1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 ** ccio-dma.c: 4 ** DMA management routines for first generation cache-coherent machines. 5 ** Program U2/Uturn in "Virtual Mode" and use the I/O MMU. 6 ** 7 ** (c) Copyright 2000 Grant Grundler 8 ** (c) Copyright 2000 Ryan Bradetich 9 ** (c) Copyright 2000 Hewlett-Packard Company 10 ** 11 ** "Real Mode" operation refers to U2/Uturn chip operation. 12 ** U2/Uturn were designed to perform coherency checks w/o using 13 ** the I/O MMU - basically what x86 does. 14 ** 15 ** Drawbacks of using Real Mode are: 16 ** o outbound DMA is slower - U2 won't prefetch data (GSC+ XQL signal). 17 ** o Inbound DMA less efficient - U2 can't use DMA_FAST attribute. 18 ** o Ability to do scatter/gather in HW is lost. 19 ** o Doesn't work under PCX-U/U+ machines since they didn't follow 20 ** the coherency design originally worked out. Only PCX-W does. 21 */ 22 23 #include <linux/types.h> 24 #include <linux/kernel.h> 25 #include <linux/init.h> 26 #include <linux/mm.h> 27 #include <linux/spinlock.h> 28 #include <linux/slab.h> 29 #include <linux/string.h> 30 #include <linux/pci.h> 31 #include <linux/reboot.h> 32 #include <linux/proc_fs.h> 33 #include <linux/seq_file.h> 34 #include <linux/dma-map-ops.h> 35 #include <linux/scatterlist.h> 36 #include <linux/iommu-helper.h> 37 #include <linux/export.h> 38 39 #include <asm/byteorder.h> 40 #include <asm/cache.h> /* for L1_CACHE_BYTES */ 41 #include <linux/uaccess.h> 42 #include <asm/page.h> 43 #include <asm/dma.h> 44 #include <asm/io.h> 45 #include <asm/hardware.h> /* for register_module() */ 46 #include <asm/parisc-device.h> 47 48 #include "iommu.h" 49 50 /* 51 ** Choose "ccio" since that's what HP-UX calls it. 52 ** Make it easier for folks to migrate from one to the other :^) 53 */ 54 #define MODULE_NAME "ccio" 55 56 #undef DEBUG_CCIO_RES 57 #undef DEBUG_CCIO_RUN 58 #undef DEBUG_CCIO_INIT 59 #undef DEBUG_CCIO_RUN_SG 60 61 #ifdef CONFIG_PROC_FS 62 /* depends on proc fs support. But costs CPU performance. */ 63 #undef CCIO_COLLECT_STATS 64 #endif 65 66 #ifdef DEBUG_CCIO_INIT 67 #define DBG_INIT(x...) printk(x) 68 #else 69 #define DBG_INIT(x...) 70 #endif 71 72 #ifdef DEBUG_CCIO_RUN 73 #define DBG_RUN(x...) printk(x) 74 #else 75 #define DBG_RUN(x...) 76 #endif 77 78 #ifdef DEBUG_CCIO_RES 79 #define DBG_RES(x...) printk(x) 80 #else 81 #define DBG_RES(x...) 82 #endif 83 84 #ifdef DEBUG_CCIO_RUN_SG 85 #define DBG_RUN_SG(x...) printk(x) 86 #else 87 #define DBG_RUN_SG(x...) 88 #endif 89 90 #define WRITE_U32(value, addr) __raw_writel(value, addr) 91 #define READ_U32(addr) __raw_readl(addr) 92 93 #define U2_IOA_RUNWAY 0x580 94 #define U2_BC_GSC 0x501 95 #define UTURN_IOA_RUNWAY 0x581 96 #define UTURN_BC_GSC 0x502 97 98 #define IOA_NORMAL_MODE 0x00020080 /* IO_CONTROL to turn on CCIO */ 99 #define CMD_TLB_DIRECT_WRITE 35 /* IO_COMMAND for I/O TLB Writes */ 100 #define CMD_TLB_PURGE 33 /* IO_COMMAND to Purge I/O TLB entry */ 101 102 struct ioa_registers { 103 /* Runway Supervisory Set */ 104 int32_t unused1[12]; 105 uint32_t io_command; /* Offset 12 */ 106 uint32_t io_status; /* Offset 13 */ 107 uint32_t io_control; /* Offset 14 */ 108 int32_t unused2[1]; 109 110 /* Runway Auxiliary Register Set */ 111 uint32_t io_err_resp; /* Offset 0 */ 112 uint32_t io_err_info; /* Offset 1 */ 113 uint32_t io_err_req; /* Offset 2 */ 114 uint32_t io_err_resp_hi; /* Offset 3 */ 115 uint32_t io_tlb_entry_m; /* Offset 4 */ 116 uint32_t io_tlb_entry_l; /* Offset 5 */ 117 uint32_t unused3[1]; 118 uint32_t io_pdir_base; /* Offset 7 */ 119 uint32_t io_io_low_hv; /* Offset 8 */ 120 uint32_t io_io_high_hv; /* Offset 9 */ 121 uint32_t unused4[1]; 122 uint32_t io_chain_id_mask; /* Offset 11 */ 123 uint32_t unused5[2]; 124 uint32_t io_io_low; /* Offset 14 */ 125 uint32_t io_io_high; /* Offset 15 */ 126 }; 127 128 /* 129 ** IOA Registers 130 ** ------------- 131 ** 132 ** Runway IO_CONTROL Register (+0x38) 133 ** 134 ** The Runway IO_CONTROL register controls the forwarding of transactions. 135 ** 136 ** | 0 ... 13 | 14 15 | 16 ... 21 | 22 | 23 24 | 25 ... 31 | 137 ** | HV | TLB | reserved | HV | mode | reserved | 138 ** 139 ** o mode field indicates the address translation of transactions 140 ** forwarded from Runway to GSC+: 141 ** Mode Name Value Definition 142 ** Off (default) 0 Opaque to matching addresses. 143 ** Include 1 Transparent for matching addresses. 144 ** Peek 3 Map matching addresses. 145 ** 146 ** + "Off" mode: Runway transactions which match the I/O range 147 ** specified by the IO_IO_LOW/IO_IO_HIGH registers will be ignored. 148 ** + "Include" mode: all addresses within the I/O range specified 149 ** by the IO_IO_LOW and IO_IO_HIGH registers are transparently 150 ** forwarded. This is the I/O Adapter's normal operating mode. 151 ** + "Peek" mode: used during system configuration to initialize the 152 ** GSC+ bus. Runway Write_Shorts in the address range specified by 153 ** IO_IO_LOW and IO_IO_HIGH are forwarded through the I/O Adapter 154 ** *AND* the GSC+ address is remapped to the Broadcast Physical 155 ** Address space by setting the 14 high order address bits of the 156 ** 32 bit GSC+ address to ones. 157 ** 158 ** o TLB field affects transactions which are forwarded from GSC+ to Runway. 159 ** "Real" mode is the poweron default. 160 ** 161 ** TLB Mode Value Description 162 ** Real 0 No TLB translation. Address is directly mapped and the 163 ** virtual address is composed of selected physical bits. 164 ** Error 1 Software fills the TLB manually. 165 ** Normal 2 IOA fetches IO TLB misses from IO PDIR (in host memory). 166 ** 167 ** 168 ** IO_IO_LOW_HV +0x60 (HV dependent) 169 ** IO_IO_HIGH_HV +0x64 (HV dependent) 170 ** IO_IO_LOW +0x78 (Architected register) 171 ** IO_IO_HIGH +0x7c (Architected register) 172 ** 173 ** IO_IO_LOW and IO_IO_HIGH set the lower and upper bounds of the 174 ** I/O Adapter address space, respectively. 175 ** 176 ** 0 ... 7 | 8 ... 15 | 16 ... 31 | 177 ** 11111111 | 11111111 | address | 178 ** 179 ** Each LOW/HIGH pair describes a disjoint address space region. 180 ** (2 per GSC+ port). Each incoming Runway transaction address is compared 181 ** with both sets of LOW/HIGH registers. If the address is in the range 182 ** greater than or equal to IO_IO_LOW and less than IO_IO_HIGH the transaction 183 ** for forwarded to the respective GSC+ bus. 184 ** Specify IO_IO_LOW equal to or greater than IO_IO_HIGH to avoid specifying 185 ** an address space region. 