.pe_array[phb->ioda.root_pe_idx];
1057
1058         /* Check if PE is determined by M64 */
1059         if (!pe)
1060                 pe = pnv_ioda_pick_m64_pe(bus, all);
1061
1062         /* The PE number isn't pinned by M64 */
1063         if (!pe)
1064                 pe = pnv_ioda_alloc_pe(phb, 1);
1065
1066         if (!pe) {
1067                 pr_warn("%s: Not enough PE# available for PCI bus %04x:%02x\n",
1068                         __func__, pci_domain_nr(bus), bus->number);
1069                 return NULL;
1070         }
1071
1072         pe->flags |= (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS);
1073         pe->pbus = bus;
1074         pe->pdev = NULL;
1075         pe->mve_number = -1;
1076         pe->rid = bus->busn_res.start << 8;
1077
1078         if (all)
1079                 pe_info(pe, "Secondary bus %pad..%pad associated with PE#%x\n",
1080                         &bus->busn_res.start, &bus->busn_res.end,
1081                         pe->pe_number);
1082         else
1083                 pe_info(pe, "Secondary bus %pad associated with PE#%x\n",
1084                         &bus->busn_res.start, pe->pe_number);
1085
1086         if (pnv_ioda_configure_pe(phb, pe)) {
1087                 /* XXX What do we do here ? */
1088                 pnv_ioda_free_pe(pe);
1089                 pe->pbus = NULL;
1090                 return NULL;
1091         }
1092
1093         /* Put PE to the list */
1094         list_add_tail(&pe->list, &phb->ioda.pe_list);
1095
1096         return pe;
1097 }
1098
1099 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb *phb,
1100                                        struct pnv_ioda_pe *pe);
1101
1102 static void pnv_pci_ioda_dma_dev_setup(struct pci_dev *pdev)
1103 {
1104         struct pnv_phb *phb = pci_bus_to_pnvhb(pdev->bus);
1105         struct pci_dn *pdn = pci_get_pdn(pdev);
1106         struct pnv_ioda_pe *pe;
1107
1108         /* Check if the BDFN for this device is associated with a PE yet */
1109         pe = pnv_pci_bdfn_to_pe(phb, pdev->devfn | (pdev->bus->number << 8));
1110         if (!pe) {
1111                 /* VF PEs should be pre-configured in pnv_pci_sriov_enable() */
1112                 if (WARN_ON(pdev->is_virtfn))
1113                         return;
1114
1115                 pnv_pci_configure_bus(pdev->bus);
1116                 pe = pnv_pci_bdfn_to_pe(phb, pdev->devfn | (pdev->bus->number << 8));
1117                 pci_info(pdev, "Configured PE#%x\n", pe ? pe->pe_number : 0xfffff);
1118
1119
1120                 /*
1121                  * If we can't setup the IODA PE something has gone horribly
1122                  * wrong and we can't enable DMA for the device.
1123                  */
1124                 if (WARN_ON(!pe))
1125                         return;
1126         } else {
1127                 pci_info(pdev, "Added to existing PE#%x\n", pe->pe_number);
1128         }
1129
1130         /*
1131          * We assume that bridges *probably* don't need to do any DMA so we can
1132          * skip allocating a TCE table, etc unless we get a non-bridge device.
1133          */
1134         if (!pe->dma_setup_done && !pci_is_bridge(pdev)) {
1135                 switch (phb->type) {
1136                 case PNV_PHB_IODA1:
1137                         pnv_pci_ioda1_setup_dma_pe(phb, pe);
1138                         break;
1139                 case PNV_PHB_IODA2:
1140                         pnv_pci_ioda2_setup_dma_pe(phb, pe);
1141                         break;
1142                 default:
1143                         pr_warn("%s: No DMA for PHB#%x (type %d)\n",
1144                                 __func__, phb->hose->global_number, phb->type);
1145                 }
1146         }
1147
1148         if (pdn)
1149                 pdn->pe_number = pe->pe_number;
1150         pe->device_count++;
1151
1152         WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops);
1153         pdev->dev.archdata.dma_offset = pe->tce_bypass_base;
1154         set_iommu_table_base(&pdev->dev, pe->table_group.tables[0]);
1155
1156         /* PEs with a DMA weight of zero won't have a group */
1157         if (pe->table_group.group)
1158                 iommu_add_device(&pe->table_group, &pdev->dev);
1159 }
1160
1161 /*
1162  * Reconfigure TVE#0 to be usable as 64-bit DMA space.
1163  *
1164  * The first 4GB of virtual memory for a PE is reserved for 32-bit accesses.
1165  * Devices can only access more than that if bit 59 of the PCI address is set
1166  * by hardware, which indicates TVE#1 should be used instead of TVE#0.
1167  * Many PCI devices are not capable of addressing that many bits, and as a
1168  * result are limited to the 4GB of virtual memory made available to 32-bit
1169  * devices in TVE#0.
1170  *
1171  * In order to work around this, reconfigure TVE#0 to be suitable for 64-bit
1172  * devices by configuring the virtual memory past the first 4GB inaccessible
1173  * by 64-bit DMAs.  This should only be used by devices that want more than
1174  * 4GB, and only on PEs that have no 32-bit devices.
1175  *
1176  * Currently this will only work on PHB3 (POWER8).
1177  */
1178 static int pnv_pci_ioda_dma_64bit_bypass(struct pnv_ioda_pe *pe)
1179 {
1180         u64 window_size, table_size, tce_count, addr;
1181         struct page *table_pages;
1182         u64 tce_order = 28; /* 256MB TCEs */
1183         __be64 *tces;
1184         s64 rc;
1185
1186         /*
1187          * Window size needs to be a power of two, but needs to account for
1188          * shifting memory by the 4GB offset required to skip 32bit space.
1189          */
1190         window_size = roundup_pow_of_two(memory_hotplug_max() + (1ULL << 32));
1191         tce_count = window_size >> tce_order;
1192         table_size = tce_count << 3;
1193
1194         if (table_size < PAGE_SIZE)
1195                 table_size = PAGE_SIZE;
1196
1197         table_pages = alloc_pages_node(pe->phb->hose->node, GFP_KERNEL,
1198                                        get_order(table_size));
1199         if (!table_pages)
1200                 goto err;
1201
1202         tces = page_address(table_pages);
1203         if (!tces)
1204                 goto err;
1205
1206         memset(tces, 0, table_size);
1207
1208         for (addr = 0; addr < memory_hotplug_max(); addr += (1 << tce_order)) {
1209                 tces[(addr + (1ULL << 32)) >> tce_order] =
1210                         cpu_to_be64(addr | TCE_PCI_READ | TCE_PCI_WRITE);
1211         }
1212
1213         rc = opal_pci_map_pe_dma_window(pe->phb->opal_id,
1214                                         pe->pe_number,
1215                                         /* reconfigure window 0 */
1216                                         (pe->pe_number << 1) + 0,
1217                                         1,
1218                                         __pa(tces),
1219                                         table_size,
1220                                         1 << tce_order);
1221         if (rc == OPAL_SUCCESS) {
1222                 pe_info(pe, "Using 64-bit DMA iommu bypass (through TVE#0)\n");
1223                 return 0;
1224         }
1225 err:
1226         pe_err(pe, "Error configuring 64-bit DMA bypass\n");
1227         return -EIO;
1228 }
1229
1230 static bool pnv_pci_ioda_iommu_bypass_supported(struct pci_dev *pdev,
1231                 u64 dma_mask)
1232 {
1233         struct pnv_phb *phb = pci_bus_to_pnvhb(pdev->bus);
1234         struct pci_dn *pdn = pci_get_pdn(pdev);
1235         struct pnv_ioda_pe *pe;
1236
1237         if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1238                 return false;
1239
1240         pe = &phb->ioda.pe_array[pdn->pe_number];
1241         if (pe->tce_bypass_enabled) {
1242                 u64 top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1;
1243                 if (dma_mask >= top)
1244                         return true;
1245         }
1246
1247         /*
1248          * If the device can't set the TCE bypass bit but still wants
1249          * to access 4GB or more, on PHB3 we can reconfigure TVE#0 to
1250          * bypass the 32-bit region and be usable for 64-bit DMAs.
1251          * The device needs to be able to address all of this space.
