1 /* 2 ** System Bus Adapter (SBA) I/O MMU manager 3 ** 4 ** (c) Copyright 2000-2004 Grant Grundler <grundler @ parisc-linux x org> 5 ** (c) Copyright 2004 Naresh Kumar Inna <knaresh at india x hp x com> 6 ** (c) Copyright 2000-2004 Hewlett-Packard Company 7 ** 8 ** Portions (c) 1999 Dave S. Miller (from sparc64 I/O MMU code) 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 ** 16 ** This module initializes the IOC (I/O Controller) found on B1000/C3000/ 17 ** J5000/J7000/N-class/L-class machines and their successors. 18 ** 19 ** FIXME: add DMA hint support programming in both sba and lba modules. 20 */ 21 22 #include <linux/types.h> 23 #include <linux/kernel.h> 24 #include <linux/spinlock.h> 25 #include <linux/slab.h> 26 #include <linux/init.h> 27 28 #include <linux/mm.h> 29 #include <linux/string.h> 30 #include <linux/pci.h> 31 #include <linux/scatterlist.h> 32 #include <linux/iommu-helper.h> 33 34 #include <asm/byteorder.h> 35 #include <asm/io.h> 36 #include <asm/dma.h> /* for DMA_CHUNK_SIZE */ 37 38 #include <asm/hardware.h> /* for register_parisc_driver() stuff */ 39 40 #include <linux/proc_fs.h> 41 #include <linux/seq_file.h> 42 #include <linux/module.h> 43 44 #include <asm/ropes.h> 45 #include <asm/mckinley.h> /* for proc_mckinley_root */ 46 #include <asm/runway.h> /* for proc_runway_root */ 47 #include <asm/page.h> /* for PAGE0 */ 48 #include <asm/pdc.h> /* for PDC_MODEL_* */ 49 #include <asm/pdcpat.h> /* for is_pdc_pat() */ 50 #include <asm/parisc-device.h> 51 52 #define MODULE_NAME "SBA" 53 54 /* 55 ** The number of debug flags is a clue - this code is fragile. 56 ** Don't even think about messing with it unless you have 57 ** plenty of 710's to sacrifice to the computer gods. :^) 58 */ 59 #undef DEBUG_SBA_INIT 60 #undef DEBUG_SBA_RUN 61 #undef DEBUG_SBA_RUN_SG 62 #undef DEBUG_SBA_RESOURCE 63 #undef ASSERT_PDIR_SANITY 64 #undef DEBUG_LARGE_SG_ENTRIES 65 #undef DEBUG_DMB_TRAP 66 67 #ifdef DEBUG_SBA_INIT 68 #define DBG_INIT(x...) printk(x) 69 #else 70 #define DBG_INIT(x...) 71 #endif 72 73 #ifdef DEBUG_SBA_RUN 74 #define DBG_RUN(x...) printk(x) 75 #else 76 #define DBG_RUN(x...) 77 #endif 78 79 #ifdef DEBUG_SBA_RUN_SG 80 #define DBG_RUN_SG(x...) printk(x) 81 #else 82 #define DBG_RUN_SG(x...) 83 #endif 84 85 86 #ifdef DEBUG_SBA_RESOURCE 87 #define DBG_RES(x...) printk(x) 88 #else 89 #define DBG_RES(x...) 90 #endif 91 92 #define SBA_INLINE __inline__ 93 94 #define DEFAULT_DMA_HINT_REG 0 95 96 struct sba_device *sba_list; 97 EXPORT_SYMBOL_GPL(sba_list); 98 99 static unsigned long ioc_needs_fdc = 0; 100 101 /* global count of IOMMUs in the system */ 102 static unsigned int global_ioc_cnt = 0; 103 104 /* PA8700 (Piranha 2.2) bug workaround */ 105 static unsigned long piranha_bad_128k = 0; 106 107 /* Looks nice and keeps the compiler happy */ 108 #define SBA_DEV(d) ((struct sba_device *) (d)) 109 110 #ifdef CONFIG_AGP_PARISC 111 #define SBA_AGP_SUPPORT 112 #endif /*CONFIG_AGP_PARISC*/ 113 114 #ifdef SBA_AGP_SUPPORT 115 static int sba_reserve_agpgart = 1; 116 module_param(sba_reserve_agpgart, int, 0444); 117 MODULE_PARM_DESC(sba_reserve_agpgart, "Reserve half of IO pdir as AGPGART"); 118 #endif 119 120 121 /************************************ 122 ** SBA register read and write support 123 ** 124 ** BE WARNED: register writes are posted. 125 ** (ie follow writes which must reach HW with a read) 126 ** 127 ** Superdome (in particular, REO) allows only 64-bit CSR accesses. 128 */ 129 #define READ_REG32(addr) readl(addr) 130 #define READ_REG64(addr) readq(addr) 131 #define WRITE_REG32(val, addr) writel((val), (addr)) 132 #define WRITE_REG64(val, addr) writeq((val), (addr)) 133 134 #ifdef CONFIG_64BIT 135 #define READ_REG(addr) READ_REG64(addr) 136 #define WRITE_REG(value, addr) WRITE_REG64(value, addr) 137 #else 138 #define READ_REG(addr) READ_REG32(addr) 139 #define WRITE_REG(value, addr) WRITE_REG32(value, addr) 140 #endif 141 142 #ifdef DEBUG_SBA_INIT 143 144 /* NOTE: When CONFIG_64BIT isn't defined, READ_REG64() is two 32-bit reads */ 145 146 /** 147 * sba_dump_ranges - debugging only - print ranges assigned to this IOA 148 * @hpa: base address of the sba 149 * 150 * Print the MMIO and IO Port address ranges forwarded by an Astro/Ike/RIO 151 * IO Adapter (aka Bus Converter). 152 */ 153 static void 154 sba_dump_ranges(void __iomem *hpa) 155 { 156 DBG_INIT("SBA at 0x%p\n", hpa); 157 DBG_INIT("IOS_DIST_BASE : %Lx\n", READ_REG64(hpa+IOS_DIST_BASE)); 158 DBG_INIT("IOS_DIST_MASK : %Lx\n", READ_REG64(hpa+IOS_DIST_MASK)); 159 DBG_INIT("IOS_DIST_ROUTE : %Lx\n", READ_REG64(hpa+IOS_DIST_ROUTE)); 160 DBG_INIT("\n"); 161 DBG_INIT("IOS_DIRECT_BASE : %Lx\n", READ_REG64(hpa+IOS_DIRECT_BASE)); 162 DBG_INIT("IOS_DIRECT_MASK : %Lx\n", READ_REG64(hpa+IOS_DIRECT_MASK)); 163 DBG_INIT("IOS_DIRECT_ROUTE: %Lx\n", READ_REG64(hpa+IOS_DIRECT_ROUTE)); 164 } 165 166 /** 167 * sba_dump_tlb - debugging only - print IOMMU operating parameters 168 * @hpa: base address of the IOMMU 169 * 170 * Print the size/location of the IO MMU PDIR. 171 */ 172 static void sba_dump_tlb(void __iomem *hpa) 173 { 174 DBG_INIT("IO TLB at 0x%p\n", hpa); 175 DBG_INIT("IOC_IBASE : 0x%Lx\n", READ_REG64(hpa+IOC_IBASE)); 176 DBG_INIT("IOC_IMASK : 0x%Lx\n", READ_REG64(hpa+IOC_IMASK)); 177 DBG_INIT("IOC_TCNFG : 0x%Lx\n", READ_REG64(hpa+IOC_TCNFG)); 178 DBG_INIT("IOC_PDIR_BASE: 0x%Lx\n", READ_REG64(hpa+IOC_PDIR_BASE)); 179 DBG_INIT("\n"); 180 } 181 #else 182 #define sba_dump_ranges(x) 183 #define sba_dump_tlb(x) 184 #endif /* DEBUG_SBA_INIT */ 185 186 187 #ifdef ASSERT_PDIR_SANITY 188 189 /** 190 * sba_dump_pdir_entry - debugging only - print one IOMMU PDIR entry 191 * @ioc: IO MMU structure which owns the pdir we are interested in. 192 * @msg: text to print ont the output line. 193 * @pide: pdir index. 194 * 195 * Print one entry of the IO MMU PDIR in human readable form. 196 */ 197 static void 198 sba_dump_pdir_entry(struct ioc *ioc, char *msg, uint pide) 199 { 200 /* start printing from lowest pde in rval */ 201 u64 *ptr = &(ioc->pdir_base[pide & (~0U * BITS_PER_LONG)]); 202 unsigned long *rptr = (unsigned long *) &(ioc->res_map[(pide >>3) & ~(sizeof(unsigned long) - 1)]); 203 uint rcnt; 204 205 printk(KERN_DEBUG "SBA: %s rp %p bit %d rval 0x%lx\n", 206 msg, 207 rptr, pide & (BITS_PER_LONG - 1), *rptr); 208 209 rcnt = 0; 210 while (rcnt < BITS_PER_LONG) { 211 printk(KERN_DEBUG "%s %2d %p %016Lx\n", 212 (rcnt == (pide & (BITS_PER_LONG - 1))) 213 ? " -->" : " ", 214 rcnt, ptr, *ptr ); 215 rcnt++; 216 ptr++; 217 } 218 printk(KERN_DEBUG "%s", msg); 219 } 220 221 222 /** 223 * sba_check_pdir - debugging only - consistency checker 224 * @ioc: IO MMU structure which owns the pdir we are interested in. 225 * @msg: text to print ont the output line. 226 * 227 * Verify the resource map and pdir state is consistent 228 */ 229 static int 230 sba_check_pdir(struct ioc *ioc, char *msg) 231 { 232 u32 *rptr_end = (u32 *) &(ioc->res_map[ioc->res_size]); 233 u32 *rptr = (u32 *) ioc->res_map; /* resource map ptr */ 234 u64 *pptr = ioc->pdir_base; /* pdir ptr */ 235 uint pide = 0; 236 237 while (rptr < rptr_end) { 238 u32 rval = *rptr; 239 int rcnt = 32; /* number of bits we might check */ 240 241 while (rcnt) { 242 /* Get last byte and highest bit from that */ 243 u32 pde = ((u32) (((char *)pptr)[7])) << 24; 244 if ((rval ^ pde) & 0x80000000) 245 { 246 /* 247 ** BUMMER! -- res_map != pdir -- 248 ** Dump rval and matching pdir entries 249 */ 250 sba_dump_pdir_entry(ioc, msg, pide); 251 return(1); 252 } 253 rcnt--; 254 rval <<= 1; /* try the next bit */ 255 pptr++; 256 pide++; 257 } 258 rptr++; /* look at next word of res_map */ 259 } 260 /* It'd be nice if we always got here :^) */ 261 return 0; 262 } 263 264 265 /** 266 * sba_dump_sg - debugging only - print Scatter-Gather list 267 * @ioc: IO MMU structure which owns the pdir we are interested in. 268 * @startsg: head of the SG list 269 * @nents: number of entries in SG list 270 * 271 * print the SG list so we can verify it's correct by hand. 272 */ 273 static void 274 sba_dump_sg( struct ioc *ioc, struct scatterlist *startsg, int nents) 275 { 276 while (nents-- > 0) { 277 printk(KERN_DEBUG " %d : %08lx/%05x %p/%05x\n", 278 nents, 279 (unsigned long) sg_dma_address(startsg), 280 sg_dma_len(startsg), 281 sg_virt(startsg), startsg->length); 282 startsg++; 283 } 284 } 285 286 #endif /* ASSERT_PDIR_SANITY */ 287 288 289 290 291 /************************************************************** 292 * 293 * I/O Pdir Resource Management 294 * 295 * Bits set in the resource map are in use. 296 * Each bit can represent a number of pages. 