1 /* 2 * Support PCI/PCIe on PowerNV platforms 3 * 4 * Copyright 2011 Benjamin Herrenschmidt, IBM Corp. 5 * 6 * This program is free software; you can redistribute it and/or 7 * modify it under the terms of the GNU General Public License 8 * as published by the Free Software Foundation; either version 9 * 2 of the License, or (at your option) any later version. 10 */ 11 12 #undef DEBUG 13 14 #include <linux/kernel.h> 15 #include <linux/pci.h> 16 #include <linux/crash_dump.h> 17 #include <linux/delay.h> 18 #include <linux/string.h> 19 #include <linux/init.h> 20 #include <linux/memblock.h> 21 #include <linux/irq.h> 22 #include <linux/io.h> 23 #include <linux/msi.h> 24 #include <linux/iommu.h> 25 #include <linux/rculist.h> 26 #include <linux/sizes.h> 27 28 #include <asm/sections.h> 29 #include <asm/io.h> 30 #include <asm/prom.h> 31 #include <asm/pci-bridge.h> 32 #include <asm/machdep.h> 33 #include <asm/msi_bitmap.h> 34 #include <asm/ppc-pci.h> 35 #include <asm/opal.h> 36 #include <asm/iommu.h> 37 #include <asm/tce.h> 38 #include <asm/xics.h> 39 #include <asm/debugfs.h> 40 #include <asm/firmware.h> 41 #include <asm/pnv-pci.h> 42 #include <asm/mmzone.h> 43 44 #include <misc/cxl-base.h> 45 46 #include "powernv.h" 47 #include "pci.h" 48 #include "../../../../drivers/pci/pci.h" 49 50 #define PNV_IODA1_M64_NUM 16 /* Number of M64 BARs */ 51 #define PNV_IODA1_M64_SEGS 8 /* Segments per M64 BAR */ 52 #define PNV_IODA1_DMA32_SEGSIZE 0x10000000 53 54 static const char * const pnv_phb_names[] = { "IODA1", "IODA2", "NPU_NVLINK", 55 "NPU_OCAPI" }; 56 57 void pe_level_printk(const struct pnv_ioda_pe *pe, const char *level, 58 const char *fmt, ...) 59 { 60 struct va_format vaf; 61 va_list args; 62 char pfix[32]; 63 64 va_start(args, fmt); 65 66 vaf.fmt = fmt; 67 vaf.va = &args; 68 69 if (pe->flags & PNV_IODA_PE_DEV) 70 strlcpy(pfix, dev_name(&pe->pdev->dev), sizeof(pfix)); 71 else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)) 72 sprintf(pfix, "%04x:%02x ", 73 pci_domain_nr(pe->pbus), pe->pbus->number); 74 #ifdef CONFIG_PCI_IOV 75 else if (pe->flags & PNV_IODA_PE_VF) 76 sprintf(pfix, "%04x:%02x:%2x.%d", 77 pci_domain_nr(pe->parent_dev->bus), 78 (pe->rid & 0xff00) >> 8, 79 PCI_SLOT(pe->rid), PCI_FUNC(pe->rid)); 80 #endif /* CONFIG_PCI_IOV*/ 81 82 printk("%spci %s: [PE# %.2x] %pV", 83 level, pfix, pe->pe_number, &vaf); 84 85 va_end(args); 86 } 87 88 static bool pnv_iommu_bypass_disabled __read_mostly; 89 static bool pci_reset_phbs __read_mostly; 90 91 static int __init iommu_setup(char *str) 92 { 93 if (!str) 94 return -EINVAL; 95 96 while (*str) { 97 if (!strncmp(str, "nobypass", 8)) { 98 pnv_iommu_bypass_disabled = true; 99 pr_info("PowerNV: IOMMU bypass window disabled.\n"); 100 break; 101 } 102 str += strcspn(str, ","); 103 if (*str == ',') 104 str++; 105 } 106 107 return 0; 108 } 109 early_param("iommu", iommu_setup); 110 111 static int __init pci_reset_phbs_setup(char *str) 112 { 113 pci_reset_phbs = true; 114 return 0; 115 } 116 117 early_param("ppc_pci_reset_phbs", pci_reset_phbs_setup); 118 119 static inline bool pnv_pci_is_m64(struct pnv_phb *phb, struct resource *r) 120 { 121 /* 122 * WARNING: We cannot rely on the resource flags. The Linux PCI 123 * allocation code sometimes decides to put a 64-bit prefetchable 124 * BAR in the 32-bit window, so we have to compare the addresses. 125 * 126 * For simplicity we only test resource start. 127 */ 128 return (r->start >= phb->ioda.m64_base && 129 r->start < (phb->ioda.m64_base + phb->ioda.m64_size)); 130 } 131 132 static inline bool pnv_pci_is_m64_flags(unsigned long resource_flags) 133 { 134 unsigned long flags = (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH); 135 136 return (resource_flags & flags) == flags; 137 } 138 139 static struct pnv_ioda_pe *pnv_ioda_init_pe(struct pnv_phb *phb, int pe_no) 140 { 141 s64 rc; 142 143 phb->ioda.pe_array[pe_no].phb = phb; 144 phb->ioda.pe_array[pe_no].pe_number = pe_no; 145 146 /* 147 * Clear the PE frozen state as it might be put into frozen state 148 * in the last PCI remove path. It's not harmful to do so when the 149 * PE is already in unfrozen state. 150 */ 151 rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, 152 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL); 153 if (rc != OPAL_SUCCESS && rc != OPAL_UNSUPPORTED) 154 pr_warn("%s: Error %lld unfreezing PHB#%x-PE#%x\n", 155 __func__, rc, phb->hose->global_number, pe_no); 156 157 return &phb->ioda.pe_array[pe_no]; 158 } 159 160 static void pnv_ioda_reserve_pe(struct pnv_phb *phb, int pe_no) 161 { 162 if (!(pe_no >= 0 && pe_no < phb->ioda.total_pe_num)) { 163 pr_warn("%s: Invalid PE %x on PHB#%x\n", 164 __func__, pe_no, phb->hose->global_number); 165 return; 166 } 167 168 if (test_and_set_bit(pe_no, phb->ioda.pe_alloc)) 169 pr_debug("%s: PE %x was reserved on PHB#%x\n", 170 __func__, pe_no, phb->hose->global_number); 171 172 pnv_ioda_init_pe(phb, pe_no); 173 } 174 175 static struct pnv_ioda_pe *pnv_ioda_alloc_pe(struct pnv_phb *phb) 176 { 177 long pe; 178 179 for (pe = phb->ioda.total_pe_num - 1; pe >= 0; pe--) { 180 if (!test_and_set_bit(pe, phb->ioda.pe_alloc)) 181 return pnv_ioda_init_pe(phb, pe); 182 } 183 184 return NULL; 185 } 186 187 static void pnv_ioda_free_pe(struct pnv_ioda_pe *pe) 188 { 189 struct pnv_phb *phb = pe->phb; 190 unsigned int pe_num = pe->pe_number; 191 192 WARN_ON(pe->pdev); 193 194 memset(pe, 0, sizeof(struct pnv_ioda_pe)); 195 clear_bit(pe_num, phb->ioda.pe_alloc); 196 } 197 198 /* The default M64 BAR is shared by all PEs */ 199 static int pnv_ioda2_init_m64(struct pnv_phb *phb) 200 { 201 const char *desc; 202 struct resource *r; 203 s64 rc; 204 205 /* Configure the default M64 BAR */ 206 rc = opal_pci_set_phb_mem_window(phb->opal_id, 207 OPAL_M64_WINDOW_TYPE, 208 phb->ioda.m64_bar_idx, 209 phb->ioda.m64_base, 210 0, /* unused */ 211 phb->ioda.m64_size); 212 if (rc != OPAL_SUCCESS) { 213 desc = "configuring"; 214 goto fail; 215 } 216 217 /* Enable the default M64 BAR */ 218 rc = opal_pci_phb_mmio_enable(phb->opal_id, 219 OPAL_M64_WINDOW_TYPE, 220 phb->ioda.m64_bar_idx, 221 OPAL_ENABLE_M64_SPLIT); 222 if (rc != OPAL_SUCCESS) { 223 desc = "enabling"; 224 goto fail; 225 } 226 227 /* 228 * Exclude the segments for reserved and root bus PE, which 229 * are first or last two PEs. 230 */ 231 r = &phb->hose->mem_resources[1]; 232 if (phb->ioda.reserved_pe_idx == 0) 233 r->start += (2 * phb->ioda.m64_segsize); 234 else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1)) 235 r->end -= (2 * phb->ioda.m64_segsize); 236 else 237 pr_warn(" Cannot strip M64 segment for reserved PE#%x\n", 238 phb->ioda.reserved_pe_idx); 239 240 return 0; 241 242 fail: 243 pr_warn(" Failure %lld %s M64 BAR#%d\n", 244 rc, desc, phb->ioda.m64_bar_idx); 245 opal_pci_phb_mmio_enable(phb->opal_id, 246 OPAL_M64_WINDOW_TYPE, 247 phb->ioda.m64_bar_idx, 248 OPAL_DISABLE_M64); 249 return -EIO; 250 } 251 252 static void pnv_ioda_reserve_dev_m64_pe(struct pci_dev *pdev, 253 unsigned long *pe_bitmap) 254 { 255 struct pci_controller *hose = pci_bus_to_host(pdev->bus); 256 struct pnv_phb *phb = hose->private_data; 257 struct resource *r; 258 resource_size_t base, sgsz, start, end; 259 int segno, i; 260 261 base = phb->ioda.m64_base; 262 sgsz = phb->ioda.m64_segsize; 263 for (i = 0; i <= PCI_ROM_RESOURCE; i++) { 264 r = &pdev->resource[i]; 265 if (!r->parent || !pnv_pci_is_m64(phb, r)) 266 continue; 267 268 start = _ALIGN_DOWN(r->start - base, sgsz); 269 end = _ALIGN_UP(r->end - base, sgsz); 270 for (segno = start / sgsz; segno < end / sgsz; segno++) { 271 if (pe_bitmap) 272 set_bit(segno, pe_bitmap); 273 else 274 pnv_ioda_reserve_pe(phb, segno); 275 } 276 } 277 } 278 279 static int pnv_ioda1_init_m64(struct pnv_phb *phb) 280 { 281 struct resource *r; 282 int index; 283 284 /* 285 * There are 16 M64 BARs, each of which has 8 segments. So 286 * there are as many M64 segments as the maximum number of 287 * PEs, which is 128. 288 */ 289 for (index = 0; index < PNV_IODA1_M64_NUM; index++) { 290 unsigned long base, segsz = phb->ioda.m64_segsize; 291 int64_t rc; 292 293 base = phb->ioda.m64_base + 294 index * PNV_IODA1_M64_SEGS * segsz; 295 rc = opal_pci_set_phb_mem_window(phb->opal_id, 296 OPAL_M64_WINDOW_TYPE, index, base, 0, 297 PNV_IODA1_M64_SEGS * segsz); 298 if (rc != OPAL_SUCCESS) { 299 pr_warn(" Error %lld setting M64 PHB#%x-BAR#%d\n", 300 rc, phb->hose->global_number, index); 301 goto fail; 302 } 303 304 rc = opal_pci_phb_mmio_enable(phb->opal_id, 305 OPAL_M64_WINDOW_TYPE, index, 306 OPAL_ENABLE_M64_SPLIT); 307 if (rc != OPAL_SUCCESS) { 308 pr_warn(" Error %lld enabling M64 PHB#%x-BAR#%d\n", 309 rc, phb->hose->global_number, index); 310 goto fail; 311 } 312 } 313 314 /* 315 * Exclude the segments for reserved and root bus PE, which 316 * are first or last two PEs. 317 */ 318 r = &phb->hose->mem_resources[1]; 319 if (phb->ioda.reserved_pe_idx == 0) 320 r->start += (2 * phb->ioda.m64_segsize); 321 else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1)) 322 r->end -= (2 * phb->ioda.m64_segsize); 323 else 324 WARN(1, "Wrong reserved PE#%x on PHB#%x\n", 325 phb->ioda.reserved_pe_idx, phb->hose->global_number); 326 327 return 0; 328 329 fail: 330 for ( ; index >= 0; index--) 331 opal_pci_phb_mmio_enable(phb->opal_id, 332 OPAL_M64_WINDOW_TYPE, index, OPAL_DISABLE_M64); 333 334 return -EIO; 335 } 336 337 static void pnv_ioda_reserve_m64_pe(struct pci_bus *bus, 338 unsigned long *pe_bitmap, 339 bool all) 340 { 341 struct pci_dev *pdev; 342 343 list_for_each_entry(pdev, &bus->devices, bus_list) { 344 pnv_ioda_reserve_dev_m64_pe(pdev, pe_bitmap); 345 346 if (all && pdev->subordinate) 347 pnv_ioda_reserve_m64_pe(pdev->subordinate, 348 pe_bitmap, all); 349 } 350 } 351 352 static struct pnv_ioda_pe *pnv_ioda_pick_m64_pe(struct pci_bus *bus, bool all) 353 { 354 struct pci_controller *hose = pci_bus_to_host(bus); 355 struct pnv_phb *phb = hose->private_data; 356 struct pnv_ioda_pe *master_pe, *pe; 357 unsigned long size, *pe_alloc; 358 int i; 359 360 /* Root bus shouldn't use M64 */ 361 if (pci_is_root_bus(bus)) 362 return NULL; 363 364 /* Allocate bitmap */ 365 size = _ALIGN_UP(phb->ioda.total_pe_num / 8, sizeof(unsigned long)); 366 pe_alloc = kzalloc(size, GFP_KERNEL); 367 if (!pe_alloc) { 368 pr_warn("%s: Out of memory !\n", 369 __func__); 370 return NULL; 371 } 372 373 /* Figure out reserved PE numbers by the PE */ 374 pnv_ioda_reserve_m64_pe(bus, pe_alloc, all); 375 376 /* 377 * the current bus might not own M64 window and that's all 378 * contributed by its child buses. For the case, we needn't 379 * pick M64 dependent PE#. 380 */ 381 if (bitmap_empty(pe_alloc, phb->ioda.total_pe_num)) { 382 kfree(pe_alloc); 383 return NULL; 384 } 385 386 /* 387 * Figure out the master PE and put all slave PEs to master 388 * PE's list to form compound PE. 389 */ 390 master_pe = NULL; 391 i = -1; 392 while ((i = find_next_bit(pe_alloc, phb->ioda.total_pe_num, i + 1)) < 393 phb->ioda.total_pe_num) { 394 pe = &phb->ioda.pe_array[i]; 395 396 phb->ioda.m64_segmap[pe->pe_number] = pe->pe_number; 397 if (!master_pe) { 398 pe->flags |= PNV_IODA_PE_MASTER; 399 INIT_LIST_HEAD(&pe->slaves); 400 master_pe = pe; 401 } else { 402 pe->flags |= PNV_IODA_PE_SLAVE; 403 pe->master = master_pe; 404 list_add_tail(&pe->list, &master_pe->slaves); 405 } 406 407 /* 408 * P7IOC supports M64DT, which helps mapping M64 segment 409 * to one particular PE#. However, PHB3 has fixed mapping 410 * between M64 segment and PE#. In order to have same logic 411 * for P7IOC and PHB3, we enforce fixed mapping between M64 412 * segment and PE# on P7IOC. 413 */ 414 if (phb->type == PNV_PHB_IODA1) { 415 int64_t rc; 416 417 rc = opal_pci_map_pe_mmio_window(phb->opal_id, 418 pe->pe_number, OPAL_M64_WINDOW_TYPE, 419 pe->pe_number / PNV_IODA1_M64_SEGS, 420 pe->pe_number % PNV_IODA1_M64_SEGS); 421 if (rc != OPAL_SUCCESS) 422 pr_warn("%s: Error %lld mapping M64 for PHB#%x-PE#%x\n", 423 __func__, rc, phb->hose->global_number, 424 pe->pe_number); 425 } 426 } 427 428 kfree(pe_alloc); 429 return master_pe; 430 } 431 432 static void __init pnv_ioda_parse_m64_window(struct pnv_phb *phb) 433 { 434 struct pci_controller *hose = phb->hose; 435 struct device_node *dn = hose->dn; 436 struct resource *res; 437 u32 m64_range[2], i; 438 const __be32 *r; 439 u64 pci_addr; 440 441 if (phb->type != PNV_PHB_IODA1 && phb->type != PNV_PHB_IODA2) { 442 pr_info(" Not support M64 window\n"); 443 return; 444 } 445 446 if (!firmware_has_feature(FW_FEATURE_OPAL)) { 447 pr_info(" Firmware too old to support M64 window\n"); 448 return; 449 } 450 451 r = of_get_property(dn, "ibm,opal-m64-window", NULL); 452 if (!r) { 453 pr_info(" No <ibm,opal-m64-window> on %pOF\n", 454 dn); 455 return; 456 } 457 458 /* 459 * Find the available M64 BAR range and pickup the last one for 460 * covering the whole 64-bits space. We support only one range. 461 */ 462 if (of_property_read_u32_array(dn, "ibm,opal-available-m64-ranges", 463 m64_range, 2)) { 464 /* In absence of the property, assume 0..15 */ 465 m64_range[0] = 0; 466 m64_range[1] = 16; 467 } 468 /* We only support 64 bits in our allocator */ 469 if (m64_range[1] > 63) { 470 pr_warn("%s: Limiting M64 range to 63 (from %d) on PHB#%x\n", 471 __func__, m64_range[1], phb->hose->global_number); 472 m64_range[1] = 63; 473 } 474 /* Empty range, no m64 */ 475 if (m64_range[1] <= m64_range[0]) { 476 pr_warn("%s: M64 empty, disabling M64 usage on PHB#%x\n", 477 __func__, phb->hose->global_number); 478 return; 479 } 480 481 /* Configure M64 informations */ 482 res = &hose->mem_resources[1]; 483 res->name = dn->full_name; 484 res->start = of_translate_address(dn, r + 2); 485 res->end = res->start + of_read_number(r + 4, 2) - 1; 486 res->flags = (IORESOURCE_MEM | IORESOURCE_MEM_64 | IORESOURCE_PREFETCH); 487 pci_addr = of_read_number(r, 2); 488 hose->mem_offset[1] = res->start - pci_addr; 489 490 phb->ioda.m64_size = resource_size(res); 491 phb->ioda.m64_segsize = phb->ioda.m64_size / phb->ioda.total_pe_num; 492 phb->ioda.m64_base = pci_addr; 493 494 /* This lines up nicely with the display from processing OF ranges */ 495 pr_info(" MEM 0x%016llx..0x%016llx -> 0x%016llx (M64 #%d..%d)\n", 496 res->start, res->end, pci_addr, m64_range[0], 497 m64_range[0] + m64_range[1] - 1); 498 499 /* Mark all M64 used up by default */ 500 phb->ioda.m64_bar_alloc = (unsigned long)-1; 501 502 /* Use last M64 BAR to cover M64 window */ 503 m64_range[1]--; 504 phb->ioda.m64_bar_idx = m64_range[0] + m64_range[1]; 505 506 pr_info(" Using M64 #%d as default window\n", phb->ioda.m64_bar_idx); 507 508 /* Mark remaining ones free */ 509 for (i = m64_range[0]; i < m64_range[1]; i++) 510 clear_bit(i, &phb->ioda.m64_bar_alloc); 511 512 /* 513 * Setup init functions for M64 based on IODA version, IODA3 uses 514 * the IODA2 code. 515 */ 516 if (phb->type == PNV_PHB_IODA1) 517 phb->init_m64 = pnv_ioda1_init_m64; 518 else 519 phb->init_m64 = pnv_ioda2_init_m64; 520 } 521 522 static void pnv_ioda_freeze_pe(struct pnv_phb *phb, int pe_no) 523 { 524 struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_no]; 525 struct pnv_ioda_pe *slave; 526 s64 rc; 527 528 /* Fetch master PE */ 529 if (pe->flags & PNV_IODA_PE_SLAVE) { 530 pe = pe->master; 531 if (WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER))) 532 return; 533 534 pe_no = pe->pe_number; 535 } 536 537 /* Freeze master PE */ 538 rc = opal_pci_eeh_freeze_set(phb->opal_id, 539 pe_no, 540 OPAL_EEH_ACTION_SET_FREEZE_ALL); 541 if (rc != OPAL_SUCCESS) { 542 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n", 543 __func__, rc, phb->hose->global_number, pe_no); 544 return; 545 } 546 547 /* Freeze slave PEs */ 548 if (!