arch/x86/kernel/pci-calgary_64.c

   1 /*
   2  * Derived from arch/powerpc/kernel/iommu.c
   3  *
   4  * Copyright IBM Corporation, 2006-2007
   5  * Copyright (C) 2006  Jon Mason <jdmason@kudzu.us>
   6  *
   7  * Author: Jon Mason <jdmason@kudzu.us>
   8  * Author: Muli Ben-Yehuda <muli@il.ibm.com>
   9
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License as published by
  12  * the Free Software Foundation; either version 2 of the License, or
  13  * (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  * GNU General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public License
  21  * along with this program; if not, write to the Free Software
  22  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
  23  */
  24
  25 #include <linux/kernel.h>
  26 #include <linux/init.h>
  27 #include <linux/types.h>
  28 #include <linux/slab.h>
  29 #include <linux/mm.h>
  30 #include <linux/spinlock.h>
  31 #include <linux/string.h>
  32 #include <linux/dma-mapping.h>
  33 #include <linux/bitops.h>
  34 #include <linux/pci_ids.h>
  35 #include <linux/pci.h>
  36 #include <linux/delay.h>
  37 #include <linux/scatterlist.h>
  38 #include <linux/iommu-helper.h>
  39 #include <asm/gart.h>
  40 #include <asm/calgary.h>
  41 #include <asm/tce.h>
  42 #include <asm/pci-direct.h>
  43 #include <asm/system.h>
  44 #include <asm/dma.h>
  45 #include <asm/rio.h>
  46
  47 #ifdef CONFIG_CALGARY_IOMMU_ENABLED_BY_DEFAULT
  48 int use_calgary __read_mostly = 1;
  49 #else
  50 int use_calgary __read_mostly = 0;
  51 #endif /* CONFIG_CALGARY_DEFAULT_ENABLED */
  52
  53 #define PCI_DEVICE_ID_IBM_CALGARY 0x02a1
  54 #define PCI_DEVICE_ID_IBM_CALIOC2 0x0308
  55
  56 /* register offsets inside the host bridge space */
  57 #define CALGARY_CONFIG_REG      0x0108
  58 #define PHB_CSR_OFFSET          0x0110 /* Channel Status */
  59 #define PHB_PLSSR_OFFSET        0x0120
  60 #define PHB_CONFIG_RW_OFFSET    0x0160
  61 #define PHB_IOBASE_BAR_LOW      0x0170
  62 #define PHB_IOBASE_BAR_HIGH     0x0180
  63 #define PHB_MEM_1_LOW           0x0190
  64 #define PHB_MEM_1_HIGH          0x01A0
  65 #define PHB_IO_ADDR_SIZE        0x01B0
  66 #define PHB_MEM_1_SIZE          0x01C0
  67 #define PHB_MEM_ST_OFFSET       0x01D0
  68 #define PHB_AER_OFFSET          0x0200
  69 #define PHB_CONFIG_0_HIGH       0x0220
  70 #define PHB_CONFIG_0_LOW        0x0230
  71 #define PHB_CONFIG_0_END        0x0240
  72 #define PHB_MEM_2_LOW           0x02B0
  73 #define PHB_MEM_2_HIGH          0x02C0
  74 #define PHB_MEM_2_SIZE_HIGH     0x02D0
  75 #define PHB_MEM_2_SIZE_LOW      0x02E0
  76 #define PHB_DOSHOLE_OFFSET      0x08E0
  77
  78 /* CalIOC2 specific */
  79 #define PHB_SAVIOR_L2           0x0DB0
  80 #define PHB_PAGE_MIG_CTRL       0x0DA8
  81 #define PHB_PAGE_MIG_DEBUG      0x0DA0
  82 #define PHB_ROOT_COMPLEX_STATUS 0x0CB0
  83
  84 /* PHB_CONFIG_RW */
  85 #define PHB_TCE_ENABLE          0x20000000
  86 #define PHB_SLOT_DISABLE        0x1C000000
  87 #define PHB_DAC_DISABLE         0x01000000
  88 #define PHB_MEM2_ENABLE         0x00400000
  89 #define PHB_MCSR_ENABLE         0x00100000
  90 /* TAR (Table Address Register) */
  91 #define TAR_SW_BITS             0x0000ffffffff800fUL
  92 #define TAR_VALID               0x0000000000000008UL
  93 /* CSR (Channel/DMA Status Register) */
  94 #define CSR_AGENT_MASK          0xffe0ffff
  95 /* CCR (Calgary Configuration Register) */
  96 #define CCR_2SEC_TIMEOUT        0x000000000000000EUL
  97 /* PMCR/PMDR (Page Migration Control/Debug Registers */
  98 #define PMR_SOFTSTOP            0x80000000
  99 #define PMR_SOFTSTOPFAULT       0x40000000
 100 #define PMR_HARDSTOP            0x20000000
 101
 102 #define MAX_NUM_OF_PHBS         8 /* how many PHBs in total? */
 103 #define MAX_NUM_CHASSIS         8 /* max number of chassis */
 104 /* MAX_PHB_BUS_NUM is the maximal possible dev->bus->number */
 105 #define MAX_PHB_BUS_NUM         (MAX_NUM_OF_PHBS * MAX_NUM_CHASSIS * 2)
 106 #define PHBS_PER_CALGARY        4
 107
 108 /* register offsets in Calgary's internal register space */
 109 static const unsigned long tar_offsets[] = {
 110         0x0580 /* TAR0 */,
 111         0x0588 /* TAR1 */,
 112         0x0590 /* TAR2 */,
 113         0x0598 /* TAR3 */
 114 };
 115
 116 static const unsigned long split_queue_offsets[] = {
 117         0x4870 /* SPLIT QUEUE 0 */,
 118         0x5870 /* SPLIT QUEUE 1 */,
 119         0x6870 /* SPLIT QUEUE 2 */,
 120         0x7870 /* SPLIT QUEUE 3 */
 121 };
 122
 123 static const unsigned long phb_offsets[] = {
 124         0x8000 /* PHB0 */,
 125         0x9000 /* PHB1 */,
 126         0xA000 /* PHB2 */,
 127         0xB000 /* PHB3 */
 128 };
 129
 130 /* PHB debug registers */
 131
 132 static const unsigned long phb_debug_offsets[] = {
 133         0x4000  /* PHB 0 DEBUG */,
 134         0x5000  /* PHB 1 DEBUG */,
 135         0x6000  /* PHB 2 DEBUG */,
 136         0x7000  /* PHB 3 DEBUG */
 137 };
 138
 139 /*
 140  * STUFF register for each debug PHB,
 141  * byte 1 = start bus number, byte 2 = end bus number
 142  */
 143
 144 #define PHB_DEBUG_STUFF_OFFSET  0x0020
 145
 146 #define EMERGENCY_PAGES 32 /* = 128KB */
 147
 148 unsigned int specified_table_size = TCE_TABLE_SIZE_UNSPECIFIED;
 149 static int translate_empty_slots __read_mostly = 0;
 150 static int calgary_detected __read_mostly = 0;
 151
 152 static struct rio_table_hdr     *rio_table_hdr __initdata;
 153 static struct scal_detail       *scal_devs[MAX_NUMNODES] __initdata;
 154 static struct rio_detail        *rio_devs[MAX_NUMNODES * 4] __initdata;
 155
 156 struct calgary_bus_info {
 157         void *tce_space;
 158         unsigned char translation_disabled;
 159         signed char phbid;
 160         void __iomem *bbar;
 161 };
 162
 163 static void calgary_handle_quirks(struct iommu_table *tbl, struct pci_dev *dev);
 164 static void calgary_tce_cache_blast(struct iommu_table *tbl);
 165 static void calgary_dump_error_regs(struct iommu_table *tbl);
 166 static void calioc2_handle_quirks(struct iommu_table *tbl, struct pci_dev *dev);
 167 static void calioc2_tce_cache_blast(struct iommu_table *tbl);
 168 static void calioc2_dump_error_regs(struct iommu_table *tbl);
 169
 170 static struct cal_chipset_ops calgary_chip_ops = {
 171         .handle_quirks = calgary_handle_quirks,
 172         .tce_cache_blast = calgary_tce_cache_blast,
 173         .dump_error_regs = calgary_dump_error_regs
 174 };
 175
 176 static struct cal_chipset_ops calioc2_chip_ops = {
 177         .handle_quirks = calioc2_handle_quirks,
 178         .tce_cache_blast = calioc2_tce_cache_blast,
 179         .dump_error_regs = calioc2_dump_error_regs
 180 };
 181
 182 static struct calgary_bus_info bus_info[MAX_PHB_BUS_NUM] = { { NULL, 0, 0 }, };
 183
 184 /* enable this to stress test the chip's TCE cache */
 185 #ifdef CONFIG_IOMMU_DEBUG
 186 static int debugging = 1;
 187
 188 static inline unsigned long verify_bit_range(unsigned long* bitmap,
 189         int expected, unsigned long start, unsigned long end)
 190 {
 191         unsigned long idx = start;
 192
 193         BUG_ON(start >= end);
 194
 195         while (idx < end) {
 196                 if (!!test_bit(idx, bitmap) != expected)
