git.oblomov.eu Git - linux-2.6/blob - 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/crash_dump.h>
  33 #include <linux/dma-mapping.h>
  34 #include <linux/bitops.h>
  35 #include <linux/pci_ids.h>
  36 #include <linux/pci.h>
  37 #include <linux/delay.h>
  38 #include <linux/scatterlist.h>
  39 #include <linux/iommu-helper.h>
  40
  41 #include <asm/iommu.h>
  42 #include <asm/calgary.h>
  43 #include <asm/tce.h>
  44 #include <asm/pci-direct.h>
  45 #include <asm/system.h>
  46 #include <asm/dma.h>
  47 #include <asm/rio.h>
  48 #include <asm/bios_ebda.h>
  49
  50 #ifdef CONFIG_CALGARY_IOMMU_ENABLED_BY_DEFAULT
  51 int use_calgary __read_mostly = 1;
  52 #else
  53 int use_calgary __read_mostly = 0;
  54 #endif /* CONFIG_CALGARY_DEFAULT_ENABLED */
  55
  56 #define PCI_DEVICE_ID_IBM_CALGARY 0x02a1
  57 #define PCI_DEVICE_ID_IBM_CALIOC2 0x0308
  58
  59 /* register offsets inside the host bridge space */
  60 #define CALGARY_CONFIG_REG      0x0108
  61 #define PHB_CSR_OFFSET          0x0110 /* Channel Status */
  62 #define PHB_PLSSR_OFFSET        0x0120
  63 #define PHB_CONFIG_RW_OFFSET    0x0160
  64 #define PHB_IOBASE_BAR_LOW      0x0170
  65 #define PHB_IOBASE_BAR_HIGH     0x0180
  66 #define PHB_MEM_1_LOW           0x0190
  67 #define PHB_MEM_1_HIGH          0x01A0
  68 #define PHB_IO_ADDR_SIZE        0x01B0
  69 #define PHB_MEM_1_SIZE          0x01C0
  70 #define PHB_MEM_ST_OFFSET       0x01D0
  71 #define PHB_AER_OFFSET          0x0200
  72 #define PHB_CONFIG_0_HIGH       0x0220
  73 #define PHB_CONFIG_0_LOW        0x0230
  74 #define PHB_CONFIG_0_END        0x0240
  75 #define PHB_MEM_2_LOW           0x02B0
  76 #define PHB_MEM_2_HIGH          0x02C0
  77 #define PHB_MEM_2_SIZE_HIGH     0x02D0
  78 #define PHB_MEM_2_SIZE_LOW      0x02E0
  79 #define PHB_DOSHOLE_OFFSET      0x08E0
  80
  81 /* CalIOC2 specific */
  82 #define PHB_SAVIOR_L2           0x0DB0
  83 #define PHB_PAGE_MIG_CTRL       0x0DA8
  84 #define PHB_PAGE_MIG_DEBUG      0x0DA0
  85 #define PHB_ROOT_COMPLEX_STATUS 0x0CB0
  86
  87 /* PHB_CONFIG_RW */
  88 #define PHB_TCE_ENABLE          0x20000000
  89 #define PHB_SLOT_DISABLE        0x1C000000
  90 #define PHB_DAC_DISABLE         0x01000000
  91 #define PHB_MEM2_ENABLE         0x00400000
  92 #define PHB_MCSR_ENABLE         0x00100000
  93 /* TAR (Table Address Register) */
  94 #define TAR_SW_BITS             0x0000ffffffff800fUL
  95 #define TAR_VALID               0x0000000000000008UL
  96 /* CSR (Channel/DMA Status Register) */
  97 #define CSR_AGENT_MASK          0xffe0ffff
  98 /* CCR (Calgary Configuration Register) */
  99 #define CCR_2SEC_TIMEOUT        0x000000000000000EUL
 100 /* PMCR/PMDR (Page Migration Control/Debug Registers */
 101 #define PMR_SOFTSTOP            0x80000000
 102 #define PMR_SOFTSTOPFAULT       0x40000000
 103 #define PMR_HARDSTOP            0x20000000
 104
 105 #define MAX_NUM_OF_PHBS         8 /* how many PHBs in total? */
 106 #define MAX_NUM_CHASSIS         8 /* max number of chassis */
 107 /* MAX_PHB_BUS_NUM is the maximal possible dev->bus->number */
 108 #define MAX_PHB_BUS_NUM         (MAX_NUM_OF_PHBS * MAX_NUM_CHASSIS * 2)
 109 #define PHBS_PER_CALGARY        4
 110
 111 /* register offsets in Calgary's internal register space */
 112 static const unsigned long tar_offsets[] = {
 113         0x0580 /* TAR0 */,
 114         0x0588 /* TAR1 */,
 115         0x0590 /* TAR2 */,
 116         0x0598 /* TAR3 */
 117 };
 118
 119 static const unsigned long split_queue_offsets[] = {
 120         0x4870 /* SPLIT QUEUE 0 */,
 121         0x5870 /* SPLIT QUEUE 1 */,
 122         0x6870 /* SPLIT QUEUE 2 */,
 123         0x7870 /* SPLIT QUEUE 3 */
 124 };
 125
 126 static const unsigned long phb_offsets[] = {
 127         0x8000 /* PHB0 */,
 128         0x9000 /* PHB1 */,
 129         0xA000 /* PHB2 */,
 130         0xB000 /* PHB3 */
 131 };
 132
 133 /* PHB debug registers */
 134
 135 static const unsigned long phb_debug_offsets[] = {
 136         0x4000  /* PHB 0 DEBUG */,
 137         0x5000  /* PHB 1 DEBUG */,
 138         0x6000  /* PHB 2 DEBUG */,
 139         0x7000  /* PHB 3 DEBUG */
 140 };
 141
 142 /*
 143  * STUFF register for each debug PHB,
 144  * byte 1 = start bus number, byte 2 = end bus number
 145  */
 146
 147 #define PHB_DEBUG_STUFF_OFFSET  0x0020
 148
 149 #define EMERGENCY_PAGES 32 /* = 128KB */
 150
 151 unsigned int specified_table_size = TCE_TABLE_SIZE_UNSPECIFIED;
 152 static int translate_empty_slots __read_mostly = 0;
 153 static int calgary_detected __read_mostly = 0;
 154
 155 static struct rio_table_hdr     *rio_table_hdr __initdata;
 156 static struct scal_detail       *scal_devs[MAX_NUMNODES] __initdata;
 157 static struct rio_detail        *rio_devs[MAX_NUMNODES * 4] __initdata;
 158
 159 struct calgary_bus_info {
 160         void *tce_space;
 161         unsigned char translation_disabled;
 162         signed char phbid;
 163         void __iomem *bbar;
 164 };
 165
 166 static void calgary_handle_quirks(struct iommu_table *tbl, struct pci_dev *dev);
 167 static void calgary_tce_cache_blast(struct iommu_table *tbl);
 168 static void calgary_dump_error_regs(struct iommu_table *tbl);
 169 static void calioc2_handle_quirks(struct iommu_table *tbl, struct pci_dev *dev);
 170 static void calioc2_tce_cache_blast(struct iommu_table *tbl);
 171 static void calioc2_dump_error_regs(struct iommu_table *tbl);
 172 static void calgary_init_bitmap_from_tce_table(struct iommu_table *tbl);
 173 static void get_tce_space_from_tar(void);
 174
 175 static struct cal_chipset_ops calgary_chip_ops = {
 176         .handle_quirks = calgary_handle_quirks,
 177         .tce_cache_blast = calgary_tce_cache_blast,
 178         .dump_error_regs = calgary_dump_error_regs
 179 };
 180
 181 static struct cal_chipset_ops calioc2_chip_ops = {
 182         .handle_quirks = calioc2_handle_quirks,
 183         .tce_cache_blast = calioc2_tce_cache_blast,
 184         .dump_error_regs = calioc2_dump_error_regs
 185 };
 186
 187 static struct calgary_bus_info bus_info[MAX_PHB_BUS_NUM] = { { NULL, 0, 0 }, };
 188
 189 /* enable this to stress test the chip's TCE cache */
 190 #ifdef CONFIG_IOMMU_DEBUG
 191 static int debugging = 1;
 192
 193 static inline unsigned long verify_bit_range(unsigned long* bitmap,
 194         int expected, unsigned long start, unsigned long end)
 195 {
 196         unsigned long idx = start;
 197
 198         BUG_ON(start >= end);
 199
 200         while (idx < end) {
 201                 if (!!test_bit(idx, bitmap) != expected)
 202                         return idx;
 203                 ++idx;
 204         }
 205
 206         /* all bits have the expected value */
 207         return ~0UL;
 208 }
 209 #else /* debugging is disabled */
 210 static int debugging;
 211
 212 static inline unsigned long verify_bit_range(unsigned long* bitmap,
 213         int expected, unsigned long start, unsigned long end)
 214 {
 215         return ~0UL;
 216 }
 217
