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 #include <linux/of_address.h>
  25 #include <linux/of_irq.h>
  26
  27 #include <asm/sections.h>
  28 #include <asm/io.h>
  29 #include <asm/pci-bridge.h>
  30 #include <asm/machdep.h>
  31 #include <asm/msi_bitmap.h>
  32 #include <asm/ppc-pci.h>
  33 #include <asm/opal.h>
  34 #include <asm/iommu.h>
  35 #include <asm/tce.h>
  36 #include <asm/xics.h>
  37 #include <asm/firmware.h>
  38 #include <asm/pnv-pci.h>
  39 #include <asm/mmzone.h>
  40 #include <asm/xive.h>
  41
  42 #include <misc/cxl-base.h>
  43
  44 #include "powernv.h"
  45 #include "pci.h"
  46 #include "../../../../drivers/pci/pci.h"
  47
  48 #define PNV_IODA1_M64_NUM       16      /* Number of M64 BARs   */
  49 #define PNV_IODA1_M64_SEGS      8       /* Segments per M64 BAR */
  50 #define PNV_IODA1_DMA32_SEGSIZE 0x10000000
  51
  52 static const char * const pnv_phb_names[] = { "IODA1", "IODA2", "NPU_OCAPI" };
  53
  54 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable);
  55 static void pnv_pci_configure_bus(struct pci_bus *bus);
  56
  57 void pe_level_printk(const struct pnv_ioda_pe *pe, const char *level,
  58                             const char *fmt, ...)
  59 {
  60         struct va_format vaf;
  61         va_list args;
  62         char pfix[32];
  63
  64         va_start(args, fmt);
  65
  66         vaf.fmt = fmt;
  67         vaf.va = &args;
  68
  69         if (pe->flags & PNV_IODA_PE_DEV)
  70                 strscpy(pfix, dev_name(&pe->pdev->dev), sizeof(pfix));
  71         else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
  72                 sprintf(pfix, "%04x:%02x     ",
  73                         pci_domain_nr(pe->pbus), pe->pbus->number);
  74 #ifdef CONFIG_PCI_IOV
  75         else if (pe->flags & PNV_IODA_PE_VF)
  76                 sprintf(pfix, "%04x:%02x:%2x.%d",
  77                         pci_domain_nr(pe->parent_dev->bus),
  78                         (pe->rid & 0xff00) >> 8,
  79                         PCI_SLOT(pe->rid), PCI_FUNC(pe->rid));
  80 #endif /* CONFIG_PCI_IOV*/
  81
  82         printk("%spci %s: [PE# %.2x] %pV",
  83                level, pfix, pe->pe_number, &vaf);
  84
  85         va_end(args);
  86 }
  87
  88 static bool pnv_iommu_bypass_disabled __read_mostly;
  89 static bool pci_reset_phbs __read_mostly;
  90
  91 static int __init iommu_setup(char *str)
  92 {
  93         if (!str)
  94                 return -EINVAL;
  95
  96         while (*str) {
  97                 if (!strncmp(str, "nobypass", 8)) {
  98                         pnv_iommu_bypass_disabled = true;
  99                         pr_info("PowerNV: IOMMU bypass window disabled.\n");
 100                         break;
 101                 }
 102                 str += strcspn(str, ",");
 103                 if (*str == ',')
 104                         str++;
 105         }
 106
 107         return 0;
 108 }
 109 early_param("iommu", iommu_setup);
 110
 111 static int __init pci_reset_phbs_setup(char *str)
 112 {
 113         pci_reset_phbs = true;
 114         return 0;
 115 }
 116
 117 early_param("ppc_pci_reset_phbs", pci_reset_phbs_setup);
 118
 119 static struct pnv_ioda_pe *pnv_ioda_init_pe(struct pnv_phb *phb, int pe_no)
 120 {
 121         s64 rc;
 122
 123         phb->ioda.pe_array[pe_no].phb = phb;
 124         phb->ioda.pe_array[pe_no].pe_number = pe_no;
 125         phb->ioda.pe_array[pe_no].dma_setup_done = false;
 126
 127         /*
 128          * Clear the PE frozen state as it might be put into frozen state
 129          * in the last PCI remove path. It's not harmful to do so when the
 130          * PE is already in unfrozen state.
 131          */
 132         rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no,
 133                                        OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 134         if (rc != OPAL_SUCCESS && rc != OPAL_UNSUPPORTED)
 135                 pr_warn("%s: Error %lld unfreezing PHB#%x-PE#%x\n",
 136                         __func__, rc, phb->hose->global_number, pe_no);
 137
 138         return &phb->ioda.pe_array[pe_no];
 139 }
 140
 141 static void pnv_ioda_reserve_pe(struct pnv_phb *phb, int pe_no)
 142 {
 143         if (!(pe_no >= 0 && pe_no < phb->ioda.total_pe_num)) {
 144                 pr_warn("%s: Invalid PE %x on PHB#%x\n",
 145                         __func__, pe_no, phb->hose->global_number);
 146                 return;
 147         }
 148
 149         mutex_lock(&phb->ioda.pe_alloc_mutex);
 150         if (test_and_set_bit(pe_no, phb->ioda.pe_alloc))
 151                 pr_debug("%s: PE %x was reserved on PHB#%x\n",
 152                          __func__, pe_no, phb->hose->global_number);
 153         mutex_unlock(&phb->ioda.pe_alloc_mutex);
 154
 155         pnv_ioda_init_pe(phb, pe_no);
 156 }
 157
 158 struct pnv_ioda_pe *pnv_ioda_alloc_pe(struct pnv_phb *phb, int count)
 159 {
 160         struct pnv_ioda_pe *ret = NULL;
 161         int run = 0, pe, i;
 162
 163         mutex_lock(&phb->ioda.pe_alloc_mutex);
 164
 165         /* scan backwards for a run of @count cleared bits */
 166         for (pe = phb->ioda.total_pe_num - 1; pe >= 0; pe--) {
 167                 if (test_bit(pe, phb->ioda.pe_alloc)) {
 168                         run = 0;
 169                         continue;
 170                 }
 171
 172                 run++;
 173                 if (run == count)
 174                         break;
 175         }
 176         if (run != count)
 177                 goto out;
 178
 179         for (i = pe; i < pe + count; i++) {
 180                 set_bit(i, phb->ioda.pe_alloc);
 181                 pnv_ioda_init_pe(phb, i);
 182         }
 183         ret = &phb->ioda.pe_array[pe];
 184
 185 out:
 186         mutex_unlock(&phb->ioda.pe_alloc_mutex);
 187         return ret;
 188 }
 189
 190 void pnv_ioda_free_pe(struct pnv_ioda_pe *pe)
 191 {
 192         struct pnv_phb *phb = pe->phb;
 193         unsigned int pe_num = pe->pe_number;
 194
 195         WARN_ON(pe->pdev);
 196         memset(pe, 0, sizeof(struct pnv_ioda_pe));
 197
 198         mutex_lock(&phb->ioda.pe_alloc_mutex);
 199         clear_bit(pe_num, phb->ioda.pe_alloc);
 200         mutex_unlock(&phb->ioda.pe_alloc_mutex);
 201 }
 202
 203 /* The default M64 BAR is shared by all PEs */
 204 static int pnv_ioda2_init_m64(struct pnv_phb *phb)
 205 {
 206         const char *desc;
 207         struct resource *r;
 208         s64 rc;
 209
 210         /* Configure the default M64 BAR */
 211         rc = opal_pci_set_phb_mem_window(phb->opal_id,
 212                                          OPAL_M64_WINDOW_TYPE,
 213                                          phb->ioda.m64_bar_idx,
 214                                          phb->ioda.m64_base,
 215                                          0, /* unused */
 216                                          phb->ioda.m64_size);
 217         if (rc != OPAL_SUCCESS) {
 218                 desc = "configuring";
 219                 goto fail;
 220         }
 221
 222         /* Enable the default M64 BAR */
 223         rc = opal_pci_phb_mmio_enable(phb->opal_id,
 224                                       OPAL_M64_WINDOW_TYPE,
 225                                       phb->ioda.m64_bar_idx,
 226                                       OPAL_ENABLE_M64_SPLIT);
 227         if (rc != OPAL_SUCCESS) {
 228                 desc = "enabling";
 229                 goto fail;
 230         }
 231
 232         /*
 233          * Exclude the segments for reserved and root bus PE, which
 234          * are first or last two PEs.
 235          */
 236         r = &phb->hose->mem_resources[1];
 237         if (phb->ioda.reserved_pe_idx == 0)
 238                 r->start += (2 * phb->ioda.m64_segsize);
 239         else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
 240                 r->end -= (2 * phb->ioda.m64_segsize);
 241         else
 242                 pr_warn("  Cannot strip M64 segment for reserved PE#%x\n",
 243                         phb->ioda.reserved_pe_idx);
 244
 245         return 0;
 246
 247 fail:
 248         pr_warn("  Failure %lld %s M64 BAR#%d\n",
 249                 rc, desc, phb->ioda.m64_bar_idx);
 250         opal_pci_phb_mmio_enable(phb->opal_id,
 251                                  OPAL_M64_WINDOW_TYPE,
 252                                  phb->ioda.m64_bar_idx,
 253                                  OPAL_DISABLE_M64);
 254         return -EIO;
 255 }
 256
 257 static void pnv_ioda_reserve_dev_m64_pe(struct pci_dev *pdev,
 258                                          unsigned long *pe_bitmap)
 259 {
 260         struct pnv_phb *phb = pci_bus_to_pnvhb(pdev->bus);
 261         struct resource *r;
 262         resource_size_t base, sgsz, start, end;
 263         int segno, i;
 264
 265         base = phb->ioda.m64_base;
 266         sgsz = phb->ioda.m64_segsize;
 267         for (i = 0; i <= PCI_ROM_RESOURCE; i++) {
 268                 r = &pdev->resource[i];
 269                 if (!r->parent || !pnv_pci_is_m64(phb, r))
 270                         continue;
 271
 272                 start = ALIGN_DOWN(r->start - base, sgsz);
 273                 end = ALIGN(r->end - base, sgsz);
 274                 for (segno = start / sgsz; segno < end / sgsz; segno++) {
 275                         if (pe_bitmap)
 276                                 set_bit(segno, pe_bitmap);
 277                         else
 278                                 pnv_ioda_reserve_pe(phb, segno);
 279                 }
 280         }
 281 }
 282
 283 static int pnv_ioda1_init_m64(struct pnv_phb *phb)
 284 {
 285         struct resource *r;
 286         int index;
 287
 288         /*
 289          * There are 16 M64 BARs, each of which has 8 segments. So
 290          * there are as many M64 segments as the maximum number of
 291          * PEs, which is 128.
 292          */
 293         for (index = 0; index < PNV_IODA1_M64_NUM; index++) {
 294                 unsigned long base, segsz = phb->ioda.m64_segsize;
 295                 int64_t rc;
 296
 297                 base = phb->ioda.m64_base +
 298                        index * PNV_IODA1_M64_SEGS * segsz;
 299                 rc = opal_pci_set_phb_mem_window(phb->opal_id,
 300                                 OPAL_M64_WINDOW_TYPE, index, base, 0,
 301                                 PNV_IODA1_M64_SEGS * segsz);
 302                 if (rc != OPAL_SUCCESS) {
 303                         pr_warn("  Error %lld setting M64 PHB#%x-BAR#%d\n",
 304                                 rc, phb->hose->global_number, index);
 305                         goto fail;
 306                 }
 307
 308                 rc = opal_pci_phb_mmio_enable(phb->opal_id,
 309                                 OPAL_M64_WINDOW_TYPE, index,
 310                                 OPAL_ENABLE_M64_SPLIT);
 311                 if (rc != OPAL_SUCCESS) {
 312                         pr_warn("  Error %lld enabling M64 PHB#%x-BAR#%d\n",
 313                                 rc, phb->hose->global_number, index);
 314                         goto fail;
 315                 }
 316         }
 317
 318         for (index = 0; index < phb->ioda.total_pe_num; index++) {
 319                 int64_t rc;
 320
 321                 /*
 322                  * P7IOC supports M64DT, which helps mapping M64 segment
 323                  * to one particular PE#. However, PHB3 has fixed mapping
 324                  * between M64 segment and PE#. In order to have same logic
 325                  * for P7IOC and PHB3, we enforce fixed mapping between M64
 326                  * segment and PE# on P7IOC.
 327                  */
 328                 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
 329                                 index, OPAL_M64_WINDOW_TYPE,
 330                                 index / PNV_IODA1_M64_SEGS,
 331                                 index % PNV_IODA1_M64_SEGS);
 332                 if (rc != OPAL_SUCCESS) {
 333                         pr_warn("%s: Error %lld mapping M64 for PHB#%x-PE#%x\n",
 334                                 __func__, rc, phb->hose->global_number,
 335                                 index);
 336                         goto fail;
 337                 }
 338         }
 339
 340         /*
 341          * Exclude the segments for reserved and root bus PE, which
 342          * are first or last two PEs.
 343          */
 344         r = &phb->hose->mem_resources[1];
 345         if (phb->ioda.reserved_pe_idx == 0)
 346                 r->start += (2 * phb->ioda.m64_segsize);
 347         else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
 348                 r->end -= (2 * phb->ioda.m64_segsize);
 349         else
 350                 WARN(1, "Wrong reserved PE#%x on PHB#%x\n",
 351                      phb->ioda.reserved_pe_idx, phb->hose->global_number);
 352
 353         return 0;
 354
 355 fail:
 356         for ( ; index >= 0; index--)
 357                 opal_pci_phb_mmio_enable(phb->opal_id,
 358                         OPAL_M64_WINDOW_TYPE, index, OPAL_DISABLE_M64);
 359
 360         return -EIO;
 361 }
 362
 363 static void pnv_ioda_reserve_m64_pe(struct pci_bus *bus,
 364                                     unsigned long *pe_bitmap,
 365                                     bool all)
 366 {
 367         struct pci_dev *pdev;
 368
 369         list_for_each_entry(pdev, &bus->devices, bus_list) {
 370                 pnv_ioda_reserve_dev_m64_pe(pdev, pe_bitmap);
 371
 372                 if (all && pdev->subordinate)
 373                         pnv_ioda_reserve_m64_pe(pdev->subordinate,
 374                                                 pe_bitmap, all);
 375         }
 376 }
 377
 378 static struct pnv_ioda_pe *pnv_ioda_pick_m64_pe(struct pci_bus *bus, bool all)
 379 {
 380         struct pnv_phb *phb = pci_bus_to_pnvhb(bus);
 381         struct pnv_ioda_pe *master_pe, *pe;
 382         unsigned long size, *pe_alloc;
 383         int i;
 384
 385         /* Root bus shouldn't use M64 */
 386         if (pci_is_root_bus(bus))
 387                 return NULL;
 388
 389         /* Allocate bitmap */
 390         size = ALIGN(phb->ioda.total_pe_num / 8, sizeof(unsigned long));
 391         pe_alloc = kzalloc(size, GFP_KERNEL);
 392         if (!pe_alloc) {
 393                 pr_warn("%s: Out of memory !\n",
 394                         __func__);
 395                 return NULL;
 396         }
 397
 398         /* Figure out reserved PE numbers by the PE */
 399         pnv_ioda_reserve_m64_pe(bus, pe_alloc, all);
 400
 401         /*
 402          * the current bus might not own M64 window and that's all
 403          * contributed by its child buses. For the case, we needn't
 404          * pick M64 dependent PE#.
 405          */
 406         if (bitmap_empty(pe_alloc, phb->ioda.total_pe_num)) {
 407                 kfree(pe_alloc);
 408                 return NULL;
 409         }
 410
 411         /*
 412          * Figure out the master PE and put all slave PEs to master
 413          * PE's list to form compound PE.
