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

   1 /*
   2  * Support PCI/PCIe on PowerNV platforms
   3  *
   4  * Copyright 2011 Benjamin Herrenschmidt, IBM Corp.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the License, or (at your option) any later version.
  10  */
  11
  12 #undef DEBUG
  13
  14 #include <linux/kernel.h>
  15 #include <linux/pci.h>
  16 #include <linux/crash_dump.h>
  17 #include <linux/delay.h>
  18 #include <linux/string.h>
  19 #include <linux/init.h>
  20 #include <linux/bootmem.h>
  21 #include <linux/irq.h>
  22 #include <linux/io.h>
  23 #include <linux/msi.h>
  24 #include <linux/memblock.h>
  25 #include <linux/iommu.h>
  26 #include <linux/rculist.h>
  27 #include <linux/sizes.h>
  28
  29 #include <asm/sections.h>
  30 #include <asm/io.h>
  31 #include <asm/prom.h>
  32 #include <asm/pci-bridge.h>
  33 #include <asm/machdep.h>
  34 #include <asm/msi_bitmap.h>
  35 #include <asm/ppc-pci.h>
  36 #include <asm/opal.h>
  37 #include <asm/iommu.h>
  38 #include <asm/tce.h>
  39 #include <asm/xics.h>
  40 #include <asm/debugfs.h>
  41 #include <asm/firmware.h>
  42 #include <asm/pnv-pci.h>
  43 #include <asm/mmzone.h>
  44
  45 #include <misc/cxl-base.h>
  46
  47 #include "powernv.h"
  48 #include "pci.h"
  49 #include "../../../../drivers/pci/pci.h"
  50
  51 #define PNV_IODA1_M64_NUM       16      /* Number of M64 BARs   */
  52 #define PNV_IODA1_M64_SEGS      8       /* Segments per M64 BAR */
  53 #define PNV_IODA1_DMA32_SEGSIZE 0x10000000
  54
  55 static const char * const pnv_phb_names[] = { "IODA1", "IODA2", "NPU_NVLINK",
  56                                               "NPU_OCAPI" };
  57
  58 void pe_level_printk(const struct pnv_ioda_pe *pe, const char *level,
  59                             const char *fmt, ...)
  60 {
  61         struct va_format vaf;
  62         va_list args;
  63         char pfix[32];
  64
  65         va_start(args, fmt);
  66
  67         vaf.fmt = fmt;
  68         vaf.va = &args;
  69
  70         if (pe->flags & PNV_IODA_PE_DEV)
  71                 strlcpy(pfix, dev_name(&pe->pdev->dev), sizeof(pfix));
  72         else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
  73                 sprintf(pfix, "%04x:%02x     ",
  74                         pci_domain_nr(pe->pbus), pe->pbus->number);
  75 #ifdef CONFIG_PCI_IOV
  76         else if (pe->flags & PNV_IODA_PE_VF)
  77                 sprintf(pfix, "%04x:%02x:%2x.%d",
  78                         pci_domain_nr(pe->parent_dev->bus),
  79                         (pe->rid & 0xff00) >> 8,
  80                         PCI_SLOT(pe->rid), PCI_FUNC(pe->rid));
  81 #endif /* CONFIG_PCI_IOV*/
  82
  83         printk("%spci %s: [PE# %.2x] %pV",
  84                level, pfix, pe->pe_number, &vaf);
  85
  86         va_end(args);
  87 }
  88
  89 static bool pnv_iommu_bypass_disabled __read_mostly;
  90 static bool pci_reset_phbs __read_mostly;
  91
  92 static int __init iommu_setup(char *str)
  93 {
  94         if (!str)
  95                 return -EINVAL;
  96
  97         while (*str) {
  98                 if (!strncmp(str, "nobypass", 8)) {
  99                         pnv_iommu_bypass_disabled = true;
 100                         pr_info("PowerNV: IOMMU bypass window disabled.\n");
 101                         break;
 102                 }
 103                 str += strcspn(str, ",");
 104                 if (*str == ',')
 105                         str++;
 106         }
 107
 108         return 0;
 109 }
 110 early_param("iommu", iommu_setup);
 111
 112 static int __init pci_reset_phbs_setup(char *str)
 113 {
 114         pci_reset_phbs = true;
 115         return 0;
 116 }
 117
 118 early_param("ppc_pci_reset_phbs", pci_reset_phbs_setup);
 119
 120 static inline bool pnv_pci_is_m64(struct pnv_phb *phb, struct resource *r)
 121 {
 122         /*
 123          * WARNING: We cannot rely on the resource flags. The Linux PCI
 124          * allocation code sometimes decides to put a 64-bit prefetchable
 125          * BAR in the 32-bit window, so we have to compare the addresses.
 126          *
 127          * For simplicity we only test resource start.
 128          */
 129         return (r->start >= phb->ioda.m64_base &&
 130                 r->start < (phb->ioda.m64_base + phb->ioda.m64_size));
 131 }
 132
 133 static inline bool pnv_pci_is_m64_flags(unsigned long resource_flags)
 134 {
 135         unsigned long flags = (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
 136
 137         return (resource_flags & flags) == flags;
 138 }
 139
 140 static struct pnv_ioda_pe *pnv_ioda_init_pe(struct pnv_phb *phb, int pe_no)
 141 {
 142         s64 rc;
 143
 144         phb->ioda.pe_array[pe_no].phb = phb;
 145         phb->ioda.pe_array[pe_no].pe_number = pe_no;
 146
 147         /*
 148          * Clear the PE frozen state as it might be put into frozen state
 149          * in the last PCI remove path. It's not harmful to do so when the
 150          * PE is already in unfrozen state.
 151          */
 152         rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no,
 153                                        OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 154         if (rc != OPAL_SUCCESS && rc != OPAL_UNSUPPORTED)
 155                 pr_warn("%s: Error %lld unfreezing PHB#%x-PE#%x\n",
 156                         __func__, rc, phb->hose->global_number, pe_no);
 157
 158         return &phb->ioda.pe_array[pe_no];
 159 }
 160
 161 static void pnv_ioda_reserve_pe(struct pnv_phb *phb, int pe_no)
 162 {
 163         if (!(pe_no >= 0 && pe_no < phb->ioda.total_pe_num)) {
 164                 pr_warn("%s: Invalid PE %x on PHB#%x\n",
 165                         __func__, pe_no, phb->hose->global_number);
 166                 return;
 167         }
 168
 169         if (test_and_set_bit(pe_no, phb->ioda.pe_alloc))
 170                 pr_debug("%s: PE %x was reserved on PHB#%x\n",
 171                          __func__, pe_no, phb->hose->global_number);
 172
 173         pnv_ioda_init_pe(phb, pe_no);
 174 }
 175
 176 static struct pnv_ioda_pe *pnv_ioda_alloc_pe(struct pnv_phb *phb)
 177 {
 178         long pe;
 179
 180         for (pe = phb->ioda.total_pe_num - 1; pe >= 0; pe--) {
 181                 if (!test_and_set_bit(pe, phb->ioda.pe_alloc))
 182                         return pnv_ioda_init_pe(phb, pe);
 183         }
 184
 185         return NULL;
 186 }
 187
 188 static void pnv_ioda_free_pe(struct pnv_ioda_pe *pe)
 189 {
 190         struct pnv_phb *phb = pe->phb;
 191         unsigned int pe_num = pe->pe_number;
 192
 193         WARN_ON(pe->pdev);
 194
 195         memset(pe, 0, sizeof(struct pnv_ioda_pe));
 196         clear_bit(pe_num, phb->ioda.pe_alloc);
 197 }
 198
 199 /* The default M64 BAR is shared by all PEs */
 200 static int pnv_ioda2_init_m64(struct pnv_phb *phb)
 201 {
 202         const char *desc;
 203         struct resource *r;
 204         s64 rc;
 205
 206         /* Configure the default M64 BAR */
 207         rc = opal_pci_set_phb_mem_window(phb->opal_id,
 208                                          OPAL_M64_WINDOW_TYPE,
 209                                          phb->ioda.m64_bar_idx,
 210                                          phb->ioda.m64_base,
 211                                          0, /* unused */
 212                                          phb->ioda.m64_size);
 213         if (rc != OPAL_SUCCESS) {
 214                 desc = "configuring";
 215                 goto fail;
 216         }
 217
 218         /* Enable the default M64 BAR */
 219         rc = opal_pci_phb_mmio_enable(phb->opal_id,
 220                                       OPAL_M64_WINDOW_TYPE,
 221                                       phb->ioda.m64_bar_idx,
 222                                       OPAL_ENABLE_M64_SPLIT);
 223         if (rc != OPAL_SUCCESS) {
 224                 desc = "enabling";
 225                 goto fail;
 226         }
 227
 228         /*
 229          * Exclude the segments for reserved and root bus PE, which
 230          * are first or last two PEs.
 231          */
 232         r = &phb->hose->mem_resources[1];
 233         if (phb->ioda.reserved_pe_idx == 0)
 234                 r->start += (2 * phb->ioda.m64_segsize);
 235         else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
 236                 r->end -= (2 * phb->ioda.m64_segsize);
 237         else
 238                 pr_warn("  Cannot strip M64 segment for reserved PE#%x\n",
 239                         phb->ioda.reserved_pe_idx);
 240
 241         return 0;
 242
 243 fail:
 244         pr_warn("  Failure %lld %s M64 BAR#%d\n",
 245                 rc, desc, phb->ioda.m64_bar_idx);
 246         opal_pci_phb_mmio_enable(phb->opal_id,
 247                                  OPAL_M64_WINDOW_TYPE,
 248                                  phb->ioda.m64_bar_idx,
 249                                  OPAL_DISABLE_M64);
 250         return -EIO;
 251 }
 252
 253 static void pnv_ioda_reserve_dev_m64_pe(struct pci_dev *pdev,
 254                                          unsigned long *pe_bitmap)
 255 {
 256         struct pci_controller *hose = pci_bus_to_host(pdev->bus);
 257         struct pnv_phb *phb = hose->private_data;
 258         struct resource *r;
 259         resource_size_t base, sgsz, start, end;
 260         int segno, i;
 261
 262         base = phb->ioda.m64_base;
 263         sgsz = phb->ioda.m64_segsize;
 264         for (i = 0; i <= PCI_ROM_RESOURCE; i++) {
 265                 r = &pdev->resource[i];
 266                 if (!r->parent || !pnv_pci_is_m64(phb, r))
 267                         continue;
 268
 269                 start = _ALIGN_DOWN(r->start - base, sgsz);
 270                 end = _ALIGN_UP(r->end - base, sgsz);
 271                 for (segno = start / sgsz; segno < end / sgsz; segno++) {
 272                         if (pe_bitmap)
 273                                 set_bit(segno, pe_bitmap);
 274                         else
 275                                 pnv_ioda_reserve_pe(phb, segno);
 276                 }
 277         }
 278 }
 279
 280 static int pnv_ioda1_init_m64(struct pnv_phb *phb)
 281 {
 282         struct resource *r;
 283         int index;
 284
 285         /*
 286          * There are 16 M64 BARs, each of which has 8 segments. So
 287          * there are as many M64 segments as the maximum number of
 288          * PEs, which is 128.
 289          */
 290         for (index = 0; index < PNV_IODA1_M64_NUM; index++) {
 291                 unsigned long base, segsz = phb->ioda.m64_segsize;
 292                 int64_t rc;
 293
 294                 base = phb->ioda.m64_base +
 295                        index * PNV_IODA1_M64_SEGS * segsz;
 296                 rc = opal_pci_set_phb_mem_window(phb->opal_id,
 297                                 OPAL_M64_WINDOW_TYPE, index, base, 0,
 298                                 PNV_IODA1_M64_SEGS * segsz);
 299                 if (rc != OPAL_SUCCESS) {
 300                         pr_warn("  Error %lld setting M64 PHB#%x-BAR#%d\n",
 301                                 rc, phb->hose->global_number, index);
 302                         goto fail;
 303                 }
 304
 305                 rc = opal_pci_phb_mmio_enable(phb->opal_id,
 306                                 OPAL_M64_WINDOW_TYPE, index,
 307                                 OPAL_ENABLE_M64_SPLIT);
 308                 if (rc != OPAL_SUCCESS) {
 309                         pr_warn("  Error %lld enabling M64 PHB#%x-BAR#%d\n",
 310                                 rc, phb->hose->global_number, index);
 311                         goto fail;
 312                 }
 313         }
 314
 315         /*
 316          * Exclude the segments for reserved and root bus PE, which
 317          * are first or last two PEs.
 318          */
 319         r = &phb->hose->mem_resources[1];
 320         if (phb->ioda.reserved_pe_idx == 0)
 321                 r->start += (2 * phb->ioda.m64_segsize);
 322         else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
 323                 r->end -= (2 * phb->ioda.m64_segsize);
 324         else
 325                 WARN(1, "Wrong reserved PE#%x on PHB#%x\n",
 326                      phb->ioda.reserved_pe_idx, phb->hose->global_number);
 327
 328         return 0;
 329
 330 fail:
 331         for ( ; index >= 0; index--)
 332                 opal_pci_phb_mmio_enable(phb->opal_id,
 333                         OPAL_M64_WINDOW_TYPE, index, OPAL_DISABLE_M64);
 334
 335         return -EIO;
 336 }
 337
 338 static void pnv_ioda_reserve_m64_pe(struct pci_bus *bus,
 339                                     unsigned long *pe_bitmap,
 340                                     bool all)
 341 {
 342         struct pci_dev *pdev;
 343
 344         list_for_each_entry(pdev, &bus->devices, bus_list) {
 345                 pnv_ioda_reserve_dev_m64_pe(pdev, pe_bitmap);
 346
 347                 if (all && pdev->subordinate)
 348                         pnv_ioda_reserve_m64_pe(pdev->subordinate,
 349                                                 pe_bitmap, all);
 350         }
 351 }
 352
 353 static struct pnv_ioda_pe *pnv_ioda_pick_m64_pe(struct pci_bus *bus, bool all)
 354 {
 355         struct pci_controller *hose = pci_bus_to_host(bus);
 356         struct pnv_phb *phb = hose->private_data;
 357         struct pnv_ioda_pe *master_pe, *pe;
 358         unsigned long size, *pe_alloc;
 359         int i;
 360
 361         /* Root bus shouldn't use M64 */
 362         if (pci_is_root_bus(bus))
 363                 return NULL;
 364
 365         /* Allocate bitmap */
 366         size = _ALIGN_UP(phb->ioda.total_pe_num / 8, sizeof(unsigned long));
 367         pe_alloc = kzalloc(size, GFP_KERNEL);
 368         if (!pe_alloc) {
 369                 pr_warn("%s: Out of memory !\n",
 370                         __func__);
 371                 return NULL;
 372         }
 373
 374         /* Figure out reserved PE numbers by the PE */
 375         pnv_ioda_reserve_m64_pe(bus, pe_alloc, all);
 376
 377         /*
 378          * the current bus might not own M64 window and that's all
 379          * contributed by its child buses. For the case, we needn't
 380          * pick M64 dependent PE#.
 381          */
 382         if (bitmap_empty(pe_alloc, phb->ioda.total_pe_num)) {
 383                 kfree(pe_alloc);
 384                 return NULL;
 385         }
 386
 387         /*
 388          * Figure out the master PE and put all slave PEs to master
 389          * PE's list to form compound PE.
 390          */
 391         master_pe = NULL;
 392         i = -1;
 393         while ((i = find_next_bit(pe_alloc, phb->ioda.total_pe_num, i + 1)) <
 394                 phb->ioda.total_pe_num) {
 395                 pe = &phb->ioda.pe_array[i];
 396
 397                 phb->ioda.m64_segmap[pe->pe_number] = pe->pe_number;
 398                 if (!master_pe) {
 399                         pe->flags |= PNV_IODA_PE_MASTER;
 400                         INIT_LIST_HEAD(&pe->slaves);
 401                         master_pe = pe;
 402                 } else {
 403                         pe->flags |= PNV_IODA_PE_SLAVE;
 404                         pe->master = master_pe;
 405                         list_add_tail(&pe->list, &master_pe->slaves);
 406                 }
 407
 408                 /*
 409                  * P7IOC supports M64DT, which helps mapping M64 segment
 410                  * to one particular PE#. However, PHB3 has fixed mapping
 411                  * between M64 segment and PE#. In order to have same logic
 412                  * for P7IOC and PHB3, we enforce fixed mapping between M64
 413                  * segment and PE# on P7IOC.
 414                  */
 415                 if (phb->type == PNV_PHB_IODA1) {
 416                         int64_t rc;
 417
 418                         rc = opal_pci_map_pe_mmio_window(phb->opal_id,
 419                                         pe->pe_number, OPAL_M64_WINDOW_TYPE,
 420                                         pe->pe_number / PNV_IODA1_M64_SEGS,
 421                                         pe->pe_number % PNV_IODA1_M64_SEGS);
 422                         if (rc != OPAL_SUCCESS)
 423                                 pr_warn("%s: Error %lld mapping M64 for PHB#%x-PE#%x\n",
 424                                         __func__, rc, phb->hose->global_number,
 425                                         pe->pe_number);
 426                 }
 427         }
 428
 429         kfree(pe_alloc);
 430         return master_pe;
 431 }
 432
 433 static void __init pnv_ioda_parse_m64_window(struct pnv_phb *phb)
 434 {
 435         struct pci_controller *hose = phb->hose;
 436         struct device_node *dn = hose->dn;
 437         struct resource *res;
 438         u32 m64_range[2], i;
 439         const __be32 *r;
 440         u64 pci_addr;
 441
 442         if (phb->type != PNV_PHB_IODA1 && phb->type != PNV_PHB_IODA2) {
 443                 pr_info("  Not support M64 window\n");
 444                 return;
 445         }
 446
 447         if (!firmware_has_feature(FW_FEATURE_OPAL)) {
 448                 pr_info("  Firmware too old to support M64 window\n");
 449                 return;
 450         }
 451
 452         r = of_get_property(dn, "ibm,opal-m64-window", NULL);
 453         if (!r) {
 454                 pr_info("  No <ibm,opal-m64-window> on %pOF\n",
 455                         dn);
 456                 return;
 457         }
 458
 459         /*
 460          * Find the available M64 BAR range and pickup the last one for
 461          * covering the whole 64-bits space. We support only one range.
 462          */
 463         if (of_property_read_u32_array(dn, "ibm,opal-available-m64-ranges",
 464                                        m64_range, 2)) {
 465                 /* In absence of the property, assume 0..15 */
 466                 m64_range[0] = 0;
 467                 m64_range[1] = 16;
 468         }
 469         /* We only support 64 bits in our allocator */
 470         if (m64_range[1] > 63) {
 471                 pr_warn("%s: Limiting M64 range to 63 (from %d) on PHB#%x\n",
 472                         __func__, m64_range[1], phb->hose->global_number);
 473                 m64_range[1] = 63;
 474         }
 475         /* Empty range, no m64 */
 476         if (m64_range[1] <= m64_range[0]) {
 477                 pr_warn("%s: M64 empty, disabling M64 usage on PHB#%x\n",
 478                         __func__, phb->hose->global_number);
 479                 return;
 480         }
 481
 482         /* Configure M64 informations */
 483         res = &hose->mem_resources[1];
 484         res->name = dn->full_name;
 485         res->start = of_translate_address(dn, r + 2);
 486         res->end = res->start + of_read_number(r + 4, 2) - 1;
 487         res->flags = (IORESOURCE_MEM | IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
 488         pci_addr = of_read_number(r, 2);
 489         hose->mem_offset[1] = res->start - pci_addr;
 490
 491         phb->ioda.m64_size = resource_size(res);
 492         phb->ioda.m64_segsize = phb->ioda.m64_size / phb->ioda.total_pe_num;
 493         phb->ioda.m64_base = pci_addr;
 494
 495         /* This lines up nicely with the display from processing OF ranges */
 496         pr_info(" MEM 0x%016llx..0x%016llx -> 0x%016llx (M64 #%d..%d)\n",
 497                 res->start, res->end, pci_addr, m64_range[0],
 498                 m64_range[0] + m64_range[1] - 1);
 499
 500         /* Mark all M64 used up by default */
 501         phb->ioda.m64_bar_alloc = (unsigned long)-1;
 502
 503         /* Use last M64 BAR to cover M64 window */
 504         m64_range[1]--;
 505         phb->ioda.m64_bar_idx = m64_range[0] + m64_range[1];
 506
 507         pr_info(" Using M64 #%d as default window\n", phb->ioda.m64_bar_idx);
 508
 509         /* Mark remaining ones free */
 510         for (i = m64_range[0]; i < m64_range[1]; i++)
 511                 clear_bit(i, &phb->ioda.m64_bar_alloc);
 512
 513         /*
 514          * Setup init functions for M64 based on IODA version, IODA3 uses
 515          * the IODA2 code.