 197                         return idx;
 198                 ++idx;
 199         }
 200
 201         /* all bits have the expected value */
 202         return ~0UL;
 203 }
 204 #else /* debugging is disabled */
 205 static int debugging;
 206
 207 static inline unsigned long verify_bit_range(unsigned long* bitmap,
 208         int expected, unsigned long start, unsigned long end)
 209 {
 210         return ~0UL;
 211 }
 212
 213 #endif /* CONFIG_IOMMU_DEBUG */
 214
 215 static inline unsigned int num_dma_pages(unsigned long dma, unsigned int dmalen)
 216 {
 217         unsigned int npages;
 218
 219         npages = PAGE_ALIGN(dma + dmalen) - (dma & PAGE_MASK);
 220         npages >>= PAGE_SHIFT;
 221
 222         return npages;
 223 }
 224
 225 static inline int translation_enabled(struct iommu_table *tbl)
 226 {
 227         /* only PHBs with translation enabled have an IOMMU table */
 228         return (tbl != NULL);
 229 }
 230
 231 static void iommu_range_reserve(struct iommu_table *tbl,
 232         unsigned long start_addr, unsigned int npages)
 233 {
 234         unsigned long index;
 235         unsigned long end;
 236         unsigned long badbit;
 237         unsigned long flags;
 238
 239         index = start_addr >> PAGE_SHIFT;
 240
 241         /* bail out if we're asked to reserve a region we don't cover */
 242         if (index >= tbl->it_size)
 243                 return;
 244
 245         end = index + npages;
 246         if (end > tbl->it_size) /* don't go off the table */
 247                 end = tbl->it_size;
 248
 249         spin_lock_irqsave(&tbl->it_lock, flags);
 250
 251         badbit = verify_bit_range(tbl->it_map, 0, index, end);
 252         if (badbit != ~0UL) {
 253                 if (printk_ratelimit())
 254                         printk(KERN_ERR "Calgary: entry already allocated at "
 255                                "0x%lx tbl %p dma 0x%lx npages %u\n",
 256                                badbit, tbl, start_addr, npages);
 257         }
 258
 259         set_bit_string(tbl->it_map, index, npages);
 260
 261         spin_unlock_irqrestore(&tbl->it_lock, flags);
 262 }
 263
 264 static unsigned long iommu_range_alloc(struct device *dev,
 265                                        struct iommu_table *tbl,
 266                                        unsigned int npages)
 267 {
 268         unsigned long flags;
 269         unsigned long offset;
 270         unsigned long boundary_size;
 271
 272         boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
 273                               PAGE_SIZE) >> PAGE_SHIFT;
 274
 275         BUG_ON(npages == 0);
 276
 277         spin_lock_irqsave(&tbl->it_lock, flags);
 278
 279         offset = iommu_area_alloc(tbl->it_map, tbl->it_size, tbl->it_hint,
 280                                   npages, 0, boundary_size, 0);
 281         if (offset == ~0UL) {
 282                 tbl->chip_ops->tce_cache_blast(tbl);
 283
 284                 offset = iommu_area_alloc(tbl->it_map, tbl->it_size, 0,
 285                                           npages, 0, boundary_size, 0);
 286                 if (offset == ~0UL) {
 287                         printk(KERN_WARNING "Calgary: IOMMU full.\n");
 288                         spin_unlock_irqrestore(&tbl->it_lock, flags);
 289                         if (panic_on_overflow)
 290                                 panic("Calgary: fix the allocator.\n");
 291                         else
 292                                 return bad_dma_address;
 293                 }
 294         }
 295
 296         tbl->it_hint = offset + npages;
 297         BUG_ON(tbl->it_hint > tbl->it_size);
 298
 299         spin_unlock_irqrestore(&tbl->it_lock, flags);
 300
 301         return offset;
 302 }
 303
 304 static dma_addr_t iommu_alloc(struct device *dev, struct iommu_table *tbl,
 305                               void *vaddr, unsigned int npages, int direction)
 306 {
 307         unsigned long entry;
 308         dma_addr_t ret = bad_dma_address;
 309
 310         entry = iommu_range_alloc(dev, tbl, npages);
 311
 312         if (unlikely(entry == bad_dma_address))
 313                 goto error;
 314
 315         /* set the return dma address */
 316         ret = (entry << PAGE_SHIFT) | ((unsigned long)vaddr & ~PAGE_MASK);
 317
 318         /* put the TCEs in the HW table */
 319         tce_build(tbl, entry, npages, (unsigned long)vaddr & PAGE_MASK,
 320                   direction);
 321
 322         return ret;
 323
 324 error:
 325         printk(KERN_WARNING "Calgary: failed to allocate %u pages in "
 326                "iommu %p\n", npages, tbl);
 327         return bad_dma_address;
 328 }
 329
 330 static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
 331         unsigned int npages)
 332 {
 333         unsigned long entry;
 334         unsigned long badbit;
 335         unsigned long badend;
 336         unsigned long flags;
 337
 338         /* were we called with bad_dma_address? */
 339         badend = bad_dma_address + (EMERGENCY_PAGES * PAGE_SIZE);
 340         if (unlikely((dma_addr >= bad_dma_address) && (dma_addr < badend))) {
 341                 printk(KERN_ERR "Calgary: driver tried unmapping bad DMA "
 342                        "address 0x%Lx\n", dma_addr);
 343                 WARN_ON(1);
 344                 return;
 345         }
 346
 347         entry = dma_addr >> PAGE_SHIFT;
 348
 349         BUG_ON(entry + npages > tbl->it_size);
 350
 351         tce_free(tbl, entry, npages);
 352
 353         spin_lock_irqsave(&tbl->it_lock, flags);
 354
 355         badbit = verify_bit_range(tbl->it_map, 1, entry, entry + npages);
 356         if (badbit != ~0UL) {
 357                 if (printk_ratelimit())
 358                         printk(KERN_ERR "Calgary: bit is off at 0x%lx "
 359                                "tbl %p dma 0x%Lx entry 0x%lx npages %u\n",
 360                                badbit, tbl, dma_addr, entry, npages);
 361         }
 362
 363         iommu_area_free(tbl->it_map, entry, npages);
 364
 365         spin_unlock_irqrestore(&tbl->it_lock, flags);
 366 }
 367
 368 static inline struct iommu_table *find_iommu_table(struct device *dev)
 369 {
 370         struct pci_dev *pdev;
 371         struct pci_bus *pbus;
 372         struct iommu_table *tbl;
 373
 374         pdev = to_pci_dev(dev);
 375
 376         pbus = pdev->bus;
 377
 378         /* is the device behind a bridge? Look for the root bus */
 379         while (pbus->parent)
 380                 pbus = pbus->parent;
 381
 382         tbl = pci_iommu(pbus);
 383
 384         BUG_ON(tbl && (tbl->it_busno != pbus->number));
 385
 386         return tbl;
 387 }
 388
 389 static void calgary_unmap_sg(struct device *dev,
 390         struct scatterlist *sglist, int nelems, int direction)
 391 {
 392         struct iommu_table *tbl = find_iommu_table(dev);
 393         struct scatterlist *s;
 394         int i;
 395
 396         if (!translation_enabled(tbl))
 397                 return;
 398
 399         for_each_sg(sglist, s, nelems, i) {
 400                 unsigned int npages;
 401                 dma_addr_t dma = s->dma_address;