(pe->flags & PNV_IODA_PE_MASTER)) 549 return; 550 551 list_for_each_entry(slave, &pe->slaves, list) { 552 rc = opal_pci_eeh_freeze_set(phb->opal_id, 553 slave->pe_number, 554 OPAL_EEH_ACTION_SET_FREEZE_ALL); 555 if (rc != OPAL_SUCCESS) 556 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n", 557 __func__, rc, phb->hose->global_number, 558 slave->pe_number); 559 } 560 } 561 562 static int pnv_ioda_unfreeze_pe(struct pnv_phb *phb, int pe_no, int opt) 563 { 564 struct pnv_ioda_pe *pe, *slave; 565 s64 rc; 566 567 /* Find master PE */ 568 pe = &phb->ioda.pe_array[pe_no]; 569 if (pe->flags & PNV_IODA_PE_SLAVE) { 570 pe = pe->master; 571 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)); 572 pe_no = pe->pe_number; 573 } 574 575 /* Clear frozen state for master PE */ 576 rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, opt); 577 if (rc != OPAL_SUCCESS) { 578 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n", 579 __func__, rc, opt, phb->hose->global_number, pe_no); 580 return -EIO; 581 } 582 583 if (!(pe->flags & PNV_IODA_PE_MASTER)) 584 return 0; 585 586 /* Clear frozen state for slave PEs */ 587 list_for_each_entry(slave, &pe->slaves, list) { 588 rc = opal_pci_eeh_freeze_clear(phb->opal_id, 589 slave->pe_number, 590 opt); 591 if (rc != OPAL_SUCCESS) { 592 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n", 593 __func__, rc, opt, phb->hose->global_number, 594 slave->pe_number); 595 return -EIO; 596 } 597 } 598 599 return 0; 600 } 601 602 static int pnv_ioda_get_pe_state(struct pnv_phb *phb, int pe_no) 603 { 604 struct pnv_ioda_pe *slave, *pe; 605 u8 fstate = 0, state; 606 __be16 pcierr = 0; 607 s64 rc; 608 609 /* Sanity check on PE number */ 610 if (pe_no < 0 || pe_no >= phb->ioda.total_pe_num) 611 return OPAL_EEH_STOPPED_PERM_UNAVAIL; 612 613 /* 614 * Fetch the master PE and the PE instance might be 615 * not initialized yet. 616 */ 617 pe = &phb->ioda.pe_array[pe_no]; 618 if (pe->flags & PNV_IODA_PE_SLAVE) { 619 pe = pe->master; 620 WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)); 621 pe_no = pe->pe_number; 622 } 623 624 /* Check the master PE */ 625 rc = opal_pci_eeh_freeze_status(phb->opal_id, pe_no, 626 &state, &pcierr, NULL); 627 if (rc != OPAL_SUCCESS) { 628 pr_warn("%s: Failure %lld getting " 629 "PHB#%x-PE#%x state\n", 630 __func__, rc, 631 phb->hose->global_number, pe_no); 632 return OPAL_EEH_STOPPED_TEMP_UNAVAIL; 633 } 634 635 /* Check the slave PE */ 636 if (!(pe->flags & PNV_IODA_PE_MASTER)) 637 return state; 638 639 list_for_each_entry(slave, &pe->slaves, list) { 640 rc = opal_pci_eeh_freeze_status(phb->opal_id, 641 slave->pe_number, 642 &fstate, 643 &pcierr, 644 NULL); 645 if (rc != OPAL_SUCCESS) { 646 pr_warn("%s: Failure %lld getting " 647 "PHB#%x-PE#%x state\n", 648 __func__, rc, 649 phb->hose->global_number, slave->pe_number); 650 return OPAL_EEH_STOPPED_TEMP_UNAVAIL; 651 } 652 653 /* 654 * Override the result based on the ascending 655 * priority. 656 */ 657 if (fstate > state) 658 state = fstate; 659 } 660 661 return state; 662 } 663 664 struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev) 665 { 666 struct pci_controller *hose = pci_bus_to_host(dev->bus); 667 struct pnv_phb *phb = hose->private_data; 668 struct pci_dn *pdn = pci_get_pdn(dev); 669 670 if (!pdn) 671 return NULL; 672 if (pdn->pe_number == IODA_INVALID_PE) 673 return NULL; 674 return &phb->ioda.pe_array[pdn->pe_number]; 675 } 676 677 static int pnv_ioda_set_one_peltv(struct pnv_phb *phb, 678 struct pnv_ioda_pe *parent, 679 struct pnv_ioda_pe *child, 680 bool is_add) 681 { 682 const char *desc = is_add ? "adding" : "removing"; 683 uint8_t op = is_add ? OPAL_ADD_PE_TO_DOMAIN : 684 OPAL_REMOVE_PE_FROM_DOMAIN; 685 struct pnv_ioda_pe *slave; 686 long rc; 687 688 /* Parent PE affects child PE */ 689 rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number, 690 child->pe_number, op); 691 if (rc != OPAL_SUCCESS) { 692 pe_warn(child, "OPAL error %ld %s to parent PELTV\n", 693 rc, desc); 694 return -ENXIO; 695 } 696 697 if (!(child->flags & PNV_IODA_PE_MASTER)) 698 return 0; 699 700 /* Compound case: parent PE affects slave PEs */ 701 list_for_each_entry(slave, &child->slaves, list) { 702 rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number, 703 slave->pe_number, op); 704 if (rc != OPAL_SUCCESS) { 705 pe_warn(slave, "OPAL error %ld %s to parent PELTV\n", 706 rc, desc); 707 return -ENXIO; 708 } 709 } 710 711 return 0; 712 } 713 714 static int pnv_ioda_set_peltv(struct pnv_phb *phb, 715 struct pnv_ioda_pe *pe, 716 bool is_add) 717 { 718 struct pnv_ioda_pe *slave; 719 struct pci_dev *pdev = NULL; 720 int ret; 721 722 /* 723 * Clear PE frozen state. If it's master PE, we need 724 * clear slave PE frozen state as well. 725 */ 726 if (is_add) { 727 opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number, 728 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL); 729 if (pe->flags & PNV_IODA_PE_MASTER) { 730 list_for_each_entry(slave, &pe->slaves, list) 731 opal_pci_eeh_freeze_clear(phb->opal_id, 732 slave->pe_number, 733 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL); 734 } 735 } 736 737 /* 738 * Associate PE in PELT. We need add the PE into the 739 * corresponding PELT-V as well. Otherwise, the error 740 * originated from the PE might contribute to other 741 * PEs. 742 */ 743 ret = pnv_ioda_set_one_peltv(phb, pe, pe, is_add); 744 if (ret) 745 return ret; 746 747 /* For compound PEs, any one affects all of them */ 748 if (pe->flags & PNV_IODA_PE_MASTER) { 749 list_for_each_entry(slave, &pe->slaves, list) { 750 ret = pnv_ioda_set_one_peltv(phb, slave, pe, is_add); 751 if (ret) 752 return ret; 753 } 754 } 755 756 if (pe->flags & (PNV_IODA_PE_BUS_ALL | PNV_IODA_PE_BUS)) 757 pdev = pe->pbus->self; 758 else if (pe->flags & PNV_IODA_PE_DEV) 759 pdev = pe->pdev->bus->self; 760 #ifdef CONFIG_PCI_IOV 761 else if (pe->flags & PNV_IODA_PE_VF) 762 pdev = pe->parent_dev; 763 #endif /* CONFIG_PCI_IOV */ 764 while (pdev) { 765 struct pci_dn *pdn = pci_get_pdn(pdev); 766 struct pnv_ioda_pe *parent; 767 768 if (pdn && pdn->pe_number != IODA_INVALID_PE) { 769 parent = &phb->ioda.pe_array[pdn->pe_number]; 770 ret = pnv_ioda_set_one_peltv(phb, parent, pe, is_add); 771 if (ret) 772 return ret; 773 } 774 775 pdev = pdev->bus->self; 776 } 777 778 return 0; 779 } 780 781 static int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe) 782 { 783 struct pci_dev *parent; 784 uint8_t bcomp, dcomp, fcomp; 785 int64_t rc; 786 long rid_end, rid; 787 788 /* Currently, we just deconfigure VF PE. Bus PE will always there.*/ 789 if (pe->pbus) { 790 int count; 791 792 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER; 793 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER; 794 parent = pe->pbus->self; 795 if (pe->flags & PNV_IODA_PE_BUS_ALL) 796 count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1; 797 else 798 count = 1; 799 800 switch(count) { 801 case 1: bcomp = OpalPciBusAll; break; 802 case 2: bcomp = OpalPciBus7Bits; break; 803 case 4: bcomp = OpalPciBus6Bits; break; 804 case 8: bcomp = OpalPciBus5Bits; break; 805 case 16: bcomp = OpalPciBus4Bits; break; 806 case 32: bcomp = OpalPciBus3Bits; break; 807 default: 808 dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n", 809 count); 810 /* Do an exact match only */ 811 bcomp = OpalPciBusAll; 812 } 813 rid_end = pe->rid + (count << 8); 814 } else { 815 #ifdef CONFIG_PCI_IOV 816 if (pe->flags & PNV_IODA_PE_VF) 817 parent = pe->parent_dev; 818 else 819 #endif 820 parent = pe->pdev->bus->self; 821 bcomp = OpalPciBusAll; 822 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER; 823 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER; 824 rid_end = pe->rid + 1; 825 } 826 827 /* Clear the reverse map */ 828 for (rid = pe->rid; rid < rid_end; rid++) 829 phb->ioda.pe_rmap[rid] = IODA_INVALID_PE; 830 831 /* Release from all parents PELT-V */ 832 while (parent) { 833 struct pci_dn *pdn = pci_get_pdn(parent); 834 if (pdn && pdn->pe_number != IODA_INVALID_PE) { 835 rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number, 836 pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN); 837 /* XXX What to do in case of error ? */ 838 } 839 parent = parent->bus->self; 840 } 841 842 opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number, 843 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL); 844 845 /* Disassociate PE in PELT */ 846 rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number, 847 pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN); 848 if (rc) 849 pe_warn(pe, "OPAL error %ld remove self from PELTV\n", rc); 850 rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid, 851 bcomp, dcomp, fcomp, OPAL_UNMAP_PE); 852 if (rc) 853 pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc); 854 855 pe->pbus = NULL; 856 pe->pdev = NULL; 857 #ifdef CONFIG_PCI_IOV 858 pe->parent_dev = NULL; 859 #endif 860 861 return 0; 862 } 863 864 static int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe) 865 { 866 struct pci_dev *parent; 867 uint8_t bcomp, dcomp, fcomp; 868 long rc, rid_end, rid; 869 870 /* Bus validation ? */ 871 if (pe->pbus) { 872 int count; 873 874 dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER; 875 fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER; 876 parent = pe->pbus->self; 877 if (pe->flags & PNV_IODA_PE_BUS_ALL) 878 count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1; 879 else 880 count = 1; 881 882 switch(count) { 883 case 1: bcomp = OpalPciBusAll; break; 884 case 2: bcomp = OpalPciBus7Bits; break; 885 case 4: bcomp = OpalPciBus6Bits; break; 886 case 8: bcomp = OpalPciBus5Bits; break; 887 case 16: bcomp = OpalPciBus4Bits; break; 888 case 32: bcomp = OpalPciBus3Bits; break; 889 default: 890 dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n", 891 count); 892 /* Do an exact match only */ 893 bcomp = OpalPciBusAll; 894 } 895 rid_end = pe->rid + (count << 8); 896 } else { 897 #ifdef CONFIG_PCI_IOV 898 if (pe->flags & PNV_IODA_PE_VF) 899 parent = pe->parent_dev; 900 else 901 #endif /* CONFIG_PCI_IOV */ 902 parent = pe->pdev->bus->self; 903 bcomp = OpalPciBusAll; 904 dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER; 905 fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER; 906 rid_end = pe->rid + 1; 907 } 908 909 /* 910 * Associate PE in PELT. We need add the PE into the 911 * corresponding PELT-V as well. Otherwise, the error 912 * originated from the PE might contribute to other 913 * PEs. 914 */ 915 rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid, 916 bcomp, dcomp, fcomp, OPAL_MAP_PE); 917 if (rc) { 918 pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc); 919 return -ENXIO; 920 } 921 922 /* 923 * Configure PELTV. NPUs don't have a PELTV table so skip 924 * configuration on them. 925 */ 926 if (phb->type != PNV_PHB_NPU_NVLINK && phb->type != PNV_PHB_NPU_OCAPI) 927 pnv_ioda_set_peltv(phb, pe, true); 928 929 /* Setup reverse map */ 930 for (rid = pe->rid; rid < rid_end; rid++) 931 phb->ioda.pe_rmap[rid] = pe->pe_number; 932 933 /* Setup one MVTs on IODA1 */ 934 if (phb->type != PNV_PHB_IODA1) { 935 pe->mve_number = 0; 936 goto out; 937 } 938 939 pe->mve_number = pe->pe_number; 940 rc = opal_pci_set_mve(phb->opal_id, pe->mve_number, pe->pe_number); 941 if (rc != OPAL_SUCCESS) { 942 pe_err(pe, "OPAL error %ld setting up MVE %x\n", 943 rc, pe->mve_number); 944 pe->mve_number = -1; 945 } else { 946 rc = opal_pci_set_mve_enable(phb->opal_id, 947 pe->mve_number, OPAL_ENABLE_MVE); 948 if (rc) { 949 pe_err(pe, "OPAL error %ld enabling MVE %x\n", 950 rc, pe->mve_number); 951 pe->mve_number = -1; 952 } 953 } 954 955 out: 956 return 0; 957 } 958 959 #ifdef CONFIG_PCI_IOV 960 static int pnv_pci_vf_resource_shift(struct pci_dev *dev, int offset) 961 { 962 struct pci_dn *pdn = pci_get_pdn(dev); 963 int i; 964 struct resource *res, res2; 965 resource_size_t size; 966 u16 num_vfs; 967 968 if (!dev->is_physfn) 969 return -EINVAL; 970 971 /* 972 * "offset" is in VFs. The M64 windows are sized so that when they 973 * are segmented, each segment is the same size as the IOV BAR. 974 * Each segment is in a separate PE, and the high order bits of the 975 * address are the PE number. Therefore, each VF's BAR is in a 976 * separate PE, and changing the IOV BAR start address changes the 977 * range of PEs the VFs are in. 978 */ 979 num_vfs = pdn->num_vfs; 980 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { 981 res = &dev->resource[i + PCI_IOV_RESOURCES]; 982 if (!res->flags || !res->parent) 983 continue; 984 985 /* 986 * The actual IOV BAR range is determined by the start address 987 * and the actual size for num_vfs VFs BAR. This check is to 988 * make sure that after shifting, the range will not overlap 989 * with another device. 990 */ 991 size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES); 992 res2.flags = res->flags; 993 res2.start = res->start + (size * offset); 994 res2.end = res2.start + (size * num_vfs) - 1; 995 996 if (res2.end > res->end) { 997 dev_err(&dev->dev, "VF BAR%d: %pR would extend past %pR (trying to enable %d VFs shifted by %d)\n", 998 i, &res2, res, num_vfs, offset); 999 return -EBUSY; 1000 } 1001 } 1002 1003 /* 1004 * Since M64 BAR shares segments among all possible 256 PEs, 1005 * we have to shift the beginning of PF IOV BAR to make it start from 1006 * the segment which belongs to the PE number assigned to the first VF. 1007 * This creates a "hole" in the /proc/iomem which could be used for 1008 * allocating other resources so we reserve this area below and 1009 * release when IOV is released. 1010 */ 1011 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { 1012 res = &dev->resource[i + PCI_IOV_RESOURCES]; 1013 if (!res->flags || !res->parent) 1014 continue; 1015 1016 size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES); 1017 res2 = *res; 1018 res->start += size * offset; 1019 1020 dev_info(&dev->dev, "VF BAR%d: %pR shifted to %pR (%sabling %d VFs shifted by %d)\n", 1021 i, &res2, res, (offset > 0) ? "En" : "Dis", 1022 num_vfs, offset); 1023 1024 if (offset < 0) { 1025 devm_release_resource(&dev->dev, &pdn->holes[i]); 1026 memset(&pdn->holes[i], 0, sizeof(pdn->holes[i])); 1027 } 1028 1029 pci_update_resource(dev, i + PCI_IOV_RESOURCES); 1030 1031 if (offset > 0) { 1032 pdn->holes[i].start = res2.start; 1033 pdn->holes[i].end = res2.start + size * offset - 1; 1034 pdn->holes[i].flags = IORESOURCE_BUS; 1035 pdn->holes[i].name = "pnv_iov_reserved"; 1036 devm_request_resource(&dev->dev, res->parent, 1037 &pdn->holes[i]); 1038 } 1039 } 1040 return 0; 1041 } 1042 #endif /* CONFIG_PCI_IOV */ 1043 1044 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev) 1045 { 1046 struct pci_controller *hose = pci_bus_to_host(dev->bus); 1047 struct pnv_phb *phb = hose->private_data; 1048 struct pci_dn *pdn = pci_get_pdn(dev); 1049 struct pnv_ioda_pe *pe; 1050 1051 if (!pdn) { 1052 pr_err("%s: Device tree node not associated properly\n", 1053 pci_name(dev)); 1054 return NULL; 1055 } 1056 if (pdn->pe_number != IODA_INVALID_PE) 1057 return NULL; 1058 1059 pe = pnv_ioda_alloc_pe(phb); 1060 if (!pe) { 1061 pr_warn("%s: Not enough PE# available, disabling device\n", 1062 pci_name(dev)); 1063 return NULL; 1064 } 1065 1066 /* NOTE: We get only one ref to the pci_dev for the pdn, not for the 1067 * pointer in the PE data structure, both should be destroyed at the 1068 * same time. However, this needs to be looked at more closely again 1069 * once we actually start removing things (Hotplug, SR-IOV, ...) 