1 /* 2 * Support PCI/PCIe on PowerNV platforms 3 * 4 * Copyright 2011 Benjamin Herrenschmidt, IBM Corp. 5 * 6 * This program is free software; you can redistribute it and/or 7 * modify it under the terms of the GNU General Public License 8 * as published by the Free Software Foundation; either version 9 * 2 of the License, or (at your option) any later version. 10 */ 11 12 #undef DEBUG 13 14 #include <linux/kernel.h> 15 #include <linux/pci.h> 16 #include <linux/crash_dump.h> 17 #include <linux/debugfs.h> 18 #include <linux/delay.h> 19 #include <linux/string.h> 20 #include <linux/init.h> 21 #include <linux/bootmem.h> 22 #include <linux/irq.h> 23 #include <linux/io.h> 24 #include <linux/msi.h> 25 #include <linux/memblock.h> 26 #include <linux/iommu.h> 27 #include <linux/rculist.h> 28 #include <linux/sizes.h> 29 30 #include <asm/sections.h> 31 #include <asm/io.h> 32 #include <asm/prom.h> 33 #include <asm/pci-bridge.h> 34 #include <asm/machdep.h> 35 #include <asm/msi_bitmap.h> 36 #include <asm/ppc-pci.h> 37 #include <asm/opal.h> 38 #include <asm/iommu.h> 39 #include <asm/tce.h> 40 #include <asm/xics.h> 41 #include <asm/debug.h> 42 #include <asm/firmware.h> 43 #include <asm/pnv-pci.h> 44 #include <asm/mmzone.h> 45 46 #include <misc/cxl-base.h> 47 48 #include "powernv.h" 49 #include "pci.h" 50 51 /* 256M DMA window, 4K TCE pages, 8 bytes TCE */ 52 #define TCE32_TABLE_SIZE ((0x10000000 / 0x1000) * 8) 53 54 #define POWERNV_IOMMU_DEFAULT_LEVELS 1 55 #define POWERNV_IOMMU_MAX_LEVELS 5 56 57 static void pnv_pci_ioda2_table_free_pages(struct iommu_table *tbl); 58 59 static void pe_level_printk(const struct pnv_ioda_pe *pe, const char *level, 60 const char *fmt, ...) 61 { 62 struct va_format vaf; 63 va_list args; 64 char pfix[32]; 65 66 va_start(args, fmt); 67 68 vaf.fmt = fmt; 69 vaf.va = &args; 70 71 if (pe->flags & PNV_IODA_PE_DEV) 72 strlcpy(pfix, dev_name(&pe->pdev->dev), sizeof(pfix)); 73 else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)) 74 sprintf(pfix, "%04x:%02x ", 75 pci_domain_nr(pe->pbus), pe->pbus->number); 76 #ifdef CONFIG_PCI_IOV 77 else if (pe->flags & PNV_IODA_PE_VF) 78 sprintf(pfix, "%04x:%02x:%2x.%d", 79 pci_domain_nr(pe->parent_dev->bus), 80 (pe->rid & 0xff00) >> 8, 81 PCI_SLOT(pe->rid), PCI_FUNC(pe->rid)); 82 #endif /* CONFIG_PCI_IOV*/ 83 84 printk("%spci %s: [PE# %.3d] %pV", 85 level, pfix, pe->pe_number, &vaf); 86 87 va_end(args); 88 } 89 90 #define pe_err(pe, fmt, ...) \ 91 pe_level_printk(pe, KERN_ERR, fmt, ##__VA_ARGS__) 92 #define pe_warn(pe, fmt, ...) \ 93 pe_level_printk(pe, KERN_WARNING, fmt, ##__VA_ARGS__) 94 #define pe_info(pe, fmt, ...) \ 95 pe_level_printk(pe, KERN_INFO, fmt, ##__VA_ARGS__) 96 97 static bool pnv_iommu_bypass_disabled __read_mostly; 98 99 static int __init iommu_setup(char *str) 100 { 101 if (!str) 102 return -EINVAL; 103 104 while (*str) { 105 if (!strncmp(str, "nobypass", 8)) { 106 pnv_iommu_bypass_disabled = true; 107 pr_info("PowerNV: IOMMU bypass window disabled.\n"); 108 break; 109 } 110 str += strcspn(str, ","); 111 if (*str == ',') 112 str++; 113 } 114 115 return 0; 116 } 117 early_param("iommu", iommu_setup); 118 119 /* 120 * stdcix is only supposed to be used in hypervisor real mode as per 121 * the architecture spec 122 */ 123 static inline void __raw_rm_writeq(u64 val, volatile void __iomem *paddr) 124 { 125 __asm__ __volatile__("stdcix %0,0,%1" 126 : : "r" (val), "r" (paddr) : "memory"); 127 } 128 129 static inline bool pnv_pci_is_mem_pref_64(unsigned long flags) 130 { 131 return ((flags & (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH)) == 132 (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH)); 133 } 134 135 static void pnv_ioda_reserve_pe(struct pnv_phb *phb, int pe_no) 136 { 137 if (!(pe_no >= 0 && pe_no < phb->ioda.total_pe)) { 138 pr_warn("%s: Invalid PE %d on PHB#%x\n", 139 __func__, pe_no, phb->hose->global_number); 140 return; 141 } 142 143 if (test_and_set_bit(pe_no, phb->ioda.pe_alloc)) { 144 pr_warn("%s: PE %d was assigned on PHB#%x\n", 145 __func__, pe_no, phb->hose->global_number); 146 return; 147 } 148 149 phb->ioda.pe_array[pe_no].phb = phb; 150 phb->ioda.pe_array[pe_no].pe_number = pe_no; 151 } 152 153 static int pnv_ioda_alloc_pe(struct pnv_phb *phb) 154 { 155 unsigned long pe; 156 157 do { 158 pe = find_next_zero_bit(phb->ioda.pe_alloc, 159 phb->ioda.total_pe, 0); 160 if (pe >= phb->ioda.total_pe) 161 return IODA_INVALID_PE; 162 } while(test_and_set_bit(pe, phb->ioda.pe_alloc)); 163 164 phb->ioda.pe_array[pe].phb = phb; 165 phb->ioda.pe_array[pe].pe_number = pe; 166 return pe; 167 } 168 169 static void pnv_ioda_free_pe(struct pnv_phb *phb, int pe) 170 { 171 WARN_ON(phb->ioda.pe_array[pe].pdev); 172 173 memset(&phb->ioda.pe_array[pe], 0, sizeof(struct pnv_ioda_pe)); 174 clear_bit(pe, phb->ioda.pe_alloc); 175 } 176 177 /* The default M64 BAR is shared by all PEs */ 178 static int pnv_ioda2_init_m64(struct pnv_phb *phb) 179 { 180 const char *desc; 181 struct resource *r; 182 s64 rc; 183 184 /* Configure the default M64 BAR */ 185 rc = opal_pci_set_phb_mem_window(phb->opal_id, 186 OPAL_M64_WINDOW_TYPE, 187 phb->ioda.m64_bar_idx, 188 phb->ioda.m64_base, 189 0, /* unused */ 190 phb->ioda.m64_size); 191 if (rc != OPAL_SUCCESS) { 192 desc = "configuring"; 193 goto fail; 194 } 195 196 /* Enable the default M64 BAR */ 197 rc = opal_pci_phb_mmio_enable(phb->opal_id, 198 OPAL_M64_WINDOW_TYPE, 199 phb->ioda.m64_bar_idx, 200 OPAL_ENABLE_M64_SPLIT); 201 if (rc != OPAL_SUCCESS) { 202 desc = "enabling"; 203 goto fail; 204 } 205 206 /* Mark the M64 BAR assigned */ 207 set_bit(phb->ioda.m64_bar_idx, &phb->ioda.m64_bar_alloc); 208 209 /* 210 * Strip off the segment used by the reserved PE, which is 211 * expected to be 0 or last one of PE capabicity. 212 */ 213 r = &phb->hose->mem_resources[1]; 214 if (phb->ioda.reserved_pe == 0) 215 r->start += phb->ioda.m64_segsize; 216 else if (phb->ioda.reserved_pe == (phb->ioda.total_pe - 1)) 217 r->end -= phb->ioda.m64_segsize; 218 else 219 pr_warn(" Cannot strip M64 segment for reserved PE#%d\n", 220 phb->ioda.reserved_pe); 221 222 return 0; 223 224 fail: 225 pr_warn(" Failure %lld %s M64 BAR#%d\n", 226 rc, desc, phb->ioda.m64_bar_idx); 227 opal_pci_phb_mmio_enable(phb->opal_id, 228 OPAL_M64_WINDOW_TYPE, 229 phb->ioda.m64_bar_idx, 230 OPAL_DISABLE_M64); 231 return -EIO; 232 } 233 234 static void pnv_ioda2_reserve_m64_pe(struct pnv_phb *phb) 235 { 236 resource_size_t sgsz = phb->ioda.m64_segsize; 237 struct pci_dev *pdev; 238 struct resource *r; 239 int base, step, i; 240 241 /* 242 * Root bus always has full M64 range and root port has 243 * M64 range used in reality. So we're checking root port 244 * instead of root bus. 245 */ 246 list_for_each_entry(pdev, &phb->hose->bus->devices, bus_list) { 247 for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++) { 248 r = &pdev->resource[PCI_BRIDGE_RESOURCES + i]; 249 if (!r->parent || 250 !pnv_pci_is_mem_pref_64(r->flags)) 251 continue; 252 253 base = (r->start - phb->ioda.m64_base) / sgsz; 254 for (step = 0; step < resource_size(r) / sgsz; step++) 255 pnv_ioda_reserve_pe(phb, base + step); 256 } 257 } 258 } 259 260 static int pnv_ioda2_pick_m64_pe(struct pnv_phb *phb, 261 struct pci_bus *bus, int all) 262 { 263 resource_size_t segsz = phb->ioda.m64_segsize; 264 struct pci_dev *pdev; 265 struct resource *r; 266 struct pnv_ioda_pe *master_pe, *pe; 267 unsigned long size, *pe_alloc; 268 bool found; 269 int start, i, j; 270 271 /* Root bus shouldn't use M64 */ 272 if (pci_is_root_bus(bus)) 273 return IODA_INVALID_PE; 274 275 /* We support only one M64 window on each bus */ 276 found = false; 277 pci_bus_for_each_resource(bus, r, i) { 278 if (r && r->parent && 279 pnv_pci_is_mem_pref_64(r->flags)) { 280 found = true; 281 break; 282 } 283 } 284 285 /* No M64 window found ? */ 286 if (!found) 287 return IODA_INVALID_PE; 288 289 /* Allocate bitmap */ 290 size = _ALIGN_UP(phb->ioda.total_pe / 8, sizeof(unsigned long)); 291 pe_alloc = kzalloc(size, GFP_KERNEL); 292 if (!pe_alloc) { 293 pr_warn("%s: Out of memory !\n", 294 __func__); 295 return IODA_INVALID_PE; 296 } 297 298 /* 299 * Figure out reserved PE numbers by the PE 300 * the its child PEs. 301 */ 302 start = (r->start - phb->ioda.m64_base) / segsz; 303 for (i = 0; i < resource_size(r) / segsz; i++) 304 set_bit(start + i, pe_alloc); 305 306 if (all) 307 goto done; 308 309 /* 310 * If the PE doesn't cover all subordinate buses, 311 * we need subtract from reserved PEs for children. 312 */ 313 list_for_each_entry(pdev, &bus->devices, bus_list) { 314 if (!pdev->subordinate) 315 continue; 316 317 pci_bus_for_each_resource(pdev->subordinate, r, i) { 318 if (!r || !r->parent || 319 !pnv_pci_is_mem_pref_64(r->flags)) 320 continue; 321 322 start = (r->start - phb->ioda.m64_base) / segsz; 323 for (j = 0; j < resource_size(r) / segsz ; j++) 324 clear_bit(start + j, pe_alloc); 325 } 326 } 327 328 /* 329 * the current bus might not own M64 window and that's all 330 * contributed by its child buses. For the case, we needn't 331 * pick M64 dependent PE#. 332 */ 333 if (bitmap_empty(pe_alloc, phb->ioda.total_pe)) { 334 kfree(pe_alloc); 335 return IODA_INVALID_PE; 336 } 337 338 /* 339 * Figure out the master PE and put all slave PEs to master 340 * PE's list to form compound PE. 341 */ 342 done: 343 master_pe = NULL; 344 i = -1; 345 while ((i = find_next_bit(pe_alloc, phb->ioda.total_pe, i + 1)) < 346 phb->ioda.total_pe) { 347 pe = &phb->ioda.pe_array[i]; 348 349 if (!master_pe) { 350 pe->flags |= PNV_IODA_PE_MASTER; 351 INIT_LIST_HEAD(&pe->slaves); 352 master_pe = pe; 353 } else { 354 pe->flags |= PNV_IODA_PE_SLAVE; 355 pe->master = master_pe; 356 list_add_tail(&pe->list, &master_pe->slaves); 357 } 358 } 359 360 kfree(pe_alloc); 361 return master_pe->pe_number; 362 } 363 364 static void __init pnv_ioda_parse_m64_window(struct pnv_phb *phb) 365 { 366 struct pci_controller *hose = phb->hose; 367 struct device_node *dn = hose->dn; 368 struct resource *res; 369 const u32 *r; 370 u64 pci_addr; 371 372 /* FIXME: Support M64 for P7IOC */ 373 if (phb->type != PNV_PHB_IODA2) { 374 pr_info(" Not support M64 window\n"); 375 return; 376 } 377 378 if (!firmware_has_feature(FW_FEATURE_OPALv3)) { 379 pr_info(" Firmware too old to support M64 window\n"); 380 return; 381 } 382 383 r = of_get_property(dn, "ibm,opal-m64-window", NULL); 384 if (!r) { 385 pr_info(" No <ibm,opal-m64-window> on %s\n", 386 dn->full_name); 387 return; 388 } 389 390 res = &hose->mem_resources[1]; 391 res->start = of_translate_address(dn, r + 2); 392 res->end = res->start + of_read_number(r + 4, 2) - 1; 393 res->flags = (IORESOURCE_MEM | IORESOURCE_MEM_64 | IORESOURCE_PREFETCH); 394 pci_addr = of_read_number(r, 2); 395 hose->mem_offset[1] = res->start - pci_addr; 396 397 phb->ioda.m64_size = resource_size(res); 398 phb->ioda.m64_segsize = phb->ioda.m64_size / phb->ioda.total_pe; 399 phb->ioda.m64_base = pci_addr; 400 401 pr_info(" MEM64 0x%016llx..0x%016llx -> 0x%016llx\n", 402 res->start, res->end, pci_addr); 403 404 /* Use last M64 BAR to cover M64 window */ 405 phb->ioda.m64_bar_idx = 15; 406 phb->init_m64 = pnv_ioda2_init_m64; 407 phb->reserve_m64_pe = pnv_ioda2_reserve_m64_pe; 408 phb->pick_m64_pe = pnv_ioda2_pick_m64_pe; 409 } 410 411 static void pnv_ioda_freeze_pe(struct pnv_phb *phb, int pe_no) 412 { 413 struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_no]; 414 struct pnv_ioda_pe *slave; 415 s64 rc; 416 417 /* Fetch master PE */ 418 if (pe->flags & PNV_IODA_PE_SLAVE) { 419 pe = pe->master; 420 if (WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER))) 421 return; 422 423 pe_no = pe->pe_number; 424 } 425 426 /* Freeze master PE */ 427 rc = opal_pci_eeh_freeze_set(phb->opal_id, 428 pe_no, 429 OPAL_EEH_ACTION_SET_FREEZE_ALL); 430 if (rc != OPAL_SUCCESS) { 431 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n", 432 __func__, rc, phb->hose->global_number, pe_no); 433 return; 434 } 435 436 /* Freeze slave PEs */ 437 if (!