1252          */
1253         if (dma_mask >> 32 &&
1254             dma_mask > (memory_hotplug_max() + (1ULL << 32)) &&
1255             /* pe->pdev should be set if it's a single device, pe->pbus if not */
1256             (pe->device_count == 1 || !pe->pbus) &&
1257             phb->model == PNV_PHB_MODEL_PHB3) {
1258                 /* Configure the bypass mode */
1259                 s64 rc = pnv_pci_ioda_dma_64bit_bypass(pe);
1260                 if (rc)
1261                         return false;
1262                 /* 4GB offset bypasses 32-bit space */
1263                 pdev->dev.archdata.dma_offset = (1ULL << 32);
1264                 return true;
1265         }
1266
1267         return false;
1268 }
1269
1270 static inline __be64 __iomem *pnv_ioda_get_inval_reg(struct pnv_phb *phb,
1271                                                      bool real_mode)
1272 {
1273         return real_mode ? (__be64 __iomem *)(phb->regs_phys + 0x210) :
1274                 (phb->regs + 0x210);
1275 }
1276
1277 static void pnv_pci_p7ioc_tce_invalidate(struct iommu_table *tbl,
1278                 unsigned long index, unsigned long npages, bool rm)
1279 {
1280         struct iommu_table_group_link *tgl = list_first_entry_or_null(
1281                         &tbl->it_group_list, struct iommu_table_group_link,
1282                         next);
1283         struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1284                         struct pnv_ioda_pe, table_group);
1285         __be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm);
1286         unsigned long start, end, inc;
1287
1288         start = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset);
1289         end = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset +
1290                         npages - 1);
1291
1292         /* p7ioc-style invalidation, 2 TCEs per write */
1293         start |= (1ull << 63);
1294         end |= (1ull << 63);
1295         inc = 16;
1296         end |= inc - 1; /* round up end to be different than start */
1297
1298         mb(); /* Ensure above stores are visible */
1299         while (start <= end) {
1300                 if (rm)
1301                         __raw_rm_writeq_be(start, invalidate);
1302                 else
1303                         __raw_writeq_be(start, invalidate);
1304
1305                 start += inc;
1306         }
1307
1308         /*
1309          * The iommu layer will do another mb() for us on build()
1310          * and we don't care on free()
1311          */
1312 }
1313
1314 static int pnv_ioda1_tce_build(struct iommu_table *tbl, long index,
1315                 long npages, unsigned long uaddr,
1316                 enum dma_data_direction direction,
1317                 unsigned long attrs)
1318 {
1319         int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1320                         attrs);
1321
1322         if (!ret)
1323                 pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false);
1324
1325         return ret;
1326 }
1327
1328 #ifdef CONFIG_IOMMU_API
1329 /* Common for IODA1 and IODA2 */
1330 static int pnv_ioda_tce_xchg_no_kill(struct iommu_table *tbl, long index,
1331                 unsigned long *hpa, enum dma_data_direction *direction,
1332                 bool realmode)
1333 {
1334         return pnv_tce_xchg(tbl, index, hpa, direction, !realmode);
1335 }
1336 #endif
1337
1338 static void pnv_ioda1_tce_free(struct iommu_table *tbl, long index,
1339                 long npages)
1340 {
1341         pnv_tce_free(tbl, index, npages);
1342
1343         pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false);
1344 }
1345
1346 static struct iommu_table_ops pnv_ioda1_iommu_ops = {
1347         .set = pnv_ioda1_tce_build,
1348 #ifdef CONFIG_IOMMU_API
1349         .xchg_no_kill = pnv_ioda_tce_xchg_no_kill,
1350         .tce_kill = pnv_pci_p7ioc_tce_invalidate,
1351         .useraddrptr = pnv_tce_useraddrptr,
1352 #endif
1353         .clear = pnv_ioda1_tce_free,
1354         .get = pnv_tce_get,
1355 };
1356
1357 #define PHB3_TCE_KILL_INVAL_ALL         PPC_BIT(0)
1358 #define PHB3_TCE_KILL_INVAL_PE          PPC_BIT(1)
1359 #define PHB3_TCE_KILL_INVAL_ONE         PPC_BIT(2)
1360
1361 static inline void pnv_pci_phb3_tce_invalidate_pe(struct pnv_ioda_pe *pe)
1362 {
1363         /* 01xb - invalidate TCEs that match the specified PE# */
1364         __be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, false);
1365         unsigned long val = PHB3_TCE_KILL_INVAL_PE | (pe->pe_number & 0xFF);
1366
1367         mb(); /* Ensure above stores are visible */
1368         __raw_writeq_be(val, invalidate);
1369 }
1370
1371 static void pnv_pci_phb3_tce_invalidate(struct pnv_ioda_pe *pe, bool rm,
1372                                         unsigned shift, unsigned long index,
1373                                         unsigned long npages)
1374 {
1375         __be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm);
1376         unsigned long start, end, inc;
1377
1378         /* We'll invalidate DMA address in PE scope */
1379         start = PHB3_TCE_KILL_INVAL_ONE;
1380         start |= (pe->pe_number & 0xFF);
1381         end = start;
1382
1383         /* Figure out the start, end and step */
1384         start |= (index << shift);
1385         end |= ((index + npages - 1) << shift);
1386         inc = (0x1ull << shift);
1387         mb();
1388
1389         while (start <= end) {
1390                 if (rm)
1391                         __raw_rm_writeq_be(start, invalidate);
1392                 else
1393                         __raw_writeq_be(start, invalidate);
1394                 start += inc;
1395         }
1396 }
1397
1398 static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe *pe)
1399 {
1400         struct pnv_phb *phb = pe->phb;
1401
1402         if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
1403                 pnv_pci_phb3_tce_invalidate_pe(pe);
1404         else
1405                 opal_pci_tce_kill(phb->opal_id, OPAL_PCI_TCE_KILL_PE,
1406                                   pe->pe_number, 0, 0, 0);
1407 }
1408
1409 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table *tbl,
1410                 unsigned long index, unsigned long npages, bool rm)
1411 {
1412         struct iommu_table_group_link *tgl;
1413
1414         list_for_each_entry_lockless(tgl, &tbl->it_group_list, next) {
1415                 struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1416                                 struct pnv_ioda_pe, table_group);
1417                 struct pnv_phb *phb = pe->phb;
1418                 unsigned int shift = tbl->it_page_shift;
1419
1420                 if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
1421                         pnv_pci_phb3_tce_invalidate(pe, rm, shift,
1422                                                     index, npages);
1423                 else
1424                         opal_pci_tce_kill(phb->opal_id,
1425                                           OPAL_PCI_TCE_KILL_PAGES,
1426                                           pe->pe_number, 1u << shift,
1427                                           index << shift, npages);
1428         }
1429 }
1430
1431 static int pnv_ioda2_tce_build(struct iommu_table *tbl, long index,
1432                 long npages, unsigned long uaddr,
1433                 enum dma_data_direction direction,
1434                 unsigned long attrs)
1435 {
1436         int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1437                         attrs);
1438
1439         if (!ret)
1440                 pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
1441
1442         return ret;
1443 }
1444
1445 static void pnv_ioda2_tce_free(struct iommu_table *tbl, long index,
1446                 long npages)
1447 {
1448         pnv_tce_free(tbl, index, npages);
1449
1450         pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
1451 }
1452
1453 static struct iommu_table_ops pnv_ioda2_iommu_ops = {
1454         .set = pnv_ioda2_tce_build,
1455 #ifdef CONFIG_IOMMU_API
1456         .xchg_no_kill = pnv_ioda_tce_xchg_no_kill,
1457         .tce_kill = pnv_pci_ioda2_tce_invalidate,
1458         .useraddrptr = pnv_tce_useraddrptr,
1459 #endif
1460         .clear = pnv_ioda2_tce_free,
1461         .get = pnv_tce_get,
1462         .free = pnv_pci_ioda2_table_free_pages,
1463 };
1464
1465 static int pnv_pci_ioda_dev_dma_weight(struct pci_dev *dev, void *data)
1466 {
1467         unsigned int *weight = (unsigned int *)data;
1468
1469         /* This is quite simplistic. The "base" weight of a device
1470          * is 10. 0 means no DMA is to be accounted for it.
1471          */
1472         if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1473                 return 0;
1474
1475         if (dev->class == PCI_CLASS_SERIAL_USB_UHCI ||
1476             dev->class == PCI_CLASS_SERIAL_USB_OHCI ||
1477             dev->class == PCI_CLASS_SERIAL_USB_EHCI)
1478                 *weight += 3;
1479         else if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID)
1480                 *weight += 15;
1481         else
1482                 *weight += 10;
1483
1484         return 0;
1485 }
1486
1487 static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe *pe)
1488 {
1489         unsigned int weight = 0;
1490
1491         /* SRIOV VF has same DMA32 weight as its PF */
1492 #ifdef CONFIG_PCI_IOV
1493         if ((pe->flags & PNV_IODA_PE_VF) && pe->parent_dev) {
1494                 pnv_pci_ioda_dev_dma_weight(pe->parent_dev, &weight);
1495                 return weight;
1496         }
1497 #endif
1498
1499         if ((pe->flags & PNV_IODA_PE_DEV) && pe->pdev) {
1500                 pnv_pci_ioda_dev_dma_weight(pe->pdev, &weight);
1501         } else if ((pe->flags & PNV_IODA_PE_BUS) && pe->pbus) {
1502                 struct pci_dev *pdev;
1503
1504                 list_for_each_entry(pdev, &pe->pbus->devices, bus_list)
1505                         pnv_pci_ioda_dev_dma_weight(pdev, &weight);
1506         } else if ((pe->flags & PNV_IODA_PE_BUS_ALL) && pe->pbus) {
1507                 pci_walk_bus(pe->pbus, pnv_pci_ioda_dev_dma_weight, &weight);
1508         }
1509
1510         return weight;
1511 }
1512
1513 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb *phb,
1514                                        struct pnv_ioda_pe *pe)
1515 {
1516
1517         struct page *tce_mem = NULL;
1518         struct iommu_table *tbl;
1519         unsigned int weight, total_weight = 0;
1520         unsigned int tce32_segsz, base, segs, avail, i;
1521         int64_t rc;
1522         void *addr;
1523
1524         /* XXX FIXME: Handle 64-bit only DMA devices */
1525         /* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
1526         /* XXX FIXME: Allocate multi-level tables on PHB3 */
1527         weight = pnv_pci_ioda_pe_dma_weight(pe);
1528         if (!weight)
1529                 return;
1530
1531         pci_walk_bus(phb->hose->bus, pnv_pci_ioda_dev_dma_weight,
1532                      &total_weight);
1533         segs = (weight * phb->ioda.dma32_count) / total_weight;
1534         if (!segs)
1535                 segs = 1;
1536
1537         /*
1538          * Allocate contiguous DMA32 segments. We begin with the expected
1539          * number of segments. With one more attempt, the number of DMA32
1540          * segments to be allocated is decreased by one until one segment
1541          * is allocated successfully.
1542          */
1543         do {
1544                 for (base = 0; base <= phb->ioda.dma32_count - segs; base++) {
1545                         for (avail = 0, i = base; i < base + segs; i++) {
1546                                 if (phb->ioda.dma32_segmap[i] ==
1547                                     IODA_INVALID_PE)
1548                                         avail++;
1549                         }
1550
1551                         if (avail == segs)
1552                                 goto found;
1553                 }
1554         } while (--segs);
1555
1556         if (!segs) {
1557                 pe_warn(pe, "No available DMA32 segments\n");
1558                 return;
1559         }
1560
1561 found:
1562         tbl = pnv_pci_table_alloc(phb->hose->node);
1563         if (WARN_ON(!tbl))
1564                 return;
1565
1566         iommu_register_group(&pe->table_group, phb->hose->global_number,
1567                         pe->pe_number);
1568         pnv_pci_link_table_and_group(phb->hose->node, 0, tbl, &pe->table_group);
1569
1570         /* Grab a 32-bit TCE table */
1571         pe_info(pe, "DMA weight %d (%d), assigned (%d) %d DMA32 segments\n",
1572                 weight, total_weight, base, segs);
1573         pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n",
1574                 base * PNV_IODA1_DMA32_SEGSIZE,
1575                 (base + segs) * PNV_IODA1_DMA32_SEGSIZE - 1);
1576
1577         /* XXX Currently, we allocate one big contiguous table for the
1578          * TCEs. We only really need one chunk per 256M of TCE space
1579          * (ie per segment) but that's an optimization for later, it
1580          * requires some added smarts with our get/put_tce implementation
1581          *
1582          * Each TCE page is 4KB in size and each TCE entry occupies 8
1583          * bytes
1584          */
1585         tce32_segsz = PNV_IODA1_DMA32_SEGSIZE >> (IOMMU_PAGE_SHIFT_4K - 3);
1586         tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
1587                                    get_order(tce32_segsz * segs));
1588         if (!tce_mem) {
1589                 pe_err(pe, " Failed to allocate a 32-bit TCE memory\n");
1590                 goto fail;
1591         }
1592         addr = page_address(tce_mem);
1593         memset(addr, 0, tce32_segsz * segs);
1594
1595         /* Configure HW */
1596         for (i = 0; i < segs; i++) {
1597                 rc = opal_pci_map_pe_dma_window(phb->opal_id,
1598                                               pe->pe_number,
1599                                               base + i, 1,
1600                                               __pa(addr) + tce32_segsz * i,
1601                                               tce32_segsz, IOMMU_PAGE_SIZE_4K);
1602                 if (rc) {
1603                         pe_err(pe, " Failed to configure 32-bit TCE table, err %lld\n",
1604                                rc);
1605                         goto fail;
1606                 }
1607         }
1608
1609         /* Setup DMA32 segment mapping */
1610         for (i = base; i < base + segs; i++)
1611                 phb->ioda.dma32_segmap[i] = pe->pe_number;
1612
1613         /* Setup linux iommu table */
1614         pnv_pci_setup_iommu_table(tbl, addr, tce32_segsz * segs,
1615                                   base * PNV_IODA1_DMA32_SEGSIZE,
1616                                   IOMMU_PAGE_SHIFT_4K);
1617
1618         tbl->it_ops = &pnv_ioda1_iommu_ops;
1619         pe->table_group.tce32_start = tbl->it_offset << tbl->it_page_shift;
1620         pe->table_group.tce32_size = tbl->it_size << tbl->it_page_shift;
1621         tbl->it_index = (phb->hose->global_number << 16) | pe->pe_number;
1622         if (!iommu_init_table(tbl, phb->hose->node, 0, 0))
1623                 panic("Failed to initialize iommu table");
1624
1625         pe->dma_setup_done = true;
1626         return;
1627  fail:
1628         /* XXX Failure: Try to fallback to 64-bit only ? */
1629         if (tce_mem)
1630                 __free_pages(tce_mem, get_order(tce32_segsz * segs));
1631         if (tbl) {
1632                 pnv_pci_unlink_table_and_group(tbl, &pe->table_group);
1633                 iommu_tce_table_put(tbl);
1634         }
1635 }
1636
1637 static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group,
1638                 int num, struct iommu_table *tbl)
1639 {
1640         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1641                         table_group);
1642         struct pnv_phb *phb = pe->phb;
1643         int64_t rc;
1644         const unsigned long size = tbl->it_indirect_levels ?