1070 * 1071 * At some point we want to remove the PDN completely anyways 1072 */ 1073 pci_dev_get(dev); 1074 pdn->pe_number = pe->pe_number; 1075 pe->flags = PNV_IODA_PE_DEV; 1076 pe->pdev = dev; 1077 pe->pbus = NULL; 1078 pe->mve_number = -1; 1079 pe->rid = dev->bus->number << 8 | pdn->devfn; 1080 1081 pe_info(pe, "Associated device to PE\n"); 1082 1083 if (pnv_ioda_configure_pe(phb, pe)) { 1084 /* XXX What do we do here ? */ 1085 pnv_ioda_free_pe(pe); 1086 pdn->pe_number = IODA_INVALID_PE; 1087 pe->pdev = NULL; 1088 pci_dev_put(dev); 1089 return NULL; 1090 } 1091 1092 /* Put PE to the list */ 1093 list_add_tail(&pe->list, &phb->ioda.pe_list); 1094 1095 return pe; 1096 } 1097 1098 static void pnv_ioda_setup_same_PE(struct pci_bus *bus, struct pnv_ioda_pe *pe) 1099 { 1100 struct pci_dev *dev; 1101 1102 list_for_each_entry(dev, &bus->devices, bus_list) { 1103 struct pci_dn *pdn = pci_get_pdn(dev); 1104 1105 if (pdn == NULL) { 1106 pr_warn("%s: No device node associated with device !\n", 1107 pci_name(dev)); 1108 continue; 1109 } 1110 1111 /* 1112 * In partial hotplug case, the PCI device might be still 1113 * associated with the PE and needn't attach it to the PE 1114 * again. 1115 */ 1116 if (pdn->pe_number != IODA_INVALID_PE) 1117 continue; 1118 1119 pe->device_count++; 1120 pdn->pe_number = pe->pe_number; 1121 if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate) 1122 pnv_ioda_setup_same_PE(dev->subordinate, pe); 1123 } 1124 } 1125 1126 /* 1127 * There're 2 types of PCI bus sensitive PEs: One that is compromised of 1128 * single PCI bus. Another one that contains the primary PCI bus and its 1129 * subordinate PCI devices and buses. The second type of PE is normally 1130 * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports. 1131 */ 1132 static struct pnv_ioda_pe *pnv_ioda_setup_bus_PE(struct pci_bus *bus, bool all) 1133 { 1134 struct pci_controller *hose = pci_bus_to_host(bus); 1135 struct pnv_phb *phb = hose->private_data; 1136 struct pnv_ioda_pe *pe = NULL; 1137 unsigned int pe_num; 1138 1139 /* 1140 * In partial hotplug case, the PE instance might be still alive. 1141 * We should reuse it instead of allocating a new one. 1142 */ 1143 pe_num = phb->ioda.pe_rmap[bus->number << 8]; 1144 if (pe_num != IODA_INVALID_PE) { 1145 pe = &phb->ioda.pe_array[pe_num]; 1146 pnv_ioda_setup_same_PE(bus, pe); 1147 return NULL; 1148 } 1149 1150 /* PE number for root bus should have been reserved */ 1151 if (pci_is_root_bus(bus) && 1152 phb->ioda.root_pe_idx != IODA_INVALID_PE) 1153 pe = &phb->ioda.pe_array[phb->ioda.root_pe_idx]; 1154 1155 /* Check if PE is determined by M64 */ 1156 if (!pe) 1157 pe = pnv_ioda_pick_m64_pe(bus, all); 1158 1159 /* The PE number isn't pinned by M64 */ 1160 if (!pe) 1161 pe = pnv_ioda_alloc_pe(phb); 1162 1163 if (!pe) { 1164 pr_warn("%s: Not enough PE# available for PCI bus %04x:%02x\n", 1165 __func__, pci_domain_nr(bus), bus->number); 1166 return NULL; 1167 } 1168 1169 pe->flags |= (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS); 1170 pe->pbus = bus; 1171 pe->pdev = NULL; 1172 pe->mve_number = -1; 1173 pe->rid = bus->busn_res.start << 8; 1174 1175 if (all) 1176 pe_info(pe, "Secondary bus %d..%d associated with PE#%x\n", 1177 bus->busn_res.start, bus->busn_res.end, pe->pe_number); 1178 else 1179 pe_info(pe, "Secondary bus %d associated with PE#%x\n", 1180 bus->busn_res.start, pe->pe_number); 1181 1182 if (pnv_ioda_configure_pe(phb, pe)) { 1183 /* XXX What do we do here ? */ 1184 pnv_ioda_free_pe(pe); 1185 pe->pbus = NULL; 1186 return NULL; 1187 } 1188 1189 /* Associate it with all child devices */ 1190 pnv_ioda_setup_same_PE(bus, pe); 1191 1192 /* Put PE to the list */ 1193 list_add_tail(&pe->list, &phb->ioda.pe_list); 1194 1195 return pe; 1196 } 1197 1198 static struct pnv_ioda_pe *pnv_ioda_setup_npu_PE(struct pci_dev *npu_pdev) 1199 { 1200 int pe_num, found_pe = false, rc; 1201 long rid; 1202 struct pnv_ioda_pe *pe; 1203 struct pci_dev *gpu_pdev; 1204 struct pci_dn *npu_pdn; 1205 struct pci_controller *hose = pci_bus_to_host(npu_pdev->bus); 1206 struct pnv_phb *phb = hose->private_data; 1207 1208 /* 1209 * Due to a hardware errata PE#0 on the NPU is reserved for 1210 * error handling. This means we only have three PEs remaining 1211 * which need to be assigned to four links, implying some 1212 * links must share PEs. 1213 * 1214 * To achieve this we assign PEs such that NPUs linking the 1215 * same GPU get assigned the same PE. 1216 */ 1217 gpu_pdev = pnv_pci_get_gpu_dev(npu_pdev); 1218 for (pe_num = 0; pe_num < phb->ioda.total_pe_num; pe_num++) { 1219 pe = &phb->ioda.pe_array[pe_num]; 1220 if (!pe->pdev) 1221 continue; 1222 1223 if (pnv_pci_get_gpu_dev(pe->pdev) == gpu_pdev) { 1224 /* 1225 * This device has the same peer GPU so should 1226 * be assigned the same PE as the existing 1227 * peer NPU. 1228 */ 1229 dev_info(&npu_pdev->dev, 1230 "Associating to existing PE %x\n", pe_num); 1231 pci_dev_get(npu_pdev); 1232 npu_pdn = pci_get_pdn(npu_pdev); 1233 rid = npu_pdev->bus->number << 8 | npu_pdn->devfn; 1234 npu_pdn->pe_number = pe_num; 1235 phb->ioda.pe_rmap[rid] = pe->pe_number; 1236 1237 /* Map the PE to this link */ 1238 rc = opal_pci_set_pe(phb->opal_id, pe_num, rid, 1239 OpalPciBusAll, 1240 OPAL_COMPARE_RID_DEVICE_NUMBER, 1241 OPAL_COMPARE_RID_FUNCTION_NUMBER, 1242 OPAL_MAP_PE); 1243 WARN_ON(rc != OPAL_SUCCESS); 1244 found_pe = true; 1245 break; 1246 } 1247 } 1248 1249 if (!found_pe) 1250 /* 1251 * Could not find an existing PE so allocate a new 1252 * one. 1253 */ 1254 return pnv_ioda_setup_dev_PE(npu_pdev); 1255 else 1256 return pe; 1257 } 1258 1259 static void pnv_ioda_setup_npu_PEs(struct pci_bus *bus) 1260 { 1261 struct pci_dev *pdev; 1262 1263 list_for_each_entry(pdev, &bus->devices, bus_list) 1264 pnv_ioda_setup_npu_PE(pdev); 1265 } 1266 1267 static void pnv_pci_ioda_setup_PEs(void) 1268 { 1269 struct pci_controller *hose; 1270 struct pnv_phb *phb; 1271 struct pci_bus *bus; 1272 struct pci_dev *pdev; 1273 struct pnv_ioda_pe *pe; 1274 1275 list_for_each_entry(hose, &hose_list, list_node) { 1276 phb = hose->private_data; 1277 if (phb->type == PNV_PHB_NPU_NVLINK) { 1278 /* PE#0 is needed for error reporting */ 1279 pnv_ioda_reserve_pe(phb, 0); 1280 pnv_ioda_setup_npu_PEs(hose->bus); 1281 if (phb->model == PNV_PHB_MODEL_NPU2) 1282 WARN_ON_ONCE(pnv_npu2_init(hose)); 1283 } 1284 if (phb->type == PNV_PHB_NPU_OCAPI) { 1285 bus = hose->bus; 1286 list_for_each_entry(pdev, &bus->devices, bus_list) 1287 pnv_ioda_setup_dev_PE(pdev); 1288 } 1289 } 1290 list_for_each_entry(hose, &hose_list, list_node) { 1291 phb = hose->private_data; 1292 if (phb->type != PNV_PHB_IODA2) 1293 continue; 1294 1295 list_for_each_entry(pe, &phb->ioda.pe_list, list) 1296 pnv_npu2_map_lpar(pe, MSR_DR | MSR_PR | MSR_HV); 1297 } 1298 } 1299 1300 #ifdef CONFIG_PCI_IOV 1301 static int pnv_pci_vf_release_m64(struct pci_dev *pdev, u16 num_vfs) 1302 { 1303 struct pci_bus *bus; 1304 struct pci_controller *hose; 1305 struct pnv_phb *phb; 1306 struct pci_dn *pdn; 1307 int i, j; 1308 int m64_bars; 1309 1310 bus = pdev->bus; 1311 hose = pci_bus_to_host(bus); 1312 phb = hose->private_data; 1313 pdn = pci_get_pdn(pdev); 1314 1315 if (pdn->m64_single_mode) 1316 m64_bars = num_vfs; 1317 else 1318 m64_bars = 1; 1319 1320 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) 1321 for (j = 0; j < m64_bars; j++) { 1322 if (pdn->m64_map[j][i] == IODA_INVALID_M64) 1323 continue; 1324 opal_pci_phb_mmio_enable(phb->opal_id, 1325 OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 0); 1326 clear_bit(pdn->m64_map[j][i], &phb->ioda.m64_bar_alloc); 1327 pdn->m64_map[j][i] = IODA_INVALID_M64; 1328 } 1329 1330 kfree(pdn->m64_map); 1331 return 0; 1332 } 1333 1334 static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs) 1335 { 1336 struct pci_bus *bus; 1337 struct pci_controller *hose; 1338 struct pnv_phb *phb; 1339 struct pci_dn *pdn; 1340 unsigned int win; 1341 struct resource *res; 1342 int i, j; 1343 int64_t rc; 1344 int total_vfs; 1345 resource_size_t size, start; 1346 int pe_num; 1347 int m64_bars; 1348 1349 bus = pdev->bus; 1350 hose = pci_bus_to_host(bus); 1351 phb = hose->private_data; 1352 pdn = pci_get_pdn(pdev); 1353 total_vfs = pci_sriov_get_totalvfs(pdev); 1354 1355 if (pdn->m64_single_mode) 1356 m64_bars = num_vfs; 1357 else 1358 m64_bars = 1; 1359 1360 pdn->m64_map = kmalloc_array(m64_bars, 1361 sizeof(*pdn->m64_map), 1362 GFP_KERNEL); 1363 if (!pdn->m64_map) 1364 return -ENOMEM; 1365 /* Initialize the m64_map to IODA_INVALID_M64 */ 1366 for (i = 0; i < m64_bars ; i++) 1367 for (j = 0; j < PCI_SRIOV_NUM_BARS; j++) 1368 pdn->m64_map[i][j] = IODA_INVALID_M64; 1369 1370 1371 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { 1372 res = &pdev->resource[i + PCI_IOV_RESOURCES]; 1373 if (!res->flags || !res->parent) 1374 continue; 1375 1376 for (j = 0; j < m64_bars; j++) { 1377 do { 1378 win = find_next_zero_bit(&phb->ioda.m64_bar_alloc, 1379 phb->ioda.m64_bar_idx + 1, 0); 1380 1381 if (win >= phb->ioda.m64_bar_idx + 1) 1382 goto m64_failed; 1383 } while (test_and_set_bit(win, &phb->ioda.m64_bar_alloc)); 1384 1385 pdn->m64_map[j][i] = win; 1386 1387 if (pdn->m64_single_mode) { 1388 size = pci_iov_resource_size(pdev, 1389 PCI_IOV_RESOURCES + i); 1390 start = res->start + size * j; 1391 } else { 1392 size = resource_size(res); 1393 start = res->start; 1394 } 1395 1396 /* Map the M64 here */ 1397 if (pdn->m64_single_mode) { 1398 pe_num = pdn->pe_num_map[j]; 1399 rc = opal_pci_map_pe_mmio_window(phb->opal_id, 1400 pe_num, OPAL_M64_WINDOW_TYPE, 1401 pdn->m64_map[j][i], 0); 1402 } 1403 1404 rc = opal_pci_set_phb_mem_window(phb->opal_id, 1405 OPAL_M64_WINDOW_TYPE, 1406 pdn->m64_map[j][i], 1407 start, 1408 0, /* unused */ 1409 size); 1410 1411 1412 if (rc != OPAL_SUCCESS) { 1413 dev_err(&pdev->dev, "Failed to map M64 window #%d: %lld\n", 1414 win, rc); 1415 goto m64_failed; 1416 } 1417 1418 if (pdn->m64_single_mode) 1419 rc = opal_pci_phb_mmio_enable(phb->opal_id, 1420 OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 2); 1421 else 1422 rc = opal_pci_phb_mmio_enable(phb->opal_id, 1423 OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 1); 1424 1425 if (rc != OPAL_SUCCESS) { 1426 dev_err(&pdev->dev, "Failed to enable M64 window #%d: %llx\n", 1427 win, rc); 1428 goto m64_failed; 1429 } 1430 } 1431 } 1432 return 0; 1433 1434 m64_failed: 1435 pnv_pci_vf_release_m64(pdev, num_vfs); 1436 return -EBUSY; 1437 } 1438 1439 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group, 1440 int num); 1441 1442 static void pnv_pci_ioda2_release_dma_pe(struct pci_dev *dev, struct pnv_ioda_pe *pe) 1443 { 1444 struct iommu_table *tbl; 1445 int64_t rc; 1446 1447 tbl = pe->table_group.tables[0]; 1448 rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0); 1449 if (rc) 1450 pe_warn(pe, "OPAL error %ld release DMA window\n", rc); 1451 1452 pnv_pci_ioda2_set_bypass(pe, false); 1453 if (pe->table_group.group) { 1454 iommu_group_put(pe->table_group.group); 1455 BUG_ON(pe->table_group.group); 1456 } 1457 iommu_tce_table_put(tbl); 1458 } 1459 1460 static void pnv_ioda_release_vf_PE(struct pci_dev *pdev) 1461 { 1462 struct pci_bus *bus; 1463 struct pci_controller *hose; 1464 struct pnv_phb *phb; 1465 struct pnv_ioda_pe *pe, *pe_n; 1466 struct pci_dn *pdn; 1467 1468 bus = pdev->bus; 1469 hose = pci_bus_to_host(bus); 1470 phb = hose->private_data; 1471 pdn = pci_get_pdn(pdev); 1472 1473 if (!pdev->is_physfn) 1474 return; 1475 1476 list_for_each_entry_safe(pe, pe_n, &phb->ioda.pe_list, list) { 1477 if (pe->parent_dev != pdev) 1478 continue; 1479 1480 pnv_pci_ioda2_release_dma_pe(pdev, pe); 1481 1482 /* Remove from list */ 1483 mutex_lock(&phb->ioda.pe_list_mutex); 1484 list_del(&pe->list); 1485 mutex_unlock(&phb->ioda.pe_list_mutex); 1486 1487 pnv_ioda_deconfigure_pe(phb, pe); 1488 1489 pnv_ioda_free_pe(pe); 1490 } 1491 } 1492 1493 void pnv_pci_sriov_disable(struct pci_dev *pdev) 1494 { 1495 struct pci_bus *bus; 1496 struct pci_controller *hose; 1497 struct pnv_phb *phb; 1498 struct pnv_ioda_pe *pe; 1499 struct pci_dn *pdn; 1500 u16 num_vfs, i; 1501 1502 bus = pdev->bus; 1503 hose = pci_bus_to_host(bus); 1504 phb = hose->private_data; 1505 pdn = pci_get_pdn(pdev); 1506 num_vfs = pdn->num_vfs; 1507 1508 /* Release VF PEs */ 1509 pnv_ioda_release_vf_PE(pdev); 1510 1511 if (phb->type == PNV_PHB_IODA2) { 1512 if (!pdn->m64_single_mode) 1513 pnv_pci_vf_resource_shift(pdev, -*pdn->pe_num_map); 1514 1515 /* Release M64 windows */ 1516 pnv_pci_vf_release_m64(pdev, num_vfs); 1517 1518 /* Release PE numbers */ 1519 if (pdn->m64_single_mode) { 1520 for (i = 0; i < num_vfs; i++) { 1521 if (pdn->pe_num_map[i] == IODA_INVALID_PE) 1522 continue; 1523 1524 pe = &phb->ioda.pe_array[pdn->pe_num_map[i]]; 1525 pnv_ioda_free_pe(pe); 1526 } 1527 } else 1528 bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs); 1529 /* Releasing pe_num_map */ 1530 kfree(pdn->pe_num_map); 1531 } 1532 } 1533 1534 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb, 1535 struct pnv_ioda_pe *pe); 1536 #ifdef CONFIG_IOMMU_API 1537 static void pnv_ioda_setup_bus_iommu_group(struct pnv_ioda_pe *pe); 1538 #endif 1539 static void pnv_ioda_setup_vf_PE(struct pci_dev *pdev, u16 num_vfs) 1540 { 1541 struct pci_bus *bus; 1542 struct pci_controller *hose; 1543 struct pnv_phb *phb; 1544 struct pnv_ioda_pe *pe; 1545 int pe_num; 1546 u16 vf_index; 1547 struct pci_dn *pdn; 1548 1549 bus = pdev->bus; 1550 hose = pci_bus_to_host(bus); 1551 phb = hose->private_data; 1552 pdn = pci_get_pdn(pdev); 1553 1554 if (!pdev->is_physfn) 1555 return; 1556 1557 /* Reserve PE for each VF */ 1558 for (vf_index = 0; vf_index < num_vfs; vf_index++) { 1559 if (pdn->m64_single_mode) 1560 pe_num = pdn->pe_num_map[vf_index]; 1561 else 1562 pe_num = *pdn->pe_num_map + vf_index; 1563 1564 pe = &phb->ioda.pe_array[pe_num]; 1565 pe->pe_number = pe_num; 1566 pe->phb = phb; 1567 pe->flags = PNV_IODA_PE_VF; 1568 pe->pbus = NULL; 1569 pe->parent_dev = pdev; 1570 pe->mve_number = -1; 1571 pe->rid = (pci_iov_virtfn_bus(pdev, vf_index) << 8) | 1572 pci_iov_virtfn_devfn(pdev, vf_index); 1573 1574 pe_info(pe, "VF %04d:%02d:%02d.