(pe->flags & PNV_IODA_PE_MASTER)) 438 return; 439 440 list_for_each_entry(slave, &pe->slaves, list) { 441 rc = opal_pci_eeh_freeze_set(phb->opal_id, 442 slave->pe_number, 443 OPAL_EEH_ACTION_SET_FREEZE_ALL); 444 if (rc != OPAL_SUCCESS) 445 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n", 446 __func__, rc, phb->hose->global_number, 447 slave->pe_number); 448 } 449 } 450 451 static int pnv_ioda_unfreeze_pe(struct pnv_phb *phb, int pe_no, int opt) 452 { 453 struct pnv_ioda_pe *pe, *slave; 454 s64 rc; 455 456 /* Find master PE */ 457 pe = &phb->ioda.pe_array[pe_no]; 458 if (pe->flags & PNV_IODA_PE_SLAVE) { 459 pe = pe->master; 460 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)); 461 pe_no = pe->pe_number; 462 } 463 464 /* Clear frozen state for master PE */ 465 rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, opt); 466 if (rc != OPAL_SUCCESS) { 467 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n", 468 __func__, rc, opt, phb->hose->global_number, pe_no); 469 return -EIO; 470 } 471 472 if (!(pe->flags & PNV_IODA_PE_MASTER)) 473 return 0; 474 475 /* Clear frozen state for slave PEs */ 476 list_for_each_entry(slave, &pe->slaves, list) { 477 rc = opal_pci_eeh_freeze_clear(phb->opal_id, 478 slave->pe_number, 479 opt); 480 if (rc != OPAL_SUCCESS) { 481 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n", 482 __func__, rc, opt, phb->hose->global_number, 483 slave->pe_number); 484 return -EIO; 485 } 486 } 487 488 return 0; 489 } 490 491 static int pnv_ioda_get_pe_state(struct pnv_phb *phb, int pe_no) 492 { 493 struct pnv_ioda_pe *slave, *pe; 494 u8 fstate, state; 495 __be16 pcierr; 496 s64 rc; 497 498 /* Sanity check on PE number */ 499 if (pe_no < 0 || pe_no >= phb->ioda.total_pe) 500 return OPAL_EEH_STOPPED_PERM_UNAVAIL; 501 502 /* 503 * Fetch the master PE and the PE instance might be 504 * not initialized yet. 505 */ 506 pe = &phb->ioda.pe_array[pe_no]; 507 if (pe->flags & PNV_IODA_PE_SLAVE) { 508 pe = pe->master; 509 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)); 510 pe_no = pe->pe_number; 511 } 512 513 /* Check the master PE */ 514 rc = opal_pci_eeh_freeze_status(phb->opal_id, pe_no, 515 &state, &pcierr, NULL); 516 if (rc != OPAL_SUCCESS) { 517 pr_warn("%s: Failure %lld getting " 518 "PHB#%x-PE#%x state\n", 519 __func__, rc, 520 phb->hose->global_number, pe_no); 521 return OPAL_EEH_STOPPED_TEMP_UNAVAIL; 522 } 523 524 /* Check the slave PE */ 525 if (!(pe->flags & PNV_IODA_PE_MASTER)) 526 return state; 527 528 list_for_each_entry(slave, &pe->slaves, list) { 529 rc = opal_pci_eeh_freeze_status(phb->opal_id, 530 slave->pe_number, 531 &fstate, 532 &pcierr, 533 NULL); 534 if (rc != OPAL_SUCCESS) { 535 pr_warn("%s: Failure %lld getting " 536 "PHB#%x-PE#%x state\n", 537 __func__, rc, 538 phb->hose->global_number, slave->pe_number); 539 return OPAL_EEH_STOPPED_TEMP_UNAVAIL; 540 } 541 542 /* 543 * Override the result based on the ascending 544 * priority. 545 */ 546 if (fstate > state) 547 state = fstate; 548 } 549 550 return state; 551 } 552 553 /* Currently those 2 are only used when MSIs are enabled, this will change 554 * but in the meantime, we need to protect them to avoid warnings 555 */ 556 #ifdef CONFIG_PCI_MSI 557 static struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev) 558 { 559 struct pci_controller *hose = pci_bus_to_host(dev->bus); 560 struct pnv_phb *phb = hose->private_data; 561 struct pci_dn *pdn = pci_get_pdn(dev); 562 563 if (!pdn) 564 return NULL; 565 if (pdn->pe_number == IODA_INVALID_PE) 566 return NULL; 567 return &phb->ioda.pe_array[pdn->pe_number]; 568 } 569 #endif /* CONFIG_PCI_MSI */ 570 571 static int pnv_ioda_set_one_peltv(struct pnv_phb *phb, 572 struct pnv_ioda_pe *parent, 573 struct pnv_ioda_pe *child, 574 bool is_add) 575 { 576 const char *desc = is_add ? "adding" : "removing"; 577 uint8_t op = is_add ? OPAL_ADD_PE_TO_DOMAIN : 578 OPAL_REMOVE_PE_FROM_DOMAIN; 579 struct pnv_ioda_pe *slave; 580 long rc; 581 582 /* Parent PE affects child PE */ 583 rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number, 584 child->pe_number, op); 585 if (rc != OPAL_SUCCESS) { 586 pe_warn(child, "OPAL error %ld %s to parent PELTV\n", 587 rc, desc); 588 return -ENXIO; 589 } 590 591 if (!(child->flags & PNV_IODA_PE_MASTER)) 592 return 0; 593 594 /* Compound case: parent PE affects slave PEs */ 595 list_for_each_entry(slave, &child->slaves, list) { 596 rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number, 597 slave->pe_number, op); 598 if (rc != OPAL_SUCCESS) { 599 pe_warn(slave, "OPAL error %ld %s to parent PELTV\n", 600 rc, desc); 601 return -ENXIO; 602 } 603 } 604 605 return 0; 606 } 607 608 static int pnv_ioda_set_peltv(struct pnv_phb *phb, 609 struct pnv_ioda_pe *pe, 610 bool is_add) 611 { 612 struct pnv_ioda_pe *slave; 613 struct pci_dev *pdev = NULL; 614 int ret; 615 616 /* 617 * Clear PE frozen state. If it's master PE, we need 618 * clear slave PE frozen state as well. 619 */ 620 if (is_add) { 621 opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number, 622 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL); 623 if (pe->flags & PNV_IODA_PE_MASTER) { 624 list_for_each_entry(slave, &pe->slaves, list) 625 opal_pci_eeh_freeze_clear(phb->opal_id, 626 slave->pe_number, 627 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL); 628 } 629 } 630 631 /* 632 * Associate PE in PELT. We need add the PE into the 633 * corresponding PELT-V as well. Otherwise, the error 634 * originated from the PE might contribute to other 635 * PEs. 636 */ 637 ret = pnv_ioda_set_one_peltv(phb, pe, pe, is_add); 638 if (ret) 639 return ret; 640 641 /* For compound PEs, any one affects all of them */ 642 if (pe->flags & PNV_IODA_PE_MASTER) { 643 list_for_each_entry(slave, &pe->slaves, list) { 644 ret = pnv_ioda_set_one_peltv(phb, slave, pe, is_add); 645 if (ret) 646 return ret; 647 } 648 } 649 650 if (pe->flags & (PNV_IODA_PE_BUS_ALL | PNV_IODA_PE_BUS)) 651 pdev = pe->pbus->self; 652 else if (pe->flags & PNV_IODA_PE_DEV) 653 pdev = pe->pdev->bus->self; 654 #ifdef CONFIG_PCI_IOV 655 else if (pe->flags & PNV_IODA_PE_VF) 656 pdev = pe->parent_dev->bus->self; 657 #endif /* CONFIG_PCI_IOV */ 658 while (pdev) { 659 struct pci_dn *pdn = pci_get_pdn(pdev); 660 struct pnv_ioda_pe *parent; 661 662 if (pdn && pdn->pe_number != IODA_INVALID_PE) { 663 parent = &phb->ioda.pe_array[pdn->pe_number]; 664 ret = pnv_ioda_set_one_peltv(phb, parent, pe, is_add); 665 if (ret) 666 return ret; 667 } 668 669 pdev = pdev->bus->self; 670 } 671 672 return 0; 673 } 674 675 #ifdef CONFIG_PCI_IOV 676 static int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe) 677 { 678 struct pci_dev *parent; 679 uint8_t bcomp, dcomp, fcomp; 680 int64_t rc; 681 long rid_end, rid; 682 683 /* Currently, we just deconfigure VF PE. Bus PE will always there.*/ 684 if (pe->pbus) { 685 int count; 686 687 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER; 688 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER; 689 parent = pe->pbus->self; 690 if (pe->flags & PNV_IODA_PE_BUS_ALL) 691 count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1; 692 else 693 count = 1; 694 695 switch(count) { 696 case 1: bcomp = OpalPciBusAll; break; 697 case 2: bcomp = OpalPciBus7Bits; break; 698 case 4: bcomp = OpalPciBus6Bits; break; 699 case 8: bcomp = OpalPciBus5Bits; break; 700 case 16: bcomp = OpalPciBus4Bits; break; 701 case 32: bcomp = OpalPciBus3Bits; break; 702 default: 703 dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n", 704 count); 705 /* Do an exact match only */ 706 bcomp = OpalPciBusAll; 707 } 708 rid_end = pe->rid + (count << 8); 709 } else { 710 if (pe->flags & PNV_IODA_PE_VF) 711 parent = pe->parent_dev; 712 else 713 parent = pe->pdev->bus->self; 714 bcomp = OpalPciBusAll; 715 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER; 716 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER; 717 rid_end = pe->rid + 1; 718 } 719 720 /* Clear the reverse map */ 721 for (rid = pe->rid; rid < rid_end; rid++) 722 phb->ioda.pe_rmap[rid] = 0; 723 724 /* Release from all parents PELT-V */ 725 while (parent) { 726 struct pci_dn *pdn = pci_get_pdn(parent); 727 if (pdn && pdn->pe_number != IODA_INVALID_PE) { 728 rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number, 729 pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN); 730 /* XXX What to do in case of error ? */ 731 } 732 parent = parent->bus->self; 733 } 734 735 opal_pci_eeh_freeze_set(phb->opal_id, pe->pe_number, 736 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL); 737 738 /* Disassociate PE in PELT */ 739 rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number, 740 pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN); 741 if (rc) 742 pe_warn(pe, "OPAL error %ld remove self from PELTV\n", rc); 743 rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid, 744 bcomp, dcomp, fcomp, OPAL_UNMAP_PE); 745 if (rc) 746 pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc); 747 748 pe->pbus = NULL; 749 pe->pdev = NULL; 750 pe->parent_dev = NULL; 751 752 return 0; 753 } 754 #endif /* CONFIG_PCI_IOV */ 755 756 static int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe) 757 { 758 struct pci_dev *parent; 759 uint8_t bcomp, dcomp, fcomp; 760 long rc, rid_end, rid; 761 762 /* Bus validation ? */ 763 if (pe->pbus) { 764 int count; 765 766 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER; 767 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER; 768 parent = pe->pbus->self; 769 if (pe->flags & PNV_IODA_PE_BUS_ALL) 770 count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1; 771 else 772 count = 1; 773 774 switch(count) { 775 case 1: bcomp = OpalPciBusAll; break; 776 case 2: bcomp = OpalPciBus7Bits; break; 777 case 4: bcomp = OpalPciBus6Bits; break; 778 case 8: bcomp = OpalPciBus5Bits; break; 779 case 16: bcomp = OpalPciBus4Bits; break; 780 case 32: bcomp = OpalPciBus3Bits; break; 781 default: 782 dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n", 783 count); 784 /* Do an exact match only */ 785 bcomp = OpalPciBusAll; 786 } 787 rid_end = pe->rid + (count << 8); 788 } else { 789 #ifdef CONFIG_PCI_IOV 790 if (pe->flags & PNV_IODA_PE_VF) 791 parent = pe->parent_dev; 792 else 793 #endif /* CONFIG_PCI_IOV */ 794 parent = pe->pdev->bus->self; 795 bcomp = OpalPciBusAll; 796 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER; 797 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER; 798 rid_end = pe->rid + 1; 799 } 800 801 /* 802 * Associate PE in PELT. We need add the PE into the 803 * corresponding PELT-V as well. Otherwise, the error 804 * originated from the PE might contribute to other 805 * PEs. 806 */ 807 rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid, 808 bcomp, dcomp, fcomp, OPAL_MAP_PE); 809 if (rc) { 810 pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc); 811 return -ENXIO; 812 } 813 814 /* Configure PELTV */ 815 pnv_ioda_set_peltv(phb, pe, true); 816 817 /* Setup reverse map */ 818 for (rid = pe->rid; rid < rid_end; rid++) 819 phb->ioda.pe_rmap[rid] = pe->pe_number; 820 821 /* Setup one MVTs on IODA1 */ 822 if (phb->type != PNV_PHB_IODA1) { 823 pe->mve_number = 0; 824 goto out; 825 } 826 827 pe->mve_number = pe->pe_number; 828 rc = opal_pci_set_mve(phb->opal_id, pe->mve_number, pe->pe_number); 829 if (rc != OPAL_SUCCESS) { 830 pe_err(pe, "OPAL error %ld setting up MVE %d\n", 831 rc, pe->mve_number); 832 pe->mve_number = -1; 833 } else { 834 rc = opal_pci_set_mve_enable(phb->opal_id, 835 pe->mve_number, OPAL_ENABLE_MVE); 836 if (rc) { 837 pe_err(pe, "OPAL error %ld enabling MVE %d\n", 838 rc, pe->mve_number); 839 pe->mve_number = -1; 840 } 841 } 842 843 out: 844 return 0; 845 } 846 847 static void pnv_ioda_link_pe_by_weight(struct pnv_phb *phb, 848 struct pnv_ioda_pe *pe) 849 { 850 struct pnv_ioda_pe *lpe; 851 852 list_for_each_entry(lpe, &phb->ioda.pe_dma_list, dma_link) { 853 if (lpe->dma_weight < pe->dma_weight) { 854 list_add_tail(&pe->dma_link, &lpe->dma_link); 855 return; 856 } 857 } 858 list_add_tail(&pe->dma_link, &phb->ioda.pe_dma_list); 859 } 860 861 static unsigned int pnv_ioda_dma_weight(struct pci_dev *dev) 862 { 863 /* This is quite simplistic. The "base" weight of a device 864 * is 10. 