297 * LSbs represent lower addresses (IOVA's). 298 * 299 ***************************************************************/ 300 #define PAGES_PER_RANGE 1 /* could increase this to 4 or 8 if needed */ 301 302 /* Convert from IOVP to IOVA and vice versa. */ 303 304 #ifdef ZX1_SUPPORT 305 /* Pluto (aka ZX1) boxes need to set or clear the ibase bits appropriately */ 306 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((ioc->ibase) | (iovp) | (offset)) 307 #define SBA_IOVP(ioc,iova) ((iova) & (ioc)->iovp_mask) 308 #else 309 /* only support Astro and ancestors. Saves a few cycles in key places */ 310 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((iovp) | (offset)) 311 #define SBA_IOVP(ioc,iova) (iova) 312 #endif 313 314 #define PDIR_INDEX(iovp) ((iovp)>>IOVP_SHIFT) 315 316 #define RESMAP_MASK(n) (~0UL << (BITS_PER_LONG - (n))) 317 #define RESMAP_IDX_MASK (sizeof(unsigned long) - 1) 318 319 static unsigned long ptr_to_pide(struct ioc *ioc, unsigned long *res_ptr, 320 unsigned int bitshiftcnt) 321 { 322 return (((unsigned long)res_ptr - (unsigned long)ioc->res_map) << 3) 323 + bitshiftcnt; 324 } 325 326 /** 327 * sba_search_bitmap - find free space in IO PDIR resource bitmap 328 * @ioc: IO MMU structure which owns the pdir we are interested in. 329 * @bits_wanted: number of entries we need. 330 * 331 * Find consecutive free bits in resource bitmap. 332 * Each bit represents one entry in the IO Pdir. 333 * Cool perf optimization: search for log2(size) bits at a time. 334 */ 335 static SBA_INLINE unsigned long 336 sba_search_bitmap(struct ioc *ioc, struct device *dev, 337 unsigned long bits_wanted) 338 { 339 unsigned long *res_ptr = ioc->res_hint; 340 unsigned long *res_end = (unsigned long *) &(ioc->res_map[ioc->res_size]); 341 unsigned long pide = ~0UL, tpide; 342 unsigned long boundary_size; 343 unsigned long shift; 344 int ret; 345 346 boundary_size = ALIGN((unsigned long long)dma_get_seg_boundary(dev) + 1, 347 1ULL << IOVP_SHIFT) >> IOVP_SHIFT; 348 349 #if defined(ZX1_SUPPORT) 350 BUG_ON(ioc->ibase & ~IOVP_MASK); 351 shift = ioc->ibase >> IOVP_SHIFT; 352 #else 353 shift = 0; 354 #endif 355 356 if (bits_wanted > (BITS_PER_LONG/2)) { 357 /* Search word at a time - no mask needed */ 358 for(; res_ptr < res_end; ++res_ptr) { 359 tpide = ptr_to_pide(ioc, res_ptr, 0); 360 ret = iommu_is_span_boundary(tpide, bits_wanted, 361 shift, 362 boundary_size); 363 if ((*res_ptr == 0) && !ret) { 364 *res_ptr = RESMAP_MASK(bits_wanted); 365 pide = tpide; 366 break; 367 } 368 } 369 /* point to the next word on next pass */ 370 res_ptr++; 371 ioc->res_bitshift = 0; 372 } else { 373 /* 374 ** Search the resource bit map on well-aligned values. 375 ** "o" is the alignment. 376 ** We need the alignment to invalidate I/O TLB using 377 ** SBA HW features in the unmap path. 378 */ 379 unsigned long o = 1 << get_order(bits_wanted << PAGE_SHIFT); 380 uint bitshiftcnt = ALIGN(ioc->res_bitshift, o); 381 unsigned long mask; 382 383 if (bitshiftcnt >= BITS_PER_LONG) { 384 bitshiftcnt = 0; 385 res_ptr++; 386 } 387 mask = RESMAP_MASK(bits_wanted) >> bitshiftcnt; 388 389 DBG_RES("%s() o %ld %p", __func__, o, res_ptr); 390 while(res_ptr < res_end) 391 { 392 DBG_RES(" %p %lx %lx\n", res_ptr, mask, *res_ptr); 393 WARN_ON(mask == 0); 394 tpide = ptr_to_pide(ioc, res_ptr, bitshiftcnt); 395 ret = iommu_is_span_boundary(tpide, bits_wanted, 396 shift, 397 boundary_size); 398 if ((((*res_ptr) & mask) == 0) && !ret) { 399 *res_ptr |= mask; /* mark resources busy! */ 400 pide = tpide; 401 break; 402 } 403 mask >>= o; 404 bitshiftcnt += o; 405 if (mask == 0) { 406 mask = RESMAP_MASK(bits_wanted); 407 bitshiftcnt=0; 408 res_ptr++; 409 } 410 } 411 /* look in the same word on the next pass */ 412 ioc->res_bitshift = bitshiftcnt + bits_wanted; 413 } 414 415 /* wrapped ? */ 416 if (res_end <= res_ptr) { 417 ioc->res_hint = (unsigned long *) ioc->res_map; 418 ioc->res_bitshift = 0; 419 } else { 420 ioc->res_hint = res_ptr; 421 } 422 return (pide); 423 } 424 425 426 /** 427 * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap 428 * @ioc: IO MMU structure which owns the pdir we are interested in. 429 * @size: number of bytes to create a mapping for 430 * 431 * Given a size, find consecutive unmarked and then mark those bits in the 432 * resource bit map. 433 */ 434 static int 435 sba_alloc_range(struct ioc *ioc, struct device *dev, size_t size) 436 { 437 unsigned int pages_needed = size >> IOVP_SHIFT; 438 #ifdef SBA_COLLECT_STATS 439 unsigned long cr_start = mfctl(16); 440 #endif 441 unsigned long pide; 442 443 pide = sba_search_bitmap(ioc, dev, pages_needed); 444 if (pide >= (ioc->res_size << 3)) { 445 pide = sba_search_bitmap(ioc, dev, pages_needed); 446 if (pide >= (ioc->res_size << 3)) 447 panic("%s: I/O MMU @ %p is out of mapping resources\n", 448 __FILE__, ioc->ioc_hpa); 449 } 450 451 #ifdef ASSERT_PDIR_SANITY 452 /* verify the first enable bit is clear */ 453 if(0x00 != ((u8 *) ioc->pdir_base)[pide*sizeof(u64) + 7]) { 454 sba_dump_pdir_entry(ioc, "sba_search_bitmap() botched it?", pide); 455 } 456 #endif 457 458 DBG_RES("%s(%x) %d -> %lx hint %x/%x\n", 459 __func__, size, pages_needed, pide, 460 (uint) ((unsigned long) ioc->res_hint - (unsigned long) ioc->res_map), 461 ioc->res_bitshift ); 462 463 #ifdef SBA_COLLECT_STATS 464 { 465 unsigned long cr_end = mfctl(16); 466 unsigned long tmp = cr_end - cr_start; 467 /* check for roll over */ 468 cr_start = (cr_end < cr_start) ? -(tmp) : (tmp); 469 } 470 ioc->avg_search[ioc->avg_idx++] = cr_start; 471 ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1; 472 473 ioc->used_pages += pages_needed; 474 #endif 475 476 return (pide); 477 } 478 479 480 /** 481 * sba_free_range - unmark bits in IO PDIR resource bitmap 482 * @ioc: IO MMU structure which owns the pdir we are interested in. 483 * @iova: IO virtual address which was previously allocated. 484 * @size: number of bytes to create a mapping for 485 * 486 * clear bits in the ioc's resource map 487 */ 488 static SBA_INLINE void 489 sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size) 490 { 491 unsigned long iovp = SBA_IOVP(ioc, iova); 492 unsigned int pide = PDIR_INDEX(iovp); 493 unsigned int ridx = pide >> 3; /* convert bit to byte address */ 494 unsigned long *res_ptr = (unsigned long *) &((ioc)->res_map[ridx & ~RESMAP_IDX_MASK]); 495 496 int bits_not_wanted = size >> IOVP_SHIFT; 497 498 /* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */ 499 unsigned long m = RESMAP_MASK(bits_not_wanted) >> (pide & (BITS_PER_LONG - 1)); 500 501 DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n", 502 __func__, (uint) iova, size, 503 bits_not_wanted, m, pide, res_ptr, *res_ptr); 504 505 #ifdef SBA_COLLECT_STATS 506 ioc->used_pages -= bits_not_wanted; 507 #endif 508 509 *res_ptr &= ~m; 510 } 511 512 513 /************************************************************** 514 * 515 * "Dynamic DMA Mapping" support (aka "Coherent I/O") 516 * 517 ***************************************************************/ 518 519 #ifdef SBA_HINT_SUPPORT 520 #define SBA_DMA_HINT(ioc, val) ((val) << (ioc)->hint_shift_pdir) 521 #endif 522 523 typedef unsigned long space_t; 524 #define KERNEL_SPACE 0 525 526 /** 527 * sba_io_pdir_entry - fill in one IO PDIR entry 528 * @pdir_ptr: pointer to IO PDIR entry 529 * @sid: process Space ID - currently only support KERNEL_SPACE 530 * @vba: Virtual CPU address of buffer to map 531 * @hint: DMA hint set to use for this mapping 532 * 533 * SBA Mapping Routine 534 * 535 * Given a virtual address (vba, arg2) and space id, (sid, arg1) 536 * sba_io_pdir_entry() loads the I/O PDIR entry pointed to by 537 * pdir_ptr (arg0). 