%d associated with PE#%x\n", 1575 hose->global_number, pdev->bus->number, 1576 PCI_SLOT(pci_iov_virtfn_devfn(pdev, vf_index)), 1577 PCI_FUNC(pci_iov_virtfn_devfn(pdev, vf_index)), pe_num); 1578 1579 if (pnv_ioda_configure_pe(phb, pe)) { 1580 /* XXX What do we do here ? */ 1581 pnv_ioda_free_pe(pe); 1582 pe->pdev = NULL; 1583 continue; 1584 } 1585 1586 /* Put PE to the list */ 1587 mutex_lock(&phb->ioda.pe_list_mutex); 1588 list_add_tail(&pe->list, &phb->ioda.pe_list); 1589 mutex_unlock(&phb->ioda.pe_list_mutex); 1590 1591 pnv_pci_ioda2_setup_dma_pe(phb, pe); 1592 #ifdef CONFIG_IOMMU_API 1593 pnv_ioda_setup_bus_iommu_group(pe); 1594 #endif 1595 } 1596 } 1597 1598 int pnv_pci_sriov_enable(struct pci_dev *pdev, u16 num_vfs) 1599 { 1600 struct pci_bus *bus; 1601 struct pci_controller *hose; 1602 struct pnv_phb *phb; 1603 struct pnv_ioda_pe *pe; 1604 struct pci_dn *pdn; 1605 int ret; 1606 u16 i; 1607 1608 bus = pdev->bus; 1609 hose = pci_bus_to_host(bus); 1610 phb = hose->private_data; 1611 pdn = pci_get_pdn(pdev); 1612 1613 if (phb->type == PNV_PHB_IODA2) { 1614 if (!pdn->vfs_expanded) { 1615 dev_info(&pdev->dev, "don't support this SRIOV device" 1616 " with non 64bit-prefetchable IOV BAR\n"); 1617 return -ENOSPC; 1618 } 1619 1620 /* 1621 * When M64 BARs functions in Single PE mode, the number of VFs 1622 * could be enabled must be less than the number of M64 BARs. 1623 */ 1624 if (pdn->m64_single_mode && num_vfs > phb->ioda.m64_bar_idx) { 1625 dev_info(&pdev->dev, "Not enough M64 BAR for VFs\n"); 1626 return -EBUSY; 1627 } 1628 1629 /* Allocating pe_num_map */ 1630 if (pdn->m64_single_mode) 1631 pdn->pe_num_map = kmalloc_array(num_vfs, 1632 sizeof(*pdn->pe_num_map), 1633 GFP_KERNEL); 1634 else 1635 pdn->pe_num_map = kmalloc(sizeof(*pdn->pe_num_map), GFP_KERNEL); 1636 1637 if (!pdn->pe_num_map) 1638 return -ENOMEM; 1639 1640 if (pdn->m64_single_mode) 1641 for (i = 0; i < num_vfs; i++) 1642 pdn->pe_num_map[i] = IODA_INVALID_PE; 1643 1644 /* Calculate available PE for required VFs */ 1645 if (pdn->m64_single_mode) { 1646 for (i = 0; i < num_vfs; i++) { 1647 pe = pnv_ioda_alloc_pe(phb); 1648 if (!pe) { 1649 ret = -EBUSY; 1650 goto m64_failed; 1651 } 1652 1653 pdn->pe_num_map[i] = pe->pe_number; 1654 } 1655 } else { 1656 mutex_lock(&phb->ioda.pe_alloc_mutex); 1657 *pdn->pe_num_map = bitmap_find_next_zero_area( 1658 phb->ioda.pe_alloc, phb->ioda.total_pe_num, 1659 0, num_vfs, 0); 1660 if (*pdn->pe_num_map >= phb->ioda.total_pe_num) { 1661 mutex_unlock(&phb->ioda.pe_alloc_mutex); 1662 dev_info(&pdev->dev, "Failed to enable VF%d\n", num_vfs); 1663 kfree(pdn->pe_num_map); 1664 return -EBUSY; 1665 } 1666 bitmap_set(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs); 1667 mutex_unlock(&phb->ioda.pe_alloc_mutex); 1668 } 1669 pdn->num_vfs = num_vfs; 1670 1671 /* Assign M64 window accordingly */ 1672 ret = pnv_pci_vf_assign_m64(pdev, num_vfs); 1673 if (ret) { 1674 dev_info(&pdev->dev, "Not enough M64 window resources\n"); 1675 goto m64_failed; 1676 } 1677 1678 /* 1679 * When using one M64 BAR to map one IOV BAR, we need to shift 1680 * the IOV BAR according to the PE# allocated to the VFs. 1681 * Otherwise, the PE# for the VF will conflict with others. 1682 */ 1683 if (!pdn->m64_single_mode) { 1684 ret = pnv_pci_vf_resource_shift(pdev, *pdn->pe_num_map); 1685 if (ret) 1686 goto m64_failed; 1687 } 1688 } 1689 1690 /* Setup VF PEs */ 1691 pnv_ioda_setup_vf_PE(pdev, num_vfs); 1692 1693 return 0; 1694 1695 m64_failed: 1696 if (pdn->m64_single_mode) { 1697 for (i = 0; i < num_vfs; i++) { 1698 if (pdn->pe_num_map[i] == IODA_INVALID_PE) 1699 continue; 1700 1701 pe = &phb->ioda.pe_array[pdn->pe_num_map[i]]; 1702 pnv_ioda_free_pe(pe); 1703 } 1704 } else 1705 bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs); 1706 1707 /* Releasing pe_num_map */ 1708 kfree(pdn->pe_num_map); 1709 1710 return ret; 1711 } 1712 1713 int pnv_pcibios_sriov_disable(struct pci_dev *pdev) 1714 { 1715 pnv_pci_sriov_disable(pdev); 1716 1717 /* Release PCI data */ 1718 remove_dev_pci_data(pdev); 1719 return 0; 1720 } 1721 1722 int pnv_pcibios_sriov_enable(struct pci_dev *pdev, u16 num_vfs) 1723 { 1724 /* Allocate PCI data */ 1725 add_dev_pci_data(pdev); 1726 1727 return pnv_pci_sriov_enable(pdev, num_vfs); 1728 } 1729 #endif /* CONFIG_PCI_IOV */ 1730 1731 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb *phb, struct pci_dev *pdev) 1732 { 1733 struct pci_dn *pdn = pci_get_pdn(pdev); 1734 struct pnv_ioda_pe *pe; 1735 1736 /* 1737 * The function can be called while the PE# 1738 * hasn't been assigned. Do nothing for the 1739 * case. 1740 */ 1741 if (!pdn || pdn->pe_number == IODA_INVALID_PE) 1742 return; 1743 1744 pe = &phb->ioda.pe_array[pdn->pe_number]; 1745 WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops); 1746 set_dma_offset(&pdev->dev, pe->tce_bypass_base); 1747 set_iommu_table_base(&pdev->dev, pe->table_group.tables[0]); 1748 /* 1749 * Note: iommu_add_device() will fail here as 1750 * for physical PE: the device is already added by now; 1751 * for virtual PE: sysfs entries are not ready yet and 1752 * tce_iommu_bus_notifier will add the device to a group later. 1753 */ 1754 } 1755 1756 static bool pnv_pci_ioda_pe_single_vendor(struct pnv_ioda_pe *pe) 1757 { 1758 unsigned short vendor = 0; 1759 struct pci_dev *pdev; 1760 1761 if (pe->device_count == 1) 1762 return true; 1763 1764 /* pe->pdev should be set if it's a single device, pe->pbus if not */ 1765 if (!pe->pbus) 1766 return true; 1767 1768 list_for_each_entry(pdev, &pe->pbus->devices, bus_list) { 1769 if (!vendor) { 1770 vendor = pdev->vendor; 1771 continue; 1772 } 1773 1774 if (pdev->vendor != vendor) 1775 return false; 1776 } 1777 1778 return true; 1779 } 1780 1781 /* 1782 * Reconfigure TVE#0 to be usable as 64-bit DMA space. 1783 * 1784 * The first 4GB of virtual memory for a PE is reserved for 32-bit accesses. 1785 * Devices can only access more than that if bit 59 of the PCI address is set 1786 * by hardware, which indicates TVE#1 should be used instead of TVE#0. 1787 * Many PCI devices are not capable of addressing that many bits, and as a 1788 * result are limited to the 4GB of virtual memory made available to 32-bit 1789 * devices in TVE#0. 1790 * 1791 * In order to work around this, reconfigure TVE#0 to be suitable for 64-bit 1792 * devices by configuring the virtual memory past the first 4GB inaccessible 1793 * by 64-bit DMAs. This should only be used by devices that want more than 1794 * 4GB, and only on PEs that have no 32-bit devices. 1795 * 1796 * Currently this will only work on PHB3 (POWER8). 1797 */ 1798 static int pnv_pci_ioda_dma_64bit_bypass(struct pnv_ioda_pe *pe) 1799 { 1800 u64 window_size, table_size, tce_count, addr; 1801 struct page *table_pages; 1802 u64 tce_order = 28; /* 256MB TCEs */ 1803 __be64 *tces; 1804 s64 rc; 1805 1806 /* 1807 * Window size needs to be a power of two, but needs to account for 1808 * shifting memory by the 4GB offset required to skip 32bit space. 1809 */ 1810 window_size = roundup_pow_of_two(memory_hotplug_max() + (1ULL << 32)); 1811 tce_count = window_size >> tce_order; 1812 table_size = tce_count << 3; 1813 1814 if (table_size < PAGE_SIZE) 1815 table_size = PAGE_SIZE; 1816 1817 table_pages = alloc_pages_node(pe->phb->hose->node, GFP_KERNEL, 1818 get_order(table_size)); 1819 if (!table_pages) 1820 goto err; 1821 1822 tces = page_address(table_pages); 1823 if (!tces) 1824 goto err; 1825 1826 memset(tces, 0, table_size); 1827 1828 for (addr = 0; addr < memory_hotplug_max(); addr += (1 << tce_order)) { 1829 tces[(addr + (1ULL << 32)) >> tce_order] = 1830 cpu_to_be64(addr | TCE_PCI_READ | TCE_PCI_WRITE); 1831 } 1832 1833 rc = opal_pci_map_pe_dma_window(pe->phb->opal_id, 1834 pe->pe_number, 1835 /* reconfigure window 0 */ 1836 (pe->pe_number << 1) + 0, 1837 1, 1838 __pa(tces), 1839 table_size, 1840 1 << tce_order); 1841 if (rc == OPAL_SUCCESS) { 1842 pe_info(pe, "Using 64-bit DMA iommu bypass (through TVE#0)\n"); 1843 return 0; 1844 } 1845 err: 1846 pe_err(pe, "Error configuring 64-bit DMA bypass\n"); 1847 return -EIO; 1848 } 1849 1850 static int pnv_pci_ioda_dma_set_mask(struct pci_dev *pdev, u64 dma_mask) 1851 { 1852 struct pci_controller *hose = pci_bus_to_host(pdev->bus); 1853 struct pnv_phb *phb = hose->private_data; 1854 struct pci_dn *pdn = pci_get_pdn(pdev); 1855 struct pnv_ioda_pe *pe; 1856 uint64_t top; 1857 bool bypass = false; 1858 s64 rc; 1859 1860 if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE)) 1861 return -ENODEV; 1862 1863 pe = &phb->ioda.pe_array[pdn->pe_number]; 1864 if (pe->tce_bypass_enabled) { 1865 top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1; 1866 bypass = (dma_mask >= top); 1867 } 1868 1869 if (bypass) { 1870 dev_info(&pdev->dev, "Using 64-bit DMA iommu bypass\n"); 1871 set_dma_ops(&pdev->dev, &dma_nommu_ops); 1872 } else { 1873 /* 1874 * If the device can't set the TCE bypass bit but still wants 1875 * to access 4GB or more, on PHB3 we can reconfigure TVE#0 to 1876 * bypass the 32-bit region and be usable for 64-bit DMAs. 1877 * The device needs to be able to address all of this space. 1878 */ 1879 if (dma_mask >> 32 && 1880 dma_mask > (memory_hotplug_max() + (1ULL << 32)) && 1881 pnv_pci_ioda_pe_single_vendor(pe) && 1882 phb->model == PNV_PHB_MODEL_PHB3) { 1883 /* Configure the bypass mode */ 1884 rc = pnv_pci_ioda_dma_64bit_bypass(pe); 1885 if (rc) 1886 return rc; 1887 /* 4GB offset bypasses 32-bit space */ 1888 set_dma_offset(&pdev->dev, (1ULL << 32)); 1889 set_dma_ops(&pdev->dev, &dma_nommu_ops); 1890 } else if (dma_mask >> 32 && dma_mask != DMA_BIT_MASK(64)) { 1891 /* 1892 * Fail the request if a DMA mask between 32 and 64 bits 1893 * was requested but couldn't be fulfilled. Ideally we 1894 * would do this for 64-bits but historically we have 1895 * always fallen back to 32-bits. 1896 */ 1897 return -ENOMEM; 1898 } else { 1899 dev_info(&pdev->dev, "Using 32-bit DMA via iommu\n"); 1900 set_dma_ops(&pdev->dev, &dma_iommu_ops); 1901 } 1902 } 1903 *pdev->dev.dma_mask = dma_mask; 1904 1905 /* Update peer npu devices */ 1906 pnv_npu_try_dma_set_bypass(pdev, bypass); 1907 1908 return 0; 1909 } 1910 1911 static u64 pnv_pci_ioda_dma_get_required_mask(struct pci_dev *pdev) 1912 { 1913 struct pci_controller *hose = pci_bus_to_host(pdev->bus); 1914 struct pnv_phb *phb = hose->private_data; 1915 struct pci_dn *pdn = pci_get_pdn(pdev); 1916 struct pnv_ioda_pe *pe; 1917 u64 end, mask; 1918 1919 if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE)) 1920 return 0; 1921 1922 pe = &phb->ioda.pe_array[pdn->pe_number]; 1923 if (!pe->tce_bypass_enabled) 1924 return __dma_get_required_mask(&pdev->dev); 1925 1926 1927 end = pe->tce_bypass_base + memblock_end_of_DRAM(); 1928 mask = 1ULL << (fls64(end) - 1); 1929 mask += mask - 1; 1930 1931 return mask; 1932 } 1933 1934 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe, struct pci_bus *bus) 1935 { 1936 struct pci_dev *dev; 1937 1938 list_for_each_entry(dev, &bus->devices, bus_list) { 1939 set_iommu_table_base(&dev->dev, pe->table_group.tables[0]); 1940 set_dma_offset(&dev->dev, pe->tce_bypass_base); 1941 1942 if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate) 1943 pnv_ioda_setup_bus_dma(pe, dev->subordinate); 1944 } 1945 } 1946 1947 static inline __be64 __iomem *pnv_ioda_get_inval_reg(struct pnv_phb *phb, 1948 bool real_mode) 1949 { 1950 return real_mode ? (__be64 __iomem *)(phb->regs_phys + 0x210) : 1951 (phb->regs + 0x210); 1952 } 1953 1954 static void pnv_pci_p7ioc_tce_invalidate(struct iommu_table *tbl, 1955 unsigned long index, unsigned long npages, bool rm) 1956 { 1957 struct iommu_table_group_link *tgl = list_first_entry_or_null( 1958 &tbl->it_group_list, struct iommu_table_group_link, 1959 next); 1960 struct pnv_ioda_pe *pe = container_of(tgl->table_group, 1961 struct pnv_ioda_pe, table_group); 1962 __be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm); 1963 unsigned long start, end, inc; 1964 1965 start = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset); 1966 end = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset + 1967 npages - 1); 1968 1969 /* p7ioc-style invalidation, 2 TCEs per write */ 1970 start |= (1ull << 63); 1971 end |= (1ull << 63); 1972 inc = 16; 1973 end |= inc - 1; /* round up end to be different than start */ 1974 1975 mb(); /* Ensure above stores are visible */ 1976 while (start <= end) { 1977 if (rm) 1978 __raw_rm_writeq_be(start, invalidate); 1979 else 1980 __raw_writeq_be(start, invalidate); 1981 1982 start += inc; 1983 } 1984 1985 /* 1986 * The iommu layer will do another mb() for us on build() 1987 * and we don't care on free() 1988 */ 1989 } 1990 1991 static int pnv_ioda1_tce_build(struct iommu_table *tbl, long index, 1992 long npages, unsigned long uaddr, 1993 enum dma_data_direction direction, 1994 unsigned long attrs) 1995 { 1996 int ret = pnv_tce_build(tbl, index, npages, uaddr, direction, 1997 attrs); 1998 1999 if (!ret) 2000 pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false); 2001 2002 return ret; 2003 } 2004 2005 #ifdef CONFIG_IOMMU_API 2006 static int pnv_ioda1_tce_xchg(struct iommu_table *tbl, long index, 2007 unsigned long *hpa, enum dma_data_direction *direction) 2008 { 2009 long ret = pnv_tce_xchg(tbl, index, hpa, direction, true); 2010 2011 if (!ret) 2012 pnv_pci_p7ioc_tce_invalidate(tbl, index, 1, false); 2013 2014 return ret; 2015 } 2016 2017 static int pnv_ioda1_tce_xchg_rm(struct iommu_table *tbl, long index, 2018 unsigned long *hpa, enum dma_data_direction *direction) 2019 { 2020 long ret = pnv_tce_xchg(tbl, index, hpa, direction, false); 2021 2022 if (!ret) 2023 pnv_pci_p7ioc_tce_invalidate(tbl, index, 1, true); 2024 2025 return ret; 2026 } 2027 #endif 2028 2029 static void pnv_ioda1_tce_free(struct iommu_table *tbl, long index, 2030 long npages) 2031 { 2032 pnv_tce_free(tbl, index, npages); 2033 2034 pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false); 2035 } 2036 2037 static struct iommu_table_ops pnv_ioda1_iommu_ops = { 2038 .set = pnv_ioda1_tce_build, 2039 #ifdef CONFIG_IOMMU_API 2040 .exchange = pnv_ioda1_tce_xchg, 2041 .exchange_rm = pnv_ioda1_tce_xchg_rm, 2042 .useraddrptr = pnv_tce_useraddrptr, 2043 #endif 2044 .clear = pnv_ioda1_tce_free, 2045 .get = pnv_tce_get, 2046 }; 2047 2048 #define PHB3_TCE_KILL_INVAL_ALL PPC_BIT(0) 2049 #define PHB3_TCE_KILL_INVAL_PE PPC_BIT(1) 2050 #define PHB3_TCE_KILL_INVAL_ONE PPC_BIT(2) 2051 2052 static void pnv_pci_phb3_tce_invalidate_entire(struct pnv_phb *phb, bool rm) 2053 { 2054 __be64 __iomem *invalidate = pnv_ioda_get_inval_reg(phb, rm); 2055 const unsigned long val = PHB3_TCE_KILL_INVAL_ALL; 2056 2057 mb(); /* Ensure previous TCE table stores are visible */ 2058 if (rm) 2059 __raw_rm_writeq_be(val, invalidate); 2060 else 2061 __raw_writeq_be(val, invalidate); 2062 } 2063 2064 static inline void pnv_pci_phb3_tce_invalidate_pe(struct pnv_ioda_pe *pe) 2065 { 2066 /* 01xb - invalidate TCEs that match the specified PE# */ 2067 __be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, false); 2068 unsigned long val = PHB3_TCE_KILL_INVAL_PE | (pe->pe_number & 0xFF); 2069 2070 mb(); /* Ensure above stores are visible */ 2071 __raw_writeq_be(val, invalidate); 2072 } 2073 2074 static void pnv_pci_phb3_tce_invalidate(struct pnv_ioda_pe *pe, bool rm, 2075 unsigned shift, unsigned long index, 2076 unsigned long npages) 2077 { 2078 __be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm); 2079 unsigned long start, end, inc; 2080 2081 /* We'll invalidate DMA address in PE scope */ 2082 start = PHB3_TCE_KILL_INVAL_ONE; 2083 start |= (pe->pe_number & 0xFF); 2084 end = start; 2085 2086 /* Figure out the start, end and step */ 2087 start |= (index << shift); 2088 end |= ((index + npages - 1) << shift); 2089 inc = (0x1ull << shift); 2090 mb(); 2091 2092 while (start <= end) { 2093 if (rm) 2094 __raw_rm_writeq_be(start, invalidate); 2095 else 2096 __raw_writeq_be(start, invalidate); 2097 start += inc; 2098 } 2099 } 2100 2101 static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe *pe) 2102 { 2103 struct pnv_phb *phb = pe->phb; 2104 2105 if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs) 2106 pnv_pci_phb3_tce_invalidate_pe(pe); 2107 else 2108 opal_pci_tce_kill(phb->opal_id, OPAL_PCI_TCE_KILL_PE, 2109 pe->pe_number, 0, 0, 0); 2110 } 2111 2112 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table *tbl, 2113 unsigned long index, unsigned long npages, bool rm) 2114 { 2115 struct iommu_table_group_link *tgl; 2116 2117 list_for_each_entry_lockless(tgl, &tbl->it_group_list, next) { 2118 struct pnv_ioda_pe *pe = container_of(tgl->table_group, 2119 struct pnv_ioda_pe, table_group); 2120 struct pnv_phb *phb = pe->phb; 2121 unsigned int shift = tbl->it_page_shift; 2122 2123 /* 2124 * NVLink1 can use the TCE kill register directly as 2125 * it's the same as PHB3. NVLink2 is different and 2126 * should go via the OPAL call. 2127 */ 2128 if (phb->model == PNV_PHB_MODEL_NPU) { 2129 /* 2130 * The NVLink hardware does not support TCE kill 2131 * per TCE entry so we have to invalidate 2132 * the entire cache for it. 2133 */ 2134 pnv_pci_phb3_tce_invalidate_entire(phb, rm); 2135 continue; 2136 } 2137 if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs) 2138 pnv_pci_phb3_tce_invalidate(pe, rm, shift, 2139 index, npages); 2140 else 2141 opal_pci_tce_kill(phb->opal_id, 2142 OPAL_PCI_TCE_KILL_PAGES, 2143 pe->pe_number, 1u << shift, 2144 index << shift, npages); 2145 } 2146 } 2147 2148 void pnv_pci_ioda2_tce_invalidate_entire(struct pnv_phb *phb, bool rm) 2149 { 2150 if (phb->model == PNV_PHB_MODEL_NPU || phb->model == PNV_PHB_MODEL_PHB3) 2151 pnv_pci_phb3_tce_invalidate_entire(phb, rm); 2152 else 2153 opal_pci_tce_kill(phb->opal_id, OPAL_PCI_TCE_KILL, 0, 0, 0, 0); 2154 } 2155 2156 static int pnv_ioda2_tce_build(struct iommu_table *tbl, long index, 2157 long npages, unsigned long uaddr, 2158 enum dma_data_direction direction, 2159 unsigned long attrs) 2160 { 2161 int ret = pnv_tce_build(tbl, index, npages, uaddr, direction, 2162 attrs); 2163 2164 if (!ret) 2165 pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false); 2166 2167 return ret; 2168 } 2169 2170 #ifdef CONFIG_IOMMU_API 2171 static int pnv_ioda2_tce_xchg(struct iommu_table *tbl, long index, 2172 unsigned long *hpa, enum dma_data_direction *direction) 2173 { 2174 long ret = pnv_tce_xchg(tbl, index, hpa, direction, true); 2175 2176 if (!ret) 2177 pnv_pci_ioda2_tce_invalidate(tbl, index, 1, false); 2178 2179 return ret; 2180 } 2181 2182 static int pnv_ioda2_tce_xchg_rm(struct iommu_table *tbl, long index, 2183 unsigned long *hpa, enum dma_data_direction *direction) 2184 { 2185 long ret = pnv_tce_xchg(tbl, index, hpa, direction, false); 2186 2187 if (!ret) 2188 pnv_pci_ioda2_tce_invalidate(tbl, index, 1, true); 2189 2190 return ret; 2191 } 2192 #endif 2193 2194 static void pnv_ioda2_tce_free(struct iommu_table *tbl, long index, 2195 long npages) 2196 { 2197 pnv_tce_free(tbl, index, npages); 2198 2199 pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false); 2200 } 2201 2202 static struct iommu_table_ops pnv_ioda2_iommu_ops = { 2203 .set = pnv_ioda2_tce_build, 2204 #ifdef CONFIG_IOMMU_API 2205 .exchange = pnv_ioda2_tce_xchg, 2206 .exchange_rm = pnv_ioda2_tce_xchg_rm, 2207 .useraddrptr = pnv_tce_useraddrptr, 2208 #endif 2209 .clear = pnv_ioda2_tce_free, 2210 .get = pnv_tce_get, 2211 .free = pnv_pci_ioda2_table_free_pages, 2212 }; 2213 2214 static int pnv_pci_ioda_dev_dma_weight(struct pci_dev *dev, void *data) 2215 { 2216 unsigned int *weight = (unsigned int *)data; 2217 2218 /* This is quite simplistic. The "base" weight of a device 2219 * is 10. 0 means no DMA is to be accounted for it. 2220 */ 2221 if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL) 2222 return 0; 2223 2224 if (dev->class == PCI_CLASS_SERIAL_USB_UHCI || 2225 dev->class == PCI_CLASS_SERIAL_USB_OHCI || 2226 dev->class == PCI_CLASS_SERIAL_USB_EHCI) 2227 *weight += 3; 2228 else if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID) 2229 *weight += 15; 2230 else 2231 *weight += 10; 2232 2233 return 0; 2234 } 2235 2236 static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe *pe) 2237 { 2238 unsigned int weight = 0; 2239 2240 /* SRIOV VF has same DMA32 weight as its PF */ 2241 #ifdef CONFIG_PCI_IOV 2242 if ((pe->flags & PNV_IODA_PE_VF) && pe->parent_dev) { 2243 pnv_pci_ioda_dev_dma_weight(pe->parent_dev, &weight); 2244 return weight; 2245 } 2246 #endif 2247 2248 if ((pe->flags & PNV_IODA_PE_DEV) && pe->pdev) { 2249 pnv_pci_ioda_dev_dma_weight(pe->pdev, &weight); 2250 } else if ((pe->flags & PNV_IODA_PE_BUS) && pe->pbus) { 2251 struct pci_dev *pdev; 2252 2253 list_for_each_entry(pdev, &pe->pbus->devices, bus_list) 2254 pnv_pci_ioda_dev_dma_weight(pdev, &weight); 2255 } else if ((pe->flags & PNV_IODA_PE_BUS_ALL) && pe->pbus) { 2256 pci_walk_bus(pe->pbus, pnv_pci_ioda_dev_dma_weight, &weight); 2257 } 2258 2259 return weight; 2260 } 2261 2262 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb *phb, 2263 struct pnv_ioda_pe *pe) 2264 { 2265 2266 struct page *tce_mem = NULL; 2267 struct iommu_table *tbl; 2268 unsigned int weight, total_weight = 0; 2269 unsigned int tce32_segsz, base, segs, avail, i; 2270 int64_t rc; 2271 void *addr; 2272 2273 /* XXX FIXME: Handle 64-bit only DMA devices */ 2274 /* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */ 2275 /* XXX FIXME: Allocate multi-level tables on PHB3 */ 2276 weight = pnv_pci_ioda_pe_dma_weight(pe); 2277 if (!weight) 2278 return; 2279 2280 pci_walk_bus(phb->hose->bus, pnv_pci_ioda_dev_dma_weight, 2281 &total_weight); 2282 segs = (weight * phb->ioda.dma32_count) / total_weight; 2283 if (!segs) 2284 segs = 1; 2285 2286 /* 2287 * Allocate contiguous DMA32 segments. We begin with the expected 2288 * number of segments. With one more attempt, the number of DMA32 2289 * segments to be allocated is decreased by one until one segment 2290 * is allocated successfully. 2291 */ 2292 do { 2293 for (base = 0; base <= phb->ioda.dma32_count - segs; base++) { 2294 for (avail = 0, i = base; i < base + segs; i++) { 2295 if (phb->ioda.dma32_segmap[i] == 2296 IODA_INVALID_PE) 2297 avail++; 2298 } 2299 2300 if (avail == segs) 2301 goto found; 2302 } 2303 } while (--segs); 2304 2305 if (!segs) { 2306 pe_warn(pe, "No available DMA32 segments\n"); 2307 return; 2308 } 2309 2310 found: 2311 tbl = pnv_pci_table_alloc(phb->hose->node); 2312 if (WARN_ON(!tbl)) 2313 return; 2314 2315 iommu_register_group(&pe->table_group, phb->hose->global_number, 2316 pe->pe_number); 2317 pnv_pci_link_table_and_group(phb->hose->node, 0, tbl, &pe->table_group); 2318 2319 /* Grab a 32-bit TCE table */ 2320 pe_info(pe, "DMA weight %d (%d), assigned (%d) %d DMA32 segments\n", 2321 weight, total_weight, base, segs); 2322 pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n", 2323 base * PNV_IODA1_DMA32_SEGSIZE, 2324 (base + segs) * PNV_IODA1_DMA32_SEGSIZE - 1); 2325 2326 /* XXX Currently, we allocate one big contiguous table for the 2327 * TCEs. We only really need one chunk per 256M of TCE space 2328 * (ie per segment) but that's an optimization for later, it 2329 * requires some added smarts with our get/put_tce implementation 2330 * 2331 * Each TCE page is 4KB in size and each TCE entry occupies 8 2332 * bytes 2333 */ 2334 tce32_segsz = PNV_IODA1_DMA32_SEGSIZE >> (IOMMU_PAGE_SHIFT_4K - 3); 2335 tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL, 2336 get_order(tce32_segsz * segs)); 2337 if (!tce_mem) { 2338 pe_err(pe, " Failed to allocate a 32-bit TCE memory\n"); 2339 goto fail; 2340 } 2341 addr = page_address(tce_mem); 2342 memset(addr, 0, tce32_segsz * segs); 2343 2344 /* Configure HW */ 2345 for (i = 0; i < segs; i++) { 2346 rc = opal_pci_map_pe_dma_window(phb->opal_id, 2347 pe->pe_number, 2348 base + i, 1, 2349 __pa(addr) + tce32_segsz * i, 2350 tce32_segsz, IOMMU_PAGE_SIZE_4K); 2351 if (rc) { 2352 pe_err(pe, " Failed to configure 32-bit TCE table," 2353 " err %ld\n", rc); 2354 goto fail; 2355 } 2356 } 2357 2358 /* Setup DMA32 segment mapping */ 2359 for (i = base; i < base + segs; i++) 2360 phb->ioda.dma32_segmap[i] = pe->pe_number; 2361 2362 /* Setup linux iommu table */ 2363 pnv_pci_setup_iommu_table(tbl, addr, tce32_segsz * segs, 2364 base * PNV_IODA1_DMA32_SEGSIZE, 2365 IOMMU_PAGE_SHIFT_4K); 2366 2367 tbl->it_ops = &pnv_ioda1_iommu_ops; 2368 pe->table_group.tce32_start = tbl->it_offset << tbl->it_page_shift; 2369 pe->table_group.tce32_size = tbl->it_size << tbl->it_page_shift; 2370 iommu_init_table(tbl, phb->hose->node); 2371 2372 if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)) 2373 pnv_ioda_setup_bus_dma(pe, pe->pbus); 2374 2375 return; 2376 fail: 2377 /* XXX Failure: Try to fallback to 64-bit only ? */ 2378 if (tce_mem) 2379 __free_pages(tce_mem, get_order(tce32_segsz * segs)); 2380 if (tbl) { 2381 pnv_pci_unlink_table_and_group(tbl, &pe->table_group); 2382 iommu_tce_table_put(tbl); 2383 } 2384 } 2385 2386 static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group, 2387 int num, struct iommu_table *tbl) 2388 { 2389 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe, 2390 table_group); 2391 struct pnv_phb *phb = pe->phb; 2392 int64_t rc; 2393 const unsigned long size = tbl->it_indirect_levels ? 2394 tbl->it_level_size : tbl->it_size; 2395 const __u64 start_addr = tbl->it_offset << tbl->it_page_shift; 2396 const __u64 win_size = tbl->it_size << tbl->it_page_shift; 2397 2398 pe_info(pe, "Setting up window#%d %llx..%llx pg=%x\n", num, 2399 start_addr, start_addr + win_size - 1, 2400 IOMMU_PAGE_SIZE(tbl)); 2401 2402 /* 2403 * Map TCE table through TVT. The TVE index is the PE number 2404 * shifted by 1 bit for 32-bits DMA space. 2405 */ 2406 rc = opal_pci_map_pe_dma_window(phb->opal_id, 2407 pe->pe_number, 2408 (pe->pe_number << 1) + num, 2409 tbl->it_indirect_levels + 1, 2410 __pa(tbl->it_base), 2411 size << 3, 2412 IOMMU_PAGE_SIZE(tbl)); 2413 if (rc) { 2414 pe_err(pe, "Failed to configure TCE table, err %ld\n", rc); 2415 return rc; 2416 } 2417 2418 pnv_pci_link_table_and_group(phb->hose->node, num, 2419 tbl, &pe->table_group); 2420 pnv_pci_ioda2_tce_invalidate_pe(pe); 2421 2422 return 0; 2423 } 2424 2425 void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable) 2426 { 2427 uint16_t window_id = (pe->pe_number << 1 ) + 1; 2428 int64_t rc; 2429 2430 pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis"); 2431 if (enable) { 2432 phys_addr_t top = memblock_end_of_DRAM(); 2433 2434 top = roundup_pow_of_two(top); 2435 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id, 2436 pe->pe_number, 2437 window_id, 2438 pe->tce_bypass_base, 2439 top); 2440 } else { 2441 rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id, 2442 pe->pe_number, 2443 window_id, 2444 pe->tce_bypass_base, 2445 0); 2446 } 2447 if (rc) 2448 pe_err(pe, "OPAL error %lld configuring bypass window\n", rc); 2449 else 2450 pe->tce_bypass_enabled = enable; 2451 } 2452 2453 static long pnv_pci_ioda2_create_table(struct iommu_table_group *table_group, 2454 int num, __u32 page_shift, __u64 window_size, __u32 levels, 2455 bool alloc_userspace_copy, struct iommu_table **ptbl) 2456 { 2457 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe, 2458 table_group); 2459 int nid = pe->phb->hose->node; 2460 __u64 bus_offset = num ? pe->tce_bypass_base : table_group->tce32_start; 2461 long ret; 2462 struct iommu_table *tbl; 2463 2464 tbl = pnv_pci_table_alloc(nid); 2465 if (!tbl) 2466 return -ENOMEM; 2467 2468 tbl->it_ops = &pnv_ioda2_iommu_ops; 2469 2470 ret = pnv_pci_ioda2_table_alloc_pages(nid, 2471 bus_offset, page_shift, window_size, 2472 levels, alloc_userspace_copy, tbl); 2473 if (ret) { 2474 iommu_tce_table_put(tbl); 2475 return ret; 2476 } 2477 2478 *ptbl = tbl; 2479 2480 return 0; 2481 } 2482 2483 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe *pe) 2484 { 2485 struct iommu_table *tbl = NULL; 2486 long rc; 2487 2488 /* 2489 * crashkernel= specifies the kdump kernel's maximum memory at 2490 * some offset and there is no guaranteed the result is a power 2491 * of 2, which will cause errors later. 2492 */ 2493 const u64 max_memory = __rounddown_pow_of_two(memory_hotplug_max()); 2494 2495 /* 2496 * In memory constrained environments, e.g. kdump kernel, the 2497 * DMA window can be larger than available memory, which will 2498 * cause errors later. 2499 */ 2500 const u64 window_size = min((u64)pe->table_group.tce32_size, max_memory); 2501 2502 rc = pnv_pci_ioda2_create_table(&pe->table_group, 0, 2503 IOMMU_PAGE_SHIFT_4K, 2504 window_size, 2505 POWERNV_IOMMU_DEFAULT_LEVELS, false, &tbl); 2506 if (rc) { 2507 pe_err(pe, "Failed to create 32-bit TCE table, err %ld", 2508 rc); 2509 return rc; 2510 } 2511 2512 iommu_init_table(tbl, pe->phb->hose->node); 2513 2514 rc = pnv_pci_ioda2_set_window(&pe->table_group, 0, tbl); 2515 if (rc) { 2516 pe_err(pe, "Failed to configure 32-bit TCE table, err %ld\n", 2517 rc); 2518 iommu_tce_table_put(tbl); 2519 return rc; 2520 } 2521 2522 if (!pnv_iommu_bypass_disabled) 2523 pnv_pci_ioda2_set_bypass(pe, true); 2524 2525 return 0; 2526 } 2527 2528 #if defined(CONFIG_IOMMU_API) || defined(CONFIG_PCI_IOV) 2529 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group, 2530 int num) 2531 { 2532 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe, 2533 table_group); 2534 struct pnv_phb *phb = pe->phb; 2535 long ret; 2536 2537 pe_info(pe, "Removing DMA window #%d\n", num); 2538 2539 ret = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number, 2540 (pe->pe_number << 1) + num, 2541 0/* levels */, 0/* table address */, 2542 0/* table size */, 0/* page size */); 2543 if (ret) 2544 pe_warn(pe, "Unmapping failed, ret = %ld\n", ret); 2545 else 2546 pnv_pci_ioda2_tce_invalidate_pe(pe); 2547 2548 pnv_pci_unlink_table_and_group(table_group->tables[num], table_group); 2549 2550 return ret; 2551 } 2552 #endif 2553 2554 #ifdef CONFIG_IOMMU_API 2555 static unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift, 2556 __u64 window_size, __u32 levels) 2557 { 2558 unsigned long bytes = 0; 2559 const unsigned window_shift = ilog2(window_size); 2560 unsigned entries_shift = window_shift - page_shift; 2561 unsigned table_shift = entries_shift + 3; 2562 unsigned long tce_table_size = max(0x1000UL, 1UL << table_shift); 2563 unsigned long direct_table_size; 2564 2565 if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS) || 2566 !is_power_of_2(window_size)) 2567 return 0; 2568 2569 /* Calculate a direct table size from window_size and levels */ 2570 entries_shift = (entries_shift + levels - 1) / levels; 2571 table_shift = entries_shift + 3; 2572 table_shift = max_t(unsigned, table_shift, PAGE_SHIFT); 2573 direct_table_size = 1UL << table_shift; 2574 2575 for ( ; levels; --levels) { 2576 bytes += _ALIGN_UP(tce_table_size, direct_table_size); 2577 2578 tce_table_size /= direct_table_size; 2579 tce_table_size <<= 3; 2580 tce_table_size = max_t(unsigned long, 2581 tce_table_size, direct_table_size); 2582 } 2583 2584 return bytes + bytes; /* one for HW table, one for userspace copy */ 2585 } 2586 2587 static long pnv_pci_ioda2_create_table_userspace( 2588 struct iommu_table_group *table_group, 2589 int num, __u32 page_shift, __u64 window_size, __u32 levels, 2590 struct iommu_table **ptbl) 2591 { 2592 return pnv_pci_ioda2_create_table(table_group, 2593 num, page_shift, window_size, levels, true, ptbl); 2594 } 2595 2596 static void pnv_ioda2_take_ownership(struct iommu_table_group *table_group) 2597 { 2598 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe, 2599 table_group); 2600 /* Store @tbl as pnv_pci_ioda2_unset_window() resets it */ 2601 struct iommu_table *tbl = pe->table_group.tables[0]; 2602 2603 pnv_pci_ioda2_set_bypass(pe, false); 2604 pnv_pci_ioda2_unset_window(&pe->table_group, 0); 2605 if (pe->pbus) 2606 pnv_ioda_setup_bus_dma(pe, pe->pbus); 2607 iommu_tce_table_put(tbl); 2608 } 2609 2610 static void pnv_ioda2_release_ownership(struct iommu_table_group *table_group) 2611 { 2612 struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe, 2613 table_group); 2614 2615 pnv_pci_ioda2_setup_default_config(pe); 2616 if (pe->pbus) 2617 pnv_ioda_setup_bus_dma(pe, pe->pbus); 2618 } 2619 2620 