1645                         tbl->it_level_size : tbl->it_size;
1646         const __u64 start_addr = tbl->it_offset << tbl->it_page_shift;
1647         const __u64 win_size = tbl->it_size << tbl->it_page_shift;
1648
1649         pe_info(pe, "Setting up window#%d %llx..%llx pg=%lx\n",
1650                 num, start_addr, start_addr + win_size - 1,
1651                 IOMMU_PAGE_SIZE(tbl));
1652
1653         /*
1654          * Map TCE table through TVT. The TVE index is the PE number
1655          * shifted by 1 bit for 32-bits DMA space.
1656          */
1657         rc = opal_pci_map_pe_dma_window(phb->opal_id,
1658                         pe->pe_number,
1659                         (pe->pe_number << 1) + num,
1660                         tbl->it_indirect_levels + 1,
1661                         __pa(tbl->it_base),
1662                         size << 3,
1663                         IOMMU_PAGE_SIZE(tbl));
1664         if (rc) {
1665                 pe_err(pe, "Failed to configure TCE table, err %lld\n", rc);
1666                 return rc;
1667         }
1668
1669         pnv_pci_link_table_and_group(phb->hose->node, num,
1670                         tbl, &pe->table_group);
1671         pnv_pci_ioda2_tce_invalidate_pe(pe);
1672
1673         return 0;
1674 }
1675
1676 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable)
1677 {
1678         uint16_t window_id = (pe->pe_number << 1 ) + 1;
1679         int64_t rc;
1680
1681         pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis");
1682         if (enable) {
1683                 phys_addr_t top = memblock_end_of_DRAM();
1684
1685                 top = roundup_pow_of_two(top);
1686                 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
1687                                                      pe->pe_number,
1688                                                      window_id,
1689                                                      pe->tce_bypass_base,
1690                                                      top);
1691         } else {
1692                 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
1693                                                      pe->pe_number,
1694                                                      window_id,
1695                                                      pe->tce_bypass_base,
1696                                                      0);
1697         }
1698         if (rc)
1699                 pe_err(pe, "OPAL error %lld configuring bypass window\n", rc);
1700         else
1701                 pe->tce_bypass_enabled = enable;
1702 }
1703
1704 static long pnv_pci_ioda2_create_table(struct iommu_table_group *table_group,
1705                 int num, __u32 page_shift, __u64 window_size, __u32 levels,
1706                 bool alloc_userspace_copy, struct iommu_table **ptbl)
1707 {
1708         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1709                         table_group);
1710         int nid = pe->phb->hose->node;
1711         __u64 bus_offset = num ? pe->tce_bypass_base : table_group->tce32_start;
1712         long ret;
1713         struct iommu_table *tbl;
1714
1715         tbl = pnv_pci_table_alloc(nid);
1716         if (!tbl)
1717                 return -ENOMEM;
1718
1719         tbl->it_ops = &pnv_ioda2_iommu_ops;
1720
1721         ret = pnv_pci_ioda2_table_alloc_pages(nid,
1722                         bus_offset, page_shift, window_size,
1723                         levels, alloc_userspace_copy, tbl);
1724         if (ret) {
1725                 iommu_tce_table_put(tbl);
1726                 return ret;
1727         }
1728
1729         *ptbl = tbl;
1730
1731         return 0;
1732 }
1733
1734 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe *pe)
1735 {
1736         struct iommu_table *tbl = NULL;
1737         long rc;
1738         unsigned long res_start, res_end;
1739
1740         /*
1741          * crashkernel= specifies the kdump kernel's maximum memory at
1742          * some offset and there is no guaranteed the result is a power
1743          * of 2, which will cause errors later.
1744          */
1745         const u64 max_memory = __rounddown_pow_of_two(memory_hotplug_max());
1746
1747         /*
1748          * In memory constrained environments, e.g. kdump kernel, the
1749          * DMA window can be larger than available memory, which will
1750          * cause errors later.
1751          */
1752         const u64 maxblock = 1UL << (PAGE_SHIFT + MAX_ORDER - 1);
1753
1754         /*
1755          * We create the default window as big as we can. The constraint is
1756          * the max order of allocation possible. The TCE table is likely to
1757          * end up being multilevel and with on-demand allocation in place,
1758          * the initial use is not going to be huge as the default window aims
1759          * to support crippled devices (i.e. not fully 64bit DMAble) only.
1760          */
1761         /* iommu_table::it_map uses 1 bit per IOMMU page, hence 8 */
1762         const u64 window_size = min((maxblock * 8) << PAGE_SHIFT, max_memory);
1763         /* Each TCE level cannot exceed maxblock so go multilevel if needed */
1764         unsigned long tces_order = ilog2(window_size >> PAGE_SHIFT);
1765         unsigned long tcelevel_order = ilog2(maxblock >> 3);
1766         unsigned int levels = tces_order / tcelevel_order;
1767
1768         if (tces_order % tcelevel_order)
1769                 levels += 1;
1770         /*
1771          * We try to stick to default levels (which is >1 at the moment) in
1772          * order to save memory by relying on on-demain TCE level allocation.
1773          */
1774         levels = max_t(unsigned int, levels, POWERNV_IOMMU_DEFAULT_LEVELS);
1775
1776         rc = pnv_pci_ioda2_create_table(&pe->table_group, 0, PAGE_SHIFT,
1777                         window_size, levels, false, &tbl);
1778         if (rc) {
1779                 pe_err(pe, "Failed to create 32-bit TCE table, err %ld",
1780                                 rc);
1781                 return rc;
1782         }
1783
1784         /* We use top part of 32bit space for MMIO so exclude it from DMA */
1785         res_start = 0;
1786         res_end = 0;
1787         if (window_size > pe->phb->ioda.m32_pci_base) {
1788                 res_start = pe->phb->ioda.m32_pci_base >> tbl->it_page_shift;
1789                 res_end = min(window_size, SZ_4G) >> tbl->it_page_shift;
1790         }
1791
1792         tbl->it_index = (pe->phb->hose->global_number << 16) | pe->pe_number;
1793         if (iommu_init_table(tbl, pe->phb->hose->node, res_start, res_end))
1794                 rc = pnv_pci_ioda2_set_window(&pe->table_group, 0, tbl);
1795         else
1796                 rc = -ENOMEM;
1797         if (rc) {
1798                 pe_err(pe, "Failed to configure 32-bit TCE table, err %ld\n", rc);
1799                 iommu_tce_table_put(tbl);
1800                 tbl = NULL; /* This clears iommu_table_base below */
1801         }
1802         if (!pnv_iommu_bypass_disabled)
1803                 pnv_pci_ioda2_set_bypass(pe, true);
1804
1805         /*
1806          * Set table base for the case of IOMMU DMA use. Usually this is done
1807          * from dma_dev_setup() which is not called when a device is returned
1808          * from VFIO so do it here.