538 * Using the bass-ackwards HP bit numbering, Each IO Pdir entry 539 * for Astro/Ike looks like: 540 * 541 * 542 * 0 19 51 55 63 543 * +-+---------------------+----------------------------------+----+--------+ 544 * |V| U | PPN[43:12] | U | VI | 545 * +-+---------------------+----------------------------------+----+--------+ 546 * 547 * Pluto is basically identical, supports fewer physical address bits: 548 * 549 * 0 23 51 55 63 550 * +-+------------------------+-------------------------------+----+--------+ 551 * |V| U | PPN[39:12] | U | VI | 552 * +-+------------------------+-------------------------------+----+--------+ 553 * 554 * V == Valid Bit (Most Significant Bit is bit 0) 555 * U == Unused 556 * PPN == Physical Page Number 557 * VI == Virtual Index (aka Coherent Index) 558 * 559 * LPA instruction output is put into PPN field. 560 * LCI (Load Coherence Index) instruction provides the "VI" bits. 561 * 562 * We pre-swap the bytes since PCX-W is Big Endian and the 563 * IOMMU uses little endian for the pdir. 564 */ 565 566 static void SBA_INLINE 567 sba_io_pdir_entry(u64 *pdir_ptr, space_t sid, unsigned long vba, 568 unsigned long hint) 569 { 570 u64 pa; /* physical address */ 571 register unsigned ci; /* coherent index */ 572 573 pa = virt_to_phys(vba); 574 pa &= IOVP_MASK; 575 576 mtsp(sid,1); 577 asm("lci 0(%%sr1, %1), %0" : "=r" (ci) : "r" (vba)); 578 pa |= (ci >> PAGE_SHIFT) & 0xff; /* move CI (8 bits) into lowest byte */ 579 580 pa |= SBA_PDIR_VALID_BIT; /* set "valid" bit */ 581 *pdir_ptr = cpu_to_le64(pa); /* swap and store into I/O Pdir */ 582 583 /* 584 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set 585 * (bit #61, big endian), we have to flush and sync every time 586 * IO-PDIR is changed in Ike/Astro. 587 */ 588 if (ioc_needs_fdc) 589 asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr)); 590 } 591 592 593 /** 594 * sba_mark_invalid - invalidate one or more IO PDIR entries 595 * @ioc: IO MMU structure which owns the pdir we are interested in. 596 * @iova: IO Virtual Address mapped earlier 597 * @byte_cnt: number of bytes this mapping covers. 598 * 599 * Marking the IO PDIR entry(ies) as Invalid and invalidate 600 * corresponding IO TLB entry. The Ike PCOM (Purge Command Register) 601 * is to purge stale entries in the IO TLB when unmapping entries. 602 * 603 * The PCOM register supports purging of multiple pages, with a minium 604 * of 1 page and a maximum of 2GB. Hardware requires the address be 605 * aligned to the size of the range being purged. The size of the range 606 * must be a power of 2. The "Cool perf optimization" in the 607 * allocation routine helps keep that true. 608 */ 609 static SBA_INLINE void 610 sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt) 611 { 612 u32 iovp = (u32) SBA_IOVP(ioc,iova); 613 u64 *pdir_ptr = &ioc->pdir_base[PDIR_INDEX(iovp)]; 614 615 #ifdef ASSERT_PDIR_SANITY 616 /* Assert first pdir entry is set. 617 ** 618 ** Even though this is a big-endian machine, the entries 619 ** in the iopdir are little endian. That's why we look at 620 ** the byte at +7 instead of at +0. 621 */ 622 if (0x80 != (((u8 *) pdir_ptr)[7])) { 623 sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp)); 624 } 625 #endif 626 627 if (byte_cnt > IOVP_SIZE) 628 { 629 #if 0 630 unsigned long entries_per_cacheline = ioc_needs_fdc ? 631 L1_CACHE_ALIGN(((unsigned long) pdir_ptr)) 632 - (unsigned long) pdir_ptr; 633 : 262144; 634 #endif 635 636 /* set "size" field for PCOM */ 637 iovp |= get_order(byte_cnt) + PAGE_SHIFT; 638 639 do { 640 /* clear I/O Pdir entry "valid" bit first */ 641 ((u8 *) pdir_ptr)[7] = 0; 642 if (ioc_needs_fdc) { 643 asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr)); 644 #if 0 645 entries_per_cacheline = L1_CACHE_SHIFT - 3; 646 #endif 647 } 648 pdir_ptr++; 649 byte_cnt -= IOVP_SIZE; 650 } while (byte_cnt > IOVP_SIZE); 651 } else 652 iovp |= IOVP_SHIFT; /* set "size" field for PCOM */ 653 654 /* 655 ** clear I/O PDIR entry "valid" bit. 656 ** We have to R/M/W the cacheline regardless how much of the 657 ** pdir entry that we clobber. 658 ** The rest of the entry would be useful for debugging if we 659 ** could dump core on HPMC. 660 */ 661 ((u8 *) pdir_ptr)[7] = 0; 662 if (ioc_needs_fdc) 663 asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr)); 664 665 WRITE_REG( SBA_IOVA(ioc, iovp, 0, 0), ioc->ioc_hpa+IOC_PCOM); 666 } 667 668 /** 669 * sba_dma_supported - PCI driver can query DMA support 670 * @dev: instance of PCI owned by the driver that's asking 671 * @mask: number of address bits this PCI device can handle 672 * 673 * See Documentation/DMA-API-HOWTO.txt 674 */ 675 static int sba_dma_supported( struct device *dev, u64 mask) 676 { 677 struct ioc *ioc; 678 679 if (dev == NULL) { 680 printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n"); 681 BUG(); 682 return(0); 683 } 684 685 /* Documentation/DMA-API-HOWTO.txt tells drivers to try 64-bit 686 * first, then fall back to 32-bit if that fails. 687 * We are just "encouraging" 32-bit DMA masks here since we can 688 * never allow IOMMU bypass unless we add special support for ZX1. 689 */ 690 if (mask > ~0U) 691 return 0; 692 693 ioc = GET_IOC(dev); 694 695 /* 696 * check if mask is >= than the current max IO Virt Address 697 * The max IO Virt address will *always* < 30 bits. 698 */ 699 return((int)(mask >= (ioc->ibase - 1 + 700 (ioc->pdir_size / sizeof(u64) * IOVP_SIZE) ))); 701 } 702 703 704 /** 705 * sba_map_single - map one buffer and return IOVA for DMA 706 * @dev: instance of PCI owned by the driver that's asking. 707 * @addr: driver buffer to map. 708 * @size: number of bytes to map in driver buffer. 709 * @direction: R/W or both. 710 * 711 * See Documentation/DMA-API-HOWTO.txt 712 */ 713 static dma_addr_t 714 sba_map_single(struct device *dev, void *addr, size_t size, 715 enum dma_data_direction direction) 716 { 717 struct ioc *ioc; 718 unsigned long flags; 719 dma_addr_t iovp; 720 dma_addr_t offset; 721 u64 *pdir_start; 722 int pide; 723 724 ioc = GET_IOC(dev); 725 726 /* save offset bits */ 727 offset = ((dma_addr_t) (long) addr) & ~IOVP_MASK; 728 729 /* round up to nearest IOVP_SIZE */ 730 size = (size + offset + ~IOVP_MASK) & IOVP_MASK; 731 732 spin_lock_irqsave(&ioc->res_lock, flags); 733 #ifdef ASSERT_PDIR_SANITY 734 sba_check_pdir(ioc,"Check before sba_map_single()"); 735 #endif 736 737 #ifdef SBA_COLLECT_STATS 738 ioc->msingle_calls++; 739 ioc->msingle_pages += size >> IOVP_SHIFT; 740 #endif 741 pide = sba_alloc_range(ioc, dev, size); 742 iovp = (dma_addr_t) pide << IOVP_SHIFT; 743 744 DBG_RUN("%s() 0x%p -> 0x%lx\n", 745 __func__, addr, (long) iovp | offset); 746 747 pdir_start = &(ioc->pdir_base[pide]); 748 749 while (size > 0) { 750 sba_io_pdir_entry(pdir_start, KERNEL_SPACE, (unsigned long) addr, 0); 751 752 DBG_RUN(" pdir 0x%p %02x%02x%02x%02x%02x%02x%02x%02x\n", 753 pdir_start, 754 (u8) (((u8 *) pdir_start)[7]), 755 (u8) (((u8 *) pdir_start)[6]), 756 (u8) (((u8 *) pdir_start)[5]), 757 (u8) (((u8 *) pdir_start)[4]), 758 (u8) (((u8 *) pdir_start)[3]), 759 (u8) (((u8 *) pdir_start)[2]), 760 (u8) (((u8 *) pdir_start)[1]), 761 (u8) (((u8 *) pdir_start)[0]) 762 ); 763 764 addr += IOVP_SIZE; 765 size -= IOVP_SIZE; 766 pdir_start++; 767 } 768 769 /* force FDC ops in io_pdir_entry() to be visible to IOMMU */ 770 if (ioc_needs_fdc) 771 asm volatile("sync" : : ); 772 773 #ifdef ASSERT_PDIR_SANITY 774 sba_check_pdir(ioc,"Check after sba_map_single()"); 775 #endif 776 spin_unlock_irqrestore(&ioc->res_lock, flags); 777 778 /* form complete address */ 779 return SBA_IOVA(ioc, iovp, offset, DEFAULT_DMA_HINT_REG); 780 } 781 782 783 /** 784 * sba_unmap_single - unmap one IOVA and free resources 785 * @dev: instance of PCI owned by the driver that's asking. 786 * @iova: IOVA of driver buffer previously mapped. 787 * @size: number of bytes mapped in driver buffer. 788 * @direction: R/W or both. 789 * 790 * See Documentation/DMA-API-HOWTO.txt 791 */ 792 static void 793 sba_unmap_single(struct device *dev, dma_addr_t iova, size_t size, 794 enum dma_data_direction direction) 795 { 796 struct ioc *ioc; 797 #if DELAYED_RESOURCE_CNT > 0 798 struct sba_dma_pair *d; 799 #endif 800 unsigned long flags; 801 dma_addr_t offset; 802 803 DBG_RUN("%s() iovp 0x%lx/%x\n", __func__, (long) iova, size); 804 805 ioc = GET_IOC(dev); 806 offset = iova & ~IOVP_MASK; 807 iova ^= offset; /* clear offset bits */ 808 size += offset; 809 size = ALIGN(size, IOVP_SIZE); 810 811 spin_lock_irqsave(&ioc->res_lock, flags); 812 813 #ifdef SBA_COLLECT_STATS 814 ioc->usingle_calls++; 815 ioc->usingle_pages += size >> IOVP_SHIFT; 816 #endif 817 818 sba_mark_invalid(ioc, iova, size); 819 820 #if DELAYED_RESOURCE_CNT > 0 821 /* Delaying when we re-use a IO Pdir entry reduces the number 822 * of MMIO reads needed to flush writes to the PCOM register. 