186 ** 187 ** In order for a Runway address to reside within GSC+ extended address space: 188 ** Runway Address [0:7] must identically compare to 8'b11111111 189 ** Runway Address [8:11] must be equal to IO_IO_LOW(_HV)[16:19] 190 ** Runway Address [12:23] must be greater than or equal to 191 ** IO_IO_LOW(_HV)[20:31] and less than IO_IO_HIGH(_HV)[20:31]. 192 ** Runway Address [24:39] is not used in the comparison. 193 ** 194 ** When the Runway transaction is forwarded to GSC+, the GSC+ address is 195 ** as follows: 196 ** GSC+ Address[0:3] 4'b1111 197 ** GSC+ Address[4:29] Runway Address[12:37] 198 ** GSC+ Address[30:31] 2'b00 199 ** 200 ** All 4 Low/High registers must be initialized (by PDC) once the lower bus 201 ** is interrogated and address space is defined. The operating system will 202 ** modify the architectural IO_IO_LOW and IO_IO_HIGH registers following 203 ** the PDC initialization. However, the hardware version dependent IO_IO_LOW 204 ** and IO_IO_HIGH registers should not be subsequently altered by the OS. 205 ** 206 ** Writes to both sets of registers will take effect immediately, bypassing 207 ** the queues, which ensures that subsequent Runway transactions are checked 208 ** against the updated bounds values. However reads are queued, introducing 209 ** the possibility of a read being bypassed by a subsequent write to the same 210 ** register. This sequence can be avoided by having software wait for read 211 ** returns before issuing subsequent writes. 212 */ 213 214 struct ioc { 215 struct ioa_registers __iomem *ioc_regs; /* I/O MMU base address */ 216 u8 *res_map; /* resource map, bit == pdir entry */ 217 __le64 *pdir_base; /* physical base address */ 218 u32 pdir_size; /* bytes, function of IOV Space size */ 219 u32 res_hint; /* next available IOVP - 220 circular search */ 221 u32 res_size; /* size of resource map in bytes */ 222 spinlock_t res_lock; 223 224 #ifdef CCIO_COLLECT_STATS 225 #define CCIO_SEARCH_SAMPLE 0x100 226 unsigned long avg_search[CCIO_SEARCH_SAMPLE]; 227 unsigned long avg_idx; /* current index into avg_search */ 228 unsigned long used_pages; 229 unsigned long msingle_calls; 230 unsigned long msingle_pages; 231 unsigned long msg_calls; 232 unsigned long msg_pages; 233 unsigned long usingle_calls; 234 unsigned long usingle_pages; 235 unsigned long usg_calls; 236 unsigned long usg_pages; 237 #endif 238 unsigned short cujo20_bug; 239 240 /* STUFF We don't need in performance path */ 241 u32 chainid_shift; /* specify bit location of chain_id */ 242 struct ioc *next; /* Linked list of discovered iocs */ 243 const char *name; /* device name from firmware */ 244 unsigned int hw_path; /* the hardware path this ioc is associatd with */ 245 struct pci_dev *fake_pci_dev; /* the fake pci_dev for non-pci devs */ 246 struct resource mmio_region[2]; /* The "routed" MMIO regions */ 247 }; 248 249 static struct ioc *ioc_list; 250 static int ioc_count; 251 252 /************************************************************** 253 * 254 * I/O Pdir Resource Management 255 * 256 * Bits set in the resource map are in use. 257 * Each bit can represent a number of pages. 258 * LSbs represent lower addresses (IOVA's). 259 * 260 * This was copied from sba_iommu.c. Don't try to unify 261 * the two resource managers unless a way to have different 262 * allocation policies is also adjusted. We'd like to avoid 263 * I/O TLB thrashing by having resource allocation policy 264 * match the I/O TLB replacement policy. 265 * 266 ***************************************************************/ 267 #define IOVP_SIZE PAGE_SIZE 268 #define IOVP_SHIFT PAGE_SHIFT 269 #define IOVP_MASK PAGE_MASK 270 271 /* Convert from IOVP to IOVA and vice versa. */ 272 #define CCIO_IOVA(iovp,offset) ((iovp) | (offset)) 273 #define CCIO_IOVP(iova) ((iova) & IOVP_MASK) 274 275 #define PDIR_INDEX(iovp) ((iovp)>>IOVP_SHIFT) 276 #define MKIOVP(pdir_idx) ((long)(pdir_idx) << IOVP_SHIFT) 277 #define MKIOVA(iovp,offset) (dma_addr_t)((long)iovp | (long)offset) 278 279 /* 280 ** Don't worry about the 150% average search length on a miss. 281 ** If the search wraps around, and passes the res_hint, it will 282 ** cause the kernel to panic anyhow. 283 */ 284 #define CCIO_SEARCH_LOOP(ioc, res_idx, mask, size) \ 285 for (; res_ptr < res_end; ++res_ptr) { \ 286 int ret;\ 287 unsigned int idx;\ 288 idx = (unsigned int)((unsigned long)res_ptr - (unsigned long)ioc->res_map); \ 289 ret = iommu_is_span_boundary(idx << 3, pages_needed, 0, boundary_size);\ 290 if ((0 == (*res_ptr & mask)) && !ret) { \ 291 *res_ptr |= mask; \ 292 res_idx = idx;\ 293 ioc->res_hint = res_idx + (size >> 3); \ 294 goto resource_found; \ 295 } \ 296 } 297 298 #define CCIO_FIND_FREE_MAPPING(ioa, res_idx, mask, size) \ 299 u##size *res_ptr = (u##size *)&((ioc)->res_map[ioa->res_hint & ~((size >> 3) - 1)]); \ 300 u##size *res_end = (u##size *)&(ioc)->res_map[ioa->res_size]; \ 301 CCIO_SEARCH_LOOP(ioc, res_idx, mask, size); \ 302 res_ptr = (u##size *)&(ioc)->res_map[0]; \ 303 CCIO_SEARCH_LOOP(ioa, res_idx, mask, size); 304 305 /* 306 ** Find available bit in this ioa's resource map. 307 ** Use a "circular" search: 308 ** o Most IOVA's are "temporary" - avg search time should be small. 309 ** o keep a history of what happened for debugging 310 ** o KISS. 311 ** 312 ** Perf optimizations: 313 ** o search for log2(size) bits at a time. 314 ** o search for available resource bits using byte/word/whatever. 315 ** o use different search for "large" (eg > 4 pages) or "very large" 316 ** (eg > 16 pages) mappings. 317 */ 318 319 /** 320 * ccio_alloc_range - Allocate pages in the ioc's resource map. 321 * @ioc: The I/O Controller. 322 * @dev: The PCI device. 323 * @size: The requested number of bytes to be mapped into the 324 * I/O Pdir... 325 * 326 * This function searches the resource map of the ioc to locate a range 327 * of available pages for the requested size. 328 */ 329 static int 330 ccio_alloc_range(struct ioc *ioc, struct device *dev, size_t size) 331 { 332 unsigned int pages_needed = size >> IOVP_SHIFT; 333 unsigned int res_idx; 334 unsigned long boundary_size; 335 #ifdef CCIO_COLLECT_STATS 336 unsigned long cr_start = mfctl(16); 337 #endif 338 339 BUG_ON(pages_needed == 0); 340 BUG_ON((pages_needed * IOVP_SIZE) > DMA_CHUNK_SIZE); 341 342 DBG_RES("%s() size: %zu pages_needed %d\n", 343 __func__, size, pages_needed); 344 345 /* 346 ** "seek and ye shall find"...praying never hurts either... 