static struct iommu_table_group_ops pnv_pci_ioda2_ops = { 2621 .get_table_size = pnv_pci_ioda2_get_table_size, 2622 .create_table = pnv_pci_ioda2_create_table_userspace, 2623 .set_window = pnv_pci_ioda2_set_window, 2624 .unset_window = pnv_pci_ioda2_unset_window, 2625 .take_ownership = pnv_ioda2_take_ownership, 2626 .release_ownership = pnv_ioda2_release_ownership, 2627 }; 2628 2629 static int gpe_table_group_to_npe_cb(struct device *dev, void *opaque) 2630 { 2631 struct pci_controller *hose; 2632 struct pnv_phb *phb; 2633 struct pnv_ioda_pe **ptmppe = opaque; 2634 struct pci_dev *pdev = container_of(dev, struct pci_dev, dev); 2635 struct pci_dn *pdn = pci_get_pdn(pdev); 2636 2637 if (!pdn || pdn->pe_number == IODA_INVALID_PE) 2638 return 0; 2639 2640 hose = pci_bus_to_host(pdev->bus); 2641 phb = hose->private_data; 2642 if (phb->type != PNV_PHB_NPU_NVLINK) 2643 return 0; 2644 2645 *ptmppe = &phb->ioda.pe_array[pdn->pe_number]; 2646 2647 return 1; 2648 } 2649 2650 /* 2651 * This returns PE of associated NPU. 2652 * This assumes that NPU is in the same IOMMU group with GPU and there is 2653 * no other PEs. 2654 */ 2655 static struct pnv_ioda_pe *gpe_table_group_to_npe( 2656 struct iommu_table_group *table_group) 2657 { 2658 struct pnv_ioda_pe *npe = NULL; 2659 int ret = iommu_group_for_each_dev(table_group->group, &npe, 2660 gpe_table_group_to_npe_cb); 2661 2662 BUG_ON(!ret || !npe); 2663 2664 return npe; 2665 } 2666 2667 static long pnv_pci_ioda2_npu_set_window(struct iommu_table_group *table_group, 2668 int num, struct iommu_table *tbl) 2669 { 2670 struct pnv_ioda_pe *npe = gpe_table_group_to_npe(table_group); 2671 long ret = pnv_pci_ioda2_set_window(table_group, num, tbl); 2672 2673 if (ret) 2674 return ret; 2675 2676 ret = npe->table_group.ops->set_window(&npe->table_group, num, tbl); 2677 if (ret) 2678 pnv_pci_ioda2_unset_window(table_group, num); 2679 2680 return ret; 2681 } 2682 2683 static long pnv_pci_ioda2_npu_unset_window( 2684 struct iommu_table_group *table_group, 2685 int num) 2686 { 2687 struct pnv_ioda_pe *npe = gpe_table_group_to_npe(table_group); 2688 long ret = pnv_pci_ioda2_unset_window(table_group, num); 2689 2690 if (ret) 2691 return ret; 2692 2693 return npe->table_group.ops->unset_window(&npe->table_group, num); 2694 } 2695 2696 static void pnv_ioda2_npu_take_ownership(struct iommu_table_group *table_group) 2697 { 2698 struct pnv_ioda_pe *npe = gpe_table_group_to_npe(table_group); 2699 2700 npe->table_group.ops->take_ownership(&npe->table_group); 2701 pnv_ioda2_take_ownership(table_group); 2702 } 2703 2704 static struct iommu_table_group_ops pnv_pci_ioda2_npu_ops = { 2705 .get_table_size = pnv_pci_ioda2_get_table_size, 2706 .create_table = pnv_pci_ioda2_create_table_userspace, 2707 .set_window = pnv_pci_ioda2_npu_set_window, 2708 .unset_window = pnv_pci_ioda2_npu_unset_window, 2709 .take_ownership = pnv_ioda2_npu_take_ownership, 2710 .release_ownership = pnv_ioda2_release_ownership, 2711 }; 2712 2713 static void pnv_ioda_setup_bus_iommu_group_add_devices(struct pnv_ioda_pe *pe, 2714 struct pci_bus *bus) 2715 { 2716 struct pci_dev *dev; 2717 2718 list_for_each_entry(dev, &bus->devices, bus_list) { 2719 iommu_add_device(&pe->table_group, &dev->dev); 2720 2721 if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate) 2722 pnv_ioda_setup_bus_iommu_group_add_devices(pe, 2723 dev->subordinate); 2724 } 2725 } 2726 2727 static void pnv_ioda_setup_bus_iommu_group(struct pnv_ioda_pe *pe) 2728 { 2729 if (!pnv_pci_ioda_pe_dma_weight(pe)) 2730 return; 2731 2732 iommu_register_group(&pe->table_group, pe->phb->hose->global_number, 2733 pe->pe_number); 2734 2735 /* 2736 * set_iommu_table_base(&pe->pdev->dev, tbl) should have been called 2737 * by now 2738 */ 2739 if (pe->flags & PNV_IODA_PE_DEV) 2740 iommu_add_device(&pe->table_group, &pe->pdev->dev); 2741 else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)) 2742 pnv_ioda_setup_bus_iommu_group_add_devices(pe, pe->pbus); 2743 } 2744 2745 static void pnv_pci_ioda_setup_iommu_api(void) 2746 { 2747 struct pci_controller *hose, *tmp; 2748 struct pnv_phb *phb; 2749 struct pnv_ioda_pe *pe, *gpe; 2750 2751 /* 2752 * There are 4 types of PEs: 2753 * - PNV_IODA_PE_BUS: a downstream port with an adapter, 2754 * created from pnv_pci_setup_bridge(); 2755 * - PNV_IODA_PE_BUS_ALL: a PCI-PCIX bridge with devices behind it, 2756 * created from pnv_pci_setup_bridge(); 2757 * - PNV_IODA_PE_VF: a SRIOV virtual function, 2758 * created from pnv_pcibios_sriov_enable(); 2759 * - PNV_IODA_PE_DEV: an NPU or OCAPI device, 2760 * created from pnv_pci_ioda_fixup(). 2761 * 2762 * Normally a PE is represented by an IOMMU group, however for 2763 * devices with side channels the groups need to be more strict. 2764 */ 2765 list_for_each_entry(hose, &hose_list, list_node) { 2766 phb = hose->private_data; 2767 2768 if (phb->type == PNV_PHB_NPU_NVLINK) 2769 continue; 2770 2771 list_for_each_entry(pe, &phb->ioda.pe_list, list) 2772 pnv_ioda_setup_bus_iommu_group(pe); 2773 } 2774 2775 /* 2776 * Now we have all PHBs discovered, time to add NPU devices to 2777 * the corresponding IOMMU groups. 2778 */ 2779 list_for_each_entry_safe(hose, tmp, &hose_list, list_node) { 2780 phb = hose->private_data; 2781 2782 if (phb->type != PNV_PHB_NPU_NVLINK) 2783 continue; 2784 2785 list_for_each_entry(pe, &phb->ioda.pe_list, list) { 2786 gpe = pnv_pci_npu_setup_iommu(pe); 2787 if (gpe) 2788 gpe->table_group.ops = &pnv_pci_ioda2_npu_ops; 2789 } 2790 } 2791 } 2792 #else /* !CONFIG_IOMMU_API */ 2793 static void pnv_pci_ioda_setup_iommu_api(void) { }; 2794 #endif 2795 2796 static unsigned long pnv_ioda_parse_tce_sizes(struct pnv_phb *phb) 2797 { 2798 struct pci_controller *hose = phb->hose; 2799 struct device_node *dn = hose->dn; 2800 unsigned long mask = 0; 2801 int i, rc, count; 2802 u32 val; 2803 2804 count = of_property_count_u32_elems(dn, "ibm,supported-tce-sizes"); 2805 if (count <= 0) { 2806 mask = SZ_4K | SZ_64K; 2807 /* Add 16M for POWER8 by default */ 2808 if (cpu_has_feature(CPU_FTR_ARCH_207S) && 2809 !cpu_has_feature(CPU_FTR_ARCH_300)) 2810 mask |= SZ_16M | SZ_256M; 2811 return mask; 2812 } 2813 2814 for (i = 0; i < count; i++) { 2815 rc = of_property_read_u32_index(dn, "ibm,supported-tce-sizes", 2816 i, &val); 2817 if (rc == 0) 2818 mask |= 1ULL << val; 2819 } 2820 2821 return mask; 2822 } 2823 2824 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb, 2825 struct pnv_ioda_pe *pe) 2826 { 2827 int64_t rc; 2828 2829 if (!pnv_pci_ioda_pe_dma_weight(pe)) 2830 return; 2831 2832 /* TVE #1 is selected by PCI address bit 59 */ 2833 pe->tce_bypass_base = 1ull << 59; 2834 2835 /* The PE will reserve all possible 32-bits space */ 2836 pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n", 2837 phb->ioda.m32_pci_base); 2838 2839 /* Setup linux iommu table */ 2840 pe->table_group.tce32_start = 0; 2841 pe->table_group.tce32_size = phb->ioda.m32_pci_base; 2842 pe->table_group.max_dynamic_windows_supported = 2843 IOMMU_TABLE_GROUP_MAX_TABLES; 2844 pe->table_group.max_levels = POWERNV_IOMMU_MAX_LEVELS; 2845 pe->table_group.pgsizes = pnv_ioda_parse_tce_sizes(phb); 2846 #ifdef CONFIG_IOMMU_API 2847 pe->table_group.ops = &pnv_pci_ioda2_ops; 2848 #endif 2849 2850 rc = pnv_pci_ioda2_setup_default_config(pe); 2851 if (rc) 2852 return; 2853 2854 if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)) 2855 pnv_ioda_setup_bus_dma(pe, pe->pbus); 2856 } 2857 2858 int64_t pnv_opal_pci_msi_eoi(struct irq_chip *chip, unsigned int hw_irq) 2859 { 2860 struct pnv_phb *phb = container_of(chip, struct pnv_phb, 2861 ioda.irq_chip); 2862 2863 return opal_pci_msi_eoi(phb->opal_id, hw_irq); 2864 } 2865 2866 static void pnv_ioda2_msi_eoi(struct irq_data *d) 2867 { 2868 int64_t rc; 2869 unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d); 2870 struct irq_chip *chip = irq_data_get_irq_chip(d); 2871 2872 rc = pnv_opal_pci_msi_eoi(chip, hw_irq); 2873 WARN_ON_ONCE(rc); 2874 2875 icp_native_eoi(d); 2876 } 2877 2878 2879 void pnv_set_msi_irq_chip(struct pnv_phb *phb, unsigned int virq) 2880 { 2881 struct irq_data *idata; 2882 struct irq_chip *ichip; 2883 2884 /* The MSI EOI OPAL call is only needed on PHB3 */ 2885 if (phb->model != PNV_PHB_MODEL_PHB3) 2886 return; 2887 2888 if (!phb->ioda.irq_chip_init) { 2889 /* 2890 * First time we setup an MSI IRQ, we need to setup the 2891 * corresponding IRQ chip to route correctly. 2892 */ 2893 idata = irq_get_irq_data(virq); 2894 ichip = irq_data_get_irq_chip(idata); 2895 phb->ioda.irq_chip_init = 1; 2896 phb->ioda.irq_chip = *ichip; 2897 phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi; 2898 } 2899 irq_set_chip(virq, &phb->ioda.irq_chip); 2900 } 2901 2902 /* 2903 * Returns true iff chip is something that we could call 2904 * pnv_opal_pci_msi_eoi for. 2905 */ 2906 bool is_pnv_opal_msi(struct irq_chip *chip) 2907 { 2908 return chip->irq_eoi == pnv_ioda2_msi_eoi; 2909 } 2910 EXPORT_SYMBOL_GPL(is_pnv_opal_msi); 2911 2912 static int pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev, 2913 unsigned int hwirq, unsigned int virq, 2914 unsigned int is_64, struct msi_msg *msg) 2915 { 2916 struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev); 2917 unsigned int xive_num = hwirq - phb->msi_base; 2918 __be32 data; 2919 int rc; 2920 2921 /* No PE assigned ? bail out ... no MSI for you ! */ 2922 if (pe == NULL) 2923 return -ENXIO; 2924 2925 /* Check if we have an MVE */ 2926 if (pe->mve_number < 0) 2927 return -ENXIO; 2928 2929 /* Force 32-bit MSI on some broken devices */ 2930 if (dev->no_64bit_msi) 2931 is_64 = 0; 2932 2933 /* Assign XIVE to PE */ 2934 rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num); 2935 if (rc) { 2936 pr_warn("%s: OPAL error %d setting XIVE %d PE\n", 2937 pci_name(dev), rc, xive_num); 2938 return -EIO; 2939 } 2940 2941 if (is_64) { 2942 __be64 addr64; 2943 2944 rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1, 2945 &addr64, &data); 2946 if (rc) { 2947 pr_warn("%s: OPAL error %d getting 64-bit MSI data\n", 2948 pci_name(dev), rc); 2949 return -EIO; 2950 } 2951 msg->address_hi = be64_to_cpu(addr64) >> 32; 2952 msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful; 2953 } else { 2954 __be32 addr32; 2955 2956 rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1, 2957 &addr32, &data); 2958 if (rc) { 2959 pr_warn("%s: OPAL error %d getting 32-bit MSI data\n", 2960 pci_name(dev), rc); 2961 return -EIO; 2962 } 2963 msg->address_hi = 0; 2964 msg->address_lo = be32_to_cpu(addr32); 2965 } 2966 msg->data = be32_to_cpu(data); 2967 2968 pnv_set_msi_irq_chip(phb, virq); 2969 2970 pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d)," 2971 " address=%x_%08x data=%x PE# %x\n", 2972 pci_name(dev), is_64 ? "64" : "32", hwirq, xive_num, 2973 msg->address_hi, msg->address_lo, data, pe->pe_number); 2974 2975 return 0; 2976 } 2977 2978 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb) 2979 { 2980 unsigned int count; 2981 const __be32 *prop = of_get_property(phb->hose->dn, 2982 "ibm,opal-msi-ranges", NULL); 2983 if (!prop) { 2984 /* BML Fallback */ 2985 prop = of_get_property(phb->hose->dn, "msi-ranges", NULL); 2986 } 2987 if (!prop) 2988 return; 2989 2990 phb->msi_base = be32_to_cpup(prop); 2991 count = be32_to_cpup(prop + 1); 2992 if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) { 2993 pr_err("PCI %d: Failed to allocate MSI bitmap !\n", 2994 phb->hose->global_number); 2995 return; 2996 } 2997 2998 phb->msi_setup = pnv_pci_ioda_msi_setup; 2999 phb->msi32_support = 1; 3000 pr_info(" Allocated bitmap for %d MSIs (base IRQ 0x%x)\n", 3001 count, phb->msi_base); 3002 } 3003 3004 #ifdef CONFIG_PCI_IOV 3005 static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev *pdev) 3006 { 3007 struct pci_controller *hose = pci_bus_to_host(pdev->bus); 3008 struct pnv_phb *phb = hose->private_data; 3009 const resource_size_t gate = phb->ioda.m64_segsize >> 2; 3010 struct resource *res; 3011 int i; 3012 resource_size_t size, total_vf_bar_sz; 3013 struct pci_dn *pdn; 3014 int mul, total_vfs; 3015 3016 if (!pdev->is_physfn || pci_dev_is_added(pdev)) 3017 return; 3018 3019 pdn = pci_get_pdn(pdev); 3020 pdn->vfs_expanded = 0; 3021 pdn->m64_single_mode = false; 3022 3023 total_vfs = pci_sriov_get_totalvfs(pdev); 3024 mul = phb->ioda.total_pe_num; 3025 total_vf_bar_sz = 0; 3026 3027 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { 3028 res = &pdev->resource[i + PCI_IOV_RESOURCES]; 3029 if (!res->flags || res->parent) 3030 continue; 3031 if (!pnv_pci_is_m64_flags(res->flags)) { 3032 dev_warn(&pdev->dev, "Don't support SR-IOV with" 3033 " non M64 VF BAR%d: %pR. \n", 3034 i, res); 3035 goto truncate_iov; 3036 } 3037 3038 total_vf_bar_sz += pci_iov_resource_size(pdev, 3039 i + PCI_IOV_RESOURCES); 3040 3041 /* 3042 * If bigger than quarter of M64 segment size, just round up 3043 * power of two. 3044 * 3045 * Generally, one M64 BAR maps one IOV BAR. To avoid conflict 3046 * with other devices, IOV BAR size is expanded to be 3047 * (total_pe * VF_BAR_size). When VF_BAR_size is half of M64 3048 * segment size , the expanded size would equal to half of the 3049 * whole M64 space size, which will exhaust the M64 Space and 3050 * limit the system flexibility. This is a design decision to 3051 * set the boundary to quarter of the M64 segment size. 3052 */ 3053 if (total_vf_bar_sz > gate) { 3054 mul = roundup_pow_of_two(total_vfs); 3055 dev_info(&pdev->dev, 3056 "VF BAR Total IOV size %llx > %llx, roundup to %d VFs\n", 3057 total_vf_bar_sz, gate, mul); 3058 pdn->m64_single_mode = true; 3059 break; 3060 } 3061 } 3062 3063 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { 3064 res = &pdev->resource[i + PCI_IOV_RESOURCES]; 3065 if (!res->flags || res->parent) 3066 continue; 3067 3068 size = pci_iov_resource_size(pdev, i + PCI_IOV_RESOURCES); 3069 /* 3070 * On PHB3, the minimum size alignment of M64 BAR in single 3071 * mode is 32MB. 3072 */ 3073 if (pdn->m64_single_mode && (size < SZ_32M)) 3074 goto truncate_iov; 3075 dev_dbg(&pdev->dev, " Fixing VF BAR%d: %pR to\n", i, res); 3076 res->end = res->start + size * mul - 1; 3077 dev_dbg(&pdev->dev, " %pR\n", res); 3078 dev_info(&pdev->dev, "VF BAR%d: %pR (expanded to %d VFs for PE alignment)", 3079 i, res, mul); 3080 } 3081 pdn->vfs_expanded = mul; 3082 3083 return; 3084 3085 truncate_iov: 3086 /* To save MMIO space, IOV BAR is truncated. */ 3087 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) { 3088 res = &pdev->resource[i + PCI_IOV_RESOURCES]; 3089 res->flags = 0; 3090 res->end = res->start - 1; 3091 } 3092 } 3093 #endif /* CONFIG_PCI_IOV */ 3094 3095 static void pnv_ioda_setup_pe_res(struct pnv_ioda_pe *pe, 3096 struct resource *res) 3097 { 3098 struct pnv_phb *phb = pe->phb; 3099 struct pci_bus_region region; 3100 int index; 3101 int64_t rc; 3102 3103 if (!res || !res->flags || res->start > res->end) 3104 return; 3105 3106 if (res->flags & IORESOURCE_IO) { 3107 region.start = res->start - phb->ioda.io_pci_base; 3108 region.end = res->end - phb->ioda.io_pci_base; 3109 index = region.start / phb->ioda.io_segsize; 3110 3111 while (index < phb->ioda.total_pe_num && 3112 region.start <= region.end) { 3113 phb->ioda.io_segmap[index] = pe->pe_number; 3114 rc = opal_pci_map_pe_mmio_window(phb->opal_id, 3115 pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index); 3116 if (rc != OPAL_SUCCESS) { 3117 pr_err("%s: Error %lld mapping IO segment#%d to PE#%x\n", 3118 __func__, rc, index, pe->pe_number); 3119 break; 3120 } 3121 3122 region.start += phb->ioda.io_segsize; 3123 index++; 3124 } 3125 } else if ((res->flags & IORESOURCE_MEM) && 3126 !pnv_pci_is_m64(phb, res)) { 3127 region.start = res->start - 3128 phb->hose->mem_offset[0] - 3129 phb->ioda.m32_pci_base; 3130 region.end = res->end - 3131 phb->hose->mem_offset[0] - 3132 phb->ioda.m32_pci_base; 3133 index = region.start / phb->ioda.m32_segsize; 3134 3135 while (index < phb->ioda.total_pe_num && 3136 region.start <= region.end) { 3137 phb->ioda.m32_segmap[index] = pe->pe_number; 3138 rc = opal_pci_map_pe_mmio_window(phb->opal_id, 3139 pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index); 3140 if (rc != OPAL_SUCCESS) { 3141 pr_err("%s: Error %lld mapping M32 segment#%d to PE#%x", 3142 __func__, rc, index, pe->pe_number); 3143 break; 3144 } 3145 3146 region.start += phb->ioda.m32_segsize; 3147 index++; 3148 } 3149 } 3150 } 3151 3152 /* 3153 * This function is supposed to be called on basis of PE from top 3154 * to bottom style. So the the I/O or MMIO segment assigned to 3155 * parent PE could be overridden by its child PEs if necessary. 