 218 #endif /* CONFIG_IOMMU_DEBUG */
 219
 220 static inline int translation_enabled(struct iommu_table *tbl)
 221 {
 222         /* only PHBs with translation enabled have an IOMMU table */
 223         return (tbl != NULL);
 224 }
 225
 226 static void iommu_range_reserve(struct iommu_table *tbl,
 227         unsigned long start_addr, unsigned int npages)
 228 {
 229         unsigned long index;
 230         unsigned long end;
 231         unsigned long badbit;
 232         unsigned long flags;
 233
 234         index = start_addr >> PAGE_SHIFT;
 235
 236         /* bail out if we're asked to reserve a region we don't cover */
 237         if (index >= tbl->it_size)
 238                 return;
 239
 240         end = index + npages;
 241         if (end > tbl->it_size) /* don't go off the table */
 242                 end = tbl->it_size;
 243
 244         spin_lock_irqsave(&tbl->it_lock, flags);
 245
 246         badbit = verify_bit_range(tbl->it_map, 0, index, end);
 247         if (badbit != ~0UL) {
 248                 if (printk_ratelimit())
 249                         printk(KERN_ERR "Calgary: entry already allocated at "
 250                                "0x%lx tbl %p dma 0x%lx npages %u\n",
 251                                badbit, tbl, start_addr, npages);
 252         }
 253
 254         iommu_area_reserve(tbl->it_map, index, npages);
 255
 256         spin_unlock_irqrestore(&tbl->it_lock, flags);
 257 }
 258
 259 static unsigned long iommu_range_alloc(struct device *dev,
 260                                        struct iommu_table *tbl,
 261                                        unsigned int npages)
 262 {
 263         unsigned long flags;
 264         unsigned long offset;
 265         unsigned long boundary_size;
 266
 267         boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
 268                               PAGE_SIZE) >> PAGE_SHIFT;
 269
 270         BUG_ON(npages == 0);
 271
 272         spin_lock_irqsave(&tbl->it_lock, flags);
 273
 274         offset = iommu_area_alloc(tbl->it_map, tbl->it_size, tbl->it_hint,
 275                                   npages, 0, boundary_size, 0);
 276         if (offset == ~0UL) {
 277                 tbl->chip_ops->tce_cache_blast(tbl);
 278
 279                 offset = iommu_area_alloc(tbl->it_map, tbl->it_size, 0,
 280                                           npages, 0, boundary_size, 0);
 281                 if (offset == ~0UL) {
 282                         printk(KERN_WARNING "Calgary: IOMMU full.\n");
 283                         spin_unlock_irqrestore(&tbl->it_lock, flags);
 284                         if (panic_on_overflow)
 285                                 panic("Calgary: fix the allocator.\n");
 286                         else
 287                                 return bad_dma_address;
 288                 }
 289         }
 290
 291         tbl->it_hint = offset + npages;
 292         BUG_ON(tbl->it_hint > tbl->it_size);
 293
 294         spin_unlock_irqrestore(&tbl->it_lock, flags);
 295
 296         return offset;
 297 }
 298
 299 static dma_addr_t iommu_alloc(struct device *dev, struct iommu_table *tbl,
 300                               void *vaddr, unsigned int npages, int direction)
 301 {
 302         unsigned long entry;
 303         dma_addr_t ret = bad_dma_address;
 304
 305         entry = iommu_range_alloc(dev, tbl, npages);
 306
 307         if (unlikely(entry == bad_dma_address))
 308                 goto error;
 309
 310         /* set the return dma address */
 311         ret = (entry << PAGE_SHIFT) | ((unsigned long)vaddr & ~PAGE_MASK);
 312
 313         /* put the TCEs in the HW table */
 314         tce_build(tbl, entry, npages, (unsigned long)vaddr & PAGE_MASK,
 315                   direction);
 316
 317         return ret;
 318
 319 error:
 320         printk(KERN_WARNING "Calgary: failed to allocate %u pages in "
 321                "iommu %p\n", npages, tbl);
 322         return bad_dma_address;
 323 }
 324
 325 static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
 326         unsigned int npages)
 327 {
 328         unsigned long entry;
 329         unsigned long badbit;
 330         unsigned long badend;
 331         unsigned long flags;
 332
 333         /* were we called with bad_dma_address? */
 334         badend = bad_dma_address + (EMERGENCY_PAGES * PAGE_SIZE);
 335         if (unlikely((dma_addr >= bad_dma_address) && (dma_addr < badend))) {
 336                 WARN(1, KERN_ERR "Calgary: driver tried unmapping bad DMA "
 337                        "address 0x%Lx\n", dma_addr);
 338                 return;
 339         }
 340
 341         entry = dma_addr >> PAGE_SHIFT;
 342
 343         BUG_ON(entry + npages > tbl->it_size);
 344
 345         tce_free(tbl, entry, npages);
 346
 347         spin_lock_irqsave(&tbl->it_lock, flags);
 348
 349         badbit = verify_bit_range(tbl->it_map, 1, entry, entry + npages);
 350         if (badbit != ~0UL) {
 351                 if (printk_ratelimit())
 352                         printk(KERN_ERR "Calgary: bit is off at 0x%lx "
 353                                "tbl %p dma 0x%Lx entry 0x%lx npages %u\n",
 354                                badbit, tbl, dma_addr, entry, npages);
 355         }
 356
 357         iommu_area_free(tbl->it_map, entry, npages);
 358
 359         spin_unlock_irqrestore(&tbl->it_lock, flags);
 360 }
 361
 362 static inline struct iommu_table *find_iommu_table(struct device *dev)
 363 {
 364         struct pci_dev *pdev;
 365         struct pci_bus *pbus;
 366         struct iommu_table *tbl;
 367
 368         pdev = to_pci_dev(dev);
 369
 370         pbus = pdev->bus;
 371
 372         /* is the device behind a bridge? Look for the root bus */
 373         while (pbus->parent)
 374                 pbus = pbus->parent;
 375
 376         tbl = pci_iommu(pbus);
 377
 378         BUG_ON(tbl && (tbl->it_busno != pbus->number));
 379
 380         return tbl;
 381 }
 382
 383 static void calgary_unmap_sg(struct device *dev, struct scatterlist *sglist,
 384                              int nelems,enum dma_data_direction dir,
 385                              struct dma_attrs *attrs)
 386 {
 387         struct iommu_table *tbl = find_iommu_table(dev);
 388         struct scatterlist *s;
 389         int i;
 390
 391         if (!translation_enabled(tbl))
 392                 return;
 393
 394         for_each_sg(sglist, s, nelems, i) {
 395                 unsigned int npages;
 396                 dma_addr_t dma = s->dma_address;
 397                 unsigned int dmalen = s->dma_length;
 398
 399                 if (dmalen == 0)
 400                         break;
 401
 402                 npages = iommu_num_pages(dma, dmalen, PAGE_SIZE);
 403                 iommu_free(tbl, dma, npages);
 404         }
 405 }
 406
 407 static int calgary_map_sg(struct device *dev, struct scatterlist *sg,
 408                           int nelems, enum dma_data_direction dir,