823 */ 824 d = &(ioc->saved[ioc->saved_cnt]); 825 d->iova = iova; 826 d->size = size; 827 if (++(ioc->saved_cnt) >= DELAYED_RESOURCE_CNT) { 828 int cnt = ioc->saved_cnt; 829 while (cnt--) { 830 sba_free_range(ioc, d->iova, d->size); 831 d--; 832 } 833 ioc->saved_cnt = 0; 834 835 READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */ 836 } 837 #else /* DELAYED_RESOURCE_CNT == 0 */ 838 sba_free_range(ioc, iova, size); 839 840 /* If fdc's were issued, force fdc's to be visible now */ 841 if (ioc_needs_fdc) 842 asm volatile("sync" : : ); 843 844 READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */ 845 #endif /* DELAYED_RESOURCE_CNT == 0 */ 846 847 spin_unlock_irqrestore(&ioc->res_lock, flags); 848 849 /* XXX REVISIT for 2.5 Linux - need syncdma for zero-copy support. 850 ** For Astro based systems this isn't a big deal WRT performance. 851 ** As long as 2.4 kernels copyin/copyout data from/to userspace, 852 ** we don't need the syncdma. The issue here is I/O MMU cachelines 853 ** are *not* coherent in all cases. May be hwrev dependent. 854 ** Need to investigate more. 855 asm volatile("syncdma"); 856 */ 857 } 858 859 860 /** 861 * sba_alloc_consistent - allocate/map shared mem for DMA 862 * @hwdev: instance of PCI owned by the driver that's asking. 863 * @size: number of bytes mapped in driver buffer. 864 * @dma_handle: IOVA of new buffer. 865 * 866 * See Documentation/DMA-API-HOWTO.txt 867 */ 868 static void *sba_alloc_consistent(struct device *hwdev, size_t size, 869 dma_addr_t *dma_handle, gfp_t gfp) 870 { 871 void *ret; 872 873 if (!hwdev) { 874 /* only support PCI */ 875 *dma_handle = 0; 876 return NULL; 877 } 878 879 ret = (void *) __get_free_pages(gfp, get_order(size)); 880 881 if (ret) { 882 memset(ret, 0, size); 883 *dma_handle = sba_map_single(hwdev, ret, size, 0); 884 } 885 886 return ret; 887 } 888 889 890 /** 891 * sba_free_consistent - free/unmap shared mem for DMA 892 * @hwdev: instance of PCI owned by the driver that's asking. 893 * @size: number of bytes mapped in driver buffer. 894 * @vaddr: virtual address IOVA of "consistent" buffer. 895 * @dma_handler: IO virtual address of "consistent" buffer. 896 * 897 * See Documentation/DMA-API-HOWTO.txt 898 */ 899 static void 900 sba_free_consistent(struct device *hwdev, size_t size, void *vaddr, 901 dma_addr_t dma_handle) 902 { 903 sba_unmap_single(hwdev, dma_handle, size, 0); 904 free_pages((unsigned long) vaddr, get_order(size)); 905 } 906 907 908 /* 909 ** Since 0 is a valid pdir_base index value, can't use that 910 ** to determine if a value is valid or not. Use a flag to indicate 911 ** the SG list entry contains a valid pdir index. 912 */ 913 #define PIDE_FLAG 0x80000000UL 914 915 #ifdef SBA_COLLECT_STATS 916 #define IOMMU_MAP_STATS 917 #endif 918 #include "iommu-helpers.h" 919 920 #ifdef DEBUG_LARGE_SG_ENTRIES 921 int dump_run_sg = 0; 922 #endif 923 924 925 /** 926 * sba_map_sg - map Scatter/Gather list 927 * @dev: instance of PCI owned by the driver that's asking. 928 * @sglist: array of buffer/length pairs 929 * @nents: number of entries in list 930 * @direction: R/W or both. 931 * 932 * See Documentation/DMA-API-HOWTO.txt 933 */ 934 static int 935 sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents, 936 enum dma_data_direction direction) 937 { 938 struct ioc *ioc; 939 int coalesced, filled = 0; 940 unsigned long flags; 941 942 DBG_RUN_SG("%s() START %d entries\n", __func__, nents); 943 944 ioc = GET_IOC(dev); 945 946 /* Fast path single entry scatterlists. */ 947 if (nents == 1) { 948 sg_dma_address(sglist) = sba_map_single(dev, sg_virt(sglist), 949 sglist->length, direction); 950 sg_dma_len(sglist) = sglist->length; 951 return 1; 952 } 953 954 spin_lock_irqsave(&ioc->res_lock, flags); 955 956 #ifdef ASSERT_PDIR_SANITY 957 if (sba_check_pdir(ioc,"Check before sba_map_sg()")) 958 { 959 sba_dump_sg(ioc, sglist, nents); 960 panic("Check before sba_map_sg()"); 961 } 962 #endif 963 964 #ifdef SBA_COLLECT_STATS 965 ioc->msg_calls++; 966 #endif 967 968 /* 969 ** First coalesce the chunks and allocate I/O pdir space 970 ** 971 ** If this is one DMA stream, we can properly map using the 972 ** correct virtual address associated with each DMA page. 973 ** w/o this association, we wouldn't have coherent DMA! 974 ** Access to the virtual address is what forces a two pass algorithm. 975 */ 976 coalesced = iommu_coalesce_chunks(ioc, dev, sglist, nents, sba_alloc_range); 977 978 /* 979 ** Program the I/O Pdir 980 ** 981 ** map the virtual addresses to the I/O Pdir 982 ** o dma_address will contain the pdir index 983 ** o dma_len will contain the number of bytes to map 984 ** o address contains the virtual address. 985 */ 986 filled = iommu_fill_pdir(ioc, sglist, nents, 0, sba_io_pdir_entry); 987 988 /* force FDC ops in io_pdir_entry() to be visible to IOMMU */ 989 if (ioc_needs_fdc) 990 asm volatile("sync" : : ); 991 992 #ifdef ASSERT_PDIR_SANITY 993 if (sba_check_pdir(ioc,"Check after sba_map_sg()")) 994 { 995 sba_dump_sg(ioc, sglist, nents); 996 panic("Check after sba_map_sg()\n"); 997 } 998 #endif 999 1000 spin_unlock_irqrestore(&ioc->res_lock, flags); 1001 1002 DBG_RUN_SG("%s() DONE %d mappings\n", __func__, filled); 1003 1004 return filled; 1005 } 1006 1007 1008 /** 1009 * sba_unmap_sg - unmap Scatter/Gather list 1010 * @dev: instance of PCI owned by the driver that's asking. 1011 * @sglist: array of buffer/length pairs 1012 * @nents: number of entries in list 1013 * @direction: R/W or both. 1014 * 1015 * See Documentation/DMA-API-HOWTO.txt 1016 */ 1017 static void 1018 sba_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents, 1019 enum dma_data_direction direction) 1020 { 1021 struct ioc *ioc; 1022 #ifdef ASSERT_PDIR_SANITY 1023 unsigned long flags; 1024 #endif 1025 1026 DBG_RUN_SG("%s() START %d entries, %p,%x\n", 1027 __func__, nents, sg_virt(sglist), sglist->length); 1028 1029 ioc = GET_IOC(dev); 1030 1031 #ifdef SBA_COLLECT_STATS 1032 ioc->usg_calls++; 1033 #endif 1034 1035 #ifdef ASSERT_PDIR_SANITY 1036 spin_lock_irqsave(&ioc->res_lock, flags); 1037 sba_check_pdir(ioc,"Check before sba_unmap_sg()"); 1038 spin_unlock_irqrestore(&ioc->res_lock, flags); 1039 #endif 1040 1041 while (sg_dma_len(sglist) && nents--) { 1042 1043 sba_unmap_single(dev, sg_dma_address(sglist), sg_dma_len(sglist), direction); 1044 #ifdef SBA_COLLECT_STATS 1045 ioc->usg_pages += ((sg_dma_address(sglist) & ~IOVP_MASK) + sg_dma_len(sglist) + IOVP_SIZE - 1) >> PAGE_SHIFT; 1046 ioc->usingle_calls--; /* kluge since call is unmap_sg() */ 1047 #endif 1048 ++sglist; 1049 } 1050 1051 DBG_RUN_SG("%s() DONE (nents %d)\n", __func__, nents); 1052 1053 #ifdef ASSERT_PDIR_SANITY 1054 spin_lock_irqsave(&ioc->res_lock, flags); 1055 sba_check_pdir(ioc,"Check after sba_unmap_sg()"); 1056 spin_unlock_irqrestore(&ioc->res_lock, flags); 1057 #endif 1058 1059 } 1060 1061 static struct hppa_dma_ops sba_ops = { 1062 .dma_supported = sba_dma_supported, 1063 .alloc_consistent = sba_alloc_consistent, 1064 .alloc_noncoherent = sba_alloc_consistent, 1065 .free_consistent = sba_free_consistent, 1066 .map_single = sba_map_single, 1067 .unmap_single = sba_unmap_single, 1068 .map_sg = sba_map_sg, 1069 .unmap_sg = sba_unmap_sg, 1070 .dma_sync_single_for_cpu = NULL, 1071 .dma_sync_single_for_device = NULL, 1072 .dma_sync_sg_for_cpu = NULL, 1073 .dma_sync_sg_for_device = NULL, 1074 }; 1075 1076 1077 /************************************************************************** 1078 ** 1079 ** SBA PAT PDC support 1080 ** 1081 ** o call pdc_pat_cell_module() 1082 ** o store ranges in PCI "resource" structures 1083 ** 1084 **************************************************************************/ 1085 1086 static void 1087 sba_get_pat_resources(struct sba_device *sba_dev) 1088 { 1089 #if 0 1090 /* 1091 ** TODO/REVISIT/FIXME: support for directed ranges requires calls to 1092 ** PAT PDC to program the SBA/LBA directed range registers...this 1093 ** burden may fall on the LBA code since it directly supports the 1094 ** PCI subsystem. It's not clear yet. - ggg 1095 */ 1096 PAT_MOD(mod)->mod_info.mod_pages = PAT_GET_MOD_PAGES(temp); 1097 FIXME : ??? 