 402                 unsigned int dmalen = s->dma_length;
 403
 404                 if (dmalen == 0)
 405                         break;
 406
 407                 npages = num_dma_pages(dma, dmalen);
 408                 iommu_free(tbl, dma, npages);
 409         }
 410 }
 411
 412 static int calgary_nontranslate_map_sg(struct device* dev,
 413         struct scatterlist *sg, int nelems, int direction)
 414 {
 415         struct scatterlist *s;
 416         int i;
 417
 418         for_each_sg(sg, s, nelems, i) {
 419                 struct page *p = sg_page(s);
 420
 421                 BUG_ON(!p);
 422                 s->dma_address = virt_to_bus(sg_virt(s));
 423                 s->dma_length = s->length;
 424         }
 425         return nelems;
 426 }
 427
 428 static int calgary_map_sg(struct device *dev, struct scatterlist *sg,
 429         int nelems, int direction)
 430 {
 431         struct iommu_table *tbl = find_iommu_table(dev);
 432         struct scatterlist *s;
 433         unsigned long vaddr;
 434         unsigned int npages;
 435         unsigned long entry;
 436         int i;
 437
 438         if (!translation_enabled(tbl))
 439                 return calgary_nontranslate_map_sg(dev, sg, nelems, direction);
 440
 441         for_each_sg(sg, s, nelems, i) {
 442                 BUG_ON(!sg_page(s));
 443
 444                 vaddr = (unsigned long) sg_virt(s);
 445                 npages = num_dma_pages(vaddr, s->length);
 446
 447                 entry = iommu_range_alloc(dev, tbl, npages);
 448                 if (entry == bad_dma_address) {
 449                         /* makes sure unmap knows to stop */
 450                         s->dma_length = 0;
 451                         goto error;
 452                 }
 453
 454                 s->dma_address = (entry << PAGE_SHIFT) | s->offset;
 455
 456                 /* insert into HW table */
 457                 tce_build(tbl, entry, npages, vaddr & PAGE_MASK,
 458                           direction);
 459
 460                 s->dma_length = s->length;
 461         }
 462
 463         return nelems;
 464 error:
 465         calgary_unmap_sg(dev, sg, nelems, direction);
 466         for_each_sg(sg, s, nelems, i) {
 467                 sg->dma_address = bad_dma_address;
 468                 sg->dma_length = 0;
 469         }
 470         return 0;
 471 }
 472
 473 static dma_addr_t calgary_map_single(struct device *dev, void *vaddr,
 474         size_t size, int direction)
 475 {
 476         dma_addr_t dma_handle = bad_dma_address;
 477         unsigned long uaddr;
 478         unsigned int npages;
 479         struct iommu_table *tbl = find_iommu_table(dev);
 480
 481         uaddr = (unsigned long)vaddr;
 482         npages = num_dma_pages(uaddr, size);
 483
 484         if (translation_enabled(tbl))
 485                 dma_handle = iommu_alloc(dev, tbl, vaddr, npages, direction);
 486         else
 487                 dma_handle = virt_to_bus(vaddr);
 488
 489         return dma_handle;
 490 }
 491
 492 static void calgary_unmap_single(struct device *dev, dma_addr_t dma_handle,
 493         size_t size, int direction)
 494 {
 495         struct iommu_table *tbl = find_iommu_table(dev);
 496         unsigned int npages;
 497
 498         if (!translation_enabled(tbl))
 499                 return;
 500
 501         npages = num_dma_pages(dma_handle, size);
 502         iommu_free(tbl, dma_handle, npages);
 503 }
 504
 505 static void* calgary_alloc_coherent(struct device *dev, size_t size,
 506         dma_addr_t *dma_handle, gfp_t flag)
 507 {
 508         void *ret = NULL;
 509         dma_addr_t mapping;
 510         unsigned int npages, order;
 511         struct iommu_table *tbl = find_iommu_table(dev);
 512
 513         size = PAGE_ALIGN(size); /* size rounded up to full pages */
 514         npages = size >> PAGE_SHIFT;
 515         order = get_order(size);
 516
 517         /* alloc enough pages (and possibly more) */
 518         ret = (void *)__get_free_pages(flag, order);
 519         if (!ret)
 520                 goto error;
 521         memset(ret, 0, size);
 522
 523         if (translation_enabled(tbl)) {
 524                 /* set up tces to cover the allocated range */
 525                 mapping = iommu_alloc(dev, tbl, ret, npages, DMA_BIDIRECTIONAL);
 526                 if (mapping == bad_dma_address)
 527                         goto free;
 528
 529                 *dma_handle = mapping;
 530         } else /* non translated slot */
 531                 *dma_handle = virt_to_bus(ret);
 532
 533         return ret;
 534
 535 free:
 536         free_pages((unsigned long)ret, get_order(size));
 537         ret = NULL;
 538 error:
 539         return ret;
 540 }
 541
 542 static const struct dma_mapping_ops calgary_dma_ops = {
 543         .alloc_coherent = calgary_alloc_coherent,
 544         .map_single = calgary_map_single,
 545         .unmap_single = calgary_unmap_single,
 546         .map_sg = calgary_map_sg,
 547         .unmap_sg = calgary_unmap_sg,
 548 };
 549
 550 static inline void __iomem * busno_to_bbar(unsigned char num)
 551 {
 552         return bus_info[num].bbar;
 553 }
 554
 555 static inline int busno_to_phbid(unsigned char num)
 556 {
 557         return bus_info[num].phbid;
 558 }
 559
 560 static inline unsigned long split_queue_offset(unsigned char num)
 561 {
 562         size_t idx = busno_to_phbid(num);
 563
 564         return split_queue_offsets[idx];
 565 }
 566
 567 static inline unsigned long tar_offset(unsigned char num)
 568 {
 569         size_t idx = busno_to_phbid(num);
 570
 571         return tar_offsets[idx];
 572 }
 573
 574 static inline unsigned long phb_offset(unsigned char num)
 575 {
 576         size_t idx = busno_to_phbid(num);
 577
 578         return phb_offsets[idx];
 579 }
 580
 581 static inline void __iomem* calgary_reg(void __iomem *bar, unsigned long offset)
 582 {
 583         unsigned long target = ((unsigned long)bar) | offset;
 584         return (void __iomem*)target;
 585 }
 586
 587 static inline int is_calioc2(unsigned short device)
 588 {
 589         return (device == PCI_DEVICE_ID_IBM_CALIOC2);
 590 }
 591
 592 static inline int is_calgary(unsigned short device)
 593 {
 594         return (device == PCI_DEVICE_ID_IBM_CALGARY);
 595 }
 596
 597 static inline int is_cal_pci_dev(unsigned short device)
 598 {
 599         return (is_calgary(device) || is_calioc2(device));
 600 }
 601
 602 static void calgary_tce_cache_blast(struct iommu_table *tbl)
 603 {
 604         u64 val;
 605         u32 aer;
 606         int i = 0;
 607         void __iomem *bbar = tbl->bbar;
 608         void __iomem *target;
 609
 610         /* disable arbitration on the bus */
 611         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_AER_OFFSET);
 612         aer = readl(target);
 613         writel(0, target);
 614
 615         /* read plssr to ensure it got there */
 616         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_PLSSR_OFFSET);
 617         val = readl(target);
 618
 619         /* poll split queues until all DMA activity is done */
 620         target = calgary_reg(bbar, split_queue_offset(tbl->it_busno));