347 ** ggg sacrifices another 710 to the computer gods. 348 */ 349 350 boundary_size = dma_get_seg_boundary_nr_pages(dev, IOVP_SHIFT); 351 352 if (pages_needed <= 8) { 353 /* 354 * LAN traffic will not thrash the TLB IFF the same NIC 355 * uses 8 adjacent pages to map separate payload data. 356 * ie the same byte in the resource bit map. 357 */ 358 #if 0 359 /* FIXME: bit search should shift it's way through 360 * an unsigned long - not byte at a time. As it is now, 361 * we effectively allocate this byte to this mapping. 362 */ 363 unsigned long mask = ~(~0UL >> pages_needed); 364 CCIO_FIND_FREE_MAPPING(ioc, res_idx, mask, 8); 365 #else 366 CCIO_FIND_FREE_MAPPING(ioc, res_idx, 0xff, 8); 367 #endif 368 } else if (pages_needed <= 16) { 369 CCIO_FIND_FREE_MAPPING(ioc, res_idx, 0xffff, 16); 370 } else if (pages_needed <= 32) { 371 CCIO_FIND_FREE_MAPPING(ioc, res_idx, ~(unsigned int)0, 32); 372 #ifdef __LP64__ 373 } else if (pages_needed <= 64) { 374 CCIO_FIND_FREE_MAPPING(ioc, res_idx, ~0UL, 64); 375 #endif 376 } else { 377 panic("%s: %s() Too many pages to map. pages_needed: %u\n", 378 __FILE__, __func__, pages_needed); 379 } 380 381 panic("%s: %s() I/O MMU is out of mapping resources.\n", __FILE__, 382 __func__); 383 384 resource_found: 385 386 DBG_RES("%s() res_idx %d res_hint: %d\n", 387 __func__, res_idx, ioc->res_hint); 388 389 #ifdef CCIO_COLLECT_STATS 390 { 391 unsigned long cr_end = mfctl(16); 392 unsigned long tmp = cr_end - cr_start; 393 /* check for roll over */ 394 cr_start = (cr_end < cr_start) ? -(tmp) : (tmp); 395 } 396 ioc->avg_search[ioc->avg_idx++] = cr_start; 397 ioc->avg_idx &= CCIO_SEARCH_SAMPLE - 1; 398 ioc->used_pages += pages_needed; 399 #endif 400 /* 401 ** return the bit address. 402 */ 403 return res_idx << 3; 404 } 405 406 #define CCIO_FREE_MAPPINGS(ioc, res_idx, mask, size) \ 407 u##size *res_ptr = (u##size *)&((ioc)->res_map[res_idx]); \ 408 BUG_ON((*res_ptr & mask) != mask); \ 409 *res_ptr &= ~(mask); 410 411 /** 412 * ccio_free_range - Free pages from the ioc's resource map. 413 * @ioc: The I/O Controller. 414 * @iova: The I/O Virtual Address. 415 * @pages_mapped: The requested number of pages to be freed from the 416 * I/O Pdir. 417 * 418 * This function frees the resouces allocated for the iova. 419 */ 420 static void 421 ccio_free_range(struct ioc *ioc, dma_addr_t iova, unsigned long pages_mapped) 422 { 423 unsigned long iovp = CCIO_IOVP(iova); 424 unsigned int res_idx = PDIR_INDEX(iovp) >> 3; 425 426 BUG_ON(pages_mapped == 0); 427 BUG_ON((pages_mapped * IOVP_SIZE) > DMA_CHUNK_SIZE); 428 BUG_ON(pages_mapped > BITS_PER_LONG); 429 430 DBG_RES("%s(): res_idx: %d pages_mapped %lu\n", 431 __func__, res_idx, pages_mapped); 432 433 #ifdef CCIO_COLLECT_STATS 434 ioc->used_pages -= pages_mapped; 435 #endif 436 437 if(pages_mapped <= 8) { 438 #if 0 439 /* see matching comments in alloc_range */ 440 unsigned long mask = ~(~0UL >> pages_mapped); 441 CCIO_FREE_MAPPINGS(ioc, res_idx, mask, 8); 442 #else 443 CCIO_FREE_MAPPINGS(ioc, res_idx, 0xffUL, 8); 444 #endif 445 } else if(pages_mapped <= 16) { 446 CCIO_FREE_MAPPINGS(ioc, res_idx, 0xffffUL, 16); 447 } else if(pages_mapped <= 32) { 448 CCIO_FREE_MAPPINGS(ioc, res_idx, ~(unsigned int)0, 32); 449 #ifdef __LP64__ 450 } else if(pages_mapped <= 64) { 451 CCIO_FREE_MAPPINGS(ioc, res_idx, ~0UL, 64); 452 #endif 453 } else { 454 panic("%s:%s() Too many pages to unmap.\n", __FILE__, 455 __func__); 456 } 457 } 458 459 /**************************************************************** 460 ** 461 ** CCIO dma_ops support routines 462 ** 463 *****************************************************************/ 464 465 typedef unsigned long space_t; 466 #define KERNEL_SPACE 0 467 468 /* 469 ** DMA "Page Type" and Hints 470 ** o if SAFE_DMA isn't set, mapping is for FAST_DMA. SAFE_DMA should be 471 ** set for subcacheline DMA transfers since we don't want to damage the 472 ** other part of a cacheline. 473 ** o SAFE_DMA must be set for "memory" allocated via pci_alloc_consistent(). 474 ** This bit tells U2 to do R/M/W for partial cachelines. "Streaming" 475 ** data can avoid this if the mapping covers full cache lines. 476 ** o STOP_MOST is needed for atomicity across cachelines. 477 ** Apparently only "some EISA devices" need this. 478 ** Using CONFIG_ISA is hack. Only the IOA with EISA under it needs 479 ** to use this hint iff the EISA devices needs this feature. 480 ** According to the U2 ERS, STOP_MOST enabled pages hurt performance. 481 ** o PREFETCH should *not* be set for cases like Multiple PCI devices 482 ** behind GSCtoPCI (dino) bus converter. Only one cacheline per GSC 483 ** device can be fetched and multiply DMA streams will thrash the 484 ** prefetch buffer and burn memory bandwidth. See 6.7.3 "Prefetch Rules 485 ** and Invalidation of Prefetch Entries". 486 ** 487 ** FIXME: the default hints need to be per GSC device - not global. 488 ** 489 ** HP-UX dorks: linux device driver programming model is totally different 490 ** than HP-UX's. HP-UX always sets HINT_PREFETCH since it's drivers 491 ** do special things to work on non-coherent platforms...linux has to 492 ** be much more careful with this. 493 */ 494 #define IOPDIR_VALID 0x01UL 495 #define HINT_SAFE_DMA 0x02UL /* used for pci_alloc_consistent() pages */ 496 #ifdef CONFIG_EISA 497 #define HINT_STOP_MOST 0x04UL /* LSL support */ 498 #else 499 #define HINT_STOP_MOST 0x00UL /* only needed for "some EISA devices" */ 500 #endif 501 #define HINT_UDPATE_ENB 0x08UL /* not used/supported by U2 */ 502 #define HINT_PREFETCH 0x10UL /* for outbound pages which are not SAFE */ 503 504 505 /* 506 ** Use direction (ie PCI_DMA_TODEVICE) to pick hint. 507 ** ccio_alloc_consistent() depends on this to get SAFE_DMA 508 ** when it passes in BIDIRECTIONAL flag. 509 */ 510 static u32 hint_lookup[] = { 511 [DMA_BIDIRECTIONAL] = HINT_STOP_MOST | HINT_SAFE_DMA | IOPDIR_VALID, 512 [DMA_TO_DEVICE] = HINT_STOP_MOST | HINT_PREFETCH | IOPDIR_VALID, 513 [DMA_FROM_DEVICE] = HINT_STOP_MOST | IOPDIR_VALID, 514 }; 515 516 /** 517 * ccio_io_pdir_entry - Initialize an I/O Pdir. 518 * @pdir_ptr: A pointer into I/O Pdir. 519 * @sid: The Space Identifier. 520 * @pba: The physical address. 