1098 PAT_MOD(mod)->mod_info.dvi = PAT_GET_DVI(temp); 1099 Tells where the dvi bits are located in the address. 1100 PAT_MOD(mod)->mod_info.ioc = PAT_GET_IOC(temp); 1101 FIXME : ??? 1102 #endif 1103 } 1104 1105 1106 /************************************************************** 1107 * 1108 * Initialization and claim 1109 * 1110 ***************************************************************/ 1111 #define PIRANHA_ADDR_MASK 0x00160000UL /* bit 17,18,20 */ 1112 #define PIRANHA_ADDR_VAL 0x00060000UL /* bit 17,18 on */ 1113 static void * 1114 sba_alloc_pdir(unsigned int pdir_size) 1115 { 1116 unsigned long pdir_base; 1117 unsigned long pdir_order = get_order(pdir_size); 1118 1119 pdir_base = __get_free_pages(GFP_KERNEL, pdir_order); 1120 if (NULL == (void *) pdir_base) { 1121 panic("%s() could not allocate I/O Page Table\n", 1122 __func__); 1123 } 1124 1125 /* If this is not PA8700 (PCX-W2) 1126 ** OR newer than ver 2.2 1127 ** OR in a system that doesn't need VINDEX bits from SBA, 1128 ** 1129 ** then we aren't exposed to the HW bug. 1130 */ 1131 if ( ((boot_cpu_data.pdc.cpuid >> 5) & 0x7f) != 0x13 1132 || (boot_cpu_data.pdc.versions > 0x202) 1133 || (boot_cpu_data.pdc.capabilities & 0x08L) ) 1134 return (void *) pdir_base; 1135 1136 /* 1137 * PA8700 (PCX-W2, aka piranha) silent data corruption fix 1138 * 1139 * An interaction between PA8700 CPU (Ver 2.2 or older) and 1140 * Ike/Astro can cause silent data corruption. This is only 1141 * a problem if the I/O PDIR is located in memory such that 1142 * (little-endian) bits 17 and 18 are on and bit 20 is off. 1143 * 1144 * Since the max IO Pdir size is 2MB, by cleverly allocating the 1145 * right physical address, we can either avoid (IOPDIR <= 1MB) 1146 * or minimize (2MB IO Pdir) the problem if we restrict the 1147 * IO Pdir to a maximum size of 2MB-128K (1902K). 1148 * 1149 * Because we always allocate 2^N sized IO pdirs, either of the 1150 * "bad" regions will be the last 128K if at all. That's easy 1151 * to test for. 1152 * 1153 */ 1154 if (pdir_order <= (19-12)) { 1155 if (((virt_to_phys(pdir_base)+pdir_size-1) & PIRANHA_ADDR_MASK) == PIRANHA_ADDR_VAL) { 1156 /* allocate a new one on 512k alignment */ 1157 unsigned long new_pdir = __get_free_pages(GFP_KERNEL, (19-12)); 1158 /* release original */ 1159 free_pages(pdir_base, pdir_order); 1160 1161 pdir_base = new_pdir; 1162 1163 /* release excess */ 1164 while (pdir_order < (19-12)) { 1165 new_pdir += pdir_size; 1166 free_pages(new_pdir, pdir_order); 1167 pdir_order +=1; 1168 pdir_size <<=1; 1169 } 1170 } 1171 } else { 1172 /* 1173 ** 1MB or 2MB Pdir 1174 ** Needs to be aligned on an "odd" 1MB boundary. 1175 */ 1176 unsigned long new_pdir = __get_free_pages(GFP_KERNEL, pdir_order+1); /* 2 or 4MB */ 1177 1178 /* release original */ 1179 free_pages( pdir_base, pdir_order); 1180 1181 /* release first 1MB */ 1182 free_pages(new_pdir, 20-12); 1183 1184 pdir_base = new_pdir + 1024*1024; 1185 1186 if (pdir_order > (20-12)) { 1187 /* 1188 ** 2MB Pdir. 1189 ** 1190 ** Flag tells init_bitmap() to mark bad 128k as used 1191 ** and to reduce the size by 128k. 1192 */ 1193 piranha_bad_128k = 1; 1194 1195 new_pdir += 3*1024*1024; 1196 /* release last 1MB */ 1197 free_pages(new_pdir, 20-12); 1198 1199 /* release unusable 128KB */ 1200 free_pages(new_pdir - 128*1024 , 17-12); 1201 1202 pdir_size -= 128*1024; 1203 } 1204 } 1205 1206 memset((void *) pdir_base, 0, pdir_size); 1207 return (void *) pdir_base; 1208 } 1209 1210 struct ibase_data_struct { 1211 struct ioc *ioc; 1212 int ioc_num; 1213 }; 1214 1215 static int setup_ibase_imask_callback(struct device *dev, void *data) 1216 { 1217 /* lba_set_iregs() is in drivers/parisc/lba_pci.c */ 1218 extern void lba_set_iregs(struct parisc_device *, u32, u32); 1219 struct parisc_device *lba = to_parisc_device(dev); 1220 struct ibase_data_struct *ibd = data; 1221 int rope_num = (lba->hpa.start >> 13) & 0xf; 1222 if (rope_num >> 3 == ibd->ioc_num) 1223 lba_set_iregs(lba, ibd->ioc->ibase, ibd->ioc->imask); 1224 return 0; 1225 } 1226 1227 /* setup Mercury or Elroy IBASE/IMASK registers. */ 1228 static void 1229 setup_ibase_imask(struct parisc_device *sba, struct ioc *ioc, int ioc_num) 1230 { 1231 struct ibase_data_struct ibase_data = { 1232 .ioc = ioc, 1233 .ioc_num = ioc_num, 1234 }; 1235 1236 device_for_each_child(&sba->dev, &ibase_data, 1237 setup_ibase_imask_callback); 1238 } 1239 1240 #ifdef SBA_AGP_SUPPORT 1241 static int 1242 sba_ioc_find_quicksilver(struct device *dev, void *data) 1243 { 1244 int *agp_found = data; 1245 struct parisc_device *lba = to_parisc_device(dev); 1246 1247 if (IS_QUICKSILVER(lba)) 1248 *agp_found = 1; 1249 return 0; 1250 } 1251 #endif 1252 1253 static void 1254 sba_ioc_init_pluto(struct parisc_device *sba, struct ioc *ioc, int ioc_num) 1255 { 1256 u32 iova_space_mask; 1257 u32 iova_space_size; 1258 int iov_order, tcnfg; 1259 #ifdef SBA_AGP_SUPPORT 1260 int agp_found = 0; 1261 #endif 1262 /* 1263 ** Firmware programs the base and size of a "safe IOVA space" 1264 ** (one that doesn't overlap memory or LMMIO space) in the 1265 ** IBASE and IMASK registers. 1266 */ 1267 ioc->ibase = READ_REG(ioc->ioc_hpa + IOC_IBASE); 1268 iova_space_size = ~(READ_REG(ioc->ioc_hpa + IOC_IMASK) & 0xFFFFFFFFUL) + 1; 1269 1270 if ((ioc->ibase < 0xfed00000UL) && ((ioc->ibase + iova_space_size) > 0xfee00000UL)) { 1271 printk("WARNING: IOV space overlaps local config and interrupt message, truncating\n"); 1272 iova_space_size /= 2; 1273 } 1274 1275 /* 1276 ** iov_order is always based on a 1GB IOVA space since we want to 1277 ** turn on the other half for AGP GART. 1278 */ 1279 iov_order = get_order(iova_space_size >> (IOVP_SHIFT - PAGE_SHIFT)); 1280 ioc->pdir_size = (iova_space_size / IOVP_SIZE) * sizeof(u64); 1281 1282 DBG_INIT("%s() hpa 0x%p IOV %dMB (%d bits)\n", 1283 __func__, ioc->ioc_hpa, iova_space_size >> 20, 1284 iov_order + PAGE_SHIFT); 1285 1286 ioc->pdir_base = (void *) __get_free_pages(GFP_KERNEL, 1287 get_order(ioc->pdir_size)); 1288 if (!ioc->pdir_base) 1289 panic("Couldn't allocate I/O Page Table\n"); 1290 1291 memset(ioc->pdir_base, 0, ioc->pdir_size); 1292 1293 DBG_INIT("%s() pdir %p size %x\n", 1294 __func__, ioc->pdir_base, ioc->pdir_size); 1295 1296 #ifdef SBA_HINT_SUPPORT 1297 ioc->hint_shift_pdir = iov_order + PAGE_SHIFT; 1298 ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT)); 1299 1300 DBG_INIT(" hint_shift_pdir %x hint_mask_pdir %lx\n", 1301 ioc->hint_shift_pdir, ioc->hint_mask_pdir); 1302 #endif 1303 1304 WARN_ON((((unsigned long) ioc->pdir_base) & PAGE_MASK) != (unsigned long) ioc->pdir_base); 1305 WRITE_REG(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE); 1306 1307 /* build IMASK for IOC and Elroy */ 1308 iova_space_mask = 0xffffffff; 1309 iova_space_mask <<= (iov_order + PAGE_SHIFT); 1310 ioc->imask = iova_space_mask; 1311 #ifdef ZX1_SUPPORT 1312 ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1); 1313 #endif 1314 sba_dump_tlb(ioc->ioc_hpa); 1315 1316 setup_ibase_imask(sba, ioc, ioc_num); 1317 1318 WRITE_REG(ioc->imask, ioc->ioc_hpa + IOC_IMASK); 1319 1320 #ifdef CONFIG_64BIT 1321 /* 1322 ** Setting the upper bits makes checking for bypass addresses 1323 ** a little faster later on. 1324 */ 1325 ioc->imask |= 0xFFFFFFFF00000000UL; 1326 #endif 1327 1328 /* Set I/O PDIR Page size to system page size */ 1329 switch (PAGE_SHIFT) { 1330 case 12: tcnfg = 0; break; /* 4K */ 1331 case 13: tcnfg = 1; break; /* 8K */ 1332 case 14: tcnfg = 2; break; /* 16K */ 1333 case 16: tcnfg = 3; break; /* 64K */ 1334 default: 1335 panic(__FILE__ "Unsupported system page size %d", 1336 1 << PAGE_SHIFT); 1337 break; 1338 } 1339 WRITE_REG(tcnfg, ioc->ioc_hpa + IOC_TCNFG); 1340 1341 /* 1342 ** Program the IOC's ibase and enable IOVA translation 1343 ** Bit zero == enable bit. 1344 */ 1345 WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa + IOC_IBASE); 1346 1347 /* 1348 ** Clear I/O TLB of any possible entries. 