 414          */
 415         master_pe = NULL;
 416         i = -1;
 417         while ((i = find_next_bit(pe_alloc, phb->ioda.total_pe_num, i + 1)) <
 418                 phb->ioda.total_pe_num) {
 419                 pe = &phb->ioda.pe_array[i];
 420
 421                 phb->ioda.m64_segmap[pe->pe_number] = pe->pe_number;
 422                 if (!master_pe) {
 423                         pe->flags |= PNV_IODA_PE_MASTER;
 424                         INIT_LIST_HEAD(&pe->slaves);
 425                         master_pe = pe;
 426                 } else {
 427                         pe->flags |= PNV_IODA_PE_SLAVE;
 428                         pe->master = master_pe;
 429                         list_add_tail(&pe->list, &master_pe->slaves);
 430                 }
 431         }
 432
 433         kfree(pe_alloc);
 434         return master_pe;
 435 }
 436
 437 static void __init pnv_ioda_parse_m64_window(struct pnv_phb *phb)
 438 {
 439         struct pci_controller *hose = phb->hose;
 440         struct device_node *dn = hose->dn;
 441         struct resource *res;
 442         u32 m64_range[2], i;
 443         const __be32 *r;
 444         u64 pci_addr;
 445
 446         if (phb->type != PNV_PHB_IODA1 && phb->type != PNV_PHB_IODA2) {
 447                 pr_info("  Not support M64 window\n");
 448                 return;
 449         }
 450
 451         if (!firmware_has_feature(FW_FEATURE_OPAL)) {
 452                 pr_info("  Firmware too old to support M64 window\n");
 453                 return;
 454         }
 455
 456         r = of_get_property(dn, "ibm,opal-m64-window", NULL);
 457         if (!r) {
 458                 pr_info("  No <ibm,opal-m64-window> on %pOF\n",
 459                         dn);
 460                 return;
 461         }
 462
 463         /*
 464          * Find the available M64 BAR range and pickup the last one for
 465          * covering the whole 64-bits space. We support only one range.
 466          */
 467         if (of_property_read_u32_array(dn, "ibm,opal-available-m64-ranges",
 468                                        m64_range, 2)) {
 469                 /* In absence of the property, assume 0..15 */
 470                 m64_range[0] = 0;
 471                 m64_range[1] = 16;
 472         }
 473         /* We only support 64 bits in our allocator */
 474         if (m64_range[1] > 63) {
 475                 pr_warn("%s: Limiting M64 range to 63 (from %d) on PHB#%x\n",
 476                         __func__, m64_range[1], phb->hose->global_number);
 477                 m64_range[1] = 63;
 478         }
 479         /* Empty range, no m64 */
 480         if (m64_range[1] <= m64_range[0]) {
 481                 pr_warn("%s: M64 empty, disabling M64 usage on PHB#%x\n",
 482                         __func__, phb->hose->global_number);
 483                 return;
 484         }
 485
 486         /* Configure M64 informations */
 487         res = &hose->mem_resources[1];
 488         res->name = dn->full_name;
 489         res->start = of_translate_address(dn, r + 2);
 490         res->end = res->start + of_read_number(r + 4, 2) - 1;
 491         res->flags = (IORESOURCE_MEM | IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
 492         pci_addr = of_read_number(r, 2);
 493         hose->mem_offset[1] = res->start - pci_addr;
 494
 495         phb->ioda.m64_size = resource_size(res);
 496         phb->ioda.m64_segsize = phb->ioda.m64_size / phb->ioda.total_pe_num;
 497         phb->ioda.m64_base = pci_addr;
 498
 499         /* This lines up nicely with the display from processing OF ranges */
 500         pr_info(" MEM 0x%016llx..0x%016llx -> 0x%016llx (M64 #%d..%d)\n",
 501                 res->start, res->end, pci_addr, m64_range[0],
 502                 m64_range[0] + m64_range[1] - 1);
 503
 504         /* Mark all M64 used up by default */
 505         phb->ioda.m64_bar_alloc = (unsigned long)-1;
 506
 507         /* Use last M64 BAR to cover M64 window */
 508         m64_range[1]--;
 509         phb->ioda.m64_bar_idx = m64_range[0] + m64_range[1];
 510
 511         pr_info(" Using M64 #%d as default window\n", phb->ioda.m64_bar_idx);
 512
 513         /* Mark remaining ones free */
 514         for (i = m64_range[0]; i < m64_range[1]; i++)
 515                 clear_bit(i, &phb->ioda.m64_bar_alloc);
 516
 517         /*
 518          * Setup init functions for M64 based on IODA version, IODA3 uses
 519          * the IODA2 code.
 520          */
 521         if (phb->type == PNV_PHB_IODA1)
 522                 phb->init_m64 = pnv_ioda1_init_m64;
 523         else
 524                 phb->init_m64 = pnv_ioda2_init_m64;
 525 }
 526
 527 static void pnv_ioda_freeze_pe(struct pnv_phb *phb, int pe_no)
 528 {
 529         struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_no];
 530         struct pnv_ioda_pe *slave;
 531         s64 rc;
 532
 533         /* Fetch master PE */
 534         if (pe->flags & PNV_IODA_PE_SLAVE) {
 535                 pe = pe->master;
 536                 if (WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)))
 537                         return;
 538
 539                 pe_no = pe->pe_number;
 540         }
 541
 542         /* Freeze master PE */
 543         rc = opal_pci_eeh_freeze_set(phb->opal_id,
 544                                      pe_no,
 545                                      OPAL_EEH_ACTION_SET_FREEZE_ALL);
 546         if (rc != OPAL_SUCCESS) {
 547                 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
 548                         __func__, rc, phb->hose->global_number, pe_no);
 549                 return;
 550         }
 551
 552         /* Freeze slave PEs */
 553         if (!(pe->flags & PNV_IODA_PE_MASTER))
 554                 return;
 555
 556         list_for_each_entry(slave, &pe->slaves, list) {
 557                 rc = opal_pci_eeh_freeze_set(phb->opal_id,
 558                                              slave->pe_number,
 559                                              OPAL_EEH_ACTION_SET_FREEZE_ALL);
 560                 if (rc != OPAL_SUCCESS)
 561                         pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
 562                                 __func__, rc, phb->hose->global_number,
 563                                 slave->pe_number);
 564         }
 565 }
 566
 567 static int pnv_ioda_unfreeze_pe(struct pnv_phb *phb, int pe_no, int opt)
 568 {
 569         struct pnv_ioda_pe *pe, *slave;
 570         s64 rc;
 571
 572         /* Find master PE */
 573         pe = &phb->ioda.pe_array[pe_no];
 574         if (pe->flags & PNV_IODA_PE_SLAVE) {
 575                 pe = pe->master;
 576                 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
 577                 pe_no = pe->pe_number;
 578         }
 579
 580         /* Clear frozen state for master PE */
 581         rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, opt);
 582         if (rc != OPAL_SUCCESS) {
 583                 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
 584                         __func__, rc, opt, phb->hose->global_number, pe_no);
 585                 return -EIO;
 586         }
 587
 588         if (!(pe->flags & PNV_IODA_PE_MASTER))
 589                 return 0;
 590
 591         /* Clear frozen state for slave PEs */
 592         list_for_each_entry(slave, &pe->slaves, list) {
 593                 rc = opal_pci_eeh_freeze_clear(phb->opal_id,
 594                                              slave->pe_number,
 595                                              opt);
 596                 if (rc != OPAL_SUCCESS) {
 597                         pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
 598                                 __func__, rc, opt, phb->hose->global_number,
 599                                 slave->pe_number);
 600                         return -EIO;
 601                 }
 602         }
 603
 604         return 0;
 605 }
 606
 607 static int pnv_ioda_get_pe_state(struct pnv_phb *phb, int pe_no)
 608 {
 609         struct pnv_ioda_pe *slave, *pe;
 610         u8 fstate = 0, state;
 611         __be16 pcierr = 0;
 612         s64 rc;
 613
 614         /* Sanity check on PE number */
 615         if (pe_no < 0 || pe_no >= phb->ioda.total_pe_num)
 616                 return OPAL_EEH_STOPPED_PERM_UNAVAIL;
 617
 618         /*
 619          * Fetch the master PE and the PE instance might be
 620          * not initialized yet.
 621          */
 622         pe = &phb->ioda.pe_array[pe_no];
 623         if (pe->flags & PNV_IODA_PE_SLAVE) {
 624                 pe = pe->master;
 625                 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
 626                 pe_no = pe->pe_number;
 627         }
 628
 629         /* Check the master PE */
 630         rc = opal_pci_eeh_freeze_status(phb->opal_id, pe_no,
 631                                         &state, &pcierr, NULL);
 632         if (rc != OPAL_SUCCESS) {
 633                 pr_warn("%s: Failure %lld getting "
 634                         "PHB#%x-PE#%x state\n",
 635                         __func__, rc,
 636                         phb->hose->global_number, pe_no);
 637                 return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
 638         }
 639
 640         /* Check the slave PE */
 641         if (!(pe->flags & PNV_IODA_PE_MASTER))
 642                 return state;
 643
 644         list_for_each_entry(slave, &pe->slaves, list) {
 645                 rc = opal_pci_eeh_freeze_status(phb->opal_id,
 646                                                 slave->pe_number,
 647                                                 &fstate,
 648                                                 &pcierr,
 649                                                 NULL);
 650                 if (rc != OPAL_SUCCESS) {
 651                         pr_warn("%s: Failure %lld getting "
 652                                 "PHB#%x-PE#%x state\n",
 653                                 __func__, rc,
 654                                 phb->hose->global_number, slave->pe_number);
 655                         return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
 656                 }
 657
 658                 /*
 659                  * Override the result based on the ascending
 660                  * priority.
 661                  */
 662                 if (fstate > state)
 663                         state = fstate;
 664         }
 665
 666         return state;
 667 }
 668
 669 struct pnv_ioda_pe *pnv_pci_bdfn_to_pe(struct pnv_phb *phb, u16 bdfn)
 670 {
 671         int pe_number = phb->ioda.pe_rmap[bdfn];
 672
 673         if (pe_number == IODA_INVALID_PE)
 674                 return NULL;
 675
 676         return &phb->ioda.pe_array[pe_number];
 677 }
 678
 679 struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev)
 680 {
 681         struct pnv_phb *phb = pci_bus_to_pnvhb(dev->bus);
 682         struct pci_dn *pdn = pci_get_pdn(dev);
 683
 684         if (!pdn)
 685                 return NULL;
 686         if (pdn->pe_number == IODA_INVALID_PE)
 687                 return NULL;
 688         return &phb->ioda.pe_array[pdn->pe_number];
 689 }
 690
 691 static int pnv_ioda_set_one_peltv(struct pnv_phb *phb,
 692                                   struct pnv_ioda_pe *parent,
 693                                   struct pnv_ioda_pe *child,
 694                                   bool is_add)
 695 {
 696         const char *desc = is_add ? "adding" : "removing";
 697         uint8_t op = is_add ? OPAL_ADD_PE_TO_DOMAIN :
 698                               OPAL_REMOVE_PE_FROM_DOMAIN;
 699         struct pnv_ioda_pe *slave;
 700         long rc;
 701
 702         /* Parent PE affects child PE */
 703         rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
 704                                 child->pe_number, op);
 705         if (rc != OPAL_SUCCESS) {
 706                 pe_warn(child, "OPAL error %ld %s to parent PELTV\n",
 707                         rc, desc);
 708                 return -ENXIO;
 709         }
 710
 711         if (!(child->flags & PNV_IODA_PE_MASTER))
 712                 return 0;
 713
 714         /* Compound case: parent PE affects slave PEs */
 715         list_for_each_entry(slave, &child->slaves, list) {
 716                 rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
 717                                         slave->pe_number, op);
 718                 if (rc != OPAL_SUCCESS) {
 719                         pe_warn(slave, "OPAL error %ld %s to parent PELTV\n",
 720                                 rc, desc);
 721                         return -ENXIO;
 722                 }
 723         }
 724
 725         return 0;
 726 }
 727
 728 static int pnv_ioda_set_peltv(struct pnv_phb *phb,
 729                               struct pnv_ioda_pe *pe,
 730                               bool is_add)
 731 {
 732         struct pnv_ioda_pe *slave;
 733         struct pci_dev *pdev = NULL;
 734         int ret;
 735
 736         /*
 737          * Clear PE frozen state. If it's master PE, we need
 738          * clear slave PE frozen state as well.
 739          */
 740         if (is_add) {
 741                 opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
 742                                           OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 743                 if (pe->flags & PNV_IODA_PE_MASTER) {
 744                         list_for_each_entry(slave, &pe->slaves, list)
 745                                 opal_pci_eeh_freeze_clear(phb->opal_id,
 746                                                           slave->pe_number,
 747                                                           OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 748                 }
 749         }
 750
 751         /*
 752          * Associate PE in PELT. We need add the PE into the
 753          * corresponding PELT-V as well. Otherwise, the error
 754          * originated from the PE might contribute to other
 755          * PEs.
 756          */
 757         ret = pnv_ioda_set_one_peltv(phb, pe, pe, is_add);
 758         if (ret)
 759                 return ret;
 760
 761         /* For compound PEs, any one affects all of them */
 762         if (pe->flags & PNV_IODA_PE_MASTER) {
 763                 list_for_each_entry(slave, &pe->slaves, list) {
 764                         ret = pnv_ioda_set_one_peltv(phb, slave, pe, is_add);
 765                         if (ret)
 766                                 return ret;
 767                 }
 768         }
 769
 770         if (pe->flags & (PNV_IODA_PE_BUS_ALL | PNV_IODA_PE_BUS))
 771                 pdev = pe->pbus->self;
 772         else if (pe->flags & PNV_IODA_PE_DEV)
 773                 pdev = pe->pdev->bus->self;
 774 #ifdef CONFIG_PCI_IOV
 775         else if (pe->flags & PNV_IODA_PE_VF)
 776                 pdev = pe->parent_dev;
 777 #endif /* CONFIG_PCI_IOV */
 778         while (pdev) {
 779                 struct pci_dn *pdn = pci_get_pdn(pdev);
 780                 struct pnv_ioda_pe *parent;
 781
 782                 if (pdn && pdn->pe_number != IODA_INVALID_PE) {
 783                         parent = &phb->ioda.pe_array[pdn->pe_number];
 784                         ret = pnv_ioda_set_one_peltv(phb, parent, pe, is_add);
 785                         if (ret)
 786                                 return ret;
 787                 }
 788
 789                 pdev = pdev->bus->self;
 790         }
 791
 792         return 0;
 793 }
 794
 795 static void pnv_ioda_unset_peltv(struct pnv_phb *phb,
 796                                  struct pnv_ioda_pe *pe,
 797                                  struct pci_dev *parent)
 798 {
 799         int64_t rc;
 800
 801         while (parent) {
 802                 struct pci_dn *pdn = pci_get_pdn(parent);
 803
 804                 if (pdn && pdn->pe_number != IODA_INVALID_PE) {
 805                         rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number,
 806                                                 pe->pe_number,
 807                                                 OPAL_REMOVE_PE_FROM_DOMAIN);
 808                         /* XXX What to do in case of error ? */
 809                 }
 810                 parent = parent->bus->self;
 811         }
 812
 813         opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
 814                                   OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 815
 816         /* Disassociate PE in PELT */
 817         rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number,
 818                                 pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
 819         if (rc)
 820                 pe_warn(pe, "OPAL error %lld remove self from PELTV\n", rc);
 821 }
 822
 823 int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
 824 {
 825         struct pci_dev *parent;
 826         uint8_t bcomp, dcomp, fcomp;
 827         int64_t rc;
 828         long rid_end, rid;
 829
 830         /* Currently, we just deconfigure VF PE. Bus PE will always there.*/
 831         if (pe->pbus) {
 832                 int count;
 833
 834                 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
 835                 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
 836                 parent = pe->pbus->self;
 837                 if (pe->flags & PNV_IODA_PE_BUS_ALL)
 838                         count = resource_size(&pe->pbus->busn_res);
 839                 else
 840                         count = 1;
 841
 842                 switch(count) {
 843                 case  1: bcomp = OpalPciBusAll;         break;
 844                 case  2: bcomp = OpalPciBus7Bits;       break;
 845                 case  4: bcomp = OpalPciBus6Bits;       break;
 846                 case  8: bcomp = OpalPciBus5Bits;       break;
 847                 case 16: bcomp = OpalPciBus4Bits;       break;
 848                 case 32: bcomp = OpalPciBus3Bits;       break;
 849                 default:
 850                         dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
 851                                 count);
 852                         /* Do an exact match only */
 853                         bcomp = OpalPciBusAll;
 854                 }
 855                 rid_end = pe->rid + (count << 8);
 856         } else {
 857 #ifdef CONFIG_PCI_IOV
 858                 if (pe->flags & PNV_IODA_PE_VF)
 859                         parent = pe->parent_dev;
 860                 else
 861 #endif
 862                         parent = pe->pdev->bus->self;
 863                 bcomp = OpalPciBusAll;
 864                 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
 865                 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
 866                 rid_end = pe->rid + 1;
 867         }
 868
 869         /* Clear the reverse map */
 870         for (rid = pe->rid; rid < rid_end; rid++)
 871                 phb->ioda.pe_rmap[rid] = IODA_INVALID_PE;
 872
 873         /*
 874          * Release from all parents PELT-V. NPUs don't have a PELTV
 875          * table
 876          */
 877         if (phb->type != PNV_PHB_NPU_OCAPI)
 878                 pnv_ioda_unset_peltv(phb, pe, parent);
 879
 880         rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
 881                              bcomp, dcomp, fcomp, OPAL_UNMAP_PE);
 882         if (rc)
 883                 pe_err(pe, "OPAL error %lld trying to setup PELT table\n", rc);
 884
 885         pe->pbus = NULL;
 886         pe->pdev = NULL;
 887 #ifdef CONFIG_PCI_IOV
 888         pe->parent_dev = NULL;
 889 #endif
 890
 891         return 0;
 892 }
 893
 894 int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
 895 {
 896         uint8_t bcomp, dcomp, fcomp;
 897         long rc, rid_end, rid;
 898
 899         /* Bus validation ? */
 900         if (pe->pbus) {
 901                 int count;
 902
 903                 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
 904                 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
 905                 if (pe->flags & PNV_IODA_PE_BUS_ALL)
 906                         count = resource_size(&pe->pbus->busn_res);
 907                 else
 908                         count = 1;
 909
 910                 switch(count) {
 911                 case  1: bcomp = OpalPciBusAll;         break;
 912                 case  2: bcomp = OpalPciBus7Bits;       break;
 913                 case  4: bcomp = OpalPciBus6Bits;       break;
 914                 case  8: bcomp = OpalPciBus5Bits;       break;
 915                 case 16: bcomp = OpalPciBus4Bits;       break;
 916                 case 32: bcomp = OpalPciBus3Bits;       break;
 917                 default:
 918                         dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
 919                                 count);
 920                         /* Do an exact match only */
 921                         bcomp = OpalPciBusAll;
 922                 }
 923                 rid_end = pe->rid + (count << 8);
 924         } else {
 925                 bcomp = OpalPciBusAll;
 926                 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
 927                 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
 928                 rid_end = pe->rid + 1;
 929         }
 930
 931         /*
 932          * Associate PE in PELT. We need add the PE into the
 933          * corresponding PELT-V as well. Otherwise, the error
 934          * originated from the PE might contribute to other
 935          * PEs.