521 * @hints: The DMA Hint. 522 * 523 * Given a physical address (pba, arg2) and space id, (sid, arg1), 524 * load the I/O PDIR entry pointed to by pdir_ptr (arg0). Each IO Pdir 525 * entry consists of 8 bytes as shown below (MSB == bit 0): 526 * 527 * 528 * WORD 0: 529 * +------+----------------+-----------------------------------------------+ 530 * | Phys | Virtual Index | Phys | 531 * | 0:3 | 0:11 | 4:19 | 532 * |4 bits| 12 bits | 16 bits | 533 * +------+----------------+-----------------------------------------------+ 534 * WORD 1: 535 * +-----------------------+-----------------------------------------------+ 536 * | Phys | Rsvd | Prefetch |Update |Rsvd |Lock |Safe |Valid | 537 * | 20:39 | | Enable |Enable | |Enable|DMA | | 538 * | 20 bits | 5 bits | 1 bit |1 bit |2 bits|1 bit |1 bit |1 bit | 539 * +-----------------------+-----------------------------------------------+ 540 * 541 * The virtual index field is filled with the results of the LCI 542 * (Load Coherence Index) instruction. The 8 bits used for the virtual 543 * index are bits 12:19 of the value returned by LCI. 544 */ 545 static void 546 ccio_io_pdir_entry(__le64 *pdir_ptr, space_t sid, phys_addr_t pba, 547 unsigned long hints) 548 { 549 register unsigned long pa; 550 register unsigned long ci; /* coherent index */ 551 552 /* We currently only support kernel addresses */ 553 BUG_ON(sid != KERNEL_SPACE); 554 555 /* 556 ** WORD 1 - low order word 557 ** "hints" parm includes the VALID bit! 558 ** "dep" clobbers the physical address offset bits as well. 559 */ 560 pa = pba; 561 asm volatile("depw %1,31,12,%0" : "+r" (pa) : "r" (hints)); 562 ((u32 *)pdir_ptr)[1] = (u32) pa; 563 564 /* 565 ** WORD 0 - high order word 566 */ 567 568 #ifdef __LP64__ 569 /* 570 ** get bits 12:15 of physical address 571 ** shift bits 16:31 of physical address 572 ** and deposit them 573 */ 574 asm volatile ("extrd,u %1,15,4,%0" : "=r" (ci) : "r" (pa)); 575 asm volatile ("extrd,u %1,31,16,%0" : "+r" (pa) : "r" (pa)); 576 asm volatile ("depd %1,35,4,%0" : "+r" (pa) : "r" (ci)); 577 #else 578 pa = 0; 579 #endif 580 /* 581 ** get CPU coherency index bits 582 ** Grab virtual index [0:11] 583 ** Deposit virt_idx bits into I/O PDIR word 584 */ 585 asm volatile ("lci %%r0(%1), %0" : "=r" (ci) : "r" (phys_to_virt(pba))); 586 asm volatile ("extru %1,19,12,%0" : "+r" (ci) : "r" (ci)); 587 asm volatile ("depw %1,15,12,%0" : "+r" (pa) : "r" (ci)); 588 589 ((u32 *)pdir_ptr)[0] = (u32) pa; 590 591 592 /* FIXME: PCX_W platforms don't need FDC/SYNC. (eg C360) 593 ** PCX-U/U+ do. (eg C200/C240) 594 ** PCX-T'? Don't know. (eg C110 or similar K-class) 595 ** 596 ** See PDC_MODEL/option 0/SW_CAP word for "Non-coherent IO-PDIR bit". 597 ** 598 ** "Since PCX-U employs an offset hash that is incompatible with 599 ** the real mode coherence index generation of U2, the PDIR entry 600 ** must be flushed to memory to retain coherence." 601 */ 602 asm_io_fdc(pdir_ptr); 603 asm_io_sync(); 604 } 605 606 /** 607 * ccio_clear_io_tlb - Remove stale entries from the I/O TLB. 608 * @ioc: The I/O Controller. 609 * @iovp: The I/O Virtual Page. 610 * @byte_cnt: The requested number of bytes to be freed from the I/O Pdir. 611 * 612 * Purge invalid I/O PDIR entries from the I/O TLB. 613 * 614 * FIXME: Can we change the byte_cnt to pages_mapped? 615 */ 616 static void 617 ccio_clear_io_tlb(struct ioc *ioc, dma_addr_t iovp, size_t byte_cnt) 618 { 619 u32 chain_size = 1 << ioc->chainid_shift; 620 621 iovp &= IOVP_MASK; /* clear offset bits, just want pagenum */ 622 byte_cnt += chain_size; 623 624 while(byte_cnt > chain_size) { 625 WRITE_U32(CMD_TLB_PURGE | iovp, &ioc->ioc_regs->io_command); 626 iovp += chain_size; 627 byte_cnt -= chain_size; 628 } 629 } 630 631 /** 632 * ccio_mark_invalid - Mark the I/O Pdir entries invalid. 633 * @ioc: The I/O Controller. 634 * @iova: The I/O Virtual Address. 635 * @byte_cnt: The requested number of bytes to be freed from the I/O Pdir. 636 * 637 * Mark the I/O Pdir entries invalid and blow away the corresponding I/O 638 * TLB entries. 639 * 640 * FIXME: at some threshold it might be "cheaper" to just blow 641 * away the entire I/O TLB instead of individual entries. 642 * 643 * FIXME: Uturn has 256 TLB entries. We don't need to purge every 644 * PDIR entry - just once for each possible TLB entry. 645 * (We do need to maker I/O PDIR entries invalid regardless). 646 * 647 * FIXME: Can we change byte_cnt to pages_mapped? 648 */ 649 static void 650 ccio_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt) 651 { 652 u32 iovp = (u32)CCIO_IOVP(iova); 653 size_t saved_byte_cnt; 654 655 /* round up to nearest page size */ 656 saved_byte_cnt = byte_cnt = ALIGN(byte_cnt, IOVP_SIZE); 657 658 while(byte_cnt > 0) { 659 /* invalidate one page at a time */ 660 unsigned int idx = PDIR_INDEX(iovp); 661 char *pdir_ptr = (char *) &(ioc->pdir_base[idx]); 662 663 BUG_ON(idx >= (ioc->pdir_size / sizeof(u64))); 664 pdir_ptr[7] = 0; /* clear only VALID bit */ 665 /* 666 ** FIXME: PCX_W platforms don't need FDC/SYNC. (eg C360) 667 ** PCX-U/U+ do. (eg C200/C240) 668 ** See PDC_MODEL/option 0/SW_CAP for "Non-coherent IO-PDIR bit". 669 */ 670 asm_io_fdc(pdir_ptr); 671 672 iovp += IOVP_SIZE; 673 byte_cnt -= IOVP_SIZE; 674 } 675 676 asm_io_sync(); 677 ccio_clear_io_tlb(ioc, CCIO_IOVP(iova), saved_byte_cnt); 678 } 679 680 /**************************************************************** 681 ** 682 ** CCIO dma_ops 683 ** 684 *****************************************************************/ 685 686 /** 687 * ccio_dma_supported - Verify the IOMMU supports the DMA address range. 688 * @dev: The PCI device. 689 * @mask: A bit mask describing the DMA address range of the device. 690 */ 691 static int 692 ccio_dma_supported(struct device *dev, u64 mask) 693 { 694 if(dev == NULL) { 695 printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n"); 696 BUG(); 697 return 0; 698 } 699 700 /* only support 32-bit or better devices (ie PCI/GSC) */ 701 return (int)(mask >= 0xffffffffUL); 702 } 703 704 /** 705 * ccio_map_single - Map an address range into the IOMMU. 706 * @dev: The PCI device. 707 * @addr: The physical address of the DMA region. 708 * @size: The length of the DMA region. 709 * @direction: The direction of the DMA transaction (to/from device). 710 * 711 * This function implements the pci_map_single function. 