0 means no DMA is to be accounted for it. 865 */ 866 867 /* If it's a bridge, no DMA */ 868 if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL) 869 return 0; 870 871 /* Reduce the weight of slow USB controllers */ 872 if (dev->class == PCI_CLASS_SERIAL_USB_UHCI || 873 dev->class == PCI_CLASS_SERIAL_USB_OHCI || 874 dev->class == PCI_CLASS_SERIAL_USB_EHCI) 875 return 3; 876 877 /* Increase the weight of RAID (includes Obsidian) */ 878 if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID) 879 return 15; 880 881 /* Default */ 882 return 10; 883 } 884 885 #ifdef CONFIG_PCI_IOV 886 static int pnv_pci_vf_resource_shift(struct pci_dev *dev, int offset) 887 { 888 struct pci_dn *pdn = pci_get_pdn(dev); 889 int i; 890 struct resource *res, res2; 891 resource_size_t size; 892 u16 num_vfs; 893 894 if (!dev->is_physfn) 895 return -EINVAL; 896 897 /* 898 * "offset" is in VFs. The M64 windows are sized so that when they 899 * are segmented, each segment is the same size as the IOV BAR. 900 * Each segment is in a separate PE, and the high order bits of the 901 * address are the PE number. Therefore, each VF's BAR is in a 902 * separate PE, and changing the IOV BAR start address changes the 903 * range of PEs the VFs are in. 904 */ 905 num_vfs = pdn->num_vfs; 906 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { 907 res = &dev->resource[i + PCI_IOV_RESOURCES]; 908 if (!res->flags || !res->parent) 909 continue; 910 911 if (!pnv_pci_is_mem_pref_64(res->flags)) 912 continue; 913 914 /* 915 * The actual IOV BAR range is determined by the start address 916 * and the actual size for num_vfs VFs BAR. This check is to 917 * make sure that after shifting, the range will not overlap 918 * with another device. 919 */ 920 size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES); 921 res2.flags = res->flags; 922 res2.start = res->start + (size * offset); 923 res2.end = res2.start + (size * num_vfs) - 1; 924 925 if (res2.end > res->end) { 926 dev_err(&dev->dev, "VF BAR%d: %pR would extend past %pR (trying to enable %d VFs shifted by %d)\n", 927 i, &res2, res, num_vfs, offset); 928 return -EBUSY; 929 } 930 } 931 932 /* 933 * After doing so, there would be a "hole" in the /proc/iomem when 934 * offset is a positive value. It looks like the device return some 935 * mmio back to the system, which actually no one could use it. 936 */ 937 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { 938 res = &dev->resource[i + PCI_IOV_RESOURCES]; 939 if (!res->flags || !res->parent) 940 continue; 941 942 if (!pnv_pci_is_mem_pref_64(res->flags)) 943 continue; 944 945 size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES); 946 res2 = *res; 947 res->start += size * offset; 948 949 dev_info(&dev->dev, "VF BAR%d: %pR shifted to %pR (enabling %d VFs shifted by %d)\n", 950 i, &res2, res, num_vfs, offset); 951 pci_update_resource(dev, i + PCI_IOV_RESOURCES); 952 } 953 return 0; 954 } 955 #endif /* CONFIG_PCI_IOV */ 956 957 #if 0 958 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev) 959 { 960 struct pci_controller *hose = pci_bus_to_host(dev->bus); 961 struct pnv_phb *phb = hose->private_data; 962 struct pci_dn *pdn = pci_get_pdn(dev); 963 struct pnv_ioda_pe *pe; 964 int pe_num; 965 966 if (!pdn) { 967 pr_err("%s: Device tree node not associated properly\n", 968 pci_name(dev)); 969 return NULL; 970 } 971 if (pdn->pe_number != IODA_INVALID_PE) 972 return NULL; 973 974 /* PE#0 has been pre-set */ 975 if (dev->bus->number == 0) 976 pe_num = 0; 977 else 978 pe_num = pnv_ioda_alloc_pe(phb); 979 if (pe_num == IODA_INVALID_PE) { 980 pr_warning("%s: Not enough PE# available, disabling device\n", 981 pci_name(dev)); 982 return NULL; 983 } 984 985 /* NOTE: We get only one ref to the pci_dev for the pdn, not for the 986 * pointer in the PE data structure, both should be destroyed at the 987 * same time. However, this needs to be looked at more closely again 988 * once we actually start removing things (Hotplug, SR-IOV, ...) 989 * 990 * At some point we want to remove the PDN completely anyways 991 */ 992 pe = &phb->ioda.pe_array[pe_num]; 993 pci_dev_get(dev); 994 pdn->pcidev = dev; 995 pdn->pe_number = pe_num; 996 pe->pdev = dev; 997 pe->pbus = NULL; 998 pe->tce32_seg = -1; 999 pe->mve_number = -1; 1000 pe->rid = dev->bus->number << 8 | pdn->devfn; 1001 1002 pe_info(pe, "Associated device to PE\n"); 1003 1004 if (pnv_ioda_configure_pe(phb, pe)) { 1005 /* XXX What do we do here ? */ 1006 if (pe_num) 1007 pnv_ioda_free_pe(phb, pe_num); 1008 pdn->pe_number = IODA_INVALID_PE; 1009 pe->pdev = NULL; 1010 pci_dev_put(dev); 1011 return NULL; 1012 } 1013 1014 /* Assign a DMA weight to the device */ 1015 pe->dma_weight = pnv_ioda_dma_weight(dev); 1016 if (pe->dma_weight != 0) { 1017 phb->ioda.dma_weight += pe->dma_weight; 1018 phb->ioda.dma_pe_count++; 1019 } 1020 1021 /* Link the PE */ 1022 pnv_ioda_link_pe_by_weight(phb, pe); 1023 1024 return pe; 1025 } 1026 #endif /* Useful for SRIOV case */ 1027 1028 static void pnv_ioda_setup_same_PE(struct pci_bus *bus, struct pnv_ioda_pe *pe) 1029 { 1030 struct pci_dev *dev; 1031 1032 list_for_each_entry(dev, &bus->devices, bus_list) { 1033 struct pci_dn *pdn = pci_get_pdn(dev); 1034 1035 if (pdn == NULL) { 1036 pr_warn("%s: No device node associated with device !\n", 1037 pci_name(dev)); 1038 continue; 1039 } 1040 pdn->pe_number = pe->pe_number; 1041 pe->dma_weight += pnv_ioda_dma_weight(dev); 1042 if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate) 1043 pnv_ioda_setup_same_PE(dev->subordinate, pe); 1044 } 1045 } 1046 1047 /* 1048 * There're 2 types of PCI bus sensitive PEs: One that is compromised of 1049 * single PCI bus. Another one that contains the primary PCI bus and its 1050 * subordinate PCI devices and buses. The second type of PE is normally 1051 * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports. 1052 */ 1053 static void pnv_ioda_setup_bus_PE(struct pci_bus *bus, int all) 1054 { 1055 struct pci_controller *hose = pci_bus_to_host(bus); 1056 struct pnv_phb *phb = hose->private_data; 1057 struct pnv_ioda_pe *pe; 1058 int pe_num = IODA_INVALID_PE; 1059 1060 /* Check if PE is determined by M64 */ 1061 if (phb->pick_m64_pe) 1062 pe_num = phb->pick_m64_pe(phb, bus, all); 1063 1064 /* The PE number isn't pinned by M64 */ 1065 if (pe_num == IODA_INVALID_PE) 1066 pe_num = pnv_ioda_alloc_pe(phb); 1067 1068 if (pe_num == IODA_INVALID_PE) { 1069 pr_warning("%s: Not enough PE# available for PCI bus %04x:%02x\n", 1070 __func__, pci_domain_nr(bus), bus->number); 1071 return; 1072 } 1073 1074 pe = &phb->ioda.pe_array[pe_num]; 1075 pe->flags |= (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS); 1076 pe->pbus = bus; 1077 pe->pdev = NULL; 1078 pe->tce32_seg = -1; 1079 pe->mve_number = -1; 1080 pe->rid = bus->busn_res.start << 8; 1081 pe->dma_weight = 0; 1082 1083 if (all) 1084 pe_info(pe, "Secondary bus %d..%d associated with PE#%d\n", 1085 bus->busn_res.start, bus->busn_res.end, pe_num); 1086 else 1087 pe_info(pe, "Secondary bus %d associated with PE#%d\n", 1088 bus->busn_res.start, pe_num); 1089 1090 if (pnv_ioda_configure_pe(phb, pe)) { 1091 /* XXX What do we do here ? */ 1092 if (pe_num) 1093 pnv_ioda_free_pe(phb, pe_num); 1094 pe->pbus = NULL; 1095 return; 1096 } 1097 1098 /* Associate it with all child devices */ 1099 pnv_ioda_setup_same_PE(bus, pe); 1100 1101 /* Put PE to the list */ 1102 list_add_tail(&pe->list, &phb->ioda.pe_list); 1103 1104 /* Account for one DMA PE if at least one DMA capable device exist 1105 * below the bridge 1106 */ 1107 if (pe->dma_weight != 0) { 1108 phb->ioda.dma_weight += pe->dma_weight; 1109 phb->ioda.dma_pe_count++; 1110 } 1111 1112 /* Link the PE */ 1113 pnv_ioda_link_pe_by_weight(phb, pe); 1114 } 1115 1116 static void pnv_ioda_setup_PEs(struct pci_bus *bus) 1117 { 1118 struct pci_dev *dev; 1119 1120 pnv_ioda_setup_bus_PE(bus, 0); 1121 1122 list_for_each_entry(dev, &bus->devices, bus_list) { 1123 if (dev->subordinate) { 1124 if (pci_pcie_type(dev) == PCI_EXP_TYPE_PCI_BRIDGE) 1125 pnv_ioda_setup_bus_PE(dev->subordinate, 1); 1126 else 1127 pnv_ioda_setup_PEs(dev->subordinate); 1128 } 1129 } 1130 } 1131 1132 /* 1133 * Configure PEs so that the downstream PCI buses and devices 1134 * could have their associated PE#. Unfortunately, we didn't 1135 * figure out the way to identify the PLX bridge yet. So we 1136 * simply put the PCI bus and the subordinate behind the root 1137 * port to PE# here. The game rule here is expected to be changed 1138 * as soon as we can detected PLX bridge correctly. 1139 */ 1140 static void pnv_pci_ioda_setup_PEs(void) 1141 { 1142 struct pci_controller *hose, *tmp; 1143 struct pnv_phb *phb; 1144 1145 list_for_each_entry_safe(hose, tmp, &hose_list, list_node) { 1146 phb = hose->private_data; 1147 1148 /* M64 layout might affect PE allocation */ 1149 if (phb->reserve_m64_pe) 1150 phb->reserve_m64_pe(phb); 1151 1152 pnv_ioda_setup_PEs(hose->bus); 1153 } 1154 } 1155 1156 #ifdef CONFIG_PCI_IOV 1157 static int pnv_pci_vf_release_m64(struct pci_dev *pdev) 1158 { 1159 struct pci_bus *bus; 1160 struct pci_controller *hose; 1161 struct pnv_phb *phb; 1162 struct pci_dn *pdn; 1163 int i, j; 1164 1165 bus = pdev->bus; 1166 hose = pci_bus_to_host(bus); 1167 phb = hose->private_data; 1168 pdn = pci_get_pdn(pdev); 1169 1170 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) 1171 for (j = 0; j < M64_PER_IOV; j++) { 1172 if (pdn->m64_wins[i][j] == IODA_INVALID_M64) 1173 continue; 1174 opal_pci_phb_mmio_enable(phb->opal_id, 1175 OPAL_M64_WINDOW_TYPE, pdn->m64_wins[i][j], 0); 1176 clear_bit(pdn->m64_wins[i][j], &phb->ioda.m64_bar_alloc); 1177 pdn->m64_wins[i][j] = IODA_INVALID_M64; 1178 } 1179 1180 return 0; 1181 } 1182 1183 static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs) 1184 { 1185 struct pci_bus *bus; 1186 struct pci_controller *hose; 1187 struct pnv_phb *phb; 1188 struct pci_dn *pdn; 1189 unsigned int win; 1190 struct resource *res; 1191 int i, j; 1192 int64_t rc; 1193 int total_vfs; 1194 resource_size_t size, start; 1195 int pe_num; 1196 int vf_groups; 1197 int vf_per_group; 1198 1199 bus = pdev->bus; 1200 hose = pci_bus_to_host(bus); 1201 phb = hose->private_data; 1202 pdn = pci_get_pdn(pdev); 1203 total_vfs = pci_sriov_get_totalvfs(pdev); 1204 1205 /* Initialize the m64_wins to IODA_INVALID_M64 */ 1206 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) 1207 for (j = 0; j < M64_PER_IOV; j++) 1208 pdn->m64_wins[i][j] = IODA_INVALID_M64; 1209 1210 if (pdn->m64_per_iov == M64_PER_IOV) { 1211 vf_groups = (num_vfs <= M64_PER_IOV) ? num_vfs: M64_PER_IOV; 1212 vf_per_group = (num_vfs <= M64_PER_IOV)? 