 936          */
 937         rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
 938                              bcomp, dcomp, fcomp, OPAL_MAP_PE);
 939         if (rc) {
 940                 pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
 941                 return -ENXIO;
 942         }
 943
 944         /*
 945          * Configure PELTV. NPUs don't have a PELTV table so skip
 946          * configuration on them.
 947          */
 948         if (phb->type != PNV_PHB_NPU_OCAPI)
 949                 pnv_ioda_set_peltv(phb, pe, true);
 950
 951         /* Setup reverse map */
 952         for (rid = pe->rid; rid < rid_end; rid++)
 953                 phb->ioda.pe_rmap[rid] = pe->pe_number;
 954
 955         /* Setup one MVTs on IODA1 */
 956         if (phb->type != PNV_PHB_IODA1) {
 957                 pe->mve_number = 0;
 958                 goto out;
 959         }
 960
 961         pe->mve_number = pe->pe_number;
 962         rc = opal_pci_set_mve(phb->opal_id, pe->mve_number, pe->pe_number);
 963         if (rc != OPAL_SUCCESS) {
 964                 pe_err(pe, "OPAL error %ld setting up MVE %x\n",
 965                        rc, pe->mve_number);
 966                 pe->mve_number = -1;
 967         } else {
 968                 rc = opal_pci_set_mve_enable(phb->opal_id,
 969                                              pe->mve_number, OPAL_ENABLE_MVE);
 970                 if (rc) {
 971                         pe_err(pe, "OPAL error %ld enabling MVE %x\n",
 972                                rc, pe->mve_number);
 973                         pe->mve_number = -1;
 974                 }
 975         }
 976
 977 out:
 978         return 0;
 979 }
 980
 981 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev)
 982 {
 983         struct pnv_phb *phb = pci_bus_to_pnvhb(dev->bus);
 984         struct pci_dn *pdn = pci_get_pdn(dev);
 985         struct pnv_ioda_pe *pe;
 986
 987         if (!pdn) {
 988                 pr_err("%s: Device tree node not associated properly\n",
 989                            pci_name(dev));
 990                 return NULL;
 991         }
 992         if (pdn->pe_number != IODA_INVALID_PE)
 993                 return NULL;
 994
 995         pe = pnv_ioda_alloc_pe(phb, 1);
 996         if (!pe) {
 997                 pr_warn("%s: Not enough PE# available, disabling device\n",
 998                         pci_name(dev));
 999                 return NULL;
1000         }
1001
1002         /* NOTE: We don't get a reference for the pointer in the PE
1003          * data structure, both the device and PE structures should be
1004          * destroyed at the same time.
1005          *
1006          * At some point we want to remove the PDN completely anyways
1007          */
1008         pdn->pe_number = pe->pe_number;
1009         pe->flags = PNV_IODA_PE_DEV;
1010         pe->pdev = dev;
1011         pe->pbus = NULL;
1012         pe->mve_number = -1;
1013         pe->rid = dev->bus->number << 8 | pdn->devfn;
1014         pe->device_count++;
1015
1016         pe_info(pe, "Associated device to PE\n");
1017
1018         if (pnv_ioda_configure_pe(phb, pe)) {
1019                 /* XXX What do we do here ? */
1020                 pnv_ioda_free_pe(pe);
1021                 pdn->pe_number = IODA_INVALID_PE;
1022                 pe->pdev = NULL;
1023                 return NULL;
1024         }
1025
1026         /* Put PE to the list */
1027         mutex_lock(&phb->ioda.pe_list_mutex);
1028         list_add_tail(&pe->list, &phb->ioda.pe_list);
1029         mutex_unlock(&phb->ioda.pe_list_mutex);
1030         return pe;
1031 }
1032
1033 /*
1034  * There're 2 types of PCI bus sensitive PEs: One that is compromised of
1035  * single PCI bus. Another one that contains the primary PCI bus and its
1036  * subordinate PCI devices and buses. The second type of PE is normally
1037  * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
1038  */
1039 static struct pnv_ioda_pe *pnv_ioda_setup_bus_PE(struct pci_bus *bus, bool all)
1040 {
1041         struct pnv_phb *phb = pci_bus_to_pnvhb(bus);
1042         struct pnv_ioda_pe *pe = NULL;
1043         unsigned int pe_num;
1044
1045         /*
1046          * In partial hotplug case, the PE instance might be still alive.
1047          * We should reuse it instead of allocating a new one.
1048          */
1049         pe_num = phb->ioda.pe_rmap[bus->number << 8];
1050         if (WARN_ON(pe_num != IODA_INVALID_PE)) {
1051                 pe = &phb->ioda.pe_array[pe_num];
1052                 return NULL;
1053         }
1054
1055         /* PE number for root bus should have been reserved */
1056         if (pci_is_root_bus(bus))
1057                 pe = &phb->ioda.pe_array[phb->ioda.root_pe_idx];
1058
1059         /* Check if PE is determined by M64 */
1060         if (!pe)
1061                 pe = pnv_ioda_pick_m64_pe(bus, all);
1062
1063         /* The PE number isn't pinned by M64 */
1064         if (!pe)
1065                 pe = pnv_ioda_alloc_pe(phb, 1);
1066
1067         if (!pe) {
1068                 pr_warn("%s: Not enough PE# available for PCI bus %04x:%02x\n",
1069                         __func__, pci_domain_nr(bus), bus->number);
1070                 return NULL;
1071         }
1072
1073         pe->flags |= (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS);
1074         pe->pbus = bus;
1075         pe->pdev = NULL;
1076         pe->mve_number = -1;
1077         pe->rid = bus->busn_res.start << 8;
1078
1079         if (all)
1080                 pe_info(pe, "Secondary bus %pad..%pad associated with PE#%x\n",
1081                         &bus->busn_res.start, &bus->busn_res.end,
1082                         pe->pe_number);
1083         else
1084                 pe_info(pe, "Secondary bus %pad associated with PE#%x\n",
1085                         &bus->busn_res.start, pe->pe_number);
1086
1087         if (pnv_ioda_configure_pe(phb, pe)) {
1088                 /* XXX What do we do here ? */
1089                 pnv_ioda_free_pe(pe);
1090                 pe->pbus = NULL;
1091                 return NULL;
1092         }
1093
1094         /* Put PE to the list */
1095         list_add_tail(&pe->list, &phb->ioda.pe_list);
1096
1097         return pe;
1098 }
1099
1100 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb *phb,
1101                                        struct pnv_ioda_pe *pe);
1102
1103 static void pnv_pci_ioda_dma_dev_setup(struct pci_dev *pdev)
1104 {
1105         struct pnv_phb *phb = pci_bus_to_pnvhb(pdev->bus);
1106         struct pci_dn *pdn = pci_get_pdn(pdev);
1107         struct pnv_ioda_pe *pe;
1108
1109         /* Check if the BDFN for this device is associated with a PE yet */
1110         pe = pnv_pci_bdfn_to_pe(phb, pdev->devfn | (pdev->bus->number << 8));
1111         if (!pe) {
1112                 /* VF PEs should be pre-configured in pnv_pci_sriov_enable() */
1113                 if (WARN_ON(pdev->is_virtfn))
1114                         return;
1115
1116                 pnv_pci_configure_bus(pdev->bus);
1117                 pe = pnv_pci_bdfn_to_pe(phb, pdev->devfn | (pdev->bus->number << 8));
1118                 pci_info(pdev, "Configured PE#%x\n", pe ? pe->pe_number : 0xfffff);
1119
1120
1121                 /*
1122                  * If we can't setup the IODA PE something has gone horribly
1123                  * wrong and we can't enable DMA for the device.
1124                  */
1125                 if (WARN_ON(!pe))
1126                         return;
1127         } else {
1128                 pci_info(pdev, "Added to existing PE#%x\n", pe->pe_number);
1129         }
1130
1131         /*
1132          * We assume that bridges *probably* don't need to do any DMA so we can
1133          * skip allocating a TCE table, etc unless we get a non-bridge device.
1134          */
1135         if (!pe->dma_setup_done && !pci_is_bridge(pdev)) {
1136                 switch (phb->type) {
1137                 case PNV_PHB_IODA1:
1138                         pnv_pci_ioda1_setup_dma_pe(phb, pe);
1139                         break;
1140                 case PNV_PHB_IODA2:
1141                         pnv_pci_ioda2_setup_dma_pe(phb, pe);
1142                         break;
1143                 default:
1144                         pr_warn("%s: No DMA for PHB#%x (type %d)\n",
1145                                 __func__, phb->hose->global_number, phb->type);
1146                 }
1147         }
1148
1149         if (pdn)
1150                 pdn->pe_number = pe->pe_number;
1151         pe->device_count++;
1152
1153         WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops);
1154         pdev->dev.archdata.dma_offset = pe->tce_bypass_base;
1155         set_iommu_table_base(&pdev->dev, pe->table_group.tables[0]);
1156
1157         /* PEs with a DMA weight of zero won't have a group */
1158         if (pe->table_group.group)
1159                 iommu_add_device(&pe->table_group, &pdev->dev);
1160 }
1161
1162 /*
1163  * Reconfigure TVE#0 to be usable as 64-bit DMA space.
1164  *
1165  * The first 4GB of virtual memory for a PE is reserved for 32-bit accesses.
1166  * Devices can only access more than that if bit 59 of the PCI address is set
1167  * by hardware, which indicates TVE#1 should be used instead of TVE#0.
1168  * Many PCI devices are not capable of addressing that many bits, and as a
1169  * result are limited to the 4GB of virtual memory made available to 32-bit
1170  * devices in TVE#0.
1171  *
1172  * In order to work around this, reconfigure TVE#0 to be suitable for 64-bit
1173  * devices by configuring the virtual memory past the first 4GB inaccessible
1174  * by 64-bit DMAs.  This should only be used by devices that want more than
1175  * 4GB, and only on PEs that have no 32-bit devices.
1176  *
1177  * Currently this will only work on PHB3 (POWER8).
1178  */
1179 static int pnv_pci_ioda_dma_64bit_bypass(struct pnv_ioda_pe *pe)
1180 {
1181         u64 window_size, table_size, tce_count, addr;
1182         struct page *table_pages;
1183         u64 tce_order = 28; /* 256MB TCEs */
1184         __be64 *tces;
1185         s64 rc;
1186
1187         /*
1188          * Window size needs to be a power of two, but needs to account for
1189          * shifting memory by the 4GB offset required to skip 32bit space.
1190          */
1191         window_size = roundup_pow_of_two(memory_hotplug_max() + (1ULL << 32));
1192         tce_count = window_size >> tce_order;
1193         table_size = tce_count << 3;
1194
1195         if (table_size < PAGE_SIZE)
1196                 table_size = PAGE_SIZE;
1197
1198         table_pages = alloc_pages_node(pe->phb->hose->node, GFP_KERNEL,
1199                                        get_order(table_size));
1200         if (!table_pages)
1201                 goto err;
1202
1203         tces = page_address(table_pages);
1204         if (!tces)
1205                 goto err;
1206
1207         memset(tces, 0, table_size);
1208
1209         for (addr = 0; addr < memory_hotplug_max(); addr += (1 << tce_order)) {
1210                 tces[(addr + (1ULL << 32)) >> tce_order] =
1211                         cpu_to_be64(addr | TCE_PCI_READ | TCE_PCI_WRITE);
1212         }
1213
1214         rc = opal_pci_map_pe_dma_window(pe->phb->opal_id,
1215                                         pe->pe_number,
1216                                         /* reconfigure window 0 */
1217                                         (pe->pe_number << 1) + 0,
1218                                         1,
1219                                         __pa(tces),
1220                                         table_size,
1221                                         1 << tce_order);
1222         if (rc == OPAL_SUCCESS) {
1223                 pe_info(pe, "Using 64-bit DMA iommu bypass (through TVE#0)\n");
1224                 return 0;
1225         }
1226 err:
1227         pe_err(pe, "Error configuring 64-bit DMA bypass\n");
1228         return -EIO;
1229 }
1230
1231 static bool pnv_pci_ioda_iommu_bypass_supported(struct pci_dev *pdev,
1232                 u64 dma_mask)
1233 {
1234         struct pnv_phb *phb = pci_bus_to_pnvhb(pdev->bus);
1235         struct pci_dn *pdn = pci_get_pdn(pdev);
1236         struct pnv_ioda_pe *pe;
1237
1238         if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1239                 return false;
1240
1241         pe = &phb->ioda.pe_array[pdn->pe_number];
1242         if (pe->tce_bypass_enabled) {
1243                 u64 top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1;
1244                 if (dma_mask >= top)
1245                         return true;
1246         }
1247
1248         /*
1249          * If the device can't set the TCE bypass bit but still wants
1250          * to access 4GB or more, on PHB3 we can reconfigure TVE#0 to
1251          * bypass the 32-bit region and be usable for 64-bit DMAs.
1252          * The device needs to be able to address all of this space.