 409                           struct dma_attrs *attrs)
 410 {
 411         struct iommu_table *tbl = find_iommu_table(dev);
 412         struct scatterlist *s;
 413         unsigned long vaddr;
 414         unsigned int npages;
 415         unsigned long entry;
 416         int i;
 417
 418         for_each_sg(sg, s, nelems, i) {
 419                 BUG_ON(!sg_page(s));
 420
 421                 vaddr = (unsigned long) sg_virt(s);
 422                 npages = iommu_num_pages(vaddr, s->length, PAGE_SIZE);
 423
 424                 entry = iommu_range_alloc(dev, tbl, npages);
 425                 if (entry == bad_dma_address) {
 426                         /* makes sure unmap knows to stop */
 427                         s->dma_length = 0;
 428                         goto error;
 429                 }
 430
 431                 s->dma_address = (entry << PAGE_SHIFT) | s->offset;
 432
 433                 /* insert into HW table */
 434                 tce_build(tbl, entry, npages, vaddr & PAGE_MASK, dir);
 435
 436                 s->dma_length = s->length;
 437         }
 438
 439         return nelems;
 440 error:
 441         calgary_unmap_sg(dev, sg, nelems, dir, NULL);
 442         for_each_sg(sg, s, nelems, i) {
 443                 sg->dma_address = bad_dma_address;
 444                 sg->dma_length = 0;
 445         }
 446         return 0;
 447 }
 448
 449 static dma_addr_t calgary_map_page(struct device *dev, struct page *page,
 450                                    unsigned long offset, size_t size,
 451                                    enum dma_data_direction dir,
 452                                    struct dma_attrs *attrs)
 453 {
 454         void *vaddr = page_address(page) + offset;
 455         unsigned long uaddr;
 456         unsigned int npages;
 457         struct iommu_table *tbl = find_iommu_table(dev);
 458
 459         uaddr = (unsigned long)vaddr;
 460         npages = iommu_num_pages(uaddr, size, PAGE_SIZE);
 461
 462         return iommu_alloc(dev, tbl, vaddr, npages, dir);
 463 }
 464
 465 static void calgary_unmap_page(struct device *dev, dma_addr_t dma_addr,
 466                                size_t size, enum dma_data_direction dir,
 467                                struct dma_attrs *attrs)
 468 {
 469         struct iommu_table *tbl = find_iommu_table(dev);
 470         unsigned int npages;
 471
 472         npages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
 473         iommu_free(tbl, dma_addr, npages);
 474 }
 475
 476 static void* calgary_alloc_coherent(struct device *dev, size_t size,
 477         dma_addr_t *dma_handle, gfp_t flag)
 478 {
 479         void *ret = NULL;
 480         dma_addr_t mapping;
 481         unsigned int npages, order;
 482         struct iommu_table *tbl = find_iommu_table(dev);
 483
 484         size = PAGE_ALIGN(size); /* size rounded up to full pages */
 485         npages = size >> PAGE_SHIFT;
 486         order = get_order(size);
 487
 488         flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
 489
 490         /* alloc enough pages (and possibly more) */
 491         ret = (void *)__get_free_pages(flag, order);
 492         if (!ret)
 493                 goto error;
 494         memset(ret, 0, size);
 495
 496         /* set up tces to cover the allocated range */
 497         mapping = iommu_alloc(dev, tbl, ret, npages, DMA_BIDIRECTIONAL);
 498         if (mapping == bad_dma_address)
 499                 goto free;
 500         *dma_handle = mapping;
 501         return ret;
 502 free:
 503         free_pages((unsigned long)ret, get_order(size));
 504         ret = NULL;
 505 error:
 506         return ret;
 507 }
 508
 509 static void calgary_free_coherent(struct device *dev, size_t size,
 510                                   void *vaddr, dma_addr_t dma_handle)
 511 {
 512         unsigned int npages;
 513         struct iommu_table *tbl = find_iommu_table(dev);
 514
 515         size = PAGE_ALIGN(size);
 516         npages = size >> PAGE_SHIFT;
 517
 518         iommu_free(tbl, dma_handle, npages);
 519         free_pages((unsigned long)vaddr, get_order(size));
 520 }
 521
 522 static struct dma_map_ops calgary_dma_ops = {
 523         .alloc_coherent = calgary_alloc_coherent,
 524         .free_coherent = calgary_free_coherent,
 525         .map_sg = calgary_map_sg,
 526         .unmap_sg = calgary_unmap_sg,
 527         .map_page = calgary_map_page,
 528         .unmap_page = calgary_unmap_page,
 529 };
 530
 531 static inline void __iomem * busno_to_bbar(unsigned char num)
 532 {
 533         return bus_info[num].bbar;
 534 }
 535
 536 static inline int busno_to_phbid(unsigned char num)
 537 {
 538         return bus_info[num].phbid;
 539 }
 540
 541 static inline unsigned long split_queue_offset(unsigned char num)
 542 {
 543         size_t idx = busno_to_phbid(num);
 544
 545         return split_queue_offsets[idx];
 546 }
 547
 548 static inline unsigned long tar_offset(unsigned char num)
 549 {
 550         size_t idx = busno_to_phbid(num);
 551
 552         return tar_offsets[idx];
 553 }
 554
 555 static inline unsigned long phb_offset(unsigned char num)
 556 {
 557         size_t idx = busno_to_phbid(num);
 558
 559         return phb_offsets[idx];
 560 }
 561
 562 static inline void __iomem* calgary_reg(void __iomem *bar, unsigned long offset)
 563 {
 564         unsigned long target = ((unsigned long)bar) | offset;
 565         return (void __iomem*)target;
 566 }
 567
 568 static inline int is_calioc2(unsigned short device)
 569 {
 570         return (device == PCI_DEVICE_ID_IBM_CALIOC2);
 571 }
 572
 573 static inline int is_calgary(unsigned short device)
 574 {
 575         return (device == PCI_DEVICE_ID_IBM_CALGARY);
 576 }
 577
 578 static inline int is_cal_pci_dev(unsigned short device)
 579 {
 580         return (is_calgary(device) || is_calioc2(device));
 581 }
 582
 583 static void calgary_tce_cache_blast(struct iommu_table *tbl)
 584 {
 585         u64 val;
 586         u32 aer;
 587         int i = 0;
 588         void __iomem *bbar = tbl->bbar;
 589         void __iomem *target;
 590
 591         /* disable arbitration on the bus */
 592         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_AER_OFFSET);
 593         aer = readl(target);
 594         writel(0, target);
 595
 596         /* read plssr to ensure it got there */
 597         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_PLSSR_OFFSET);
 598         val = readl(target);
 599
 600         /* poll split queues until all DMA activity is done */
 601         target = calgary_reg(bbar, split_queue_offset(tbl->it_busno));
 602         do {
 603                 val = readq(target);
 604                 i++;
 605         } while ((val & 0xff) != 0xff && i < 100);
 606         if (i == 100)
 607                 printk(KERN_WARNING "Calgary: PCI bus not quiesced, "
 608                        "continuing anyway\n");
 609
 610         /* invalidate TCE cache */