 516          */
 517         if (phb->type == PNV_PHB_IODA1)
 518                 phb->init_m64 = pnv_ioda1_init_m64;
 519         else
 520                 phb->init_m64 = pnv_ioda2_init_m64;
 521         phb->reserve_m64_pe = pnv_ioda_reserve_m64_pe;
 522         phb->pick_m64_pe = pnv_ioda_pick_m64_pe;
 523 }
 524
 525 static void pnv_ioda_freeze_pe(struct pnv_phb *phb, int pe_no)
 526 {
 527         struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_no];
 528         struct pnv_ioda_pe *slave;
 529         s64 rc;
 530
 531         /* Fetch master PE */
 532         if (pe->flags & PNV_IODA_PE_SLAVE) {
 533                 pe = pe->master;
 534                 if (WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)))
 535                         return;
 536
 537                 pe_no = pe->pe_number;
 538         }
 539
 540         /* Freeze master PE */
 541         rc = opal_pci_eeh_freeze_set(phb->opal_id,
 542                                      pe_no,
 543                                      OPAL_EEH_ACTION_SET_FREEZE_ALL);
 544         if (rc != OPAL_SUCCESS) {
 545                 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
 546                         __func__, rc, phb->hose->global_number, pe_no);
 547                 return;
 548         }
 549
 550         /* Freeze slave PEs */
 551         if (!(pe->flags & PNV_IODA_PE_MASTER))
 552                 return;
 553
 554         list_for_each_entry(slave, &pe->slaves, list) {
 555                 rc = opal_pci_eeh_freeze_set(phb->opal_id,
 556                                              slave->pe_number,
 557                                              OPAL_EEH_ACTION_SET_FREEZE_ALL);
 558                 if (rc != OPAL_SUCCESS)
 559                         pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
 560                                 __func__, rc, phb->hose->global_number,
 561                                 slave->pe_number);
 562         }
 563 }
 564
 565 static int pnv_ioda_unfreeze_pe(struct pnv_phb *phb, int pe_no, int opt)
 566 {
 567         struct pnv_ioda_pe *pe, *slave;
 568         s64 rc;
 569
 570         /* Find master PE */
 571         pe = &phb->ioda.pe_array[pe_no];
 572         if (pe->flags & PNV_IODA_PE_SLAVE) {
 573                 pe = pe->master;
 574                 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
 575                 pe_no = pe->pe_number;
 576         }
 577
 578         /* Clear frozen state for master PE */
 579         rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, opt);
 580         if (rc != OPAL_SUCCESS) {
 581                 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
 582                         __func__, rc, opt, phb->hose->global_number, pe_no);
 583                 return -EIO;
 584         }
 585
 586         if (!(pe->flags & PNV_IODA_PE_MASTER))
 587                 return 0;
 588
 589         /* Clear frozen state for slave PEs */
 590         list_for_each_entry(slave, &pe->slaves, list) {
 591                 rc = opal_pci_eeh_freeze_clear(phb->opal_id,
 592                                              slave->pe_number,
 593                                              opt);
 594                 if (rc != OPAL_SUCCESS) {
 595                         pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
 596                                 __func__, rc, opt, phb->hose->global_number,
 597                                 slave->pe_number);
 598                         return -EIO;
 599                 }
 600         }
 601
 602         return 0;
 603 }
 604
 605 static int pnv_ioda_get_pe_state(struct pnv_phb *phb, int pe_no)
 606 {
 607         struct pnv_ioda_pe *slave, *pe;
 608         u8 fstate = 0, state;
 609         __be16 pcierr = 0;
 610         s64 rc;
 611
 612         /* Sanity check on PE number */
 613         if (pe_no < 0 || pe_no >= phb->ioda.total_pe_num)
 614                 return OPAL_EEH_STOPPED_PERM_UNAVAIL;
 615
 616         /*
 617          * Fetch the master PE and the PE instance might be
 618          * not initialized yet.
 619          */
 620         pe = &phb->ioda.pe_array[pe_no];
 621         if (pe->flags & PNV_IODA_PE_SLAVE) {
 622                 pe = pe->master;
 623                 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
 624                 pe_no = pe->pe_number;
 625         }
 626
 627         /* Check the master PE */
 628         rc = opal_pci_eeh_freeze_status(phb->opal_id, pe_no,
 629                                         &state, &pcierr, NULL);
 630         if (rc != OPAL_SUCCESS) {
 631                 pr_warn("%s: Failure %lld getting "
 632                         "PHB#%x-PE#%x state\n",
 633                         __func__, rc,
 634                         phb->hose->global_number, pe_no);
 635                 return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
 636         }
 637
 638         /* Check the slave PE */
 639         if (!(pe->flags & PNV_IODA_PE_MASTER))
 640                 return state;
 641
 642         list_for_each_entry(slave, &pe->slaves, list) {
 643                 rc = opal_pci_eeh_freeze_status(phb->opal_id,
 644                                                 slave->pe_number,
 645                                                 &fstate,
 646                                                 &pcierr,
 647                                                 NULL);
 648                 if (rc != OPAL_SUCCESS) {
 649                         pr_warn("%s: Failure %lld getting "
 650                                 "PHB#%x-PE#%x state\n",
 651                                 __func__, rc,
 652                                 phb->hose->global_number, slave->pe_number);
 653                         return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
 654                 }
 655
 656                 /*
 657                  * Override the result based on the ascending
 658                  * priority.
 659                  */
 660                 if (fstate > state)
 661                         state = fstate;
 662         }
 663
 664         return state;
 665 }
 666
 667 /* Currently those 2 are only used when MSIs are enabled, this will change
 668  * but in the meantime, we need to protect them to avoid warnings
 669  */
 670 #ifdef CONFIG_PCI_MSI
 671 struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev)
 672 {
 673         struct pci_controller *hose = pci_bus_to_host(dev->bus);
 674         struct pnv_phb *phb = hose->private_data;
 675         struct pci_dn *pdn = pci_get_pdn(dev);
 676
 677         if (!pdn)
 678                 return NULL;
 679         if (pdn->pe_number == IODA_INVALID_PE)
 680                 return NULL;
 681         return &phb->ioda.pe_array[pdn->pe_number];
 682 }
 683 #endif /* CONFIG_PCI_MSI */
 684
 685 static int pnv_ioda_set_one_peltv(struct pnv_phb *phb,
 686                                   struct pnv_ioda_pe *parent,
 687                                   struct pnv_ioda_pe *child,
 688                                   bool is_add)
 689 {
 690         const char *desc = is_add ? "adding" : "removing";
 691         uint8_t op = is_add ? OPAL_ADD_PE_TO_DOMAIN :
 692                               OPAL_REMOVE_PE_FROM_DOMAIN;
 693         struct pnv_ioda_pe *slave;
 694         long rc;
 695
 696         /* Parent PE affects child PE */
 697         rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
 698                                 child->pe_number, op);
 699         if (rc != OPAL_SUCCESS) {
 700                 pe_warn(child, "OPAL error %ld %s to parent PELTV\n",
 701                         rc, desc);
 702                 return -ENXIO;
 703         }
 704
 705         if (!(child->flags & PNV_IODA_PE_MASTER))
 706                 return 0;
 707
 708         /* Compound case: parent PE affects slave PEs */
 709         list_for_each_entry(slave, &child->slaves, list) {
 710                 rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
 711                                         slave->pe_number, op);
 712                 if (rc != OPAL_SUCCESS) {
 713                         pe_warn(slave, "OPAL error %ld %s to parent PELTV\n",
 714                                 rc, desc);
 715                         return -ENXIO;
 716                 }
 717         }
 718
 719         return 0;
 720 }
 721
 722 static int pnv_ioda_set_peltv(struct pnv_phb *phb,
 723                               struct pnv_ioda_pe *pe,
 724                               bool is_add)
 725 {
 726         struct pnv_ioda_pe *slave;
 727         struct pci_dev *pdev = NULL;
 728         int ret;
 729
 730         /*
 731          * Clear PE frozen state. If it's master PE, we need
 732          * clear slave PE frozen state as well.
 733          */
 734         if (is_add) {
 735                 opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
 736                                           OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 737                 if (pe->flags & PNV_IODA_PE_MASTER) {
 738                         list_for_each_entry(slave, &pe->slaves, list)
 739                                 opal_pci_eeh_freeze_clear(phb->opal_id,
 740                                                           slave->pe_number,
 741                                                           OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 742                 }
 743         }
 744
 745         /*
 746          * Associate PE in PELT. We need add the PE into the
 747          * corresponding PELT-V as well. Otherwise, the error
 748          * originated from the PE might contribute to other
 749          * PEs.
 750          */
 751         ret = pnv_ioda_set_one_peltv(phb, pe, pe, is_add);
 752         if (ret)
 753                 return ret;
 754
 755         /* For compound PEs, any one affects all of them */
 756         if (pe->flags & PNV_IODA_PE_MASTER) {
 757                 list_for_each_entry(slave, &pe->slaves, list) {
 758                         ret = pnv_ioda_set_one_peltv(phb, slave, pe, is_add);
 759                         if (ret)
 760                                 return ret;
 761                 }
 762         }
 763
 764         if (pe->flags & (PNV_IODA_PE_BUS_ALL | PNV_IODA_PE_BUS))
 765                 pdev = pe->pbus->self;
 766         else if (pe->flags & PNV_IODA_PE_DEV)
 767                 pdev = pe->pdev->bus->self;
 768 #ifdef CONFIG_PCI_IOV
 769         else if (pe->flags & PNV_IODA_PE_VF)
 770                 pdev = pe->parent_dev;
 771 #endif /* CONFIG_PCI_IOV */
 772         while (pdev) {
 773                 struct pci_dn *pdn = pci_get_pdn(pdev);
 774                 struct pnv_ioda_pe *parent;
 775
 776                 if (pdn && pdn->pe_number != IODA_INVALID_PE) {
 777                         parent = &phb->ioda.pe_array[pdn->pe_number];
 778                         ret = pnv_ioda_set_one_peltv(phb, parent, pe, is_add);
 779                         if (ret)
 780                                 return ret;
 781                 }
 782
 783                 pdev = pdev->bus->self;
 784         }
 785
 786         return 0;
 787 }
 788
 789 static int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
 790 {
 791         struct pci_dev *parent;
 792         uint8_t bcomp, dcomp, fcomp;
 793         int64_t rc;
 794         long rid_end, rid;
 795
 796         /* Currently, we just deconfigure VF PE. Bus PE will always there.*/
 797         if (pe->pbus) {
 798                 int count;
 799
 800                 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
 801                 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
 802                 parent = pe->pbus->self;
 803                 if (pe->flags & PNV_IODA_PE_BUS_ALL)
 804                         count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
 805                 else
 806                         count = 1;
 807
 808                 switch(count) {
 809                 case  1: bcomp = OpalPciBusAll;         break;
 810                 case  2: bcomp = OpalPciBus7Bits;       break;
 811                 case  4: bcomp = OpalPciBus6Bits;       break;
 812                 case  8: bcomp = OpalPciBus5Bits;       break;
 813                 case 16: bcomp = OpalPciBus4Bits;       break;
 814                 case 32: bcomp = OpalPciBus3Bits;       break;
 815                 default:
 816                         dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
 817                                 count);
 818                         /* Do an exact match only */
 819                         bcomp = OpalPciBusAll;
 820                 }
 821                 rid_end = pe->rid + (count << 8);
 822         } else {
 823 #ifdef CONFIG_PCI_IOV
 824                 if (pe->flags & PNV_IODA_PE_VF)
 825                         parent = pe->parent_dev;
 826                 else
 827 #endif
 828                         parent = pe->pdev->bus->self;
 829                 bcomp = OpalPciBusAll;
 830                 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
 831                 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
 832                 rid_end = pe->rid + 1;
 833         }
 834
 835         /* Clear the reverse map */
 836         for (rid = pe->rid; rid < rid_end; rid++)
 837                 phb->ioda.pe_rmap[rid] = IODA_INVALID_PE;
 838
 839         /* Release from all parents PELT-V */
 840         while (parent) {
 841                 struct pci_dn *pdn = pci_get_pdn(parent);
 842                 if (pdn && pdn->pe_number != IODA_INVALID_PE) {
 843                         rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number,
 844                                                 pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
 845                         /* XXX What to do in case of error ? */
 846                 }
 847                 parent = parent->bus->self;
 848         }
 849
 850         opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
 851                                   OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
 852
 853         /* Disassociate PE in PELT */
 854         rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number,
 855                                 pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
 856         if (rc)
 857                 pe_warn(pe, "OPAL error %ld remove self from PELTV\n", rc);
 858         rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
 859                              bcomp, dcomp, fcomp, OPAL_UNMAP_PE);
 860         if (rc)
 861                 pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
 862
 863         pe->pbus = NULL;
 864         pe->pdev = NULL;
 865 #ifdef CONFIG_PCI_IOV
 866         pe->parent_dev = NULL;
 867 #endif
 868
 869         return 0;
 870 }
 871
 872 static int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
 873 {
 874         struct pci_dev *parent;
 875         uint8_t bcomp, dcomp, fcomp;
 876         long rc, rid_end, rid;
 877
 878         /* Bus validation ? */
 879         if (pe->pbus) {
 880                 int count;
 881
 882                 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
 883                 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
 884                 parent = pe->pbus->self;
 885                 if (pe->flags & PNV_IODA_PE_BUS_ALL)
 886                         count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
 887                 else
 888                         count = 1;
 889
 890                 switch(count) {
 891                 case  1: bcomp = OpalPciBusAll;         break;
 892                 case  2: bcomp = OpalPciBus7Bits;       break;
 893                 case  4: bcomp = OpalPciBus6Bits;       break;
 894                 case  8: bcomp = OpalPciBus5Bits;       break;
 895                 case 16: bcomp = OpalPciBus4Bits;       break;
 896                 case 32: bcomp = OpalPciBus3Bits;       break;
 897                 default:
 898                         dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
 899                                 count);
 900                         /* Do an exact match only */
 901                         bcomp = OpalPciBusAll;
 902                 }
 903                 rid_end = pe->rid + (count << 8);
 904         } else {
 905 #ifdef CONFIG_PCI_IOV
 906                 if (pe->flags & PNV_IODA_PE_VF)
 907                         parent = pe->parent_dev;
 908                 else
 909 #endif /* CONFIG_PCI_IOV */
 910                         parent = pe->pdev->bus->self;
 911                 bcomp = OpalPciBusAll;
 912                 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
 913                 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
 914                 rid_end = pe->rid + 1;
 915         }
 916
 917         /*
 918          * Associate PE in PELT. We need add the PE into the
 919          * corresponding PELT-V as well. Otherwise, the error
 920          * originated from the PE might contribute to other
 921          * PEs.
 922          */
 923         rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
 924                              bcomp, dcomp, fcomp, OPAL_MAP_PE);
 925         if (rc) {
 926                 pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
 927                 return -ENXIO;
 928         }
 929
 930         /*
 931          * Configure PELTV. NPUs don't have a PELTV table so skip
 932          * configuration on them.
 933          */
 934         if (phb->type != PNV_PHB_NPU_NVLINK && phb->type != PNV_PHB_NPU_OCAPI)
 935                 pnv_ioda_set_peltv(phb, pe, true);
 936
 937         /* Setup reverse map */
 938         for (rid = pe->rid; rid < rid_end; rid++)
 939                 phb->ioda.pe_rmap[rid] = pe->pe_number;
 940
 941         /* Setup one MVTs on IODA1 */
 942         if (phb->type != PNV_PHB_IODA1) {
 943                 pe->mve_number = 0;
 944                 goto out;
 945         }
 946
 947         pe->mve_number = pe->pe_number;
 948         rc = opal_pci_set_mve(phb->opal_id, pe->mve_number, pe->pe_number);
 949         if (rc != OPAL_SUCCESS) {
 950                 pe_err(pe, "OPAL error %ld setting up MVE %x\n",
 951                        rc, pe->mve_number);
 952                 pe->mve_number = -1;
 953         } else {
 954                 rc = opal_pci_set_mve_enable(phb->opal_id,
 955                                              pe->mve_number, OPAL_ENABLE_MVE);
 956                 if (rc) {
 957                         pe_err(pe, "OPAL error %ld enabling MVE %x\n",
 958                                rc, pe->mve_number);
 959                         pe->mve_number = -1;
 960                 }
 961         }
 962
 963 out:
 964         return 0;
 965 }
 966
 967 #ifdef CONFIG_PCI_IOV
 968 static int pnv_pci_vf_resource_shift(struct pci_dev *dev, int offset)
 969 {
 970         struct pci_dn *pdn = pci_get_pdn(dev);
 971         int i;
 972         struct resource *res, res2;
 973         resource_size_t size;
 974         u16 num_vfs;
 975
 976         if (!dev->is_physfn)
 977                 return -EINVAL;
 978
 979         /*
 980          * "offset" is in VFs.  The M64 windows are sized so that when they
 981          * are segmented, each segment is the same size as the IOV BAR.
 982          * Each segment is in a separate PE, and the high order bits of the
 983          * address are the PE number.  Therefore, each VF's BAR is in a
 984          * separate PE, and changing the IOV BAR start address changes the
 985          * range of PEs the VFs are in.
 986          */
 987         num_vfs = pdn->num_vfs;
 988         for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
 989                 res = &dev->resource[i + PCI_IOV_RESOURCES];
 990                 if (!res->flags || !res->parent)
 991                         continue;
 992
 993                 /*
 994                  * The actual IOV BAR range is determined by the start address
 995                  * and the actual size for num_vfs VFs BAR.  This check is to
 996                  * make sure that after shifting, the range will not overlap
 997                  * with another device.
 998                  */
 999                 size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
1000                 res2.flags = res->flags;
1001                 res2.start = res->start + (size * offset);
1002                 res2.end = res2.start + (size * num_vfs) - 1;
1003
1004                 if (res2.end > res->end) {
1005                         dev_err(&dev->dev, "VF BAR%d: %pR would extend past %pR (trying to enable %d VFs shifted by %d)\n",
1006                                 i, &res2, res, num_vfs, offset);
1007                         return -EBUSY;
1008                 }
1009         }
1010
1011         /*
1012          * Since M64 BAR shares segments among all possible 256 PEs,
1013          * we have to shift the beginning of PF IOV BAR to make it start from
1014          * the segment which belongs to the PE number assigned to the first VF.
1015          * This creates a "hole" in the /proc/iomem which could be used for
1016          * allocating other resources so we reserve this area below and
1017          * release when IOV is released.