712 */ 713 static dma_addr_t 714 ccio_map_single(struct device *dev, phys_addr_t addr, size_t size, 715 enum dma_data_direction direction) 716 { 717 int idx; 718 struct ioc *ioc; 719 unsigned long flags; 720 dma_addr_t iovp; 721 dma_addr_t offset; 722 __le64 *pdir_start; 723 unsigned long hint = hint_lookup[(int)direction]; 724 725 BUG_ON(!dev); 726 ioc = GET_IOC(dev); 727 if (!ioc) 728 return DMA_MAPPING_ERROR; 729 730 BUG_ON(size <= 0); 731 732 /* save offset bits */ 733 offset = offset_in_page(addr); 734 735 /* round up to nearest IOVP_SIZE */ 736 size = ALIGN(size + offset, IOVP_SIZE); 737 spin_lock_irqsave(&ioc->res_lock, flags); 738 739 #ifdef CCIO_COLLECT_STATS 740 ioc->msingle_calls++; 741 ioc->msingle_pages += size >> IOVP_SHIFT; 742 #endif 743 744 idx = ccio_alloc_range(ioc, dev, size); 745 iovp = (dma_addr_t)MKIOVP(idx); 746 747 pdir_start = &(ioc->pdir_base[idx]); 748 749 DBG_RUN("%s() %pa -> %#lx size: %zu\n", 750 __func__, &addr, (long)(iovp | offset), size); 751 752 /* If not cacheline aligned, force SAFE_DMA on the whole mess */ 753 if ((size % L1_CACHE_BYTES) || (addr % L1_CACHE_BYTES)) 754 hint |= HINT_SAFE_DMA; 755 756 while(size > 0) { 757 ccio_io_pdir_entry(pdir_start, KERNEL_SPACE, addr, hint); 758 759 DBG_RUN(" pdir %p %08x%08x\n", 760 pdir_start, 761 (u32) (((u32 *) pdir_start)[0]), 762 (u32) (((u32 *) pdir_start)[1])); 763 ++pdir_start; 764 addr += IOVP_SIZE; 765 size -= IOVP_SIZE; 766 } 767 768 spin_unlock_irqrestore(&ioc->res_lock, flags); 769 770 /* form complete address */ 771 return CCIO_IOVA(iovp, offset); 772 } 773 774 775 static dma_addr_t 776 ccio_map_phys(struct device *dev, phys_addr_t phys, size_t size, 777 enum dma_data_direction direction, unsigned long attrs) 778 { 779 if (unlikely(attrs & DMA_ATTR_MMIO)) 780 return DMA_MAPPING_ERROR; 781 782 return ccio_map_single(dev, phys, size, direction); 783 } 784 785 786 /** 787 * ccio_unmap_phys - Unmap an address range from the IOMMU. 788 * @dev: The PCI device. 789 * @iova: The start address of the DMA region. 790 * @size: The length of the DMA region. 791 * @direction: The direction of the DMA transaction (to/from device). 792 * @attrs: attributes 793 */ 794 static void 795 ccio_unmap_phys(struct device *dev, dma_addr_t iova, size_t size, 796 enum dma_data_direction direction, unsigned long attrs) 797 { 798 struct ioc *ioc; 799 unsigned long flags; 800 dma_addr_t offset = iova & ~IOVP_MASK; 801 802 BUG_ON(!dev); 803 ioc = GET_IOC(dev); 804 if (!ioc) { 805 WARN_ON(!ioc); 806 return; 807 } 808 809 DBG_RUN("%s() iovp %#lx/%zx\n", 810 __func__, (long)iova, size); 811 812 iova ^= offset; /* clear offset bits */ 813 size += offset; 814 size = ALIGN(size, IOVP_SIZE); 815 816 spin_lock_irqsave(&ioc->res_lock, flags); 817 818 #ifdef CCIO_COLLECT_STATS 819 ioc->usingle_calls++; 820 ioc->usingle_pages += size >> IOVP_SHIFT; 821 #endif 822 823 ccio_mark_invalid(ioc, iova, size); 824 ccio_free_range(ioc, iova, (size >> IOVP_SHIFT)); 825 spin_unlock_irqrestore(&ioc->res_lock, flags); 826 } 827 828 /** 829 * ccio_alloc - Allocate a consistent DMA mapping. 830 * @dev: The PCI device. 831 * @size: The length of the DMA region. 832 * @dma_handle: The DMA address handed back to the device (not the cpu). 833 * @flag: allocation flags 834 * @attrs: attributes 835 * 836 * This function implements the pci_alloc_consistent function. 837 */ 838 static void * 839 ccio_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle, gfp_t flag, 840 unsigned long attrs) 841 { 842 void *ret; 843 #if 0 844 /* GRANT Need to establish hierarchy for non-PCI devs as well 845 ** and then provide matching gsc_map_xxx() functions for them as well. 846 */ 847 if(!hwdev) { 848 /* only support PCI */ 849 *dma_handle = 0; 850 return 0; 851 } 852 #endif 853 ret = (void *) __get_free_pages(flag, get_order(size)); 854 855 if (ret) { 856 memset(ret, 0, size); 857 *dma_handle = ccio_map_single(dev, virt_to_phys(ret), size, 858 DMA_BIDIRECTIONAL); 859 } 860 861 return ret; 862 } 863 864 /** 865 * ccio_free - Free a consistent DMA mapping. 866 * @dev: The PCI device. 867 * @size: The length of the DMA region. 868 * @cpu_addr: The cpu address returned from the ccio_alloc_consistent. 869 * @dma_handle: The device address returned from the ccio_alloc_consistent. 870 * @attrs: attributes 871 * 872 * This function implements the pci_free_consistent function. 873 */ 874 static void 875 ccio_free(struct device *dev, size_t size, void *cpu_addr, 876 dma_addr_t dma_handle, unsigned long attrs) 877 { 878 ccio_unmap_phys(dev, dma_handle, size, 0, 0); 879 free_pages((unsigned long)cpu_addr, get_order(size)); 880 } 881 882 /* 883 ** Since 0 is a valid pdir_base index value, can't use that 884 ** to determine if a value is valid or not. Use a flag to indicate 885 ** the SG list entry contains a valid pdir index. 886 */ 887 #define PIDE_FLAG 0x80000000UL 888 889 #ifdef CCIO_COLLECT_STATS 890 #define IOMMU_MAP_STATS 891 #endif 892 #include "iommu-helpers.h" 893 894 /** 895 * ccio_map_sg - Map the scatter/gather list into the IOMMU. 896 * @dev: The PCI device. 897 * @sglist: The scatter/gather list to be mapped in the IOMMU. 898 * @nents: The number of entries in the scatter/gather list. 899 * @direction: The direction of the DMA transaction (to/from device). 900 * @attrs: attributes 901 * 902 * This function implements the pci_map_sg function. 903 */ 904 static int 905 ccio_map_sg(struct device *dev, struct scatterlist *sglist, int nents, 906 enum dma_data_direction direction, unsigned long attrs) 907 { 908 struct ioc *ioc; 909 int coalesced, filled = 0; 910 unsigned long flags; 911 unsigned long hint = hint_lookup[(int)direction]; 912 unsigned long prev_len = 0, current_len = 0; 913 int i; 914 915 BUG_ON(!dev); 916 ioc = GET_IOC(dev); 917 if (!ioc) 918 return -EINVAL; 919 920 DBG_RUN_SG("%s() START %d entries\n", __func__, nents); 921 922 /* Fast path single entry scatterlists. */ 923 if (nents == 1) { 924 sg_dma_address(sglist) = ccio_map_single(dev, 925 sg_phys(sglist), sglist->length, 926 direction); 927 sg_dma_len(sglist) = sglist->length; 928 return 1; 929 } 930 931 for(i = 0; i < nents; i++) 932 prev_len += sglist[i].length; 933 934 spin_lock_irqsave(&ioc->res_lock, flags); 935 936 #ifdef CCIO_COLLECT_STATS 937 ioc->msg_calls++; 938 #endif 939 940 /* 941 ** First coalesce the chunks and allocate I/O pdir space 942 ** 943 ** If this is one DMA stream, we can properly map using the 944 ** correct virtual address associated with each DMA page. 945 ** w/o this association, we wouldn't have coherent DMA! 946 ** Access to the virtual address is what forces a two pass algorithm. 947 */ 948 coalesced = iommu_coalesce_chunks(ioc, dev, sglist, nents, ccio_alloc_range); 949 950 /* 951 ** Program the I/O Pdir 952 ** 953 ** map the virtual addresses to the I/O Pdir 954 ** o dma_address will contain the pdir index 955 ** o dma_len will contain the number of bytes to map 956 ** o page/offset contain the virtual address. 