 611         target = calgary_reg(bbar, tar_offset(tbl->it_busno));
 612         writeq(tbl->tar_val, target);
 613
 614         /* enable arbitration */
 615         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_AER_OFFSET);
 616         writel(aer, target);
 617         (void)readl(target); /* flush */
 618 }
 619
 620 static void calioc2_tce_cache_blast(struct iommu_table *tbl)
 621 {
 622         void __iomem *bbar = tbl->bbar;
 623         void __iomem *target;
 624         u64 val64;
 625         u32 val;
 626         int i = 0;
 627         int count = 1;
 628         unsigned char bus = tbl->it_busno;
 629
 630 begin:
 631         printk(KERN_DEBUG "Calgary: CalIOC2 bus 0x%x entering tce cache blast "
 632                "sequence - count %d\n", bus, count);
 633
 634         /* 1. using the Page Migration Control reg set SoftStop */
 635         target = calgary_reg(bbar, phb_offset(bus) | PHB_PAGE_MIG_CTRL);
 636         val = be32_to_cpu(readl(target));
 637         printk(KERN_DEBUG "1a. read 0x%x [LE] from %p\n", val, target);
 638         val |= PMR_SOFTSTOP;
 639         printk(KERN_DEBUG "1b. writing 0x%x [LE] to %p\n", val, target);
 640         writel(cpu_to_be32(val), target);
 641
 642         /* 2. poll split queues until all DMA activity is done */
 643         printk(KERN_DEBUG "2a. starting to poll split queues\n");
 644         target = calgary_reg(bbar, split_queue_offset(bus));
 645         do {
 646                 val64 = readq(target);
 647                 i++;
 648         } while ((val64 & 0xff) != 0xff && i < 100);
 649         if (i == 100)
 650                 printk(KERN_WARNING "CalIOC2: PCI bus not quiesced, "
 651                        "continuing anyway\n");
 652
 653         /* 3. poll Page Migration DEBUG for SoftStopFault */
 654         target = calgary_reg(bbar, phb_offset(bus) | PHB_PAGE_MIG_DEBUG);
 655         val = be32_to_cpu(readl(target));
 656         printk(KERN_DEBUG "3. read 0x%x [LE] from %p\n", val, target);
 657
 658         /* 4. if SoftStopFault - goto (1) */
 659         if (val & PMR_SOFTSTOPFAULT) {
 660                 if (++count < 100)
 661                         goto begin;
 662                 else {
 663                         printk(KERN_WARNING "CalIOC2: too many SoftStopFaults, "
 664                                "aborting TCE cache flush sequence!\n");
 665                         return; /* pray for the best */
 666                 }
 667         }
 668
 669         /* 5. Slam into HardStop by reading PHB_PAGE_MIG_CTRL */
 670         target = calgary_reg(bbar, phb_offset(bus) | PHB_PAGE_MIG_CTRL);
 671         printk(KERN_DEBUG "5a. slamming into HardStop by reading %p\n", target);
 672         val = be32_to_cpu(readl(target));
 673         printk(KERN_DEBUG "5b. read 0x%x [LE] from %p\n", val, target);
 674         target = calgary_reg(bbar, phb_offset(bus) | PHB_PAGE_MIG_DEBUG);
 675         val = be32_to_cpu(readl(target));
 676         printk(KERN_DEBUG "5c. read 0x%x [LE] from %p (debug)\n", val, target);
 677
 678         /* 6. invalidate TCE cache */
 679         printk(KERN_DEBUG "6. invalidating TCE cache\n");
 680         target = calgary_reg(bbar, tar_offset(bus));
 681         writeq(tbl->tar_val, target);
 682
 683         /* 7. Re-read PMCR */
 684         printk(KERN_DEBUG "7a. Re-reading PMCR\n");
 685         target = calgary_reg(bbar, phb_offset(bus) | PHB_PAGE_MIG_CTRL);
 686         val = be32_to_cpu(readl(target));
 687         printk(KERN_DEBUG "7b. read 0x%x [LE] from %p\n", val, target);
 688
 689         /* 8. Remove HardStop */
 690         printk(KERN_DEBUG "8a. removing HardStop from PMCR\n");
 691         target = calgary_reg(bbar, phb_offset(bus) | PHB_PAGE_MIG_CTRL);
 692         val = 0;
 693         printk(KERN_DEBUG "8b. writing 0x%x [LE] to %p\n", val, target);
 694         writel(cpu_to_be32(val), target);
 695         val = be32_to_cpu(readl(target));
 696         printk(KERN_DEBUG "8c. read 0x%x [LE] from %p\n", val, target);
 697 }
 698
 699 static void __init calgary_reserve_mem_region(struct pci_dev *dev, u64 start,
 700         u64 limit)
 701 {
 702         unsigned int numpages;
 703
 704         limit = limit | 0xfffff;
 705         limit++;
 706
 707         numpages = ((limit - start) >> PAGE_SHIFT);
 708         iommu_range_reserve(pci_iommu(dev->bus), start, numpages);
 709 }
 710
 711 static void __init calgary_reserve_peripheral_mem_1(struct pci_dev *dev)
 712 {
 713         void __iomem *target;
 714         u64 low, high, sizelow;
 715         u64 start, limit;
 716         struct iommu_table *tbl = pci_iommu(dev->bus);
 717         unsigned char busnum = dev->bus->number;
 718         void __iomem *bbar = tbl->bbar;
 719
 720         /* peripheral MEM_1 region */
 721         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_1_LOW);
 722         low = be32_to_cpu(readl(target));
 723         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_1_HIGH);
 724         high = be32_to_cpu(readl(target));
 725         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_1_SIZE);
 726         sizelow = be32_to_cpu(readl(target));
 727
 728         start = (high << 32) | low;
 729         limit = sizelow;
 730
 731         calgary_reserve_mem_region(dev, start, limit);
 732 }
 733
 734 static void __init calgary_reserve_peripheral_mem_2(struct pci_dev *dev)
 735 {
 736         void __iomem *target;
 737         u32 val32;
 738         u64 low, high, sizelow, sizehigh;
 739         u64 start, limit;
 740         struct iommu_table *tbl = pci_iommu(dev->bus);
 741         unsigned char busnum = dev->bus->number;
 742         void __iomem *bbar = tbl->bbar;
 743
 744         /* is it enabled? */
 745         target = calgary_reg(bbar, phb_offset(busnum) | PHB_CONFIG_RW_OFFSET);
 746         val32 = be32_to_cpu(readl(target));
 747         if (!(val32 & PHB_MEM2_ENABLE))
 748                 return;
 749
 750         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_2_LOW);
 751         low = be32_to_cpu(readl(target));
 752         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_2_HIGH);
 753         high = be32_to_cpu(readl(target));
 754         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_2_SIZE_LOW);
 755         sizelow = be32_to_cpu(readl(target));
 756         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_2_SIZE_HIGH);
 757         sizehigh = be32_to_cpu(readl(target));
 758
 759         start = (high << 32) | low;
 760         limit = (sizehigh << 32) | sizelow;
 761
 762         calgary_reserve_mem_region(dev, start, limit);
 763 }
 764
 765 /*
 766  * some regions of the IO address space do not get translated, so we
 767  * must not give devices IO addresses in those regions. The regions
 768  * are the 640KB-1MB region and the two PCI peripheral memory holes.
 769  * Reserve all of them in the IOMMU bitmap to avoid giving them out
 770  * later.