1: 1213 roundup_pow_of_two(num_vfs) / pdn->m64_per_iov; 1214 } else { 1215 vf_groups = 1; 1216 vf_per_group = 1; 1217 } 1218 1219 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { 1220 res = &pdev->resource[i + PCI_IOV_RESOURCES]; 1221 if (!res->flags || !res->parent) 1222 continue; 1223 1224 if (!pnv_pci_is_mem_pref_64(res->flags)) 1225 continue; 1226 1227 for (j = 0; j < vf_groups; j++) { 1228 do { 1229 win = find_next_zero_bit(&phb->ioda.m64_bar_alloc, 1230 phb->ioda.m64_bar_idx + 1, 0); 1231 1232 if (win >= phb->ioda.m64_bar_idx + 1) 1233 goto m64_failed; 1234 } while (test_and_set_bit(win, &phb->ioda.m64_bar_alloc)); 1235 1236 pdn->m64_wins[i][j] = win; 1237 1238 if (pdn->m64_per_iov == M64_PER_IOV) { 1239 size = pci_iov_resource_size(pdev, 1240 PCI_IOV_RESOURCES + i); 1241 size = size * vf_per_group; 1242 start = res->start + size * j; 1243 } else { 1244 size = resource_size(res); 1245 start = res->start; 1246 } 1247 1248 /* Map the M64 here */ 1249 if (pdn->m64_per_iov == M64_PER_IOV) { 1250 pe_num = pdn->offset + j; 1251 rc = opal_pci_map_pe_mmio_window(phb->opal_id, 1252 pe_num, OPAL_M64_WINDOW_TYPE, 1253 pdn->m64_wins[i][j], 0); 1254 } 1255 1256 rc = opal_pci_set_phb_mem_window(phb->opal_id, 1257 OPAL_M64_WINDOW_TYPE, 1258 pdn->m64_wins[i][j], 1259 start, 1260 0, /* unused */ 1261 size); 1262 1263 1264 if (rc != OPAL_SUCCESS) { 1265 dev_err(&pdev->dev, "Failed to map M64 window #%d: %lld\n", 1266 win, rc); 1267 goto m64_failed; 1268 } 1269 1270 if (pdn->m64_per_iov == M64_PER_IOV) 1271 rc = opal_pci_phb_mmio_enable(phb->opal_id, 1272 OPAL_M64_WINDOW_TYPE, pdn->m64_wins[i][j], 2); 1273 else 1274 rc = opal_pci_phb_mmio_enable(phb->opal_id, 1275 OPAL_M64_WINDOW_TYPE, pdn->m64_wins[i][j], 1); 1276 1277 if (rc != OPAL_SUCCESS) { 1278 dev_err(&pdev->dev, "Failed to enable M64 window #%d: %llx\n", 1279 win, rc); 1280 goto m64_failed; 1281 } 1282 } 1283 } 1284 return 0; 1285 1286 m64_failed: 1287 pnv_pci_vf_release_m64(pdev); 1288 return -EBUSY; 1289 } 1290 1291 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group, 1292 int num); 1293 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable); 1294 1295 static void pnv_pci_ioda2_release_dma_pe(struct pci_dev *dev, struct pnv_ioda_pe *pe) 1296 { 1297 struct iommu_table *tbl; 1298 int64_t rc; 1299 1300 tbl = pe->table_group.tables[0]; 1301 rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0); 1302 if (rc) 1303 pe_warn(pe, "OPAL error %ld release DMA window\n", rc); 1304 1305 pnv_pci_ioda2_set_bypass(pe, false); 1306 if (pe->table_group.group) { 1307 iommu_group_put(pe->table_group.group); 1308 BUG_ON(pe->table_group.group); 1309 } 1310 pnv_pci_ioda2_table_free_pages(tbl); 1311 iommu_free_table(tbl, of_node_full_name(dev->dev.of_node)); 1312 } 1313 1314 static void pnv_ioda_release_vf_PE(struct pci_dev *pdev, u16 num_vfs) 1315 { 1316 struct pci_bus *bus; 1317 struct pci_controller *hose; 1318 struct pnv_phb *phb; 1319 struct pnv_ioda_pe *pe, *pe_n; 1320 struct pci_dn *pdn; 1321 u16 vf_index; 1322 int64_t rc; 1323 1324 bus = pdev->bus; 1325 hose = pci_bus_to_host(bus); 1326 phb = hose->private_data; 1327 pdn = pci_get_pdn(pdev); 1328 1329 if (!pdev->is_physfn) 1330 return; 1331 1332 if (pdn->m64_per_iov == M64_PER_IOV && num_vfs > M64_PER_IOV) { 1333 int vf_group; 1334 int vf_per_group; 1335 int vf_index1; 1336 1337 vf_per_group = roundup_pow_of_two(num_vfs) / pdn->m64_per_iov; 1338 1339 for (vf_group = 0; vf_group < M64_PER_IOV; vf_group++) 1340 for (vf_index = vf_group * vf_per_group; 1341 vf_index < (vf_group + 1) * vf_per_group && 1342 vf_index < num_vfs; 1343 vf_index++) 1344 for (vf_index1 = vf_group * vf_per_group; 1345 vf_index1 < (vf_group + 1) * vf_per_group && 1346 vf_index1 < num_vfs; 1347 vf_index1++){ 1348 1349 rc = opal_pci_set_peltv(phb->opal_id, 1350 pdn->offset + vf_index, 1351 pdn->offset + vf_index1, 1352 OPAL_REMOVE_PE_FROM_DOMAIN); 1353 1354 if (rc) 1355 dev_warn(&pdev->dev, "%s: Failed to unlink same group PE#%d(%lld)\n", 1356 __func__, 1357 pdn->offset + vf_index1, rc); 1358 } 1359 } 1360 1361 list_for_each_entry_safe(pe, pe_n, &phb->ioda.pe_list, list) { 1362 if (pe->parent_dev != pdev) 1363 continue; 1364 1365 pnv_pci_ioda2_release_dma_pe(pdev, pe); 1366 1367 /* Remove from list */ 1368 mutex_lock(&phb->ioda.pe_list_mutex); 1369 list_del(&pe->list); 1370 mutex_unlock(&phb->ioda.pe_list_mutex); 1371 1372 pnv_ioda_deconfigure_pe(phb, pe); 1373 1374 pnv_ioda_free_pe(phb, pe->pe_number); 1375 } 1376 } 1377 1378 void pnv_pci_sriov_disable(struct pci_dev *pdev) 1379 { 1380 struct pci_bus *bus; 1381 struct pci_controller *hose; 1382 struct pnv_phb *phb; 1383 struct pci_dn *pdn; 1384 struct pci_sriov *iov; 1385 u16 num_vfs; 1386 1387 bus = pdev->bus; 1388 hose = pci_bus_to_host(bus); 1389 phb = hose->private_data; 1390 pdn = pci_get_pdn(pdev); 1391 iov = pdev->sriov; 1392 num_vfs = pdn->num_vfs; 1393 1394 /* Release VF PEs */ 1395 pnv_ioda_release_vf_PE(pdev, num_vfs); 1396 1397 if (phb->type == PNV_PHB_IODA2) { 1398 if (pdn->m64_per_iov == 1) 1399 pnv_pci_vf_resource_shift(pdev, -pdn->offset); 1400 1401 /* Release M64 windows */ 1402 pnv_pci_vf_release_m64(pdev); 1403 1404 /* Release PE numbers */ 1405 bitmap_clear(phb->ioda.pe_alloc, pdn->offset, num_vfs); 1406 pdn->offset = 0; 1407 } 1408 } 1409 1410 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb, 1411 struct pnv_ioda_pe *pe); 1412 static void pnv_ioda_setup_vf_PE(struct pci_dev *pdev, u16 num_vfs) 1413 { 1414 struct pci_bus *bus; 1415 struct pci_controller *hose; 1416 struct pnv_phb *phb; 1417 struct pnv_ioda_pe *pe; 1418 int pe_num; 1419 u16 vf_index; 1420 struct pci_dn *pdn; 1421 int64_t rc; 1422 1423 bus = pdev->bus; 1424 hose = pci_bus_to_host(bus); 1425 phb = hose->private_data; 1426 pdn = pci_get_pdn(pdev); 1427 1428 if (!pdev->is_physfn) 1429 return; 1430 1431 /* Reserve PE for each VF */ 1432 for (vf_index = 0; vf_index < num_vfs; vf_index++) { 1433 pe_num = pdn->offset + vf_index; 1434 1435 pe = &phb->ioda.pe_array[pe_num]; 1436 pe->pe_number = pe_num; 1437 pe->phb = phb; 1438 pe->flags = PNV_IODA_PE_VF; 1439 pe->pbus = NULL; 1440 pe->parent_dev = pdev; 1441 pe->tce32_seg = -1; 1442 pe->mve_number = -1; 1443 pe->rid = (pci_iov_virtfn_bus(pdev, vf_index) << 8) | 1444 pci_iov_virtfn_devfn(pdev, vf_index); 1445 1446 pe_info(pe, "VF %04d:%02d:%02d.%d associated with PE#%d\n", 1447 hose->global_number, pdev->bus->number, 1448 PCI_SLOT(pci_iov_virtfn_devfn(pdev, vf_index)), 1449 PCI_FUNC(pci_iov_virtfn_devfn(pdev, vf_index)), pe_num); 1450 1451 if (pnv_ioda_configure_pe(phb, pe)) { 1452 /* XXX What do we do here ? */ 1453 if (pe_num) 1454 pnv_ioda_free_pe(phb, pe_num); 1455 pe->pdev = NULL; 1456 continue; 1457 } 1458 1459 /* Put PE to the list */ 1460 mutex_lock(&phb->ioda.pe_list_mutex); 1461 list_add_tail(&pe->list, &phb->ioda.pe_list); 1462 mutex_unlock(&phb->ioda.pe_list_mutex); 1463 1464 pnv_pci_ioda2_setup_dma_pe(phb, pe); 1465 } 1466 1467 if (pdn->m64_per_iov == M64_PER_IOV && num_vfs > M64_PER_IOV) { 1468 int vf_group; 1469 int vf_per_group; 1470 int vf_index1; 1471 1472 vf_per_group = roundup_pow_of_two(num_vfs) / pdn->m64_per_iov; 1473 1474 for (vf_group = 0; vf_group < M64_PER_IOV; vf_group++) { 1475 for (vf_index = vf_group * vf_per_group; 1476 vf_index < (vf_group + 1) * vf_per_group && 1477 vf_index < num_vfs; 1478 vf_index++) { 1479 for (vf_index1 = vf_group * vf_per_group; 1480 vf_index1 < (vf_group + 1) * vf_per_group && 1481 vf_index1 < num_vfs; 1482 vf_index1++) { 1483 1484 rc = opal_pci_set_peltv(phb->opal_id, 1485 pdn->offset + vf_index, 1486 pdn->offset + vf_index1, 1487 OPAL_ADD_PE_TO_DOMAIN); 1488 1489 if (rc) 1490 dev_warn(&pdev->dev, "%s: Failed to link same group PE#%d(%lld)\n", 1491 __func__, 1492 pdn->offset + vf_index1, rc); 1493 } 1494 } 1495 } 1496 } 1497 } 1498 1499 int pnv_pci_sriov_enable(struct pci_dev *pdev, u16 num_vfs) 1500 { 1501 struct pci_bus *bus; 1502 struct pci_controller *hose; 1503 struct pnv_phb *phb; 1504 struct pci_dn *pdn; 1505 int ret; 1506 1507 bus = pdev->bus; 1508 hose = pci_bus_to_host(bus); 1509 phb = hose->private_data; 1510 pdn = pci_get_pdn(pdev); 1511 1512 if (phb->type == PNV_PHB_IODA2) { 1513 /* Calculate available PE for required VFs */ 1514 mutex_lock(&phb->ioda.pe_alloc_mutex); 1515 pdn->offset = bitmap_find_next_zero_area( 1516 phb->ioda.pe_alloc, phb->ioda.total_pe, 1517 0, num_vfs, 0); 1518 if (pdn->offset >= phb->ioda.total_pe) { 1519 mutex_unlock(&phb->ioda.pe_alloc_mutex); 1520 dev_info(&pdev->dev, "Failed to enable VF%d\n", num_vfs); 1521 pdn->offset = 0; 1522 return -EBUSY; 1523 } 1524 bitmap_set(phb->ioda.pe_alloc, pdn->offset, num_vfs); 1525 pdn->num_vfs = num_vfs; 1526 mutex_unlock(&phb->ioda.pe_alloc_mutex); 1527 1528 /* Assign M64 window accordingly */ 1529 ret = pnv_pci_vf_assign_m64(pdev, num_vfs); 1530 if (ret) { 1531 dev_info(&pdev->dev, "Not enough M64 window resources\n"); 1532 goto m64_failed; 1533 } 1534 1535 /* 1536 * When using one M64 BAR to map one IOV BAR, we need to shift 1537 * the IOV BAR according to the PE# allocated to the VFs. 1538 * Otherwise, the PE# for the VF will conflict with others. 1539 */ 1540 if (pdn->m64_per_iov == 1) { 1541 ret = pnv_pci_vf_resource_shift(pdev, pdn->offset); 1542 if (ret) 1543 goto m64_failed; 1544 } 1545 } 1546 1547 /* Setup VF PEs */ 1548 pnv_ioda_setup_vf_PE(pdev, num_vfs); 1549 1550 return 0; 1551 1552 m64_failed: 1553 bitmap_clear(phb->ioda.pe_alloc, pdn->offset, num_vfs); 1554 pdn->offset = 0; 1555 1556 return ret; 1557 } 1558 1559 int pcibios_sriov_disable(struct pci_dev *pdev) 1560 { 1561 pnv_pci_sriov_disable(pdev); 1562 1563 /* Release PCI data */ 1564 remove_dev_pci_data(pdev); 1565 return 0; 1566 } 1567 1568 int pcibios_sriov_enable(struct pci_dev *pdev, u16 num_vfs) 1569 { 1570 /* Allocate PCI data */ 1571 add_dev_pci_data(pdev); 1572 1573 pnv_pci_sriov_enable(pdev, num_vfs); 1574 return 0; 1575 } 1576 #endif /* CONFIG_PCI_IOV */ 1577 1578 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb *phb, struct pci_dev *pdev) 1579 { 1580 struct pci_dn *pdn = pci_get_pdn(pdev); 1581 struct pnv_ioda_pe *pe; 1582 1583 /* 1584 * The function can be called while the PE# 1585 * hasn't been assigned. Do nothing for the 1586 * case. 1587 */ 1588 if (!pdn || pdn->pe_number == IODA_INVALID_PE) 1589 return; 1590 1591 pe = &phb->ioda.pe_array[pdn->pe_number]; 1592 WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops); 1593 set_iommu_table_base(&pdev->dev, pe->table_group.tables[0]); 1594 /* 1595 * Note: iommu_add_device() will fail here as 1596 * for physical PE: the device is already added by now; 1597 * for virtual PE: sysfs entries are not ready yet and 1598 * tce_iommu_bus_notifier will add the device to a group later. 1599 */ 1600 } 1601 1602 static int pnv_pci_ioda_dma_set_mask(struct pci_dev *pdev, u64 dma_mask) 1603 { 1604 struct pci_controller *hose = pci_bus_to_host(pdev->bus); 1605 struct pnv_phb *phb = hose->private_data; 1606 struct pci_dn *pdn = pci_get_pdn(pdev); 1607 struct pnv_ioda_pe *pe; 1608 uint64_t top; 1609 bool bypass = false; 1610 1611 if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE)) 1612 return -ENODEV;; 1613 1614 pe = &phb->ioda.pe_array[pdn->pe_number]; 1615 if (pe->tce_bypass_enabled) { 1616 top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1; 1617 bypass = (dma_mask >= top); 1618 } 1619 1620 if (bypass) { 1621 dev_info(&pdev->dev, "Using 64-bit DMA iommu bypass\n"); 1622 set_dma_ops(&pdev->dev, &dma_direct_ops); 1623 set_dma_offset(&pdev->dev, pe->tce_bypass_base); 1624 } else { 1625 dev_info(&pdev->dev, "Using 32-bit DMA via iommu\n"); 1626 set_dma_ops(&pdev->dev, &dma_iommu_ops); 1627 set_iommu_table_base(&pdev->dev, pe->table_group.tables[0]); 1628 } 1629 *pdev->dev.dma_mask = dma_mask; 1630 return 0; 1631 } 1632 1633 static u64 pnv_pci_ioda_dma_get_required_mask(struct pnv_phb *phb, 1634 struct pci_dev *pdev) 1635 { 1636 struct pci_dn *pdn = pci_get_pdn(pdev); 1637 struct pnv_ioda_pe *pe; 1638 u64 end, mask; 1639 1640 if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE)) 1641 return 0; 1642 1643 pe = &phb->ioda.pe_array[pdn->pe_number]; 1644 if (!pe->tce_bypass_enabled) 1645 return __dma_get_required_mask(&pdev->dev); 1646 1647 1648 end = pe->tce_bypass_base + memblock_end_of_DRAM(); 1649 mask = 1ULL << (fls64(end) - 1); 1650 mask += mask - 1; 1651 1652 return mask; 1653 } 1654 1655 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe, 1656 struct pci_bus *bus) 1657 { 1658 struct pci_dev *dev; 1659 1660 list_for_each_entry(dev, &bus->devices, bus_list) { 1661 set_iommu_table_base(&dev->dev, pe->table_group.tables[0]); 1662 iommu_add_device(&dev->dev); 1663 1664 if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate) 1665 pnv_ioda_setup_bus_dma(pe, dev->subordinate); 1666 } 1667 } 1668 1669 static void pnv_pci_ioda1_tce_invalidate(struct iommu_table *tbl, 1670 unsigned long index, unsigned long npages, bool rm) 1671 { 1672 struct iommu_table_group_link *tgl = list_first_entry_or_null( 1673 &tbl->it_group_list, struct iommu_table_group_link, 1674 next); 1675 struct pnv_ioda_pe *pe = container_of(tgl->table_group, 1676 struct pnv_ioda_pe, table_group); 1677 __be64 __iomem *invalidate = rm ? 