1018          */
1019         for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
1020                 res = &dev->resource[i + PCI_IOV_RESOURCES];
1021                 if (!res->flags || !res->parent)
1022                         continue;
1023
1024                 size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
1025                 res2 = *res;
1026                 res->start += size * offset;
1027
1028                 dev_info(&dev->dev, "VF BAR%d: %pR shifted to %pR (%sabling %d VFs shifted by %d)\n",
1029                          i, &res2, res, (offset > 0) ? "En" : "Dis",
1030                          num_vfs, offset);
1031
1032                 if (offset < 0) {
1033                         devm_release_resource(&dev->dev, &pdn->holes[i]);
1034                         memset(&pdn->holes[i], 0, sizeof(pdn->holes[i]));
1035                 }
1036
1037                 pci_update_resource(dev, i + PCI_IOV_RESOURCES);
1038
1039                 if (offset > 0) {
1040                         pdn->holes[i].start = res2.start;
1041                         pdn->holes[i].end = res2.start + size * offset - 1;
1042                         pdn->holes[i].flags = IORESOURCE_BUS;
1043                         pdn->holes[i].name = "pnv_iov_reserved";
1044                         devm_request_resource(&dev->dev, res->parent,
1045                                         &pdn->holes[i]);
1046                 }
1047         }
1048         return 0;
1049 }
1050 #endif /* CONFIG_PCI_IOV */
1051
1052 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev)
1053 {
1054         struct pci_controller *hose = pci_bus_to_host(dev->bus);
1055         struct pnv_phb *phb = hose->private_data;
1056         struct pci_dn *pdn = pci_get_pdn(dev);
1057         struct pnv_ioda_pe *pe;
1058
1059         if (!pdn) {
1060                 pr_err("%s: Device tree node not associated properly\n",
1061                            pci_name(dev));
1062                 return NULL;
1063         }
1064         if (pdn->pe_number != IODA_INVALID_PE)
1065                 return NULL;
1066
1067         pe = pnv_ioda_alloc_pe(phb);
1068         if (!pe) {
1069                 pr_warn("%s: Not enough PE# available, disabling device\n",
1070                         pci_name(dev));
1071                 return NULL;
1072         }
1073
1074         /* NOTE: We get only one ref to the pci_dev for the pdn, not for the
1075          * pointer in the PE data structure, both should be destroyed at the
1076          * same time. However, this needs to be looked at more closely again
1077          * once we actually start removing things (Hotplug, SR-IOV, ...)
1078          *
1079          * At some point we want to remove the PDN completely anyways
1080          */
1081         pci_dev_get(dev);
1082         pdn->pe_number = pe->pe_number;
1083         pe->flags = PNV_IODA_PE_DEV;
1084         pe->pdev = dev;
1085         pe->pbus = NULL;
1086         pe->mve_number = -1;
1087         pe->rid = dev->bus->number << 8 | pdn->devfn;
1088
1089         pe_info(pe, "Associated device to PE\n");
1090
1091         if (pnv_ioda_configure_pe(phb, pe)) {
1092                 /* XXX What do we do here ? */
1093                 pnv_ioda_free_pe(pe);
1094                 pdn->pe_number = IODA_INVALID_PE;
1095                 pe->pdev = NULL;
1096                 pci_dev_put(dev);
1097                 return NULL;
1098         }
1099
1100         /* Put PE to the list */
1101         list_add_tail(&pe->list, &phb->ioda.pe_list);
1102
1103         return pe;
1104 }
1105
1106 static void pnv_ioda_setup_same_PE(struct pci_bus *bus, struct pnv_ioda_pe *pe)
1107 {
1108         struct pci_dev *dev;
1109
1110         list_for_each_entry(dev, &bus->devices, bus_list) {
1111                 struct pci_dn *pdn = pci_get_pdn(dev);
1112
1113                 if (pdn == NULL) {
1114                         pr_warn("%s: No device node associated with device !\n",
1115                                 pci_name(dev));
1116                         continue;
1117                 }
1118
1119                 /*
1120                  * In partial hotplug case, the PCI device might be still
1121                  * associated with the PE and needn't attach it to the PE
1122                  * again.
1123                  */
1124                 if (pdn->pe_number != IODA_INVALID_PE)
1125                         continue;
1126
1127                 pe->device_count++;
1128                 pdn->pe_number = pe->pe_number;
1129                 if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1130                         pnv_ioda_setup_same_PE(dev->subordinate, pe);
1131         }
1132 }
1133
1134 /*
1135  * There're 2 types of PCI bus sensitive PEs: One that is compromised of
1136  * single PCI bus. Another one that contains the primary PCI bus and its
1137  * subordinate PCI devices and buses. The second type of PE is normally
1138  * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
1139  */
1140 static struct pnv_ioda_pe *pnv_ioda_setup_bus_PE(struct pci_bus *bus, bool all)
1141 {
1142         struct pci_controller *hose = pci_bus_to_host(bus);
1143         struct pnv_phb *phb = hose->private_data;
1144         struct pnv_ioda_pe *pe = NULL;
1145         unsigned int pe_num;
1146
1147         /*
1148          * In partial hotplug case, the PE instance might be still alive.
1149          * We should reuse it instead of allocating a new one.
1150          */
1151         pe_num = phb->ioda.pe_rmap[bus->number << 8];
1152         if (pe_num != IODA_INVALID_PE) {
1153                 pe = &phb->ioda.pe_array[pe_num];
1154                 pnv_ioda_setup_same_PE(bus, pe);
1155                 return NULL;
1156         }
1157
1158         /* PE number for root bus should have been reserved */
1159         if (pci_is_root_bus(bus) &&
1160             phb->ioda.root_pe_idx != IODA_INVALID_PE)
1161                 pe = &phb->ioda.pe_array[phb->ioda.root_pe_idx];
1162
1163         /* Check if PE is determined by M64 */
1164         if (!pe && phb->pick_m64_pe)
1165                 pe = phb->pick_m64_pe(bus, all);
1166
1167         /* The PE number isn't pinned by M64 */
1168         if (!pe)
1169                 pe = pnv_ioda_alloc_pe(phb);
1170
1171         if (!pe) {
1172                 pr_warn("%s: Not enough PE# available for PCI bus %04x:%02x\n",
1173                         __func__, pci_domain_nr(bus), bus->number);
1174                 return NULL;
1175         }
1176
1177         pe->flags |= (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS);
1178         pe->pbus = bus;
1179         pe->pdev = NULL;
1180         pe->mve_number = -1;
1181         pe->rid = bus->busn_res.start << 8;
1182
1183         if (all)
1184                 pe_info(pe, "Secondary bus %d..%d associated with PE#%x\n",
1185                         bus->busn_res.start, bus->busn_res.end, pe->pe_number);
1186         else
1187                 pe_info(pe, "Secondary bus %d associated with PE#%x\n",
1188                         bus->busn_res.start, pe->pe_number);
1189
1190         if (pnv_ioda_configure_pe(phb, pe)) {
1191                 /* XXX What do we do here ? */
1192                 pnv_ioda_free_pe(pe);
1193                 pe->pbus = NULL;
1194                 return NULL;
1195         }
1196
1197         /* Associate it with all child devices */
1198         pnv_ioda_setup_same_PE(bus, pe);
1199
1200         /* Put PE to the list */
1201         list_add_tail(&pe->list, &phb->ioda.pe_list);
1202
1203         return pe;
1204 }
1205
1206 static struct pnv_ioda_pe *pnv_ioda_setup_npu_PE(struct pci_dev *npu_pdev)
1207 {
1208         int pe_num, found_pe = false, rc;
1209         long rid;
1210         struct pnv_ioda_pe *pe;
1211         struct pci_dev *gpu_pdev;
1212         struct pci_dn *npu_pdn;
1213         struct pci_controller *hose = pci_bus_to_host(npu_pdev->bus);
1214         struct pnv_phb *phb = hose->private_data;
1215
1216         /*
1217          * Due to a hardware errata PE#0 on the NPU is reserved for
1218          * error handling. This means we only have three PEs remaining
1219          * which need to be assigned to four links, implying some
1220          * links must share PEs.
1221          *
1222          * To achieve this we assign PEs such that NPUs linking the
1223          * same GPU get assigned the same PE.
1224          */
1225         gpu_pdev = pnv_pci_get_gpu_dev(npu_pdev);
1226         for (pe_num = 0; pe_num < phb->ioda.total_pe_num; pe_num++) {
1227                 pe = &phb->ioda.pe_array[pe_num];
1228                 if (!pe->pdev)
1229                         continue;
1230
1231                 if (pnv_pci_get_gpu_dev(pe->pdev) == gpu_pdev) {
1232                         /*
1233                          * This device has the same peer GPU so should
1234                          * be assigned the same PE as the existing
1235                          * peer NPU.
1236                          */
1237                         dev_info(&npu_pdev->dev,
1238                                 "Associating to existing PE %x\n", pe_num);
1239                         pci_dev_get(npu_pdev);
1240                         npu_pdn = pci_get_pdn(npu_pdev);
1241                         rid = npu_pdev->bus->number << 8 | npu_pdn->devfn;
1242                         npu_pdn->pe_number = pe_num;
1243                         phb->ioda.pe_rmap[rid] = pe->pe_number;
1244
1245                         /* Map the PE to this link */
1246                         rc = opal_pci_set_pe(phb->opal_id, pe_num, rid,
1247                                         OpalPciBusAll,
1248                                         OPAL_COMPARE_RID_DEVICE_NUMBER,
1249                                         OPAL_COMPARE_RID_FUNCTION_NUMBER,
1250                                         OPAL_MAP_PE);
1251                         WARN_ON(rc != OPAL_SUCCESS);
1252                         found_pe = true;
1253                         break;
1254                 }
1255         }
1256
1257         if (!found_pe)
1258                 /*
1259                  * Could not find an existing PE so allocate a new
1260                  * one.
1261                  */
1262                 return pnv_ioda_setup_dev_PE(npu_pdev);
1263         else
1264                 return pe;
1265 }
1266
1267 static void pnv_ioda_setup_npu_PEs(struct pci_bus *bus)
1268 {
1269         struct pci_dev *pdev;
1270
1271         list_for_each_entry(pdev, &bus->devices, bus_list)
1272                 pnv_ioda_setup_npu_PE(pdev);
1273 }
1274
1275 static void pnv_pci_ioda_setup_PEs(void)
1276 {
1277         struct pci_controller *hose, *tmp;
1278         struct pnv_phb *phb;
1279         struct pci_bus *bus;
1280         struct pci_dev *pdev;
1281
1282         list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1283                 phb = hose->private_data;
1284                 if (phb->type == PNV_PHB_NPU_NVLINK) {
1285                         /* PE#0 is needed for error reporting */
1286                         pnv_ioda_reserve_pe(phb, 0);
1287                         pnv_ioda_setup_npu_PEs(hose->bus);
1288                         if (phb->model == PNV_PHB_MODEL_NPU2)
1289                                 pnv_npu2_init(phb);
1290                 }
1291                 if (phb->type == PNV_PHB_NPU_OCAPI) {
1292                         bus = hose->bus;
1293                         list_for_each_entry(pdev, &bus->devices, bus_list)
1294                                 pnv_ioda_setup_dev_PE(pdev);
1295                 }
1296         }
1297 }
1298
1299 #ifdef CONFIG_PCI_IOV
1300 static int pnv_pci_vf_release_m64(struct pci_dev *pdev, u16 num_vfs)
1301 {
1302         struct pci_bus        *bus;
1303         struct pci_controller *hose;
1304         struct pnv_phb        *phb;
1305         struct pci_dn         *pdn;
1306         int                    i, j;
1307         int                    m64_bars;
1308
1309         bus = pdev->bus;
1310         hose = pci_bus_to_host(bus);
1311         phb = hose->private_data;
1312         pdn = pci_get_pdn(pdev);
1313
1314         if (pdn->m64_single_mode)
1315                 m64_bars = num_vfs;
1316         else
1317                 m64_bars = 1;
1318
1319         for (i = 0; i < PCI_SRIOV_NUM_BARS; i++)
1320                 for (j = 0; j < m64_bars; j++) {
1321                         if (pdn->m64_map[j][i] == IODA_INVALID_M64)
1322                                 continue;
1323                         opal_pci_phb_mmio_enable(phb->opal_id,
1324                                 OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 0);
1325                         clear_bit(pdn->m64_map[j][i], &phb->ioda.m64_bar_alloc);
1326                         pdn->m64_map[j][i] = IODA_INVALID_M64;
1327                 }
1328
1329         kfree(pdn->m64_map);
1330         return 0;
1331 }
1332
1333 static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)
1334 {
1335         struct pci_bus        *bus;
1336         struct pci_controller *hose;
1337         struct pnv_phb        *phb;
1338         struct pci_dn         *pdn;
1339         unsigned int           win;
1340         struct resource       *res;
1341         int                    i, j;
1342         int64_t                rc;
1343         int                    total_vfs;
1344         resource_size_t        size, start;
1345         int                    pe_num;
1346         int                    m64_bars;
1347
1348         bus = pdev->bus;
1349         hose = pci_bus_to_host(bus);
1350         phb = hose->private_data;
1351         pdn = pci_get_pdn(pdev);
1352         total_vfs = pci_sriov_get_totalvfs(pdev);
1353
1354         if (pdn->m64_single_mode)
1355                 m64_bars = num_vfs;
1356         else
1357                 m64_bars = 1;
1358
1359         pdn->m64_map = kmalloc_array(m64_bars,
1360                                      sizeof(*pdn->m64_map),
1361                                      GFP_KERNEL);
1362         if (!pdn->m64_map)
1363                 return -ENOMEM;
1364         /* Initialize the m64_map to IODA_INVALID_M64 */
1365         for (i = 0; i < m64_bars ; i++)
1366                 for (j = 0; j < PCI_SRIOV_NUM_BARS; j++)
1367                         pdn->m64_map[i][j] = IODA_INVALID_M64;
1368
1369
1370         for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
1371                 res = &pdev->resource[i + PCI_IOV_RESOURCES];
1372                 if (!res->flags || !res->parent)
1373                         continue;
1374
1375                 for (j = 0; j < m64_bars; j++) {
1376                         do {
1377                                 win = find_next_zero_bit(&phb->ioda.m64_bar_alloc,
1378                                                 phb->ioda.m64_bar_idx + 1, 0);
1379
1380                                 if (win >= phb->ioda.m64_bar_idx + 1)
1381                                         goto m64_failed;
1382                         } while (test_and_set_bit(win, &phb->ioda.m64_bar_alloc));
1383
1384                         pdn->m64_map[j][i] = win;
1385
1386                         if (pdn->m64_single_mode) {
1387                                 size = pci_iov_resource_size(pdev,
1388                                                         PCI_IOV_RESOURCES + i);
1389                                 start = res->start + size * j;
1390                         } else {
1391                                 size = resource_size(res);
1392                                 start = res->start;
1393                         }
1394
1395                         /* Map the M64 here */
1396                         if (pdn->m64_single_mode) {
1397                                 pe_num = pdn->pe_num_map[j];
1398                                 rc = opal_pci_map_pe_mmio_window(phb->opal_id,
1399                                                 pe_num, OPAL_M64_WINDOW_TYPE,
1400                                                 pdn->m64_map[j][i], 0);
1401                         }
1402
1403                         rc = opal_pci_set_phb_mem_window(phb->opal_id,
1404                                                  OPAL_M64_WINDOW_TYPE,
1405                                                  pdn->m64_map[j][i],
1406                                                  start,
1407                                                  0, /* unused */
1408                                                  size);
1409
1410
1411                         if (rc != OPAL_SUCCESS) {
1412                                 dev_err(&pdev->dev, "Failed to map M64 window #%d: %lld\n",
1413                                         win, rc);
1414                                 goto m64_failed;
1415                         }
1416
1417                         if (pdn->m64_single_mode)
1418                                 rc = opal_pci_phb_mmio_enable(phb->opal_id,
1419                                      OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 2);
1420                         else
1421                                 rc = opal_pci_phb_mmio_enable(phb->opal_id,
1422                                      OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 1);
1423
1424                         if (rc != OPAL_SUCCESS) {
1425                                 dev_err(&pdev->dev, "Failed to enable M64 window #%d: %llx\n",
1426                                         win, rc);
1427                                 goto m64_failed;
1428                         }
1429                 }
1430         }
1431         return 0;
1432
1433 m64_failed:
1434         pnv_pci_vf_release_m64(pdev, num_vfs);
1435         return -EBUSY;
1436 }
1437
1438 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
1439                 int num);
1440
1441 static void pnv_pci_ioda2_release_dma_pe(struct pci_dev *dev, struct pnv_ioda_pe *pe)
1442 {
1443         struct iommu_table    *tbl;
1444         int64_t               rc;
1445
1446         tbl = pe->table_group.tables[0];
1447         rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
1448         if (rc)
1449                 pe_warn(pe, "OPAL error %ld release DMA window\n", rc);
1450
1451         pnv_pci_ioda2_set_bypass(pe, false);
1452         if (pe->table_group.group) {
1453                 iommu_group_put(pe->table_group.group);
1454                 BUG_ON(pe->table_group.group);
1455         }
1456         iommu_tce_table_put(tbl);
1457 }
1458
1459 static void pnv_ioda_release_vf_PE(struct pci_dev *pdev)
1460 {
1461         struct pci_bus        *bus;
1462         struct pci_controller *hose;
1463         struct pnv_phb        *phb;
1464         struct pnv_ioda_pe    *pe, *pe_n;
1465         struct pci_dn         *pdn;
1466
1467         bus = pdev->bus;
1468         hose = pci_bus_to_host(bus);
1469         phb = hose->private_data;
1470         pdn = pci_get_pdn(pdev);
1471
1472         if (!pdev->is_physfn)
1473                 return;
1474
1475         list_for_each_entry_safe(pe, pe_n, &phb->ioda.pe_list, list) {
1476                 if (pe->parent_dev != pdev)
1477                         continue;
1478
1479                 pnv_pci_ioda2_release_dma_pe(pdev, pe);
1480
1481                 /* Remove from list */
1482                 mutex_lock(&phb->ioda.pe_list_mutex);
1483                 list_del(&pe->list);
1484                 mutex_unlock(&phb->ioda.pe_list_mutex);
1485
1486                 pnv_ioda_deconfigure_pe(phb, pe);
1487
1488                 pnv_ioda_free_pe(pe);
1489         }
1490 }
1491
1492 void pnv_pci_sriov_disable(struct pci_dev *pdev)
1493 {
1494         struct pci_bus        *bus;
1495         struct pci_controller *hose;
1496         struct pnv_phb        *phb;
1497         struct pnv_ioda_pe    *pe;
1498         struct pci_dn         *pdn;
1499         u16                    num_vfs, i;
1500
1501         bus = pdev->bus;
1502         hose = pci_bus_to_host(bus);
1503         phb = hose->private_data;
1504         pdn = pci_get_pdn(pdev);
1505         num_vfs = pdn->num_vfs;
1506
1507         /* Release VF PEs */
1508         pnv_ioda_release_vf_PE(pdev);
1509
1510         if (phb->type == PNV_PHB_IODA2) {
1511                 if (!pdn->m64_single_mode)
1512                         pnv_pci_vf_resource_shift(pdev, -*pdn->pe_num_map);
1513
1514                 /* Release M64 windows */
1515                 pnv_pci_vf_release_m64(pdev, num_vfs);
1516
1517                 /* Release PE numbers */
1518                 if (pdn->m64_single_mode) {
1519                         for (i = 0; i < num_vfs; i++) {
1520                                 if (pdn->pe_num_map[i] == IODA_INVALID_PE)
1521                                         continue;
1522
1523                                 pe = &phb->ioda.pe_array[pdn->pe_num_map[i]];
1524                                 pnv_ioda_free_pe(pe);
1525                         }
1526                 } else
1527                         bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1528                 /* Releasing pe_num_map */
1529                 kfree(pdn->pe_num_map);
1530         }
1531 }
1532
1533 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
1534                                        struct pnv_ioda_pe *pe);
1535 static void pnv_ioda_setup_vf_PE(struct pci_dev *pdev, u16 num_vfs)
1536 {
1537         struct pci_bus        *bus;
1538         struct pci_controller *hose;
1539         struct pnv_phb        *phb;
1540         struct pnv_ioda_pe    *pe;
1541         int                    pe_num;
1542         u16                    vf_index;
1543         struct pci_dn         *pdn;
1544
1545         bus = pdev->bus;
1546         hose = pci_bus_to_host(bus);
1547         phb = hose->private_data;
1548         pdn = pci_get_pdn(pdev);
1549
1550         if (!pdev->is_physfn)
1551                 return;
1552
1553         /* Reserve PE for each VF */
1554         for (vf_index = 0; vf_index < num_vfs; vf_index++) {
1555                 if (pdn->m64_single_mode)
1556                         pe_num = pdn->pe_num_map[vf_index];
1557                 else
1558                         pe_num = *pdn->pe_num_map + vf_index;
1559
1560                 pe = &phb->ioda.pe_array[pe_num];
1561                 pe->pe_number = pe_num;
1562                 pe->phb = phb;
1563                 pe->flags = PNV_IODA_PE_VF;
1564                 pe->pbus = NULL;
1565                 pe->parent_dev = pdev;
1566                 pe->mve_number = -1;
1567                 pe->rid = (pci_iov_virtfn_bus(pdev, vf_index) << 8) |
1568                            pci_iov_virtfn_devfn(pdev, vf_index);
1569
1570                 pe_info(pe, "VF %04d:%02d:%02d.%d associated with PE#%x\n",
1571                         hose->global_number, pdev->bus->number,
1572                         PCI_SLOT(pci_iov_virtfn_devfn(pdev, vf_index)),
1573                         PCI_FUNC(pci_iov_virtfn_devfn(pdev, vf_index)), pe_num);
1574
1575                 if (pnv_ioda_configure_pe(phb, pe)) {
1576                         /* XXX What do we do here ? */
1577                         pnv_ioda_free_pe(pe);
1578                         pe->pdev = NULL;
1579                         continue;
1580                 }
1581
1582                 /* Put PE to the list */
1583                 mutex_lock(&phb->ioda.pe_list_mutex);
1584                 list_add_tail(&pe->list, &phb->ioda.pe_list);
1585                 mutex_unlock(&phb->ioda.pe_list_mutex);
1586
1587                 pnv_pci_ioda2_setup_dma_pe(phb, pe);
1588         }
1589 }
1590
1591 int pnv_pci_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1592 {
1593         struct pci_bus        *bus;
1594         struct pci_controller *hose;
1595         struct pnv_phb        *phb;
1596         struct pnv_ioda_pe    *pe;
1597         struct pci_dn         *pdn;
1598         int                    ret;
1599         u16                    i;
1600
1601         bus = pdev->bus;
1602         hose = pci_bus_to_host(bus);
1603         phb = hose->private_data;
1604         pdn = pci_get_pdn(pdev);
1605
1606         if (phb->type == PNV_PHB_IODA2) {
1607                 if (!pdn->vfs_expanded) {
1608                         dev_info(&pdev->dev, "don't support this SRIOV device"
1609                                 " with non 64bit-prefetchable IOV BAR\n");
1610                         return -ENOSPC;
1611                 }
1612
1613                 /*
1614                  * When M64 BARs functions in Single PE mode, the number of VFs
1615                  * could be enabled must be less than the number of M64 BARs.