1253          */
1254         if (dma_mask >> 32 &&
1255             dma_mask > (memory_hotplug_max() + (1ULL << 32)) &&
1256             /* pe->pdev should be set if it's a single device, pe->pbus if not */
1257             (pe->device_count == 1 || !pe->pbus) &&
1258             phb->model == PNV_PHB_MODEL_PHB3) {
1259                 /* Configure the bypass mode */
1260                 s64 rc = pnv_pci_ioda_dma_64bit_bypass(pe);
1261                 if (rc)
1262                         return false;
1263                 /* 4GB offset bypasses 32-bit space */
1264                 pdev->dev.archdata.dma_offset = (1ULL << 32);
1265                 return true;
1266         }
1267
1268         return false;
1269 }
1270
1271 static inline __be64 __iomem *pnv_ioda_get_inval_reg(struct pnv_phb *phb)
1272 {
1273         return phb->regs + 0x210;
1274 }
1275
1276 static void pnv_pci_p7ioc_tce_invalidate(struct iommu_table *tbl,
1277                 unsigned long index, unsigned long npages)
1278 {
1279         struct iommu_table_group_link *tgl = list_first_entry_or_null(
1280                         &tbl->it_group_list, struct iommu_table_group_link,
1281                         next);
1282         struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1283                         struct pnv_ioda_pe, table_group);
1284         __be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb);
1285         unsigned long start, end, inc;
1286
1287         start = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset);
1288         end = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset +
1289                         npages - 1);
1290
1291         /* p7ioc-style invalidation, 2 TCEs per write */
1292         start |= (1ull << 63);
1293         end |= (1ull << 63);
1294         inc = 16;
1295         end |= inc - 1; /* round up end to be different than start */
1296
1297         mb(); /* Ensure above stores are visible */
1298         while (start <= end) {
1299                 __raw_writeq_be(start, invalidate);
1300                 start += inc;
1301         }
1302
1303         /*
1304          * The iommu layer will do another mb() for us on build()
1305          * and we don't care on free()
1306          */
1307 }
1308
1309 static int pnv_ioda1_tce_build(struct iommu_table *tbl, long index,
1310                 long npages, unsigned long uaddr,
1311                 enum dma_data_direction direction,
1312                 unsigned long attrs)
1313 {
1314         int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1315                         attrs);
1316
1317         if (!ret)
1318                 pnv_pci_p7ioc_tce_invalidate(tbl, index, npages);
1319
1320         return ret;
1321 }
1322
1323 #ifdef CONFIG_IOMMU_API
1324 /* Common for IODA1 and IODA2 */
1325 static int pnv_ioda_tce_xchg_no_kill(struct iommu_table *tbl, long index,
1326                 unsigned long *hpa, enum dma_data_direction *direction)
1327 {
1328         return pnv_tce_xchg(tbl, index, hpa, direction);
1329 }
1330 #endif
1331
1332 static void pnv_ioda1_tce_free(struct iommu_table *tbl, long index,
1333                 long npages)
1334 {
1335         pnv_tce_free(tbl, index, npages);
1336
1337         pnv_pci_p7ioc_tce_invalidate(tbl, index, npages);
1338 }
1339
1340 static struct iommu_table_ops pnv_ioda1_iommu_ops = {
1341         .set = pnv_ioda1_tce_build,
1342 #ifdef CONFIG_IOMMU_API
1343         .xchg_no_kill = pnv_ioda_tce_xchg_no_kill,
1344         .tce_kill = pnv_pci_p7ioc_tce_invalidate,
1345         .useraddrptr = pnv_tce_useraddrptr,
1346 #endif
1347         .clear = pnv_ioda1_tce_free,
1348         .get = pnv_tce_get,
1349 };
1350
1351 #define PHB3_TCE_KILL_INVAL_ALL         PPC_BIT(0)
1352 #define PHB3_TCE_KILL_INVAL_PE          PPC_BIT(1)
1353 #define PHB3_TCE_KILL_INVAL_ONE         PPC_BIT(2)
1354
1355 static inline void pnv_pci_phb3_tce_invalidate_pe(struct pnv_ioda_pe *pe)
1356 {
1357         /* 01xb - invalidate TCEs that match the specified PE# */
1358         __be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb);
1359         unsigned long val = PHB3_TCE_KILL_INVAL_PE | (pe->pe_number & 0xFF);
1360
1361         mb(); /* Ensure above stores are visible */
1362         __raw_writeq_be(val, invalidate);
1363 }
1364
1365 static void pnv_pci_phb3_tce_invalidate(struct pnv_ioda_pe *pe,
1366                                         unsigned shift, unsigned long index,
1367                                         unsigned long npages)
1368 {
1369         __be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb);
1370         unsigned long start, end, inc;
1371
1372         /* We'll invalidate DMA address in PE scope */
1373         start = PHB3_TCE_KILL_INVAL_ONE;
1374         start |= (pe->pe_number & 0xFF);
1375         end = start;
1376
1377         /* Figure out the start, end and step */
1378         start |= (index << shift);
1379         end |= ((index + npages - 1) << shift);
1380         inc = (0x1ull << shift);
1381         mb();
1382
1383         while (start <= end) {
1384                 __raw_writeq_be(start, invalidate);
1385                 start += inc;
1386         }
1387 }
1388
1389 static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe *pe)
1390 {
1391         struct pnv_phb *phb = pe->phb;
1392
1393         if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
1394                 pnv_pci_phb3_tce_invalidate_pe(pe);
1395         else
1396                 opal_pci_tce_kill(phb->opal_id, OPAL_PCI_TCE_KILL_PE,
1397                                   pe->pe_number, 0, 0, 0);
1398 }
1399
1400 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table *tbl,
1401                 unsigned long index, unsigned long npages)
1402 {
1403         struct iommu_table_group_link *tgl;
1404
1405         list_for_each_entry_lockless(tgl, &tbl->it_group_list, next) {
1406                 struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1407                                 struct pnv_ioda_pe, table_group);
1408                 struct pnv_phb *phb = pe->phb;
1409                 unsigned int shift = tbl->it_page_shift;
1410
1411                 if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
1412                         pnv_pci_phb3_tce_invalidate(pe, shift,
1413                                                     index, npages);
1414                 else
1415                         opal_pci_tce_kill(phb->opal_id,
1416                                           OPAL_PCI_TCE_KILL_PAGES,
1417                                           pe->pe_number, 1u << shift,
1418                                           index << shift, npages);
1419         }
1420 }
1421
1422 static int pnv_ioda2_tce_build(struct iommu_table *tbl, long index,
1423                 long npages, unsigned long uaddr,
1424                 enum dma_data_direction direction,
1425                 unsigned long attrs)
1426 {
1427         int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1428                         attrs);
1429
1430         if (!ret)
1431                 pnv_pci_ioda2_tce_invalidate(tbl, index, npages);
1432
1433         return ret;
1434 }
1435
1436 static void pnv_ioda2_tce_free(struct iommu_table *tbl, long index,
1437                 long npages)
1438 {
1439         pnv_tce_free(tbl, index, npages);
1440
1441         pnv_pci_ioda2_tce_invalidate(tbl, index, npages);
1442 }
1443
1444 static struct iommu_table_ops pnv_ioda2_iommu_ops = {
1445         .set = pnv_ioda2_tce_build,
1446 #ifdef CONFIG_IOMMU_API
1447         .xchg_no_kill = pnv_ioda_tce_xchg_no_kill,
1448         .tce_kill = pnv_pci_ioda2_tce_invalidate,
1449         .useraddrptr = pnv_tce_useraddrptr,
1450 #endif
1451         .clear = pnv_ioda2_tce_free,
1452         .get = pnv_tce_get,
1453         .free = pnv_pci_ioda2_table_free_pages,
1454 };
1455
1456 static int pnv_pci_ioda_dev_dma_weight(struct pci_dev *dev, void *data)
1457 {
1458         unsigned int *weight = (unsigned int *)data;
1459
1460         /* This is quite simplistic. The "base" weight of a device
1461          * is 10. 0 means no DMA is to be accounted for it.
1462          */
1463         if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1464                 return 0;
1465
1466         if (dev->class == PCI_CLASS_SERIAL_USB_UHCI ||
1467             dev->class == PCI_CLASS_SERIAL_USB_OHCI ||
1468             dev->class == PCI_CLASS_SERIAL_USB_EHCI)
1469                 *weight += 3;
1470         else if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID)
1471                 *weight += 15;
1472         else
1473                 *weight += 10;
1474
1475         return 0;
1476 }
1477
1478 static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe *pe)
1479 {
1480         unsigned int weight = 0;
1481
1482         /* SRIOV VF has same DMA32 weight as its PF */
1483 #ifdef CONFIG_PCI_IOV
1484         if ((pe->flags & PNV_IODA_PE_VF) && pe->parent_dev) {
1485                 pnv_pci_ioda_dev_dma_weight(pe->parent_dev, &weight);
1486                 return weight;
1487         }
1488 #endif
1489
1490         if ((pe->flags & PNV_IODA_PE_DEV) && pe->pdev) {
1491                 pnv_pci_ioda_dev_dma_weight(pe->pdev, &weight);
1492         } else if ((pe->flags & PNV_IODA_PE_BUS) && pe->pbus) {
1493                 struct pci_dev *pdev;
1494
1495                 list_for_each_entry(pdev, &pe->pbus->devices, bus_list)
1496                         pnv_pci_ioda_dev_dma_weight(pdev, &weight);
1497         } else if ((pe->flags & PNV_IODA_PE_BUS_ALL) && pe->pbus) {
1498                 pci_walk_bus(pe->pbus, pnv_pci_ioda_dev_dma_weight, &weight);
1499         }
1500
1501         return weight;
1502 }
1503
1504 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb *phb,
1505                                        struct pnv_ioda_pe *pe)
1506 {
1507
1508         struct page *tce_mem = NULL;
1509         struct iommu_table *tbl;
1510         unsigned int weight, total_weight = 0;
1511         unsigned int tce32_segsz, base, segs, avail, i;
1512         int64_t rc;
1513         void *addr;
1514
1515         /* XXX FIXME: Handle 64-bit only DMA devices */
1516         /* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
1517         /* XXX FIXME: Allocate multi-level tables on PHB3 */
1518         weight = pnv_pci_ioda_pe_dma_weight(pe);
1519         if (!weight)
1520                 return;
1521
1522         pci_walk_bus(phb->hose->bus, pnv_pci_ioda_dev_dma_weight,
1523                      &total_weight);
1524         segs = (weight * phb->ioda.dma32_count) / total_weight;
1525         if (!segs)
1526                 segs = 1;
1527
1528         /*
1529          * Allocate contiguous DMA32 segments. We begin with the expected
1530          * number of segments. With one more attempt, the number of DMA32
1531          * segments to be allocated is decreased by one until one segment
1532          * is allocated successfully.
1533          */
1534         do {
1535                 for (base = 0; base <= phb->ioda.dma32_count - segs; base++) {
1536                         for (avail = 0, i = base; i < base + segs; i++) {
1537                                 if (phb->ioda.dma32_segmap[i] ==
1538                                     IODA_INVALID_PE)
1539                                         avail++;
1540                         }
1541
1542                         if (avail == segs)
1543                                 goto found;
1544                 }
1545         } while (--segs);
1546
1547         if (!segs) {
1548                 pe_warn(pe, "No available DMA32 segments\n");
1549                 return;
1550         }
1551
1552 found:
1553         tbl = pnv_pci_table_alloc(phb->hose->node);
1554         if (WARN_ON(!tbl))
1555                 return;
1556
1557 #ifdef CONFIG_IOMMU_API
1558         pe->table_group.ops = &spapr_tce_table_group_ops;
1559         pe->table_group.pgsizes = SZ_4K;
1560 #endif
1561         iommu_register_group(&pe->table_group, phb->hose->global_number,
1562                         pe->pe_number);
1563         pnv_pci_link_table_and_group(phb->hose->node, 0, tbl, &pe->table_group);
1564
1565         /* Grab a 32-bit TCE table */
1566         pe_info(pe, "DMA weight %d (%d), assigned (%d) %d DMA32 segments\n",
1567                 weight, total_weight, base, segs);
1568         pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n",
1569                 base * PNV_IODA1_DMA32_SEGSIZE,
1570                 (base + segs) * PNV_IODA1_DMA32_SEGSIZE - 1);
1571
1572         /* XXX Currently, we allocate one big contiguous table for the
1573          * TCEs. We only really need one chunk per 256M of TCE space
1574          * (ie per segment) but that's an optimization for later, it
1575          * requires some added smarts with our get/put_tce implementation
1576          *
1577          * Each TCE page is 4KB in size and each TCE entry occupies 8
1578          * bytes
1579          */
1580         tce32_segsz = PNV_IODA1_DMA32_SEGSIZE >> (IOMMU_PAGE_SHIFT_4K - 3);
1581         tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
1582                                    get_order(tce32_segsz * segs));
1583         if (!tce_mem) {
1584                 pe_err(pe, " Failed to allocate a 32-bit TCE memory\n");
1585                 goto fail;
1586         }
1587         addr = page_address(tce_mem);
1588         memset(addr, 0, tce32_segsz * segs);
1589
1590         /* Configure HW */
1591         for (i = 0; i < segs; i++) {
1592                 rc = opal_pci_map_pe_dma_window(phb->opal_id,
1593                                               pe->pe_number,
1594                                               base + i, 1,
1595                                               __pa(addr) + tce32_segsz * i,
1596                                               tce32_segsz, IOMMU_PAGE_SIZE_4K);
1597                 if (rc) {
1598                         pe_err(pe, " Failed to configure 32-bit TCE table, err %lld\n",
1599                                rc);
1600                         goto fail;
1601                 }
1602         }
1603
1604         /* Setup DMA32 segment mapping */
1605         for (i = base; i < base + segs; i++)
1606                 phb->ioda.dma32_segmap[i] = pe->pe_number;
1607
1608         /* Setup linux iommu table */
1609         pnv_pci_setup_iommu_table(tbl, addr, tce32_segsz * segs,
1610                                   base * PNV_IODA1_DMA32_SEGSIZE,
1611                                   IOMMU_PAGE_SHIFT_4K);
1612
1613         tbl->it_ops = &pnv_ioda1_iommu_ops;
1614         pe->table_group.tce32_start = tbl->it_offset << tbl->it_page_shift;
1615         pe->table_group.tce32_size = tbl->it_size << tbl->it_page_shift;
1616         tbl->it_index = (phb->hose->global_number << 16) | pe->pe_number;
1617         if (!iommu_init_table(tbl, phb->hose->node, 0, 0))
1618                 panic("Failed to initialize iommu table");
1619
1620         pe->dma_setup_done = true;
1621         return;
1622  fail:
1623         /* XXX Failure: Try to fallback to 64-bit only ? */
1624         if (tce_mem)
1625                 __free_pages(tce_mem, get_order(tce32_segsz * segs));
1626         if (tbl) {
1627                 pnv_pci_unlink_table_and_group(tbl, &pe->table_group);
1628                 iommu_tce_table_put(tbl);
1629         }
1630 }
1631
1632 static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group,
1633                 int num, struct iommu_table *tbl)
1634 {
1635         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1636                         table_group);
1637         struct pnv_phb *phb = pe->phb;
1638         int64_t rc;
1639         const unsigned long size = tbl->it_indirect_levels ?
1640                         tbl->it_level_size : tbl->it_size;
1641         const __u64 start_addr = tbl->it_offset << tbl->it_page_shift;
1642         const __u64 win_size = tbl->it_size << tbl->it_page_shift;
1643
1644         pe_info(pe, "Setting up window#%d %llx..%llx pg=%lx\n",
1645                 num, start_addr, start_addr + win_size - 1,
1646                 IOMMU_PAGE_SIZE(tbl));
1647
1648         /*
1649          * Map TCE table through TVT. The TVE index is the PE number
1650          * shifted by 1 bit for 32-bits DMA space.