957 */ 958 filled = iommu_fill_pdir(ioc, sglist, nents, hint, ccio_io_pdir_entry); 959 960 spin_unlock_irqrestore(&ioc->res_lock, flags); 961 962 BUG_ON(coalesced != filled); 963 964 DBG_RUN_SG("%s() DONE %d mappings\n", __func__, filled); 965 966 for (i = 0; i < filled; i++) 967 current_len += sg_dma_len(sglist + i); 968 969 BUG_ON(current_len != prev_len); 970 971 return filled; 972 } 973 974 /** 975 * ccio_unmap_sg - Unmap the scatter/gather list from the IOMMU. 976 * @dev: The PCI device. 977 * @sglist: The scatter/gather list to be unmapped from the IOMMU. 978 * @nents: The number of entries in the scatter/gather list. 979 * @direction: The direction of the DMA transaction (to/from device). 980 * @attrs: attributes 981 * 982 * This function implements the pci_unmap_sg function. 983 */ 984 static void 985 ccio_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents, 986 enum dma_data_direction direction, unsigned long attrs) 987 { 988 struct ioc *ioc; 989 990 BUG_ON(!dev); 991 ioc = GET_IOC(dev); 992 if (!ioc) { 993 WARN_ON(!ioc); 994 return; 995 } 996 997 DBG_RUN_SG("%s() START %d entries, %p,%x\n", 998 __func__, nents, sg_virt(sglist), sglist->length); 999 1000 #ifdef CCIO_COLLECT_STATS 1001 ioc->usg_calls++; 1002 #endif 1003 1004 while (nents && sg_dma_len(sglist)) { 1005 1006 #ifdef CCIO_COLLECT_STATS 1007 ioc->usg_pages += sg_dma_len(sglist) >> PAGE_SHIFT; 1008 #endif 1009 ccio_unmap_phys(dev, sg_dma_address(sglist), 1010 sg_dma_len(sglist), direction, 0); 1011 ++sglist; 1012 nents--; 1013 } 1014 1015 DBG_RUN_SG("%s() DONE (nents %d)\n", __func__, nents); 1016 } 1017 1018 static const struct dma_map_ops ccio_ops = { 1019 .dma_supported = ccio_dma_supported, 1020 .alloc = ccio_alloc, 1021 .free = ccio_free, 1022 .map_phys = ccio_map_phys, 1023 .unmap_phys = ccio_unmap_phys, 1024 .map_sg = ccio_map_sg, 1025 .unmap_sg = ccio_unmap_sg, 1026 .get_sgtable = dma_common_get_sgtable, 1027 .alloc_pages_op = dma_common_alloc_pages, 1028 .free_pages = dma_common_free_pages, 1029 }; 1030 1031 #ifdef CONFIG_PROC_FS 1032 static int ccio_proc_info(struct seq_file *m, void *p) 1033 { 1034 struct ioc *ioc = ioc_list; 1035 1036 while (ioc != NULL) { 1037 unsigned int total_pages = ioc->res_size << 3; 1038 #ifdef CCIO_COLLECT_STATS 1039 unsigned long avg = 0, min, max; 1040 int j; 1041 #endif 1042 1043 seq_printf(m, "%s\n", ioc->name); 1044 1045 seq_printf(m, "Cujo 2.0 bug : %s\n", 1046 (ioc->cujo20_bug ? "yes" : "no")); 1047 1048 seq_printf(m, "IO PDIR size : %d bytes (%d entries)\n", 1049 total_pages * 8, total_pages); 1050 1051 #ifdef CCIO_COLLECT_STATS 1052 seq_printf(m, "IO PDIR entries : %ld free %ld used (%d%%)\n", 1053 total_pages - ioc->used_pages, ioc->used_pages, 1054 (int)(ioc->used_pages * 100 / total_pages)); 1055 #endif 1056 1057 seq_printf(m, "Resource bitmap : %d bytes (%d pages)\n", 1058 ioc->res_size, total_pages); 1059 1060 #ifdef CCIO_COLLECT_STATS 1061 min = max = ioc->avg_search[0]; 1062 for(j = 0; j < CCIO_SEARCH_SAMPLE; ++j) { 1063 avg += ioc->avg_search[j]; 1064 if(ioc->avg_search[j] > max) 1065 max = ioc->avg_search[j]; 1066 if(ioc->avg_search[j] < min) 1067 min = ioc->avg_search[j]; 1068 } 1069 avg /= CCIO_SEARCH_SAMPLE; 1070 seq_printf(m, " Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles)\n", 1071 min, avg, max); 1072 1073 seq_printf(m, "pci_map_single(): %8ld calls %8ld pages (avg %d/1000)\n", 1074 ioc->msingle_calls, ioc->msingle_pages, 1075 (int)((ioc->msingle_pages * 1000)/ioc->msingle_calls)); 1076 1077 /* KLUGE - unmap_sg calls unmap_phys for each mapped page */ 1078 min = ioc->usingle_calls - ioc->usg_calls; 1079 max = ioc->usingle_pages - ioc->usg_pages; 1080 seq_printf(m, "pci_unmap_single: %8ld calls %8ld pages (avg %d/1000)\n", 1081 min, max, (int)((max * 1000)/min)); 1082 1083 seq_printf(m, "pci_map_sg() : %8ld calls %8ld pages (avg %d/1000)\n", 1084 ioc->msg_calls, ioc->msg_pages, 1085 (int)((ioc->msg_pages * 1000)/ioc->msg_calls)); 1086 1087 seq_printf(m, "pci_unmap_sg() : %8ld calls %8ld pages (avg %d/1000)\n\n\n", 1088 ioc->usg_calls, ioc->usg_pages, 1089 (int)((ioc->usg_pages * 1000)/ioc->usg_calls)); 1090 #endif /* CCIO_COLLECT_STATS */ 1091 1092 ioc = ioc->next; 1093 } 1094 1095 return 0; 1096 } 1097 1098 static int ccio_proc_bitmap_info(struct seq_file *m, void *p) 1099 { 1100 struct ioc *ioc = ioc_list; 1101 1102 while (ioc != NULL) { 1103 seq_hex_dump(m, " ", DUMP_PREFIX_NONE, 32, 4, ioc->res_map, 1104 ioc->res_size, false); 1105 seq_putc(m, '\n'); 1106 ioc = ioc->next; 1107 break; /* XXX - remove me */ 1108 } 1109 1110 return 0; 1111 } 1112 #endif /* CONFIG_PROC_FS */ 1113 1114 /** 1115 * ccio_find_ioc - Find the ioc in the ioc_list 1116 * @hw_path: The hardware path of the ioc. 1117 * 1118 * This function searches the ioc_list for an ioc that matches 1119 * the provide hardware path. 1120 */ 1121 static struct ioc * ccio_find_ioc(int hw_path) 1122 { 1123 int i; 1124 struct ioc *ioc; 1125 1126 ioc = ioc_list; 1127 for (i = 0; i < ioc_count; i++) { 1128 if (ioc->hw_path == hw_path) 1129 return ioc; 1130 1131 ioc = ioc->next; 1132 } 1133 1134 return NULL; 1135 } 1136 1137 /** 1138 * ccio_get_iommu - Find the iommu which controls this device 1139 * @dev: The parisc device. 1140 * 1141 * This function searches through the registered IOMMU's and returns 1142 * the appropriate IOMMU for the device based on its hardware path. 1143 */ 1144 void * ccio_get_iommu(const struct parisc_device *dev) 1145 { 1146 dev = find_pa_parent_type(dev, HPHW_IOA); 1147 if (!dev) 1148 return NULL; 1149 1150 return ccio_find_ioc(dev->hw_path); 1151 } 1152 1153 #define CUJO_20_STEP 0x10000000 /* inc upper nibble */ 1154 1155 /* Cujo 2.0 has a bug which will silently corrupt data being transferred 1156 * to/from certain pages. To avoid this happening, we mark these pages 1157 * as `used', and ensure that nothing will try to allocate from them. 1158 */ 1159 void __init ccio_cujo20_fixup(struct parisc_device *cujo, u32 iovp) 1160 { 1161 unsigned int idx; 1162 struct parisc_device *dev = parisc_parent(cujo); 1163 struct ioc *ioc = ccio_get_iommu(dev); 1164 u8 *res_ptr; 1165 1166 ioc->cujo20_bug = 1; 1167 res_ptr = ioc->res_map; 1168 idx = PDIR_INDEX(iovp) >> 3; 1169 1170 while (idx < ioc->res_size) { 1171 res_ptr[idx] |= 0xff; 1172 idx += PDIR_INDEX(CUJO_20_STEP) >> 3; 1173 } 1174 } 1175 1176 #if 0 1177 /* GRANT - is this needed for U2 or not? */ 1178 1179 /* 1180 ** Get the size of the I/O TLB for this I/O MMU. 1181 ** 1182 ** If spa_shift is non-zero (ie probably U2), 1183 ** then calculate the I/O TLB size using spa_shift. 