1349 ** (Yes. This is a bit paranoid...but so what) 1350 */ 1351 WRITE_REG(ioc->ibase | 31, ioc->ioc_hpa + IOC_PCOM); 1352 1353 #ifdef SBA_AGP_SUPPORT 1354 1355 /* 1356 ** If an AGP device is present, only use half of the IOV space 1357 ** for PCI DMA. Unfortunately we can't know ahead of time 1358 ** whether GART support will actually be used, for now we 1359 ** can just key on any AGP device found in the system. 1360 ** We program the next pdir index after we stop w/ a key for 1361 ** the GART code to handshake on. 1362 */ 1363 device_for_each_child(&sba->dev, &agp_found, sba_ioc_find_quicksilver); 1364 1365 if (agp_found && sba_reserve_agpgart) { 1366 printk(KERN_INFO "%s: reserving %dMb of IOVA space for agpgart\n", 1367 __func__, (iova_space_size/2) >> 20); 1368 ioc->pdir_size /= 2; 1369 ioc->pdir_base[PDIR_INDEX(iova_space_size/2)] = SBA_AGPGART_COOKIE; 1370 } 1371 #endif /*SBA_AGP_SUPPORT*/ 1372 } 1373 1374 static void 1375 sba_ioc_init(struct parisc_device *sba, struct ioc *ioc, int ioc_num) 1376 { 1377 u32 iova_space_size, iova_space_mask; 1378 unsigned int pdir_size, iov_order, tcnfg; 1379 1380 /* 1381 ** Determine IOVA Space size from memory size. 1382 ** 1383 ** Ideally, PCI drivers would register the maximum number 1384 ** of DMA they can have outstanding for each device they 1385 ** own. Next best thing would be to guess how much DMA 1386 ** can be outstanding based on PCI Class/sub-class. Both 1387 ** methods still require some "extra" to support PCI 1388 ** Hot-Plug/Removal of PCI cards. (aka PCI OLARD). 1389 ** 1390 ** While we have 32-bits "IOVA" space, top two 2 bits are used 1391 ** for DMA hints - ergo only 30 bits max. 1392 */ 1393 1394 iova_space_size = (u32) (totalram_pages/global_ioc_cnt); 1395 1396 /* limit IOVA space size to 1MB-1GB */ 1397 if (iova_space_size < (1 << (20 - PAGE_SHIFT))) { 1398 iova_space_size = 1 << (20 - PAGE_SHIFT); 1399 } 1400 else if (iova_space_size > (1 << (30 - PAGE_SHIFT))) { 1401 iova_space_size = 1 << (30 - PAGE_SHIFT); 1402 } 1403 1404 /* 1405 ** iova space must be log2() in size. 1406 ** thus, pdir/res_map will also be log2(). 1407 ** PIRANHA BUG: Exception is when IO Pdir is 2MB (gets reduced) 1408 */ 1409 iov_order = get_order(iova_space_size << PAGE_SHIFT); 1410 1411 /* iova_space_size is now bytes, not pages */ 1412 iova_space_size = 1 << (iov_order + PAGE_SHIFT); 1413 1414 ioc->pdir_size = pdir_size = (iova_space_size/IOVP_SIZE) * sizeof(u64); 1415 1416 DBG_INIT("%s() hpa 0x%lx mem %ldMB IOV %dMB (%d bits)\n", 1417 __func__, 1418 ioc->ioc_hpa, 1419 (unsigned long) totalram_pages >> (20 - PAGE_SHIFT), 1420 iova_space_size>>20, 1421 iov_order + PAGE_SHIFT); 1422 1423 ioc->pdir_base = sba_alloc_pdir(pdir_size); 1424 1425 DBG_INIT("%s() pdir %p size %x\n", 1426 __func__, ioc->pdir_base, pdir_size); 1427 1428 #ifdef SBA_HINT_SUPPORT 1429 /* FIXME : DMA HINTs not used */ 1430 ioc->hint_shift_pdir = iov_order + PAGE_SHIFT; 1431 ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT)); 1432 1433 DBG_INIT(" hint_shift_pdir %x hint_mask_pdir %lx\n", 1434 ioc->hint_shift_pdir, ioc->hint_mask_pdir); 1435 #endif 1436 1437 WRITE_REG64(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE); 1438 1439 /* build IMASK for IOC and Elroy */ 1440 iova_space_mask = 0xffffffff; 1441 iova_space_mask <<= (iov_order + PAGE_SHIFT); 1442 1443 /* 1444 ** On C3000 w/512MB mem, HP-UX 10.20 reports: 1445 ** ibase=0, imask=0xFE000000, size=0x2000000. 1446 */ 1447 ioc->ibase = 0; 1448 ioc->imask = iova_space_mask; /* save it */ 1449 #ifdef ZX1_SUPPORT 1450 ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1); 1451 #endif 1452 1453 DBG_INIT("%s() IOV base 0x%lx mask 0x%0lx\n", 1454 __func__, ioc->ibase, ioc->imask); 1455 1456 /* 1457 ** FIXME: Hint registers are programmed with default hint 1458 ** values during boot, so hints should be sane even if we 1459 ** can't reprogram them the way drivers want. 1460 */ 1461 1462 setup_ibase_imask(sba, ioc, ioc_num); 1463 1464 /* 1465 ** Program the IOC's ibase and enable IOVA translation 1466 */ 1467 WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa+IOC_IBASE); 1468 WRITE_REG(ioc->imask, ioc->ioc_hpa+IOC_IMASK); 1469 1470 /* Set I/O PDIR Page size to system page size */ 1471 switch (PAGE_SHIFT) { 1472 case 12: tcnfg = 0; break; /* 4K */ 1473 case 13: tcnfg = 1; break; /* 8K */ 1474 case 14: tcnfg = 2; break; /* 16K */ 1475 case 16: tcnfg = 3; break; /* 64K */ 1476 default: 1477 panic(__FILE__ "Unsupported system page size %d", 1478 1 << PAGE_SHIFT); 1479 break; 1480 } 1481 /* Set I/O PDIR Page size to PAGE_SIZE (4k/16k/...) */ 1482 WRITE_REG(tcnfg, ioc->ioc_hpa+IOC_TCNFG); 1483 1484 /* 1485 ** Clear I/O TLB of any possible entries. 1486 ** (Yes. This is a bit paranoid...but so what) 1487 */ 1488 WRITE_REG(0 | 31, ioc->ioc_hpa+IOC_PCOM); 1489 1490 ioc->ibase = 0; /* used by SBA_IOVA and related macros */ 1491 1492 DBG_INIT("%s() DONE\n", __func__); 1493 } 1494 1495 1496 1497 /************************************************************************** 1498 ** 1499 ** SBA initialization code (HW and SW) 1500 ** 1501 ** o identify SBA chip itself 1502 ** o initialize SBA chip modes (HardFail) 1503 ** o initialize SBA chip modes (HardFail) 1504 ** o FIXME: initialize DMA hints for reasonable defaults 1505 ** 1506 **************************************************************************/ 1507 1508 static void __iomem *ioc_remap(struct sba_device *sba_dev, unsigned int offset) 1509 { 1510 return ioremap_nocache(sba_dev->dev->hpa.start + offset, SBA_FUNC_SIZE); 1511 } 1512 1513 static void sba_hw_init(struct sba_device *sba_dev) 1514 { 1515 int i; 1516 int num_ioc; 1517 u64 ioc_ctl; 1518 1519 if (!is_pdc_pat()) { 1520 /* Shutdown the USB controller on Astro-based workstations. 1521 ** Once we reprogram the IOMMU, the next DMA performed by 1522 ** USB will HPMC the box. USB is only enabled if a 1523 ** keyboard is present and found. 1524 ** 1525 ** With serial console, j6k v5.0 firmware says: 1526 ** mem_kbd hpa 0xfee003f8 sba 0x0 pad 0x0 cl_class 0x7 1527 ** 1528 ** FIXME: Using GFX+USB console at power up but direct 1529 ** linux to serial console is still broken. 1530 ** USB could generate DMA so we must reset USB. 1531 ** The proper sequence would be: 1532 ** o block console output 1533 ** o reset USB device 1534 ** o reprogram serial port 1535 ** o unblock console output 1536 */ 1537 if (PAGE0->mem_kbd.cl_class == CL_KEYBD) { 1538 pdc_io_reset_devices(); 1539 } 1540 1541 } 1542 1543 1544 #if 0 1545 printk("sba_hw_init(): mem_boot 0x%x 0x%x 0x%x 0x%x\n", PAGE0->mem_boot.hpa, 1546 PAGE0->mem_boot.spa, PAGE0->mem_boot.pad, PAGE0->mem_boot.cl_class); 1547 1548 /* 1549 ** Need to deal with DMA from LAN. 1550 ** Maybe use page zero boot device as a handle to talk 1551 ** to PDC about which device to shutdown. 1552 ** 1553 ** Netbooting, j6k v5.0 firmware says: 1554 ** mem_boot hpa 0xf4008000 sba 0x0 pad 0x0 cl_class 0x1002 1555 ** ARGH! invalid class. 1556 */ 1557 if ((PAGE0->mem_boot.cl_class != CL_RANDOM) 1558 && (PAGE0->mem_boot.cl_class != CL_SEQU)) { 1559 pdc_io_reset(); 1560 } 1561 #endif 1562 1563 if (!IS_PLUTO(sba_dev->dev)) { 1564 ioc_ctl = READ_REG(sba_dev->sba_hpa+IOC_CTRL); 1565 DBG_INIT("%s() hpa 0x%lx ioc_ctl 0x%Lx ->", 1566 __func__, sba_dev->sba_hpa, ioc_ctl); 1567 ioc_ctl &= ~(IOC_CTRL_RM | IOC_CTRL_NC | IOC_CTRL_CE); 1568 ioc_ctl |= IOC_CTRL_DD | IOC_CTRL_D4 | IOC_CTRL_TC; 1569 /* j6700 v1.6 firmware sets 0x294f */ 1570 /* A500 firmware sets 0x4d */ 1571 1572 WRITE_REG(ioc_ctl, sba_dev->sba_hpa+IOC_CTRL); 1573 1574 #ifdef DEBUG_SBA_INIT 1575 ioc_ctl = READ_REG64(sba_dev->sba_hpa+IOC_CTRL); 1576 DBG_INIT(" 0x%Lx\n", ioc_ctl); 1577 #endif 1578 } /* if !PLUTO */ 1579 1580 if (IS_ASTRO(sba_dev->dev)) { 1581 int err; 1582 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, ASTRO_IOC_OFFSET); 1583 num_ioc = 1; 1584 1585 sba_dev->chip_resv.name = "Astro Intr Ack"; 1586 sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfef00000UL; 1587 sba_dev->chip_resv.end = PCI_F_EXTEND | (0xff000000UL - 1) ; 1588 err = request_resource(&iomem_resource, &(sba_dev->chip_resv)); 1589 BUG_ON(err < 0); 1590 1591 } else if (IS_PLUTO(sba_dev->dev)) { 1592 int err; 1593 1594 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, PLUTO_IOC_OFFSET); 1595 num_ioc = 1; 1596 1597 sba_dev->chip_resv.name = "Pluto Intr/PIOP/VGA"; 1598 sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfee00000UL; 1599 sba_dev->chip_resv.end = PCI_F_EXTEND | (0xff200000UL - 1); 1600 err = request_resource(&iomem_resource, &(sba_dev->chip_resv)); 1601 WARN_ON(err < 0); 1602 1603 sba_dev->iommu_resv.name = "IOVA Space"; 1604 sba_dev->iommu_resv.start = 0x40000000UL; 1605 sba_dev->iommu_resv.end = 0x50000000UL - 1; 1606 err = request_resource(&iomem_resource, &(sba_dev->iommu_resv)); 1607 WARN_ON(err < 0); 1608 } else { 1609 /* IKE, REO */ 1610 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(0)); 1611 sba_dev->ioc[1].