1651          */
1652         rc = opal_pci_map_pe_dma_window(phb->opal_id,
1653                         pe->pe_number,
1654                         (pe->pe_number << 1) + num,
1655                         tbl->it_indirect_levels + 1,
1656                         __pa(tbl->it_base),
1657                         size << 3,
1658                         IOMMU_PAGE_SIZE(tbl));
1659         if (rc) {
1660                 pe_err(pe, "Failed to configure TCE table, err %lld\n", rc);
1661                 return rc;
1662         }
1663
1664         pnv_pci_link_table_and_group(phb->hose->node, num,
1665                         tbl, &pe->table_group);
1666         pnv_pci_ioda2_tce_invalidate_pe(pe);
1667
1668         return 0;
1669 }
1670
1671 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable)
1672 {
1673         uint16_t window_id = (pe->pe_number << 1 ) + 1;
1674         int64_t rc;
1675
1676         pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis");
1677         if (enable) {
1678                 phys_addr_t top = memblock_end_of_DRAM();
1679
1680                 top = roundup_pow_of_two(top);
1681                 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
1682                                                      pe->pe_number,
1683                                                      window_id,
1684                                                      pe->tce_bypass_base,
1685                                                      top);
1686         } else {
1687                 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
1688                                                      pe->pe_number,
1689                                                      window_id,
1690                                                      pe->tce_bypass_base,
1691                                                      0);
1692         }
1693         if (rc)
1694                 pe_err(pe, "OPAL error %lld configuring bypass window\n", rc);
1695         else
1696                 pe->tce_bypass_enabled = enable;
1697 }
1698
1699 static long pnv_pci_ioda2_create_table(struct iommu_table_group *table_group,
1700                 int num, __u32 page_shift, __u64 window_size, __u32 levels,
1701                 bool alloc_userspace_copy, struct iommu_table **ptbl)
1702 {
1703         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1704                         table_group);
1705         int nid = pe->phb->hose->node;
1706         __u64 bus_offset = num ? pe->tce_bypass_base : table_group->tce32_start;
1707         long ret;
1708         struct iommu_table *tbl;
1709
1710         tbl = pnv_pci_table_alloc(nid);
1711         if (!tbl)
1712                 return -ENOMEM;
1713
1714         tbl->it_ops = &pnv_ioda2_iommu_ops;
1715
1716         ret = pnv_pci_ioda2_table_alloc_pages(nid,
1717                         bus_offset, page_shift, window_size,
1718                         levels, alloc_userspace_copy, tbl);
1719         if (ret) {
1720                 iommu_tce_table_put(tbl);
1721                 return ret;
1722         }
1723
1724         *ptbl = tbl;
1725
1726         return 0;
1727 }
1728
1729 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe *pe)
1730 {
1731         struct iommu_table *tbl = NULL;
1732         long rc;
1733         unsigned long res_start, res_end;
1734
1735         /*
1736          * crashkernel= specifies the kdump kernel's maximum memory at
1737          * some offset and there is no guaranteed the result is a power
1738          * of 2, which will cause errors later.
1739          */
1740         const u64 max_memory = __rounddown_pow_of_two(memory_hotplug_max());
1741
1742         /*
1743          * In memory constrained environments, e.g. kdump kernel, the
1744          * DMA window can be larger than available memory, which will
1745          * cause errors later.
1746          */
1747         const u64 maxblock = 1UL << (PAGE_SHIFT + MAX_ORDER);
1748
1749         /*
1750          * We create the default window as big as we can. The constraint is
1751          * the max order of allocation possible. The TCE table is likely to
1752          * end up being multilevel and with on-demand allocation in place,
1753          * the initial use is not going to be huge as the default window aims
1754          * to support crippled devices (i.e. not fully 64bit DMAble) only.
1755          */
1756         /* iommu_table::it_map uses 1 bit per IOMMU page, hence 8 */
1757         const u64 window_size = min((maxblock * 8) << PAGE_SHIFT, max_memory);
1758         /* Each TCE level cannot exceed maxblock so go multilevel if needed */
1759         unsigned long tces_order = ilog2(window_size >> PAGE_SHIFT);
1760         unsigned long tcelevel_order = ilog2(maxblock >> 3);
1761         unsigned int levels = tces_order / tcelevel_order;
1762
1763         if (tces_order % tcelevel_order)
1764                 levels += 1;
1765         /*
1766          * We try to stick to default levels (which is >1 at the moment) in
1767          * order to save memory by relying on on-demain TCE level allocation.
1768          */
1769         levels = max_t(unsigned int, levels, POWERNV_IOMMU_DEFAULT_LEVELS);
1770
1771         rc = pnv_pci_ioda2_create_table(&pe->table_group, 0, PAGE_SHIFT,
1772                         window_size, levels, false, &tbl);
1773         if (rc) {
1774                 pe_err(pe, "Failed to create 32-bit TCE table, err %ld",
1775                                 rc);
1776                 return rc;
1777         }
1778
1779         /* We use top part of 32bit space for MMIO so exclude it from DMA */
1780         res_start = 0;
1781         res_end = 0;
1782         if (window_size > pe->phb->ioda.m32_pci_base) {
1783                 res_start = pe->phb->ioda.m32_pci_base >> tbl->it_page_shift;
1784                 res_end = min(window_size, SZ_4G) >> tbl->it_page_shift;
1785         }
1786
1787         tbl->it_index = (pe->phb->hose->global_number << 16) | pe->pe_number;
1788         if (iommu_init_table(tbl, pe->phb->hose->node, res_start, res_end))
1789                 rc = pnv_pci_ioda2_set_window(&pe->table_group, 0, tbl);
1790         else
1791                 rc = -ENOMEM;
1792         if (rc) {
1793                 pe_err(pe, "Failed to configure 32-bit TCE table, err %ld\n", rc);
1794                 iommu_tce_table_put(tbl);
1795                 tbl = NULL; /* This clears iommu_table_base below */
1796         }
1797         if (!pnv_iommu_bypass_disabled)
1798                 pnv_pci_ioda2_set_bypass(pe, true);
1799
1800         /*
1801          * Set table base for the case of IOMMU DMA use. Usually this is done
1802          * from dma_dev_setup() which is not called when a device is returned
1803          * from VFIO so do it here.
1804          */
1805         if (pe->pdev)
1806                 set_iommu_table_base(&pe->pdev->dev, tbl);
1807
1808         return 0;
1809 }
1810
1811 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
1812                 int num)
1813 {
1814         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1815                         table_group);
1816         struct pnv_phb *phb = pe->phb;
1817         long ret;
1818
1819         pe_info(pe, "Removing DMA window #%d\n", num);
1820
1821         ret = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
1822                         (pe->pe_number << 1) + num,
1823                         0/* levels */, 0/* table address */,
1824                         0/* table size */, 0/* page size */);
1825         if (ret)
1826                 pe_warn(pe, "Unmapping failed, ret = %ld\n", ret);
1827         else
1828                 pnv_pci_ioda2_tce_invalidate_pe(pe);
1829
1830         pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
1831
1832         return ret;
1833 }
1834
1835 #ifdef CONFIG_IOMMU_API
1836 unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift,
1837                 __u64 window_size, __u32 levels)
1838 {
1839         unsigned long bytes = 0;
1840         const unsigned window_shift = ilog2(window_size);
1841         unsigned entries_shift = window_shift - page_shift;
1842         unsigned table_shift = entries_shift + 3;
1843         unsigned long tce_table_size = max(0x1000UL, 1UL << table_shift);
1844         unsigned long direct_table_size;
1845
1846         if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS) ||
1847                         !is_power_of_2(window_size))
1848                 return 0;
1849
1850         /* Calculate a direct table size from window_size and levels */
1851         entries_shift = (entries_shift + levels - 1) / levels;
1852         table_shift = entries_shift + 3;
1853         table_shift = max_t(unsigned, table_shift, PAGE_SHIFT);
1854         direct_table_size =  1UL << table_shift;
1855
1856         for ( ; levels; --levels) {
1857                 bytes += ALIGN(tce_table_size, direct_table_size);
1858
1859                 tce_table_size /= direct_table_size;
1860                 tce_table_size <<= 3;
1861                 tce_table_size = max_t(unsigned long,
1862                                 tce_table_size, direct_table_size);
1863         }
1864
1865         return bytes + bytes; /* one for HW table, one for userspace copy */
1866 }
1867
1868 static long pnv_pci_ioda2_create_table_userspace(
1869                 struct iommu_table_group *table_group,
1870                 int num, __u32 page_shift, __u64 window_size, __u32 levels,
1871                 struct iommu_table **ptbl)
1872 {
1873         long ret = pnv_pci_ioda2_create_table(table_group,
1874                         num, page_shift, window_size, levels, true, ptbl);
1875
1876         if (!ret)
1877                 (*ptbl)->it_allocated_size = pnv_pci_ioda2_get_table_size(
1878                                 page_shift, window_size, levels);
1879         return ret;
1880 }
1881
1882 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe, struct pci_bus *bus)
1883 {
1884         struct pci_dev *dev;
1885
1886         list_for_each_entry(dev, &bus->devices, bus_list) {
1887                 set_iommu_table_base(&dev->dev, pe->table_group.tables[0]);
1888                 dev->dev.archdata.dma_offset = pe->tce_bypass_base;
1889
1890                 if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1891                         pnv_ioda_setup_bus_dma(pe, dev->subordinate);
1892         }
1893 }
1894
1895 static long pnv_ioda2_take_ownership(struct iommu_table_group *table_group)
1896 {
1897         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1898                                                 table_group);
1899         /* Store @tbl as pnv_pci_ioda2_unset_window() resets it */
1900         struct iommu_table *tbl = pe->table_group.tables[0];
1901
1902         /*
1903          * iommu_ops transfers the ownership per a device and we mode
1904          * the group ownership with the first device in the group.
1905          */
1906         if (!tbl)
1907                 return 0;
1908
1909         pnv_pci_ioda2_set_bypass(pe, false);
1910         pnv_pci_ioda2_unset_window(&pe->table_group, 0);
1911         if (pe->pbus)
1912                 pnv_ioda_setup_bus_dma(pe, pe->pbus);
1913         else if (pe->pdev)
1914                 set_iommu_table_base(&pe->pdev->dev, NULL);
1915         iommu_tce_table_put(tbl);
1916
1917         return 0;
1918 }
1919
1920 static void pnv_ioda2_release_ownership(struct iommu_table_group *table_group)
1921 {
1922         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1923                                                 table_group);
1924
1925         /* See the comment about iommu_ops above */
1926         if (pe->table_group.tables[0])
1927                 return;
1928         pnv_pci_ioda2_setup_default_config(pe);
1929         if (pe->pbus)
1930                 pnv_ioda_setup_bus_dma(pe, pe->pbus);
1931 }
1932
1933 static struct iommu_table_group_ops pnv_pci_ioda2_ops = {
1934         .get_table_size = pnv_pci_ioda2_get_table_size,
1935         .create_table = pnv_pci_ioda2_create_table_userspace,
1936         .set_window = pnv_pci_ioda2_set_window,
1937         .unset_window = pnv_pci_ioda2_unset_window,
1938         .take_ownership = pnv_ioda2_take_ownership,
1939         .release_ownership = pnv_ioda2_release_ownership,
1940 };
1941 #endif
1942
1943 void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
1944                                 struct pnv_ioda_pe *pe)
1945 {
1946         int64_t rc;
1947
1948         /* TVE #1 is selected by PCI address bit 59 */
1949         pe->tce_bypass_base = 1ull << 59;
1950
1951         /* The PE will reserve all possible 32-bits space */
1952         pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n",
1953                 phb->ioda.m32_pci_base);
1954
1955         /* Setup linux iommu table */
1956         pe->table_group.tce32_start = 0;
1957         pe->table_group.tce32_size = phb->ioda.m32_pci_base;
1958         pe->table_group.max_dynamic_windows_supported =
1959                         IOMMU_TABLE_GROUP_MAX_TABLES;
1960         pe->table_group.max_levels = POWERNV_IOMMU_MAX_LEVELS;
1961         pe->table_group.pgsizes = pnv_ioda_parse_tce_sizes(phb);
1962
1963         rc = pnv_pci_ioda2_setup_default_config(pe);
1964         if (rc)
1965                 return;
1966
1967 #ifdef CONFIG_IOMMU_API
1968         pe->table_group.ops = &pnv_pci_ioda2_ops;
1969         iommu_register_group(&pe->table_group, phb->hose->global_number,
1970                              pe->pe_number);
1971 #endif
1972         pe->dma_setup_done = true;
1973 }
1974
1975 /*
1976  * Called from KVM in real mode to EOI passthru interrupts. The ICP
1977  * EOI is handled directly in KVM in kvmppc_deliver_irq_passthru().
1978  *
1979  * The IRQ data is mapped in the PCI-MSI domain and the EOI OPAL call
1980  * needs an HW IRQ number mapped in the XICS IRQ domain. The HW IRQ
1981  * numbers of the in-the-middle MSI domain are vector numbers and it's
1982  * good enough for OPAL. Use that.
1983  */
1984 int64_t pnv_opal_pci_msi_eoi(struct irq_data *d)
1985 {
1986         struct pci_controller *hose = irq_data_get_irq_chip_data(d->parent_data);
1987         struct pnv_phb *phb = hose->private_data;
1988
1989         return opal_pci_msi_eoi(phb->opal_id, d->parent_data->hwirq);
1990 }
1991
1992 /*
1993  * The IRQ data is mapped in the XICS domain, with OPAL HW IRQ numbers
1994  */
1995 static void pnv_ioda2_msi_eoi(struct irq_data *d)
1996 {
1997         int64_t rc;
1998         unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
1999         struct pci_controller *hose = irq_data_get_irq_chip_data(d);
2000         struct pnv_phb *phb = hose->private_data;
2001
2002         rc = opal_pci_msi_eoi(phb->opal_id, hw_irq);
2003         WARN_ON_ONCE(rc);
2004
2005         icp_native_eoi(d);
2006 }
2007
2008 /* P8/CXL only */
2009 void pnv_set_msi_irq_chip(struct pnv_phb *phb, unsigned int virq)
2010 {
2011         struct irq_data *idata;
2012         struct irq_chip *ichip;
2013
2014         /* The MSI EOI OPAL call is only needed on PHB3 */
2015         if (phb->model != PNV_PHB_MODEL_PHB3)
2016                 return;
2017
2018         if (!phb->ioda.irq_chip_init) {
2019                 /*
2020                  * First time we setup an MSI IRQ, we need to setup the
2021                  * corresponding IRQ chip to route correctly.
2022                  */
2023                 idata = irq_get_irq_data(virq);
2024                 ichip = irq_data_get_irq_chip(idata);
2025                 phb->ioda.irq_chip_init = 1;
2026                 phb->ioda.irq_chip = *ichip;
2027                 phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi;
2028         }
2029         irq_set_chip(virq, &phb->ioda.irq_chip);
2030         irq_set_chip_data(virq, phb->hose);
2031 }
2032
2033 static struct irq_chip pnv_pci_msi_irq_chip;
2034
2035 /*
2036  * Returns true iff chip is something that we could call
2037  * pnv_opal_pci_msi_eoi for.
2038  */
2039 bool is_pnv_opal_msi(struct irq_chip *chip)
2040 {
2041         return chip == &pnv_pci_msi_irq_chip;
2042 }
2043 EXPORT_SYMBOL_GPL(is_pnv_opal_msi);
2044
2045 static int __pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev,
2046                                     unsigned int xive_num,
2047                                     unsigned int is_64, struct msi_msg *msg)
2048 {
2049         struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev);
2050         __be32 data;
2051         int rc;
2052
2053         dev_dbg(&dev->dev, "%s: setup %s-bit MSI for vector #%d\n", __func__,
2054                 is_64 ? "64" : "32", xive_num);
2055
2056         /* No PE assigned ? bail out ... no MSI for you ! */
2057         if (pe == NULL)
2058                 return -ENXIO;
2059
2060         /* Check if we have an MVE */
2061         if (pe->mve_number < 0)
2062                 return -ENXIO;
2063
2064         /* Force 32-bit MSI on some broken devices */
2065         if (dev->no_64bit_msi)
2066                 is_64 = 0;
2067
2068         /* Assign XIVE to PE */
2069         rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
2070         if (rc) {
2071                 pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
2072                         pci_name(dev), rc, xive_num);
2073                 return -EIO;
2074         }
2075
2076         if (is_64) {
2077                 __be64 addr64;
2078
2079                 rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1,
2080                                      &addr64, &data);
2081                 if (rc) {
2082                         pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
2083                                 pci_name(dev), rc);
2084                         return -EIO;
2085                 }
2086                 msg->address_hi = be64_to_cpu(addr64) >> 32;
2087                 msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful;
2088         } else {
2089                 __be32 addr32;
2090
2091                 rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1,
2092                                      &addr32, &data);
2093                 if (rc) {
2094                         pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
2095                                 pci_name(dev), rc);
2096                         return -EIO;
2097                 }
2098                 msg->address_hi = 0;
2099                 msg->address_lo = be32_to_cpu(addr32);
2100         }
2101         msg->data = be32_to_cpu(data);
2102
2103         return 0;
2104 }
2105
2106 /*
2107  * The msi_free() op is called before irq_domain_free_irqs_top() when
2108  * the handler data is still available. Use that to clear the XIVE
2109  * controller.