1616                  */
1617                 if (pdn->m64_single_mode && num_vfs > phb->ioda.m64_bar_idx) {
1618                         dev_info(&pdev->dev, "Not enough M64 BAR for VFs\n");
1619                         return -EBUSY;
1620                 }
1621
1622                 /* Allocating pe_num_map */
1623                 if (pdn->m64_single_mode)
1624                         pdn->pe_num_map = kmalloc_array(num_vfs,
1625                                                         sizeof(*pdn->pe_num_map),
1626                                                         GFP_KERNEL);
1627                 else
1628                         pdn->pe_num_map = kmalloc(sizeof(*pdn->pe_num_map), GFP_KERNEL);
1629
1630                 if (!pdn->pe_num_map)
1631                         return -ENOMEM;
1632
1633                 if (pdn->m64_single_mode)
1634                         for (i = 0; i < num_vfs; i++)
1635                                 pdn->pe_num_map[i] = IODA_INVALID_PE;
1636
1637                 /* Calculate available PE for required VFs */
1638                 if (pdn->m64_single_mode) {
1639                         for (i = 0; i < num_vfs; i++) {
1640                                 pe = pnv_ioda_alloc_pe(phb);
1641                                 if (!pe) {
1642                                         ret = -EBUSY;
1643                                         goto m64_failed;
1644                                 }
1645
1646                                 pdn->pe_num_map[i] = pe->pe_number;
1647                         }
1648                 } else {
1649                         mutex_lock(&phb->ioda.pe_alloc_mutex);
1650                         *pdn->pe_num_map = bitmap_find_next_zero_area(
1651                                 phb->ioda.pe_alloc, phb->ioda.total_pe_num,
1652                                 0, num_vfs, 0);
1653                         if (*pdn->pe_num_map >= phb->ioda.total_pe_num) {
1654                                 mutex_unlock(&phb->ioda.pe_alloc_mutex);
1655                                 dev_info(&pdev->dev, "Failed to enable VF%d\n", num_vfs);
1656                                 kfree(pdn->pe_num_map);
1657                                 return -EBUSY;
1658                         }
1659                         bitmap_set(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1660                         mutex_unlock(&phb->ioda.pe_alloc_mutex);
1661                 }
1662                 pdn->num_vfs = num_vfs;
1663
1664                 /* Assign M64 window accordingly */
1665                 ret = pnv_pci_vf_assign_m64(pdev, num_vfs);
1666                 if (ret) {
1667                         dev_info(&pdev->dev, "Not enough M64 window resources\n");
1668                         goto m64_failed;
1669                 }
1670
1671                 /*
1672                  * When using one M64 BAR to map one IOV BAR, we need to shift
1673                  * the IOV BAR according to the PE# allocated to the VFs.
1674                  * Otherwise, the PE# for the VF will conflict with others.
1675                  */
1676                 if (!pdn->m64_single_mode) {
1677                         ret = pnv_pci_vf_resource_shift(pdev, *pdn->pe_num_map);
1678                         if (ret)
1679                                 goto m64_failed;
1680                 }
1681         }
1682
1683         /* Setup VF PEs */
1684         pnv_ioda_setup_vf_PE(pdev, num_vfs);
1685
1686         return 0;
1687
1688 m64_failed:
1689         if (pdn->m64_single_mode) {
1690                 for (i = 0; i < num_vfs; i++) {
1691                         if (pdn->pe_num_map[i] == IODA_INVALID_PE)
1692                                 continue;
1693
1694                         pe = &phb->ioda.pe_array[pdn->pe_num_map[i]];
1695                         pnv_ioda_free_pe(pe);
1696                 }
1697         } else
1698                 bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1699
1700         /* Releasing pe_num_map */
1701         kfree(pdn->pe_num_map);
1702
1703         return ret;
1704 }
1705
1706 int pnv_pcibios_sriov_disable(struct pci_dev *pdev)
1707 {
1708         pnv_pci_sriov_disable(pdev);
1709
1710         /* Release PCI data */
1711         remove_dev_pci_data(pdev);
1712         return 0;
1713 }
1714
1715 int pnv_pcibios_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1716 {
1717         /* Allocate PCI data */
1718         add_dev_pci_data(pdev);
1719
1720         return pnv_pci_sriov_enable(pdev, num_vfs);
1721 }
1722 #endif /* CONFIG_PCI_IOV */
1723
1724 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb *phb, struct pci_dev *pdev)
1725 {
1726         struct pci_dn *pdn = pci_get_pdn(pdev);
1727         struct pnv_ioda_pe *pe;
1728
1729         /*
1730          * The function can be called while the PE#
1731          * hasn't been assigned. Do nothing for the
1732          * case.
1733          */
1734         if (!pdn || pdn->pe_number == IODA_INVALID_PE)
1735                 return;
1736
1737         pe = &phb->ioda.pe_array[pdn->pe_number];
1738         WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops);
1739         set_dma_offset(&pdev->dev, pe->tce_bypass_base);
1740         set_iommu_table_base(&pdev->dev, pe->table_group.tables[0]);
1741         /*
1742          * Note: iommu_add_device() will fail here as
1743          * for physical PE: the device is already added by now;
1744          * for virtual PE: sysfs entries are not ready yet and
1745          * tce_iommu_bus_notifier will add the device to a group later.
1746          */
1747 }
1748
1749 static bool pnv_pci_ioda_pe_single_vendor(struct pnv_ioda_pe *pe)
1750 {
1751         unsigned short vendor = 0;
1752         struct pci_dev *pdev;
1753
1754         if (pe->device_count == 1)
1755                 return true;
1756
1757         /* pe->pdev should be set if it's a single device, pe->pbus if not */
1758         if (!pe->pbus)
1759                 return true;
1760
1761         list_for_each_entry(pdev, &pe->pbus->devices, bus_list) {
1762                 if (!vendor) {
1763                         vendor = pdev->vendor;
1764                         continue;
1765                 }
1766
1767                 if (pdev->vendor != vendor)
1768                         return false;
1769         }
1770
1771         return true;
1772 }
1773
1774 /*
1775  * Reconfigure TVE#0 to be usable as 64-bit DMA space.
1776  *
1777  * The first 4GB of virtual memory for a PE is reserved for 32-bit accesses.
1778  * Devices can only access more than that if bit 59 of the PCI address is set
1779  * by hardware, which indicates TVE#1 should be used instead of TVE#0.
1780  * Many PCI devices are not capable of addressing that many bits, and as a
1781  * result are limited to the 4GB of virtual memory made available to 32-bit
1782  * devices in TVE#0.
1783  *
1784  * In order to work around this, reconfigure TVE#0 to be suitable for 64-bit
1785  * devices by configuring the virtual memory past the first 4GB inaccessible
1786  * by 64-bit DMAs.  This should only be used by devices that want more than
1787  * 4GB, and only on PEs that have no 32-bit devices.
1788  *
1789  * Currently this will only work on PHB3 (POWER8).
1790  */
1791 static int pnv_pci_ioda_dma_64bit_bypass(struct pnv_ioda_pe *pe)
1792 {
1793         u64 window_size, table_size, tce_count, addr;
1794         struct page *table_pages;
1795         u64 tce_order = 28; /* 256MB TCEs */
1796         __be64 *tces;
1797         s64 rc;
1798
1799         /*
1800          * Window size needs to be a power of two, but needs to account for
1801          * shifting memory by the 4GB offset required to skip 32bit space.
1802          */
1803         window_size = roundup_pow_of_two(memory_hotplug_max() + (1ULL << 32));
1804         tce_count = window_size >> tce_order;
1805         table_size = tce_count << 3;
1806
1807         if (table_size < PAGE_SIZE)
1808                 table_size = PAGE_SIZE;
1809
1810         table_pages = alloc_pages_node(pe->phb->hose->node, GFP_KERNEL,
1811                                        get_order(table_size));
1812         if (!table_pages)
1813                 goto err;
1814
1815         tces = page_address(table_pages);
1816         if (!tces)
1817                 goto err;
1818
1819         memset(tces, 0, table_size);
1820
1821         for (addr = 0; addr < memory_hotplug_max(); addr += (1 << tce_order)) {
1822                 tces[(addr + (1ULL << 32)) >> tce_order] =
1823                         cpu_to_be64(addr | TCE_PCI_READ | TCE_PCI_WRITE);
1824         }
1825
1826         rc = opal_pci_map_pe_dma_window(pe->phb->opal_id,
1827                                         pe->pe_number,
1828                                         /* reconfigure window 0 */
1829                                         (pe->pe_number << 1) + 0,
1830                                         1,
1831                                         __pa(tces),
1832                                         table_size,
1833                                         1 << tce_order);
1834         if (rc == OPAL_SUCCESS) {
1835                 pe_info(pe, "Using 64-bit DMA iommu bypass (through TVE#0)\n");
1836                 return 0;
1837         }
1838 err:
1839         pe_err(pe, "Error configuring 64-bit DMA bypass\n");
1840         return -EIO;
1841 }
1842
1843 static int pnv_pci_ioda_dma_set_mask(struct pci_dev *pdev, u64 dma_mask)
1844 {
1845         struct pci_controller *hose = pci_bus_to_host(pdev->bus);
1846         struct pnv_phb *phb = hose->private_data;
1847         struct pci_dn *pdn = pci_get_pdn(pdev);
1848         struct pnv_ioda_pe *pe;
1849         uint64_t top;
1850         bool bypass = false;
1851         s64 rc;
1852
1853         if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1854                 return -ENODEV;
1855
1856         pe = &phb->ioda.pe_array[pdn->pe_number];
1857         if (pe->tce_bypass_enabled) {
1858                 top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1;
1859                 bypass = (dma_mask >= top);
1860         }
1861
1862         if (bypass) {
1863                 dev_info(&pdev->dev, "Using 64-bit DMA iommu bypass\n");
1864                 set_dma_ops(&pdev->dev, &dma_nommu_ops);
1865         } else {
1866                 /*
1867                  * If the device can't set the TCE bypass bit but still wants
1868                  * to access 4GB or more, on PHB3 we can reconfigure TVE#0 to
1869                  * bypass the 32-bit region and be usable for 64-bit DMAs.
1870                  * The device needs to be able to address all of this space.
1871                  */
1872                 if (dma_mask >> 32 &&
1873                     dma_mask > (memory_hotplug_max() + (1ULL << 32)) &&
1874                     pnv_pci_ioda_pe_single_vendor(pe) &&
1875                     phb->model == PNV_PHB_MODEL_PHB3) {
1876                         /* Configure the bypass mode */
1877                         rc = pnv_pci_ioda_dma_64bit_bypass(pe);
1878                         if (rc)
1879                                 return rc;
1880                         /* 4GB offset bypasses 32-bit space */
1881                         set_dma_offset(&pdev->dev, (1ULL << 32));
1882                         set_dma_ops(&pdev->dev, &dma_nommu_ops);
1883                 } else if (dma_mask >> 32 && dma_mask != DMA_BIT_MASK(64)) {
1884                         /*
1885                          * Fail the request if a DMA mask between 32 and 64 bits
1886                          * was requested but couldn't be fulfilled. Ideally we
1887                          * would do this for 64-bits but historically we have
1888                          * always fallen back to 32-bits.
1889                          */
1890                         return -ENOMEM;
1891                 } else {
1892                         dev_info(&pdev->dev, "Using 32-bit DMA via iommu\n");
1893                         set_dma_ops(&pdev->dev, &dma_iommu_ops);
1894                 }
1895         }
1896         *pdev->dev.dma_mask = dma_mask;
1897
1898         /* Update peer npu devices */
1899         pnv_npu_try_dma_set_bypass(pdev, bypass);
1900
1901         return 0;
1902 }
1903
1904 static u64 pnv_pci_ioda_dma_get_required_mask(struct pci_dev *pdev)
1905 {
1906         struct pci_controller *hose = pci_bus_to_host(pdev->bus);
1907         struct pnv_phb *phb = hose->private_data;
1908         struct pci_dn *pdn = pci_get_pdn(pdev);
1909         struct pnv_ioda_pe *pe;
1910         u64 end, mask;
1911
1912         if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1913                 return 0;
1914
1915         pe = &phb->ioda.pe_array[pdn->pe_number];
1916         if (!pe->tce_bypass_enabled)
1917                 return __dma_get_required_mask(&pdev->dev);
1918
1919
1920         end = pe->tce_bypass_base + memblock_end_of_DRAM();
1921         mask = 1ULL << (fls64(end) - 1);
1922         mask += mask - 1;
1923
1924         return mask;
1925 }
1926
1927 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe,
1928                                    struct pci_bus *bus,
1929                                    bool add_to_group)
1930 {
1931         struct pci_dev *dev;
1932
1933         list_for_each_entry(dev, &bus->devices, bus_list) {
1934                 set_iommu_table_base(&dev->dev, pe->table_group.tables[0]);
1935                 set_dma_offset(&dev->dev, pe->tce_bypass_base);
1936                 if (add_to_group)
1937                         iommu_add_device(&dev->dev);
1938
1939                 if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1940                         pnv_ioda_setup_bus_dma(pe, dev->subordinate,
1941                                         add_to_group);
1942         }
1943 }
1944
1945 static inline __be64 __iomem *pnv_ioda_get_inval_reg(struct pnv_phb *phb,
1946                                                      bool real_mode)
1947 {
1948         return real_mode ? (__be64 __iomem *)(phb->regs_phys + 0x210) :
1949                 (phb->regs + 0x210);
1950 }
1951
1952 static void pnv_pci_p7ioc_tce_invalidate(struct iommu_table *tbl,
1953                 unsigned long index, unsigned long npages, bool rm)
1954 {
1955         struct iommu_table_group_link *tgl = list_first_entry_or_null(
1956                         &tbl->it_group_list, struct iommu_table_group_link,
1957                         next);
1958         struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1959                         struct pnv_ioda_pe, table_group);
1960         __be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm);
1961         unsigned long start, end, inc;
1962
1963         start = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset);
1964         end = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset +
1965                         npages - 1);
1966
1967         /* p7ioc-style invalidation, 2 TCEs per write */
1968         start |= (1ull << 63);
1969         end |= (1ull << 63);
1970         inc = 16;
1971         end |= inc - 1; /* round up end to be different than start */
1972
1973         mb(); /* Ensure above stores are visible */
1974         while (start <= end) {
1975                 if (rm)
1976                         __raw_rm_writeq_be(start, invalidate);
1977                 else
1978                         __raw_writeq_be(start, invalidate);
1979
1980                 start += inc;
1981         }
1982
1983         /*
1984          * The iommu layer will do another mb() for us on build()
1985          * and we don't care on free()
1986          */
1987 }
1988
1989 static int pnv_ioda1_tce_build(struct iommu_table *tbl, long index,
1990                 long npages, unsigned long uaddr,
1991                 enum dma_data_direction direction,
1992                 unsigned long attrs)
1993 {
1994         int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1995                         attrs);
1996
1997         if (!ret)
1998                 pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false);
1999
2000         return ret;
2001 }
2002
2003 #ifdef CONFIG_IOMMU_API
2004 static int pnv_ioda1_tce_xchg(struct iommu_table *tbl, long index,
2005                 unsigned long *hpa, enum dma_data_direction *direction)
2006 {
2007         long ret = pnv_tce_xchg(tbl, index, hpa, direction, true);
2008
2009         if (!ret)
2010                 pnv_pci_p7ioc_tce_invalidate(tbl, index, 1, false);
2011
2012         return ret;
2013 }
2014
2015 static int pnv_ioda1_tce_xchg_rm(struct iommu_table *tbl, long index,
2016                 unsigned long *hpa, enum dma_data_direction *direction)
2017 {
2018         long ret = pnv_tce_xchg(tbl, index, hpa, direction, false);
2019
2020         if (!ret)
2021                 pnv_pci_p7ioc_tce_invalidate(tbl, index, 1, true);
2022
2023         return ret;
2024 }
2025 #endif
2026
2027 static void pnv_ioda1_tce_free(struct iommu_table *tbl, long index,
2028                 long npages)
2029 {
2030         pnv_tce_free(tbl, index, npages);
2031
2032         pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false);
2033 }
2034
2035 static struct iommu_table_ops pnv_ioda1_iommu_ops = {
2036         .set = pnv_ioda1_tce_build,
2037 #ifdef CONFIG_IOMMU_API
2038         .exchange = pnv_ioda1_tce_xchg,
2039         .exchange_rm = pnv_ioda1_tce_xchg_rm,
2040         .useraddrptr = pnv_tce_useraddrptr,
2041 #endif
2042         .clear = pnv_ioda1_tce_free,
2043         .get = pnv_tce_get,
2044 };
2045
2046 #define PHB3_TCE_KILL_INVAL_ALL         PPC_BIT(0)
2047 #define PHB3_TCE_KILL_INVAL_PE          PPC_BIT(1)
2048 #define PHB3_TCE_KILL_INVAL_ONE         PPC_BIT(2)
2049
2050 static void pnv_pci_phb3_tce_invalidate_entire(struct pnv_phb *phb, bool rm)
2051 {
2052         __be64 __iomem *invalidate = pnv_ioda_get_inval_reg(phb, rm);
2053         const unsigned long val = PHB3_TCE_KILL_INVAL_ALL;
2054
2055         mb(); /* Ensure previous TCE table stores are visible */
2056         if (rm)
2057                 __raw_rm_writeq_be(val, invalidate);
2058         else
2059                 __raw_writeq_be(val, invalidate);
2060 }
2061
2062 static inline void pnv_pci_phb3_tce_invalidate_pe(struct pnv_ioda_pe *pe)
2063 {
2064         /* 01xb - invalidate TCEs that match the specified PE# */
2065         __be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, false);
2066         unsigned long val = PHB3_TCE_KILL_INVAL_PE | (pe->pe_number & 0xFF);
2067
2068         mb(); /* Ensure above stores are visible */
2069         __raw_writeq_be(val, invalidate);
2070 }
2071
2072 static void pnv_pci_phb3_tce_invalidate(struct pnv_ioda_pe *pe, bool rm,
2073                                         unsigned shift, unsigned long index,
2074                                         unsigned long npages)
2075 {
2076         __be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm);
2077         unsigned long start, end, inc;
2078
2079         /* We'll invalidate DMA address in PE scope */
2080         start = PHB3_TCE_KILL_INVAL_ONE;
2081         start |= (pe->pe_number & 0xFF);
2082         end = start;
2083
2084         /* Figure out the start, end and step */
2085         start |= (index << shift);
2086         end |= ((index + npages - 1) << shift);
2087         inc = (0x1ull << shift);
2088         mb();
2089
2090         while (start <= end) {
2091                 if (rm)
2092                         __raw_rm_writeq_be(start, invalidate);
2093                 else
2094                         __raw_writeq_be(start, invalidate);
2095                 start += inc;
2096         }
2097 }
2098
2099 static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe *pe)
2100 {
2101         struct pnv_phb *phb = pe->phb;
2102
2103         if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
2104                 pnv_pci_phb3_tce_invalidate_pe(pe);
2105         else
2106                 opal_pci_tce_kill(phb->opal_id, OPAL_PCI_TCE_KILL_PE,
2107                                   pe->pe_number, 0, 0, 0);
2108 }
2109
2110 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table *tbl,
2111                 unsigned long index, unsigned long npages, bool rm)
2112 {
2113         struct iommu_table_group_link *tgl;
2114
2115         list_for_each_entry_lockless(tgl, &tbl->it_group_list, next) {
2116                 struct pnv_ioda_pe *pe = container_of(tgl->table_group,
2117                                 struct pnv_ioda_pe, table_group);
2118                 struct pnv_phb *phb = pe->phb;
2119                 unsigned int shift = tbl->it_page_shift;
2120
2121                 /*
2122                  * NVLink1 can use the TCE kill register directly as
2123                  * it's the same as PHB3. NVLink2 is different and
2124                  * should go via the OPAL call.