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(1)); 1612 num_ioc = 2; 1613 1614 /* TODO - LOOKUP Ike/Stretch chipset mem map */ 1615 } 1616 /* XXX: What about Reo Grande? */ 1617 1618 sba_dev->num_ioc = num_ioc; 1619 for (i = 0; i < num_ioc; i++) { 1620 void __iomem *ioc_hpa = sba_dev->ioc[i].ioc_hpa; 1621 unsigned int j; 1622 1623 for (j=0; j < sizeof(u64) * ROPES_PER_IOC; j+=sizeof(u64)) { 1624 1625 /* 1626 * Clear ROPE(N)_CONFIG AO bit. 1627 * Disables "NT Ordering" (~= !"Relaxed Ordering") 1628 * Overrides bit 1 in DMA Hint Sets. 1629 * Improves netperf UDP_STREAM by ~10% for bcm5701. 1630 */ 1631 if (IS_PLUTO(sba_dev->dev)) { 1632 void __iomem *rope_cfg; 1633 unsigned long cfg_val; 1634 1635 rope_cfg = ioc_hpa + IOC_ROPE0_CFG + j; 1636 cfg_val = READ_REG(rope_cfg); 1637 cfg_val &= ~IOC_ROPE_AO; 1638 WRITE_REG(cfg_val, rope_cfg); 1639 } 1640 1641 /* 1642 ** Make sure the box crashes on rope errors. 1643 */ 1644 WRITE_REG(HF_ENABLE, ioc_hpa + ROPE0_CTL + j); 1645 } 1646 1647 /* flush out the last writes */ 1648 READ_REG(sba_dev->ioc[i].ioc_hpa + ROPE7_CTL); 1649 1650 DBG_INIT(" ioc[%d] ROPE_CFG 0x%Lx ROPE_DBG 0x%Lx\n", 1651 i, 1652 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x40), 1653 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x50) 1654 ); 1655 DBG_INIT(" STATUS_CONTROL 0x%Lx FLUSH_CTRL 0x%Lx\n", 1656 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x108), 1657 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x400) 1658 ); 1659 1660 if (IS_PLUTO(sba_dev->dev)) { 1661 sba_ioc_init_pluto(sba_dev->dev, &(sba_dev->ioc[i]), i); 1662 } else { 1663 sba_ioc_init(sba_dev->dev, &(sba_dev->ioc[i]), i); 1664 } 1665 } 1666 } 1667 1668 static void 1669 sba_common_init(struct sba_device *sba_dev) 1670 { 1671 int i; 1672 1673 /* add this one to the head of the list (order doesn't matter) 1674 ** This will be useful for debugging - especially if we get coredumps 1675 */ 1676 sba_dev->next = sba_list; 1677 sba_list = sba_dev; 1678 1679 for(i=0; i< sba_dev->num_ioc; i++) { 1680 int res_size; 1681 #ifdef DEBUG_DMB_TRAP 1682 extern void iterate_pages(unsigned long , unsigned long , 1683 void (*)(pte_t * , unsigned long), 1684 unsigned long ); 1685 void set_data_memory_break(pte_t * , unsigned long); 1686 #endif 1687 /* resource map size dictated by pdir_size */ 1688 res_size = sba_dev->ioc[i].pdir_size/sizeof(u64); /* entries */ 1689 1690 /* Second part of PIRANHA BUG */ 1691 if (piranha_bad_128k) { 1692 res_size -= (128*1024)/sizeof(u64); 1693 } 1694 1695 res_size >>= 3; /* convert bit count to byte count */ 1696 DBG_INIT("%s() res_size 0x%x\n", 1697 __func__, res_size); 1698 1699 sba_dev->ioc[i].res_size = res_size; 1700 sba_dev->ioc[i].res_map = (char *) __get_free_pages(GFP_KERNEL, get_order(res_size)); 1701 1702 #ifdef DEBUG_DMB_TRAP 1703 iterate_pages( sba_dev->ioc[i].res_map, res_size, 1704 set_data_memory_break, 0); 1705 #endif 1706 1707 if (NULL == sba_dev->ioc[i].res_map) 1708 { 1709 panic("%s:%s() could not allocate resource map\n", 1710 __FILE__, __func__ ); 1711 } 1712 1713 memset(sba_dev->ioc[i].res_map, 0, res_size); 1714 /* next available IOVP - circular search */ 1715 sba_dev->ioc[i].res_hint = (unsigned long *) 1716 &(sba_dev->ioc[i].res_map[L1_CACHE_BYTES]); 1717 1718 #ifdef ASSERT_PDIR_SANITY 1719 /* Mark first bit busy - ie no IOVA 0 */ 1720 sba_dev->ioc[i].res_map[0] = 0x80; 1721 sba_dev->ioc[i].pdir_base[0] = 0xeeffc0addbba0080ULL; 1722 #endif 1723 1724 /* Third (and last) part of PIRANHA BUG */ 1725 if (piranha_bad_128k) { 1726 /* region from +1408K to +1536 is un-usable. */ 1727 1728 int idx_start = (1408*1024/sizeof(u64)) >> 3; 1729 int idx_end = (1536*1024/sizeof(u64)) >> 3; 1730 long *p_start = (long *) &(sba_dev->ioc[i].res_map[idx_start]); 1731 long *p_end = (long *) &(sba_dev->ioc[i].res_map[idx_end]); 1732 1733 /* mark that part of the io pdir busy */ 1734 while (p_start < p_end) 1735 *p_start++ = -1; 1736 1737 } 1738 1739 #ifdef DEBUG_DMB_TRAP 1740 iterate_pages( sba_dev->ioc[i].res_map, res_size, 1741 set_data_memory_break, 0); 1742 iterate_pages( sba_dev->ioc[i].pdir_base, sba_dev->ioc[i].pdir_size, 1743 set_data_memory_break, 0); 1744 #endif 1745 1746 DBG_INIT("%s() %d res_map %x %p\n", 1747 __func__, i, res_size, sba_dev->ioc[i].res_map); 1748 } 1749 1750 spin_lock_init(&sba_dev->sba_lock); 1751 ioc_needs_fdc = boot_cpu_data.pdc.capabilities & PDC_MODEL_IOPDIR_FDC; 1752 1753 #ifdef DEBUG_SBA_INIT 1754 /* 1755 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set 1756 * (bit #61, big endian), we have to flush and sync every time 1757 * IO-PDIR is changed in Ike/Astro. 1758 */ 1759 if (ioc_needs_fdc) { 1760 printk(KERN_INFO MODULE_NAME " FDC/SYNC required.\n"); 1761 } else { 1762 printk(KERN_INFO MODULE_NAME " IOC has cache coherent PDIR.\n"); 1763 } 1764 #endif 1765 } 1766 1767 #ifdef CONFIG_PROC_FS 1768 static int sba_proc_info(struct seq_file *m, void *p) 1769 { 1770 struct sba_device *sba_dev = sba_list; 1771 struct ioc *ioc = &sba_dev->ioc[0]; /* FIXME: Multi-IOC support! */ 1772 int total_pages = (int) (ioc->res_size << 3); /* 8 bits per byte */ 1773 #ifdef SBA_COLLECT_STATS 1774 unsigned long avg = 0, min, max; 1775 #endif 1776 int i; 1777 1778 seq_printf(m, "%s rev %d.%d\n", 1779 sba_dev->name, 1780 (sba_dev->hw_rev & 0x7) + 1, 1781 (sba_dev->hw_rev & 0x18) >> 3); 1782 seq_printf(m, "IO PDIR size : %d bytes (%d entries)\n", 1783 (int)((ioc->res_size << 3) * sizeof(u64)), /* 8 bits/byte */ 1784 total_pages); 1785 1786 seq_printf(m, "Resource bitmap : %d bytes (%d pages)\n", 1787 ioc->res_size, ioc->res_size << 3); /* 8 bits per byte */ 1788 1789 seq_printf(m, "LMMIO_BASE/MASK/ROUTE %08x %08x %08x\n", 1790 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_BASE), 1791 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_MASK), 1792 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_ROUTE)); 1793 1794 for (i=0; i<4; i++) 1795 seq_printf(m, "DIR%d_BASE/MASK/ROUTE %08x %08x %08x\n", 1796 i, 1797 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_BASE + i*0x18), 1798 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_MASK + i*0x18), 1799 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_ROUTE + i*0x18)); 1800 1801 #ifdef SBA_COLLECT_STATS 1802 seq_printf(m, "IO PDIR entries : %ld free %ld used (%d%%)\n", 1803 total_pages - ioc->used_pages, ioc->used_pages, 1804 (int)(ioc->used_pages * 100 / total_pages)); 1805 1806 min = max = ioc->avg_search[0]; 1807 for (i = 0; i < SBA_SEARCH_SAMPLE; i++) { 1808 avg += ioc->avg_search[i]; 1809 if (ioc->avg_search[i] > max) max = ioc->avg_search[i]; 1810 if (ioc->avg_search[i] < min) min = ioc->avg_search[i]; 1811 } 1812 avg /= SBA_SEARCH_SAMPLE; 1813 seq_printf(m, " Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles)\n", 1814 min, avg, max); 1815 1816 seq_printf(m, "pci_map_single(): %12ld calls %12ld pages (avg %d/1000)\n", 1817 ioc->msingle_calls, ioc->msingle_pages, 1818 (int)((ioc->msingle_pages * 1000)/ioc->msingle_calls)); 1819 1820 /* KLUGE - unmap_sg calls unmap_single for each mapped page */ 1821 min = ioc->usingle_calls; 1822 max = ioc->usingle_pages - ioc->usg_pages; 1823 seq_printf(m, "pci_unmap_single: %12ld calls %12ld pages (avg %d/1000)\n", 1824 min, max, (int)((max * 1000)/min)); 1825 1826 seq_printf(m, "pci_map_sg() : %12ld calls %12ld pages (avg %d/1000)\n", 1827 ioc->msg_calls, ioc->msg_pages, 1828 (int)((ioc->msg_pages * 1000)/ioc->msg_calls)); 1829 1830 seq_printf(m, "pci_unmap_sg() : %12ld calls %12ld pages (avg %d/1000)\n", 1831 ioc->usg_calls, ioc->usg_pages, 1832 (int)((ioc->usg_pages * 1000)/ioc->usg_calls)); 1833 #endif 1834 