1809          */
1810         if (pe->pdev)
1811                 set_iommu_table_base(&pe->pdev->dev, tbl);
1812
1813         return 0;
1814 }
1815
1816 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
1817                 int num)
1818 {
1819         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1820                         table_group);
1821         struct pnv_phb *phb = pe->phb;
1822         long ret;
1823
1824         pe_info(pe, "Removing DMA window #%d\n", num);
1825
1826         ret = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
1827                         (pe->pe_number << 1) + num,
1828                         0/* levels */, 0/* table address */,
1829                         0/* table size */, 0/* page size */);
1830         if (ret)
1831                 pe_warn(pe, "Unmapping failed, ret = %ld\n", ret);
1832         else
1833                 pnv_pci_ioda2_tce_invalidate_pe(pe);
1834
1835         pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
1836
1837         return ret;
1838 }
1839
1840 #ifdef CONFIG_IOMMU_API
1841 unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift,
1842                 __u64 window_size, __u32 levels)
1843 {
1844         unsigned long bytes = 0;
1845         const unsigned window_shift = ilog2(window_size);
1846         unsigned entries_shift = window_shift - page_shift;
1847         unsigned table_shift = entries_shift + 3;
1848         unsigned long tce_table_size = max(0x1000UL, 1UL << table_shift);
1849         unsigned long direct_table_size;
1850
1851         if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS) ||
1852                         !is_power_of_2(window_size))
1853                 return 0;
1854
1855         /* Calculate a direct table size from window_size and levels */
1856         entries_shift = (entries_shift + levels - 1) / levels;
1857         table_shift = entries_shift + 3;
1858         table_shift = max_t(unsigned, table_shift, PAGE_SHIFT);
1859         direct_table_size =  1UL << table_shift;
1860
1861         for ( ; levels; --levels) {
1862                 bytes += ALIGN(tce_table_size, direct_table_size);
1863
1864                 tce_table_size /= direct_table_size;
1865                 tce_table_size <<= 3;
1866                 tce_table_size = max_t(unsigned long,
1867                                 tce_table_size, direct_table_size);
1868         }
1869
1870         return bytes + bytes; /* one for HW table, one for userspace copy */
1871 }
1872
1873 static long pnv_pci_ioda2_create_table_userspace(
1874                 struct iommu_table_group *table_group,
1875                 int num, __u32 page_shift, __u64 window_size, __u32 levels,
1876                 struct iommu_table **ptbl)
1877 {
1878         long ret = pnv_pci_ioda2_create_table(table_group,
1879                         num, page_shift, window_size, levels, true, ptbl);
1880
1881         if (!ret)
1882                 (*ptbl)->it_allocated_size = pnv_pci_ioda2_get_table_size(
1883                                 page_shift, window_size, levels);
1884         return ret;
1885 }
1886
1887 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe, struct pci_bus *bus)
1888 {
1889         struct pci_dev *dev;
1890
1891         list_for_each_entry(dev, &bus->devices, bus_list) {
1892                 set_iommu_table_base(&dev->dev, pe->table_group.tables[0]);
1893                 dev->dev.archdata.dma_offset = pe->tce_bypass_base;
1894
1895                 if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1896                         pnv_ioda_setup_bus_dma(pe, dev->subordinate);
1897         }
1898 }
1899
1900 static void pnv_ioda2_take_ownership(struct iommu_table_group *table_group)
1901 {
1902         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1903                                                 table_group);
1904         /* Store @tbl as pnv_pci_ioda2_unset_window() resets it */
1905         struct iommu_table *tbl = pe->table_group.tables[0];
1906
1907         pnv_pci_ioda2_set_bypass(pe, false);
1908         pnv_pci_ioda2_unset_window(&pe->table_group, 0);
1909         if (pe->pbus)
1910                 pnv_ioda_setup_bus_dma(pe, pe->pbus);
1911         else if (pe->pdev)
1912                 set_iommu_table_base(&pe->pdev->dev, NULL);
1913         iommu_tce_table_put(tbl);
1914 }
1915
1916 static void pnv_ioda2_release_ownership(struct iommu_table_group *table_group)
1917 {
1918         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1919                                                 table_group);
1920
1921         pnv_pci_ioda2_setup_default_config(pe);
1922         if (pe->pbus)
1923                 pnv_ioda_setup_bus_dma(pe, pe->pbus);
1924 }
1925
1926 static struct iommu_table_group_ops pnv_pci_ioda2_ops = {
1927         .get_table_size = pnv_pci_ioda2_get_table_size,
1928         .create_table = pnv_pci_ioda2_create_table_userspace,
1929         .set_window = pnv_pci_ioda2_set_window,
1930         .unset_window = pnv_pci_ioda2_unset_window,
1931         .take_ownership = pnv_ioda2_take_ownership,
1932         .release_ownership = pnv_ioda2_release_ownership,
1933 };
1934 #endif
1935
1936 void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
1937                                 struct pnv_ioda_pe *pe)
1938 {
1939         int64_t rc;
1940
1941         /* TVE #1 is selected by PCI address bit 59 */
1942         pe->tce_bypass_base = 1ull << 59;
1943
1944         /* The PE will reserve all possible 32-bits space */
1945         pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n",
1946                 phb->ioda.m32_pci_base);
1947
1948         /* Setup linux iommu table */
1949         pe->table_group.tce32_start = 0;
1950         pe->table_group.tce32_size = phb->ioda.m32_pci_base;
1951         pe->table_group.max_dynamic_windows_supported =
1952                         IOMMU_TABLE_GROUP_MAX_TABLES;
1953         pe->table_group.max_levels = POWERNV_IOMMU_MAX_LEVELS;
1954         pe->table_group.pgsizes = pnv_ioda_parse_tce_sizes(phb);
1955
1956         rc = pnv_pci_ioda2_setup_default_config(pe);
1957         if (rc)
1958                 return;
1959
1960 #ifdef CONFIG_IOMMU_API
1961         pe->table_group.ops = &pnv_pci_ioda2_ops;
1962         iommu_register_group(&pe->table_group, phb->hose->global_number,
1963                              pe->pe_number);
1964 #endif
1965         pe->dma_setup_done = true;
1966 }
1967
1968 /*
1969  * Called from KVM in real mode to EOI passthru interrupts. The ICP
1970  * EOI is handled directly in KVM in kvmppc_deliver_irq_passthru().
1971  *
1972  * The IRQ data is mapped in the PCI-MSI domain and the EOI OPAL call
1973  * needs an HW IRQ number mapped in the XICS IRQ domain. The HW IRQ
1974  * numbers of the in-the-middle MSI domain are vector numbers and it's
1975  * good enough for OPAL. Use that.
1976  */
1977 int64_t pnv_opal_pci_msi_eoi(struct irq_data *d)
1978 {
1979         struct pci_controller *hose = irq_data_get_irq_chip_data(d->parent_data);
1980         struct pnv_phb *phb = hose->private_data;
1981
1982         return opal_pci_msi_eoi(phb->opal_id, d->parent_data->hwirq);
1983 }
1984
1985 /*
1986  * The IRQ data is mapped in the XICS domain, with OPAL HW IRQ numbers
1987  */
1988 static void pnv_ioda2_msi_eoi(struct irq_data *d)
1989 {
1990         int64_t rc;
1991         unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
1992         struct pci_controller *hose = irq_data_get_irq_chip_data(d);
1993         struct pnv_phb *phb = hose->private_data;
1994
1995         rc = opal_pci_msi_eoi(phb->opal_id, hw_irq);
1996         WARN_ON_ONCE(rc);
1997
1998         icp_native_eoi(d);
1999 }
2000
2001 /* P8/CXL only */
2002 void pnv_set_msi_irq_chip(struct pnv_phb *phb, unsigned int virq)
2003 {
2004         struct irq_data *idata;
2005         struct irq_chip *ichip;
2006
2007         /* The MSI EOI OPAL call is only needed on PHB3 */
2008         if (phb->model != PNV_PHB_MODEL_PHB3)
2009                 return;
2010
2011         if (!phb->ioda.irq_chip_init) {
2012                 /*
2013                  * First time we setup an MSI IRQ, we need to setup the
2014                  * corresponding IRQ chip to route correctly.
2015                  */
2016                 idata = irq_get_irq_data(virq);
2017                 ichip = irq_data_get_irq_chip(idata);
2018                 phb->ioda.irq_chip_init = 1;
2019                 phb->ioda.irq_chip = *ichip;
2020                 phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi;
2021         }
2022         irq_set_chip(virq, &phb->ioda.irq_chip);
2023         irq_set_chip_data(virq, phb->hose);
2024 }
2025
2026 static struct irq_chip pnv_pci_msi_irq_chip;
2027
2028 /*
2029  * Returns true iff chip is something that we could call
2030  * pnv_opal_pci_msi_eoi for.
2031  */
2032 bool is_pnv_opal_msi(struct irq_chip *chip)
2033 {
2034         return chip == &pnv_pci_msi_irq_chip;
2035 }
2036 EXPORT_SYMBOL_GPL(is_pnv_opal_msi);
2037
2038 static int __pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev,
2039                                     unsigned int xive_num,
2040                                     unsigned int is_64, struct msi_msg *msg)
2041 {
2042         struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev);
2043         __be32 data;
2044         int rc;
2045
2046         dev_dbg(&dev->dev, "%s: setup %s-bit MSI for vector #%d\n", __func__,
2047                 is_64 ? "64" : "32", xive_num);
2048
2049         /* No PE assigned ? bail out ... no MSI for you ! */
2050         if (pe == NULL)
2051                 return -ENXIO;
2052
2053         /* Check if we have an MVE */
2054         if (pe->mve_number < 0)
2055                 return -ENXIO;
2056
2057         /* Force 32-bit MSI on some broken devices */
2058         if (dev->no_64bit_msi)
2059                 is_64 = 0;
2060
2061         /* Assign XIVE to PE */
2062         rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
2063         if (rc) {
2064                 pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
2065                         pci_name(dev), rc, xive_num);
2066                 return -EIO;
2067         }
2068
2069         if (is_64) {
2070                 __be64 addr64;
2071
2072                 rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1,
2073                                      &addr64, &data);
2074                 if (rc) {
2075                         pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
2076                                 pci_name(dev), rc);
2077                         return -EIO;
2078                 }
2079                 msg->address_hi = be64_to_cpu(addr64) >> 32;
2080                 msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful;
2081         } else {
2082                 __be32 addr32;
2083
2084                 rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1,
2085                                      &addr32, &data);
2086                 if (rc) {
2087                         pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
2088                                 pci_name(dev), rc);
2089                         return -EIO;
2090                 }
2091                 msg->address_hi = 0;
2092                 msg->address_lo = be32_to_cpu(addr32);
2093         }
2094         msg->data = be32_to_cpu(data);
2095
2096         return 0;
2097 }
2098
2099 /*
2100  * The msi_free() op is called before irq_domain_free_irqs_top() when
2101  * the handler data is still available. Use that to clear the XIVE
2102  * controller.
2103  */
2104 static void pnv_msi_ops_msi_free(struct irq_domain *domain,
2105                                  struct msi_domain_info *info,
2106                                  unsigned int irq)
2107 {
2108         if (xive_enabled())
2109                 xive_irq_free_data(irq);
2110 }
2111
2112 static struct msi_domain_ops pnv_pci_msi_domain_ops = {
2113         .msi_free       = pnv_msi_ops_msi_free,
2114 };
2115
2116 static void pnv_msi_shutdown(struct irq_data *d)
2117 {
2118         d = d->parent_data;
2119         if (d->chip->irq_shutdown)
2120                 d->chip->irq_shutdown(d);
2121 }
2122
2123 static void pnv_msi_mask(struct irq_data *d)
2124 {
2125         pci_msi_mask_irq(d);
2126         irq_chip_mask_parent(d);
2127 }
2128
2129 static void pnv_msi_unmask(struct irq_data *d)
2130 {
2131         pci_msi_unmask_irq(d);
2132         irq_chip_unmask_parent(d);
2133 }
2134
2135 static struct irq_chip pnv_pci_msi_irq_chip = {
2136         .name           = "PNV-PCI-MSI",
2137         .irq_shutdown   = pnv_msi_shutdown,
2138         .irq_mask       = pnv_msi_mask,
2139         .irq_unmask     = pnv_msi_unmask,
2140         .irq_eoi        = irq_chip_eoi_parent,
2141 };
2142
2143 static struct msi_domain_info pnv_msi_domain_info = {
2144         .flags = (MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
2145                   MSI_FLAG_MULTI_PCI_MSI  | MSI_FLAG_PCI_MSIX),
2146         .ops   = &pnv_pci_msi_domain_ops,
2147         .chip  = &pnv_pci_msi_irq_chip,
2148 };
2149
2150 static void pnv_msi_compose_msg(struct irq_data *d, struct msi_msg *msg)
2151 {
2152         struct msi_desc *entry = irq_data_get_msi_desc(d);
2153         struct pci_dev *pdev = msi_desc_to_pci_dev(entry);
2154         struct pci_controller *hose = irq_data_get_irq_chip_data(d);
2155         struct pnv_phb *phb = hose->private_data;
2156         int rc;
2157
2158         rc = __pnv_pci_ioda_msi_setup(phb, pdev, d->hwirq,
2159                                       entry->msi_attrib.is_64, msg);
2160         if (rc)
2161                 dev_err(&pdev->dev, "Failed to setup %s-bit MSI #%ld : %d\n",
2162                         entry->msi_attrib.is_64 ? "64" : "32", d->hwirq, rc);
2163 }
2164
2165 /*
2166  * The IRQ data is mapped in the MSI domain in which HW IRQ numbers
2167  * correspond to vector numbers.
2168  */
2169 static void pnv_msi_eoi(struct irq_data *d)
2170 {
2171         struct pci_controller *hose = irq_data_get_irq_chip_data(d);
2172         struct pnv_phb *phb = hose->private_data;
2173
2174         if (phb->model == PNV_PHB_MODEL_PHB3) {
2175                 /*
2176                  * The EOI OPAL call takes an OPAL HW IRQ number but
2177                  * since it is translated into a vector number in
2178                  * OPAL, use that directly.