1678 (__be64 __iomem *)pe->phb->ioda.tce_inval_reg_phys : 1679 pe->phb->ioda.tce_inval_reg; 1680 unsigned long start, end, inc; 1681 const unsigned shift = tbl->it_page_shift; 1682 1683 start = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset); 1684 end = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset + 1685 npages - 1); 1686 1687 /* BML uses this case for p6/p7/galaxy2: Shift addr and put in node */ 1688 if (tbl->it_busno) { 1689 start <<= shift; 1690 end <<= shift; 1691 inc = 128ull << shift; 1692 start |= tbl->it_busno; 1693 end |= tbl->it_busno; 1694 } else if (tbl->it_type & TCE_PCI_SWINV_PAIR) { 1695 /* p7ioc-style invalidation, 2 TCEs per write */ 1696 start |= (1ull << 63); 1697 end |= (1ull << 63); 1698 inc = 16; 1699 } else { 1700 /* Default (older HW) */ 1701 inc = 128; 1702 } 1703 1704 end |= inc - 1; /* round up end to be different than start */ 1705 1706 mb(); /* Ensure above stores are visible */ 1707 while (start <= end) { 1708 if (rm) 1709 __raw_rm_writeq(cpu_to_be64(start), invalidate); 1710 else 1711 __raw_writeq(cpu_to_be64(start), invalidate); 1712 start += inc; 1713 } 1714 1715 /* 1716 * The iommu layer will do another mb() for us on build() 1717 * and we don't care on free() 1718 */ 1719 } 1720 1721 static int pnv_ioda1_tce_build(struct iommu_table *tbl, long index, 1722 long npages, unsigned long uaddr, 1723 enum dma_data_direction direction, 1724 struct dma_attrs *attrs) 1725 { 1726 int ret = pnv_tce_build(tbl, index, npages, uaddr, direction, 1727 attrs); 1728 1729 if (!ret && (tbl->it_type & TCE_PCI_SWINV_CREATE)) 1730 pnv_pci_ioda1_tce_invalidate(tbl, index, npages, false); 1731 1732 return ret; 1733 } 1734 1735 #ifdef CONFIG_IOMMU_API 1736 static int pnv_ioda1_tce_xchg(struct iommu_table *tbl, long index, 1737 unsigned long *hpa, enum dma_data_direction *direction) 1738 { 1739 long ret = pnv_tce_xchg(tbl, index, hpa, direction); 1740 1741 if (!ret && (tbl->it_type & 1742 (TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE))) 1743 pnv_pci_ioda1_tce_invalidate(tbl, index, 1, false); 1744 1745 return ret; 1746 } 1747 #endif 1748 1749 static void pnv_ioda1_tce_free(struct iommu_table *tbl, long index, 1750 long npages) 1751 { 1752 pnv_tce_free(tbl, index, npages); 1753 1754 if (tbl->it_type & TCE_PCI_SWINV_FREE) 1755 pnv_pci_ioda1_tce_invalidate(tbl, index, npages, false); 1756 } 1757 1758 static struct iommu_table_ops pnv_ioda1_iommu_ops = { 1759 .set = pnv_ioda1_tce_build, 1760 #ifdef CONFIG_IOMMU_API 1761 .exchange = pnv_ioda1_tce_xchg, 1762 #endif 1763 .clear = pnv_ioda1_tce_free, 1764 .get = pnv_tce_get, 1765 }; 1766 1767 static inline void pnv_pci_ioda2_tce_invalidate_entire(struct pnv_ioda_pe *pe) 1768 { 1769 /* 01xb - invalidate TCEs that match the specified PE# */ 1770 unsigned long val = (0x4ull << 60) | (pe->pe_number & 0xFF); 1771 struct pnv_phb *phb = pe->phb; 1772 1773 if (!phb->ioda.tce_inval_reg) 1774 return; 1775 1776 mb(); /* Ensure above stores are visible */ 1777 __raw_writeq(cpu_to_be64(val), phb->ioda.tce_inval_reg); 1778 } 1779 1780 static void pnv_pci_ioda2_do_tce_invalidate(unsigned pe_number, bool rm, 1781 __be64 __iomem *invalidate, unsigned shift, 1782 unsigned long index, unsigned long npages) 1783 { 1784 unsigned long start, end, inc; 1785 1786 /* We'll invalidate DMA address in PE scope */ 1787 start = 0x2ull << 60; 1788 start |= (pe_number & 0xFF); 1789 end = start; 1790 1791 /* Figure out the start, end and step */ 1792 start |= (index << shift); 1793 end |= ((index + npages - 1) << shift); 1794 inc = (0x1ull << shift); 1795 mb(); 1796 1797 while (start <= end) { 1798 if (rm) 1799 __raw_rm_writeq(cpu_to_be64(start), invalidate); 1800 else 1801 __raw_writeq(cpu_to_be64(start), invalidate); 1802 start += inc; 1803 } 1804 } 1805 1806 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table *tbl, 1807 unsigned long index, unsigned long npages, bool rm) 1808 { 1809 struct iommu_table_group_link *tgl; 1810 1811 list_for_each_entry_rcu(tgl, &tbl->it_group_list, next) { 1812 struct pnv_ioda_pe *pe = container_of(tgl->table_group, 1813 struct pnv_ioda_pe, table_group); 1814 __be64 __iomem *invalidate = rm ? 1815 (__be64 __iomem *)pe->phb->ioda.tce_inval_reg_phys : 1816 pe->phb->ioda.tce_inval_reg; 1817 1818 pnv_pci_ioda2_do_tce_invalidate(pe->pe_number, rm, 1819 invalidate, tbl->it_page_shift, 1820 index, npages); 1821 } 1822 } 1823 1824 static int pnv_ioda2_tce_build(struct iommu_table *tbl, long index, 1825 long npages, unsigned long uaddr, 1826 enum dma_data_direction direction, 1827 struct dma_attrs *attrs) 1828 { 1829 int ret = pnv_tce_build(tbl, index, npages, uaddr, direction, 1830 attrs); 1831 1832 if (!ret && (tbl->it_type & TCE_PCI_SWINV_CREATE)) 1833 pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false); 1834 1835 return ret; 1836 } 1837 1838 #ifdef CONFIG_IOMMU_API 1839 static int pnv_ioda2_tce_xchg(struct iommu_table *tbl, long index, 1840 unsigned long *hpa, enum dma_data_direction *direction) 1841 { 1842 long ret = pnv_tce_xchg(tbl, index, hpa, direction); 1843 1844 if (!ret && (tbl->it_type & 1845 (TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE))) 1846 pnv_pci_ioda2_tce_invalidate(tbl, index, 1, false); 1847 1848 return ret; 1849 } 1850 #endif 1851 1852 static void pnv_ioda2_tce_free(struct iommu_table *tbl, long index, 1853 long npages) 1854 { 1855 pnv_tce_free(tbl, index, npages); 1856 1857 if (tbl->it_type & TCE_PCI_SWINV_FREE) 1858 pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false); 1859 } 1860 1861 static void pnv_ioda2_table_free(struct iommu_table *tbl) 1862 { 1863 pnv_pci_ioda2_table_free_pages(tbl); 1864 iommu_free_table(tbl, "pnv"); 1865 } 1866 1867 static struct iommu_table_ops pnv_ioda2_iommu_ops = { 1868 .set = pnv_ioda2_tce_build, 1869 #ifdef CONFIG_IOMMU_API 1870 .exchange = pnv_ioda2_tce_xchg, 1871 #endif 1872 .clear = pnv_ioda2_tce_free, 1873 .get = pnv_tce_get, 1874 .free = pnv_ioda2_table_free, 1875 }; 1876 1877 static void pnv_pci_ioda_setup_dma_pe(struct pnv_phb *phb, 1878 struct pnv_ioda_pe *pe, unsigned int base, 1879 unsigned int segs) 1880 { 1881 1882 struct page *tce_mem = NULL; 1883 struct iommu_table *tbl; 1884 unsigned int i; 1885 int64_t rc; 1886 void *addr; 1887 1888 /* XXX FIXME: Handle 64-bit only DMA devices */ 1889 /* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */ 1890 /* XXX FIXME: Allocate multi-level tables on PHB3 */ 1891 1892 /* We shouldn't already have a 32-bit DMA associated */ 1893 if (WARN_ON(pe->tce32_seg >= 0)) 1894 return; 1895 1896 tbl = pnv_pci_table_alloc(phb->hose->node); 1897 iommu_register_group(&pe->table_group, phb->hose->global_number, 1898 pe->pe_number); 1899 pnv_pci_link_table_and_group(phb->hose->node, 0, tbl, &pe->table_group); 1900 1901 /* Grab a 32-bit TCE table */ 1902 pe->tce32_seg = base; 1903 pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n", 1904 (base << 28), ((base + segs) << 28) - 1); 1905 1906 /* XXX Currently, we allocate one big contiguous table for the 1907 * TCEs. We only really need one chunk per 256M of TCE space 1908 * (ie per segment) but that's an optimization for later, it 1909 * requires some added smarts with our get/put_tce implementation 1910 */ 1911 tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL, 1912 get_order(TCE32_TABLE_SIZE * segs)); 1913 if (!tce_mem) { 1914 pe_err(pe, " Failed to allocate a 32-bit TCE memory\n"); 1915 goto fail; 1916 } 1917 addr = page_address(tce_mem); 1918 memset(addr, 0, TCE32_TABLE_SIZE * segs); 1919 1920 /* Configure HW */ 1921 for (i = 0; i < segs; i++) { 1922 rc = opal_pci_map_pe_dma_window(phb->opal_id, 1923 pe->pe_number, 1924 base + i, 1, 1925 __pa(addr) + TCE32_TABLE_SIZE * i, 1926 TCE32_TABLE_SIZE, 0x1000); 1927 if (rc) { 1928 pe_err(pe, " Failed to configure 32-bit TCE table," 1929 " err %ld\n", rc); 1930 goto fail; 1931 } 1932 } 1933 1934 /* Setup linux iommu table */ 1935 pnv_pci_setup_iommu_table(tbl, addr, TCE32_TABLE_SIZE * segs, 1936 base << 28, IOMMU_PAGE_SHIFT_4K); 1937 1938 /* OPAL variant of P7IOC SW invalidated TCEs */ 1939 if (phb->ioda.tce_inval_reg) 1940 tbl->it_type |= (TCE_PCI_SWINV_CREATE | 1941 TCE_PCI_SWINV_FREE | 1942 TCE_PCI_SWINV_PAIR); 1943 1944 tbl->it_ops = &pnv_ioda1_iommu_ops; 1945 pe->table_group.tce32_start = tbl->it_offset << tbl->it_page_shift; 1946 pe->table_group.tce32_size = tbl->it_size << tbl->it_page_shift; 1947 iommu_init_table(tbl, phb->hose->node); 1948 1949 if (pe->flags & PNV_IODA_PE_DEV) { 1950 /* 1951 * Setting table base here only for carrying iommu_group 1952 * further down to let iommu_add_device() do the job. 1953 * pnv_pci_ioda_dma_dev_setup will override it later anyway. 1954 */ 1955 set_iommu_table_base(&pe->pdev->dev, tbl); 1956 iommu_add_device(&pe->pdev->dev); 1957 } else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)) 1958 pnv_ioda_setup_bus_dma(pe, pe->pbus); 1959 1960 return; 1961 fail: 1962 /* XXX Failure: Try to fallback to 64-bit only ? */ 1963 if (pe->tce32_seg >= 0) 1964 pe->tce32_seg = -1; 1965 if (tce_mem) 1966 __free_pages(tce_mem, get_order(TCE32_TABLE_SIZE * segs)); 1967 if (tbl) { 1968 pnv_pci_unlink_table_and_group(tbl, &pe->table_group); 1969 iommu_free_table(tbl, "pnv"); 1970 } 1971 } 1972 1973 static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group, 1974 int num, struct iommu_table *tbl) 1975 { 1976 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe, 1977 table_group); 1978 struct pnv_phb *phb = pe->phb; 1979 int64_t rc; 1980 const unsigned long size = tbl->it_indirect_levels ? 1981 tbl->it_level_size : tbl->it_size; 1982 const __u64 start_addr = tbl->it_offset << tbl->it_page_shift; 1983 const __u64 win_size = tbl->it_size << tbl->it_page_shift; 1984 1985 pe_info(pe, "Setting up window#%d %llx..%llx pg=%x\n", num, 1986 start_addr, start_addr + win_size - 1, 1987 IOMMU_PAGE_SIZE(tbl)); 1988 1989 /* 1990 * Map TCE table through TVT. The TVE index is the PE number 1991 * shifted by 1 bit for 32-bits DMA space. 1992 */ 1993 rc = opal_pci_map_pe_dma_window(phb->opal_id, 1994 pe->pe_number, 1995 (pe->pe_number << 1) + num, 1996 tbl->it_indirect_levels + 1, 1997 __pa(tbl->it_base), 1998 size << 3, 1999 IOMMU_PAGE_SIZE(tbl)); 2000 if (rc) { 2001 pe_err(pe, "Failed to configure TCE table, err %ld\n", rc); 2002 return rc; 2003 } 2004 2005 pnv_pci_link_table_and_group(phb->hose->node, num, 2006 tbl, &pe->table_group); 2007 pnv_pci_ioda2_tce_invalidate_entire(pe); 2008 2009 return 0; 2010 } 2011 2012 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable) 2013 { 2014 uint16_t window_id = (pe->pe_number << 1 ) + 1; 2015 int64_t rc; 2016 2017 pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis"); 2018 if (enable) { 2019 phys_addr_t top = memblock_end_of_DRAM(); 2020 2021 top = roundup_pow_of_two(top); 2022 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id, 2023 pe->pe_number, 2024 window_id, 2025 pe->tce_bypass_base, 2026 top); 2027 } else { 2028 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id, 2029 pe->pe_number, 2030 window_id, 2031 pe->tce_bypass_base, 2032 0); 2033 } 2034 if (rc) 2035 pe_err(pe, "OPAL error %lld configuring bypass window\n", rc); 2036 else 2037 pe->tce_bypass_enabled = enable; 2038 } 2039 2040 static long pnv_pci_ioda2_table_alloc_pages(int nid, __u64 bus_offset, 2041 __u32 page_shift, __u64 window_size, __u32 levels, 2042 struct iommu_table *tbl); 2043 2044 static long pnv_pci_ioda2_create_table(struct iommu_table_group *table_group, 2045 int num, __u32 page_shift, __u64 window_size, __u32 levels, 2046 struct iommu_table **ptbl) 2047 { 2048 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe, 2049 table_group); 2050 int nid = pe->phb->hose->node; 2051 __u64 bus_offset = num ? pe->tce_bypass_base : table_group->tce32_start; 2052 long ret; 2053 struct iommu_table *tbl; 2054 2055 tbl = pnv_pci_table_alloc(nid); 2056 if (!tbl) 2057 return -ENOMEM; 2058 2059 ret = pnv_pci_ioda2_table_alloc_pages(nid, 2060 bus_offset, page_shift, window_size, 2061 levels, tbl); 2062 if (ret) { 2063 iommu_free_table(tbl, "pnv"); 2064 return ret; 2065 } 2066 2067 tbl->it_ops = &pnv_ioda2_iommu_ops; 2068 if (pe->phb->ioda.tce_inval_reg) 2069 tbl->it_type |= (TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE); 2070 2071 *ptbl = tbl; 2072 2073 return 0; 2074 } 2075 2076 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe *pe) 2077 { 2078 struct iommu_table *tbl = NULL; 2079 long rc; 2080 2081 rc = pnv_pci_ioda2_create_table(&pe->table_group, 0, 2082 IOMMU_PAGE_SHIFT_4K, 2083 pe->table_group.tce32_size, 2084 POWERNV_IOMMU_DEFAULT_LEVELS, &tbl); 2085 if (rc) { 2086 pe_err(pe, "Failed to create 32-bit TCE table, err %ld", 2087 rc); 2088 return rc; 2089 } 2090 2091 iommu_init_table(tbl, pe->phb->hose->node); 2092 2093 rc = pnv_pci_ioda2_set_window(&pe->table_group, 0, tbl); 2094 if (rc) { 2095 pe_err(pe, "Failed to configure 32-bit TCE table, err %ld\n", 2096 rc); 2097 pnv_ioda2_table_free(tbl); 2098 return rc; 2099 } 2100 2101 if (!pnv_iommu_bypass_disabled) 2102 pnv_pci_ioda2_set_bypass(pe, true); 2103 2104 /* OPAL variant of PHB3 invalidated TCEs */ 2105 if (pe->phb->ioda.tce_inval_reg) 2106 tbl->it_type |= (TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE); 2107 2108 /* 2109 * Setting table base here only for carrying iommu_group 2110 * further down to let iommu_add_device() do the job. 2111 * pnv_pci_ioda_dma_dev_setup will override it later anyway. 