 771  */
 772 static void __init calgary_reserve_regions(struct pci_dev *dev)
 773 {
 774         unsigned int npages;
 775         u64 start;
 776         struct iommu_table *tbl = pci_iommu(dev->bus);
 777
 778         /* reserve EMERGENCY_PAGES from bad_dma_address and up */
 779         iommu_range_reserve(tbl, bad_dma_address, EMERGENCY_PAGES);
 780
 781         /* avoid the BIOS/VGA first 640KB-1MB region */
 782         /* for CalIOC2 - avoid the entire first MB */
 783         if (is_calgary(dev->device)) {
 784                 start = (640 * 1024);
 785                 npages = ((1024 - 640) * 1024) >> PAGE_SHIFT;
 786         } else { /* calioc2 */
 787                 start = 0;
 788                 npages = (1 * 1024 * 1024) >> PAGE_SHIFT;
 789         }
 790         iommu_range_reserve(tbl, start, npages);
 791
 792         /* reserve the two PCI peripheral memory regions in IO space */
 793         calgary_reserve_peripheral_mem_1(dev);
 794         calgary_reserve_peripheral_mem_2(dev);
 795 }
 796
 797 static int __init calgary_setup_tar(struct pci_dev *dev, void __iomem *bbar)
 798 {
 799         u64 val64;
 800         u64 table_phys;
 801         void __iomem *target;
 802         int ret;
 803         struct iommu_table *tbl;
 804
 805         /* build TCE tables for each PHB */
 806         ret = build_tce_table(dev, bbar);
 807         if (ret)
 808                 return ret;
 809
 810         tbl = pci_iommu(dev->bus);
 811         tbl->it_base = (unsigned long)bus_info[dev->bus->number].tce_space;
 812
 813         if (is_kdump_kernel())
 814                 calgary_init_bitmap_from_tce_table(tbl);
 815         else
 816                 tce_free(tbl, 0, tbl->it_size);
 817
 818         if (is_calgary(dev->device))
 819                 tbl->chip_ops = &calgary_chip_ops;
 820         else if (is_calioc2(dev->device))
 821                 tbl->chip_ops = &calioc2_chip_ops;
 822         else
 823                 BUG();
 824
 825         calgary_reserve_regions(dev);
 826
 827         /* set TARs for each PHB */
 828         target = calgary_reg(bbar, tar_offset(dev->bus->number));
 829         val64 = be64_to_cpu(readq(target));
 830
 831         /* zero out all TAR bits under sw control */
 832         val64 &= ~TAR_SW_BITS;
 833         table_phys = (u64)__pa(tbl->it_base);
 834
 835         val64 |= table_phys;
 836
 837         BUG_ON(specified_table_size > TCE_TABLE_SIZE_8M);
 838         val64 |= (u64) specified_table_size;
 839
 840         tbl->tar_val = cpu_to_be64(val64);
 841
 842         writeq(tbl->tar_val, target);
 843         readq(target); /* flush */
 844
 845         return 0;
 846 }
 847
 848 static void __init calgary_free_bus(struct pci_dev *dev)
 849 {
 850         u64 val64;
 851         struct iommu_table *tbl = pci_iommu(dev->bus);
 852         void __iomem *target;
 853         unsigned int bitmapsz;
 854
 855         target = calgary_reg(tbl->bbar, tar_offset(dev->bus->number));
 856         val64 = be64_to_cpu(readq(target));
 857         val64 &= ~TAR_SW_BITS;
 858         writeq(cpu_to_be64(val64), target);
 859         readq(target); /* flush */
 860
 861         bitmapsz = tbl->it_size / BITS_PER_BYTE;
 862         free_pages((unsigned long)tbl->it_map, get_order(bitmapsz));
 863         tbl->it_map = NULL;
 864
 865         kfree(tbl);
 866
 867         set_pci_iommu(dev->bus, NULL);
 868
 869         /* Can't free bootmem allocated memory after system is up :-( */
 870         bus_info[dev->bus->number].tce_space = NULL;
 871 }
 872
 873 static void calgary_dump_error_regs(struct iommu_table *tbl)
 874 {
 875         void __iomem *bbar = tbl->bbar;
 876         void __iomem *target;
 877         u32 csr, plssr;
 878
 879         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_CSR_OFFSET);
 880         csr = be32_to_cpu(readl(target));
 881
 882         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_PLSSR_OFFSET);
 883         plssr = be32_to_cpu(readl(target));
 884
 885         /* If no error, the agent ID in the CSR is not valid */
 886         printk(KERN_EMERG "Calgary: DMA error on Calgary PHB 0x%x, "
 887                "0x%08x@CSR 0x%08x@PLSSR\n", tbl->it_busno, csr, plssr);
 888 }
 889
 890 static void calioc2_dump_error_regs(struct iommu_table *tbl)
 891 {
 892         void __iomem *bbar = tbl->bbar;
 893         u32 csr, csmr, plssr, mck, rcstat;
 894         void __iomem *target;
 895         unsigned long phboff = phb_offset(tbl->it_busno);
 896         unsigned long erroff;
 897         u32 errregs[7];
 898         int i;
 899
 900         /* dump CSR */
 901         target = calgary_reg(bbar, phboff | PHB_CSR_OFFSET);
 902         csr = be32_to_cpu(readl(target));
 903         /* dump PLSSR */
 904         target = calgary_reg(bbar, phboff | PHB_PLSSR_OFFSET);
 905         plssr = be32_to_cpu(readl(target));
 906         /* dump CSMR */
 907         target = calgary_reg(bbar, phboff | 0x290);
 908         csmr = be32_to_cpu(readl(target));
 909         /* dump mck */
 910         target = calgary_reg(bbar, phboff | 0x800);
 911         mck = be32_to_cpu(readl(target));
 912
 913         printk(KERN_EMERG "Calgary: DMA error on CalIOC2 PHB 0x%x\n",
 914                tbl->it_busno);
 915
 916         printk(KERN_EMERG "Calgary: 0x%08x@CSR 0x%08x@PLSSR 0x%08x@CSMR 0x%08x@MCK\n",
 917                csr, plssr, csmr, mck);
 918
 919         /* dump rest of error regs */
 920         printk(KERN_EMERG "Calgary: ");
 921         for (i = 0; i < ARRAY_SIZE(errregs); i++) {
 922                 /* err regs are at 0x810 - 0x870 */
 923                 erroff = (0x810 + (i * 0x10));
 924                 target = calgary_reg(bbar, phboff | erroff);
 925                 errregs[i] = be32_to_cpu(readl(target));
 926                 printk("0x%08x@0x%lx ", errregs[i], erroff);
 927         }
 928         printk("\n");
 929
 930         /* root complex status */
 931         target = calgary_reg(bbar, phboff | PHB_ROOT_COMPLEX_STATUS);
 932         rcstat = be32_to_cpu(readl(target));
 933         printk(KERN_EMERG "Calgary: 0x%08x@0x%x\n", rcstat,
 934                PHB_ROOT_COMPLEX_STATUS);
 935 }
 936
 937 static void calgary_watchdog(unsigned long data)
 938 {
 939         struct pci_dev *dev = (struct pci_dev *)data;
 940         struct iommu_table *tbl = pci_iommu(dev->bus);
 941         void __iomem *bbar = tbl->bbar;
 942         u32 val32;
 943         void __iomem *target;
 944
 945         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_CSR_OFFSET);
 946         val32 = be32_to_cpu(readl(target));
 947
 948         /* If no error, the agent ID in the CSR is not valid */
 949         if (val32 & CSR_AGENT_MASK) {
 950                 tbl->chip_ops->dump_error_regs(tbl);
 951
 952                 /* reset error */
 953                 writel(0, target);
 954
 955                 /* Disable bus that caused the error */
 956                 target = calgary_reg(bbar, phb_offset(tbl->it_busno) |
 957                                      PHB_CONFIG_RW_OFFSET);
 958                 val32 = be32_to_cpu(readl(target));
 959                 val32 |= PHB_SLOT_DISABLE;
 960                 writel(cpu_to_be32(val32), target);
 961                 readl(target); /* flush */
 962         } else {
 963                 /* Reset the timer */
 964                 mod_timer(&tbl->watchdog_timer, jiffies + 2 * HZ);
 965         }
 966 }
 967
 968 static void __init calgary_set_split_completion_timeout(void __iomem *bbar,
 969         unsigned char busnum, unsigned long timeout)
 970 {
 971         u64 val64;
 972         void __iomem *target;
 973         unsigned int phb_shift = ~0; /* silence gcc */
 974         u64 mask;
 975
 976         switch (busno_to_phbid(busnum)) {
 977         case 0: phb_shift = (63 - 19);
 978                 break;
 979         case 1: phb_shift = (63 - 23);
 980                 break;
 981         case 2: phb_shift = (63 - 27);
 982                 break;