 621         do {
 622                 val = readq(target);
 623                 i++;
 624         } while ((val & 0xff) != 0xff && i < 100);
 625         if (i == 100)
 626                 printk(KERN_WARNING "Calgary: PCI bus not quiesced, "
 627                        "continuing anyway\n");
 628
 629         /* invalidate TCE cache */
 630         target = calgary_reg(bbar, tar_offset(tbl->it_busno));
 631         writeq(tbl->tar_val, target);
 632
 633         /* enable arbitration */
 634         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_AER_OFFSET);
 635         writel(aer, target);
 636         (void)readl(target); /* flush */
 637 }
 638
 639 static void calioc2_tce_cache_blast(struct iommu_table *tbl)
 640 {
 641         void __iomem *bbar = tbl->bbar;
 642         void __iomem *target;
 643         u64 val64;
 644         u32 val;
 645         int i = 0;
 646         int count = 1;
 647         unsigned char bus = tbl->it_busno;
 648
 649 begin:
 650         printk(KERN_DEBUG "Calgary: CalIOC2 bus 0x%x entering tce cache blast "
 651                "sequence - count %d\n", bus, count);
 652
 653         /* 1. using the Page Migration Control reg set SoftStop */
 654         target = calgary_reg(bbar, phb_offset(bus) | PHB_PAGE_MIG_CTRL);
 655         val = be32_to_cpu(readl(target));
 656         printk(KERN_DEBUG "1a. read 0x%x [LE] from %p\n", val, target);
 657         val |= PMR_SOFTSTOP;
 658         printk(KERN_DEBUG "1b. writing 0x%x [LE] to %p\n", val, target);
 659         writel(cpu_to_be32(val), target);
 660
 661         /* 2. poll split queues until all DMA activity is done */
 662         printk(KERN_DEBUG "2a. starting to poll split queues\n");
 663         target = calgary_reg(bbar, split_queue_offset(bus));
 664         do {
 665                 val64 = readq(target);
 666                 i++;
 667         } while ((val64 & 0xff) != 0xff && i < 100);
 668         if (i == 100)
 669                 printk(KERN_WARNING "CalIOC2: PCI bus not quiesced, "
 670                        "continuing anyway\n");
 671
 672         /* 3. poll Page Migration DEBUG for SoftStopFault */
 673         target = calgary_reg(bbar, phb_offset(bus) | PHB_PAGE_MIG_DEBUG);
 674         val = be32_to_cpu(readl(target));
 675         printk(KERN_DEBUG "3. read 0x%x [LE] from %p\n", val, target);
 676
 677         /* 4. if SoftStopFault - goto (1) */
 678         if (val & PMR_SOFTSTOPFAULT) {
 679                 if (++count < 100)
 680                         goto begin;
 681                 else {
 682                         printk(KERN_WARNING "CalIOC2: too many SoftStopFaults, "
 683                                "aborting TCE cache flush sequence!\n");
 684                         return; /* pray for the best */
 685                 }
 686         }
 687
 688         /* 5. Slam into HardStop by reading PHB_PAGE_MIG_CTRL */
 689         target = calgary_reg(bbar, phb_offset(bus) | PHB_PAGE_MIG_CTRL);
 690         printk(KERN_DEBUG "5a. slamming into HardStop by reading %p\n", target);
 691         val = be32_to_cpu(readl(target));
 692         printk(KERN_DEBUG "5b. read 0x%x [LE] from %p\n", val, target);
 693         target = calgary_reg(bbar, phb_offset(bus) | PHB_PAGE_MIG_DEBUG);
 694         val = be32_to_cpu(readl(target));
 695         printk(KERN_DEBUG "5c. read 0x%x [LE] from %p (debug)\n", val, target);
 696
 697         /* 6. invalidate TCE cache */
 698         printk(KERN_DEBUG "6. invalidating TCE cache\n");
 699         target = calgary_reg(bbar, tar_offset(bus));
 700         writeq(tbl->tar_val, target);
 701
 702         /* 7. Re-read PMCR */
 703         printk(KERN_DEBUG "7a. Re-reading PMCR\n");
 704         target = calgary_reg(bbar, phb_offset(bus) | PHB_PAGE_MIG_CTRL);
 705         val = be32_to_cpu(readl(target));
 706         printk(KERN_DEBUG "7b. read 0x%x [LE] from %p\n", val, target);
 707
 708         /* 8. Remove HardStop */
 709         printk(KERN_DEBUG "8a. removing HardStop from PMCR\n");
 710         target = calgary_reg(bbar, phb_offset(bus) | PHB_PAGE_MIG_CTRL);
 711         val = 0;
 712         printk(KERN_DEBUG "8b. writing 0x%x [LE] to %p\n", val, target);
 713         writel(cpu_to_be32(val), target);
 714         val = be32_to_cpu(readl(target));
 715         printk(KERN_DEBUG "8c. read 0x%x [LE] from %p\n", val, target);
 716 }
 717
 718 static void __init calgary_reserve_mem_region(struct pci_dev *dev, u64 start,
 719         u64 limit)
 720 {
 721         unsigned int numpages;
 722
 723         limit = limit | 0xfffff;
 724         limit++;
 725
 726         numpages = ((limit - start) >> PAGE_SHIFT);
 727         iommu_range_reserve(pci_iommu(dev->bus), start, numpages);
 728 }
 729
 730 static void __init calgary_reserve_peripheral_mem_1(struct pci_dev *dev)
 731 {
 732         void __iomem *target;
 733         u64 low, high, sizelow;
 734         u64 start, limit;
 735         struct iommu_table *tbl = pci_iommu(dev->bus);
 736         unsigned char busnum = dev->bus->number;
 737         void __iomem *bbar = tbl->bbar;
 738
 739         /* peripheral MEM_1 region */
 740         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_1_LOW);
 741         low = be32_to_cpu(readl(target));
 742         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_1_HIGH);
 743         high = be32_to_cpu(readl(target));
 744         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_1_SIZE);
 745         sizelow = be32_to_cpu(readl(target));
 746
 747         start = (high << 32) | low;
 748         limit = sizelow;
 749
 750         calgary_reserve_mem_region(dev, start, limit);
 751 }
 752
 753 static void __init calgary_reserve_peripheral_mem_2(struct pci_dev *dev)
 754 {
 755         void __iomem *target;
 756         u32 val32;
 757         u64 low, high, sizelow, sizehigh;
 758         u64 start, limit;
 759         struct iommu_table *tbl = pci_iommu(dev->bus);
 760         unsigned char busnum = dev->bus->number;
 761         void __iomem *bbar = tbl->bbar;
 762
 763         /* is it enabled? */
 764         target = calgary_reg(bbar, phb_offset(busnum) | PHB_CONFIG_RW_OFFSET);
 765         val32 = be32_to_cpu(readl(target));
 766         if (!(val32 & PHB_MEM2_ENABLE))
 767                 return;
 768
 769         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_2_LOW);
 770         low = be32_to_cpu(readl(target));
 771         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_2_HIGH);
 772         high = be32_to_cpu(readl(target));
 773         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_2_SIZE_LOW);
 774         sizelow = be32_to_cpu(readl(target));
 775         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_2_SIZE_HIGH);
 776         sizehigh = be32_to_cpu(readl(target));
 777
 778         start = (high << 32) | low;
 779         limit = (sizehigh << 32) | sizelow;
 780
 781         calgary_reserve_mem_region(dev, start, limit);
 782 }
 783
 784 /*
 785  * some regions of the IO address space do not get translated, so we
 786  * must not give devices IO addresses in those regions. The regions
 787  * are the 640KB-1MB region and the two PCI peripheral memory holes.