1835 return 0; 1836 } 1837 1838 static int 1839 sba_proc_open(struct inode *i, struct file *f) 1840 { 1841 return single_open(f, &sba_proc_info, NULL); 1842 } 1843 1844 static const struct file_operations sba_proc_fops = { 1845 .owner = THIS_MODULE, 1846 .open = sba_proc_open, 1847 .read = seq_read, 1848 .llseek = seq_lseek, 1849 .release = single_release, 1850 }; 1851 1852 static int 1853 sba_proc_bitmap_info(struct seq_file *m, void *p) 1854 { 1855 struct sba_device *sba_dev = sba_list; 1856 struct ioc *ioc = &sba_dev->ioc[0]; /* FIXME: Multi-IOC support! */ 1857 1858 seq_hex_dump(m, " ", DUMP_PREFIX_NONE, 32, 4, ioc->res_map, 1859 ioc->res_size, false); 1860 seq_putc(m, '\n'); 1861 1862 return 0; 1863 } 1864 1865 static int 1866 sba_proc_bitmap_open(struct inode *i, struct file *f) 1867 { 1868 return single_open(f, &sba_proc_bitmap_info, NULL); 1869 } 1870 1871 static const struct file_operations sba_proc_bitmap_fops = { 1872 .owner = THIS_MODULE, 1873 .open = sba_proc_bitmap_open, 1874 .read = seq_read, 1875 .llseek = seq_lseek, 1876 .release = single_release, 1877 }; 1878 #endif /* CONFIG_PROC_FS */ 1879 1880 static struct parisc_device_id sba_tbl[] = { 1881 { HPHW_IOA, HVERSION_REV_ANY_ID, ASTRO_RUNWAY_PORT, 0xb }, 1882 { HPHW_BCPORT, HVERSION_REV_ANY_ID, IKE_MERCED_PORT, 0xc }, 1883 { HPHW_BCPORT, HVERSION_REV_ANY_ID, REO_MERCED_PORT, 0xc }, 1884 { HPHW_BCPORT, HVERSION_REV_ANY_ID, REOG_MERCED_PORT, 0xc }, 1885 { HPHW_IOA, HVERSION_REV_ANY_ID, PLUTO_MCKINLEY_PORT, 0xc }, 1886 { 0, } 1887 }; 1888 1889 static int sba_driver_callback(struct parisc_device *); 1890 1891 static struct parisc_driver sba_driver = { 1892 .name = MODULE_NAME, 1893 .id_table = sba_tbl, 1894 .probe = sba_driver_callback, 1895 }; 1896 1897 /* 1898 ** Determine if sba should claim this chip (return 0) or not (return 1). 1899 ** If so, initialize the chip and tell other partners in crime they 1900 ** have work to do. 1901 */ 1902 static int sba_driver_callback(struct parisc_device *dev) 1903 { 1904 struct sba_device *sba_dev; 1905 u32 func_class; 1906 int i; 1907 char *version; 1908 void __iomem *sba_addr = ioremap_nocache(dev->hpa.start, SBA_FUNC_SIZE); 1909 #ifdef CONFIG_PROC_FS 1910 struct proc_dir_entry *root; 1911 #endif 1912 1913 sba_dump_ranges(sba_addr); 1914 1915 /* Read HW Rev First */ 1916 func_class = READ_REG(sba_addr + SBA_FCLASS); 1917 1918 if (IS_ASTRO(dev)) { 1919 unsigned long fclass; 1920 static char astro_rev[]="Astro ?.?"; 1921 1922 /* Astro is broken...Read HW Rev First */ 1923 fclass = READ_REG(sba_addr); 1924 1925 astro_rev[6] = '1' + (char) (fclass & 0x7); 1926 astro_rev[8] = '0' + (char) ((fclass & 0x18) >> 3); 1927 version = astro_rev; 1928 1929 } else if (IS_IKE(dev)) { 1930 static char ike_rev[] = "Ike rev ?"; 1931 ike_rev[8] = '0' + (char) (func_class & 0xff); 1932 version = ike_rev; 1933 } else if (IS_PLUTO(dev)) { 1934 static char pluto_rev[]="Pluto ?.?"; 1935 pluto_rev[6] = '0' + (char) ((func_class & 0xf0) >> 4); 1936 pluto_rev[8] = '0' + (char) (func_class & 0x0f); 1937 version = pluto_rev; 1938 } else { 1939 static char reo_rev[] = "REO rev ?"; 1940 reo_rev[8] = '0' + (char) (func_class & 0xff); 1941 version = reo_rev; 1942 } 1943 1944 if (!global_ioc_cnt) { 1945 global_ioc_cnt = count_parisc_driver(&sba_driver); 1946 1947 /* Astro and Pluto have one IOC per SBA */ 1948 if ((!IS_ASTRO(dev)) || (!IS_PLUTO(dev))) 1949 global_ioc_cnt *= 2; 1950 } 1951 1952 printk(KERN_INFO "%s found %s at 0x%llx\n", 1953 MODULE_NAME, version, (unsigned long long)dev->hpa.start); 1954 1955 sba_dev = kzalloc(sizeof(struct sba_device), GFP_KERNEL); 1956 if (!sba_dev) { 1957 printk(KERN_ERR MODULE_NAME " - couldn't alloc sba_device\n"); 1958 return -ENOMEM; 1959 } 1960 1961 parisc_set_drvdata(dev, sba_dev); 1962 1963 for(i=0; i<MAX_IOC; i++) 1964 spin_lock_init(&(sba_dev->ioc[i].res_lock)); 1965 1966 sba_dev->dev = dev; 1967 sba_dev->hw_rev = func_class; 1968 sba_dev->name = dev->name; 1969 sba_dev->sba_hpa = sba_addr; 1970 1971 sba_get_pat_resources(sba_dev); 1972 sba_hw_init(sba_dev); 1973 sba_common_init(sba_dev); 1974 1975 hppa_dma_ops = &sba_ops; 1976 1977 #ifdef CONFIG_PROC_FS 1978 switch (dev->id.hversion) { 1979 case PLUTO_MCKINLEY_PORT: 1980 root = proc_mckinley_root; 1981 break; 1982 case ASTRO_RUNWAY_PORT: 1983 case IKE_MERCED_PORT: 1984 default: 1985 root = proc_runway_root; 1986 break; 1987 } 1988 1989 proc_create("sba_iommu", 0, root, &sba_proc_fops); 1990 proc_create("sba_iommu-bitmap", 0, root, &sba_proc_bitmap_fops); 1991 #endif 1992 1993 parisc_has_iommu(); 1994 return 0; 1995 } 1996 1997 /* 1998 ** One time initialization to let the world know the SBA was found. 1999 ** This is the only routine which is NOT static. 2000 ** Must be called exactly once before pci_init(). 2001 */ 2002 void __init sba_init(void) 2003 { 2004 register_parisc_driver(&sba_driver); 2005 } 2006 2007 2008 /** 2009 * sba_get_iommu - Assign the iommu pointer for the pci bus controller. 2010 * @dev: The parisc device. 2011 * 2012 * Returns the appropriate IOMMU data for the given parisc PCI controller. 2013 * This is cached and used later for PCI DMA Mapping. 2014 */ 2015 void * sba_get_iommu(struct parisc_device *pci_hba) 2016 { 2017 struct parisc_device *sba_dev = parisc_parent(pci_hba); 2018 struct sba_device *sba = dev_get_drvdata(&sba_dev->dev); 2019 char t = sba_dev->id.hw_type; 2020 int iocnum = (pci_hba->hw_path >> 3); /* rope # */ 2021 2022 WARN_ON((t != HPHW_IOA) && (t != HPHW_BCPORT)); 2023 2024 return &(sba->ioc[iocnum]); 2025 } 2026 2027 2028 /** 2029 * sba_directed_lmmio - return first directed LMMIO range routed to rope 2030 * @pa_dev: The parisc device. 2031 * @r: resource PCI host controller wants start/end fields assigned. 2032 * 2033 * For the given parisc PCI controller, determine if any direct ranges 2034 * are routed down the corresponding rope. 2035 */ 2036 void sba_directed_lmmio(struct parisc_device *pci_hba, struct resource *r) 2037 { 2038 struct parisc_device *sba_dev = parisc_parent(pci_hba); 2039 struct sba_device *sba = dev_get_drvdata(&sba_dev->dev); 2040 char t = sba_dev->id.hw_type; 2041 int i; 2042 int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1)); /* rope # */ 2043 2044 BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT)); 2045 2046 r->start = r->end = 0; 2047 2048 /* Astro has 4 directed ranges. Not sure about Ike/Pluto/et al */ 2049 for (i=0; i<4; i++) { 2050 int base, size; 2051 void __iomem *reg = sba->sba_hpa + i*0x18; 2052 2053 base = READ_REG32(reg + LMMIO_DIRECT0_BASE); 2054 if ((base & 1) == 0) 2055 continue; /* not enabled */ 2056 2057 size = READ_REG32(reg + LMMIO_DIRECT0_ROUTE); 2058 2059 if ((size & (ROPES_PER_IOC-1)) != rope) 2060 continue; /* directed down different rope */ 2061 2062 r->start = (base & ~1UL) | PCI_F_EXTEND; 2063 size = ~ READ_REG32(reg + LMMIO_DIRECT0_MASK); 2064 r->end = r->start + size; 2065 r->flags = IORESOURCE_MEM; 2066 } 2067 } 2068 2069 2070 /** 2071 * sba_distributed_lmmio - return portion of distributed LMMIO range 2072 * @pa_dev: The parisc device. 2073 * @r: resource PCI host controller wants start/end fields assigned. 2074 * 2075 * For the given parisc PCI controller, return portion of distributed LMMIO 2076 * range. The distributed LMMIO is always present and it's just a question 2077 * of the base address and size of the range. 2078 */ 2079 void sba_distributed_lmmio(struct parisc_device *pci_hba, struct resource *r ) 2080 { 2081 struct parisc_device *sba_dev = parisc_parent(pci_hba); 2082 struct sba_device *sba = dev_get_drvdata(&sba_dev->dev); 2083 char t = sba_dev->id.hw_type; 2084 int base, size; 2085 int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1)); /* rope # */ 2086 2087 BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT)); 2088 2089 r->start = r->end = 0; 2090 2091 base = READ_REG32(sba->sba_hpa + LMMIO_DIST_BASE); 2092 if ((base & 1) == 0) { 2093 BUG(); /* Gah! Distr Range wasn't enabled! */ 2094 return; 2095 } 2096 2097 r->start = (base & ~1UL) | PCI_F_EXTEND; 2098 2099 size = (~READ_REG32(sba->sba_hpa + LMMIO_DIST_MASK)) / ROPES_PER_IOC; 2100 r->start += rope * (size + 1); /* adjust base for this rope */ 2101 r->end = r->start + size; 2102 r->flags = IORESOURCE_MEM; 2103 } 2104