2179                  */
2180                 WARN_ON_ONCE(opal_pci_msi_eoi(phb->opal_id, d->hwirq));
2181         }
2182
2183         irq_chip_eoi_parent(d);
2184 }
2185
2186 static struct irq_chip pnv_msi_irq_chip = {
2187         .name                   = "PNV-MSI",
2188         .irq_shutdown           = pnv_msi_shutdown,
2189         .irq_mask               = irq_chip_mask_parent,
2190         .irq_unmask             = irq_chip_unmask_parent,
2191         .irq_eoi                = pnv_msi_eoi,
2192         .irq_set_affinity       = irq_chip_set_affinity_parent,
2193         .irq_compose_msi_msg    = pnv_msi_compose_msg,
2194 };
2195
2196 static int pnv_irq_parent_domain_alloc(struct irq_domain *domain,
2197                                        unsigned int virq, int hwirq)
2198 {
2199         struct irq_fwspec parent_fwspec;
2200         int ret;
2201
2202         parent_fwspec.fwnode = domain->parent->fwnode;
2203         parent_fwspec.param_count = 2;
2204         parent_fwspec.param[0] = hwirq;
2205         parent_fwspec.param[1] = IRQ_TYPE_EDGE_RISING;
2206
2207         ret = irq_domain_alloc_irqs_parent(domain, virq, 1, &parent_fwspec);
2208         if (ret)
2209                 return ret;
2210
2211         return 0;
2212 }
2213
2214 static int pnv_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
2215                                 unsigned int nr_irqs, void *arg)
2216 {
2217         struct pci_controller *hose = domain->host_data;
2218         struct pnv_phb *phb = hose->private_data;
2219         msi_alloc_info_t *info = arg;
2220         struct pci_dev *pdev = msi_desc_to_pci_dev(info->desc);
2221         int hwirq;
2222         int i, ret;
2223
2224         hwirq = msi_bitmap_alloc_hwirqs(&phb->msi_bmp, nr_irqs);
2225         if (hwirq < 0) {
2226                 dev_warn(&pdev->dev, "failed to find a free MSI\n");
2227                 return -ENOSPC;
2228         }
2229
2230         dev_dbg(&pdev->dev, "%s bridge %pOF %d/%x #%d\n", __func__,
2231                 hose->dn, virq, hwirq, nr_irqs);
2232
2233         for (i = 0; i < nr_irqs; i++) {
2234                 ret = pnv_irq_parent_domain_alloc(domain, virq + i,
2235                                                   phb->msi_base + hwirq + i);
2236                 if (ret)
2237                         goto out;
2238
2239                 irq_domain_set_hwirq_and_chip(domain, virq + i, hwirq + i,
2240                                               &pnv_msi_irq_chip, hose);
2241         }
2242
2243         return 0;
2244
2245 out:
2246         irq_domain_free_irqs_parent(domain, virq, i - 1);
2247         msi_bitmap_free_hwirqs(&phb->msi_bmp, hwirq, nr_irqs);
2248         return ret;
2249 }
2250
2251 static void pnv_irq_domain_free(struct irq_domain *domain, unsigned int virq,
2252                                 unsigned int nr_irqs)
2253 {
2254         struct irq_data *d = irq_domain_get_irq_data(domain, virq);
2255         struct pci_controller *hose = irq_data_get_irq_chip_data(d);
2256         struct pnv_phb *phb = hose->private_data;
2257
2258         pr_debug("%s bridge %pOF %d/%lx #%d\n", __func__, hose->dn,
2259                  virq, d->hwirq, nr_irqs);
2260
2261         msi_bitmap_free_hwirqs(&phb->msi_bmp, d->hwirq, nr_irqs);
2262         /* XIVE domain is cleared through ->msi_free() */
2263 }
2264
2265 static const struct irq_domain_ops pnv_irq_domain_ops = {
2266         .alloc  = pnv_irq_domain_alloc,
2267         .free   = pnv_irq_domain_free,
2268 };
2269
2270 static int pnv_msi_allocate_domains(struct pci_controller *hose, unsigned int count)
2271 {
2272         struct pnv_phb *phb = hose->private_data;
2273         struct irq_domain *parent = irq_get_default_host();
2274
2275         hose->fwnode = irq_domain_alloc_named_id_fwnode("PNV-MSI", phb->opal_id);
2276         if (!hose->fwnode)
2277                 return -ENOMEM;
2278
2279         hose->dev_domain = irq_domain_create_hierarchy(parent, 0, count,
2280                                                        hose->fwnode,
2281                                                        &pnv_irq_domain_ops, hose);
2282         if (!hose->dev_domain) {
2283                 pr_err("PCI: failed to create IRQ domain bridge %pOF (domain %d)\n",
2284                        hose->dn, hose->global_number);
2285                 irq_domain_free_fwnode(hose->fwnode);
2286                 return -ENOMEM;
2287         }
2288
2289         hose->msi_domain = pci_msi_create_irq_domain(of_node_to_fwnode(hose->dn),
2290                                                      &pnv_msi_domain_info,
2291                                                      hose->dev_domain);
2292         if (!hose->msi_domain) {
2293                 pr_err("PCI: failed to create MSI IRQ domain bridge %pOF (domain %d)\n",
2294                        hose->dn, hose->global_number);
2295                 irq_domain_free_fwnode(hose->fwnode);
2296                 irq_domain_remove(hose->dev_domain);
2297                 return -ENOMEM;
2298         }
2299
2300         return 0;
2301 }
2302
2303 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb)
2304 {
2305         unsigned int count;
2306         const __be32 *prop = of_get_property(phb->hose->dn,
2307                                              "ibm,opal-msi-ranges", NULL);
2308         if (!prop) {
2309                 /* BML Fallback */
2310                 prop = of_get_property(phb->hose->dn, "msi-ranges", NULL);
2311         }
2312         if (!prop)
2313                 return;
2314
2315         phb->msi_base = be32_to_cpup(prop);
2316         count = be32_to_cpup(prop + 1);
2317         if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) {
2318                 pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
2319                        phb->hose->global_number);
2320                 return;
2321         }
2322
2323         pr_info("  Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
2324                 count, phb->msi_base);
2325
2326         pnv_msi_allocate_domains(phb->hose, count);
2327 }
2328
2329 static void pnv_ioda_setup_pe_res(struct pnv_ioda_pe *pe,
2330                                   struct resource *res)
2331 {
2332         struct pnv_phb *phb = pe->phb;
2333         struct pci_bus_region region;
2334         int index;
2335         int64_t rc;
2336
2337         if (!res || !res->flags || res->start > res->end)
2338                 return;
2339
2340         if (res->flags & IORESOURCE_IO) {
2341                 region.start = res->start - phb->ioda.io_pci_base;
2342                 region.end   = res->end - phb->ioda.io_pci_base;
2343                 index = region.start / phb->ioda.io_segsize;
2344
2345                 while (index < phb->ioda.total_pe_num &&
2346                        region.start <= region.end) {
2347                         phb->ioda.io_segmap[index] = pe->pe_number;
2348                         rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2349                                 pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index);
2350                         if (rc != OPAL_SUCCESS) {
2351                                 pr_err("%s: Error %lld mapping IO segment#%d to PE#%x\n",
2352                                        __func__, rc, index, pe->pe_number);
2353                                 break;
2354                         }
2355
2356                         region.start += phb->ioda.io_segsize;
2357                         index++;
2358                 }
2359         } else if ((res->flags & IORESOURCE_MEM) &&
2360                    !pnv_pci_is_m64(phb, res)) {
2361                 region.start = res->start -
2362                                phb->hose->mem_offset[0] -
2363                                phb->ioda.m32_pci_base;
2364                 region.end   = res->end -
2365                                phb->hose->mem_offset[0] -
2366                                phb->ioda.m32_pci_base;
2367                 index = region.start / phb->ioda.m32_segsize;
2368
2369                 while (index < phb->ioda.total_pe_num &&
2370                        region.start <= region.end) {
2371                         phb->ioda.m32_segmap[index] = pe->pe_number;
2372                         rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2373                                 pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index);
2374                         if (rc != OPAL_SUCCESS) {
2375                                 pr_err("%s: Error %lld mapping M32 segment#%d to PE#%x",
2376                                        __func__, rc, index, pe->pe_number);
2377                                 break;
2378                         }
2379
2380                         region.start += phb->ioda.m32_segsize;
2381                         index++;
2382                 }
2383         }
2384 }
2385
2386 /*
2387  * This function is supposed to be called on basis of PE from top
2388  * to bottom style. So the the I/O or MMIO segment assigned to
2389  * parent PE could be overridden by its child PEs if necessary.
2390  */
2391 static void pnv_ioda_setup_pe_seg(struct pnv_ioda_pe *pe)
2392 {
2393         struct pci_dev *pdev;
2394         int i;
2395
2396         /*
2397          * NOTE: We only care PCI bus based PE for now. For PCI
2398          * device based PE, for example SRIOV sensitive VF should
2399          * be figured out later.
2400          */
2401         BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)));
2402
2403         list_for_each_entry(pdev, &pe->pbus->devices, bus_list) {
2404                 for (i = 0; i <= PCI_ROM_RESOURCE; i++)
2405                         pnv_ioda_setup_pe_res(pe, &pdev->resource[i]);
2406
2407                 /*
2408                  * If the PE contains all subordinate PCI buses, the
2409                  * windows of the child bridges should be mapped to
2410                  * the PE as well.