 983         case 3: phb_shift = (63 - 35);
 984                 break;
 985         default:
 986                 BUG_ON(busno_to_phbid(busnum));
 987         }
 988
 989         target = calgary_reg(bbar, CALGARY_CONFIG_REG);
 990         val64 = be64_to_cpu(readq(target));
 991
 992         /* zero out this PHB's timer bits */
 993         mask = ~(0xFUL << phb_shift);
 994         val64 &= mask;
 995         val64 |= (timeout << phb_shift);
 996         writeq(cpu_to_be64(val64), target);
 997         readq(target); /* flush */
 998 }
 999
1000 static void __init calioc2_handle_quirks(struct iommu_table *tbl, struct pci_dev *dev)
1001 {
1002         unsigned char busnum = dev->bus->number;
1003         void __iomem *bbar = tbl->bbar;
1004         void __iomem *target;
1005         u32 val;
1006
1007         /*
1008          * CalIOC2 designers recommend setting bit 8 in 0xnDB0 to 1
1009          */
1010         target = calgary_reg(bbar, phb_offset(busnum) | PHB_SAVIOR_L2);
1011         val = cpu_to_be32(readl(target));
1012         val |= 0x00800000;
1013         writel(cpu_to_be32(val), target);
1014 }
1015
1016 static void __init calgary_handle_quirks(struct iommu_table *tbl, struct pci_dev *dev)
1017 {
1018         unsigned char busnum = dev->bus->number;
1019
1020         /*
1021          * Give split completion a longer timeout on bus 1 for aic94xx
1022          * http://bugzilla.kernel.org/show_bug.cgi?id=7180
1023          */
1024         if (is_calgary(dev->device) && (busnum == 1))
1025                 calgary_set_split_completion_timeout(tbl->bbar, busnum,
1026                                                      CCR_2SEC_TIMEOUT);
1027 }
1028
1029 static void __init calgary_enable_translation(struct pci_dev *dev)
1030 {
1031         u32 val32;
1032         unsigned char busnum;
1033         void __iomem *target;
1034         void __iomem *bbar;
1035         struct iommu_table *tbl;
1036
1037         busnum = dev->bus->number;
1038         tbl = pci_iommu(dev->bus);
1039         bbar = tbl->bbar;
1040
1041         /* enable TCE in PHB Config Register */
1042         target = calgary_reg(bbar, phb_offset(busnum) | PHB_CONFIG_RW_OFFSET);
1043         val32 = be32_to_cpu(readl(target));
1044         val32 |= PHB_TCE_ENABLE | PHB_DAC_DISABLE | PHB_MCSR_ENABLE;
1045
1046         printk(KERN_INFO "Calgary: enabling translation on %s PHB %#x\n",
1047                (dev->device == PCI_DEVICE_ID_IBM_CALGARY) ?
1048                "Calgary" : "CalIOC2", busnum);
1049         printk(KERN_INFO "Calgary: errant DMAs will now be prevented on this "
1050                "bus.\n");
1051
1052         writel(cpu_to_be32(val32), target);
1053         readl(target); /* flush */
1054
1055         init_timer(&tbl->watchdog_timer);
1056         tbl->watchdog_timer.function = &calgary_watchdog;
1057         tbl->watchdog_timer.data = (unsigned long)dev;
1058         mod_timer(&tbl->watchdog_timer, jiffies);
1059 }
1060
1061 static void __init calgary_disable_translation(struct pci_dev *dev)
1062 {
1063         u32 val32;
1064         unsigned char busnum;
1065         void __iomem *target;
1066         void __iomem *bbar;
1067         struct iommu_table *tbl;
1068
1069         busnum = dev->bus->number;
1070         tbl = pci_iommu(dev->bus);
1071         bbar = tbl->bbar;
1072
1073         /* disable TCE in PHB Config Register */
1074         target = calgary_reg(bbar, phb_offset(busnum) | PHB_CONFIG_RW_OFFSET);
1075         val32 = be32_to_cpu(readl(target));
1076         val32 &= ~(PHB_TCE_ENABLE | PHB_DAC_DISABLE | PHB_MCSR_ENABLE);
1077
1078         printk(KERN_INFO "Calgary: disabling translation on PHB %#x!\n", busnum);
1079         writel(cpu_to_be32(val32), target);
1080         readl(target); /* flush */
1081
1082         del_timer_sync(&tbl->watchdog_timer);
1083 }
1084
1085 static void __init calgary_init_one_nontraslated(struct pci_dev *dev)
1086 {
1087         pci_dev_get(dev);
1088         set_pci_iommu(dev->bus, NULL);
1089
1090         /* is the device behind a bridge? */
1091         if (dev->bus->parent)
1092                 dev->bus->parent->self = dev;
1093         else
1094                 dev->bus->self = dev;
1095 }
1096
1097 static int __init calgary_init_one(struct pci_dev *dev)
1098 {
1099         void __iomem *bbar;
1100         struct iommu_table *tbl;
1101         int ret;
1102
1103         BUG_ON(dev->bus->number >= MAX_PHB_BUS_NUM);
1104
1105         bbar = busno_to_bbar(dev->bus->number);
1106         ret = calgary_setup_tar(dev, bbar);
1107         if (ret)
1108                 goto done;
1109
1110         pci_dev_get(dev);
1111
1112         if (dev->bus->parent) {
1113                 if (dev->bus->parent->self)
1114                         printk(KERN_WARNING "Calgary: IEEEE, dev %p has "
1115                                "bus->parent->self!\n", dev);
1116                 dev->bus->parent->self = dev;
1117         } else
1118                 dev->bus->self = dev;
1119
1120         tbl = pci_iommu(dev->bus);
1121         tbl->chip_ops->handle_quirks(tbl, dev);
1122
1123         calgary_enable_translation(dev);
1124
1125         return 0;
1126
1127 done:
1128         return ret;
1129 }
1130
1131 static int __init calgary_locate_bbars(void)
1132 {
1133         int ret;
1134         int rioidx, phb, bus;
1135         void __iomem *bbar;
1136         void __iomem *target;
1137         unsigned long offset;
1138         u8 start_bus, end_bus;
1139         u32 val;
1140
1141         ret = -ENODATA;
1142         for (rioidx = 0; rioidx < rio_table_hdr->num_rio_dev; rioidx++) {
1143                 struct rio_detail *rio = rio_devs[rioidx];
1144
1145                 if ((rio->type != COMPAT_CALGARY) && (rio->type != ALT_CALGARY))
1146                         continue;
1147
1148                 /* map entire 1MB of Calgary config space */
1149                 bbar = ioremap_nocache(rio->BBAR, 1024 * 1024);
1150                 if (!bbar)
1151                         goto error;
1152
1153                 for (phb = 0; phb < PHBS_PER_CALGARY; phb++) {
1154                         offset = phb_debug_offsets[phb] | PHB_DEBUG_STUFF_OFFSET;
1155                         target = calgary_reg(bbar, offset);
1156
1157                         val = be32_to_cpu(readl(target));
1158
1159                         start_bus = (u8)((val & 0x00FF0000) >> 16);
1160                         end_bus = (u8)((val & 0x0000FF00) >> 8);
1161
1162                         if (end_bus) {
1163                                 for (bus = start_bus; bus <= end_bus; bus++) {
1164                                         bus_info[bus].bbar = bbar;
1165                                         bus_info[bus].phbid = phb;
1166                                 }
1167                         } else {
1168                                 bus_info[start_bus].bbar = bbar;
1169                                 bus_info[start_bus].phbid = phb;
1170                         }
1171                 }
1172         }
1173
1174         return 0;
1175
1176 error:
1177         /* scan bus_info and iounmap any bbars we previously ioremap'd */
1178         for (bus = 0; bus < ARRAY_SIZE(bus_info); bus++)
1179                 if (bus_info[bus].bbar)
1180                         iounmap(bus_info[bus].bbar);
1181
1182         return ret;
1183 }
1184
1185 static int __init calgary_init(void)
1186 {
1187         int ret;
1188         struct pci_dev *dev = NULL;
1189         struct calgary_bus_info *info;
1190
1191         ret = calgary_locate_bbars();
1192         if (ret)
1193                 return ret;
1194
1195         /* Purely for kdump kernel case */
1196         if (is_kdump_kernel())
1197                 get_tce_space_from_tar();
1198
1199         do {
1200                 dev = pci_get_device(PCI_VENDOR_ID_IBM, PCI_ANY_ID, dev);
1201                 if (!dev)
1202                         break;
1203                 if (!is_cal_pci_dev(dev->device))
1204                         continue;
1205