 788  * Reserve all of them in the IOMMU bitmap to avoid giving them out
 789  * later.
 790  */
 791 static void __init calgary_reserve_regions(struct pci_dev *dev)
 792 {
 793         unsigned int npages;
 794         u64 start;
 795         struct iommu_table *tbl = pci_iommu(dev->bus);
 796
 797         /* reserve EMERGENCY_PAGES from bad_dma_address and up */
 798         iommu_range_reserve(tbl, bad_dma_address, EMERGENCY_PAGES);
 799
 800         /* avoid the BIOS/VGA first 640KB-1MB region */
 801         /* for CalIOC2 - avoid the entire first MB */
 802         if (is_calgary(dev->device)) {
 803                 start = (640 * 1024);
 804                 npages = ((1024 - 640) * 1024) >> PAGE_SHIFT;
 805         } else { /* calioc2 */
 806                 start = 0;
 807                 npages = (1 * 1024 * 1024) >> PAGE_SHIFT;
 808         }
 809         iommu_range_reserve(tbl, start, npages);
 810
 811         /* reserve the two PCI peripheral memory regions in IO space */
 812         calgary_reserve_peripheral_mem_1(dev);
 813         calgary_reserve_peripheral_mem_2(dev);
 814 }
 815
 816 static int __init calgary_setup_tar(struct pci_dev *dev, void __iomem *bbar)
 817 {
 818         u64 val64;
 819         u64 table_phys;
 820         void __iomem *target;
 821         int ret;
 822         struct iommu_table *tbl;
 823
 824         /* build TCE tables for each PHB */
 825         ret = build_tce_table(dev, bbar);
 826         if (ret)
 827                 return ret;
 828
 829         tbl = pci_iommu(dev->bus);
 830         tbl->it_base = (unsigned long)bus_info[dev->bus->number].tce_space;
 831         tce_free(tbl, 0, tbl->it_size);
 832
 833         if (is_calgary(dev->device))
 834                 tbl->chip_ops = &calgary_chip_ops;
 835         else if (is_calioc2(dev->device))
 836                 tbl->chip_ops = &calioc2_chip_ops;
 837         else
 838                 BUG();
 839
 840         calgary_reserve_regions(dev);
 841
 842         /* set TARs for each PHB */
 843         target = calgary_reg(bbar, tar_offset(dev->bus->number));
 844         val64 = be64_to_cpu(readq(target));
 845
 846         /* zero out all TAR bits under sw control */
 847         val64 &= ~TAR_SW_BITS;
 848         table_phys = (u64)__pa(tbl->it_base);
 849
 850         val64 |= table_phys;
 851
 852         BUG_ON(specified_table_size > TCE_TABLE_SIZE_8M);
 853         val64 |= (u64) specified_table_size;
 854
 855         tbl->tar_val = cpu_to_be64(val64);
 856
 857         writeq(tbl->tar_val, target);
 858         readq(target); /* flush */
 859
 860         return 0;
 861 }
 862
 863 static void __init calgary_free_bus(struct pci_dev *dev)
 864 {
 865         u64 val64;
 866         struct iommu_table *tbl = pci_iommu(dev->bus);
 867         void __iomem *target;
 868         unsigned int bitmapsz;
 869
 870         target = calgary_reg(tbl->bbar, tar_offset(dev->bus->number));
 871         val64 = be64_to_cpu(readq(target));
 872         val64 &= ~TAR_SW_BITS;
 873         writeq(cpu_to_be64(val64), target);
 874         readq(target); /* flush */
 875
 876         bitmapsz = tbl->it_size / BITS_PER_BYTE;
 877         free_pages((unsigned long)tbl->it_map, get_order(bitmapsz));
 878         tbl->it_map = NULL;
 879
 880         kfree(tbl);
 881
 882         set_pci_iommu(dev->bus, NULL);
 883
 884         /* Can't free bootmem allocated memory after system is up :-( */
 885         bus_info[dev->bus->number].tce_space = NULL;
 886 }
 887
 888 static void calgary_dump_error_regs(struct iommu_table *tbl)
 889 {
 890         void __iomem *bbar = tbl->bbar;
 891         void __iomem *target;
 892         u32 csr, plssr;
 893
 894         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_CSR_OFFSET);
 895         csr = be32_to_cpu(readl(target));
 896
 897         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_PLSSR_OFFSET);
 898         plssr = be32_to_cpu(readl(target));
 899
 900         /* If no error, the agent ID in the CSR is not valid */
 901         printk(KERN_EMERG "Calgary: DMA error on Calgary PHB 0x%x, "
 902                "0x%08x@CSR 0x%08x@PLSSR\n", tbl->it_busno, csr, plssr);
 903 }
 904
 905 static void calioc2_dump_error_regs(struct iommu_table *tbl)
 906 {
 907         void __iomem *bbar = tbl->bbar;
 908         u32 csr, csmr, plssr, mck, rcstat;
 909         void __iomem *target;
 910         unsigned long phboff = phb_offset(tbl->it_busno);
 911         unsigned long erroff;
 912         u32 errregs[7];
 913         int i;
 914
 915         /* dump CSR */
 916         target = calgary_reg(bbar, phboff | PHB_CSR_OFFSET);
 917         csr = be32_to_cpu(readl(target));
 918         /* dump PLSSR */
 919         target = calgary_reg(bbar, phboff | PHB_PLSSR_OFFSET);
 920         plssr = be32_to_cpu(readl(target));
 921         /* dump CSMR */
 922         target = calgary_reg(bbar, phboff | 0x290);
 923         csmr = be32_to_cpu(readl(target));
 924         /* dump mck */
 925         target = calgary_reg(bbar, phboff | 0x800);
 926         mck = be32_to_cpu(readl(target));
 927
 928         printk(KERN_EMERG "Calgary: DMA error on CalIOC2 PHB 0x%x\n",
 929                tbl->it_busno);
 930
 931         printk(KERN_EMERG "Calgary: 0x%08x@CSR 0x%08x@PLSSR 0x%08x@CSMR 0x%08x@MCK\n",
 932                csr, plssr, csmr, mck);
 933
 934         /* dump rest of error regs */
 935         printk(KERN_EMERG "Calgary: ");
 936         for (i = 0; i < ARRAY_SIZE(errregs); i++) {
 937                 /* err regs are at 0x810 - 0x870 */
 938                 erroff = (0x810 + (i * 0x10));
 939                 target = calgary_reg(bbar, phboff | erroff);
 940                 errregs[i] = be32_to_cpu(readl(target));
 941                 printk("0x%08x@0x%lx ", errregs[i], erroff);
 942         }
 943         printk("\n");
 944
 945         /* root complex status */
 946         target = calgary_reg(bbar, phboff | PHB_ROOT_COMPLEX_STATUS);
 947         rcstat = be32_to_cpu(readl(target));
 948         printk(KERN_EMERG "Calgary: 0x%08x@0x%x\n", rcstat,
 949                PHB_ROOT_COMPLEX_STATUS);
 950 }
 951
 952 static void calgary_watchdog(unsigned long data)
 953 {
 954         struct pci_dev *dev = (struct pci_dev *)data;
 955         struct iommu_table *tbl = pci_iommu(dev->bus);
 956         void __iomem *bbar = tbl->bbar;
 957         u32 val32;
 958         void __iomem *target;
 959
 960         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_CSR_OFFSET);
 961         val32 = be32_to_cpu(readl(target));
 962
 963         /* If no error, the agent ID in the CSR is not valid */
 964         if (val32 & CSR_AGENT_MASK) {
 965                 tbl->chip_ops->dump_error_regs(tbl);
 966
 967                 /* reset error */
 968                 writel(0, target);
 969
 970                 /* Disable bus that caused the error */
 971                 target = calgary_reg(bbar, phb_offset(tbl->it_busno) |
 972                                      PHB_CONFIG_RW_OFFSET);
 973                 val32 = be32_to_cpu(readl(target));