2125                  */
2126                 if (phb->model == PNV_PHB_MODEL_NPU) {
2127                         /*
2128                          * The NVLink hardware does not support TCE kill
2129                          * per TCE entry so we have to invalidate
2130                          * the entire cache for it.
2131                          */
2132                         pnv_pci_phb3_tce_invalidate_entire(phb, rm);
2133                         continue;
2134                 }
2135                 if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
2136                         pnv_pci_phb3_tce_invalidate(pe, rm, shift,
2137                                                     index, npages);
2138                 else
2139                         opal_pci_tce_kill(phb->opal_id,
2140                                           OPAL_PCI_TCE_KILL_PAGES,
2141                                           pe->pe_number, 1u << shift,
2142                                           index << shift, npages);
2143         }
2144 }
2145
2146 void pnv_pci_ioda2_tce_invalidate_entire(struct pnv_phb *phb, bool rm)
2147 {
2148         if (phb->model == PNV_PHB_MODEL_NPU || phb->model == PNV_PHB_MODEL_PHB3)
2149                 pnv_pci_phb3_tce_invalidate_entire(phb, rm);
2150         else
2151                 opal_pci_tce_kill(phb->opal_id, OPAL_PCI_TCE_KILL, 0, 0, 0, 0);
2152 }
2153
2154 static int pnv_ioda2_tce_build(struct iommu_table *tbl, long index,
2155                 long npages, unsigned long uaddr,
2156                 enum dma_data_direction direction,
2157                 unsigned long attrs)
2158 {
2159         int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
2160                         attrs);
2161
2162         if (!ret)
2163                 pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
2164
2165         return ret;
2166 }
2167
2168 #ifdef CONFIG_IOMMU_API
2169 static int pnv_ioda2_tce_xchg(struct iommu_table *tbl, long index,
2170                 unsigned long *hpa, enum dma_data_direction *direction)
2171 {
2172         long ret = pnv_tce_xchg(tbl, index, hpa, direction, true);
2173
2174         if (!ret)
2175                 pnv_pci_ioda2_tce_invalidate(tbl, index, 1, false);
2176
2177         return ret;
2178 }
2179
2180 static int pnv_ioda2_tce_xchg_rm(struct iommu_table *tbl, long index,
2181                 unsigned long *hpa, enum dma_data_direction *direction)
2182 {
2183         long ret = pnv_tce_xchg(tbl, index, hpa, direction, false);
2184
2185         if (!ret)
2186                 pnv_pci_ioda2_tce_invalidate(tbl, index, 1, true);
2187
2188         return ret;
2189 }
2190 #endif
2191
2192 static void pnv_ioda2_tce_free(struct iommu_table *tbl, long index,
2193                 long npages)
2194 {
2195         pnv_tce_free(tbl, index, npages);
2196
2197         pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
2198 }
2199
2200 static struct iommu_table_ops pnv_ioda2_iommu_ops = {
2201         .set = pnv_ioda2_tce_build,
2202 #ifdef CONFIG_IOMMU_API
2203         .exchange = pnv_ioda2_tce_xchg,
2204         .exchange_rm = pnv_ioda2_tce_xchg_rm,
2205         .useraddrptr = pnv_tce_useraddrptr,
2206 #endif
2207         .clear = pnv_ioda2_tce_free,
2208         .get = pnv_tce_get,
2209         .free = pnv_pci_ioda2_table_free_pages,
2210 };
2211
2212 static int pnv_pci_ioda_dev_dma_weight(struct pci_dev *dev, void *data)
2213 {
2214         unsigned int *weight = (unsigned int *)data;
2215
2216         /* This is quite simplistic. The "base" weight of a device
2217          * is 10. 0 means no DMA is to be accounted for it.
2218          */
2219         if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
2220                 return 0;
2221
2222         if (dev->class == PCI_CLASS_SERIAL_USB_UHCI ||
2223             dev->class == PCI_CLASS_SERIAL_USB_OHCI ||
2224             dev->class == PCI_CLASS_SERIAL_USB_EHCI)
2225                 *weight += 3;
2226         else if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID)
2227                 *weight += 15;
2228         else
2229                 *weight += 10;
2230
2231         return 0;
2232 }
2233
2234 static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe *pe)
2235 {
2236         unsigned int weight = 0;
2237
2238         /* SRIOV VF has same DMA32 weight as its PF */
2239 #ifdef CONFIG_PCI_IOV
2240         if ((pe->flags & PNV_IODA_PE_VF) && pe->parent_dev) {
2241                 pnv_pci_ioda_dev_dma_weight(pe->parent_dev, &weight);
2242                 return weight;
2243         }
2244 #endif
2245
2246         if ((pe->flags & PNV_IODA_PE_DEV) && pe->pdev) {
2247                 pnv_pci_ioda_dev_dma_weight(pe->pdev, &weight);
2248         } else if ((pe->flags & PNV_IODA_PE_BUS) && pe->pbus) {
2249                 struct pci_dev *pdev;
2250
2251                 list_for_each_entry(pdev, &pe->pbus->devices, bus_list)
2252                         pnv_pci_ioda_dev_dma_weight(pdev, &weight);
2253         } else if ((pe->flags & PNV_IODA_PE_BUS_ALL) && pe->pbus) {
2254                 pci_walk_bus(pe->pbus, pnv_pci_ioda_dev_dma_weight, &weight);
2255         }
2256
2257         return weight;
2258 }
2259
2260 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb *phb,
2261                                        struct pnv_ioda_pe *pe)
2262 {
2263
2264         struct page *tce_mem = NULL;
2265         struct iommu_table *tbl;
2266         unsigned int weight, total_weight = 0;
2267         unsigned int tce32_segsz, base, segs, avail, i;
2268         int64_t rc;
2269         void *addr;
2270
2271         /* XXX FIXME: Handle 64-bit only DMA devices */
2272         /* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
2273         /* XXX FIXME: Allocate multi-level tables on PHB3 */
2274         weight = pnv_pci_ioda_pe_dma_weight(pe);
2275         if (!weight)
2276                 return;
2277
2278         pci_walk_bus(phb->hose->bus, pnv_pci_ioda_dev_dma_weight,
2279                      &total_weight);
2280         segs = (weight * phb->ioda.dma32_count) / total_weight;
2281         if (!segs)
2282                 segs = 1;
2283
2284         /*
2285          * Allocate contiguous DMA32 segments. We begin with the expected
2286          * number of segments. With one more attempt, the number of DMA32
2287          * segments to be allocated is decreased by one until one segment
2288          * is allocated successfully.
2289          */
2290         do {
2291                 for (base = 0; base <= phb->ioda.dma32_count - segs; base++) {
2292                         for (avail = 0, i = base; i < base + segs; i++) {
2293                                 if (phb->ioda.dma32_segmap[i] ==
2294                                     IODA_INVALID_PE)
2295                                         avail++;
2296                         }
2297
2298                         if (avail == segs)
2299                                 goto found;
2300                 }
2301         } while (--segs);
2302
2303         if (!segs) {
2304                 pe_warn(pe, "No available DMA32 segments\n");
2305                 return;
2306         }
2307
2308 found:
2309         tbl = pnv_pci_table_alloc(phb->hose->node);
2310         if (WARN_ON(!tbl))
2311                 return;
2312
2313         iommu_register_group(&pe->table_group, phb->hose->global_number,
2314                         pe->pe_number);
2315         pnv_pci_link_table_and_group(phb->hose->node, 0, tbl, &pe->table_group);
2316
2317         /* Grab a 32-bit TCE table */
2318         pe_info(pe, "DMA weight %d (%d), assigned (%d) %d DMA32 segments\n",
2319                 weight, total_weight, base, segs);
2320         pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n",
2321                 base * PNV_IODA1_DMA32_SEGSIZE,
2322                 (base + segs) * PNV_IODA1_DMA32_SEGSIZE - 1);
2323
2324         /* XXX Currently, we allocate one big contiguous table for the
2325          * TCEs. We only really need one chunk per 256M of TCE space
2326          * (ie per segment) but that's an optimization for later, it
2327          * requires some added smarts with our get/put_tce implementation
2328          *
2329          * Each TCE page is 4KB in size and each TCE entry occupies 8
2330          * bytes
2331          */
2332         tce32_segsz = PNV_IODA1_DMA32_SEGSIZE >> (IOMMU_PAGE_SHIFT_4K - 3);
2333         tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
2334                                    get_order(tce32_segsz * segs));
2335         if (!tce_mem) {
2336                 pe_err(pe, " Failed to allocate a 32-bit TCE memory\n");
2337                 goto fail;
2338         }
2339         addr = page_address(tce_mem);
2340         memset(addr, 0, tce32_segsz * segs);
2341
2342         /* Configure HW */
2343         for (i = 0; i < segs; i++) {
2344                 rc = opal_pci_map_pe_dma_window(phb->opal_id,
2345                                               pe->pe_number,
2346                                               base + i, 1,
2347                                               __pa(addr) + tce32_segsz * i,
2348                                               tce32_segsz, IOMMU_PAGE_SIZE_4K);
2349                 if (rc) {
2350                         pe_err(pe, " Failed to configure 32-bit TCE table,"
2351                                " err %ld\n", rc);
2352                         goto fail;
2353                 }
2354         }
2355
2356         /* Setup DMA32 segment mapping */
2357         for (i = base; i < base + segs; i++)
2358                 phb->ioda.dma32_segmap[i] = pe->pe_number;
2359
2360         /* Setup linux iommu table */
2361         pnv_pci_setup_iommu_table(tbl, addr, tce32_segsz * segs,
2362                                   base * PNV_IODA1_DMA32_SEGSIZE,
2363                                   IOMMU_PAGE_SHIFT_4K);
2364
2365         tbl->it_ops = &pnv_ioda1_iommu_ops;
2366         pe->table_group.tce32_start = tbl->it_offset << tbl->it_page_shift;
2367         pe->table_group.tce32_size = tbl->it_size << tbl->it_page_shift;
2368         iommu_init_table(tbl, phb->hose->node);
2369
2370         if (pe->flags & PNV_IODA_PE_DEV) {
2371                 /*
2372                  * Setting table base here only for carrying iommu_group
2373                  * further down to let iommu_add_device() do the job.
2374                  * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2375                  */
2376                 set_iommu_table_base(&pe->pdev->dev, tbl);
2377                 iommu_add_device(&pe->pdev->dev);
2378         } else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
2379                 pnv_ioda_setup_bus_dma(pe, pe->pbus, true);
2380
2381         return;
2382  fail:
2383         /* XXX Failure: Try to fallback to 64-bit only ? */
2384         if (tce_mem)
2385                 __free_pages(tce_mem, get_order(tce32_segsz * segs));
2386         if (tbl) {
2387                 pnv_pci_unlink_table_and_group(tbl, &pe->table_group);
2388                 iommu_tce_table_put(tbl);
2389         }
2390 }
2391
2392 static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group,
2393                 int num, struct iommu_table *tbl)
2394 {
2395         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2396                         table_group);
2397         struct pnv_phb *phb = pe->phb;
2398         int64_t rc;
2399         const unsigned long size = tbl->it_indirect_levels ?
2400                         tbl->it_level_size : tbl->it_size;
2401         const __u64 start_addr = tbl->it_offset << tbl->it_page_shift;
2402         const __u64 win_size = tbl->it_size << tbl->it_page_shift;
2403
2404         pe_info(pe, "Setting up window#%d %llx..%llx pg=%x\n", num,
2405                         start_addr, start_addr + win_size - 1,
2406                         IOMMU_PAGE_SIZE(tbl));
2407
2408         /*
2409          * Map TCE table through TVT. The TVE index is the PE number
2410          * shifted by 1 bit for 32-bits DMA space.
2411          */
2412         rc = opal_pci_map_pe_dma_window(phb->opal_id,
2413                         pe->pe_number,
2414                         (pe->pe_number << 1) + num,
2415                         tbl->it_indirect_levels + 1,
2416                         __pa(tbl->it_base),
2417                         size << 3,
2418                         IOMMU_PAGE_SIZE(tbl));
2419         if (rc) {
2420                 pe_err(pe, "Failed to configure TCE table, err %ld\n", rc);
2421                 return rc;
2422         }
2423
2424         pnv_pci_link_table_and_group(phb->hose->node, num,
2425                         tbl, &pe->table_group);
2426         pnv_pci_ioda2_tce_invalidate_pe(pe);
2427
2428         return 0;
2429 }
2430
2431 void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable)
2432 {
2433         uint16_t window_id = (pe->pe_number << 1 ) + 1;
2434         int64_t rc;
2435
2436         pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis");
2437         if (enable) {
2438                 phys_addr_t top = memblock_end_of_DRAM();
2439
2440                 top = roundup_pow_of_two(top);
2441                 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
2442                                                      pe->pe_number,
2443                                                      window_id,
2444                                                      pe->tce_bypass_base,
2445                                                      top);
2446         } else {
2447                 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
2448                                                      pe->pe_number,
2449                                                      window_id,
2450                                                      pe->tce_bypass_base,
2451                                                      0);
2452         }
2453         if (rc)
2454                 pe_err(pe, "OPAL error %lld configuring bypass window\n", rc);
2455         else
2456                 pe->tce_bypass_enabled = enable;
2457 }
2458
2459 static long pnv_pci_ioda2_create_table(struct iommu_table_group *table_group,
2460                 int num, __u32 page_shift, __u64 window_size, __u32 levels,
2461                 bool alloc_userspace_copy, struct iommu_table **ptbl)
2462 {
2463         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2464                         table_group);
2465         int nid = pe->phb->hose->node;
2466         __u64 bus_offset = num ? pe->tce_bypass_base : table_group->tce32_start;
2467         long ret;
2468         struct iommu_table *tbl;
2469
2470         tbl = pnv_pci_table_alloc(nid);
2471         if (!tbl)
2472                 return -ENOMEM;
2473
2474         tbl->it_ops = &pnv_ioda2_iommu_ops;
2475
2476         ret = pnv_pci_ioda2_table_alloc_pages(nid,
2477                         bus_offset, page_shift, window_size,
2478                         levels, alloc_userspace_copy, tbl);
2479         if (ret) {
2480                 iommu_tce_table_put(tbl);
2481                 return ret;
2482         }
2483
2484         *ptbl = tbl;
2485
2486         return 0;
2487 }
2488
2489 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe *pe)
2490 {
2491         struct iommu_table *tbl = NULL;
2492         long rc;
2493
2494         /*
2495          * crashkernel= specifies the kdump kernel's maximum memory at
2496          * some offset and there is no guaranteed the result is a power
2497          * of 2, which will cause errors later.
2498          */
2499         const u64 max_memory = __rounddown_pow_of_two(memory_hotplug_max());
2500
2501         /*
2502          * In memory constrained environments, e.g. kdump kernel, the
2503          * DMA window can be larger than available memory, which will
2504          * cause errors later.
2505          */
2506         const u64 window_size = min((u64)pe->table_group.tce32_size, max_memory);
2507
2508         rc = pnv_pci_ioda2_create_table(&pe->table_group, 0,
2509                         IOMMU_PAGE_SHIFT_4K,
2510                         window_size,
2511                         POWERNV_IOMMU_DEFAULT_LEVELS, false, &tbl);
2512         if (rc) {
2513                 pe_err(pe, "Failed to create 32-bit TCE table, err %ld",
2514                                 rc);
2515                 return rc;
2516         }
2517
2518         iommu_init_table(tbl, pe->phb->hose->node);
2519
2520         rc = pnv_pci_ioda2_set_window(&pe->table_group, 0, tbl);
2521         if (rc) {
2522                 pe_err(pe, "Failed to configure 32-bit TCE table, err %ld\n",
2523                                 rc);
2524                 iommu_tce_table_put(tbl);
2525                 return rc;
2526         }
2527
2528         if (!pnv_iommu_bypass_disabled)
2529                 pnv_pci_ioda2_set_bypass(pe, true);
2530
2531         /*
2532          * Setting table base here only for carrying iommu_group
2533          * further down to let iommu_add_device() do the job.