3156 */ 3157 static void pnv_ioda_setup_pe_seg(struct pnv_ioda_pe *pe) 3158 { 3159 struct pci_dev *pdev; 3160 int i; 3161 3162 /* 3163 * NOTE: We only care PCI bus based PE for now. For PCI 3164 * device based PE, for example SRIOV sensitive VF should 3165 * be figured out later. 3166 */ 3167 BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))); 3168 3169 list_for_each_entry(pdev, &pe->pbus->devices, bus_list) { 3170 for (i = 0; i <= PCI_ROM_RESOURCE; i++) 3171 pnv_ioda_setup_pe_res(pe, &pdev->resource[i]); 3172 3173 /* 3174 * If the PE contains all subordinate PCI buses, the 3175 * windows of the child bridges should be mapped to 3176 * the PE as well. 3177 */ 3178 if (!(pe->flags & PNV_IODA_PE_BUS_ALL) || !pci_is_bridge(pdev)) 3179 continue; 3180 for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++) 3181 pnv_ioda_setup_pe_res(pe, 3182 &pdev->resource[PCI_BRIDGE_RESOURCES + i]); 3183 } 3184 } 3185 3186 #ifdef CONFIG_DEBUG_FS 3187 static int pnv_pci_diag_data_set(void *data, u64 val) 3188 { 3189 struct pci_controller *hose; 3190 struct pnv_phb *phb; 3191 s64 ret; 3192 3193 if (val != 1ULL) 3194 return -EINVAL; 3195 3196 hose = (struct pci_controller *)data; 3197 if (!hose || !hose->private_data) 3198 return -ENODEV; 3199 3200 phb = hose->private_data; 3201 3202 /* Retrieve the diag data from firmware */ 3203 ret = opal_pci_get_phb_diag_data2(phb->opal_id, phb->diag_data, 3204 phb->diag_data_size); 3205 if (ret != OPAL_SUCCESS) 3206 return -EIO; 3207 3208 /* Print the diag data to the kernel log */ 3209 pnv_pci_dump_phb_diag_data(phb->hose, phb->diag_data); 3210 return 0; 3211 } 3212 3213 DEFINE_SIMPLE_ATTRIBUTE(pnv_pci_diag_data_fops, NULL, 3214 pnv_pci_diag_data_set, "%llu\n"); 3215 3216 #endif /* CONFIG_DEBUG_FS */ 3217 3218 static void pnv_pci_ioda_create_dbgfs(void) 3219 { 3220 #ifdef CONFIG_DEBUG_FS 3221 struct pci_controller *hose, *tmp; 3222 struct pnv_phb *phb; 3223 char name[16]; 3224 3225 list_for_each_entry_safe(hose, tmp, &hose_list, list_node) { 3226 phb = hose->private_data; 3227 3228 /* Notify initialization of PHB done */ 3229 phb->initialized = 1; 3230 3231 sprintf(name, "PCI%04x", hose->global_number); 3232 phb->dbgfs = debugfs_create_dir(name, powerpc_debugfs_root); 3233 if (!phb->dbgfs) { 3234 pr_warn("%s: Error on creating debugfs on PHB#%x\n", 3235 __func__, hose->global_number); 3236 continue; 3237 } 3238 3239 debugfs_create_file("dump_diag_regs", 0200, phb->dbgfs, hose, 3240 &pnv_pci_diag_data_fops); 3241 } 3242 #endif /* CONFIG_DEBUG_FS */ 3243 } 3244 3245 static void pnv_pci_enable_bridge(struct pci_bus *bus) 3246 { 3247 struct pci_dev *dev = bus->self; 3248 struct pci_bus *child; 3249 3250 /* Empty bus ? bail */ 3251 if (list_empty(&bus->devices)) 3252 return; 3253 3254 /* 3255 * If there's a bridge associated with that bus enable it. This works 3256 * around races in the generic code if the enabling is done during 3257 * parallel probing. This can be removed once those races have been 3258 * fixed. 3259 */ 3260 if (dev) { 3261 int rc = pci_enable_device(dev); 3262 if (rc) 3263 pci_err(dev, "Error enabling bridge (%d)\n", rc); 3264 pci_set_master(dev); 3265 } 3266 3267 /* Perform the same to child busses */ 3268 list_for_each_entry(child, &bus->children, node) 3269 pnv_pci_enable_bridge(child); 3270 } 3271 3272 static void pnv_pci_enable_bridges(void) 3273 { 3274 struct pci_controller *hose; 3275 3276 list_for_each_entry(hose, &hose_list, list_node) 3277 pnv_pci_enable_bridge(hose->bus); 3278 } 3279 3280 static void pnv_pci_ioda_fixup(void) 3281 { 3282 pnv_pci_ioda_setup_PEs(); 3283 pnv_pci_ioda_setup_iommu_api(); 3284 pnv_pci_ioda_create_dbgfs(); 3285 3286 pnv_pci_enable_bridges(); 3287 3288 #ifdef CONFIG_EEH 3289 pnv_eeh_post_init(); 3290 #endif 3291 } 3292 3293 /* 3294 * Returns the alignment for I/O or memory windows for P2P 3295 * bridges. That actually depends on how PEs are segmented. 3296 * For now, we return I/O or M32 segment size for PE sensitive 3297 * P2P bridges. Otherwise, the default values (4KiB for I/O, 3298 * 1MiB for memory) will be returned. 3299 * 3300 * The current PCI bus might be put into one PE, which was 3301 * create against the parent PCI bridge. For that case, we 3302 * needn't enlarge the alignment so that we can save some 3303 * resources. 3304 */ 3305 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus, 3306 unsigned long type) 3307 { 3308 struct pci_dev *bridge; 3309 struct pci_controller *hose = pci_bus_to_host(bus); 3310 struct pnv_phb *phb = hose->private_data; 3311 int num_pci_bridges = 0; 3312 3313 bridge = bus->self; 3314 while (bridge) { 3315 if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) { 3316 num_pci_bridges++; 3317 if (num_pci_bridges >= 2) 3318 return 1; 3319 } 3320 3321 bridge = bridge->bus->self; 3322 } 3323 3324 /* 3325 * We fall back to M32 if M64 isn't supported. We enforce the M64 3326 * alignment for any 64-bit resource, PCIe doesn't care and 3327 * bridges only do 64-bit prefetchable anyway. 3328 */ 3329 if (phb->ioda.m64_segsize && pnv_pci_is_m64_flags(type)) 3330 return phb->ioda.m64_segsize; 3331 if (type & IORESOURCE_MEM) 3332 return phb->ioda.m32_segsize; 3333 3334 return phb->ioda.io_segsize; 3335 } 3336 3337 /* 3338 * We are updating root port or the upstream port of the 3339 * bridge behind the root port with PHB's windows in order 3340 * to accommodate the changes on required resources during 3341 * PCI (slot) hotplug, which is connected to either root 3342 * port or the downstream ports of PCIe switch behind the 3343 * root port. 3344 */ 3345 static void pnv_pci_fixup_bridge_resources(struct pci_bus *bus, 3346 unsigned long type) 3347 { 3348 struct pci_controller *hose = pci_bus_to_host(bus); 3349 struct pnv_phb *phb = hose->private_data; 3350 struct pci_dev *bridge = bus->self; 3351 struct resource *r, *w; 3352 bool msi_region = false; 3353 int i; 3354 3355 /* Check if we need apply fixup to the bridge's windows */ 3356 if (!pci_is_root_bus(bridge->bus) && 3357 !pci_is_root_bus(bridge->bus->self->bus)) 3358 return; 3359 3360 /* Fixup the resources */ 3361 for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++) { 3362 r = &bridge->resource[PCI_BRIDGE_RESOURCES + i]; 3363 if (!r->flags || !r->parent) 3364 continue; 3365 3366 w = NULL; 3367 if (r->flags & type & IORESOURCE_IO) 3368 w = &hose->io_resource; 3369 else if (pnv_pci_is_m64(phb, r) && 3370 (type & IORESOURCE_PREFETCH) && 3371 phb->ioda.m64_segsize) 3372 w = &hose->mem_resources[1]; 3373 else if (r->flags & type & IORESOURCE_MEM) { 3374 w = &hose->mem_resources[0]; 3375 msi_region = true; 3376 } 3377 3378 r->start = w->start; 3379 r->end = w->end; 3380 3381 /* The 64KB 32-bits MSI region shouldn't be included in 3382 * the 32-bits bridge window. Otherwise, we can see strange 3383 * issues. One of them is EEH error observed on Garrison. 3384 * 3385 * Exclude top 1MB region which is the minimal alignment of 3386 * 32-bits bridge window. 3387 */ 3388 if (msi_region) { 3389 r->end += 0x10000; 3390 r->end -= 0x100000; 3391 } 3392 } 3393 } 3394 3395 static void pnv_pci_setup_bridge(struct pci_bus *bus, unsigned long type) 3396 { 3397 struct pci_controller *hose = pci_bus_to_host(bus); 3398 struct pnv_phb *phb = hose->private_data; 3399 struct pci_dev *bridge = bus->self; 3400 struct pnv_ioda_pe *pe; 3401 bool all = (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE); 3402 3403 /* Extend bridge's windows if necessary */ 3404 pnv_pci_fixup_bridge_resources(bus, type); 3405 3406 /* The PE for root bus should be realized before any one else */ 3407 if (!phb->ioda.root_pe_populated) { 3408 pe = pnv_ioda_setup_bus_PE(phb->hose->bus, false); 3409 if (pe) { 3410 phb->ioda.root_pe_idx = pe->pe_number; 3411 phb->ioda.root_pe_populated = true; 3412 } 3413 } 3414 3415 /* Don't assign PE to PCI bus, which doesn't have subordinate devices */ 3416 if (list_empty(&bus->devices)) 3417 return; 3418 3419 /* Reserve PEs according to used M64 resources */ 3420 pnv_ioda_reserve_m64_pe(bus, NULL, all); 3421 3422 /* 3423 * Assign PE. We might run here because of partial hotplug. 3424 * For the case, we just pick up the existing PE and should 3425 * not allocate resources again. 3426 */ 3427 pe = pnv_ioda_setup_bus_PE(bus, all); 3428 if (!pe) 3429 return; 3430 3431 pnv_ioda_setup_pe_seg(pe); 3432 switch (phb->type) { 3433 case PNV_PHB_IODA1: 3434 pnv_pci_ioda1_setup_dma_pe(phb, pe); 3435 break; 3436 case PNV_PHB_IODA2: 3437 pnv_pci_ioda2_setup_dma_pe(phb, pe); 3438 break; 3439 default: 3440 pr_warn("%s: No DMA for PHB#%x (type %d)\n", 3441 __func__, phb->hose->global_number, phb->type); 3442 } 3443 } 3444 3445 static resource_size_t pnv_pci_default_alignment(void) 3446 { 3447 return PAGE_SIZE; 3448 } 3449 3450 #ifdef CONFIG_PCI_IOV 3451 static resource_size_t pnv_pci_iov_resource_alignment(struct pci_dev *pdev, 3452 int resno) 3453 { 3454 struct pci_controller *hose = pci_bus_to_host(pdev->bus); 3455 struct pnv_phb *phb = hose->private_data; 3456 struct pci_dn *pdn = pci_get_pdn(pdev); 3457 resource_size_t align; 3458 3459 /* 3460 * On PowerNV platform, IOV BAR is mapped by M64 BAR to enable the 3461 * SR-IOV. While from hardware perspective, the range mapped by M64 3462 * BAR should be size aligned. 3463 * 3464 * When IOV BAR is mapped with M64 BAR in Single PE mode, the extra 3465 * powernv-specific hardware restriction is gone. But if just use the 3466 * VF BAR size as the alignment, PF BAR / VF BAR may be allocated with 3467 * in one segment of M64 #15, which introduces the PE conflict between 3468 * PF and VF. Based on this, the minimum alignment of an IOV BAR is 3469 * m64_segsize. 3470 * 3471 * This function returns the total IOV BAR size if M64 BAR is in 3472 * Shared PE mode or just VF BAR size if not. 3473 * If the M64 BAR is in Single PE mode, return the VF BAR size or 3474 * M64 segment size if IOV BAR size is less. 3475 */ 3476 align = pci_iov_resource_size(pdev, resno); 3477 if (!pdn->vfs_expanded) 3478 return align; 3479 if (pdn->m64_single_mode) 3480 return max(align, (resource_size_t)phb->ioda.m64_segsize); 3481 3482 return pdn->vfs_expanded * align; 3483 } 3484 #endif /* CONFIG_PCI_IOV */ 3485 3486 /* Prevent enabling devices for which we couldn't properly 3487 * assign a PE 3488 */ 3489 static bool pnv_pci_enable_device_hook(struct pci_dev *dev) 3490 { 3491 struct pci_controller *hose = pci_bus_to_host(dev->bus); 3492 struct pnv_phb *phb = hose->private_data; 3493 struct pci_dn *pdn; 3494 3495 /* The function is probably called while the PEs have 3496 * not be created yet. For example, resource reassignment 3497 * during PCI probe period. We just skip the check if 3498 * PEs isn't ready. 3499 */ 3500 if (!phb->initialized) 3501 return true; 3502 3503 pdn = pci_get_pdn(dev); 3504 if (!pdn || pdn->pe_number == IODA_INVALID_PE) 3505 return false; 3506 3507 return true; 3508 } 3509 3510 static long pnv_pci_ioda1_unset_window(struct iommu_table_group *table_group, 3511 int num) 3512 { 3513 struct pnv_ioda_pe *pe = container_of(table_group, 3514 struct pnv_ioda_pe, table_group); 3515 struct pnv_phb *phb = pe->phb; 3516 unsigned int idx; 3517 long rc; 3518 3519 pe_info(pe, "Removing DMA window #%d\n", num); 3520 for (idx = 0; idx < phb->ioda.dma32_count; idx++) { 3521 if (phb->ioda.dma32_segmap[idx] != pe->pe_number) 3522 continue; 3523 3524 rc = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number, 3525 idx, 0, 0ul, 0ul, 0ul); 3526 if (rc != OPAL_SUCCESS) { 3527 pe_warn(pe, "Failure %ld unmapping DMA32 segment#%d\n", 3528 rc, idx); 3529 return rc; 3530 } 3531 3532 phb->ioda.dma32_segmap[idx] = IODA_INVALID_PE; 3533 } 3534 3535 pnv_pci_unlink_table_and_group(table_group->tables[num], table_group); 3536 return OPAL_SUCCESS; 3537 } 3538 3539 static void pnv_pci_ioda1_release_pe_dma(struct pnv_ioda_pe *pe) 3540 { 3541 unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe); 3542 struct iommu_table *tbl = pe->table_group.tables[0]; 3543 int64_t rc; 3544 3545 if (!weight) 3546 return; 3547 3548 rc = pnv_pci_ioda1_unset_window(&pe->table_group, 0); 3549 if (rc != OPAL_SUCCESS) 3550 return; 3551 3552 pnv_pci_p7ioc_tce_invalidate(tbl, tbl->it_offset, tbl->it_size, false); 3553 if (pe->table_group.group) { 3554 iommu_group_put(pe->table_group.group); 3555 WARN_ON(pe->table_group.group); 3556 } 3557 3558 free_pages(tbl->it_base, get_order(tbl->it_size << 3)); 3559 iommu_tce_table_put(tbl); 3560 } 3561 3562 static void pnv_pci_ioda2_release_pe_dma(struct pnv_ioda_pe *pe) 3563 { 3564 struct iommu_table *tbl = pe->table_group.tables[0]; 3565 unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe); 3566 #ifdef CONFIG_IOMMU_API 3567 int64_t rc; 3568 #endif 3569 3570 if (!weight) 3571 return; 3572 3573 #ifdef CONFIG_IOMMU_API 3574 rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0); 3575 if (rc) 3576 pe_warn(pe, "OPAL error %ld release DMA window\n", rc); 3577 #endif 3578 3579 pnv_pci_ioda2_set_bypass(pe, false); 3580 if (pe->table_group.group) { 3581 iommu_group_put(pe->table_group.group); 3582 WARN_ON(pe->table_group.group); 3583 } 3584 3585 iommu_tce_table_put(tbl); 3586 } 3587 3588 static void pnv_ioda_free_pe_seg(struct pnv_ioda_pe *pe, 3589 unsigned short win, 3590 unsigned int *map) 3591 { 3592 struct pnv_phb *phb = pe->phb; 3593 int idx; 3594 int64_t rc; 3595 3596 for (idx = 0; idx < phb->ioda.total_pe_num; idx++) { 3597 if (map[idx] != pe->pe_number) 3598 continue; 3599 3600 if (win == OPAL_M64_WINDOW_TYPE) 3601 rc = opal_pci_map_pe_mmio_window(phb->opal_id, 3602 phb->ioda.reserved_pe_idx, win, 3603 idx / PNV_IODA1_M64_SEGS, 3604 idx % PNV_IODA1_M64_SEGS); 3605 else 3606 rc = opal_pci_map_pe_mmio_window(phb->opal_id, 3607 phb->ioda.reserved_pe_idx, win, 0, idx); 3608 3609 if (rc != OPAL_SUCCESS) 3610 pe_warn(pe, "Error %ld unmapping (%d) segment#%d\n", 3611 rc, win, idx); 3612 3613 map[idx] = IODA_INVALID_PE; 3614 } 3615 } 3616 3617 static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe *pe) 3618 { 3619 struct pnv_phb *phb = pe->phb; 3620 3621 if (phb->type == PNV_PHB_IODA1) { 3622 pnv_ioda_free_pe_seg(pe, OPAL_IO_WINDOW_TYPE, 3623 phb->ioda.io_segmap); 3624 pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE, 3625 phb->ioda.m32_segmap); 3626 pnv_ioda_free_pe_seg(pe, OPAL_M64_WINDOW_TYPE, 3627 phb->ioda.m64_segmap); 3628 } else if (phb->type == PNV_PHB_IODA2) { 3629 pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE, 3630 phb->ioda.m32_segmap); 3631 } 3632 } 3633 3634 static void pnv_ioda_release_pe(struct pnv_ioda_pe *pe) 3635 { 3636 struct pnv_phb *phb = pe->phb; 3637 struct pnv_ioda_pe *slave, *tmp; 3638 3639 list_del(&pe->list); 3640 switch (phb->type) { 3641 case PNV_PHB_IODA1: 3642 pnv_pci_ioda1_release_pe_dma(pe); 3643 break; 3644 case PNV_PHB_IODA2: 3645 pnv_pci_ioda2_release_pe_dma(pe); 3646 break; 3647 default: 3648 WARN_ON(1); 3649 } 3650 3651 pnv_ioda_release_pe_seg(pe); 3652 pnv_ioda_deconfigure_pe(pe->phb, pe); 3653 3654 /* Release slave PEs in the compound PE */ 3655 if (pe->flags & PNV_IODA_PE_MASTER) { 3656 list_for_each_entry_safe(slave, tmp, &pe->slaves, list) { 3657 list_del(&slave->list); 3658 pnv_ioda_free_pe(slave); 3659 } 3660 } 3661 3662 /* 3663 * The PE for root bus can be removed because of hotplug in EEH 3664 * recovery for fenced PHB error. We need to mark the PE dead so 3665 * that it can be populated again in PCI hot add path. The PE 3666 * shouldn't be destroyed as it's the global reserved resource. 