2110  */
2111 static void pnv_msi_ops_msi_free(struct irq_domain *domain,
2112                                  struct msi_domain_info *info,
2113                                  unsigned int irq)
2114 {
2115         if (xive_enabled())
2116                 xive_irq_free_data(irq);
2117 }
2118
2119 static struct msi_domain_ops pnv_pci_msi_domain_ops = {
2120         .msi_free       = pnv_msi_ops_msi_free,
2121 };
2122
2123 static void pnv_msi_shutdown(struct irq_data *d)
2124 {
2125         d = d->parent_data;
2126         if (d->chip->irq_shutdown)
2127                 d->chip->irq_shutdown(d);
2128 }
2129
2130 static void pnv_msi_mask(struct irq_data *d)
2131 {
2132         pci_msi_mask_irq(d);
2133         irq_chip_mask_parent(d);
2134 }
2135
2136 static void pnv_msi_unmask(struct irq_data *d)
2137 {
2138         pci_msi_unmask_irq(d);
2139         irq_chip_unmask_parent(d);
2140 }
2141
2142 static struct irq_chip pnv_pci_msi_irq_chip = {
2143         .name           = "PNV-PCI-MSI",
2144         .irq_shutdown   = pnv_msi_shutdown,
2145         .irq_mask       = pnv_msi_mask,
2146         .irq_unmask     = pnv_msi_unmask,
2147         .irq_eoi        = irq_chip_eoi_parent,
2148 };
2149
2150 static struct msi_domain_info pnv_msi_domain_info = {
2151         .flags = (MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
2152                   MSI_FLAG_MULTI_PCI_MSI  | MSI_FLAG_PCI_MSIX),
2153         .ops   = &pnv_pci_msi_domain_ops,
2154         .chip  = &pnv_pci_msi_irq_chip,
2155 };
2156
2157 static void pnv_msi_compose_msg(struct irq_data *d, struct msi_msg *msg)
2158 {
2159         struct msi_desc *entry = irq_data_get_msi_desc(d);
2160         struct pci_dev *pdev = msi_desc_to_pci_dev(entry);
2161         struct pci_controller *hose = irq_data_get_irq_chip_data(d);
2162         struct pnv_phb *phb = hose->private_data;
2163         int rc;
2164
2165         rc = __pnv_pci_ioda_msi_setup(phb, pdev, d->hwirq,
2166                                       entry->pci.msi_attrib.is_64, msg);
2167         if (rc)
2168                 dev_err(&pdev->dev, "Failed to setup %s-bit MSI #%ld : %d\n",
2169                         entry->pci.msi_attrib.is_64 ? "64" : "32", d->hwirq, rc);
2170 }
2171
2172 /*
2173  * The IRQ data is mapped in the MSI domain in which HW IRQ numbers
2174  * correspond to vector numbers.
2175  */
2176 static void pnv_msi_eoi(struct irq_data *d)
2177 {
2178         struct pci_controller *hose = irq_data_get_irq_chip_data(d);
2179         struct pnv_phb *phb = hose->private_data;
2180
2181         if (phb->model == PNV_PHB_MODEL_PHB3) {
2182                 /*
2183                  * The EOI OPAL call takes an OPAL HW IRQ number but
2184                  * since it is translated into a vector number in
2185                  * OPAL, use that directly.
2186                  */
2187                 WARN_ON_ONCE(opal_pci_msi_eoi(phb->opal_id, d->hwirq));
2188         }
2189
2190         irq_chip_eoi_parent(d);
2191 }
2192
2193 static struct irq_chip pnv_msi_irq_chip = {
2194         .name                   = "PNV-MSI",
2195         .irq_shutdown           = pnv_msi_shutdown,
2196         .irq_mask               = irq_chip_mask_parent,
2197         .irq_unmask             = irq_chip_unmask_parent,
2198         .irq_eoi                = pnv_msi_eoi,
2199         .irq_set_affinity       = irq_chip_set_affinity_parent,
2200         .irq_compose_msi_msg    = pnv_msi_compose_msg,
2201 };
2202
2203 static int pnv_irq_parent_domain_alloc(struct irq_domain *domain,
2204                                        unsigned int virq, int hwirq)
2205 {
2206         struct irq_fwspec parent_fwspec;
2207         int ret;
2208
2209         parent_fwspec.fwnode = domain->parent->fwnode;
2210         parent_fwspec.param_count = 2;
2211         parent_fwspec.param[0] = hwirq;
2212         parent_fwspec.param[1] = IRQ_TYPE_EDGE_RISING;
2213
2214         ret = irq_domain_alloc_irqs_parent(domain, virq, 1, &parent_fwspec);
2215         if (ret)
2216                 return ret;
2217
2218         return 0;
2219 }
2220
2221 static int pnv_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
2222                                 unsigned int nr_irqs, void *arg)
2223 {
2224         struct pci_controller *hose = domain->host_data;
2225         struct pnv_phb *phb = hose->private_data;
2226         msi_alloc_info_t *info = arg;
2227         struct pci_dev *pdev = msi_desc_to_pci_dev(info->desc);
2228         int hwirq;
2229         int i, ret;
2230
2231         hwirq = msi_bitmap_alloc_hwirqs(&phb->msi_bmp, nr_irqs);
2232         if (hwirq < 0) {
2233                 dev_warn(&pdev->dev, "failed to find a free MSI\n");
2234                 return -ENOSPC;
2235         }
2236
2237         dev_dbg(&pdev->dev, "%s bridge %pOF %d/%x #%d\n", __func__,
2238                 hose->dn, virq, hwirq, nr_irqs);
2239
2240         for (i = 0; i < nr_irqs; i++) {
2241                 ret = pnv_irq_parent_domain_alloc(domain, virq + i,
2242                                                   phb->msi_base + hwirq + i);
2243                 if (ret)
2244                         goto out;
2245
2246                 irq_domain_set_hwirq_and_chip(domain, virq + i, hwirq + i,
2247                                               &pnv_msi_irq_chip, hose);
2248         }
2249
2250         return 0;
2251
2252 out:
2253         irq_domain_free_irqs_parent(domain, virq, i - 1);
2254         msi_bitmap_free_hwirqs(&phb->msi_bmp, hwirq, nr_irqs);
2255         return ret;
2256 }
2257
2258 static void pnv_irq_domain_free(struct irq_domain *domain, unsigned int virq,
2259                                 unsigned int nr_irqs)
2260 {
2261         struct irq_data *d = irq_domain_get_irq_data(domain, virq);
2262         struct pci_controller *hose = irq_data_get_irq_chip_data(d);
2263         struct pnv_phb *phb = hose->private_data;
2264
2265         pr_debug("%s bridge %pOF %d/%lx #%d\n", __func__, hose->dn,
2266                  virq, d->hwirq, nr_irqs);
2267
2268         msi_bitmap_free_hwirqs(&phb->msi_bmp, d->hwirq, nr_irqs);
2269         /* XIVE domain is cleared through ->msi_free() */
2270 }
2271
2272 static const struct irq_domain_ops pnv_irq_domain_ops = {
2273         .alloc  = pnv_irq_domain_alloc,
2274         .free   = pnv_irq_domain_free,
2275 };
2276
2277 static int __init pnv_msi_allocate_domains(struct pci_controller *hose, unsigned int count)
2278 {
2279         struct pnv_phb *phb = hose->private_data;
2280         struct irq_domain *parent = irq_get_default_host();
2281
2282         hose->fwnode = irq_domain_alloc_named_id_fwnode("PNV-MSI", phb->opal_id);
2283         if (!hose->fwnode)
2284                 return -ENOMEM;
2285
2286         hose->dev_domain = irq_domain_create_hierarchy(parent, 0, count,
2287                                                        hose->fwnode,
2288                                                        &pnv_irq_domain_ops, hose);
2289         if (!hose->dev_domain) {
2290                 pr_err("PCI: failed to create IRQ domain bridge %pOF (domain %d)\n",
2291                        hose->dn, hose->global_number);
2292                 irq_domain_free_fwnode(hose->fwnode);
2293                 return -ENOMEM;
2294         }
2295
2296         hose->msi_domain = pci_msi_create_irq_domain(of_node_to_fwnode(hose->dn),
2297                                                      &pnv_msi_domain_info,
2298                                                      hose->dev_domain);
2299         if (!hose->msi_domain) {
2300                 pr_err("PCI: failed to create MSI IRQ domain bridge %pOF (domain %d)\n",
2301                        hose->dn, hose->global_number);
2302                 irq_domain_free_fwnode(hose->fwnode);
2303                 irq_domain_remove(hose->dev_domain);
2304                 return -ENOMEM;
2305         }
2306
2307         return 0;
2308 }
2309
2310 static void __init pnv_pci_init_ioda_msis(struct pnv_phb *phb)
2311 {
2312         unsigned int count;
2313         const __be32 *prop = of_get_property(phb->hose->dn,
2314                                              "ibm,opal-msi-ranges", NULL);
2315         if (!prop) {
2316                 /* BML Fallback */
2317                 prop = of_get_property(phb->hose->dn, "msi-ranges", NULL);
2318         }
2319         if (!prop)
2320                 return;
2321
2322         phb->msi_base = be32_to_cpup(prop);
2323         count = be32_to_cpup(prop + 1);
2324         if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) {
2325                 pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
2326                        phb->hose->global_number);
2327                 return;
2328         }
2329
2330         pr_info("  Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
2331                 count, phb->msi_base);
2332
2333         pnv_msi_allocate_domains(phb->hose, count);
2334 }
2335
2336 static void pnv_ioda_setup_pe_res(struct pnv_ioda_pe *pe,
2337                                   struct resource *res)
2338 {
2339         struct pnv_phb *phb = pe->phb;
2340         struct pci_bus_region region;
2341         int index;
2342         int64_t rc;
2343
2344         if (!res || !res->flags || res->start > res->end ||
2345             res->flags & IORESOURCE_UNSET)
2346                 return;
2347
2348         if (res->flags & IORESOURCE_IO) {
2349                 region.start = res->start - phb->ioda.io_pci_base;
2350                 region.end   = res->end - phb->ioda.io_pci_base;
2351                 index = region.start / phb->ioda.io_segsize;
2352
2353                 while (index < phb->ioda.total_pe_num &&
2354                        region.start <= region.end) {
2355                         phb->ioda.io_segmap[index] = pe->pe_number;
2356                         rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2357                                 pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index);
2358                         if (rc != OPAL_SUCCESS) {
2359                                 pr_err("%s: Error %lld mapping IO segment#%d to PE#%x\n",
2360                                        __func__, rc, index, pe->pe_number);
2361                                 break;
2362                         }
2363
2364                         region.start += phb->ioda.io_segsize;
2365                         index++;
2366                 }
2367         } else if ((res->flags & IORESOURCE_MEM) &&
2368                    !pnv_pci_is_m64(phb, res)) {
2369                 region.start = res->start -
2370                                phb->hose->mem_offset[0] -
2371                                phb->ioda.m32_pci_base;
2372                 region.end   = res->end -
2373                                phb->hose->mem_offset[0] -
2374                                phb->ioda.m32_pci_base;
2375                 index = region.start / phb->ioda.m32_segsize;
2376
2377                 while (index < phb->ioda.total_pe_num &&
2378                        region.start <= region.end) {
2379                         phb->ioda.m32_segmap[index] = pe->pe_number;
2380                         rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2381                                 pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index);
2382                         if (rc != OPAL_SUCCESS) {
2383                                 pr_err("%s: Error %lld mapping M32 segment#%d to PE#%x",
2384                                        __func__, rc, index, pe->pe_number);
2385                                 break;
2386                         }
2387
2388                         region.start += phb->ioda.m32_segsize;
2389                         index++;
2390                 }
2391         }
2392 }
2393
2394 /*
2395  * This function is supposed to be called on basis of PE from top
2396  * to bottom style. So the I/O or MMIO segment assigned to
2397  * parent PE could be overridden by its child PEs if necessary.
2398  */
2399 static void pnv_ioda_setup_pe_seg(struct pnv_ioda_pe *pe)
2400 {
2401         struct pci_dev *pdev;
2402         int i;
2403
2404         /*
2405          * NOTE: We only care PCI bus based PE for now. For PCI
2406          * device based PE, for example SRIOV sensitive VF should
2407          * be figured out later.
2408          */
2409         BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)));
2410
2411         list_for_each_entry(pdev, &pe->pbus->devices, bus_list) {
2412                 for (i = 0; i <= PCI_ROM_RESOURCE; i++)
2413                         pnv_ioda_setup_pe_res(pe, &pdev->resource[i]);
2414
2415                 /*
2416                  * If the PE contains all subordinate PCI buses, the
2417                  * windows of the child bridges should be mapped to
2418                  * the PE as well.
2419                  */
2420                 if (!(pe->flags & PNV_IODA_PE_BUS_ALL) || !pci_is_bridge(pdev))
2421                         continue;
2422                 for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++)
2423                         pnv_ioda_setup_pe_res(pe,
2424                                 &pdev->resource[PCI_BRIDGE_RESOURCES + i]);
2425         }
2426 }
2427
2428 #ifdef CONFIG_DEBUG_FS
2429 static int pnv_pci_diag_data_set(void *data, u64 val)
2430 {
2431         struct pnv_phb *phb = data;
2432         s64 ret;
2433
2434         /* Retrieve the diag data from firmware */
2435         ret = opal_pci_get_phb_diag_data2(phb->opal_id, phb->diag_data,
2436                                           phb->diag_data_size);
2437         if (ret != OPAL_SUCCESS)
2438                 return -EIO;
2439
2440         /* Print the diag data to the kernel log */
2441         pnv_pci_dump_phb_diag_data(phb->hose, phb->diag_data);
2442         return 0;
2443 }
2444
2445 DEFINE_DEBUGFS_ATTRIBUTE(pnv_pci_diag_data_fops, NULL, pnv_pci_diag_data_set,
2446                          "%llu\n");
2447
2448 static int pnv_pci_ioda_pe_dump(void *data, u64 val)
2449 {
2450         struct pnv_phb *phb = data;
2451         int pe_num;
2452
2453         for (pe_num = 0; pe_num < phb->ioda.total_pe_num; pe_num++) {
2454                 struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_num];
2455
2456                 if (!test_bit(pe_num, phb->ioda.pe_alloc))
2457                         continue;
2458
2459                 pe_warn(pe, "rid: %04x dev count: %2d flags: %s%s%s%s%s%s\n",
2460                         pe->rid, pe->device_count,
2461                         (pe->flags & PNV_IODA_PE_DEV) ? "dev " : "",
2462                         (pe->flags & PNV_IODA_PE_BUS) ? "bus " : "",
2463                         (pe->flags & PNV_IODA_PE_BUS_ALL) ? "all " : "",
2464                         (pe->flags & PNV_IODA_PE_MASTER) ? "master " : "",
2465                         (pe->flags & PNV_IODA_PE_SLAVE) ? "slave " : "",
2466                         (pe->flags & PNV_IODA_PE_VF) ? "vf " : "");
2467         }
2468
2469         return 0;
2470 }
2471
2472 DEFINE_DEBUGFS_ATTRIBUTE(pnv_pci_ioda_pe_dump_fops, NULL,
2473                          pnv_pci_ioda_pe_dump, "%llu\n");
2474
2475 #endif /* CONFIG_DEBUG_FS */
2476
2477 static void pnv_pci_ioda_create_dbgfs(void)
2478 {
2479 #ifdef CONFIG_DEBUG_FS
2480         struct pci_controller *hose, *tmp;
2481         struct pnv_phb *phb;
2482         char name[16];
2483
2484         list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
2485                 phb = hose->private_data;
2486
2487                 sprintf(name, "PCI%04x", hose->global_number);
2488                 phb->dbgfs = debugfs_create_dir(name, arch_debugfs_dir);
2489
2490                 debugfs_create_file_unsafe("dump_diag_regs", 0200, phb->dbgfs,
2491                                            phb, &pnv_pci_diag_data_fops);
2492                 debugfs_create_file_unsafe("dump_ioda_pe_state", 0200, phb->dbgfs,
2493                                            phb, &pnv_pci_ioda_pe_dump_fops);
2494         }
2495 #endif /* CONFIG_DEBUG_FS */
2496 }
2497
2498 static void pnv_pci_enable_bridge(struct pci_bus *bus)
2499 {
2500         struct pci_dev *dev = bus->self;
2501         struct pci_bus *child;
2502
2503         /* Empty bus ? bail */
2504         if (list_empty(&bus->devices))
2505                 return;
2506
2507         /*
2508          * If there's a bridge associated with that bus enable it. This works
2509          * around races in the generic code if the enabling is done during
2510          * parallel probing. This can be removed once those races have been
2511          * fixed.