1206                 info = &bus_info[dev->bus->number];
1207                 if (info->translation_disabled) {
1208                         calgary_init_one_nontraslated(dev);
1209                         continue;
1210                 }
1211
1212                 if (!info->tce_space && !translate_empty_slots)
1213                         continue;
1214
1215                 ret = calgary_init_one(dev);
1216                 if (ret)
1217                         goto error;
1218         } while (1);
1219
1220         dev = NULL;
1221         for_each_pci_dev(dev) {
1222                 struct iommu_table *tbl;
1223
1224                 tbl = find_iommu_table(&dev->dev);
1225
1226                 if (translation_enabled(tbl))
1227                         dev->dev.archdata.dma_ops = &calgary_dma_ops;
1228         }
1229
1230         return ret;
1231
1232 error:
1233         do {
1234                 dev = pci_get_device(PCI_VENDOR_ID_IBM, PCI_ANY_ID, dev);
1235                 if (!dev)
1236                         break;
1237                 if (!is_cal_pci_dev(dev->device))
1238                         continue;
1239
1240                 info = &bus_info[dev->bus->number];
1241                 if (info->translation_disabled) {
1242                         pci_dev_put(dev);
1243                         continue;
1244                 }
1245                 if (!info->tce_space && !translate_empty_slots)
1246                         continue;
1247
1248                 calgary_disable_translation(dev);
1249                 calgary_free_bus(dev);
1250                 pci_dev_put(dev); /* Undo calgary_init_one()'s pci_dev_get() */
1251                 dev->dev.archdata.dma_ops = NULL;
1252         } while (1);
1253
1254         return ret;
1255 }
1256
1257 static inline int __init determine_tce_table_size(u64 ram)
1258 {
1259         int ret;
1260
1261         if (specified_table_size != TCE_TABLE_SIZE_UNSPECIFIED)
1262                 return specified_table_size;
1263
1264         /*
1265          * Table sizes are from 0 to 7 (TCE_TABLE_SIZE_64K to
1266          * TCE_TABLE_SIZE_8M). Table size 0 has 8K entries and each
1267          * larger table size has twice as many entries, so shift the
1268          * max ram address by 13 to divide by 8K and then look at the
1269          * order of the result to choose between 0-7.
1270          */
1271         ret = get_order(ram >> 13);
1272         if (ret > TCE_TABLE_SIZE_8M)
1273                 ret = TCE_TABLE_SIZE_8M;
1274
1275         return ret;
1276 }
1277
1278 static int __init build_detail_arrays(void)
1279 {
1280         unsigned long ptr;
1281         unsigned numnodes, i;
1282         int scal_detail_size, rio_detail_size;
1283
1284         numnodes = rio_table_hdr->num_scal_dev;
1285         if (numnodes > MAX_NUMNODES){
1286                 printk(KERN_WARNING
1287                         "Calgary: MAX_NUMNODES too low! Defined as %d, "
1288                         "but system has %d nodes.\n",
1289                         MAX_NUMNODES, numnodes);
1290                 return -ENODEV;
1291         }
1292
1293         switch (rio_table_hdr->version){
1294         case 2:
1295                 scal_detail_size = 11;
1296                 rio_detail_size = 13;
1297                 break;
1298         case 3:
1299                 scal_detail_size = 12;
1300                 rio_detail_size = 15;
1301                 break;
1302         default:
1303                 printk(KERN_WARNING
1304                        "Calgary: Invalid Rio Grande Table Version: %d\n",
1305                        rio_table_hdr->version);
1306                 return -EPROTO;
1307         }
1308
1309         ptr = ((unsigned long)rio_table_hdr) + 3;
1310         for (i = 0; i < numnodes; 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 /*
1342  * calgary_init_bitmap_from_tce_table():
1343  * Funtion for kdump case. In the second/kdump kernel initialize
1344  * the bitmap based on the tce table entries obtained from first kernel
1345  */
1346 static void calgary_init_bitmap_from_tce_table(struct iommu_table *tbl)
1347 {
1348         u64 *tp;
1349         unsigned int index;
1350         tp = ((u64 *)tbl->it_base);
1351         for (index = 0 ; index < tbl->it_size; index++) {
1352                 if (*tp != 0x0)
1353                         set_bit(index, tbl->it_map);
1354                 tp++;
1355         }
1356 }
1357
1358 /*
1359  * get_tce_space_from_tar():
1360  * Function for kdump case. Get the tce tables from first kernel
1361  * by reading the contents of the base adress register of calgary iommu
1362  */
1363 static void __init get_tce_space_from_tar(void)
1364 {
1365         int bus;
1366         void __iomem *target;
1367         unsigned long tce_space;
1368
1369         for (bus = 0; bus < MAX_PHB_BUS_NUM; bus++) {
1370                 struct calgary_bus_info *info = &bus_info[bus];
1371                 unsigned short pci_device;
1372                 u32 val;
1373
1374                 val = read_pci_config(bus, 0, 0, 0);
1375                 pci_device = (val & 0xFFFF0000) >> 16;
1376
1377                 if (!is_cal_pci_dev(pci_device))
1378                         continue;
1379                 if (info->translation_disabled)
1380                         continue;
1381
1382                 if (calgary_bus_has_devices(bus, pci_device) ||
1383                                                 translate_empty_slots) {
1384                         target = calgary_reg(bus_info[bus].bbar,
1385                                                 tar_offset(bus));
1386                         tce_space = be64_to_cpu(readq(target));
1387                         tce_space = tce_space & TAR_SW_BITS;
1388
1389                         tce_space = tce_space & (~specified_table_size);
1390                         info->tce_space = (u64 *)__va(tce_space);
1391                 }
1392         }
1393         return;
1394 }
1395
1396 void __init detect_calgary(void)
1397 {
1398         int bus;
1399         void *tbl;
1400         int calgary_found = 0;
1401         unsigned long ptr;
1402         unsigned int offset, prev_offset;
1403         int ret;
1404
1405         /*
1406          * if the user specified iommu=off or iommu=soft or we found
1407          * another HW IOMMU already, bail out.
1408          */
1409         if (swiotlb || no_iommu || iommu_detected)
1410                 return;
1411
1412         if (!use_calgary)
1413                 return;
1414
1415         if (!early_pci_allowed())
1416                 return;
1417
1418         printk(KERN_DEBUG "Calgary: detecting Calgary via BIOS EBDA area\n");
1419
1420         ptr = (unsigned long)phys_to_virt(get_bios_ebda());
1421
1422         rio_table_hdr = NULL;
1423         prev_offset = 0;
1424         offset = 0x180;
1425         /*
1426          * The next offset is stored in the 1st word.
1427          * Only parse up until the offset increases:
1428          */
1429         while (offset > prev_offset) {
1430                 /* The block id is stored in the 2nd word */
1431                 if (*((unsigned short *)(ptr + offset + 2)) == 0x4752){
1432                         /* set the pointer past the offset & block id */
1433                         rio_table_hdr = (struct rio_table_hdr *)(ptr + offset + 4);
1434                         break;
1435                 }
1436                 prev_offset = offset;
1437                 offset = *((unsigned short *)(ptr + offset));
1438         }
1439         if (!rio_table_hdr) {
1440                 printk(KERN_DEBUG "Calgary: Unable to locate Rio Grande table "
1441                        "in EBDA - bailing!\n");
1442                 return;
1443         }
1444
1445         ret = build_detail_arrays();
1446         if (ret) {
1447                 printk(KERN_DEBUG "Calgary: build_detail_arrays ret %d\n", ret);
1448                 return;
1449         }
1450
1451         specified_table_size = determine_tce_table_size((is_kdump_kernel() ?