3667 */ 3668 if (phb->ioda.root_pe_populated && 3669 phb->ioda.root_pe_idx == pe->pe_number) 3670 phb->ioda.root_pe_populated = false; 3671 else 3672 pnv_ioda_free_pe(pe); 3673 } 3674 3675 static void pnv_pci_release_device(struct pci_dev *pdev) 3676 { 3677 struct pci_controller *hose = pci_bus_to_host(pdev->bus); 3678 struct pnv_phb *phb = hose->private_data; 3679 struct pci_dn *pdn = pci_get_pdn(pdev); 3680 struct pnv_ioda_pe *pe; 3681 3682 if (pdev->is_virtfn) 3683 return; 3684 3685 if (!pdn || pdn->pe_number == IODA_INVALID_PE) 3686 return; 3687 3688 /* 3689 * PCI hotplug can happen as part of EEH error recovery. The @pdn 3690 * isn't removed and added afterwards in this scenario. We should 3691 * set the PE number in @pdn to an invalid one. Otherwise, the PE's 3692 * device count is decreased on removing devices while failing to 3693 * be increased on adding devices. It leads to unbalanced PE's device 3694 * count and eventually make normal PCI hotplug path broken. 3695 */ 3696 pe = &phb->ioda.pe_array[pdn->pe_number]; 3697 pdn->pe_number = IODA_INVALID_PE; 3698 3699 WARN_ON(--pe->device_count < 0); 3700 if (pe->device_count == 0) 3701 pnv_ioda_release_pe(pe); 3702 } 3703 3704 static void pnv_npu_disable_device(struct pci_dev *pdev) 3705 { 3706 struct eeh_dev *edev = pci_dev_to_eeh_dev(pdev); 3707 struct eeh_pe *eehpe = edev ? edev->pe : NULL; 3708 3709 if (eehpe && eeh_ops && eeh_ops->reset) 3710 eeh_ops->reset(eehpe, EEH_RESET_HOT); 3711 } 3712 3713 static void pnv_pci_ioda_shutdown(struct pci_controller *hose) 3714 { 3715 struct pnv_phb *phb = hose->private_data; 3716 3717 opal_pci_reset(phb->opal_id, OPAL_RESET_PCI_IODA_TABLE, 3718 OPAL_ASSERT_RESET); 3719 } 3720 3721 static const struct pci_controller_ops pnv_pci_ioda_controller_ops = { 3722 .dma_dev_setup = pnv_pci_dma_dev_setup, 3723 .dma_bus_setup = pnv_pci_dma_bus_setup, 3724 .setup_msi_irqs = pnv_setup_msi_irqs, 3725 .teardown_msi_irqs = pnv_teardown_msi_irqs, 3726 .enable_device_hook = pnv_pci_enable_device_hook, 3727 .release_device = pnv_pci_release_device, 3728 .window_alignment = pnv_pci_window_alignment, 3729 .setup_bridge = pnv_pci_setup_bridge, 3730 .reset_secondary_bus = pnv_pci_reset_secondary_bus, 3731 .dma_set_mask = pnv_pci_ioda_dma_set_mask, 3732 .dma_get_required_mask = pnv_pci_ioda_dma_get_required_mask, 3733 .shutdown = pnv_pci_ioda_shutdown, 3734 }; 3735 3736 static int pnv_npu_dma_set_mask(struct pci_dev *npdev, u64 dma_mask) 3737 { 3738 dev_err_once(&npdev->dev, 3739 "%s operation unsupported for NVLink devices\n", 3740 __func__); 3741 return -EPERM; 3742 } 3743 3744 static const struct pci_controller_ops pnv_npu_ioda_controller_ops = { 3745 .dma_dev_setup = pnv_pci_dma_dev_setup, 3746 .setup_msi_irqs = pnv_setup_msi_irqs, 3747 .teardown_msi_irqs = pnv_teardown_msi_irqs, 3748 .enable_device_hook = pnv_pci_enable_device_hook, 3749 .window_alignment = pnv_pci_window_alignment, 3750 .reset_secondary_bus = pnv_pci_reset_secondary_bus, 3751 .dma_set_mask = pnv_npu_dma_set_mask, 3752 .shutdown = pnv_pci_ioda_shutdown, 3753 .disable_device = pnv_npu_disable_device, 3754 }; 3755 3756 static const struct pci_controller_ops pnv_npu_ocapi_ioda_controller_ops = { 3757 .enable_device_hook = pnv_pci_enable_device_hook, 3758 .window_alignment = pnv_pci_window_alignment, 3759 .reset_secondary_bus = pnv_pci_reset_secondary_bus, 3760 .shutdown = pnv_pci_ioda_shutdown, 3761 }; 3762 3763 static void __init pnv_pci_init_ioda_phb(struct device_node *np, 3764 u64 hub_id, int ioda_type) 3765 { 3766 struct pci_controller *hose; 3767 struct pnv_phb *phb; 3768 unsigned long size, m64map_off, m32map_off, pemap_off; 3769 unsigned long iomap_off = 0, dma32map_off = 0; 3770 struct resource r; 3771 const __be64 *prop64; 3772 const __be32 *prop32; 3773 int len; 3774 unsigned int segno; 3775 u64 phb_id; 3776 void *aux; 3777 long rc; 3778 3779 if (!of_device_is_available(np)) 3780 return; 3781 3782 pr_info("Initializing %s PHB (%pOF)\n", pnv_phb_names[ioda_type], np); 3783 3784 prop64 = of_get_property(np, "ibm,opal-phbid", NULL); 3785 if (!prop64) { 3786 pr_err(" Missing \"ibm,opal-phbid\" property !\n"); 3787 return; 3788 } 3789 phb_id = be64_to_cpup(prop64); 3790 pr_debug(" PHB-ID : 0x%016llx\n", phb_id); 3791 3792 phb = memblock_alloc(sizeof(*phb), SMP_CACHE_BYTES); 3793 3794 /* Allocate PCI controller */ 3795 phb->hose = hose = pcibios_alloc_controller(np); 3796 if (!phb->hose) { 3797 pr_err(" Can't allocate PCI controller for %pOF\n", 3798 np); 3799 memblock_free(__pa(phb), sizeof(struct pnv_phb)); 3800 return; 3801 } 3802 3803 spin_lock_init(&phb->lock); 3804 prop32 = of_get_property(np, "bus-range", &len); 3805 if (prop32 && len == 8) { 3806 hose->first_busno = be32_to_cpu(prop32[0]); 3807 hose->last_busno = be32_to_cpu(prop32[1]); 3808 } else { 3809 pr_warn(" Broken <bus-range> on %pOF\n", np); 3810 hose->first_busno = 0; 3811 hose->last_busno = 0xff; 3812 } 3813 hose->private_data = phb; 3814 phb->hub_id = hub_id; 3815 phb->opal_id = phb_id; 3816 phb->type = ioda_type; 3817 mutex_init(&phb->ioda.pe_alloc_mutex); 3818 3819 /* Detect specific models for error handling */ 3820 if (of_device_is_compatible(np, "ibm,p7ioc-pciex")) 3821 phb->model = PNV_PHB_MODEL_P7IOC; 3822 else if (of_device_is_compatible(np, "ibm,power8-pciex")) 3823 phb->model = PNV_PHB_MODEL_PHB3; 3824 else if (of_device_is_compatible(np, "ibm,power8-npu-pciex")) 3825 phb->model = PNV_PHB_MODEL_NPU; 3826 else if (of_device_is_compatible(np, "ibm,power9-npu-pciex")) 3827 phb->model = PNV_PHB_MODEL_NPU2; 3828 else 3829 phb->model = PNV_PHB_MODEL_UNKNOWN; 3830 3831 /* Initialize diagnostic data buffer */ 3832 prop32 = of_get_property(np, "ibm,phb-diag-data-size", NULL); 3833 if (prop32) 3834 phb->diag_data_size = be32_to_cpup(prop32); 3835 else 3836 phb->diag_data_size = PNV_PCI_DIAG_BUF_SIZE; 3837 3838 phb->diag_data = memblock_alloc(phb->diag_data_size, SMP_CACHE_BYTES); 3839 3840 /* Parse 32-bit and IO ranges (if any) */ 3841 pci_process_bridge_OF_ranges(hose, np, !hose->global_number); 3842 3843 /* Get registers */ 3844 if (!of_address_to_resource(np, 0, &r)) { 3845 phb->regs_phys = r.start; 3846 phb->regs = ioremap(r.start, resource_size(&r)); 3847 if (phb->regs == NULL) 3848 pr_err(" Failed to map registers !\n"); 3849 } 3850 3851 /* Initialize more IODA stuff */ 3852 phb->ioda.total_pe_num = 1; 3853 prop32 = of_get_property(np, "ibm,opal-num-pes", NULL); 3854 if (prop32) 3855 phb->ioda.total_pe_num = be32_to_cpup(prop32); 3856 prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL); 3857 if (prop32) 3858 phb->ioda.reserved_pe_idx = be32_to_cpup(prop32); 3859 3860 /* Invalidate RID to PE# mapping */ 3861 for (segno = 0; segno < ARRAY_SIZE(phb->ioda.pe_rmap); segno++) 3862 phb->ioda.pe_rmap[segno] = IODA_INVALID_PE; 3863 3864 /* Parse 64-bit MMIO range */ 3865 pnv_ioda_parse_m64_window(phb); 3866 3867 phb->ioda.m32_size = resource_size(&hose->mem_resources[0]); 3868 /* FW Has already off top 64k of M32 space (MSI space) */ 3869 phb->ioda.m32_size += 0x10000; 3870 3871 phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe_num; 3872 phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0]; 3873 phb->ioda.io_size = hose->pci_io_size; 3874 phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe_num; 3875 phb->ioda.io_pci_base = 0; /* XXX calculate this ? */ 3876 3877 /* Calculate how many 32-bit TCE segments we have */ 3878 phb->ioda.dma32_count = phb->ioda.m32_pci_base / 3879 PNV_IODA1_DMA32_SEGSIZE; 3880 3881 /* Allocate aux data & arrays. We don't have IO ports on PHB3 */ 3882 size = _ALIGN_UP(max_t(unsigned, phb->ioda.total_pe_num, 8) / 8, 3883 sizeof(unsigned long)); 3884 m64map_off = size; 3885 size += phb->ioda.total_pe_num * sizeof(phb->ioda.m64_segmap[0]); 3886 m32map_off = size; 3887 size += phb->ioda.total_pe_num * sizeof(phb->ioda.m32_segmap[0]); 3888 if (phb->type == PNV_PHB_IODA1) { 3889 iomap_off = size; 3890 size += phb->ioda.total_pe_num * sizeof(phb->ioda.io_segmap[0]); 3891 dma32map_off = size; 3892 size += phb->ioda.dma32_count * 3893 sizeof(phb->ioda.dma32_segmap[0]); 3894 } 3895 pemap_off = size; 3896 size += phb->ioda.total_pe_num * sizeof(struct pnv_ioda_pe); 3897 aux = memblock_alloc(size, SMP_CACHE_BYTES); 3898 phb->ioda.pe_alloc = aux; 3899 phb->ioda.m64_segmap = aux + m64map_off; 3900 phb->ioda.m32_segmap = aux + m32map_off; 3901 for (segno = 0; segno < phb->ioda.total_pe_num; segno++) { 3902 phb->ioda.m64_segmap[segno] = IODA_INVALID_PE; 3903 phb->ioda.m32_segmap[segno] = IODA_INVALID_PE; 3904 } 3905 if (phb->type == PNV_PHB_IODA1) { 3906 phb->ioda.io_segmap = aux + iomap_off; 3907 for (segno = 0; segno < phb->ioda.total_pe_num; segno++) 3908 phb->ioda.io_segmap[segno] = IODA_INVALID_PE; 3909 3910 phb->ioda.dma32_segmap = aux + dma32map_off; 3911 for (segno = 0; segno < phb->ioda.dma32_count; segno++) 3912 phb->ioda.dma32_segmap[segno] = IODA_INVALID_PE; 3913 } 3914 phb->ioda.pe_array = aux + pemap_off; 3915 3916 /* 3917 * Choose PE number for root bus, which shouldn't have 3918 * M64 resources consumed by its child devices. To pick 3919 * the PE number adjacent to the reserved one if possible. 3920 */ 3921 pnv_ioda_reserve_pe(phb, phb->ioda.reserved_pe_idx); 3922 if (phb->ioda.reserved_pe_idx == 0) { 3923 phb->ioda.root_pe_idx = 1; 3924 pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx); 3925 } else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1)) { 3926 phb->ioda.root_pe_idx = phb->ioda.reserved_pe_idx - 1; 3927 pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx); 3928 } else { 3929 phb->ioda.root_pe_idx = IODA_INVALID_PE; 3930 } 3931 3932 INIT_LIST_HEAD(&phb->ioda.pe_list); 3933 mutex_init(&phb->ioda.pe_list_mutex); 3934 3935 /* Calculate how many 32-bit TCE segments we have */ 3936 phb->ioda.dma32_count = phb->ioda.m32_pci_base / 3937 PNV_IODA1_DMA32_SEGSIZE; 3938 3939 #if 0 /* We should really do that ... */ 3940 rc = opal_pci_set_phb_mem_window(opal->phb_id, 3941 window_type, 3942 window_num, 3943 starting_real_address, 3944 starting_pci_address, 3945 segment_size); 3946 #endif 3947 3948 pr_info(" %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n", 3949 phb->ioda.total_pe_num, phb->ioda.reserved_pe_idx, 3950 phb->ioda.m32_size, phb->ioda.m32_segsize); 3951 if (phb->ioda.m64_size) 3952 pr_info(" M64: 0x%lx [segment=0x%lx]\n", 3953 phb->ioda.m64_size, phb->ioda.m64_segsize); 3954 if (phb->ioda.io_size) 3955 pr_info(" IO: 0x%x [segment=0x%x]\n", 3956 phb->ioda.io_size, phb->ioda.io_segsize); 3957 3958 3959 phb->hose->ops = &pnv_pci_ops; 3960 phb->get_pe_state = pnv_ioda_get_pe_state; 3961 phb->freeze_pe = pnv_ioda_freeze_pe; 3962 phb->unfreeze_pe = pnv_ioda_unfreeze_pe; 3963 3964 /* Setup MSI support */ 3965 pnv_pci_init_ioda_msis(phb); 3966 3967 /* 3968 * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here 3969 * to let the PCI core do resource assignment. It's supposed 3970 * that the PCI core will do correct I/O and MMIO alignment 3971 * for the P2P bridge bars so that each PCI bus (excluding 3972 * the child P2P bridges) can form individual PE. 3973 */ 3974 ppc_md.pcibios_fixup = pnv_pci_ioda_fixup; 3975 3976 switch (phb->type) { 3977 case PNV_PHB_NPU_NVLINK: 3978 hose->controller_ops = pnv_npu_ioda_controller_ops; 3979 break; 3980 case PNV_PHB_NPU_OCAPI: 3981 hose->controller_ops = pnv_npu_ocapi_ioda_controller_ops; 3982 break; 3983 default: 3984 phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup; 3985 hose->controller_ops = pnv_pci_ioda_controller_ops; 3986 } 3987 3988 ppc_md.pcibios_default_alignment = pnv_pci_default_alignment; 3989 3990 #ifdef CONFIG_PCI_IOV 3991 ppc_md.pcibios_fixup_sriov = pnv_pci_ioda_fixup_iov_resources; 3992 ppc_md.pcibios_iov_resource_alignment = pnv_pci_iov_resource_alignment; 3993 ppc_md.pcibios_sriov_enable = pnv_pcibios_sriov_enable; 3994 ppc_md.pcibios_sriov_disable = pnv_pcibios_sriov_disable; 3995 #endif 3996 3997 pci_add_flags(PCI_REASSIGN_ALL_RSRC); 3998 3999 /* Reset IODA tables to a clean state */ 4000 rc = opal_pci_reset(phb_id, OPAL_RESET_PCI_IODA_TABLE, OPAL_ASSERT_RESET); 4001 if (rc) 4002 pr_warn(" OPAL Error %ld performing IODA table reset !\n", rc); 4003 4004 /* 4005 * If we're running in kdump kernel, the previous kernel never 4006 * shutdown PCI devices correctly. We already got IODA table 4007 * cleaned out. So we have to issue PHB reset to stop all PCI 4008 * transactions from previous kernel. The ppc_pci_reset_phbs 4009 * kernel parameter will force this reset too. 4010 */ 4011 if (is_kdump_kernel() || pci_reset_phbs) { 4012 pr_info(" Issue PHB reset ...\n"); 4013 pnv_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL); 4014 pnv_eeh_phb_reset(hose, EEH_RESET_DEACTIVATE); 4015 } 4016 4017 /* Remove M64 resource if we can't configure it successfully */ 4018 if (!phb->init_m64 || phb->init_m64(phb)) 4019 hose->mem_resources[1].flags = 0; 4020 } 4021 4022 void __init pnv_pci_init_ioda2_phb(struct device_node *np) 4023 { 4024 pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2); 4025 } 4026 4027 void __init pnv_pci_init_npu_phb(struct device_node *np) 4028 { 4029 pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU_NVLINK); 4030 } 4031 4032 void __init pnv_pci_init_npu2_opencapi_phb(struct device_node *np) 4033 { 4034 pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU_OCAPI); 4035 } 4036 4037 static void pnv_npu2_opencapi_cfg_size_fixup(struct pci_dev *dev) 4038 { 4039 struct pci_controller *hose = pci_bus_to_host(dev->bus); 4040 struct pnv_phb *phb = hose->private_data; 4041 4042 if (!machine_is(powernv)) 4043 return; 4044 4045 if (phb->type == PNV_PHB_NPU_OCAPI) 4046 dev->cfg_size = PCI_CFG_SPACE_EXP_SIZE; 4047 } 4048 DECLARE_PCI_FIXUP_EARLY(PCI_ANY_ID, PCI_ANY_ID, pnv_npu2_opencapi_cfg_size_fixup); 4049 4050 void __init pnv_pci_init_ioda_hub(struct device_node *np) 4051 { 4052 struct device_node *phbn; 4053 const __be64 *prop64; 4054 u64 hub_id; 4055 4056 pr_info("Probing IODA IO-Hub %pOF\n", np); 4057 4058 prop64 = of_get_property(np, "ibm,opal-hubid", NULL); 4059 if (!prop64) { 4060 pr_err(" Missing \"ibm,opal-hubid\" property !\n"); 4061 return; 4062 } 4063 hub_id = be64_to_cpup(prop64); 4064 pr_devel(" HUB-ID : 0x%016llx\n", hub_id); 4065 4066 /* Count child PHBs */ 4067 for_each_child_of_node(np, phbn) { 4068 /* Look for IODA1 PHBs */ 4069 if (of_device_is_compatible(phbn, "ibm,ioda-phb")) 4070 pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1); 4071 } 4072 } 4073