 974                 val32 |= PHB_SLOT_DISABLE;
 975                 writel(cpu_to_be32(val32), target);
 976                 readl(target); /* flush */
 977         } else {
 978                 /* Reset the timer */
 979                 mod_timer(&tbl->watchdog_timer, jiffies + 2 * HZ);
 980         }
 981 }
 982
 983 static void __init calgary_set_split_completion_timeout(void __iomem *bbar,
 984         unsigned char busnum, unsigned long timeout)
 985 {
 986         u64 val64;
 987         void __iomem *target;
 988         unsigned int phb_shift = ~0; /* silence gcc */
 989         u64 mask;
 990
 991         switch (busno_to_phbid(busnum)) {
 992         case 0: phb_shift = (63 - 19);
 993                 break;
 994         case 1: phb_shift = (63 - 23);
 995                 break;
 996         case 2: phb_shift = (63 - 27);
 997                 break;
 998         case 3: phb_shift = (63 - 35);
 999                 break;
1000         default:
1001                 BUG_ON(busno_to_phbid(busnum));
1002         }
1003
1004         target = calgary_reg(bbar, CALGARY_CONFIG_REG);
1005         val64 = be64_to_cpu(readq(target));
1006
1007         /* zero out this PHB's timer bits */
1008         mask = ~(0xFUL << phb_shift);
1009         val64 &= mask;
1010         val64 |= (timeout << phb_shift);
1011         writeq(cpu_to_be64(val64), target);
1012         readq(target); /* flush */
1013 }
1014
1015 static void __init calioc2_handle_quirks(struct iommu_table *tbl, struct pci_dev *dev)
1016 {
1017         unsigned char busnum = dev->bus->number;
1018         void __iomem *bbar = tbl->bbar;
1019         void __iomem *target;
1020         u32 val;
1021
1022         /*
1023          * CalIOC2 designers recommend setting bit 8 in 0xnDB0 to 1
1024          */
1025         target = calgary_reg(bbar, phb_offset(busnum) | PHB_SAVIOR_L2);
1026         val = cpu_to_be32(readl(target));
1027         val |= 0x00800000;
1028         writel(cpu_to_be32(val), target);
1029 }
1030
1031 static void __init calgary_handle_quirks(struct iommu_table *tbl, struct pci_dev *dev)
1032 {
1033         unsigned char busnum = dev->bus->number;
1034
1035         /*
1036          * Give split completion a longer timeout on bus 1 for aic94xx
1037          * http://bugzilla.kernel.org/show_bug.cgi?id=7180
1038          */
1039         if (is_calgary(dev->device) && (busnum == 1))
1040                 calgary_set_split_completion_timeout(tbl->bbar, busnum,
1041                                                      CCR_2SEC_TIMEOUT);
1042 }
1043
1044 static void __init calgary_enable_translation(struct pci_dev *dev)
1045 {
1046         u32 val32;
1047         unsigned char busnum;
1048         void __iomem *target;
1049         void __iomem *bbar;
1050         struct iommu_table *tbl;
1051
1052         busnum = dev->bus->number;
1053         tbl = pci_iommu(dev->bus);
1054         bbar = tbl->bbar;
1055
1056         /* enable TCE in PHB Config Register */
1057         target = calgary_reg(bbar, phb_offset(busnum) | PHB_CONFIG_RW_OFFSET);
1058         val32 = be32_to_cpu(readl(target));
1059         val32 |= PHB_TCE_ENABLE | PHB_DAC_DISABLE | PHB_MCSR_ENABLE;
1060
1061         printk(KERN_INFO "Calgary: enabling translation on %s PHB %#x\n",
1062                (dev->device == PCI_DEVICE_ID_IBM_CALGARY) ?
1063                "Calgary" : "CalIOC2", busnum);
1064         printk(KERN_INFO "Calgary: errant DMAs will now be prevented on this "
1065                "bus.\n");
1066
1067         writel(cpu_to_be32(val32), target);
1068         readl(target); /* flush */
1069
1070         init_timer(&tbl->watchdog_timer);
1071         tbl->watchdog_timer.function = &calgary_watchdog;
1072         tbl->watchdog_timer.data = (unsigned long)dev;
1073         mod_timer(&tbl->watchdog_timer, jiffies);
1074 }
1075
1076 static void __init calgary_disable_translation(struct pci_dev *dev)
1077 {
1078         u32 val32;
1079         unsigned char busnum;
1080         void __iomem *target;
1081         void __iomem *bbar;
1082         struct iommu_table *tbl;
1083
1084         busnum = dev->bus->number;
1085         tbl = pci_iommu(dev->bus);
1086         bbar = tbl->bbar;
1087
1088         /* disable TCE in PHB Config Register */
1089         target = calgary_reg(bbar, phb_offset(busnum) | PHB_CONFIG_RW_OFFSET);
1090         val32 = be32_to_cpu(readl(target));
1091         val32 &= ~(PHB_TCE_ENABLE | PHB_DAC_DISABLE | PHB_MCSR_ENABLE);
1092
1093         printk(KERN_INFO "Calgary: disabling translation on PHB %#x!\n", busnum);
1094         writel(cpu_to_be32(val32), target);
1095         readl(target); /* flush */
1096
1097         del_timer_sync(&tbl->watchdog_timer);
1098 }
1099
1100 static void __init calgary_init_one_nontraslated(struct pci_dev *dev)
1101 {
1102         pci_dev_get(dev);
1103         set_pci_iommu(dev->bus, NULL);
1104
1105         /* is the device behind a bridge? */
1106         if (dev->bus->parent)
1107                 dev->bus->parent->self = dev;
1108         else
1109                 dev->bus->self = dev;
1110 }
1111
1112 static int __init calgary_init_one(struct pci_dev *dev)
1113 {
1114         void __iomem *bbar;
1115         struct iommu_table *tbl;
1116         int ret;
1117
1118         BUG_ON(dev->bus->number >= MAX_PHB_BUS_NUM);
1119
1120         bbar = busno_to_bbar(dev->bus->number);
1121         ret = calgary_setup_tar(dev, bbar);
1122         if (ret)
1123                 goto done;
1124
1125         pci_dev_get(dev);
1126
1127         if (dev->bus->parent) {
1128                 if (dev->bus->parent->self)
1129                         printk(KERN_WARNING "Calgary: IEEEE, dev %p has "
1130                                "bus->parent->self!\n", dev);
1131                 dev->bus->parent->self = dev;
1132         } else
1133                 dev->bus->self = dev;
1134
1135         tbl = pci_iommu(dev->bus);
1136         tbl->chip_ops->handle_quirks(tbl, dev);
1137
1138         calgary_enable_translation(dev);
1139
1140         return 0;
1141
1142 done:
1143         return ret;
1144 }
1145
1146 static int __init calgary_locate_bbars(void)
1147 {
1148         int ret;
1149         int rioidx, phb, bus;
1150         void __iomem *bbar;
1151         void __iomem *target;
1152         unsigned long offset;
1153         u8 start_bus, end_bus;
1154         u32 val;
1155
1156         ret = -ENODATA;
1157         for (rioidx = 0; rioidx < rio_table_hdr->num_rio_dev; rioidx++) {
1158                 struct rio_detail *rio = rio_devs[rioidx];
1159
1160                 if ((rio->type != COMPAT_CALGARY) && (rio->type != ALT_CALGARY))
1161                         continue;
1162
1163                 /* map entire 1MB of Calgary config space */
1164                 bbar = ioremap_nocache(rio->BBAR, 1024 * 1024);
1165                 if (!bbar)
1166                         goto error;
1167
1168                 for (phb = 0; phb < PHBS_PER_CALGARY; phb++) {
1169                         offset = phb_debug_offsets[phb] | PHB_DEBUG_STUFF_OFFSET;