2534          * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2535          */
2536         if (pe->flags & PNV_IODA_PE_DEV)
2537                 set_iommu_table_base(&pe->pdev->dev, tbl);
2538
2539         return 0;
2540 }
2541
2542 #if defined(CONFIG_IOMMU_API) || defined(CONFIG_PCI_IOV)
2543 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
2544                 int num)
2545 {
2546         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2547                         table_group);
2548         struct pnv_phb *phb = pe->phb;
2549         long ret;
2550
2551         pe_info(pe, "Removing DMA window #%d\n", num);
2552
2553         ret = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
2554                         (pe->pe_number << 1) + num,
2555                         0/* levels */, 0/* table address */,
2556                         0/* table size */, 0/* page size */);
2557         if (ret)
2558                 pe_warn(pe, "Unmapping failed, ret = %ld\n", ret);
2559         else
2560                 pnv_pci_ioda2_tce_invalidate_pe(pe);
2561
2562         pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
2563
2564         return ret;
2565 }
2566 #endif
2567
2568 #ifdef CONFIG_IOMMU_API
2569 static unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift,
2570                 __u64 window_size, __u32 levels)
2571 {
2572         unsigned long bytes = 0;
2573         const unsigned window_shift = ilog2(window_size);
2574         unsigned entries_shift = window_shift - page_shift;
2575         unsigned table_shift = entries_shift + 3;
2576         unsigned long tce_table_size = max(0x1000UL, 1UL << table_shift);
2577         unsigned long direct_table_size;
2578
2579         if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS) ||
2580                         !is_power_of_2(window_size))
2581                 return 0;
2582
2583         /* Calculate a direct table size from window_size and levels */
2584         entries_shift = (entries_shift + levels - 1) / levels;
2585         table_shift = entries_shift + 3;
2586         table_shift = max_t(unsigned, table_shift, PAGE_SHIFT);
2587         direct_table_size =  1UL << table_shift;
2588
2589         for ( ; levels; --levels) {
2590                 bytes += _ALIGN_UP(tce_table_size, direct_table_size);
2591
2592                 tce_table_size /= direct_table_size;
2593                 tce_table_size <<= 3;
2594                 tce_table_size = max_t(unsigned long,
2595                                 tce_table_size, direct_table_size);
2596         }
2597
2598         return bytes + bytes; /* one for HW table, one for userspace copy */
2599 }
2600
2601 static long pnv_pci_ioda2_create_table_userspace(
2602                 struct iommu_table_group *table_group,
2603                 int num, __u32 page_shift, __u64 window_size, __u32 levels,
2604                 struct iommu_table **ptbl)
2605 {
2606         long ret = pnv_pci_ioda2_create_table(table_group,
2607                         num, page_shift, window_size, levels, true, ptbl);
2608
2609         if (!ret)
2610                 (*ptbl)->it_allocated_size = pnv_pci_ioda2_get_table_size(
2611                                 page_shift, window_size, levels);
2612         return ret;
2613 }
2614
2615 static void pnv_ioda2_take_ownership(struct iommu_table_group *table_group)
2616 {
2617         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2618                                                 table_group);
2619         /* Store @tbl as pnv_pci_ioda2_unset_window() resets it */
2620         struct iommu_table *tbl = pe->table_group.tables[0];
2621
2622         pnv_pci_ioda2_set_bypass(pe, false);
2623         pnv_pci_ioda2_unset_window(&pe->table_group, 0);
2624         if (pe->pbus)
2625                 pnv_ioda_setup_bus_dma(pe, pe->pbus, false);
2626         iommu_tce_table_put(tbl);
2627 }
2628
2629 static void pnv_ioda2_release_ownership(struct iommu_table_group *table_group)
2630 {
2631         struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2632                                                 table_group);
2633
2634         pnv_pci_ioda2_setup_default_config(pe);
2635         if (pe->pbus)
2636                 pnv_ioda_setup_bus_dma(pe, pe->pbus, false);
2637 }
2638
2639 static struct iommu_table_group_ops pnv_pci_ioda2_ops = {
2640         .get_table_size = pnv_pci_ioda2_get_table_size,
2641         .create_table = pnv_pci_ioda2_create_table_userspace,
2642         .set_window = pnv_pci_ioda2_set_window,
2643         .unset_window = pnv_pci_ioda2_unset_window,
2644         .take_ownership = pnv_ioda2_take_ownership,
2645         .release_ownership = pnv_ioda2_release_ownership,
2646 };
2647
2648 static int gpe_table_group_to_npe_cb(struct device *dev, void *opaque)
2649 {
2650         struct pci_controller *hose;
2651         struct pnv_phb *phb;
2652         struct pnv_ioda_pe **ptmppe = opaque;
2653         struct pci_dev *pdev = container_of(dev, struct pci_dev, dev);
2654         struct pci_dn *pdn = pci_get_pdn(pdev);
2655
2656         if (!pdn || pdn->pe_number == IODA_INVALID_PE)
2657                 return 0;
2658
2659         hose = pci_bus_to_host(pdev->bus);
2660         phb = hose->private_data;
2661         if (phb->type != PNV_PHB_NPU_NVLINK)
2662                 return 0;
2663
2664         *ptmppe = &phb->ioda.pe_array[pdn->pe_number];
2665
2666         return 1;
2667 }
2668
2669 /*
2670  * This returns PE of associated NPU.
2671  * This assumes that NPU is in the same IOMMU group with GPU and there is
2672  * no other PEs.
2673  */
2674 static struct pnv_ioda_pe *gpe_table_group_to_npe(
2675                 struct iommu_table_group *table_group)
2676 {
2677         struct pnv_ioda_pe *npe = NULL;
2678         int ret = iommu_group_for_each_dev(table_group->group, &npe,
2679                         gpe_table_group_to_npe_cb);
2680
2681         BUG_ON(!ret || !npe);
2682
2683         return npe;
2684 }
2685
2686 static long pnv_pci_ioda2_npu_set_window(struct iommu_table_group *table_group,
2687                 int num, struct iommu_table *tbl)
2688 {
2689         struct pnv_ioda_pe *npe = gpe_table_group_to_npe(table_group);
2690         int num2 = (num == 0) ? 1 : 0;
2691         long ret = pnv_pci_ioda2_set_window(table_group, num, tbl);
2692
2693         if (ret)
2694                 return ret;
2695
2696         if (table_group->tables[num2])
2697                 pnv_npu_unset_window(npe, num2);
2698
2699         ret = pnv_npu_set_window(npe, num, tbl);
2700         if (ret) {
2701                 pnv_pci_ioda2_unset_window(table_group, num);
2702                 if (table_group->tables[num2])
2703                         pnv_npu_set_window(npe, num2,
2704                                         table_group->tables[num2]);
2705         }
2706
2707         return ret;
2708 }
2709
2710 static long pnv_pci_ioda2_npu_unset_window(
2711                 struct iommu_table_group *table_group,
2712                 int num)
2713 {
2714         struct pnv_ioda_pe *npe = gpe_table_group_to_npe(table_group);
2715         int num2 = (num == 0) ? 1 : 0;
2716         long ret = pnv_pci_ioda2_unset_window(table_group, num);
2717
2718         if (ret)
2719                 return ret;
2720
2721         if (!npe->table_group.tables[num])
2722                 return 0;
2723
2724         ret = pnv_npu_unset_window(npe, num);
2725         if (ret)
2726                 return ret;
2727
2728         if (table_group->tables[num2])
2729                 ret = pnv_npu_set_window(npe, num2, table_group->tables[num2]);
2730
2731         return ret;
2732 }
2733
2734 static void pnv_ioda2_npu_take_ownership(struct iommu_table_group *table_group)
2735 {
2736         /*
2737          * Detach NPU first as pnv_ioda2_take_ownership() will destroy
2738          * the iommu_table if 32bit DMA is enabled.
2739          */
2740         pnv_npu_take_ownership(gpe_table_group_to_npe(table_group));
2741         pnv_ioda2_take_ownership(table_group);
2742 }
2743
2744 static struct iommu_table_group_ops pnv_pci_ioda2_npu_ops = {
2745         .get_table_size = pnv_pci_ioda2_get_table_size,
2746         .create_table = pnv_pci_ioda2_create_table_userspace,
2747         .set_window = pnv_pci_ioda2_npu_set_window,
2748         .unset_window = pnv_pci_ioda2_npu_unset_window,
2749         .take_ownership = pnv_ioda2_npu_take_ownership,
2750         .release_ownership = pnv_ioda2_release_ownership,
2751 };
2752
2753 static void pnv_pci_ioda_setup_iommu_api(void)
2754 {
2755         struct pci_controller *hose, *tmp;
2756         struct pnv_phb *phb;
2757         struct pnv_ioda_pe *pe, *gpe;
2758
2759         /*
2760          * Now we have all PHBs discovered, time to add NPU devices to
2761          * the corresponding IOMMU groups.
2762          */
2763         list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
2764                 phb = hose->private_data;
2765
2766                 if (phb->type != PNV_PHB_NPU_NVLINK)
2767                         continue;
2768
2769                 list_for_each_entry(pe, &phb->ioda.pe_list, list) {
2770                         gpe = pnv_pci_npu_setup_iommu(pe);
2771                         if (gpe)
2772                                 gpe->table_group.ops = &pnv_pci_ioda2_npu_ops;
2773                 }
2774         }
2775 }
2776 #else /* !CONFIG_IOMMU_API */
2777 static void pnv_pci_ioda_setup_iommu_api(void) { };
2778 #endif
2779
2780 static unsigned long pnv_ioda_parse_tce_sizes(struct pnv_phb *phb)
2781 {
2782         struct pci_controller *hose = phb->hose;
2783         struct device_node *dn = hose->dn;
2784         unsigned long mask = 0;
2785         int i, rc, count;
2786         u32 val;
2787
2788         count = of_property_count_u32_elems(dn, "ibm,supported-tce-sizes");
2789         if (count <= 0) {
2790                 mask = SZ_4K | SZ_64K;
2791                 /* Add 16M for POWER8 by default */
2792                 if (cpu_has_feature(CPU_FTR_ARCH_207S) &&
2793                                 !cpu_has_feature(CPU_FTR_ARCH_300))
2794                         mask |= SZ_16M | SZ_256M;
2795                 return mask;
2796         }
2797
2798         for (i = 0; i < count; i++) {
2799                 rc = of_property_read_u32_index(dn, "ibm,supported-tce-sizes",
2800                                                 i, &val);
2801                 if (rc == 0)
2802                         mask |= 1ULL << val;
2803         }
2804
2805         return mask;
2806 }
2807
2808 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
2809                                        struct pnv_ioda_pe *pe)
2810 {
2811         int64_t rc;
2812
2813         if (!pnv_pci_ioda_pe_dma_weight(pe))
2814                 return;
2815
2816         /* TVE #1 is selected by PCI address bit 59 */
2817         pe->tce_bypass_base = 1ull << 59;
2818
2819         iommu_register_group(&pe->table_group, phb->hose->global_number,
2820                         pe->pe_number);
2821
2822         /* The PE will reserve all possible 32-bits space */
2823         pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n",
2824                 phb->ioda.m32_pci_base);
2825
2826         /* Setup linux iommu table */
2827         pe->table_group.tce32_start = 0;
2828         pe->table_group.tce32_size = phb->ioda.m32_pci_base;
2829         pe->table_group.max_dynamic_windows_supported =
2830                         IOMMU_TABLE_GROUP_MAX_TABLES;
2831         pe->table_group.max_levels = POWERNV_IOMMU_MAX_LEVELS;
2832         pe->table_group.pgsizes = pnv_ioda_parse_tce_sizes(phb);
2833 #ifdef CONFIG_IOMMU_API
2834         pe->table_group.ops = &pnv_pci_ioda2_ops;
2835 #endif
2836
2837         rc = pnv_pci_ioda2_setup_default_config(pe);
2838         if (rc)
2839                 return;
2840
2841         if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
2842                 pnv_ioda_setup_bus_dma(pe, pe->pbus, true);
2843 }
2844
2845 #ifdef CONFIG_PCI_MSI
2846 int64_t pnv_opal_pci_msi_eoi(struct irq_chip *chip, unsigned int hw_irq)
2847 {
2848         struct pnv_phb *phb = container_of(chip, struct pnv_phb,
2849                                            ioda.irq_chip);
2850
2851         return opal_pci_msi_eoi(phb->opal_id, hw_irq);
2852 }
2853
2854 static void pnv_ioda2_msi_eoi(struct irq_data *d)
2855 {
2856         int64_t rc;
2857         unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
2858         struct irq_chip *chip = irq_data_get_irq_chip(d);
2859
2860         rc = pnv_opal_pci_msi_eoi(chip, hw_irq);
2861         WARN_ON_ONCE(rc);
2862
2863         icp_native_eoi(d);
2864 }
2865
2866
2867 void pnv_set_msi_irq_chip(struct pnv_phb *phb, unsigned int virq)
2868 {
2869         struct irq_data *idata;
2870         struct irq_chip *ichip;
2871
2872         /* The MSI EOI OPAL call is only needed on PHB3 */
2873         if (phb->model != PNV_PHB_MODEL_PHB3)
2874                 return;
2875
2876         if (!phb->ioda.irq_chip_init) {
2877                 /*
2878                  * First time we setup an MSI IRQ, we need to setup the
2879                  * corresponding IRQ chip to route correctly.
2880                  */
2881                 idata = irq_get_irq_data(virq);
2882                 ichip = irq_data_get_irq_chip(idata);
2883                 phb->ioda.irq_chip_init = 1;
2884                 phb->ioda.irq_chip = *ichip;
2885                 phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi;
2886         }
2887         irq_set_chip(virq, &phb->ioda.irq_chip);
2888 }
2889
2890 /*
2891  * Returns true iff chip is something that we could call
2892  * pnv_opal_pci_msi_eoi for.
2893  */
2894 bool is_pnv_opal_msi(struct irq_chip *chip)
2895 {
2896         return chip->irq_eoi == pnv_ioda2_msi_eoi;
2897 }
2898 EXPORT_SYMBOL_GPL(is_pnv_opal_msi);
2899
2900 static int pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev,
2901                                   unsigned int hwirq, unsigned int virq,
2902                                   unsigned int is_64, struct msi_msg *msg)
2903 {
2904         struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev);
2905         unsigned int xive_num = hwirq - phb->msi_base;
2906         __be32 data;
2907         int rc;
2908
2909         /* No PE assigned ? bail out ... no MSI for you ! */
2910         if (pe == NULL)
2911                 return -ENXIO;
2912
2913         /* Check if we have an MVE */
2914         if (pe->mve_number < 0)
2915                 return -ENXIO;
2916
2917         /* Force 32-bit MSI on some broken devices */
2918         if (dev->no_64bit_msi)
2919                 is_64 = 0;
2920
2921         /* Assign XIVE to PE */
2922         rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
2923         if (rc) {
2924                 pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
2925                         pci_name(dev), rc, xive_num);
2926                 return -EIO;
2927         }
2928
2929         if (is_64) {
2930                 __be64 addr64;
2931
2932                 rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1,
2933                                      &addr64, &data);
2934                 if (rc) {
2935                         pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
2936                                 pci_name(dev), rc);
2937                         return -EIO;
2938                 }
2939                 msg->address_hi = be64_to_cpu(addr64) >> 32;
2940                 msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful;
2941         } else {
2942                 __be32 addr32;
2943
2944                 rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1,
2945                                      &addr32, &data);
2946                 if (rc) {
2947                         pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
2948                                 pci_name(dev), rc);
2949                         return -EIO;
2950                 }
2951                 msg->address_hi = 0;
2952                 msg->address_lo = be32_to_cpu(addr32);
2953         }
2954         msg->data = be32_to_cpu(data);
2955
2956         pnv_set_msi_irq_chip(phb, virq);
2957
2958         pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d),"
2959                  " address=%x_%08x data=%x PE# %x\n",
2960                  pci_name(dev), is_64 ? "64" : "32", hwirq, xive_num,
2961                  msg->address_hi, msg->address_lo, data, pe->pe_number);
2962
2963         return 0;
2964 }
2965
2966 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb)
2967 {
2968         unsigned int count;
2969         const __be32 *prop = of_get_property(phb->hose->dn,
2970                                              "ibm,opal-msi-ranges", NULL);
2971         if (!prop) {
2972                 /* BML Fallback */
2973                 prop = of_get_property(phb->hose->dn, "msi-ranges", NULL);
2974         }
2975         if (!prop)
2976                 return;
2977
2978         phb->msi_base = be32_to_cpup(prop);
2979         count = be32_to_cpup(prop + 1);
2980         if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) {
2981                 pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
2982                        phb->hose->global_number);
2983                 return;
2984         }
2985
2986         phb->msi_setup = pnv_pci_ioda_msi_setup;
2987         phb->msi32_support = 1;
2988         pr_info("  Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
2989                 count, phb->msi_base);
2990 }
2991 #else
2992 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb) { }
2993 #endif /* CONFIG_PCI_MSI */
2994
2995 #ifdef CONFIG_PCI_IOV
2996 static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev *pdev)
2997 {
2998         struct pci_controller *hose = pci_bus_to_host(pdev->bus);
2999         struct pnv_phb *phb = hose->private_data;
3000         const resource_size_t gate = phb->ioda.m64_segsize >> 2;
3001         struct resource *res;
3002         int i;
3003         resource_size_t size, total_vf_bar_sz;
3004         struct pci_dn *pdn;
3005         int mul, total_vfs;
3006
3007         if (!pdev->is_physfn || pci_dev_is_added(pdev))
3008                 return;
3009
3010         pdn = pci_get_pdn(pdev);
3011         pdn->vfs_expanded = 0;
3012         pdn->m64_single_mode = false;
3013
3014         total_vfs = pci_sriov_get_totalvfs(pdev);
3015         mul = phb->ioda.total_pe_num;
3016         total_vf_bar_sz = 0;
3017
3018         for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
3019                 res = &pdev->resource[i + PCI_IOV_RESOURCES];
3020                 if (!res->flags || res->parent)
3021                         continue;
3022                 if (!pnv_pci_is_m64_flags(res->flags)) {
3023                         dev_warn(&pdev->dev, "Don't support SR-IOV with"
3024                                         " non M64 VF BAR%d: %pR. \n",
3025                                  i, res);
3026                         goto truncate_iov;
3027                 }
3028
3029                 total_vf_bar_sz += pci_iov_resource_size(pdev,
3030                                 i + PCI_IOV_RESOURCES);
3031
3032                 /*
3033                  * If bigger than quarter of M64 segment size, just round up
3034                  * power of two.
3035                  *
3036                  * Generally, one M64 BAR maps one IOV BAR. To avoid conflict
3037                  * with other devices, IOV BAR size is expanded to be
3038                  * (total_pe * VF_BAR_size).  When VF_BAR_size is half of M64
3039                  * segment size , the expanded size would equal to half of the
3040                  * whole M64 space size, which will exhaust the M64 Space and
3041                  * limit the system flexibility.  This is a design decision to
3042                  * set the boundary to quarter of the M64 segment size.
3043                  */
3044                 if (total_vf_bar_sz > gate) {
3045                         mul = roundup_pow_of_two(total_vfs);
3046                         dev_info(&pdev->dev,
3047                                 "VF BAR Total IOV size %llx > %llx, roundup to %d VFs\n",
3048                                 total_vf_bar_sz, gate, mul);
3049                         pdn->m64_single_mode = true;
3050                         break;
3051                 }
3052         }
3053
3054         for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
3055                 res = &pdev->resource[i + PCI_IOV_RESOURCES];
3056                 if (!res->flags || res->parent)
3057                         continue;
3058
3059                 size = pci_iov_resource_size(pdev, i + PCI_IOV_RESOURCES);
3060                 /*
3061                  * On PHB3, the minimum size alignment of M64 BAR in single
3062                  * mode is 32MB.