2512          */
2513         if (dev) {
2514                 int rc = pci_enable_device(dev);
2515                 if (rc)
2516                         pci_err(dev, "Error enabling bridge (%d)\n", rc);
2517                 pci_set_master(dev);
2518         }
2519
2520         /* Perform the same to child busses */
2521         list_for_each_entry(child, &bus->children, node)
2522                 pnv_pci_enable_bridge(child);
2523 }
2524
2525 static void pnv_pci_enable_bridges(void)
2526 {
2527         struct pci_controller *hose;
2528
2529         list_for_each_entry(hose, &hose_list, list_node)
2530                 pnv_pci_enable_bridge(hose->bus);
2531 }
2532
2533 static void pnv_pci_ioda_fixup(void)
2534 {
2535         pnv_pci_ioda_create_dbgfs();
2536
2537         pnv_pci_enable_bridges();
2538
2539 #ifdef CONFIG_EEH
2540         pnv_eeh_post_init();
2541 #endif
2542 }
2543
2544 /*
2545  * Returns the alignment for I/O or memory windows for P2P
2546  * bridges. That actually depends on how PEs are segmented.
2547  * For now, we return I/O or M32 segment size for PE sensitive
2548  * P2P bridges. Otherwise, the default values (4KiB for I/O,
2549  * 1MiB for memory) will be returned.
2550  *
2551  * The current PCI bus might be put into one PE, which was
2552  * create against the parent PCI bridge. For that case, we
2553  * needn't enlarge the alignment so that we can save some
2554  * resources.
2555  */
2556 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus,
2557                                                 unsigned long type)
2558 {
2559         struct pnv_phb *phb = pci_bus_to_pnvhb(bus);
2560         int num_pci_bridges = 0;
2561         struct pci_dev *bridge;
2562
2563         bridge = bus->self;
2564         while (bridge) {
2565                 if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) {
2566                         num_pci_bridges++;
2567                         if (num_pci_bridges >= 2)
2568                                 return 1;
2569                 }
2570
2571                 bridge = bridge->bus->self;
2572         }
2573
2574         /*
2575          * We fall back to M32 if M64 isn't supported. We enforce the M64
2576          * alignment for any 64-bit resource, PCIe doesn't care and
2577          * bridges only do 64-bit prefetchable anyway.
2578          */
2579         if (phb->ioda.m64_segsize && pnv_pci_is_m64_flags(type))
2580                 return phb->ioda.m64_segsize;
2581         if (type & IORESOURCE_MEM)
2582                 return phb->ioda.m32_segsize;
2583
2584         return phb->ioda.io_segsize;
2585 }
2586
2587 /*
2588  * We are updating root port or the upstream port of the
2589  * bridge behind the root port with PHB's windows in order
2590  * to accommodate the changes on required resources during
2591  * PCI (slot) hotplug, which is connected to either root
2592  * port or the downstream ports of PCIe switch behind the
2593  * root port.
2594  */
2595 static void pnv_pci_fixup_bridge_resources(struct pci_bus *bus,
2596                                            unsigned long type)
2597 {
2598         struct pci_controller *hose = pci_bus_to_host(bus);
2599         struct pnv_phb *phb = hose->private_data;
2600         struct pci_dev *bridge = bus->self;
2601         struct resource *r, *w;
2602         bool msi_region = false;
2603         int i;
2604
2605         /* Check if we need apply fixup to the bridge's windows */
2606         if (!pci_is_root_bus(bridge->bus) &&
2607             !pci_is_root_bus(bridge->bus->self->bus))
2608                 return;
2609
2610         /* Fixup the resources */
2611         for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++) {
2612                 r = &bridge->resource[PCI_BRIDGE_RESOURCES + i];
2613                 if (!r->flags || !r->parent)
2614                         continue;
2615
2616                 w = NULL;
2617                 if (r->flags & type & IORESOURCE_IO)
2618                         w = &hose->io_resource;
2619                 else if (pnv_pci_is_m64(phb, r) &&
2620                          (type & IORESOURCE_PREFETCH) &&
2621                          phb->ioda.m64_segsize)
2622                         w = &hose->mem_resources[1];
2623                 else if (r->flags & type & IORESOURCE_MEM) {
2624                         w = &hose->mem_resources[0];
2625                         msi_region = true;
2626                 }
2627
2628                 r->start = w->start;
2629                 r->end = w->end;
2630
2631                 /* The 64KB 32-bits MSI region shouldn't be included in
2632                  * the 32-bits bridge window. Otherwise, we can see strange
2633                  * issues. One of them is EEH error observed on Garrison.
2634                  *
2635                  * Exclude top 1MB region which is the minimal alignment of
2636                  * 32-bits bridge window.
2637                  */
2638                 if (msi_region) {
2639                         r->end += 0x10000;
2640                         r->end -= 0x100000;
2641                 }
2642         }
2643 }
2644
2645 static void pnv_pci_configure_bus(struct pci_bus *bus)
2646 {
2647         struct pci_dev *bridge = bus->self;
2648         struct pnv_ioda_pe *pe;
2649         bool all = (bridge && pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE);
2650
2651         dev_info(&bus->dev, "Configuring PE for bus\n");
2652
2653         /* Don't assign PE to PCI bus, which doesn't have subordinate devices */
2654         if (WARN_ON(list_empty(&bus->devices)))
2655                 return;
2656
2657         /* Reserve PEs according to used M64 resources */
2658         pnv_ioda_reserve_m64_pe(bus, NULL, all);
2659
2660         /*
2661          * Assign PE. We might run here because of partial hotplug.
2662          * For the case, we just pick up the existing PE and should
2663          * not allocate resources again.
2664          */
2665         pe = pnv_ioda_setup_bus_PE(bus, all);
2666         if (!pe)
2667                 return;
2668
2669         pnv_ioda_setup_pe_seg(pe);
2670 }
2671
2672 static resource_size_t pnv_pci_default_alignment(void)
2673 {
2674         return PAGE_SIZE;
2675 }
2676
2677 /* Prevent enabling devices for which we couldn't properly
2678  * assign a PE
2679  */
2680 static bool pnv_pci_enable_device_hook(struct pci_dev *dev)
2681 {
2682         struct pci_dn *pdn;
2683
2684         pdn = pci_get_pdn(dev);
2685         if (!pdn || pdn->pe_number == IODA_INVALID_PE) {
2686                 pci_err(dev, "pci_enable_device() blocked, no PE assigned.\n");
2687                 return false;
2688         }
2689
2690         return true;
2691 }
2692
2693 static bool pnv_ocapi_enable_device_hook(struct pci_dev *dev)
2694 {
2695         struct pci_dn *pdn;
2696         struct pnv_ioda_pe *pe;
2697
2698         pdn = pci_get_pdn(dev);
2699         if (!pdn)
2700                 return false;
2701
2702         if (pdn->pe_number == IODA_INVALID_PE) {
2703                 pe = pnv_ioda_setup_dev_PE(dev);
2704                 if (!pe)
2705                         return false;
2706         }
2707         return true;
2708 }
2709
2710 static long pnv_pci_ioda1_unset_window(struct iommu_table_group *table_group,
2711                                        int num)
2712 {
2713         struct pnv_ioda_pe *pe = container_of(table_group,
2714                                               struct pnv_ioda_pe, table_group);
2715         struct pnv_phb *phb = pe->phb;
2716         unsigned int idx;
2717         long rc;
2718
2719         pe_info(pe, "Removing DMA window #%d\n", num);
2720         for (idx = 0; idx < phb->ioda.dma32_count; idx++) {
2721                 if (phb->ioda.dma32_segmap[idx] != pe->pe_number)
2722                         continue;
2723
2724                 rc = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
2725                                                 idx, 0, 0ul, 0ul, 0ul);
2726                 if (rc != OPAL_SUCCESS) {
2727                         pe_warn(pe, "Failure %ld unmapping DMA32 segment#%d\n",
2728                                 rc, idx);
2729                         return rc;
2730                 }
2731
2732                 phb->ioda.dma32_segmap[idx] = IODA_INVALID_PE;
2733         }
2734
2735         pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
2736         return OPAL_SUCCESS;
2737 }
2738
2739 static void pnv_pci_ioda1_release_pe_dma(struct pnv_ioda_pe *pe)
2740 {
2741         struct iommu_table *tbl = pe->table_group.tables[0];
2742         int64_t rc;
2743
2744         if (!pe->dma_setup_done)
2745                 return;
2746
2747         rc = pnv_pci_ioda1_unset_window(&pe->table_group, 0);
2748         if (rc != OPAL_SUCCESS)
2749                 return;
2750
2751         pnv_pci_p7ioc_tce_invalidate(tbl, tbl->it_offset, tbl->it_size);
2752         if (pe->table_group.group) {
2753                 iommu_group_put(pe->table_group.group);
2754                 WARN_ON(pe->table_group.group);
2755         }
2756
2757         free_pages(tbl->it_base, get_order(tbl->it_size << 3));
2758         iommu_tce_table_put(tbl);
2759 }
2760
2761 void pnv_pci_ioda2_release_pe_dma(struct pnv_ioda_pe *pe)
2762 {
2763         struct iommu_table *tbl = pe->table_group.tables[0];
2764         int64_t rc;
2765
2766         if (!pe->dma_setup_done)
2767                 return;
2768
2769         rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
2770         if (rc)
2771                 pe_warn(pe, "OPAL error %lld release DMA window\n", rc);
2772
2773         pnv_pci_ioda2_set_bypass(pe, false);
2774         if (pe->table_group.group) {
2775                 iommu_group_put(pe->table_group.group);
2776                 WARN_ON(pe->table_group.group);
2777         }
2778
2779         iommu_tce_table_put(tbl);
2780 }
2781
2782 static void pnv_ioda_free_pe_seg(struct pnv_ioda_pe *pe,
2783                                  unsigned short win,
2784                                  unsigned int *map)
2785 {
2786         struct pnv_phb *phb = pe->phb;
2787         int idx;
2788         int64_t rc;
2789
2790         for (idx = 0; idx < phb->ioda.total_pe_num; idx++) {
2791                 if (map[idx] != pe->pe_number)
2792                         continue;
2793
2794                 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2795                                 phb->ioda.reserved_pe_idx, win, 0, idx);
2796
2797                 if (rc != OPAL_SUCCESS)
2798                         pe_warn(pe, "Error %lld unmapping (%d) segment#%d\n",
2799                                 rc, win, idx);
2800
2801                 map[idx] = IODA_INVALID_PE;
2802         }
2803 }
2804
2805 static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe *pe)
2806 {
2807         struct pnv_phb *phb = pe->phb;
2808
2809         if (phb->type == PNV_PHB_IODA1) {
2810                 pnv_ioda_free_pe_seg(pe, OPAL_IO_WINDOW_TYPE,
2811                                      phb->ioda.io_segmap);
2812                 pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
2813                                      phb->ioda.m32_segmap);
2814                 /* M64 is pre-configured by pnv_ioda1_init_m64() */
2815         } else if (phb->type == PNV_PHB_IODA2) {
2816                 pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
2817                                      phb->ioda.m32_segmap);
2818         }
2819 }
2820
2821 static void pnv_ioda_release_pe(struct pnv_ioda_pe *pe)
2822 {
2823         struct pnv_phb *phb = pe->phb;
2824         struct pnv_ioda_pe *slave, *tmp;
2825
2826         pe_info(pe, "Releasing PE\n");
2827
2828         mutex_lock(&phb->ioda.pe_list_mutex);
2829         list_del(&pe->list);
2830         mutex_unlock(&phb->ioda.pe_list_mutex);
2831
2832         switch (phb->type) {
2833         case PNV_PHB_IODA1:
2834                 pnv_pci_ioda1_release_pe_dma(pe);
2835                 break;
2836         case PNV_PHB_IODA2:
2837                 pnv_pci_ioda2_release_pe_dma(pe);
2838                 break;
2839         case PNV_PHB_NPU_OCAPI:
2840                 break;
2841         default:
2842                 WARN_ON(1);
2843         }
2844
2845         pnv_ioda_release_pe_seg(pe);
2846         pnv_ioda_deconfigure_pe(pe->phb, pe);
2847
2848         /* Release slave PEs in the compound PE */
2849         if (pe->flags & PNV_IODA_PE_MASTER) {
2850                 list_for_each_entry_safe(slave, tmp, &pe->slaves, list) {
2851                         list_del(&slave->list);
2852                         pnv_ioda_free_pe(slave);
2853                 }
2854         }
2855
2856         /*
2857          * The PE for root bus can be removed because of hotplug in EEH
2858          * recovery for fenced PHB error. We need to mark the PE dead so
2859          * that it can be populated again in PCI hot add path. The PE
2860          * shouldn't be destroyed as it's the global reserved resource.
2861          */
2862         if (phb->ioda.root_pe_idx == pe->pe_number)
2863                 return;
2864
2865         pnv_ioda_free_pe(pe);
2866 }
2867
2868 static void pnv_pci_release_device(struct pci_dev *pdev)
2869 {
2870         struct pnv_phb *phb = pci_bus_to_pnvhb(pdev->bus);
2871         struct pci_dn *pdn = pci_get_pdn(pdev);
2872         struct pnv_ioda_pe *pe;
2873
2874         /* The VF PE state is torn down when sriov_disable() is called */
2875         if (pdev->is_virtfn)
2876                 return;
2877
2878         if (!pdn || pdn->pe_number == IODA_INVALID_PE)
2879                 return;
2880
2881 #ifdef CONFIG_PCI_IOV
2882         /*
2883          * FIXME: Try move this to sriov_disable(). It's here since we allocate
2884          * the iov state at probe time since we need to fiddle with the IOV
2885          * resources.
2886          */
2887         if (pdev->is_physfn)
2888                 kfree(pdev->dev.archdata.iov_data);
2889 #endif
2890
2891         /*
2892          * PCI hotplug can happen as part of EEH error recovery. The @pdn
2893          * isn't removed and added afterwards in this scenario. We should
2894          * set the PE number in @pdn to an invalid one. Otherwise, the PE's
2895          * device count is decreased on removing devices while failing to
2896          * be increased on adding devices. It leads to unbalanced PE's device
2897          * count and eventually make normal PCI hotplug path broken.