1184 ** 1185 ** Otherwise we are supposed to get the IODC entry point ENTRY TLB 1186 ** and execute it. However, both U2 and Uturn firmware supplies spa_shift. 1187 ** I think only Java (K/D/R-class too?) systems don't do this. 1188 */ 1189 static int 1190 ccio_get_iotlb_size(struct parisc_device *dev) 1191 { 1192 if (dev->spa_shift == 0) { 1193 panic("%s() : Can't determine I/O TLB size.\n", __func__); 1194 } 1195 return (1 << dev->spa_shift); 1196 } 1197 #else 1198 1199 /* Uturn supports 256 TLB entries */ 1200 #define CCIO_CHAINID_SHIFT 8 1201 #define CCIO_CHAINID_MASK 0xff 1202 #endif /* 0 */ 1203 1204 /* We *can't* support JAVA (T600). Venture there at your own risk. */ 1205 static const struct parisc_device_id ccio_tbl[] __initconst = { 1206 { HPHW_IOA, HVERSION_REV_ANY_ID, U2_IOA_RUNWAY, 0xb }, /* U2 */ 1207 { HPHW_IOA, HVERSION_REV_ANY_ID, UTURN_IOA_RUNWAY, 0xb }, /* UTurn */ 1208 { 0, } 1209 }; 1210 1211 static int ccio_probe(struct parisc_device *dev); 1212 1213 static struct parisc_driver ccio_driver __refdata = { 1214 .name = "ccio", 1215 .id_table = ccio_tbl, 1216 .probe = ccio_probe, 1217 }; 1218 1219 /** 1220 * ccio_ioc_init - Initialize the I/O Controller 1221 * @ioc: The I/O Controller. 1222 * 1223 * Initialize the I/O Controller which includes setting up the 1224 * I/O Page Directory, the resource map, and initalizing the 1225 * U2/Uturn chip into virtual mode. 1226 */ 1227 static void __init 1228 ccio_ioc_init(struct ioc *ioc) 1229 { 1230 int i; 1231 unsigned int iov_order; 1232 u32 iova_space_size; 1233 1234 /* 1235 ** Determine IOVA Space size from memory size. 1236 ** 1237 ** Ideally, PCI drivers would register the maximum number 1238 ** of DMA they can have outstanding for each device they 1239 ** own. Next best thing would be to guess how much DMA 1240 ** can be outstanding based on PCI Class/sub-class. Both 1241 ** methods still require some "extra" to support PCI 1242 ** Hot-Plug/Removal of PCI cards. (aka PCI OLARD). 1243 */ 1244 1245 iova_space_size = (u32) (totalram_pages() / count_parisc_driver(&ccio_driver)); 1246 1247 /* limit IOVA space size to 1MB-1GB */ 1248 1249 if (iova_space_size < (1 << (20 - PAGE_SHIFT))) { 1250 iova_space_size = 1 << (20 - PAGE_SHIFT); 1251 #ifdef __LP64__ 1252 } else if (iova_space_size > (1 << (30 - PAGE_SHIFT))) { 1253 iova_space_size = 1 << (30 - PAGE_SHIFT); 1254 #endif 1255 } 1256 1257 /* 1258 ** iova space must be log2() in size. 1259 ** thus, pdir/res_map will also be log2(). 1260 */ 1261 1262 /* We could use larger page sizes in order to *decrease* the number 1263 ** of mappings needed. (ie 8k pages means 1/2 the mappings). 1264 ** 1265 ** Note: Grant Grunder says "Using 8k I/O pages isn't trivial either 1266 ** since the pages must also be physically contiguous - typically 1267 ** this is the case under linux." 1268 */ 1269 1270 iov_order = get_order(iova_space_size << PAGE_SHIFT); 1271 1272 /* iova_space_size is now bytes, not pages */ 1273 iova_space_size = 1 << (iov_order + PAGE_SHIFT); 1274 1275 ioc->pdir_size = (iova_space_size / IOVP_SIZE) * sizeof(u64); 1276 1277 BUG_ON(ioc->pdir_size > 8 * 1024 * 1024); /* max pdir size <= 8MB */ 1278 1279 /* Verify it's a power of two */ 1280 BUG_ON((1 << get_order(ioc->pdir_size)) != (ioc->pdir_size >> PAGE_SHIFT)); 1281 1282 DBG_INIT("%s() hpa 0x%p mem %luMB IOV %dMB (%d bits)\n", 1283 __func__, ioc->ioc_regs, 1284 (unsigned long) totalram_pages() >> (20 - PAGE_SHIFT), 1285 iova_space_size>>20, 1286 iov_order + PAGE_SHIFT); 1287 1288 ioc->pdir_base = (__le64 *)__get_free_pages(GFP_KERNEL, 1289 get_order(ioc->pdir_size)); 1290 if(NULL == ioc->pdir_base) { 1291 panic("%s() could not allocate I/O Page Table\n", __func__); 1292 } 1293 memset(ioc->pdir_base, 0, ioc->pdir_size); 1294 1295 BUG_ON((((unsigned long)ioc->pdir_base) & PAGE_MASK) != (unsigned long)ioc->pdir_base); 1296 DBG_INIT(" base %p\n", ioc->pdir_base); 1297 1298 /* resource map size dictated by pdir_size */ 1299 ioc->res_size = (ioc->pdir_size / sizeof(u64)) >> 3; 1300 DBG_INIT("%s() res_size 0x%x\n", __func__, ioc->res_size); 1301 1302 ioc->res_map = (u8 *)__get_free_pages(GFP_KERNEL, 1303 get_order(ioc->res_size)); 1304 if(NULL == ioc->res_map) { 1305 panic("%s() could not allocate resource map\n", __func__); 1306 } 1307 memset(ioc->res_map, 0, ioc->res_size); 1308 1309 /* Initialize the res_hint to 16 */ 1310 ioc->res_hint = 16; 1311 1312 /* Initialize the spinlock */ 1313 spin_lock_init(&ioc->res_lock); 1314 1315 /* 1316 ** Chainid is the upper most bits of an IOVP used to determine 1317 ** which TLB entry an IOVP will use. 1318 */ 1319 ioc->chainid_shift = get_order(iova_space_size) + PAGE_SHIFT - CCIO_CHAINID_SHIFT; 1320 DBG_INIT(" chainid_shift 0x%x\n", ioc->chainid_shift); 1321 1322 /* 1323 ** Initialize IOA hardware 1324 */ 1325 WRITE_U32(CCIO_CHAINID_MASK << ioc->chainid_shift, 1326 &ioc->ioc_regs->io_chain_id_mask); 1327 1328 WRITE_U32(virt_to_phys(ioc->pdir_base), 1329 &ioc->ioc_regs->io_pdir_base); 1330 1331 /* 1332 ** Go to "Virtual Mode" 1333 */ 1334 WRITE_U32(IOA_NORMAL_MODE, &ioc->ioc_regs->io_control); 1335 1336 /* 1337 ** Initialize all I/O TLB entries to 0 (Valid bit off). 1338 */ 1339 WRITE_U32(0, &ioc->ioc_regs->io_tlb_entry_m); 1340 WRITE_U32(0, &ioc->ioc_regs->io_tlb_entry_l); 1341 1342 for(i = 1 << CCIO_CHAINID_SHIFT; i ; i--) { 1343 WRITE_U32((CMD_TLB_DIRECT_WRITE | (i << ioc->chainid_shift)), 1344 &ioc->ioc_regs->io_command); 1345 } 1346 } 1347 1348 static void __init 1349 ccio_init_resource(struct resource *res, char *name, void __iomem *ioaddr) 1350 { 1351 int result; 1352 1353 res->parent = NULL; 1354 res->flags = IORESOURCE_MEM; 1355 /* 1356 * bracing ((signed) ...) are required for 64bit kernel because 1357 * we only want to sign extend the lower 16 bits of the register. 1358 * The upper 16-bits of range registers are hardcoded to 0xffff. 1359 */ 1360 res->start = (unsigned long)((signed) READ_U32(ioaddr) << 16); 1361 res->end = (unsigned long)((signed) (READ_U32(ioaddr + 4) << 16) - 1); 1362 res->name = name; 1363 /* 1364 * Check if this MMIO range is disable 1365 */ 1366 if (res->end + 1 == res->start) 1367 return; 1368 1369 /* On some platforms (e.g. K-Class), we have already registered 1370 * resources for devices reported by firmware. Some are children 1371 * of ccio. 1372 * "insert" ccio ranges in the mmio hierarchy (/proc/iomem). 