2411                  */
2412                 if (!(pe->flags & PNV_IODA_PE_BUS_ALL) || !pci_is_bridge(pdev))
2413                         continue;
2414                 for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++)
2415                         pnv_ioda_setup_pe_res(pe,
2416                                 &pdev->resource[PCI_BRIDGE_RESOURCES + i]);
2417         }
2418 }
2419
2420 #ifdef CONFIG_DEBUG_FS
2421 static int pnv_pci_diag_data_set(void *data, u64 val)
2422 {
2423         struct pnv_phb *phb = data;
2424         s64 ret;
2425
2426         /* Retrieve the diag data from firmware */
2427         ret = opal_pci_get_phb_diag_data2(phb->opal_id, phb->diag_data,
2428                                           phb->diag_data_size);
2429         if (ret != OPAL_SUCCESS)
2430                 return -EIO;
2431
2432         /* Print the diag data to the kernel log */
2433         pnv_pci_dump_phb_diag_data(phb->hose, phb->diag_data);
2434         return 0;
2435 }
2436
2437 DEFINE_DEBUGFS_ATTRIBUTE(pnv_pci_diag_data_fops, NULL, pnv_pci_diag_data_set,
2438                          "%llu\n");
2439
2440 static int pnv_pci_ioda_pe_dump(void *data, u64 val)
2441 {
2442         struct pnv_phb *phb = data;
2443         int pe_num;
2444
2445         for (pe_num = 0; pe_num < phb->ioda.total_pe_num; pe_num++) {
2446                 struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_num];
2447
2448                 if (!test_bit(pe_num, phb->ioda.pe_alloc))
2449                         continue;
2450
2451                 pe_warn(pe, "rid: %04x dev count: %2d flags: %s%s%s%s%s%s\n",
2452                         pe->rid, pe->device_count,
2453                         (pe->flags & PNV_IODA_PE_DEV) ? "dev " : "",
2454                         (pe->flags & PNV_IODA_PE_BUS) ? "bus " : "",
2455                         (pe->flags & PNV_IODA_PE_BUS_ALL) ? "all " : "",
2456                         (pe->flags & PNV_IODA_PE_MASTER) ? "master " : "",
2457                         (pe->flags & PNV_IODA_PE_SLAVE) ? "slave " : "",
2458                         (pe->flags & PNV_IODA_PE_VF) ? "vf " : "");
2459         }
2460
2461         return 0;
2462 }
2463
2464 DEFINE_DEBUGFS_ATTRIBUTE(pnv_pci_ioda_pe_dump_fops, NULL,
2465                          pnv_pci_ioda_pe_dump, "%llu\n");
2466
2467 #endif /* CONFIG_DEBUG_FS */
2468
2469 static void pnv_pci_ioda_create_dbgfs(void)
2470 {
2471 #ifdef CONFIG_DEBUG_FS
2472         struct pci_controller *hose, *tmp;
2473         struct pnv_phb *phb;
2474         char name[16];
2475
2476         list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
2477                 phb = hose->private_data;
2478
2479                 sprintf(name, "PCI%04x", hose->global_number);
2480                 phb->dbgfs = debugfs_create_dir(name, arch_debugfs_dir);
2481
2482                 debugfs_create_file_unsafe("dump_diag_regs", 0200, phb->dbgfs,
2483                                            phb, &pnv_pci_diag_data_fops);
2484                 debugfs_create_file_unsafe("dump_ioda_pe_state", 0200, phb->dbgfs,
2485                                            phb, &pnv_pci_ioda_pe_dump_fops);
2486         }
2487 #endif /* CONFIG_DEBUG_FS */
2488 }
2489
2490 static void pnv_pci_enable_bridge(struct pci_bus *bus)
2491 {
2492         struct pci_dev *dev = bus->self;
2493         struct pci_bus *child;
2494
2495         /* Empty bus ? bail */
2496         if (list_empty(&bus->devices))
2497                 return;
2498
2499         /*
2500          * If there's a bridge associated with that bus enable it. This works
2501          * around races in the generic code if the enabling is done during
2502          * parallel probing. This can be removed once those races have been
2503          * fixed.
2504          */
2505         if (dev) {
2506                 int rc = pci_enable_device(dev);
2507                 if (rc)
2508                         pci_err(dev, "Error enabling bridge (%d)\n", rc);
2509                 pci_set_master(dev);
2510         }
2511
2512         /* Perform the same to child busses */
2513         list_for_each_entry(child, &bus->children, node)
2514                 pnv_pci_enable_bridge(child);
2515 }
2516
2517 static void pnv_pci_enable_bridges(void)
2518 {
2519         struct pci_controller *hose;
2520
2521         list_for_each_entry(hose, &hose_list, list_node)
2522                 pnv_pci_enable_bridge(hose->bus);
2523 }
2524
2525 static void pnv_pci_ioda_fixup(void)
2526 {
2527         pnv_pci_ioda_create_dbgfs();
2528
2529         pnv_pci_enable_bridges();
2530
2531 #ifdef CONFIG_EEH
2532         pnv_eeh_post_init();
2533 #endif
2534 }
2535
2536 /*
2537  * Returns the alignment for I/O or memory windows for P2P
2538  * bridges. That actually depends on how PEs are segmented.
2539  * For now, we return I/O or M32 segment size for PE sensitive
2540  * P2P bridges. Otherwise, the default values (4KiB for I/O,
2541  * 1MiB for memory) will be returned.
2542  *
2543  * The current PCI bus might be put into one PE, which was
2544  * create against the parent PCI bridge. For that case, we
2545  * needn't enlarge the alignment so that we can save some
2546  * resources.
2547  */
2548 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus,
2549                                                 unsigned long type)
2550 {
2551         struct pnv_phb *phb = pci_bus_to_pnvhb(bus);
2552         int num_pci_bridges = 0;
2553         struct pci_dev *bridge;
2554
2555         bridge = bus->self;
2556         while (bridge) {
2557                 if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) {
2558                         num_pci_bridges++;
2559                         if (num_pci_bridges >= 2)
2560                                 return 1;
2561                 }
2562
2563                 bridge = bridge->bus->self;
2564         }
2565
2566         /*
2567          * We fall back to M32 if M64 isn't supported. We enforce the M64
2568          * alignment for any 64-bit resource, PCIe doesn't care and
2569          * bridges only do 64-bit prefetchable anyway.
2570          */
2571         if (phb->ioda.m64_segsize && pnv_pci_is_m64_flags(type))
2572                 return phb->ioda.m64_segsize;
2573         if (type & IORESOURCE_MEM)
2574                 return phb->ioda.m32_segsize;
2575
2576         return phb->ioda.io_segsize;
2577 }
2578
2579 /*
2580  * We are updating root port or the upstream port of the
2581  * bridge behind the root port with PHB's windows in order
2582  * to accommodate the changes on required resources during
2583  * PCI (slot) hotplug, which is connected to either root
2584  * port or the downstream ports of PCIe switch behind the
2585  * root port.
2586  */
2587 static void pnv_pci_fixup_bridge_resources(struct pci_bus *bus,
2588                                            unsigned long type)
2589 {
2590         struct pci_controller *hose = pci_bus_to_host(bus);
2591         struct pnv_phb *phb = hose->private_data;
2592         struct pci_dev *bridge = bus->self;
2593         struct resource *r, *w;
2594         bool msi_region = false;
2595         int i;
2596
2597         /* Check if we need apply fixup to the bridge's windows */
2598         if (!pci_is_root_bus(bridge->bus) &&
2599             !pci_is_root_bus(bridge->bus->self->bus))
2600                 return;
2601
2602         /* Fixup the resources */
2603         for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++) {
2604                 r = &bridge->resource[PCI_BRIDGE_RESOURCES + i];
2605                 if (!r->flags || !r->parent)
2606                         continue;
2607
2608                 w = NULL;
2609                 if (r->flags & type & IORESOURCE_IO)
2610                         w = &hose->io_resource;
2611                 else if (pnv_pci_is_m64(phb, r) &&
2612                          (type & IORESOURCE_PREFETCH) &&
2613                          phb->ioda.m64_segsize)
2614                         w = &hose->mem_resources[1];
2615                 else if (r->flags & type & IORESOURCE_MEM) {
2616                         w = &hose->mem_resources[0];
2617                         msi_region = true;
2618                 }
2619
2620                 r->start = w->start;
2621                 r->end = w->end;
2622
2623                 /* The 64KB 32-bits MSI region shouldn't be included in
2624                  * the 32-bits bridge window. Otherwise, we can see strange
2625                  * issues. One of them is EEH error observed on Garrison.
2626                  *
2627                  * Exclude top 1MB region which is the minimal alignment of
2628                  * 32-bits bridge window.
2629                  */
2630                 if (msi_region) {
2631                         r->end += 0x10000;
2632                         r->end -= 0x100000;
2633                 }
2634         }
2635 }
2636
2637 static void pnv_pci_configure_bus(struct pci_bus *bus)
2638 {
2639         struct pci_dev *bridge = bus->self;
2640         struct pnv_ioda_pe *pe;
2641         bool all = (bridge && pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE);
2642
2643         dev_info(&bus->dev, "Configuring PE for bus\n");
2644
2645         /* Don't assign PE to PCI bus, which doesn't have subordinate devices */
2646         if (WARN_ON(list_empty(&bus->devices)))
2647                 return;
2648
2649         /* Reserve PEs according to used M64 resources */
2650         pnv_ioda_reserve_m64_pe(bus, NULL, all);
2651
2652         /*
2653          * Assign PE. We might run here because of partial hotplug.
2654          * For the case, we just pick up the existing PE and should
2655          * not allocate resources again.
2656          */
2657         pe = pnv_ioda_setup_bus_PE(bus, all);
2658         if (!pe)
2659                 return;
2660
2661         pnv_ioda_setup_pe_seg(pe);
2662 }
2663
2664 static resource_size_t pnv_pci_default_alignment(void)
2665 {
2666         return PAGE_SIZE;
2667 }
2668
2669 /* Prevent enabling devices for which we couldn't properly
2670  * assign a PE
2671  */
2672 static bool pnv_pci_enable_device_hook(struct pci_dev *dev)
2673 {
2674         struct pci_dn *pdn;
2675
2676         pdn = pci_get_pdn(dev);
2677         if (!pdn || pdn->pe_number == IODA_INVALID_PE) {
2678                 pci_err(dev, "pci_enable_device() blocked, no PE assigned.\n");
2679                 return false;
2680         }
2681
2682         return true;
2683 }
2684
2685 static bool pnv_ocapi_enable_device_hook(struct pci_dev *dev)
2686 {
2687         struct pci_dn *pdn;
2688         struct pnv_ioda_pe *pe;
2689
2690         pdn = pci_get_pdn(dev);
2691         if (!pdn)
2692                 return false;
2693
2694         if (pdn->pe_number == IODA_INVALID_PE) {
2695                 pe = pnv_ioda_setup_dev_PE(dev);
2696                 if (!pe)
2697                         return false;
2698         }
2699         return true;
2700 }
2701
2702 static long pnv_pci_ioda1_unset_window(struct iommu_table_group *table_group,
2703                                        int num)
2704 {
2705         struct pnv_ioda_pe *pe = container_of(table_group,
2706                                               struct pnv_ioda_pe, table_group);
2707         struct pnv_phb *phb = pe->phb;
2708         unsigned int idx;
2709         long rc;
2710
2711         pe_info(pe, "Removing DMA window #%d\n", num);
2712         for (idx = 0; idx < phb->ioda.dma32_count; idx++) {