2112 */ 2113 if (pe->flags & PNV_IODA_PE_DEV) 2114 set_iommu_table_base(&pe->pdev->dev, tbl); 2115 2116 return 0; 2117 } 2118 2119 #if defined(CONFIG_IOMMU_API) || defined(CONFIG_PCI_IOV) 2120 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group, 2121 int num) 2122 { 2123 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe, 2124 table_group); 2125 struct pnv_phb *phb = pe->phb; 2126 long ret; 2127 2128 pe_info(pe, "Removing DMA window #%d\n", num); 2129 2130 ret = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number, 2131 (pe->pe_number << 1) + num, 2132 0/* levels */, 0/* table address */, 2133 0/* table size */, 0/* page size */); 2134 if (ret) 2135 pe_warn(pe, "Unmapping failed, ret = %ld\n", ret); 2136 else 2137 pnv_pci_ioda2_tce_invalidate_entire(pe); 2138 2139 pnv_pci_unlink_table_and_group(table_group->tables[num], table_group); 2140 2141 return ret; 2142 } 2143 #endif 2144 2145 #ifdef CONFIG_IOMMU_API 2146 static unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift, 2147 __u64 window_size, __u32 levels) 2148 { 2149 unsigned long bytes = 0; 2150 const unsigned window_shift = ilog2(window_size); 2151 unsigned entries_shift = window_shift - page_shift; 2152 unsigned table_shift = entries_shift + 3; 2153 unsigned long tce_table_size = max(0x1000UL, 1UL << table_shift); 2154 unsigned long direct_table_size; 2155 2156 if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS) || 2157 (window_size > memory_hotplug_max()) || 2158 !is_power_of_2(window_size)) 2159 return 0; 2160 2161 /* Calculate a direct table size from window_size and levels */ 2162 entries_shift = (entries_shift + levels - 1) / levels; 2163 table_shift = entries_shift + 3; 2164 table_shift = max_t(unsigned, table_shift, PAGE_SHIFT); 2165 direct_table_size = 1UL << table_shift; 2166 2167 for ( ; levels; --levels) { 2168 bytes += _ALIGN_UP(tce_table_size, direct_table_size); 2169 2170 tce_table_size /= direct_table_size; 2171 tce_table_size <<= 3; 2172 tce_table_size = _ALIGN_UP(tce_table_size, direct_table_size); 2173 } 2174 2175 return bytes; 2176 } 2177 2178 static void pnv_ioda2_take_ownership(struct iommu_table_group *table_group) 2179 { 2180 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe, 2181 table_group); 2182 /* Store @tbl as pnv_pci_ioda2_unset_window() resets it */ 2183 struct iommu_table *tbl = pe->table_group.tables[0]; 2184 2185 pnv_pci_ioda2_set_bypass(pe, false); 2186 pnv_pci_ioda2_unset_window(&pe->table_group, 0); 2187 pnv_ioda2_table_free(tbl); 2188 } 2189 2190 static void pnv_ioda2_release_ownership(struct iommu_table_group *table_group) 2191 { 2192 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe, 2193 table_group); 2194 2195 pnv_pci_ioda2_setup_default_config(pe); 2196 } 2197 2198 static struct iommu_table_group_ops pnv_pci_ioda2_ops = { 2199 .get_table_size = pnv_pci_ioda2_get_table_size, 2200 .create_table = pnv_pci_ioda2_create_table, 2201 .set_window = pnv_pci_ioda2_set_window, 2202 .unset_window = pnv_pci_ioda2_unset_window, 2203 .take_ownership = pnv_ioda2_take_ownership, 2204 .release_ownership = pnv_ioda2_release_ownership, 2205 }; 2206 #endif 2207 2208 static void pnv_pci_ioda_setup_opal_tce_kill(struct pnv_phb *phb) 2209 { 2210 const __be64 *swinvp; 2211 2212 /* OPAL variant of PHB3 invalidated TCEs */ 2213 swinvp = of_get_property(phb->hose->dn, "ibm,opal-tce-kill", NULL); 2214 if (!swinvp) 2215 return; 2216 2217 phb->ioda.tce_inval_reg_phys = be64_to_cpup(swinvp); 2218 phb->ioda.tce_inval_reg = ioremap(phb->ioda.tce_inval_reg_phys, 8); 2219 } 2220 2221 static __be64 *pnv_pci_ioda2_table_do_alloc_pages(int nid, unsigned shift, 2222 unsigned levels, unsigned long limit, 2223 unsigned long *current_offset) 2224 { 2225 struct page *tce_mem = NULL; 2226 __be64 *addr, *tmp; 2227 unsigned order = max_t(unsigned, shift, PAGE_SHIFT) - PAGE_SHIFT; 2228 unsigned long allocated = 1UL << (order + PAGE_SHIFT); 2229 unsigned entries = 1UL << (shift - 3); 2230 long i; 2231 2232 tce_mem = alloc_pages_node(nid, GFP_KERNEL, order); 2233 if (!tce_mem) { 2234 pr_err("Failed to allocate a TCE memory, order=%d\n", order); 2235 return NULL; 2236 } 2237 addr = page_address(tce_mem); 2238 memset(addr, 0, allocated); 2239 2240 --levels; 2241 if (!levels) { 2242 *current_offset += allocated; 2243 return addr; 2244 } 2245 2246 for (i = 0; i < entries; ++i) { 2247 tmp = pnv_pci_ioda2_table_do_alloc_pages(nid, shift, 2248 levels, limit, current_offset); 2249 if (!tmp) 2250 break; 2251 2252 addr[i] = cpu_to_be64(__pa(tmp) | 2253 TCE_PCI_READ | TCE_PCI_WRITE); 2254 2255 if (*current_offset >= limit) 2256 break; 2257 } 2258 2259 return addr; 2260 } 2261 2262 static void pnv_pci_ioda2_table_do_free_pages(__be64 *addr, 2263 unsigned long size, unsigned level); 2264 2265 static long pnv_pci_ioda2_table_alloc_pages(int nid, __u64 bus_offset, 2266 __u32 page_shift, __u64 window_size, __u32 levels, 2267 struct iommu_table *tbl) 2268 { 2269 void *addr; 2270 unsigned long offset = 0, level_shift; 2271 const unsigned window_shift = ilog2(window_size); 2272 unsigned entries_shift = window_shift - page_shift; 2273 unsigned table_shift = max_t(unsigned, entries_shift + 3, PAGE_SHIFT); 2274 const unsigned long tce_table_size = 1UL << table_shift; 2275 2276 if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS)) 2277 return -EINVAL; 2278 2279 if ((window_size > memory_hotplug_max()) || !is_power_of_2(window_size)) 2280 return -EINVAL; 2281 2282 /* Adjust direct table size from window_size and levels */ 2283 entries_shift = (entries_shift + levels - 1) / levels; 2284 level_shift = entries_shift + 3; 2285 level_shift = max_t(unsigned, level_shift, PAGE_SHIFT); 2286 2287 /* Allocate TCE table */ 2288 addr = pnv_pci_ioda2_table_do_alloc_pages(nid, level_shift, 2289 levels, tce_table_size, &offset); 2290 2291 /* addr==NULL means that the first level allocation failed */ 2292 if (!addr) 2293 return -ENOMEM; 2294 2295 /* 2296 * First level was allocated but some lower level failed as 2297 * we did not allocate as much as we wanted, 2298 * release partially allocated table. 2299 */ 2300 if (offset < tce_table_size) { 2301 pnv_pci_ioda2_table_do_free_pages(addr, 2302 1ULL << (level_shift - 3), levels - 1); 2303 return -ENOMEM; 2304 } 2305 2306 /* Setup linux iommu table */ 2307 pnv_pci_setup_iommu_table(tbl, addr, tce_table_size, bus_offset, 2308 page_shift); 2309 tbl->it_level_size = 1ULL << (level_shift - 3); 2310 tbl->it_indirect_levels = levels - 1; 2311 tbl->it_allocated_size = offset; 2312 2313 pr_devel("Created TCE table: ws=%08llx ts=%lx @%08llx\n", 2314 window_size, tce_table_size, bus_offset); 2315 2316 return 0; 2317 } 2318 2319 static void pnv_pci_ioda2_table_do_free_pages(__be64 *addr, 2320 unsigned long size, unsigned level) 2321 { 2322 const unsigned long addr_ul = (unsigned long) addr & 2323 ~(TCE_PCI_READ | TCE_PCI_WRITE); 2324 2325 if (level) { 2326 long i; 2327 u64 *tmp = (u64 *) addr_ul; 2328 2329 for (i = 0; i < size; ++i) { 2330 unsigned long hpa = be64_to_cpu(tmp[i]); 2331 2332 if (!(hpa & (TCE_PCI_READ | TCE_PCI_WRITE))) 2333 continue; 2334 2335 pnv_pci_ioda2_table_do_free_pages(__va(hpa), size, 2336 level - 1); 2337 } 2338 } 2339 2340 free_pages(addr_ul, get_order(size << 3)); 2341 } 2342 2343 static void pnv_pci_ioda2_table_free_pages(struct iommu_table *tbl) 2344 { 2345 const unsigned long size = tbl->it_indirect_levels ? 2346 tbl->it_level_size : tbl->it_size; 2347 2348 if (!tbl->it_size) 2349 return; 2350 2351 pnv_pci_ioda2_table_do_free_pages((__be64 *)tbl->it_base, size, 2352 tbl->it_indirect_levels); 2353 } 2354 2355 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb, 2356 struct pnv_ioda_pe *pe) 2357 { 2358 int64_t rc; 2359 2360 /* We shouldn't already have a 32-bit DMA associated */ 2361 if (WARN_ON(pe->tce32_seg >= 0)) 2362 return; 2363 2364 /* TVE #1 is selected by PCI address bit 59 */ 2365 pe->tce_bypass_base = 1ull << 59; 2366 2367 iommu_register_group(&pe->table_group, phb->hose->global_number, 2368 pe->pe_number); 2369 2370 /* The PE will reserve all possible 32-bits space */ 2371 pe->tce32_seg = 0; 2372 pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n", 2373 phb->ioda.m32_pci_base); 2374 2375 /* Setup linux iommu table */ 2376 pe->table_group.tce32_start = 0; 2377 pe->table_group.tce32_size = phb->ioda.m32_pci_base; 2378 pe->table_group.max_dynamic_windows_supported = 2379 IOMMU_TABLE_GROUP_MAX_TABLES; 2380 pe->table_group.max_levels = POWERNV_IOMMU_MAX_LEVELS; 2381 pe->table_group.pgsizes = SZ_4K | SZ_64K | SZ_16M; 2382 #ifdef CONFIG_IOMMU_API 2383 pe->table_group.ops = &pnv_pci_ioda2_ops; 2384 #endif 2385 2386 rc = pnv_pci_ioda2_setup_default_config(pe); 2387 if (rc) { 2388 if (pe->tce32_seg >= 0) 2389 pe->tce32_seg = -1; 2390 return; 2391 } 2392 2393 if (pe->flags & PNV_IODA_PE_DEV) 2394 iommu_add_device(&pe->pdev->dev); 2395 else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)) 2396 pnv_ioda_setup_bus_dma(pe, pe->pbus); 2397 } 2398 2399 static void pnv_ioda_setup_dma(struct pnv_phb *phb) 2400 { 2401 struct pci_controller *hose = phb->hose; 2402 unsigned int residual, remaining, segs, tw, base; 2403 struct pnv_ioda_pe *pe; 2404 2405 /* If we have more PE# than segments available, hand out one 2406 * per PE until we run out and let the rest fail. If not, 2407 * then we assign at least one segment per PE, plus more based 2408 * on the amount of devices under that PE 2409 */ 2410 if (phb->ioda.dma_pe_count > phb->ioda.tce32_count) 2411 residual = 0; 2412 else 2413 residual = phb->ioda.tce32_count - 2414 phb->ioda.dma_pe_count; 2415 2416 pr_info("PCI: Domain %04x has %ld available 32-bit DMA segments\n", 2417 hose->global_number, phb->ioda.tce32_count); 2418 pr_info("PCI: %d PE# for a total weight of %d\n", 2419 phb->ioda.dma_pe_count, phb->ioda.dma_weight); 2420 2421 pnv_pci_ioda_setup_opal_tce_kill(phb); 2422 2423 /* Walk our PE list and configure their DMA segments, hand them 2424 * out one base segment plus any residual segments based on 2425 * weight 2426 */ 2427 remaining = phb->ioda.tce32_count; 2428 tw = phb->ioda.dma_weight; 2429 base = 0; 2430 list_for_each_entry(pe, &phb->ioda.pe_dma_list, dma_link) { 2431 if (!pe->dma_weight) 2432 continue; 2433 if (!remaining) { 2434 pe_warn(pe, "No DMA32 resources available\n"); 2435 continue; 2436 } 2437 segs = 1; 2438 if (residual) { 2439 segs += ((pe->dma_weight * residual) + (tw / 2)) / tw; 2440 if (segs > remaining) 2441 segs = remaining; 2442 } 2443 2444 /* 2445 * For IODA2 compliant PHB3, we needn't care about the weight. 2446 * The all available 32-bits DMA space will be assigned to 2447 * the specific PE. 2448 */ 2449 if (phb->type == PNV_PHB_IODA1) { 2450 pe_info(pe, "DMA weight %d, assigned %d DMA32 segments\n", 2451 pe->dma_weight, segs); 2452 pnv_pci_ioda_setup_dma_pe(phb, pe, base, segs); 2453 } else { 2454 pe_info(pe, "Assign DMA32 space\n"); 2455 segs = 0; 2456 pnv_pci_ioda2_setup_dma_pe(phb, pe); 2457 } 2458 2459 remaining -= segs; 2460 base += segs; 2461 } 2462 } 2463 2464 #ifdef CONFIG_PCI_MSI 2465 static void pnv_ioda2_msi_eoi(struct irq_data *d) 2466 { 2467 unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d); 2468 struct irq_chip *chip = irq_data_get_irq_chip(d); 2469 struct pnv_phb *phb = container_of(chip, struct pnv_phb, 2470 ioda.irq_chip); 2471 int64_t rc; 2472 2473 rc = opal_pci_msi_eoi(phb->opal_id, hw_irq); 2474 WARN_ON_ONCE(rc); 2475 2476 icp_native_eoi(d); 2477 } 2478 2479 2480 static void set_msi_irq_chip(struct pnv_phb *phb, unsigned int virq) 2481 { 2482 struct irq_data *idata; 2483 struct irq_chip *ichip; 2484 2485 if (phb->type != PNV_PHB_IODA2) 2486 return; 2487 2488 if (!phb->ioda.irq_chip_init) { 2489 /* 2490 * First time we setup an MSI IRQ, we need to setup the 2491 * corresponding IRQ chip to route correctly. 