2898          */
2899         pe = &phb->ioda.pe_array[pdn->pe_number];
2900         pdn->pe_number = IODA_INVALID_PE;
2901
2902         WARN_ON(--pe->device_count < 0);
2903         if (pe->device_count == 0)
2904                 pnv_ioda_release_pe(pe);
2905 }
2906
2907 static void pnv_pci_ioda_shutdown(struct pci_controller *hose)
2908 {
2909         struct pnv_phb *phb = hose->private_data;
2910
2911         opal_pci_reset(phb->opal_id, OPAL_RESET_PCI_IODA_TABLE,
2912                        OPAL_ASSERT_RESET);
2913 }
2914
2915 static void pnv_pci_ioda_dma_bus_setup(struct pci_bus *bus)
2916 {
2917         struct pnv_phb *phb = pci_bus_to_pnvhb(bus);
2918         struct pnv_ioda_pe *pe;
2919
2920         list_for_each_entry(pe, &phb->ioda.pe_list, list) {
2921                 if (!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)))
2922                         continue;
2923
2924                 if (!pe->pbus)
2925                         continue;
2926
2927                 if (bus->number == ((pe->rid >> 8) & 0xFF)) {
2928                         pe->pbus = bus;
2929                         break;
2930                 }
2931         }
2932 }
2933
2934 #ifdef CONFIG_IOMMU_API
2935 static struct iommu_group *pnv_pci_device_group(struct pci_controller *hose,
2936                                                 struct pci_dev *pdev)
2937 {
2938         struct pnv_phb *phb = hose->private_data;
2939         struct pnv_ioda_pe *pe;
2940
2941         if (WARN_ON(!phb))
2942                 return ERR_PTR(-ENODEV);
2943
2944         pe = pnv_pci_bdfn_to_pe(phb, pdev->devfn | (pdev->bus->number << 8));
2945         if (!pe)
2946                 return ERR_PTR(-ENODEV);
2947
2948         if (!pe->table_group.group)
2949                 return ERR_PTR(-ENODEV);
2950
2951         return iommu_group_ref_get(pe->table_group.group);
2952 }
2953 #endif
2954
2955 static const struct pci_controller_ops pnv_pci_ioda_controller_ops = {
2956         .dma_dev_setup          = pnv_pci_ioda_dma_dev_setup,
2957         .dma_bus_setup          = pnv_pci_ioda_dma_bus_setup,
2958         .iommu_bypass_supported = pnv_pci_ioda_iommu_bypass_supported,
2959         .enable_device_hook     = pnv_pci_enable_device_hook,
2960         .release_device         = pnv_pci_release_device,
2961         .window_alignment       = pnv_pci_window_alignment,
2962         .setup_bridge           = pnv_pci_fixup_bridge_resources,
2963         .reset_secondary_bus    = pnv_pci_reset_secondary_bus,
2964         .shutdown               = pnv_pci_ioda_shutdown,
2965 #ifdef CONFIG_IOMMU_API
2966         .device_group           = pnv_pci_device_group,
2967 #endif
2968 };
2969
2970 static const struct pci_controller_ops pnv_npu_ocapi_ioda_controller_ops = {
2971         .enable_device_hook     = pnv_ocapi_enable_device_hook,
2972         .release_device         = pnv_pci_release_device,
2973         .window_alignment       = pnv_pci_window_alignment,
2974         .reset_secondary_bus    = pnv_pci_reset_secondary_bus,
2975         .shutdown               = pnv_pci_ioda_shutdown,
2976 };
2977
2978 static void __init pnv_pci_init_ioda_phb(struct device_node *np,
2979                                          u64 hub_id, int ioda_type)
2980 {
2981         struct pci_controller *hose;
2982         struct pnv_phb *phb;
2983         unsigned long size, m64map_off, m32map_off, pemap_off;
2984         unsigned long iomap_off = 0, dma32map_off = 0;
2985         struct pnv_ioda_pe *root_pe;
2986         struct resource r;
2987         const __be64 *prop64;
2988         const __be32 *prop32;
2989         int len;
2990         unsigned int segno;
2991         u64 phb_id;
2992         void *aux;
2993         long rc;
2994
2995         if (!of_device_is_available(np))
2996                 return;
2997
2998         pr_info("Initializing %s PHB (%pOF)\n", pnv_phb_names[ioda_type], np);
2999
3000         prop64 = of_get_property(np, "ibm,opal-phbid", NULL);
3001         if (!prop64) {
3002                 pr_err("  Missing \"ibm,opal-phbid\" property !\n");
3003                 return;
3004         }
3005         phb_id = be64_to_cpup(prop64);
3006         pr_debug("  PHB-ID  : 0x%016llx\n", phb_id);
3007
3008         phb = kzalloc(sizeof(*phb), GFP_KERNEL);
3009         if (!phb)
3010                 panic("%s: Failed to allocate %zu bytes\n", __func__,
3011                       sizeof(*phb));
3012
3013         /* Allocate PCI controller */
3014         phb->hose = hose = pcibios_alloc_controller(np);
3015         if (!phb->hose) {
3016                 pr_err("  Can't allocate PCI controller for %pOF\n",
3017                        np);
3018                 memblock_free(phb, sizeof(struct pnv_phb));
3019                 return;
3020         }
3021
3022         spin_lock_init(&phb->lock);
3023         prop32 = of_get_property(np, "bus-range", &len);
3024         if (prop32 && len == 8) {
3025                 hose->first_busno = be32_to_cpu(prop32[0]);
3026                 hose->last_busno = be32_to_cpu(prop32[1]);
3027         } else {
3028                 pr_warn("  Broken <bus-range> on %pOF\n", np);
3029                 hose->first_busno = 0;
3030                 hose->last_busno = 0xff;
3031         }
3032         hose->private_data = phb;
3033         phb->hub_id = hub_id;
3034         phb->opal_id = phb_id;
3035         phb->type = ioda_type;
3036         mutex_init(&phb->ioda.pe_alloc_mutex);
3037
3038         /* Detect specific models for error handling */
3039         if (of_device_is_compatible(np, "ibm,p7ioc-pciex"))
3040                 phb->model = PNV_PHB_MODEL_P7IOC;
3041         else if (of_device_is_compatible(np, "ibm,power8-pciex"))
3042                 phb->model = PNV_PHB_MODEL_PHB3;
3043         else
3044                 phb->model = PNV_PHB_MODEL_UNKNOWN;
3045
3046         /* Initialize diagnostic data buffer */
3047         prop32 = of_get_property(np, "ibm,phb-diag-data-size", NULL);
3048         if (prop32)
3049                 phb->diag_data_size = be32_to_cpup(prop32);
3050         else
3051                 phb->diag_data_size = PNV_PCI_DIAG_BUF_SIZE;
3052
3053         phb->diag_data = kzalloc(phb->diag_data_size, GFP_KERNEL);
3054         if (!phb->diag_data)
3055                 panic("%s: Failed to allocate %u bytes\n", __func__,
3056                       phb->diag_data_size);
3057
3058         /* Parse 32-bit and IO ranges (if any) */
3059         pci_process_bridge_OF_ranges(hose, np, !hose->global_number);
3060
3061         /* Get registers */
3062         if (!of_address_to_resource(np, 0, &r)) {
3063                 phb->regs_phys = r.start;
3064                 phb->regs = ioremap(r.start, resource_size(&r));
3065                 if (phb->regs == NULL)
3066                         pr_err("  Failed to map registers !\n");
3067         }
3068
3069         /* Initialize more IODA stuff */
3070         phb->ioda.total_pe_num = 1;
3071         prop32 = of_get_property(np, "ibm,opal-num-pes", NULL);
3072         if (prop32)
3073                 phb->ioda.total_pe_num = be32_to_cpup(prop32);
3074         prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL);
3075         if (prop32)
3076                 phb->ioda.reserved_pe_idx = be32_to_cpup(prop32);
3077
3078         /* Invalidate RID to PE# mapping */
3079         for (segno = 0; segno < ARRAY_SIZE(phb->ioda.pe_rmap); segno++)
3080                 phb->ioda.pe_rmap[segno] = IODA_INVALID_PE;
3081
3082         /* Parse 64-bit MMIO range */
3083         pnv_ioda_parse_m64_window(phb);
3084
3085         phb->ioda.m32_size = resource_size(&hose->mem_resources[0]);
3086         /* FW Has already off top 64k of M32 space (MSI space) */
3087         phb->ioda.m32_size += 0x10000;
3088
3089         phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe_num;
3090         phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0];
3091         phb->ioda.io_size = hose->pci_io_size;
3092         phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe_num;
3093         phb->ioda.io_pci_base = 0; /* XXX calculate this ? */
3094
3095         /* Calculate how many 32-bit TCE segments we have */
3096         phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3097                                 PNV_IODA1_DMA32_SEGSIZE;
3098
3099         /* Allocate aux data & arrays. We don't have IO ports on PHB3 */
3100         size = ALIGN(max_t(unsigned, phb->ioda.total_pe_num, 8) / 8,
3101                         sizeof(unsigned long));
3102         m64map_off = size;
3103         size += phb->ioda.total_pe_num * sizeof(phb->ioda.m64_segmap[0]);
3104         m32map_off = size;
3105         size += phb->ioda.total_pe_num * sizeof(phb->ioda.m32_segmap[0]);
3106         if (phb->type == PNV_PHB_IODA1) {
3107                 iomap_off = size;
3108                 size += phb->ioda.total_pe_num * sizeof(phb->ioda.io_segmap[0]);
3109                 dma32map_off = size;
3110                 size += phb->ioda.dma32_count *
3111                         sizeof(phb->ioda.dma32_segmap[0]);
3112         }
3113         pemap_off = size;
3114         size += phb->ioda.total_pe_num * sizeof(struct pnv_ioda_pe);
3115         aux = kzalloc(size, GFP_KERNEL);
3116         if (!aux)
3117                 panic("%s: Failed to allocate %lu bytes\n", __func__, size);
3118
3119         phb->ioda.pe_alloc = aux;
3120         phb->ioda.m64_segmap = aux + m64map_off;
3121         phb->ioda.m32_segmap = aux + m32map_off;
3122         for (segno = 0; segno < phb->ioda.total_pe_num; segno++) {
3123                 phb->ioda.m64_segmap[segno] = IODA_INVALID_PE;
3124                 phb->ioda.m32_segmap[segno] = IODA_INVALID_PE;
3125         }
3126         if (phb->type == PNV_PHB_IODA1) {
3127                 phb->ioda.io_segmap = aux + iomap_off;
3128                 for (segno = 0; segno < phb->ioda.total_pe_num; segno++)
3129                         phb->ioda.io_segmap[segno] = IODA_INVALID_PE;
3130
3131                 phb->ioda.dma32_segmap = aux + dma32map_off;
3132                 for (segno = 0; segno < phb->ioda.dma32_count; segno++)
3133                         phb->ioda.dma32_segmap[segno] = IODA_INVALID_PE;
3134         }
3135         phb->ioda.pe_array = aux + pemap_off;
3136
3137         /*
3138          * Choose PE number for root bus, which shouldn't have
3139          * M64 resources consumed by its child devices. To pick
3140          * the PE number adjacent to the reserved one if possible.
3141          */
3142         pnv_ioda_reserve_pe(phb, phb->ioda.reserved_pe_idx);
3143         if (phb->ioda.reserved_pe_idx == 0) {
3144                 phb->ioda.root_pe_idx = 1;
3145                 pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3146         } else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1)) {
3147                 phb->ioda.root_pe_idx = phb->ioda.reserved_pe_idx - 1;
3148                 pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3149         } else {
3150                 /* otherwise just allocate one */
3151                 root_pe = pnv_ioda_alloc_pe(phb, 1);
3152                 phb->ioda.root_pe_idx = root_pe->pe_number;
3153         }
3154
3155         INIT_LIST_HEAD(&phb->ioda.pe_list);
3156         mutex_init(&phb->ioda.pe_list_mutex);
3157
3158         /* Calculate how many 32-bit TCE segments we have */
3159         phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3160                                 PNV_IODA1_DMA32_SEGSIZE;
3161
3162 #if 0 /* We should really do that ... */
3163         rc = opal_pci_set_phb_mem_window(opal->phb_id,
3164                                          window_type,
3165                                          window_num,
3166                                          starting_real_address,
3167                                          starting_pci_address,
3168                                          segment_size);
3169 #endif
3170
3171         pr_info("  %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
3172                 phb->ioda.total_pe_num, phb->ioda.reserved_pe_idx,
3173                 phb->ioda.m32_size, phb->ioda.m32_segsize);
3174         if (phb->ioda.m64_size)
3175                 pr_info("                 M64: 0x%lx [segment=0x%lx]\n",
3176                         phb->ioda.m64_size, phb->ioda.m64_segsize);
3177         if (phb->ioda.io_size)
3178                 pr_info("                  IO: 0x%x [segment=0x%x]\n",
3179                         phb->ioda.io_size, phb->ioda.io_segsize);
3180
3181
3182         phb->hose->ops = &pnv_pci_ops;
3183         phb->get_pe_state = pnv_ioda_get_pe_state;
3184         phb->freeze_pe = pnv_ioda_freeze_pe;
3185         phb->unfreeze_pe = pnv_ioda_unfreeze_pe;
3186
3187         /* Setup MSI support */
3188         pnv_pci_init_ioda_msis(phb);
3189
3190         /*
3191          * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
3192          * to let the PCI core do resource assignment. It's supposed
3193          * that the PCI core will do correct I/O and MMIO alignment
3194          * for the P2P bridge bars so that each PCI bus (excluding
3195          * the child P2P bridges) can form individual PE.
3196          */
3197         ppc_md.pcibios_fixup = pnv_pci_ioda_fixup;
3198
3199         switch (phb->type) {
3200         case PNV_PHB_NPU_OCAPI:
3201                 hose->controller_ops = pnv_npu_ocapi_ioda_controller_ops;
3202                 break;
3203         default:
3204                 hose->controller_ops = pnv_pci_ioda_controller_ops;
3205         }
3206
3207         ppc_md.pcibios_default_alignment = pnv_pci_default_alignment;
3208
3209 #ifdef CONFIG_PCI_IOV
3210         ppc_md.pcibios_fixup_sriov = pnv_pci_ioda_fixup_iov;
3211         ppc_md.pcibios_iov_resource_alignment = pnv_pci_iov_resource_alignment;
3212         ppc_md.pcibios_sriov_enable = pnv_pcibios_sriov_enable;
3213         ppc_md.pcibios_sriov_disable = pnv_pcibios_sriov_disable;
3214 #endif
3215
3216         pci_add_flags(PCI_REASSIGN_ALL_RSRC);
3217
3218         /* Reset IODA tables to a clean state */
3219         rc = opal_pci_reset(phb_id, OPAL_RESET_PCI_IODA_TABLE, OPAL_ASSERT_RESET);
3220         if (rc)
3221                 pr_warn("  OPAL Error %ld performing IODA table reset !\n", rc);
3222
3223         /*
3224          * If we're running in kdump kernel, the previous kernel never
3225          * shutdown PCI devices correctly. We already got IODA table
3226          * cleaned out. So we have to issue PHB reset to stop all PCI
3227          * transactions from previous kernel. The ppc_pci_reset_phbs
3228          * kernel parameter will force this reset too. Additionally,
3229          * if the IODA reset above failed then use a bigger hammer.
3230          * This can happen if we get a PHB fatal error in very early
3231          * boot.
3232          */
3233         if (is_kdump_kernel() || pci_reset_phbs || rc) {
3234                 pr_info("  Issue PHB reset ...\n");
3235                 pnv_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL);
3236                 pnv_eeh_phb_reset(hose, EEH_RESET_DEACTIVATE);
3237         }
3238
3239         /* Remove M64 resource if we can't configure it successfully */
3240         if (!phb->init_m64 || phb->init_m64(phb))
3241                 hose->mem_resources[1].flags = 0;
3242
3243         /* create pci_dn's for DT nodes under this PHB */
3244         pci_devs_phb_init_dynamic(hose);
3245 }
3246
3247 void __init pnv_pci_init_ioda2_phb(struct device_node *np)
3248 {
3249         pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2);
3250 }
3251
3252 void __init pnv_pci_init_npu2_opencapi_phb(struct device_node *np)
3253 {
3254         pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU_OCAPI);
3255 }
3256
3257 static void pnv_npu2_opencapi_cfg_size_fixup(struct pci_dev *dev)
3258 {
3259         struct pnv_phb *phb = pci_bus_to_pnvhb(dev->bus);
3260
3261         if (!machine_is(powernv))
3262                 return;
3263
3264         if (phb->type == PNV_PHB_NPU_OCAPI)
3265                 dev->cfg_size = PCI_CFG_SPACE_EXP_SIZE;
3266 }
3267 DECLARE_PCI_FIXUP_EARLY(PCI_ANY_ID, PCI_ANY_ID, pnv_npu2_opencapi_cfg_size_fixup);
3268
3269 void __init pnv_pci_init_ioda_hub(struct device_node *np)
3270 {
3271         struct device_node *phbn;
3272         const __be64 *prop64;
3273         u64 hub_id;
3274
3275         pr_info("Probing IODA IO-Hub %pOF\n", np);
3276
3277         prop64 = of_get_property(np, "ibm,opal-hubid", NULL);
3278         if (!prop64) {
3279                 pr_err(" Missing \"ibm,opal-hubid\" property !\n");
3280                 return;
3281         }
3282         hub_id = be64_to_cpup(prop64);
3283         pr_devel(" HUB-ID : 0x%016llx\n", hub_id);
3284
3285         /* Count child PHBs */
3286         for_each_child_of_node(np, phbn) {
3287                 /* Look for IODA1 PHBs */
3288                 if (of_device_is_compatible(phbn, "ibm,ioda-phb"))
3289                         pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1);
3290         }
3291 }