2713                 if (phb->ioda.dma32_segmap[idx] != pe->pe_number)
2714                         continue;
2715
2716                 rc = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
2717                                                 idx, 0, 0ul, 0ul, 0ul);
2718                 if (rc != OPAL_SUCCESS) {
2719                         pe_warn(pe, "Failure %ld unmapping DMA32 segment#%d\n",
2720                                 rc, idx);
2721                         return rc;
2722                 }
2723
2724                 phb->ioda.dma32_segmap[idx] = IODA_INVALID_PE;
2725         }
2726
2727         pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
2728         return OPAL_SUCCESS;
2729 }
2730
2731 static void pnv_pci_ioda1_release_pe_dma(struct pnv_ioda_pe *pe)
2732 {
2733         struct iommu_table *tbl = pe->table_group.tables[0];
2734         int64_t rc;
2735
2736         if (!pe->dma_setup_done)
2737                 return;
2738
2739         rc = pnv_pci_ioda1_unset_window(&pe->table_group, 0);
2740         if (rc != OPAL_SUCCESS)
2741                 return;
2742
2743         pnv_pci_p7ioc_tce_invalidate(tbl, tbl->it_offset, tbl->it_size, false);
2744         if (pe->table_group.group) {
2745                 iommu_group_put(pe->table_group.group);
2746                 WARN_ON(pe->table_group.group);
2747         }
2748
2749         free_pages(tbl->it_base, get_order(tbl->it_size << 3));
2750         iommu_tce_table_put(tbl);
2751 }
2752
2753 void pnv_pci_ioda2_release_pe_dma(struct pnv_ioda_pe *pe)
2754 {
2755         struct iommu_table *tbl = pe->table_group.tables[0];
2756         int64_t rc;
2757
2758         if (!pe->dma_setup_done)
2759                 return;
2760
2761         rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
2762         if (rc)
2763                 pe_warn(pe, "OPAL error %lld release DMA window\n", rc);
2764
2765         pnv_pci_ioda2_set_bypass(pe, false);
2766         if (pe->table_group.group) {
2767                 iommu_group_put(pe->table_group.group);
2768                 WARN_ON(pe->table_group.group);
2769         }
2770
2771         iommu_tce_table_put(tbl);
2772 }
2773
2774 static void pnv_ioda_free_pe_seg(struct pnv_ioda_pe *pe,
2775                                  unsigned short win,
2776                                  unsigned int *map)
2777 {
2778         struct pnv_phb *phb = pe->phb;
2779         int idx;
2780         int64_t rc;
2781
2782         for (idx = 0; idx < phb->ioda.total_pe_num; idx++) {
2783                 if (map[idx] != pe->pe_number)
2784                         continue;
2785
2786                 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2787                                 phb->ioda.reserved_pe_idx, win, 0, idx);
2788
2789                 if (rc != OPAL_SUCCESS)
2790                         pe_warn(pe, "Error %lld unmapping (%d) segment#%d\n",
2791                                 rc, win, idx);
2792
2793                 map[idx] = IODA_INVALID_PE;
2794         }
2795 }
2796
2797 static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe *pe)
2798 {
2799         struct pnv_phb *phb = pe->phb;
2800
2801         if (phb->type == PNV_PHB_IODA1) {
2802                 pnv_ioda_free_pe_seg(pe, OPAL_IO_WINDOW_TYPE,
2803                                      phb->ioda.io_segmap);
2804                 pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
2805                                      phb->ioda.m32_segmap);
2806                 /* M64 is pre-configured by pnv_ioda1_init_m64() */
2807         } else if (phb->type == PNV_PHB_IODA2) {
2808                 pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
2809                                      phb->ioda.m32_segmap);
2810         }
2811 }
2812
2813 static void pnv_ioda_release_pe(struct pnv_ioda_pe *pe)
2814 {
2815         struct pnv_phb *phb = pe->phb;
2816         struct pnv_ioda_pe *slave, *tmp;
2817
2818         pe_info(pe, "Releasing PE\n");
2819
2820         mutex_lock(&phb->ioda.pe_list_mutex);
2821         list_del(&pe->list);
2822         mutex_unlock(&phb->ioda.pe_list_mutex);
2823
2824         switch (phb->type) {
2825         case PNV_PHB_IODA1:
2826                 pnv_pci_ioda1_release_pe_dma(pe);
2827                 break;
2828         case PNV_PHB_IODA2:
2829                 pnv_pci_ioda2_release_pe_dma(pe);
2830                 break;
2831         case PNV_PHB_NPU_OCAPI:
2832                 break;
2833         default:
2834                 WARN_ON(1);
2835         }
2836
2837         pnv_ioda_release_pe_seg(pe);
2838         pnv_ioda_deconfigure_pe(pe->phb, pe);
2839
2840         /* Release slave PEs in the compound PE */
2841         if (pe->flags & PNV_IODA_PE_MASTER) {
2842                 list_for_each_entry_safe(slave, tmp, &pe->slaves, list) {
2843                         list_del(&slave->list);
2844                         pnv_ioda_free_pe(slave);
2845                 }
2846         }
2847
2848         /*
2849          * The PE for root bus can be removed because of hotplug in EEH
2850          * recovery for fenced PHB error. We need to mark the PE dead so
2851          * that it can be populated again in PCI hot add path. The PE
2852          * shouldn't be destroyed as it's the global reserved resource.
2853          */
2854         if (phb->ioda.root_pe_idx == pe->pe_number)
2855                 return;
2856
2857         pnv_ioda_free_pe(pe);
2858 }
2859
2860 static void pnv_pci_release_device(struct pci_dev *pdev)
2861 {
2862         struct pnv_phb *phb = pci_bus_to_pnvhb(pdev->bus);
2863         struct pci_dn *pdn = pci_get_pdn(pdev);
2864         struct pnv_ioda_pe *pe;
2865
2866         /* The VF PE state is torn down when sriov_disable() is called */
2867         if (pdev->is_virtfn)
2868                 return;
2869
2870         if (!pdn || pdn->pe_number == IODA_INVALID_PE)
2871                 return;
2872
2873 #ifdef CONFIG_PCI_IOV
2874         /*
2875          * FIXME: Try move this to sriov_disable(). It's here since we allocate
2876          * the iov state at probe time since we need to fiddle with the IOV
2877          * resources.
2878          */
2879         if (pdev->is_physfn)
2880                 kfree(pdev->dev.archdata.iov_data);
2881 #endif
2882
2883         /*
2884          * PCI hotplug can happen as part of EEH error recovery. The @pdn
2885          * isn't removed and added afterwards in this scenario. We should
2886          * set the PE number in @pdn to an invalid one. Otherwise, the PE's
2887          * device count is decreased on removing devices while failing to
2888          * be increased on adding devices. It leads to unbalanced PE's device
2889          * count and eventually make normal PCI hotplug path broken.
2890          */
2891         pe = &phb->ioda.pe_array[pdn->pe_number];
2892         pdn->pe_number = IODA_INVALID_PE;
2893
2894         WARN_ON(--pe->device_count < 0);
2895         if (pe->device_count == 0)
2896                 pnv_ioda_release_pe(pe);
2897 }
2898
2899 static void pnv_pci_ioda_shutdown(struct pci_controller *hose)
2900 {
2901         struct pnv_phb *phb = hose->private_data;
2902
2903         opal_pci_reset(phb->opal_id, OPAL_RESET_PCI_IODA_TABLE,
2904                        OPAL_ASSERT_RESET);
2905 }
2906
2907 static void pnv_pci_ioda_dma_bus_setup(struct pci_bus *bus)
2908 {
2909         struct pnv_phb *phb = pci_bus_to_pnvhb(bus);
2910         struct pnv_ioda_pe *pe;
2911
2912         list_for_each_entry(pe, &phb->ioda.pe_list, list) {
2913                 if (!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)))
2914                         continue;
2915
2916                 if (!pe->pbus)
2917                         continue;
2918
2919                 if (bus->number == ((pe->rid >> 8) & 0xFF)) {
2920                         pe->pbus = bus;
2921                         break;
2922                 }
2923         }
2924 }
2925
2926 static const struct pci_controller_ops pnv_pci_ioda_controller_ops = {
2927         .dma_dev_setup          = pnv_pci_ioda_dma_dev_setup,
2928         .dma_bus_setup          = pnv_pci_ioda_dma_bus_setup,
2929         .iommu_bypass_supported = pnv_pci_ioda_iommu_bypass_supported,
2930         .enable_device_hook     = pnv_pci_enable_device_hook,
2931         .release_device         = pnv_pci_release_device,
2932         .window_alignment       = pnv_pci_window_alignment,
2933         .setup_bridge           = pnv_pci_fixup_bridge_resources,
2934         .reset_secondary_bus    = pnv_pci_reset_secondary_bus,
2935         .shutdown               = pnv_pci_ioda_shutdown,
2936 };
2937
2938 static const struct pci_controller_ops pnv_npu_ocapi_ioda_controller_ops = {
2939         .enable_device_hook     = pnv_ocapi_enable_device_hook,
2940         .release_device         = pnv_pci_release_device,
2941         .window_alignment       = pnv_pci_window_alignment,
2942         .reset_secondary_bus    = pnv_pci_reset_secondary_bus,
2943         .shutdown               = pnv_pci_ioda_shutdown,
2944 };
2945
2946 static void __init pnv_pci_init_ioda_phb(struct device_node *np,
2947                                          u64 hub_id, int ioda_type)
2948 {
2949         struct pci_controller *hose;
2950         struct pnv_phb *phb;
2951         unsigned long size, m64map_off, m32map_off, pemap_off;
2952         unsigned long iomap_off = 0, dma32map_off = 0;
2953         struct pnv_ioda_pe *root_pe;
2954         struct resource r;
2955         const __be64 *prop64;
2956         const __be32 *prop32;
2957         int len;
2958         unsigned int segno;
2959         u64 phb_id;
2960         void *aux;
2961         long rc;
2962
2963         if (!of_device_is_available(np))
2964                 return;
2965
2966         pr_info("Initializing %s PHB (%pOF)\n", pnv_phb_names[ioda_type], np);
2967
2968         prop64 = of_get_property(np, "ibm,opal-phbid", NULL);
2969         if (!prop64) {
2970                 pr_err("  Missing \"ibm,opal-phbid\" property !\n");
2971                 return;
2972         }
2973         phb_id = be64_to_cpup(prop64);
2974         pr_debug("  PHB-ID  : 0x%016llx\n", phb_id);
2975
2976         phb = kzalloc(sizeof(*phb), GFP_KERNEL);
2977         if (!phb)
2978                 panic("%s: Failed to allocate %zu bytes\n", __func__,
2979                       sizeof(*phb));
2980
2981         /* Allocate PCI controller */
2982         phb->hose = hose = pcibios_alloc_controller(np);
2983         if (!phb->hose) {
2984                 pr_err("  Can't allocate PCI controller for %pOF\n",
2985                        np);
2986                 memblock_free(__pa(phb), sizeof(struct pnv_phb));
2987                 return;
2988         }
2989
2990         spin_lock_init(&phb->lock);
2991         prop32 = of_get_property(np, "bus-range", &len);
2992         if (prop32 && len == 8) {
2993                 hose->first_busno = be32_to_cpu(prop32[0]);
2994                 hose->last_busno = be32_to_cpu(prop32[1]);
2995         } else {
2996                 pr_warn("  Broken <bus-range> on %pOF\n", np);
2997                 hose->first_busno = 0;
2998                 hose->last_busno = 0xff;
2999         }
3000         hose->private_data = phb;
3001         phb->hub_id = hub_id;
3002         phb->opal_id = phb_id;
3003         phb->type = ioda_type;
3004         mutex_init(&phb->ioda.pe_alloc_mutex);
3005
3006         /* Detect specific models for error handling */
3007         if (of_device_is_compatible(np, "ibm,p7ioc-pciex"))