2492 */ 2493 idata = irq_get_irq_data(virq); 2494 ichip = irq_data_get_irq_chip(idata); 2495 phb->ioda.irq_chip_init = 1; 2496 phb->ioda.irq_chip = *ichip; 2497 phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi; 2498 } 2499 irq_set_chip(virq, &phb->ioda.irq_chip); 2500 } 2501 2502 #ifdef CONFIG_CXL_BASE 2503 2504 struct device_node *pnv_pci_get_phb_node(struct pci_dev *dev) 2505 { 2506 struct pci_controller *hose = pci_bus_to_host(dev->bus); 2507 2508 return of_node_get(hose->dn); 2509 } 2510 EXPORT_SYMBOL(pnv_pci_get_phb_node); 2511 2512 int pnv_phb_to_cxl_mode(struct pci_dev *dev, uint64_t mode) 2513 { 2514 struct pci_controller *hose = pci_bus_to_host(dev->bus); 2515 struct pnv_phb *phb = hose->private_data; 2516 struct pnv_ioda_pe *pe; 2517 int rc; 2518 2519 pe = pnv_ioda_get_pe(dev); 2520 if (!pe) 2521 return -ENODEV; 2522 2523 pe_info(pe, "Switching PHB to CXL\n"); 2524 2525 rc = opal_pci_set_phb_cxl_mode(phb->opal_id, mode, pe->pe_number); 2526 if (rc) 2527 dev_err(&dev->dev, "opal_pci_set_phb_cxl_mode failed: %i\n", rc); 2528 2529 return rc; 2530 } 2531 EXPORT_SYMBOL(pnv_phb_to_cxl_mode); 2532 2533 /* Find PHB for cxl dev and allocate MSI hwirqs? 2534 * Returns the absolute hardware IRQ number 2535 */ 2536 int pnv_cxl_alloc_hwirqs(struct pci_dev *dev, int num) 2537 { 2538 struct pci_controller *hose = pci_bus_to_host(dev->bus); 2539 struct pnv_phb *phb = hose->private_data; 2540 int hwirq = msi_bitmap_alloc_hwirqs(&phb->msi_bmp, num); 2541 2542 if (hwirq < 0) { 2543 dev_warn(&dev->dev, "Failed to find a free MSI\n"); 2544 return -ENOSPC; 2545 } 2546 2547 return phb->msi_base + hwirq; 2548 } 2549 EXPORT_SYMBOL(pnv_cxl_alloc_hwirqs); 2550 2551 void pnv_cxl_release_hwirqs(struct pci_dev *dev, int hwirq, int num) 2552 { 2553 struct pci_controller *hose = pci_bus_to_host(dev->bus); 2554 struct pnv_phb *phb = hose->private_data; 2555 2556 msi_bitmap_free_hwirqs(&phb->msi_bmp, hwirq - phb->msi_base, num); 2557 } 2558 EXPORT_SYMBOL(pnv_cxl_release_hwirqs); 2559 2560 void pnv_cxl_release_hwirq_ranges(struct cxl_irq_ranges *irqs, 2561 struct pci_dev *dev) 2562 { 2563 struct pci_controller *hose = pci_bus_to_host(dev->bus); 2564 struct pnv_phb *phb = hose->private_data; 2565 int i, hwirq; 2566 2567 for (i = 1; i < CXL_IRQ_RANGES; i++) { 2568 if (!irqs->range[i]) 2569 continue; 2570 pr_devel("cxl release irq range 0x%x: offset: 0x%lx limit: %ld\n", 2571 i, irqs->offset[i], 2572 irqs->range[i]); 2573 hwirq = irqs->offset[i] - phb->msi_base; 2574 msi_bitmap_free_hwirqs(&phb->msi_bmp, hwirq, 2575 irqs->range[i]); 2576 } 2577 } 2578 EXPORT_SYMBOL(pnv_cxl_release_hwirq_ranges); 2579 2580 int pnv_cxl_alloc_hwirq_ranges(struct cxl_irq_ranges *irqs, 2581 struct pci_dev *dev, int num) 2582 { 2583 struct pci_controller *hose = pci_bus_to_host(dev->bus); 2584 struct pnv_phb *phb = hose->private_data; 2585 int i, hwirq, try; 2586 2587 memset(irqs, 0, sizeof(struct cxl_irq_ranges)); 2588 2589 /* 0 is reserved for the multiplexed PSL DSI interrupt */ 2590 for (i = 1; i < CXL_IRQ_RANGES && num; i++) { 2591 try = num; 2592 while (try) { 2593 hwirq = msi_bitmap_alloc_hwirqs(&phb->msi_bmp, try); 2594 if (hwirq >= 0) 2595 break; 2596 try /= 2; 2597 } 2598 if (!try) 2599 goto fail; 2600 2601 irqs->offset[i] = phb->msi_base + hwirq; 2602 irqs->range[i] = try; 2603 pr_devel("cxl alloc irq range 0x%x: offset: 0x%lx limit: %li\n", 2604 i, irqs->offset[i], irqs->range[i]); 2605 num -= try; 2606 } 2607 if (num) 2608 goto fail; 2609 2610 return 0; 2611 fail: 2612 pnv_cxl_release_hwirq_ranges(irqs, dev); 2613 return -ENOSPC; 2614 } 2615 EXPORT_SYMBOL(pnv_cxl_alloc_hwirq_ranges); 2616 2617 int pnv_cxl_get_irq_count(struct pci_dev *dev) 2618 { 2619 struct pci_controller *hose = pci_bus_to_host(dev->bus); 2620 struct pnv_phb *phb = hose->private_data; 2621 2622 return phb->msi_bmp.irq_count; 2623 } 2624 EXPORT_SYMBOL(pnv_cxl_get_irq_count); 2625 2626 int pnv_cxl_ioda_msi_setup(struct pci_dev *dev, unsigned int hwirq, 2627 unsigned int virq) 2628 { 2629 struct pci_controller *hose = pci_bus_to_host(dev->bus); 2630 struct pnv_phb *phb = hose->private_data; 2631 unsigned int xive_num = hwirq - phb->msi_base; 2632 struct pnv_ioda_pe *pe; 2633 int rc; 2634 2635 if (!(pe = pnv_ioda_get_pe(dev))) 2636 return -ENODEV; 2637 2638 /* Assign XIVE to PE */ 2639 rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num); 2640 if (rc) { 2641 pe_warn(pe, "%s: OPAL error %d setting msi_base 0x%x " 2642 "hwirq 0x%x XIVE 0x%x PE\n", 2643 pci_name(dev), rc, phb->msi_base, hwirq, xive_num); 2644 return -EIO; 2645 } 2646 set_msi_irq_chip(phb, virq); 2647 2648 return 0; 2649 } 2650 EXPORT_SYMBOL(pnv_cxl_ioda_msi_setup); 2651 #endif 2652 2653 static int pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev, 2654 unsigned int hwirq, unsigned int virq, 2655 unsigned int is_64, struct msi_msg *msg) 2656 { 2657 struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev); 2658 unsigned int xive_num = hwirq - phb->msi_base; 2659 __be32 data; 2660 int rc; 2661 2662 /* No PE assigned ? bail out ... no MSI for you ! */ 2663 if (pe == NULL) 2664 return -ENXIO; 2665 2666 /* Check if we have an MVE */ 2667 if (pe->mve_number < 0) 2668 return -ENXIO; 2669 2670 /* Force 32-bit MSI on some broken devices */ 2671 if (dev->no_64bit_msi) 2672 is_64 = 0; 2673 2674 /* Assign XIVE to PE */ 2675 rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num); 2676 if (rc) { 2677 pr_warn("%s: OPAL error %d setting XIVE %d PE\n", 2678 pci_name(dev), rc, xive_num); 2679 return -EIO; 2680 } 2681 2682 if (is_64) { 2683 __be64 addr64; 2684 2685 rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1, 2686 &addr64, &data); 2687 if (rc) { 2688 pr_warn("%s: OPAL error %d getting 64-bit MSI data\n", 2689 pci_name(dev), rc); 2690 return -EIO; 2691 } 2692 msg->address_hi = be64_to_cpu(addr64) >> 32; 2693 msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful; 2694 } else { 2695 __be32 addr32; 2696 2697 rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1, 2698 &addr32, &data); 2699 if (rc) { 2700 pr_warn("%s: OPAL error %d getting 32-bit MSI data\n", 2701 pci_name(dev), rc); 2702 return -EIO; 2703 } 2704 msg->address_hi = 0; 2705 msg->address_lo = be32_to_cpu(addr32); 2706 } 2707 msg->data = be32_to_cpu(data); 2708 2709 set_msi_irq_chip(phb, virq); 2710 2711 pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d)," 2712 " address=%x_%08x data=%x PE# %d\n", 2713 pci_name(dev), is_64 ? "64" : "32", hwirq, xive_num, 2714 msg->address_hi, msg->address_lo, data, pe->pe_number); 2715 2716 return 0; 2717 } 2718 2719 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb) 2720 { 2721 unsigned int count; 2722 const __be32 *prop = of_get_property(phb->hose->dn, 2723 "ibm,opal-msi-ranges", NULL); 2724 if (!prop) { 2725 /* BML Fallback */ 2726 prop = of_get_property(phb->hose->dn, "msi-ranges", NULL); 2727 } 2728 if (!prop) 2729 return; 2730 2731 phb->msi_base = be32_to_cpup(prop); 2732 count = be32_to_cpup(prop + 1); 2733 if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) { 2734 pr_err("PCI %d: Failed to allocate MSI bitmap !\n", 2735 phb->hose->global_number); 2736 return; 2737 } 2738 2739 phb->msi_setup = pnv_pci_ioda_msi_setup; 2740 phb->msi32_support = 1; 2741 pr_info(" Allocated bitmap for %d MSIs (base IRQ 0x%x)\n", 2742 count, phb->msi_base); 2743 } 2744 #else 2745 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb) { } 2746 #endif /* CONFIG_PCI_MSI */ 2747 2748 #ifdef CONFIG_PCI_IOV 2749 static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev *pdev) 2750 { 2751 struct pci_controller *hose; 2752 struct pnv_phb *phb; 2753 struct resource *res; 2754 int i; 2755 resource_size_t size; 2756 struct pci_dn *pdn; 2757 int mul, total_vfs; 2758 2759 if (!pdev->is_physfn || pdev->is_added) 2760 return; 2761 2762 hose = pci_bus_to_host(pdev->bus); 2763 phb = hose->private_data; 2764 2765 pdn = pci_get_pdn(pdev); 2766 pdn->vfs_expanded = 0; 2767 2768 total_vfs = pci_sriov_get_totalvfs(pdev); 2769 pdn->m64_per_iov = 1; 2770 mul = phb->ioda.total_pe; 2771 2772 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { 2773 res = &pdev->resource[i + PCI_IOV_RESOURCES]; 2774 if (!res->flags || res->parent) 2775 continue; 2776 if (!pnv_pci_is_mem_pref_64(res->flags)) { 2777 dev_warn(&pdev->dev, " non M64 VF BAR%d: %pR\n", 2778 i, res); 2779 continue; 2780 } 2781 2782 size = pci_iov_resource_size(pdev, i + PCI_IOV_RESOURCES); 2783 2784 /* bigger than 64M */ 2785 if (size > (1 << 26)) { 2786 dev_info(&pdev->dev, "PowerNV: VF BAR%d: %pR IOV size is bigger than 64M, roundup power2\n", 2787 i, res); 2788 pdn->m64_per_iov = M64_PER_IOV; 2789 mul = roundup_pow_of_two(total_vfs); 2790 break; 2791 } 2792 } 2793 2794 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { 2795 res = &pdev->resource[i + PCI_IOV_RESOURCES]; 2796 if (!res->flags || res->parent) 2797 continue; 2798 if (!pnv_pci_is_mem_pref_64(res->flags)) { 2799 dev_warn(&pdev->dev, "Skipping expanding VF BAR%d: %pR\n", 2800 i, res); 2801 continue; 2802 } 2803 2804 dev_dbg(&pdev->dev, " Fixing VF BAR%d: %pR to\n", i, res); 2805 size = pci_iov_resource_size(pdev, i + PCI_IOV_RESOURCES); 2806 res->end = res->start + size * mul - 1; 2807 dev_dbg(&pdev->dev, " %pR\n", res); 2808 dev_info(&pdev->dev, "VF BAR%d: %pR (expanded to %d VFs for PE alignment)", 2809 i, res, mul); 2810 } 2811 pdn->vfs_expanded = mul; 2812 } 2813 #endif /* CONFIG_PCI_IOV */ 2814 2815 /* 2816 * This function is supposed to be called on basis of PE from top 2817 * to bottom style. So the the I/O or MMIO segment assigned to 2818 * parent PE could be overrided by its child PEs if necessary. 2819 */ 2820 static void pnv_ioda_setup_pe_seg(struct pci_controller *hose, 2821 struct pnv_ioda_pe *pe) 2822 { 2823 struct pnv_phb *phb = hose->private_data; 2824 struct pci_bus_region region; 2825 struct resource *res; 2826 int i, index; 2827 int rc; 2828 2829 /* 2830 * NOTE: We only care PCI bus based PE for now. For PCI 2831 * device based PE, for example SRIOV sensitive VF should 2832 * be figured out later. 2833 */ 2834 BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))); 2835 2836 pci_bus_for_each_resource(pe->pbus, res, i) { 2837 if (!res || !res->flags || 2838 res->start > res->end) 2839 continue; 2840 2841 if (res->flags & IORESOURCE_IO) { 2842 region.start = res->start - phb->ioda.io_pci_base; 2843 region.end = res->end - phb->ioda.io_pci_base; 2844 index = region.start / phb->ioda.io_segsize; 2845 2846 while (index < phb->ioda.total_pe && 2847 region.start <= region.end) { 2848 phb->ioda.io_segmap[index] = pe->pe_number; 2849 rc = opal_pci_map_pe_mmio_window(phb->opal_id, 2850 pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index); 2851 if (rc != OPAL_SUCCESS) { 2852 pr_err("%s: OPAL error %d when mapping IO " 2853 "segment #%d to PE#%d\n", 2854 __func__, rc, index, pe->pe_number); 2855 break; 2856 } 2857 2858 region.start += phb->ioda.io_segsize; 2859 index++; 2860 } 2861 } else if ((res->flags & IORESOURCE_MEM) && 2862 !pnv_pci_is_mem_pref_64(res->flags)) { 2863 region.start = res->start - 2864 hose->mem_offset[0] - 2865 phb->ioda.m32_pci_base; 2866 region.end = res->end - 2867 hose->mem_offset[0] - 2868 phb->ioda.m32_pci_base; 2869 index = region.start / phb->ioda.m32_segsize; 2870 2871 while (index < phb->ioda.total_pe && 2872 region.start <= region.end) { 2873 phb->ioda.m32_segmap[index] = pe->pe_number; 2874 rc = opal_pci_map_pe_mmio_window(phb->opal_id, 2875 pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index); 2876 if (rc != OPAL_SUCCESS) { 2877 pr_err("%s: OPAL error %d when mapping M32 " 2878 "segment#%d to PE#%d", 2879 __func__, rc, index, pe->pe_number); 2880 break; 2881 } 2882 2883 region.start += phb->ioda.m32_segsize; 2884 index++; 2885 } 2886 } 2887 } 2888 } 2889 2890 static void pnv_pci_ioda_setup_seg(void) 2891 { 2892 struct pci_controller *tmp, *hose; 2893 struct pnv_phb *phb; 2894 struct pnv_ioda_pe *pe; 2895 2896 list_for_each_entry_safe(hose, tmp, &hose_list, list_node) { 2897 phb = hose->private_data; 2898 list_for_each_entry(pe, &phb->ioda.pe_list, list) { 2899 pnv_ioda_setup_pe_seg(hose, pe); 2900 } 2901 } 2902 } 2903 2904 static void pnv_pci_ioda_setup_DMA(void) 2905 { 2906 struct pci_controller *hose, *tmp; 2907 struct pnv_phb *phb; 2908 2909 list_for_each_entry_safe(hose, tmp, &hose_list, list_node) { 2910 pnv_ioda_setup_dma(hose->private_data); 2911 2912 /* Mark the PHB initialization done */ 2913 phb = hose->private_data; 2914 phb->initialized = 1; 2915 } 2916 } 2917 2918 static void pnv_pci_ioda_create_dbgfs(void) 2919 { 2920 #ifdef CONFIG_DEBUG_FS 2921 struct pci_controller *hose, *tmp; 2922 struct pnv_phb *phb; 2923 char name[16]; 2924 2925 list_for_each_entry_safe(hose, tmp, &hose_list, list_node) { 2926 phb = hose->private_data; 2927 2928 sprintf(name, "PCI%04x", hose->global_number); 2929 phb->dbgfs = debugfs_create_dir(name, powerpc_debugfs_root); 2930 if (!phb->dbgfs) 2931 pr_warning("%s: Error on creating debugfs on PHB#%x\n", 2932 __func__, hose->global_number); 2933 } 2934 #endif /* CONFIG_DEBUG_FS */ 2935 } 2936 2937 static void pnv_pci_ioda_fixup(void) 2938 { 2939 pnv_pci_ioda_setup_PEs(); 2940 pnv_pci_ioda_setup_seg(); 2941 pnv_pci_ioda_setup_DMA(); 2942 2943 pnv_pci_ioda_create_dbgfs(); 2944 2945 #ifdef CONFIG_EEH 2946 eeh_init(); 2947 eeh_addr_cache_build(); 2948 #endif 2949 } 2950 2951 /* 2952 * Returns the alignment for I/O or memory windows for P2P 2953 * bridges. That actually depends on how PEs are segmented. 