1170                         target = calgary_reg(bbar, offset);
1171
1172                         val = be32_to_cpu(readl(target));
1173
1174                         start_bus = (u8)((val & 0x00FF0000) >> 16);
1175                         end_bus = (u8)((val & 0x0000FF00) >> 8);
1176
1177                         if (end_bus) {
1178                                 for (bus = start_bus; bus <= end_bus; bus++) {
1179                                         bus_info[bus].bbar = bbar;
1180                                         bus_info[bus].phbid = phb;
1181                                 }
1182                         } else {
1183                                 bus_info[start_bus].bbar = bbar;
1184                                 bus_info[start_bus].phbid = phb;
1185                         }
1186                 }
1187         }
1188
1189         return 0;
1190
1191 error:
1192         /* scan bus_info and iounmap any bbars we previously ioremap'd */
1193         for (bus = 0; bus < ARRAY_SIZE(bus_info); bus++)
1194                 if (bus_info[bus].bbar)
1195                         iounmap(bus_info[bus].bbar);
1196
1197         return ret;
1198 }
1199
1200 static int __init calgary_init(void)
1201 {
1202         int ret;
1203         struct pci_dev *dev = NULL;
1204         struct calgary_bus_info *info;
1205
1206         ret = calgary_locate_bbars();
1207         if (ret)
1208                 return ret;
1209
1210         do {
1211                 dev = pci_get_device(PCI_VENDOR_ID_IBM, PCI_ANY_ID, dev);
1212                 if (!dev)
1213                         break;
1214                 if (!is_cal_pci_dev(dev->device))
1215                         continue;
1216
1217                 info = &bus_info[dev->bus->number];
1218                 if (info->translation_disabled) {
1219                         calgary_init_one_nontraslated(dev);
1220                         continue;
1221                 }
1222
1223                 if (!info->tce_space && !translate_empty_slots)
1224                         continue;
1225
1226                 ret = calgary_init_one(dev);
1227                 if (ret)
1228                         goto error;
1229         } while (1);
1230
1231         return ret;
1232
1233 error:
1234         do {
1235                 dev = pci_get_device_reverse(PCI_VENDOR_ID_IBM,
1236                                              PCI_ANY_ID, dev);
1237                 if (!dev)
1238                         break;
1239                 if (!is_cal_pci_dev(dev->device))
1240                         continue;
1241
1242                 info = &bus_info[dev->bus->number];
1243                 if (info->translation_disabled) {
1244                         pci_dev_put(dev);
1245                         continue;
1246                 }
1247                 if (!info->tce_space && !translate_empty_slots)
1248                         continue;
1249
1250                 calgary_disable_translation(dev);
1251                 calgary_free_bus(dev);
1252                 pci_dev_put(dev); /* Undo calgary_init_one()'s pci_dev_get() */
1253         } while (1);
1254
1255         return ret;
1256 }
1257
1258 static inline int __init determine_tce_table_size(u64 ram)
1259 {
1260         int ret;
1261
1262         if (specified_table_size != TCE_TABLE_SIZE_UNSPECIFIED)
1263                 return specified_table_size;
1264
1265         /*
1266          * Table sizes are from 0 to 7 (TCE_TABLE_SIZE_64K to
1267          * TCE_TABLE_SIZE_8M). Table size 0 has 8K entries and each
1268          * larger table size has twice as many entries, so shift the
1269          * max ram address by 13 to divide by 8K and then look at the
1270          * order of the result to choose between 0-7.
1271          */
1272         ret = get_order(ram >> 13);
1273         if (ret > TCE_TABLE_SIZE_8M)
1274                 ret = TCE_TABLE_SIZE_8M;
1275
1276         return ret;
1277 }
1278
1279 static int __init build_detail_arrays(void)
1280 {
1281         unsigned long ptr;
1282         int i, scal_detail_size, rio_detail_size;
1283
1284         if (rio_table_hdr->num_scal_dev > MAX_NUMNODES){
1285                 printk(KERN_WARNING
1286                         "Calgary: MAX_NUMNODES too low! Defined as %d, "
1287                         "but system has %d nodes.\n",
1288                         MAX_NUMNODES, rio_table_hdr->num_scal_dev);
1289                 return -ENODEV;
1290         }
1291
1292         switch (rio_table_hdr->version){
1293         case 2:
1294                 scal_detail_size = 11;
1295                 rio_detail_size = 13;
1296                 break;
1297         case 3:
1298                 scal_detail_size = 12;
1299                 rio_detail_size = 15;
1300                 break;
1301         default:
1302                 printk(KERN_WARNING
1303                        "Calgary: Invalid Rio Grande Table Version: %d\n",
1304                        rio_table_hdr->version);
1305                 return -EPROTO;
1306         }
1307
1308         ptr = ((unsigned long)rio_table_hdr) + 3;
1309         for (i = 0; i < rio_table_hdr->num_scal_dev;
1310                     i++, ptr += scal_detail_size)
1311                 scal_devs[i] = (struct scal_detail *)ptr;
1312
1313         for (i = 0; i < rio_table_hdr->num_rio_dev;
1314                     i++, ptr += rio_detail_size)
1315                 rio_devs[i] = (struct rio_detail *)ptr;
1316
1317         return 0;
1318 }
1319
1320 static int __init calgary_bus_has_devices(int bus, unsigned short pci_dev)
1321 {
1322         int dev;
1323         u32 val;
1324
1325         if (pci_dev == PCI_DEVICE_ID_IBM_CALIOC2) {
1326                 /*
1327                  * FIXME: properly scan for devices accross the
1328                  * PCI-to-PCI bridge on every CalIOC2 port.
1329                  */
1330                 return 1;
1331         }
1332
1333         for (dev = 1; dev < 8; dev++) {
1334                 val = read_pci_config(bus, dev, 0, 0);
1335                 if (val != 0xffffffff)
1336                         break;
1337         }
1338         return (val != 0xffffffff);
1339 }
1340
1341 void __init detect_calgary(void)
1342 {
1343         int bus;
1344         void *tbl;
1345         int calgary_found = 0;
1346         unsigned long ptr;
1347         unsigned int offset, prev_offset;
1348         int ret;
1349
1350         /*
1351          * if the user specified iommu=off or iommu=soft or we found
1352          * another HW IOMMU already, bail out.
1353          */
1354         if (swiotlb || no_iommu || iommu_detected)
1355                 return;
1356
1357         if (!use_calgary)
1358                 return;
1359
1360         if (!early_pci_allowed())
1361                 return;
1362
1363         printk(KERN_DEBUG "Calgary: detecting Calgary via BIOS EBDA area\n");
1364
1365         ptr = (unsigned long)phys_to_virt(get_bios_ebda());
1366
1367         rio_table_hdr = NULL;
1368         prev_offset = 0;
1369         offset = 0x180;
1370         /*
1371          * The next offset is stored in the 1st word.