3063                  */
3064                 if (pdn->m64_single_mode && (size < SZ_32M))
3065                         goto truncate_iov;
3066                 dev_dbg(&pdev->dev, " Fixing VF BAR%d: %pR to\n", i, res);
3067                 res->end = res->start + size * mul - 1;
3068                 dev_dbg(&pdev->dev, "                       %pR\n", res);
3069                 dev_info(&pdev->dev, "VF BAR%d: %pR (expanded to %d VFs for PE alignment)",
3070                          i, res, mul);
3071         }
3072         pdn->vfs_expanded = mul;
3073
3074         return;
3075
3076 truncate_iov:
3077         /* To save MMIO space, IOV BAR is truncated. */
3078         for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
3079                 res = &pdev->resource[i + PCI_IOV_RESOURCES];
3080                 res->flags = 0;
3081                 res->end = res->start - 1;
3082         }
3083 }
3084 #endif /* CONFIG_PCI_IOV */
3085
3086 static void pnv_ioda_setup_pe_res(struct pnv_ioda_pe *pe,
3087                                   struct resource *res)
3088 {
3089         struct pnv_phb *phb = pe->phb;
3090         struct pci_bus_region region;
3091         int index;
3092         int64_t rc;
3093
3094         if (!res || !res->flags || res->start > res->end)
3095                 return;
3096
3097         if (res->flags & IORESOURCE_IO) {
3098                 region.start = res->start - phb->ioda.io_pci_base;
3099                 region.end   = res->end - phb->ioda.io_pci_base;
3100                 index = region.start / phb->ioda.io_segsize;
3101
3102                 while (index < phb->ioda.total_pe_num &&
3103                        region.start <= region.end) {
3104                         phb->ioda.io_segmap[index] = pe->pe_number;
3105                         rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3106                                 pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index);
3107                         if (rc != OPAL_SUCCESS) {
3108                                 pr_err("%s: Error %lld mapping IO segment#%d to PE#%x\n",
3109                                        __func__, rc, index, pe->pe_number);
3110                                 break;
3111                         }
3112
3113                         region.start += phb->ioda.io_segsize;
3114                         index++;
3115                 }
3116         } else if ((res->flags & IORESOURCE_MEM) &&
3117                    !pnv_pci_is_m64(phb, res)) {
3118                 region.start = res->start -
3119                                phb->hose->mem_offset[0] -
3120                                phb->ioda.m32_pci_base;
3121                 region.end   = res->end -
3122                                phb->hose->mem_offset[0] -
3123                                phb->ioda.m32_pci_base;
3124                 index = region.start / phb->ioda.m32_segsize;
3125
3126                 while (index < phb->ioda.total_pe_num &&
3127                        region.start <= region.end) {
3128                         phb->ioda.m32_segmap[index] = pe->pe_number;
3129                         rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3130                                 pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index);
3131                         if (rc != OPAL_SUCCESS) {
3132                                 pr_err("%s: Error %lld mapping M32 segment#%d to PE#%x",
3133                                        __func__, rc, index, pe->pe_number);
3134                                 break;
3135                         }
3136
3137                         region.start += phb->ioda.m32_segsize;
3138                         index++;
3139                 }
3140         }
3141 }
3142
3143 /*
3144  * This function is supposed to be called on basis of PE from top
3145  * to bottom style. So the the I/O or MMIO segment assigned to
3146  * parent PE could be overridden by its child PEs if necessary.
3147  */
3148 static void pnv_ioda_setup_pe_seg(struct pnv_ioda_pe *pe)
3149 {
3150         struct pci_dev *pdev;
3151         int i;
3152
3153         /*
3154          * NOTE: We only care PCI bus based PE for now. For PCI
3155          * device based PE, for example SRIOV sensitive VF should
3156          * be figured out later.
3157          */
3158         BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)));
3159
3160         list_for_each_entry(pdev, &pe->pbus->devices, bus_list) {
3161                 for (i = 0; i <= PCI_ROM_RESOURCE; i++)
3162                         pnv_ioda_setup_pe_res(pe, &pdev->resource[i]);
3163
3164                 /*
3165                  * If the PE contains all subordinate PCI buses, the
3166                  * windows of the child bridges should be mapped to
3167                  * the PE as well.
3168                  */
3169                 if (!(pe->flags & PNV_IODA_PE_BUS_ALL) || !pci_is_bridge(pdev))
3170                         continue;
3171                 for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++)
3172                         pnv_ioda_setup_pe_res(pe,
3173                                 &pdev->resource[PCI_BRIDGE_RESOURCES + i]);
3174         }
3175 }
3176
3177 #ifdef CONFIG_DEBUG_FS
3178 static int pnv_pci_diag_data_set(void *data, u64 val)
3179 {
3180         struct pci_controller *hose;
3181         struct pnv_phb *phb;
3182         s64 ret;
3183
3184         if (val != 1ULL)
3185                 return -EINVAL;
3186
3187         hose = (struct pci_controller *)data;
3188         if (!hose || !hose->private_data)
3189                 return -ENODEV;
3190
3191         phb = hose->private_data;
3192
3193         /* Retrieve the diag data from firmware */
3194         ret = opal_pci_get_phb_diag_data2(phb->opal_id, phb->diag_data,
3195                                           phb->diag_data_size);
3196         if (ret != OPAL_SUCCESS)
3197                 return -EIO;
3198
3199         /* Print the diag data to the kernel log */
3200         pnv_pci_dump_phb_diag_data(phb->hose, phb->diag_data);
3201         return 0;
3202 }
3203
3204 DEFINE_SIMPLE_ATTRIBUTE(pnv_pci_diag_data_fops, NULL,
3205                         pnv_pci_diag_data_set, "%llu\n");
3206
3207 #endif /* CONFIG_DEBUG_FS */
3208
3209 static void pnv_pci_ioda_create_dbgfs(void)
3210 {
3211 #ifdef CONFIG_DEBUG_FS
3212         struct pci_controller *hose, *tmp;
3213         struct pnv_phb *phb;
3214         char name[16];
3215
3216         list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
3217                 phb = hose->private_data;
3218
3219                 /* Notify initialization of PHB done */
3220                 phb->initialized = 1;
3221
3222                 sprintf(name, "PCI%04x", hose->global_number);
3223                 phb->dbgfs = debugfs_create_dir(name, powerpc_debugfs_root);
3224                 if (!phb->dbgfs) {
3225                         pr_warn("%s: Error on creating debugfs on PHB#%x\n",
3226                                 __func__, hose->global_number);
3227                         continue;
3228                 }
3229
3230                 debugfs_create_file("dump_diag_regs", 0200, phb->dbgfs, hose,
3231                                     &pnv_pci_diag_data_fops);
3232         }
3233 #endif /* CONFIG_DEBUG_FS */
3234 }
3235
3236 static void pnv_pci_enable_bridge(struct pci_bus *bus)
3237 {
3238         struct pci_dev *dev = bus->self;
3239         struct pci_bus *child;
3240
3241         /* Empty bus ? bail */
3242         if (list_empty(&bus->devices))
3243                 return;
3244
3245         /*
3246          * If there's a bridge associated with that bus enable it. This works
3247          * around races in the generic code if the enabling is done during
3248          * parallel probing. This can be removed once those races have been
3249          * fixed.
3250          */
3251         if (dev) {
3252                 int rc = pci_enable_device(dev);
3253                 if (rc)
3254                         pci_err(dev, "Error enabling bridge (%d)\n", rc);
3255                 pci_set_master(dev);
3256         }
3257
3258         /* Perform the same to child busses */
3259         list_for_each_entry(child, &bus->children, node)
3260                 pnv_pci_enable_bridge(child);
3261 }
3262
3263 static void pnv_pci_enable_bridges(void)
3264 {
3265         struct pci_controller *hose;
3266
3267         list_for_each_entry(hose, &hose_list, list_node)
3268                 pnv_pci_enable_bridge(hose->bus);
3269 }
3270
3271 static void pnv_pci_ioda_fixup(void)
3272 {
3273         pnv_pci_ioda_setup_PEs();
3274         pnv_pci_ioda_setup_iommu_api();
3275         pnv_pci_ioda_create_dbgfs();
3276
3277         pnv_pci_enable_bridges();
3278
3279 #ifdef CONFIG_EEH
3280         pnv_eeh_post_init();
3281 #endif
3282 }
3283
3284 /*
3285  * Returns the alignment for I/O or memory windows for P2P
3286  * bridges. That actually depends on how PEs are segmented.
3287  * For now, we return I/O or M32 segment size for PE sensitive
3288  * P2P bridges. Otherwise, the default values (4KiB for I/O,
3289  * 1MiB for memory) will be returned.
3290  *
3291  * The current PCI bus might be put into one PE, which was
3292  * create against the parent PCI bridge. For that case, we
3293  * needn't enlarge the alignment so that we can save some
3294  * resources.
3295  */
3296 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus,
3297                                                 unsigned long type)
3298 {
3299         struct pci_dev *bridge;
3300         struct pci_controller *hose = pci_bus_to_host(bus);
3301         struct pnv_phb *phb = hose->private_data;
3302         int num_pci_bridges = 0;
3303
3304         bridge = bus->self;
3305         while (bridge) {
3306                 if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) {
3307                         num_pci_bridges++;
3308                         if (num_pci_bridges >= 2)
3309                                 return 1;
3310                 }
3311
3312                 bridge = bridge->bus->self;
3313         }
3314
3315         /*
3316          * We fall back to M32 if M64 isn't supported. We enforce the M64
3317          * alignment for any 64-bit resource, PCIe doesn't care and
3318          * bridges only do 64-bit prefetchable anyway.
3319          */
3320         if (phb->ioda.m64_segsize && pnv_pci_is_m64_flags(type))
3321                 return phb->ioda.m64_segsize;
3322         if (type & IORESOURCE_MEM)
3323                 return phb->ioda.m32_segsize;
3324
3325         return phb->ioda.io_segsize;
3326 }
3327
3328 /*
3329  * We are updating root port or the upstream port of the
3330  * bridge behind the root port with PHB's windows in order
3331  * to accommodate the changes on required resources during
3332  * PCI (slot) hotplug, which is connected to either root
3333  * port or the downstream ports of PCIe switch behind the
3334  * root port.
3335  */
3336 static void pnv_pci_fixup_bridge_resources(struct pci_bus *bus,
3337                                            unsigned long type)
3338 {
3339         struct pci_controller *hose = pci_bus_to_host(bus);
3340         struct pnv_phb *phb = hose->private_data;
3341         struct pci_dev *bridge = bus->self;
3342         struct resource *r, *w;
3343         bool msi_region = false;
3344         int i;
3345
3346         /* Check if we need apply fixup to the bridge's windows */
3347         if (!pci_is_root_bus(bridge->bus) &&
3348             !pci_is_root_bus(bridge->bus->self->bus))
3349                 return;
3350
3351         /* Fixup the resources */
3352         for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++) {
3353                 r = &bridge->resource[PCI_BRIDGE_RESOURCES + i];
3354                 if (!r->flags || !r->parent)
3355                         continue;
3356
3357                 w = NULL;
3358                 if (r->flags & type & IORESOURCE_IO)
3359                         w = &hose->io_resource;
3360                 else if (pnv_pci_is_m64(phb, r) &&
3361                          (type & IORESOURCE_PREFETCH) &&
3362                          phb->ioda.m64_segsize)
3363                         w = &hose->mem_resources[1];
3364                 else if (r->flags & type & IORESOURCE_MEM) {
3365                         w = &hose->mem_resources[0];
3366                         msi_region = true;
3367                 }
3368
3369                 r->start = w->start;
3370                 r->end = w->end;
3371
3372                 /* The 64KB 32-bits MSI region shouldn't be included in
3373                  * the 32-bits bridge window. Otherwise, we can see strange
3374                  * issues. One of them is EEH error observed on Garrison.
3375                  *
3376                  * Exclude top 1MB region which is the minimal alignment of
3377                  * 32-bits bridge window.
3378                  */
3379                 if (msi_region) {
3380                         r->end += 0x10000;
3381                         r->end -= 0x100000;
3382                 }
3383         }
3384 }
3385
3386 static void pnv_pci_setup_bridge(struct pci_bus *bus, unsigned long type)
3387 {
3388         struct pci_controller *hose = pci_bus_to_host(bus);
3389         struct pnv_phb *phb = hose->private_data;
3390         struct pci_dev *bridge = bus->self;
3391         struct pnv_ioda_pe *pe;
3392         bool all = (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE);
3393
3394         /* Extend bridge's windows if necessary */
3395         pnv_pci_fixup_bridge_resources(bus, type);
3396
3397         /* The PE for root bus should be realized before any one else */
3398         if (!phb->ioda.root_pe_populated) {
3399                 pe = pnv_ioda_setup_bus_PE(phb->hose->bus, false);
3400                 if (pe) {
3401                         phb->ioda.root_pe_idx = pe->pe_number;
3402                         phb->ioda.root_pe_populated = true;
3403                 }
3404         }
3405
3406         /* Don't assign PE to PCI bus, which doesn't have subordinate devices */
3407         if (list_empty(&bus->devices))
3408                 return;
3409
3410         /* Reserve PEs according to used M64 resources */
3411         if (phb->reserve_m64_pe)
3412                 phb->reserve_m64_pe(bus, NULL, all);
3413
3414         /*
3415          * Assign PE. We might run here because of partial hotplug.
3416          * For the case, we just pick up the existing PE and should
3417          * not allocate resources again.
3418          */
3419         pe = pnv_ioda_setup_bus_PE(bus, all);
3420         if (!pe)
3421                 return;
3422
3423         pnv_ioda_setup_pe_seg(pe);
3424         switch (phb->type) {
3425         case PNV_PHB_IODA1:
3426                 pnv_pci_ioda1_setup_dma_pe(phb, pe);
3427                 break;
3428         case PNV_PHB_IODA2:
3429                 pnv_pci_ioda2_setup_dma_pe(phb, pe);
3430                 break;
3431         default:
3432                 pr_warn("%s: No DMA for PHB#%x (type %d)\n",
3433                         __func__, phb->hose->global_number, phb->type);
3434         }
3435 }
3436
3437 static resource_size_t pnv_pci_default_alignment(void)
3438 {
3439         return PAGE_SIZE;
3440 }
3441
3442 #ifdef CONFIG_PCI_IOV
3443 static resource_size_t pnv_pci_iov_resource_alignment(struct pci_dev *pdev,
3444                                                       int resno)
3445 {
3446         struct pci_controller *hose = pci_bus_to_host(pdev->bus);
3447         struct pnv_phb *phb = hose->private_data;
3448         struct pci_dn *pdn = pci_get_pdn(pdev);
3449         resource_size_t align;
3450
3451         /*
3452          * On PowerNV platform, IOV BAR is mapped by M64 BAR to enable the
3453          * SR-IOV. While from hardware perspective, the range mapped by M64
3454          * BAR should be size aligned.
3455          *
3456          * When IOV BAR is mapped with M64 BAR in Single PE mode, the extra
3457          * powernv-specific hardware restriction is gone. But if just use the
3458          * VF BAR size as the alignment, PF BAR / VF BAR may be allocated with
3459          * in one segment of M64 #15, which introduces the PE conflict between
3460          * PF and VF. Based on this, the minimum alignment of an IOV BAR is
3461          * m64_segsize.
3462          *
3463          * This function returns the total IOV BAR size if M64 BAR is in
3464          * Shared PE mode or just VF BAR size if not.
3465          * If the M64 BAR is in Single PE mode, return the VF BAR size or
3466          * M64 segment size if IOV BAR size is less.
3467          */
3468         align = pci_iov_resource_size(pdev, resno);
3469         if (!pdn->vfs_expanded)
3470                 return align;
3471         if (pdn->m64_single_mode)
3472                 return max(align, (resource_size_t)phb->ioda.m64_segsize);
3473
3474         return pdn->vfs_expanded * align;
3475 }
3476 #endif /* CONFIG_PCI_IOV */
3477
3478 /* Prevent enabling devices for which we couldn't properly
3479  * assign a PE
3480  */
3481 static bool pnv_pci_enable_device_hook(struct pci_dev *dev)
3482 {
3483         struct pci_controller *hose = pci_bus_to_host(dev->bus);
3484         struct pnv_phb *phb = hose->private_data;
3485         struct pci_dn *pdn;
3486
3487         /* The function is probably called while the PEs have
3488          * not be created yet. For example, resource reassignment
3489          * during PCI probe period. We just skip the check if
3490          * PEs isn't ready.
3491          */
3492         if (!phb->initialized)
3493                 return true;
3494
3495         pdn = pci_get_pdn(dev);
3496         if (!pdn || pdn->pe_number == IODA_INVALID_PE)
3497                 return false;
3498
3499         return true;
3500 }
3501
3502 static long pnv_pci_ioda1_unset_window(struct iommu_table_group *table_group,
3503                                        int num)
3504 {
3505         struct pnv_ioda_pe *pe = container_of(table_group,
3506                                               struct pnv_ioda_pe, table_group);
3507         struct pnv_phb *phb = pe->phb;
3508         unsigned int idx;
3509         long rc;
3510
3511         pe_info(pe, "Removing DMA window #%d\n", num);
3512         for (idx = 0; idx < phb->ioda.dma32_count; idx++) {
3513                 if (phb->ioda.dma32_segmap[idx] != pe->pe_number)
3514                         continue;
3515
3516                 rc = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
3517                                                 idx, 0, 0ul, 0ul, 0ul);
3518                 if (rc != OPAL_SUCCESS) {
3519                         pe_warn(pe, "Failure %ld unmapping DMA32 segment#%d\n",
3520                                 rc, idx);
3521                         return rc;
3522                 }
3523
3524                 phb->ioda.dma32_segmap[idx] = IODA_INVALID_PE;
3525         }
3526
3527         pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
3528         return OPAL_SUCCESS;
3529 }
3530
3531 static void pnv_pci_ioda1_release_pe_dma(struct pnv_ioda_pe *pe)
3532 {
3533         unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe);
3534         struct iommu_table *tbl = pe->table_group.tables[0];
3535         int64_t rc;
3536
3537         if (!weight)
3538                 return;
3539
3540         rc = pnv_pci_ioda1_unset_window(&pe->table_group, 0);
3541         if (rc != OPAL_SUCCESS)
3542                 return;
3543
3544         pnv_pci_p7ioc_tce_invalidate(tbl, tbl->it_offset, tbl->it_size, false);
3545         if (pe->table_group.group) {
3546                 iommu_group_put(pe->table_group.group);
3547                 WARN_ON(pe->table_group.group);
3548         }
3549
3550         free_pages(tbl->it_base, get_order(tbl->it_size << 3));
3551         iommu_tce_table_put(tbl);
3552 }
3553
3554 static void pnv_pci_ioda2_release_pe_dma(struct pnv_ioda_pe *pe)
3555 {
3556         struct iommu_table *tbl = pe->table_group.tables[0];
3557         unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe);
3558 #ifdef CONFIG_IOMMU_API
3559         int64_t rc;
3560 #endif
3561
3562         if (!weight)
3563                 return;
3564
3565 #ifdef CONFIG_IOMMU_API
3566         rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
3567         if (rc)
3568                 pe_warn(pe, "OPAL error %ld release DMA window\n", rc);
3569 #endif
3570
3571         pnv_pci_ioda2_set_bypass(pe, false);
3572         if (pe->table_group.group) {
3573                 iommu_group_put(pe->table_group.group);
3574                 WARN_ON(pe->table_group.group);
3575         }
3576
3577         iommu_tce_table_put(tbl);
3578 }
3579
3580 static void pnv_ioda_free_pe_seg(struct pnv_ioda_pe *pe,
3581                                  unsigned short win,
3582                                  unsigned int *map)
3583 {
3584         struct pnv_phb *phb = pe->phb;
3585         int idx;
3586         int64_t rc;
3587
3588         for (idx = 0; idx < phb->ioda.total_pe_num; idx++) {
3589                 if (map[idx] != pe->pe_number)
3590                         continue;
3591
3592                 if (win == OPAL_M64_WINDOW_TYPE)
3593                         rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3594                                         phb->ioda.reserved_pe_idx, win,
3595                                         idx / PNV_IODA1_M64_SEGS,
3596                                         idx % PNV_IODA1_M64_SEGS);
3597                 else
3598                         rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3599                                         phb->ioda.reserved_pe_idx, win, 0, idx);
3600
3601                 if (rc != OPAL_SUCCESS)
3602                         pe_warn(pe, "Error %ld unmapping (%d) segment#%d\n",
3603                                 rc, win, idx);
3604
3605                 map[idx] = IODA_INVALID_PE;
3606         }
3607 }
3608
3609 static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe *pe)
3610 {
3611         struct pnv_phb *phb = pe->phb;
3612
3613         if (phb->type == PNV_PHB_IODA1) {
3614                 pnv_ioda_free_pe_seg(pe, OPAL_IO_WINDOW_TYPE,
3615                                      phb->ioda.io_segmap);
3616                 pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
3617                                      phb->ioda.m32_segmap);
3618                 pnv_ioda_free_pe_seg(pe, OPAL_M64_WINDOW_TYPE,
3619                                      phb->ioda.m64_segmap);
3620         } else if (phb->type == PNV_PHB_IODA2) {
3621                 pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
3622                                      phb->ioda.m32_segmap);
3623         }
3624 }
3625
3626 static void pnv_ioda_release_pe(struct pnv_ioda_pe *pe)
3627 {
3628         struct pnv_phb *phb = pe->phb;
3629         struct pnv_ioda_pe *slave, *tmp;
3630
3631         list_del(&pe->list);
3632         switch (phb->type) {
3633         case PNV_PHB_IODA1:
3634                 pnv_pci_ioda1_release_pe_dma(pe);
3635                 break;
3636         case PNV_PHB_IODA2:
3637                 pnv_pci_ioda2_release_pe_dma(pe);
3638                 break;
3639         default:
3640                 WARN_ON(1);
3641         }
3642
3643         pnv_ioda_release_pe_seg(pe);
3644         pnv_ioda_deconfigure_pe(pe->phb, pe);
3645
3646         /* Release slave PEs in the compound PE */
3647         if (pe->flags & PNV_IODA_PE_MASTER) {
3648                 list_for_each_entry_safe(slave, tmp, &pe->slaves, list) {
3649                         list_del(&slave->list);
3650                         pnv_ioda_free_pe(slave);
3651                 }
3652         }
3653
3654         /*
3655          * The PE for root bus can be removed because of hotplug in EEH
3656          * recovery for fenced PHB error. We need to mark the PE dead so
3657          * that it can be populated again in PCI hot add path. The PE
3658          * shouldn't be destroyed as it's the global reserved resource.