2954 * For now, we return I/O or M32 segment size for PE sensitive 2955 * P2P bridges. Otherwise, the default values (4KiB for I/O, 2956 * 1MiB for memory) will be returned. 2957 * 2958 * The current PCI bus might be put into one PE, which was 2959 * create against the parent PCI bridge. For that case, we 2960 * needn't enlarge the alignment so that we can save some 2961 * resources. 2962 */ 2963 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus, 2964 unsigned long type) 2965 { 2966 struct pci_dev *bridge; 2967 struct pci_controller *hose = pci_bus_to_host(bus); 2968 struct pnv_phb *phb = hose->private_data; 2969 int num_pci_bridges = 0; 2970 2971 bridge = bus->self; 2972 while (bridge) { 2973 if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) { 2974 num_pci_bridges++; 2975 if (num_pci_bridges >= 2) 2976 return 1; 2977 } 2978 2979 bridge = bridge->bus->self; 2980 } 2981 2982 /* We fail back to M32 if M64 isn't supported */ 2983 if (phb->ioda.m64_segsize && 2984 pnv_pci_is_mem_pref_64(type)) 2985 return phb->ioda.m64_segsize; 2986 if (type & IORESOURCE_MEM) 2987 return phb->ioda.m32_segsize; 2988 2989 return phb->ioda.io_segsize; 2990 } 2991 2992 #ifdef CONFIG_PCI_IOV 2993 static resource_size_t pnv_pci_iov_resource_alignment(struct pci_dev *pdev, 2994 int resno) 2995 { 2996 struct pci_dn *pdn = pci_get_pdn(pdev); 2997 resource_size_t align, iov_align; 2998 2999 iov_align = resource_size(&pdev->resource[resno]); 3000 if (iov_align) 3001 return iov_align; 3002 3003 align = pci_iov_resource_size(pdev, resno); 3004 if (pdn->vfs_expanded) 3005 return pdn->vfs_expanded * align; 3006 3007 return align; 3008 } 3009 #endif /* CONFIG_PCI_IOV */ 3010 3011 /* Prevent enabling devices for which we couldn't properly 3012 * assign a PE 3013 */ 3014 static bool pnv_pci_enable_device_hook(struct pci_dev *dev) 3015 { 3016 struct pci_controller *hose = pci_bus_to_host(dev->bus); 3017 struct pnv_phb *phb = hose->private_data; 3018 struct pci_dn *pdn; 3019 3020 /* The function is probably called while the PEs have 3021 * not be created yet. For example, resource reassignment 3022 * during PCI probe period. We just skip the check if 3023 * PEs isn't ready. 3024 */ 3025 if (!phb->initialized) 3026 return true; 3027 3028 pdn = pci_get_pdn(dev); 3029 if (!pdn || pdn->pe_number == IODA_INVALID_PE) 3030 return false; 3031 3032 return true; 3033 } 3034 3035 static u32 pnv_ioda_bdfn_to_pe(struct pnv_phb *phb, struct pci_bus *bus, 3036 u32 devfn) 3037 { 3038 return phb->ioda.pe_rmap[(bus->number << 8) | devfn]; 3039 } 3040 3041 static void pnv_pci_ioda_shutdown(struct pci_controller *hose) 3042 { 3043 struct pnv_phb *phb = hose->private_data; 3044 3045 opal_pci_reset(phb->opal_id, OPAL_RESET_PCI_IODA_TABLE, 3046 OPAL_ASSERT_RESET); 3047 } 3048 3049 static const struct pci_controller_ops pnv_pci_ioda_controller_ops = { 3050 .dma_dev_setup = pnv_pci_dma_dev_setup, 3051 #ifdef CONFIG_PCI_MSI 3052 .setup_msi_irqs = pnv_setup_msi_irqs, 3053 .teardown_msi_irqs = pnv_teardown_msi_irqs, 3054 #endif 3055 .enable_device_hook = pnv_pci_enable_device_hook, 3056 .window_alignment = pnv_pci_window_alignment, 3057 .reset_secondary_bus = pnv_pci_reset_secondary_bus, 3058 .dma_set_mask = pnv_pci_ioda_dma_set_mask, 3059 .shutdown = pnv_pci_ioda_shutdown, 3060 }; 3061 3062 static void __init pnv_pci_init_ioda_phb(struct device_node *np, 3063 u64 hub_id, int ioda_type) 3064 { 3065 struct pci_controller *hose; 3066 struct pnv_phb *phb; 3067 unsigned long size, m32map_off, pemap_off, iomap_off = 0; 3068 const __be64 *prop64; 3069 const __be32 *prop32; 3070 int len; 3071 u64 phb_id; 3072 void *aux; 3073 long rc; 3074 3075 pr_info("Initializing IODA%d OPAL PHB %s\n", ioda_type, np->full_name); 3076 3077 prop64 = of_get_property(np, "ibm,opal-phbid", NULL); 3078 if (!prop64) { 3079 pr_err(" Missing \"ibm,opal-phbid\" property !\n"); 3080 return; 3081 } 3082 phb_id = be64_to_cpup(prop64); 3083 pr_debug(" PHB-ID : 0x%016llx\n", phb_id); 3084 3085 phb = memblock_virt_alloc(sizeof(struct pnv_phb), 0); 3086 3087 /* Allocate PCI controller */ 3088 phb->hose = hose = pcibios_alloc_controller(np); 3089 if (!phb->hose) { 3090 pr_err(" Can't allocate PCI controller for %s\n", 3091 np->full_name); 3092 memblock_free(__pa(phb), sizeof(struct pnv_phb)); 3093 return; 3094 } 3095 3096 spin_lock_init(&phb->lock); 3097 prop32 = of_get_property(np, "bus-range", &len); 3098 if (prop32 && len == 8) { 3099 hose->first_busno = be32_to_cpu(prop32[0]); 3100 hose->last_busno = be32_to_cpu(prop32[1]); 3101 } else { 3102 pr_warn(" Broken <bus-range> on %s\n", np->full_name); 3103 hose->first_busno = 0; 3104 hose->last_busno = 0xff; 3105 } 3106 hose->private_data = phb; 3107 phb->hub_id = hub_id; 3108 phb->opal_id = phb_id; 3109 phb->type = ioda_type; 3110 mutex_init(&phb->ioda.pe_alloc_mutex); 3111 3112 /* Detect specific models for error handling */ 3113 if (of_device_is_compatible(np, "ibm,p7ioc-pciex")) 3114 phb->model = PNV_PHB_MODEL_P7IOC; 3115 else if (of_device_is_compatible(np, "ibm,power8-pciex")) 3116 phb->model = PNV_PHB_MODEL_PHB3; 3117 else 3118 phb->model = PNV_PHB_MODEL_UNKNOWN; 3119 3120 /* Parse 32-bit and IO ranges (if any) */ 3121 pci_process_bridge_OF_ranges(hose, np, !hose->global_number); 3122 3123 /* Get registers */ 3124 phb->regs = of_iomap(np, 0); 3125 if (phb->regs == NULL) 3126 pr_err(" Failed to map registers !\n"); 3127 3128 /* Initialize more IODA stuff */ 3129 phb->ioda.total_pe = 1; 3130 prop32 = of_get_property(np, "ibm,opal-num-pes", NULL); 3131 if (prop32) 3132 phb->ioda.total_pe = be32_to_cpup(prop32); 3133 prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL); 3134 if (prop32) 3135 phb->ioda.reserved_pe = be32_to_cpup(prop32); 3136 3137 /* Parse 64-bit MMIO range */ 3138 pnv_ioda_parse_m64_window(phb); 3139 3140 phb->ioda.m32_size = resource_size(&hose->mem_resources[0]); 3141 /* FW Has already off top 64k of M32 space (MSI space) */ 3142 phb->ioda.m32_size += 0x10000; 3143 3144 phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe; 3145 phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0]; 3146 phb->ioda.io_size = hose->pci_io_size; 3147 phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe; 3148 phb->ioda.io_pci_base = 0; /* XXX calculate this ? */ 3149 3150 /* Allocate aux data & arrays. We don't have IO ports on PHB3 */ 3151 size = _ALIGN_UP(phb->ioda.total_pe / 8, sizeof(unsigned long)); 3152 m32map_off = size; 3153 size += phb->ioda.total_pe * sizeof(phb->ioda.m32_segmap[0]); 3154 if (phb->type == PNV_PHB_IODA1) { 3155 iomap_off = size; 3156 size += phb->ioda.total_pe * sizeof(phb->ioda.io_segmap[0]); 3157 } 3158 pemap_off = size; 3159 size += phb->ioda.total_pe * sizeof(struct pnv_ioda_pe); 3160 aux = memblock_virt_alloc(size, 0); 3161 phb->ioda.pe_alloc = aux; 3162 phb->ioda.m32_segmap = aux + m32map_off; 3163 if (phb->type == PNV_PHB_IODA1) 3164 phb->ioda.io_segmap = aux + iomap_off; 3165 phb->ioda.pe_array = aux + pemap_off; 3166 set_bit(phb->ioda.reserved_pe, phb->ioda.pe_alloc); 3167 3168 INIT_LIST_HEAD(&phb->ioda.pe_dma_list); 3169 INIT_LIST_HEAD(&phb->ioda.pe_list); 3170 mutex_init(&phb->ioda.pe_list_mutex); 3171 3172 /* Calculate how many 32-bit TCE segments we have */ 3173 phb->ioda.tce32_count = phb->ioda.m32_pci_base >> 28; 3174 3175 #if 0 /* We should really do that ... */ 3176 rc = opal_pci_set_phb_mem_window(opal->phb_id, 3177 window_type, 3178 window_num, 3179 starting_real_address, 3180 starting_pci_address, 3181 segment_size); 3182 #endif 3183 3184 pr_info(" %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n", 3185 phb->ioda.total_pe, phb->ioda.reserved_pe, 3186 phb->ioda.m32_size, phb->ioda.m32_segsize); 3187 if (phb->ioda.m64_size) 3188 pr_info(" M64: 0x%lx [segment=0x%lx]\n", 3189 phb->ioda.m64_size, phb->ioda.m64_segsize); 3190 if (phb->ioda.io_size) 3191 pr_info(" IO: 0x%x [segment=0x%x]\n", 3192 phb->ioda.io_size, phb->ioda.io_segsize); 3193 3194 3195 phb->hose->ops = &pnv_pci_ops; 3196 phb->get_pe_state = pnv_ioda_get_pe_state; 3197 phb->freeze_pe = pnv_ioda_freeze_pe; 3198 phb->unfreeze_pe = pnv_ioda_unfreeze_pe; 3199 3200 /* Setup RID -> PE mapping function */ 3201 phb->bdfn_to_pe = pnv_ioda_bdfn_to_pe; 3202 3203 /* Setup TCEs */ 3204 phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup; 3205 phb->dma_get_required_mask = pnv_pci_ioda_dma_get_required_mask; 3206 3207 /* Setup MSI support */ 3208 pnv_pci_init_ioda_msis(phb); 3209 3210 /* 3211 * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here 3212 * to let the PCI core do resource assignment. It's supposed 3213 * that the PCI core will do correct I/O and MMIO alignment 3214 * for the P2P bridge bars so that each PCI bus (excluding 3215 * the child P2P bridges) can form individual PE. 3216 */ 3217 ppc_md.pcibios_fixup = pnv_pci_ioda_fixup; 3218 hose->controller_ops = pnv_pci_ioda_controller_ops; 3219 3220 #ifdef CONFIG_PCI_IOV 3221 ppc_md.pcibios_fixup_sriov = pnv_pci_ioda_fixup_iov_resources; 3222 ppc_md.pcibios_iov_resource_alignment = pnv_pci_iov_resource_alignment; 3223 #endif 3224 3225 pci_add_flags(PCI_REASSIGN_ALL_RSRC); 3226 3227 /* Reset IODA tables to a clean state */ 3228 rc = opal_pci_reset(phb_id, OPAL_RESET_PCI_IODA_TABLE, OPAL_ASSERT_RESET); 3229 if (rc) 3230 pr_warning(" OPAL Error %ld performing IODA table reset !\n", rc); 3231 3232 /* If we're running in kdump kerenl, the previous kerenl never 3233 * shutdown PCI devices correctly. We already got IODA table 3234 * cleaned out. So we have to issue PHB reset to stop all PCI 3235 * transactions from previous kerenl. 3236 */ 3237 if (is_kdump_kernel()) { 3238 pr_info(" Issue PHB reset ...\n"); 3239 pnv_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL); 3240 pnv_eeh_phb_reset(hose, EEH_RESET_DEACTIVATE); 3241 } 3242 3243 /* Remove M64 resource if we can't configure it successfully */ 3244 if (!phb->init_m64 || phb->init_m64(phb)) 3245 hose->mem_resources[1].flags = 0; 3246 } 3247 3248 void __init pnv_pci_init_ioda2_phb(struct device_node *np) 3249 { 3250 pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2); 3251 } 3252 3253 void __init pnv_pci_init_ioda_hub(struct device_node *np) 3254 { 3255 struct device_node *phbn; 3256 const __be64 *prop64; 3257 u64 hub_id; 3258 3259 pr_info("Probing IODA IO-Hub %s\n", np->full_name); 3260 3261 prop64 = of_get_property(np, "ibm,opal-hubid", NULL); 3262 if (!prop64) { 3263 pr_err(" Missing \"ibm,opal-hubid\" property !\n"); 3264 return; 3265 } 3266 hub_id = be64_to_cpup(prop64); 3267 pr_devel(" HUB-ID : 0x%016llx\n", hub_id); 3268 3269 /* Count child PHBs */ 3270 for_each_child_of_node(np, phbn) { 3271 /* Look for IODA1 PHBs */ 3272 if (of_device_is_compatible(phbn, "ibm,ioda-phb")) 3273 pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1); 3274 } 3275 } 3276