1372          * Only parse up until the offset increases:
1373          */
1374         while (offset > prev_offset) {
1375                 /* The block id is stored in the 2nd word */
1376                 if (*((unsigned short *)(ptr + offset + 2)) == 0x4752){
1377                         /* set the pointer past the offset & block id */
1378                         rio_table_hdr = (struct rio_table_hdr *)(ptr + offset + 4);
1379                         break;
1380                 }
1381                 prev_offset = offset;
1382                 offset = *((unsigned short *)(ptr + offset));
1383         }
1384         if (!rio_table_hdr) {
1385                 printk(KERN_DEBUG "Calgary: Unable to locate Rio Grande table "
1386                        "in EBDA - bailing!\n");
1387                 return;
1388         }
1389
1390         ret = build_detail_arrays();
1391         if (ret) {
1392                 printk(KERN_DEBUG "Calgary: build_detail_arrays ret %d\n", ret);
1393                 return;
1394         }
1395
1396         specified_table_size = determine_tce_table_size(end_pfn * PAGE_SIZE);
1397
1398         for (bus = 0; bus < MAX_PHB_BUS_NUM; bus++) {
1399                 struct calgary_bus_info *info = &bus_info[bus];
1400                 unsigned short pci_device;
1401                 u32 val;
1402
1403                 val = read_pci_config(bus, 0, 0, 0);
1404                 pci_device = (val & 0xFFFF0000) >> 16;
1405
1406                 if (!is_cal_pci_dev(pci_device))
1407                         continue;
1408
1409                 if (info->translation_disabled)
1410                         continue;
1411
1412                 if (calgary_bus_has_devices(bus, pci_device) ||
1413                     translate_empty_slots) {
1414                         tbl = alloc_tce_table();
1415                         if (!tbl)
1416                                 goto cleanup;
1417                         info->tce_space = tbl;
1418                         calgary_found = 1;
1419                 }
1420         }
1421
1422         printk(KERN_DEBUG "Calgary: finished detection, Calgary %s\n",
1423                calgary_found ? "found" : "not found");
1424
1425         if (calgary_found) {
1426                 iommu_detected = 1;
1427                 calgary_detected = 1;
1428                 printk(KERN_INFO "PCI-DMA: Calgary IOMMU detected.\n");
1429                 printk(KERN_INFO "PCI-DMA: Calgary TCE table spec is %d, "
1430                        "CONFIG_IOMMU_DEBUG is %s.\n", specified_table_size,
1431                        debugging ? "enabled" : "disabled");
1432         }
1433         return;
1434
1435 cleanup:
1436         for (--bus; bus >= 0; --bus) {
1437                 struct calgary_bus_info *info = &bus_info[bus];
1438
1439                 if (info->tce_space)
1440                         free_tce_table(info->tce_space);
1441         }
1442 }
1443
1444 int __init calgary_iommu_init(void)
1445 {
1446         int ret;
1447
1448         if (no_iommu || swiotlb)
1449                 return -ENODEV;
1450
1451         if (!calgary_detected)
1452                 return -ENODEV;
1453
1454         /* ok, we're trying to use Calgary - let's roll */
1455         printk(KERN_INFO "PCI-DMA: Using Calgary IOMMU\n");
1456
1457         ret = calgary_init();
1458         if (ret) {
1459                 printk(KERN_ERR "PCI-DMA: Calgary init failed %d, "
1460                        "falling back to no_iommu\n", ret);
1461                 if (end_pfn > MAX_DMA32_PFN)
1462                         printk(KERN_ERR "WARNING more than 4GB of memory, "
1463                                         "32bit PCI may malfunction.\n");
1464                 return ret;
1465         }
1466
1467         force_iommu = 1;
1468         bad_dma_address = 0x0;
1469         dma_ops = &calgary_dma_ops;
1470
1471         return 0;
1472 }
1473
1474 static int __init calgary_parse_options(char *p)
1475 {
1476         unsigned int bridge;
1477         size_t len;
1478         char* endp;
1479
1480         while (*p) {
1481                 if (!strncmp(p, "64k", 3))
1482                         specified_table_size = TCE_TABLE_SIZE_64K;
1483                 else if (!strncmp(p, "128k", 4))
1484                         specified_table_size = TCE_TABLE_SIZE_128K;
1485                 else if (!strncmp(p, "256k", 4))
1486                         specified_table_size = TCE_TABLE_SIZE_256K;
1487                 else if (!strncmp(p, "512k", 4))
1488                         specified_table_size = TCE_TABLE_SIZE_512K;
1489                 else if (!strncmp(p, "1M", 2))
1490                         specified_table_size = TCE_TABLE_SIZE_1M;
1491                 else if (!strncmp(p, "2M", 2))
1492                         specified_table_size = TCE_TABLE_SIZE_2M;
1493                 else if (!strncmp(p, "4M", 2))
1494                         specified_table_size = TCE_TABLE_SIZE_4M;
1495                 else if (!strncmp(p, "8M", 2))
1496                         specified_table_size = TCE_TABLE_SIZE_8M;
1497
1498                 len = strlen("translate_empty_slots");
1499                 if (!strncmp(p, "translate_empty_slots", len))
1500                         translate_empty_slots = 1;
1501
1502                 len = strlen("disable");
1503                 if (!strncmp(p, "disable", len)) {
1504                         p += len;
1505                         if (*p == '=')
1506                                 ++p;
1507                         if (*p == '\0')
1508                                 break;
1509                         bridge = simple_strtol(p, &endp, 0);
1510                         if (p == endp)
1511                                 break;
1512
1513                         if (bridge < MAX_PHB_BUS_NUM) {
1514                                 printk(KERN_INFO "Calgary: disabling "
1515                                        "translation for PHB %#x\n", bridge);
1516                                 bus_info[bridge].translation_disabled = 1;
1517                         }
1518                 }
1519
1520                 p = strpbrk(p, ",");
1521                 if (!p)
1522                         break;
1523
1524                 p++; /* skip ',' */
1525         }
1526         return 1;
1527 }
1528 __setup("calgary=", calgary_parse_options);
1529
1530 static void __init calgary_fixup_one_tce_space(struct pci_dev *dev)
1531 {
1532         struct iommu_table *tbl;
1533         unsigned int npages;
1534         int i;
1535
1536         tbl = pci_iommu(dev->bus);
1537
1538         for (i = 0; i < 4; i++) {
1539                 struct resource *r = &dev->resource[PCI_BRIDGE_RESOURCES + i];
1540
1541                 /* Don't give out TCEs that map MEM resources */
1542                 if (!(r->flags & IORESOURCE_MEM))
1543                         continue;
1544
1545                 /* 0-based? we reserve the whole 1st MB anyway */
1546                 if (!r->start)
1547                         continue;
1548
1549                 /* cover the whole region */
1550                 npages = (r->end - r->start) >> PAGE_SHIFT;
1551                 npages++;
1552
1553                 iommu_range_reserve(tbl, r->start, npages);
1554         }
1555 }
1556
1557 static int __init calgary_fixup_tce_spaces(void)
1558 {
1559         struct pci_dev *dev = NULL;
1560         struct calgary_bus_info *info;
1561
1562         if (no_iommu || swiotlb || !calgary_detected)
1563                 return -ENODEV;
1564
1565         printk(KERN_DEBUG "Calgary: fixing up tce spaces\n");
1566
1567         do {
1568                 dev = pci_get_device(PCI_VENDOR_ID_IBM, PCI_ANY_ID, dev);
1569                 if (!dev)
1570                         break;
1571                 if (!is_cal_pci_dev(dev->device))
1572                         continue;
1573
1574                 info = &bus_info[dev->bus->number];
1575                 if (info->translation_disabled)
1576                         continue;
1577
1578                 if (!info->tce_space)
1579                         continue;
1580
1581                 calgary_fixup_one_tce_space(dev);
1582
1583         } while (1);
1584
1585         return 0;
1586 }
1587
1588 /*
1589  * We need to be call after pcibios_assign_resources (fs_initcall level)
1590  * and before device_initcall.
1591  */
1592 rootfs_initcall(calgary_fixup_tce_spaces);