3659          */
3660         if (phb->ioda.root_pe_populated &&
3661             phb->ioda.root_pe_idx == pe->pe_number)
3662                 phb->ioda.root_pe_populated = false;
3663         else
3664                 pnv_ioda_free_pe(pe);
3665 }
3666
3667 static void pnv_pci_release_device(struct pci_dev *pdev)
3668 {
3669         struct pci_controller *hose = pci_bus_to_host(pdev->bus);
3670         struct pnv_phb *phb = hose->private_data;
3671         struct pci_dn *pdn = pci_get_pdn(pdev);
3672         struct pnv_ioda_pe *pe;
3673
3674         if (pdev->is_virtfn)
3675                 return;
3676
3677         if (!pdn || pdn->pe_number == IODA_INVALID_PE)
3678                 return;
3679
3680         /*
3681          * PCI hotplug can happen as part of EEH error recovery. The @pdn
3682          * isn't removed and added afterwards in this scenario. We should
3683          * set the PE number in @pdn to an invalid one. Otherwise, the PE's
3684          * device count is decreased on removing devices while failing to
3685          * be increased on adding devices. It leads to unbalanced PE's device
3686          * count and eventually make normal PCI hotplug path broken.
3687          */
3688         pe = &phb->ioda.pe_array[pdn->pe_number];
3689         pdn->pe_number = IODA_INVALID_PE;
3690
3691         WARN_ON(--pe->device_count < 0);
3692         if (pe->device_count == 0)
3693                 pnv_ioda_release_pe(pe);
3694 }
3695
3696 static void pnv_pci_ioda_shutdown(struct pci_controller *hose)
3697 {
3698         struct pnv_phb *phb = hose->private_data;
3699
3700         opal_pci_reset(phb->opal_id, OPAL_RESET_PCI_IODA_TABLE,
3701                        OPAL_ASSERT_RESET);
3702 }
3703
3704 static const struct pci_controller_ops pnv_pci_ioda_controller_ops = {
3705         .dma_dev_setup          = pnv_pci_dma_dev_setup,
3706         .dma_bus_setup          = pnv_pci_dma_bus_setup,
3707 #ifdef CONFIG_PCI_MSI
3708         .setup_msi_irqs         = pnv_setup_msi_irqs,
3709         .teardown_msi_irqs      = pnv_teardown_msi_irqs,
3710 #endif
3711         .enable_device_hook     = pnv_pci_enable_device_hook,
3712         .release_device         = pnv_pci_release_device,
3713         .window_alignment       = pnv_pci_window_alignment,
3714         .setup_bridge           = pnv_pci_setup_bridge,
3715         .reset_secondary_bus    = pnv_pci_reset_secondary_bus,
3716         .dma_set_mask           = pnv_pci_ioda_dma_set_mask,
3717         .dma_get_required_mask  = pnv_pci_ioda_dma_get_required_mask,
3718         .shutdown               = pnv_pci_ioda_shutdown,
3719 };
3720
3721 static int pnv_npu_dma_set_mask(struct pci_dev *npdev, u64 dma_mask)
3722 {
3723         dev_err_once(&npdev->dev,
3724                         "%s operation unsupported for NVLink devices\n",
3725                         __func__);
3726         return -EPERM;
3727 }
3728
3729 static const struct pci_controller_ops pnv_npu_ioda_controller_ops = {
3730         .dma_dev_setup          = pnv_pci_dma_dev_setup,
3731 #ifdef CONFIG_PCI_MSI
3732         .setup_msi_irqs         = pnv_setup_msi_irqs,
3733         .teardown_msi_irqs      = pnv_teardown_msi_irqs,
3734 #endif
3735         .enable_device_hook     = pnv_pci_enable_device_hook,
3736         .window_alignment       = pnv_pci_window_alignment,
3737         .reset_secondary_bus    = pnv_pci_reset_secondary_bus,
3738         .dma_set_mask           = pnv_npu_dma_set_mask,
3739         .shutdown               = pnv_pci_ioda_shutdown,
3740 };
3741
3742 static const struct pci_controller_ops pnv_npu_ocapi_ioda_controller_ops = {
3743         .enable_device_hook     = pnv_pci_enable_device_hook,
3744         .window_alignment       = pnv_pci_window_alignment,
3745         .reset_secondary_bus    = pnv_pci_reset_secondary_bus,
3746         .shutdown               = pnv_pci_ioda_shutdown,
3747 };
3748
3749 static void __init pnv_pci_init_ioda_phb(struct device_node *np,
3750                                          u64 hub_id, int ioda_type)
3751 {
3752         struct pci_controller *hose;
3753         struct pnv_phb *phb;
3754         unsigned long size, m64map_off, m32map_off, pemap_off;
3755         unsigned long iomap_off = 0, dma32map_off = 0;
3756         struct resource r;
3757         const __be64 *prop64;
3758         const __be32 *prop32;
3759         int len;
3760         unsigned int segno;
3761         u64 phb_id;
3762         void *aux;
3763         long rc;
3764
3765         if (!of_device_is_available(np))
3766                 return;
3767
3768         pr_info("Initializing %s PHB (%pOF)\n", pnv_phb_names[ioda_type], np);
3769
3770         prop64 = of_get_property(np, "ibm,opal-phbid", NULL);
3771         if (!prop64) {
3772                 pr_err("  Missing \"ibm,opal-phbid\" property !\n");
3773                 return;
3774         }
3775         phb_id = be64_to_cpup(prop64);
3776         pr_debug("  PHB-ID  : 0x%016llx\n", phb_id);
3777
3778         phb = memblock_virt_alloc(sizeof(*phb), 0);
3779
3780         /* Allocate PCI controller */
3781         phb->hose = hose = pcibios_alloc_controller(np);
3782         if (!phb->hose) {
3783                 pr_err("  Can't allocate PCI controller for %pOF\n",
3784                        np);
3785                 memblock_free(__pa(phb), sizeof(struct pnv_phb));
3786                 return;
3787         }
3788
3789         spin_lock_init(&phb->lock);
3790         prop32 = of_get_property(np, "bus-range", &len);
3791         if (prop32 && len == 8) {
3792                 hose->first_busno = be32_to_cpu(prop32[0]);
3793                 hose->last_busno = be32_to_cpu(prop32[1]);
3794         } else {
3795                 pr_warn("  Broken <bus-range> on %pOF\n", np);
3796                 hose->first_busno = 0;
3797                 hose->last_busno = 0xff;
3798         }
3799         hose->private_data = phb;
3800         phb->hub_id = hub_id;
3801         phb->opal_id = phb_id;
3802         phb->type = ioda_type;
3803         mutex_init(&phb->ioda.pe_alloc_mutex);
3804
3805         /* Detect specific models for error handling */
3806         if (of_device_is_compatible(np, "ibm,p7ioc-pciex"))
3807                 phb->model = PNV_PHB_MODEL_P7IOC;
3808         else if (of_device_is_compatible(np, "ibm,power8-pciex"))
3809                 phb->model = PNV_PHB_MODEL_PHB3;
3810         else if (of_device_is_compatible(np, "ibm,power8-npu-pciex"))
3811                 phb->model = PNV_PHB_MODEL_NPU;
3812         else if (of_device_is_compatible(np, "ibm,power9-npu-pciex"))
3813                 phb->model = PNV_PHB_MODEL_NPU2;
3814         else
3815                 phb->model = PNV_PHB_MODEL_UNKNOWN;
3816
3817         /* Initialize diagnostic data buffer */
3818         prop32 = of_get_property(np, "ibm,phb-diag-data-size", NULL);
3819         if (prop32)
3820                 phb->diag_data_size = be32_to_cpup(prop32);
3821         else
3822                 phb->diag_data_size = PNV_PCI_DIAG_BUF_SIZE;
3823
3824         phb->diag_data = memblock_virt_alloc(phb->diag_data_size, 0);
3825
3826         /* Parse 32-bit and IO ranges (if any) */
3827         pci_process_bridge_OF_ranges(hose, np, !hose->global_number);
3828
3829         /* Get registers */
3830         if (!of_address_to_resource(np, 0, &r)) {
3831                 phb->regs_phys = r.start;
3832                 phb->regs = ioremap(r.start, resource_size(&r));
3833                 if (phb->regs == NULL)
3834                         pr_err("  Failed to map registers !\n");
3835         }
3836
3837         /* Initialize more IODA stuff */
3838         phb->ioda.total_pe_num = 1;
3839         prop32 = of_get_property(np, "ibm,opal-num-pes", NULL);
3840         if (prop32)
3841                 phb->ioda.total_pe_num = be32_to_cpup(prop32);
3842         prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL);
3843         if (prop32)
3844                 phb->ioda.reserved_pe_idx = be32_to_cpup(prop32);
3845
3846         /* Invalidate RID to PE# mapping */
3847         for (segno = 0; segno < ARRAY_SIZE(phb->ioda.pe_rmap); segno++)
3848                 phb->ioda.pe_rmap[segno] = IODA_INVALID_PE;
3849
3850         /* Parse 64-bit MMIO range */
3851         pnv_ioda_parse_m64_window(phb);
3852
3853         phb->ioda.m32_size = resource_size(&hose->mem_resources[0]);
3854         /* FW Has already off top 64k of M32 space (MSI space) */
3855         phb->ioda.m32_size += 0x10000;
3856
3857         phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe_num;
3858         phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0];
3859         phb->ioda.io_size = hose->pci_io_size;
3860         phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe_num;
3861         phb->ioda.io_pci_base = 0; /* XXX calculate this ? */
3862
3863         /* Calculate how many 32-bit TCE segments we have */
3864         phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3865                                 PNV_IODA1_DMA32_SEGSIZE;
3866
3867         /* Allocate aux data & arrays. We don't have IO ports on PHB3 */
3868         size = _ALIGN_UP(max_t(unsigned, phb->ioda.total_pe_num, 8) / 8,
3869                         sizeof(unsigned long));
3870         m64map_off = size;
3871         size += phb->ioda.total_pe_num * sizeof(phb->ioda.m64_segmap[0]);
3872         m32map_off = size;
3873         size += phb->ioda.total_pe_num * sizeof(phb->ioda.m32_segmap[0]);
3874         if (phb->type == PNV_PHB_IODA1) {
3875                 iomap_off = size;
3876                 size += phb->ioda.total_pe_num * sizeof(phb->ioda.io_segmap[0]);
3877                 dma32map_off = size;
3878                 size += phb->ioda.dma32_count *
3879                         sizeof(phb->ioda.dma32_segmap[0]);
3880         }
3881         pemap_off = size;
3882         size += phb->ioda.total_pe_num * sizeof(struct pnv_ioda_pe);
3883         aux = memblock_virt_alloc(size, 0);
3884         phb->ioda.pe_alloc = aux;
3885         phb->ioda.m64_segmap = aux + m64map_off;
3886         phb->ioda.m32_segmap = aux + m32map_off;
3887         for (segno = 0; segno < phb->ioda.total_pe_num; segno++) {
3888                 phb->ioda.m64_segmap[segno] = IODA_INVALID_PE;
3889                 phb->ioda.m32_segmap[segno] = IODA_INVALID_PE;
3890         }
3891         if (phb->type == PNV_PHB_IODA1) {
3892                 phb->ioda.io_segmap = aux + iomap_off;
3893                 for (segno = 0; segno < phb->ioda.total_pe_num; segno++)
3894                         phb->ioda.io_segmap[segno] = IODA_INVALID_PE;
3895
3896                 phb->ioda.dma32_segmap = aux + dma32map_off;
3897                 for (segno = 0; segno < phb->ioda.dma32_count; segno++)
3898                         phb->ioda.dma32_segmap[segno] = IODA_INVALID_PE;
3899         }
3900         phb->ioda.pe_array = aux + pemap_off;
3901
3902         /*
3903          * Choose PE number for root bus, which shouldn't have
3904          * M64 resources consumed by its child devices. To pick
3905          * the PE number adjacent to the reserved one if possible.
3906          */
3907         pnv_ioda_reserve_pe(phb, phb->ioda.reserved_pe_idx);
3908         if (phb->ioda.reserved_pe_idx == 0) {
3909                 phb->ioda.root_pe_idx = 1;
3910                 pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3911         } else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1)) {
3912                 phb->ioda.root_pe_idx = phb->ioda.reserved_pe_idx - 1;
3913                 pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3914         } else {
3915                 phb->ioda.root_pe_idx = IODA_INVALID_PE;
3916         }
3917
3918         INIT_LIST_HEAD(&phb->ioda.pe_list);
3919         mutex_init(&phb->ioda.pe_list_mutex);
3920
3921         /* Calculate how many 32-bit TCE segments we have */
3922         phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3923                                 PNV_IODA1_DMA32_SEGSIZE;
3924
3925 #if 0 /* We should really do that ... */
3926         rc = opal_pci_set_phb_mem_window(opal->phb_id,
3927                                          window_type,
3928                                          window_num,
3929                                          starting_real_address,
3930                                          starting_pci_address,
3931                                          segment_size);
3932 #endif
3933
3934         pr_info("  %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
3935                 phb->ioda.total_pe_num, phb->ioda.reserved_pe_idx,
3936                 phb->ioda.m32_size, phb->ioda.m32_segsize);
3937         if (phb->ioda.m64_size)
3938                 pr_info("                 M64: 0x%lx [segment=0x%lx]\n",
3939                         phb->ioda.m64_size, phb->ioda.m64_segsize);
3940         if (phb->ioda.io_size)
3941                 pr_info("                  IO: 0x%x [segment=0x%x]\n",
3942                         phb->ioda.io_size, phb->ioda.io_segsize);
3943
3944
3945         phb->hose->ops = &pnv_pci_ops;
3946         phb->get_pe_state = pnv_ioda_get_pe_state;
3947         phb->freeze_pe = pnv_ioda_freeze_pe;
3948         phb->unfreeze_pe = pnv_ioda_unfreeze_pe;
3949
3950         /* Setup MSI support */
3951         pnv_pci_init_ioda_msis(phb);
3952
3953         /*
3954          * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
3955          * to let the PCI core do resource assignment. It's supposed
3956          * that the PCI core will do correct I/O and MMIO alignment
3957          * for the P2P bridge bars so that each PCI bus (excluding
3958          * the child P2P bridges) can form individual PE.
3959          */
3960         ppc_md.pcibios_fixup = pnv_pci_ioda_fixup;
3961
3962         switch (phb->type) {
3963         case PNV_PHB_NPU_NVLINK:
3964                 hose->controller_ops = pnv_npu_ioda_controller_ops;
3965                 break;
3966         case PNV_PHB_NPU_OCAPI:
3967                 hose->controller_ops = pnv_npu_ocapi_ioda_controller_ops;
3968                 break;
3969         default:
3970                 phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup;
3971                 hose->controller_ops = pnv_pci_ioda_controller_ops;
3972         }
3973
3974         ppc_md.pcibios_default_alignment = pnv_pci_default_alignment;
3975
3976 #ifdef CONFIG_PCI_IOV
3977         ppc_md.pcibios_fixup_sriov = pnv_pci_ioda_fixup_iov_resources;
3978         ppc_md.pcibios_iov_resource_alignment = pnv_pci_iov_resource_alignment;
3979         ppc_md.pcibios_sriov_enable = pnv_pcibios_sriov_enable;
3980         ppc_md.pcibios_sriov_disable = pnv_pcibios_sriov_disable;
3981 #endif
3982
3983         pci_add_flags(PCI_REASSIGN_ALL_RSRC);
3984
3985         /* Reset IODA tables to a clean state */
3986         rc = opal_pci_reset(phb_id, OPAL_RESET_PCI_IODA_TABLE, OPAL_ASSERT_RESET);
3987         if (rc)
3988                 pr_warn("  OPAL Error %ld performing IODA table reset !\n", rc);
3989
3990         /*
3991          * If we're running in kdump kernel, the previous kernel never
3992          * shutdown PCI devices correctly. We already got IODA table
3993          * cleaned out. So we have to issue PHB reset to stop all PCI
3994          * transactions from previous kernel. The ppc_pci_reset_phbs
3995          * kernel parameter will force this reset too.
3996          */
3997         if (is_kdump_kernel() || pci_reset_phbs) {
3998                 pr_info("  Issue PHB reset ...\n");
3999                 pnv_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL);
4000                 pnv_eeh_phb_reset(hose, EEH_RESET_DEACTIVATE);
4001         }
4002
4003         /* Remove M64 resource if we can't configure it successfully */
4004         if (!phb->init_m64 || phb->init_m64(phb))
4005                 hose->mem_resources[1].flags = 0;
4006 }
4007
4008 void __init pnv_pci_init_ioda2_phb(struct device_node *np)
4009 {
4010         pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2);
4011 }
4012
4013 void __init pnv_pci_init_npu_phb(struct device_node *np)
4014 {
4015         pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU_NVLINK);
4016 }
4017
4018 void __init pnv_pci_init_npu2_opencapi_phb(struct device_node *np)
4019 {
4020         pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU_OCAPI);
4021 }
4022
4023 static void pnv_npu2_opencapi_cfg_size_fixup(struct pci_dev *dev)
4024 {
4025         struct pci_controller *hose = pci_bus_to_host(dev->bus);
4026         struct pnv_phb *phb = hose->private_data;
4027
4028         if (!machine_is(powernv))
4029                 return;
4030
4031         if (phb->type == PNV_PHB_NPU_OCAPI)
4032                 dev->cfg_size = PCI_CFG_SPACE_EXP_SIZE;
4033 }
4034 DECLARE_PCI_FIXUP_EARLY(PCI_ANY_ID, PCI_ANY_ID, pnv_npu2_opencapi_cfg_size_fixup);
4035
4036 void __init pnv_pci_init_ioda_hub(struct device_node *np)
4037 {
4038         struct device_node *phbn;
4039         const __be64 *prop64;
4040         u64 hub_id;
4041
4042         pr_info("Probing IODA IO-Hub %pOF\n", np);
4043
4044         prop64 = of_get_property(np, "ibm,opal-hubid", NULL);
4045         if (!prop64) {
4046                 pr_err(" Missing \"ibm,opal-hubid\" property !\n");
4047                 return;
4048         }
4049         hub_id = be64_to_cpup(prop64);
4050         pr_devel(" HUB-ID : 0x%016llx\n", hub_id);
4051
4052         /* Count child PHBs */
4053         for_each_child_of_node(np, phbn) {
4054                 /* Look for IODA1 PHBs */
4055                 if (of_device_is_compatible(phbn, "ibm,ioda-phb"))
4056                         pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1);
4057         }
4058 }