1452                                         saved_max_pfn : max_pfn) * PAGE_SIZE);
1453
1454         for (bus = 0; bus < MAX_PHB_BUS_NUM; bus++) {
1455                 struct calgary_bus_info *info = &bus_info[bus];
1456                 unsigned short pci_device;
1457                 u32 val;
1458
1459                 val = read_pci_config(bus, 0, 0, 0);
1460                 pci_device = (val & 0xFFFF0000) >> 16;
1461
1462                 if (!is_cal_pci_dev(pci_device))
1463                         continue;
1464
1465                 if (info->translation_disabled)
1466                         continue;
1467
1468                 if (calgary_bus_has_devices(bus, pci_device) ||
1469                     translate_empty_slots) {
1470                         /*
1471                          * If it is kdump kernel, find and use tce tables
1472                          * from first kernel, else allocate tce tables here
1473                          */
1474                         if (!is_kdump_kernel()) {
1475                                 tbl = alloc_tce_table();
1476                                 if (!tbl)
1477                                         goto cleanup;
1478                                 info->tce_space = tbl;
1479                         }
1480                         calgary_found = 1;
1481                 }
1482         }
1483
1484         printk(KERN_DEBUG "Calgary: finished detection, Calgary %s\n",
1485                calgary_found ? "found" : "not found");
1486
1487         if (calgary_found) {
1488                 iommu_detected = 1;
1489                 calgary_detected = 1;
1490                 printk(KERN_INFO "PCI-DMA: Calgary IOMMU detected.\n");
1491                 printk(KERN_INFO "PCI-DMA: Calgary TCE table spec is %d, "
1492                        "CONFIG_IOMMU_DEBUG is %s.\n", specified_table_size,
1493                        debugging ? "enabled" : "disabled");
1494
1495                 /* swiotlb for devices that aren't behind the Calgary. */
1496                 if (max_pfn > MAX_DMA32_PFN)
1497                         swiotlb = 1;
1498         }
1499         return;
1500
1501 cleanup:
1502         for (--bus; bus >= 0; --bus) {
1503                 struct calgary_bus_info *info = &bus_info[bus];
1504
1505                 if (info->tce_space)
1506                         free_tce_table(info->tce_space);
1507         }
1508 }
1509
1510 int __init calgary_iommu_init(void)
1511 {
1512         int ret;
1513
1514         if (no_iommu || (swiotlb && !calgary_detected))
1515                 return -ENODEV;
1516
1517         if (!calgary_detected)
1518                 return -ENODEV;
1519
1520         /* ok, we're trying to use Calgary - let's roll */
1521         printk(KERN_INFO "PCI-DMA: Using Calgary IOMMU\n");
1522
1523         ret = calgary_init();
1524         if (ret) {
1525                 printk(KERN_ERR "PCI-DMA: Calgary init failed %d, "
1526                        "falling back to no_iommu\n", ret);
1527                 return ret;
1528         }
1529
1530         force_iommu = 1;
1531         bad_dma_address = 0x0;
1532         /* dma_ops is set to swiotlb or nommu */
1533         if (!dma_ops)
1534                 dma_ops = &nommu_dma_ops;
1535
1536         return 0;
1537 }
1538
1539 static int __init calgary_parse_options(char *p)
1540 {
1541         unsigned int bridge;
1542         size_t len;
1543         char* endp;
1544
1545         while (*p) {
1546                 if (!strncmp(p, "64k", 3))
1547                         specified_table_size = TCE_TABLE_SIZE_64K;
1548                 else if (!strncmp(p, "128k", 4))
1549                         specified_table_size = TCE_TABLE_SIZE_128K;
1550                 else if (!strncmp(p, "256k", 4))
1551                         specified_table_size = TCE_TABLE_SIZE_256K;
1552                 else if (!strncmp(p, "512k", 4))
1553                         specified_table_size = TCE_TABLE_SIZE_512K;
1554                 else if (!strncmp(p, "1M", 2))
1555                         specified_table_size = TCE_TABLE_SIZE_1M;
1556                 else if (!strncmp(p, "2M", 2))
1557                         specified_table_size = TCE_TABLE_SIZE_2M;
1558                 else if (!strncmp(p, "4M", 2))
1559                         specified_table_size = TCE_TABLE_SIZE_4M;
1560                 else if (!strncmp(p, "8M", 2))
1561                         specified_table_size = TCE_TABLE_SIZE_8M;
1562
1563                 len = strlen("translate_empty_slots");
1564                 if (!strncmp(p, "translate_empty_slots", len))
1565                         translate_empty_slots = 1;
1566
1567                 len = strlen("disable");
1568                 if (!strncmp(p, "disable", len)) {
1569                         p += len;
1570                         if (*p == '=')
1571                                 ++p;
1572                         if (*p == '\0')
1573                                 break;
1574                         bridge = simple_strtoul(p, &endp, 0);
1575                         if (p == endp)
1576                                 break;
1577
1578                         if (bridge < MAX_PHB_BUS_NUM) {
1579                                 printk(KERN_INFO "Calgary: disabling "
1580                                        "translation for PHB %#x\n", bridge);
1581                                 bus_info[bridge].translation_disabled = 1;
1582                         }
1583                 }
1584
1585                 p = strpbrk(p, ",");
1586                 if (!p)
1587                         break;
1588
1589                 p++; /* skip ',' */
1590         }
1591         return 1;
1592 }
1593 __setup("calgary=", calgary_parse_options);
1594
1595 static void __init calgary_fixup_one_tce_space(struct pci_dev *dev)
1596 {
1597         struct iommu_table *tbl;
1598         unsigned int npages;
1599         int i;
1600
1601         tbl = pci_iommu(dev->bus);
1602
1603         for (i = 0; i < 4; i++) {
1604                 struct resource *r = &dev->resource[PCI_BRIDGE_RESOURCES + i];
1605
1606                 /* Don't give out TCEs that map MEM resources */
1607                 if (!(r->flags & IORESOURCE_MEM))
1608                         continue;
1609
1610                 /* 0-based? we reserve the whole 1st MB anyway */
1611                 if (!r->start)
1612                         continue;
1613
1614                 /* cover the whole region */
1615                 npages = (r->end - r->start) >> PAGE_SHIFT;
1616                 npages++;
1617
1618                 iommu_range_reserve(tbl, r->start, npages);
1619         }
1620 }
1621
1622 static int __init calgary_fixup_tce_spaces(void)
1623 {
1624         struct pci_dev *dev = NULL;
1625         struct calgary_bus_info *info;
1626
1627         if (no_iommu || swiotlb || !calgary_detected)
1628                 return -ENODEV;
1629
1630         printk(KERN_DEBUG "Calgary: fixing up tce spaces\n");
1631
1632         do {
1633                 dev = pci_get_device(PCI_VENDOR_ID_IBM, PCI_ANY_ID, dev);
1634                 if (!dev)
1635                         break;
1636                 if (!is_cal_pci_dev(dev->device))
1637                         continue;
1638
1639                 info = &bus_info[dev->bus->number];
1640                 if (info->translation_disabled)
1641                         continue;
1642
1643                 if (!info->tce_space)
1644                         continue;
1645
1646                 calgary_fixup_one_tce_space(dev);
1647
1648         } while (1);
1649
1650         return 0;
1651 }
1652
1653 /*
1654  * We need to be call after pcibios_assign_resources (fs_initcall level)
1655  * and before device_initcall.
1656  */
1657 rootfs_initcall(calgary_fixup_tce_spaces);