3008                 phb->model = PNV_PHB_MODEL_P7IOC;
3009         else if (of_device_is_compatible(np, "ibm,power8-pciex"))
3010                 phb->model = PNV_PHB_MODEL_PHB3;
3011         else
3012                 phb->model = PNV_PHB_MODEL_UNKNOWN;
3013
3014         /* Initialize diagnostic data buffer */
3015         prop32 = of_get_property(np, "ibm,phb-diag-data-size", NULL);
3016         if (prop32)
3017                 phb->diag_data_size = be32_to_cpup(prop32);
3018         else
3019                 phb->diag_data_size = PNV_PCI_DIAG_BUF_SIZE;
3020
3021         phb->diag_data = kzalloc(phb->diag_data_size, GFP_KERNEL);
3022         if (!phb->diag_data)
3023                 panic("%s: Failed to allocate %u bytes\n", __func__,
3024                       phb->diag_data_size);
3025
3026         /* Parse 32-bit and IO ranges (if any) */
3027         pci_process_bridge_OF_ranges(hose, np, !hose->global_number);
3028
3029         /* Get registers */
3030         if (!of_address_to_resource(np, 0, &r)) {
3031                 phb->regs_phys = r.start;
3032                 phb->regs = ioremap(r.start, resource_size(&r));
3033                 if (phb->regs == NULL)
3034                         pr_err("  Failed to map registers !\n");
3035         }
3036
3037         /* Initialize more IODA stuff */
3038         phb->ioda.total_pe_num = 1;
3039         prop32 = of_get_property(np, "ibm,opal-num-pes", NULL);
3040         if (prop32)
3041                 phb->ioda.total_pe_num = be32_to_cpup(prop32);
3042         prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL);
3043         if (prop32)
3044                 phb->ioda.reserved_pe_idx = be32_to_cpup(prop32);
3045
3046         /* Invalidate RID to PE# mapping */
3047         for (segno = 0; segno < ARRAY_SIZE(phb->ioda.pe_rmap); segno++)
3048                 phb->ioda.pe_rmap[segno] = IODA_INVALID_PE;
3049
3050         /* Parse 64-bit MMIO range */
3051         pnv_ioda_parse_m64_window(phb);
3052
3053         phb->ioda.m32_size = resource_size(&hose->mem_resources[0]);
3054         /* FW Has already off top 64k of M32 space (MSI space) */
3055         phb->ioda.m32_size += 0x10000;
3056
3057         phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe_num;
3058         phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0];
3059         phb->ioda.io_size = hose->pci_io_size;
3060         phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe_num;
3061         phb->ioda.io_pci_base = 0; /* XXX calculate this ? */
3062
3063         /* Calculate how many 32-bit TCE segments we have */
3064         phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3065                                 PNV_IODA1_DMA32_SEGSIZE;
3066
3067         /* Allocate aux data & arrays. We don't have IO ports on PHB3 */
3068         size = ALIGN(max_t(unsigned, phb->ioda.total_pe_num, 8) / 8,
3069                         sizeof(unsigned long));
3070         m64map_off = size;
3071         size += phb->ioda.total_pe_num * sizeof(phb->ioda.m64_segmap[0]);
3072         m32map_off = size;
3073         size += phb->ioda.total_pe_num * sizeof(phb->ioda.m32_segmap[0]);
3074         if (phb->type == PNV_PHB_IODA1) {
3075                 iomap_off = size;
3076                 size += phb->ioda.total_pe_num * sizeof(phb->ioda.io_segmap[0]);
3077                 dma32map_off = size;
3078                 size += phb->ioda.dma32_count *
3079                         sizeof(phb->ioda.dma32_segmap[0]);
3080         }
3081         pemap_off = size;
3082         size += phb->ioda.total_pe_num * sizeof(struct pnv_ioda_pe);
3083         aux = kzalloc(size, GFP_KERNEL);
3084         if (!aux)
3085                 panic("%s: Failed to allocate %lu bytes\n", __func__, size);
3086
3087         phb->ioda.pe_alloc = aux;
3088         phb->ioda.m64_segmap = aux + m64map_off;
3089         phb->ioda.m32_segmap = aux + m32map_off;
3090         for (segno = 0; segno < phb->ioda.total_pe_num; segno++) {
3091                 phb->ioda.m64_segmap[segno] = IODA_INVALID_PE;
3092                 phb->ioda.m32_segmap[segno] = IODA_INVALID_PE;
3093         }
3094         if (phb->type == PNV_PHB_IODA1) {
3095                 phb->ioda.io_segmap = aux + iomap_off;
3096                 for (segno = 0; segno < phb->ioda.total_pe_num; segno++)
3097                         phb->ioda.io_segmap[segno] = IODA_INVALID_PE;
3098
3099                 phb->ioda.dma32_segmap = aux + dma32map_off;
3100                 for (segno = 0; segno < phb->ioda.dma32_count; segno++)
3101                         phb->ioda.dma32_segmap[segno] = IODA_INVALID_PE;
3102         }
3103         phb->ioda.pe_array = aux + pemap_off;
3104
3105         /*
3106          * Choose PE number for root bus, which shouldn't have
3107          * M64 resources consumed by its child devices. To pick
3108          * the PE number adjacent to the reserved one if possible.
3109          */
3110         pnv_ioda_reserve_pe(phb, phb->ioda.reserved_pe_idx);
3111         if (phb->ioda.reserved_pe_idx == 0) {
3112                 phb->ioda.root_pe_idx = 1;
3113                 pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3114         } else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1)) {
3115                 phb->ioda.root_pe_idx = phb->ioda.reserved_pe_idx - 1;
3116                 pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3117         } else {
3118                 /* otherwise just allocate one */
3119                 root_pe = pnv_ioda_alloc_pe(phb, 1);
3120                 phb->ioda.root_pe_idx = root_pe->pe_number;
3121         }
3122
3123         INIT_LIST_HEAD(&phb->ioda.pe_list);
3124         mutex_init(&phb->ioda.pe_list_mutex);
3125
3126         /* Calculate how many 32-bit TCE segments we have */
3127         phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3128                                 PNV_IODA1_DMA32_SEGSIZE;
3129
3130 #if 0 /* We should really do that ... */
3131         rc = opal_pci_set_phb_mem_window(opal->phb_id,
3132                                          window_type,
3133                                          window_num,
3134                                          starting_real_address,
3135                                          starting_pci_address,
3136                                          segment_size);
3137 #endif
3138
3139         pr_info("  %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
3140                 phb->ioda.total_pe_num, phb->ioda.reserved_pe_idx,
3141                 phb->ioda.m32_size, phb->ioda.m32_segsize);
3142         if (phb->ioda.m64_size)
3143                 pr_info("                 M64: 0x%lx [segment=0x%lx]\n",
3144                         phb->ioda.m64_size, phb->ioda.m64_segsize);
3145         if (phb->ioda.io_size)
3146                 pr_info("                  IO: 0x%x [segment=0x%x]\n",
3147                         phb->ioda.io_size, phb->ioda.io_segsize);
3148
3149
3150         phb->hose->ops = &pnv_pci_ops;
3151         phb->get_pe_state = pnv_ioda_get_pe_state;
3152         phb->freeze_pe = pnv_ioda_freeze_pe;
3153         phb->unfreeze_pe = pnv_ioda_unfreeze_pe;
3154
3155         /* Setup MSI support */
3156         pnv_pci_init_ioda_msis(phb);
3157
3158         /*
3159          * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
3160          * to let the PCI core do resource assignment. It's supposed
3161          * that the PCI core will do correct I/O and MMIO alignment
3162          * for the P2P bridge bars so that each PCI bus (excluding
3163          * the child P2P bridges) can form individual PE.
3164          */
3165         ppc_md.pcibios_fixup = pnv_pci_ioda_fixup;
3166
3167         switch (phb->type) {
3168         case PNV_PHB_NPU_OCAPI:
3169                 hose->controller_ops = pnv_npu_ocapi_ioda_controller_ops;
3170                 break;
3171         default:
3172                 hose->controller_ops = pnv_pci_ioda_controller_ops;
3173         }
3174
3175         ppc_md.pcibios_default_alignment = pnv_pci_default_alignment;
3176
3177 #ifdef CONFIG_PCI_IOV
3178         ppc_md.pcibios_fixup_sriov = pnv_pci_ioda_fixup_iov;
3179         ppc_md.pcibios_iov_resource_alignment = pnv_pci_iov_resource_alignment;
3180         ppc_md.pcibios_sriov_enable = pnv_pcibios_sriov_enable;
3181         ppc_md.pcibios_sriov_disable = pnv_pcibios_sriov_disable;
3182 #endif
3183
3184         pci_add_flags(PCI_REASSIGN_ALL_RSRC);
3185
3186         /* Reset IODA tables to a clean state */
3187         rc = opal_pci_reset(phb_id, OPAL_RESET_PCI_IODA_TABLE, OPAL_ASSERT_RESET);
3188         if (rc)
3189                 pr_warn("  OPAL Error %ld performing IODA table reset !\n", rc);
3190
3191         /*
3192          * If we're running in kdump kernel, the previous kernel never
3193          * shutdown PCI devices correctly. We already got IODA table
3194          * cleaned out. So we have to issue PHB reset to stop all PCI
3195          * transactions from previous kernel. The ppc_pci_reset_phbs
3196          * kernel parameter will force this reset too. Additionally,
3197          * if the IODA reset above failed then use a bigger hammer.
3198          * This can happen if we get a PHB fatal error in very early
3199          * boot.
3200          */
3201         if (is_kdump_kernel() || pci_reset_phbs || rc) {
3202                 pr_info("  Issue PHB reset ...\n");
3203                 pnv_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL);
3204                 pnv_eeh_phb_reset(hose, EEH_RESET_DEACTIVATE);
3205         }
3206
3207         /* Remove M64 resource if we can't configure it successfully */
3208         if (!phb->init_m64 || phb->init_m64(phb))
3209                 hose->mem_resources[1].flags = 0;
3210
3211         /* create pci_dn's for DT nodes under this PHB */
3212         pci_devs_phb_init_dynamic(hose);
3213 }
3214
3215 void __init pnv_pci_init_ioda2_phb(struct device_node *np)
3216 {
3217         pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2);
3218 }
3219
3220 void __init pnv_pci_init_npu2_opencapi_phb(struct device_node *np)
3221 {
3222         pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU_OCAPI);
3223 }
3224
3225 static void pnv_npu2_opencapi_cfg_size_fixup(struct pci_dev *dev)
3226 {
3227         struct pnv_phb *phb = pci_bus_to_pnvhb(dev->bus);
3228
3229         if (!machine_is(powernv))
3230                 return;
3231
3232         if (phb->type == PNV_PHB_NPU_OCAPI)
3233                 dev->cfg_size = PCI_CFG_SPACE_EXP_SIZE;
3234 }
3235 DECLARE_PCI_FIXUP_EARLY(PCI_ANY_ID, PCI_ANY_ID, pnv_npu2_opencapi_cfg_size_fixup);
3236
3237 void __init pnv_pci_init_ioda_hub(struct device_node *np)
3238 {
3239         struct device_node *phbn;
3240         const __be64 *prop64;
3241         u64 hub_id;
3242
3243         pr_info("Probing IODA IO-Hub %pOF\n", np);
3244
3245         prop64 = of_get_property(np, "ibm,opal-hubid", NULL);
3246         if (!prop64) {
3247                 pr_err(" Missing \"ibm,opal-hubid\" property !\n");
3248                 return;
3249         }
3250         hub_id = be64_to_cpup(prop64);
3251         pr_devel(" HUB-ID : 0x%016llx\n", hub_id);
3252
3253         /* Count child PHBs */
3254         for_each_child_of_node(np, phbn) {
3255                 /* Look for IODA1 PHBs */
3256                 if (of_device_is_compatible(phbn, "ibm,ioda-phb"))
3257                         pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1);
3258         }
3259 }