1373 */ 1374 result = insert_resource(&iomem_resource, res); 1375 if (result < 0) { 1376 printk(KERN_ERR "%s() failed to claim CCIO bus address space (%08lx,%08lx)\n", 1377 __func__, (unsigned long)res->start, (unsigned long)res->end); 1378 } 1379 } 1380 1381 static int __init ccio_init_resources(struct ioc *ioc) 1382 { 1383 struct resource *res = ioc->mmio_region; 1384 char *name = kmalloc(14, GFP_KERNEL); 1385 if (unlikely(!name)) 1386 return -ENOMEM; 1387 snprintf(name, 14, "GSC Bus [%d/]", ioc->hw_path); 1388 1389 ccio_init_resource(res, name, &ioc->ioc_regs->io_io_low); 1390 ccio_init_resource(res + 1, name, &ioc->ioc_regs->io_io_low_hv); 1391 return 0; 1392 } 1393 1394 static int new_ioc_area(struct resource *res, unsigned long size, 1395 unsigned long min, unsigned long max, unsigned long align) 1396 { 1397 if (max <= min) 1398 return -EBUSY; 1399 1400 res->start = (max - size + 1) &~ (align - 1); 1401 res->end = res->start + size; 1402 1403 /* We might be trying to expand the MMIO range to include 1404 * a child device that has already registered it's MMIO space. 1405 * Use "insert" instead of request_resource(). 1406 */ 1407 if (!insert_resource(&iomem_resource, res)) 1408 return 0; 1409 1410 return new_ioc_area(res, size, min, max - size, align); 1411 } 1412 1413 static int expand_ioc_area(struct resource *res, unsigned long size, 1414 unsigned long min, unsigned long max, unsigned long align) 1415 { 1416 unsigned long start, len; 1417 1418 if (!res->parent) 1419 return new_ioc_area(res, size, min, max, align); 1420 1421 start = (res->start - size) &~ (align - 1); 1422 len = res->end - start + 1; 1423 if (start >= min) { 1424 if (!adjust_resource(res, start, len)) 1425 return 0; 1426 } 1427 1428 start = res->start; 1429 len = ((size + res->end + align) &~ (align - 1)) - start; 1430 if (start + len <= max) { 1431 if (!adjust_resource(res, start, len)) 1432 return 0; 1433 } 1434 1435 return -EBUSY; 1436 } 1437 1438 /* 1439 * Dino calls this function. Beware that we may get called on systems 1440 * which have no IOC (725, B180, C160L, etc) but do have a Dino. 1441 * So it's legal to find no parent IOC. 1442 * 1443 * Some other issues: one of the resources in the ioc may be unassigned. 1444 */ 1445 int ccio_allocate_resource(const struct parisc_device *dev, 1446 struct resource *res, unsigned long size, 1447 unsigned long min, unsigned long max, unsigned long align) 1448 { 1449 struct resource *parent = &iomem_resource; 1450 struct ioc *ioc = ccio_get_iommu(dev); 1451 if (!ioc) 1452 goto out; 1453 1454 parent = ioc->mmio_region; 1455 if (parent->parent && 1456 !allocate_resource(parent, res, size, min, max, align, NULL, NULL)) 1457 return 0; 1458 1459 if ((parent + 1)->parent && 1460 !allocate_resource(parent + 1, res, size, min, max, align, 1461 NULL, NULL)) 1462 return 0; 1463 1464 if (!expand_ioc_area(parent, size, min, max, align)) { 1465 __raw_writel(((parent->start)>>16) | 0xffff0000, 1466 &ioc->ioc_regs->io_io_low); 1467 __raw_writel(((parent->end)>>16) | 0xffff0000, 1468 &ioc->ioc_regs->io_io_high); 1469 } else if (!expand_ioc_area(parent + 1, size, min, max, align)) { 1470 parent++; 1471 __raw_writel(((parent->start)>>16) | 0xffff0000, 1472 &ioc->ioc_regs->io_io_low_hv); 1473 __raw_writel(((parent->end)>>16) | 0xffff0000, 1474 &ioc->ioc_regs->io_io_high_hv); 1475 } else { 1476 return -EBUSY; 1477 } 1478 1479 out: 1480 return allocate_resource(parent, res, size, min, max, align, NULL,NULL); 1481 } 1482 1483 int ccio_request_resource(const struct parisc_device *dev, 1484 struct resource *res) 1485 { 1486 struct resource *parent; 1487 struct ioc *ioc = ccio_get_iommu(dev); 1488 1489 if (!ioc) { 1490 parent = &iomem_resource; 1491 } else if ((ioc->mmio_region->start <= res->start) && 1492 (res->end <= ioc->mmio_region->end)) { 1493 parent = ioc->mmio_region; 1494 } else if (((ioc->mmio_region + 1)->start <= res->start) && 1495 (res->end <= (ioc->mmio_region + 1)->end)) { 1496 parent = ioc->mmio_region + 1; 1497 } else { 1498 return -EBUSY; 1499 } 1500 1501 /* "transparent" bus bridges need to register MMIO resources 1502 * firmware assigned them. e.g. children of hppb.c (e.g. K-class) 1503 * registered their resources in the PDC "bus walk" (See 1504 * arch/parisc/kernel/inventory.c). 1505 */ 1506 return insert_resource(parent, res); 1507 } 1508 1509 /** 1510 * ccio_probe - Determine if ccio should claim this device. 1511 * @dev: The device which has been found 1512 * 1513 * Determine if ccio should claim this chip (return 0) or not (return 1). 1514 * If so, initialize the chip and tell other partners in crime they 1515 * have work to do. 1516 */ 1517 static int __init ccio_probe(struct parisc_device *dev) 1518 { 1519 int i; 1520 struct ioc *ioc, **ioc_p = &ioc_list; 1521 struct pci_hba_data *hba; 1522 1523 ioc = kzalloc(sizeof(struct ioc), GFP_KERNEL); 1524 if (ioc == NULL) { 1525 printk(KERN_ERR MODULE_NAME ": memory allocation failure\n"); 1526 return -ENOMEM; 1527 } 1528 1529 ioc->name = dev->id.hversion == U2_IOA_RUNWAY ? "U2" : "UTurn"; 1530 1531 printk(KERN_INFO "Found %s at 0x%lx\n", ioc->name, 1532 (unsigned long)dev->hpa.start); 1533 1534 for (i = 0; i < ioc_count; i++) { 1535 ioc_p = &(*ioc_p)->next; 1536 } 1537 *ioc_p = ioc; 1538 1539 ioc->hw_path = dev->hw_path; 1540 ioc->ioc_regs = ioremap(dev->hpa.start, 4096); 1541 if (!ioc->ioc_regs) { 1542 kfree(ioc); 1543 return -ENOMEM; 1544 } 1545 ccio_ioc_init(ioc); 1546 if (ccio_init_resources(ioc)) { 1547 iounmap(ioc->ioc_regs); 1548 kfree(ioc); 1549 return -ENOMEM; 1550 } 1551 hppa_dma_ops = &ccio_ops; 1552 1553 hba = kzalloc(sizeof(*hba), GFP_KERNEL); 1554 /* if this fails, no I/O cards will work, so may as well bug */ 1555 BUG_ON(hba == NULL); 1556 1557 hba->iommu = ioc; 1558 dev->dev.platform_data = hba; 1559 1560 #ifdef CONFIG_PROC_FS 1561 if (ioc_count == 0) { 1562 struct proc_dir_entry *runway; 1563 1564 runway = proc_mkdir("bus/runway", NULL); 1565 if (runway) { 1566 proc_create_single(MODULE_NAME, 0, runway, 1567 ccio_proc_info); 1568 proc_create_single(MODULE_NAME"-bitmap", 0, runway, 1569 ccio_proc_bitmap_info); 1570 } 1571 } 1572 #endif 1573 ioc_count++; 1574 return 0; 1575 } 1576 1577 /** 1578 * ccio_init - ccio initialization procedure. 1579 * 1580 * Register this driver. 1581 */ 1582 static int __init ccio_init(void) 1583 { 1584 return register_parisc_driver(&ccio_driver); 1585 } 1586 arch_initcall(ccio_init); 1587