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 ** 12 ** 13 ** "Real Mode" operation refers to U2/Uturn chip operation. 14 ** U2/Uturn were designed to perform coherency checks w/o using 15 ** the I/O MMU - basically what x86 does. 16 ** 17 ** Philipp Rumpf has a "Real Mode" driver for PCX-W machines at: 18 ** CVSROOT=:pserver:anonymous@198.186.203.37:/cvsroot/linux-parisc 19 ** cvs -z3 co linux/arch/parisc/kernel/dma-rm.c 20 ** 21 ** I've rewritten his code to work under TPG's tree. See ccio-rm-dma.c. 22 ** 23 ** Drawbacks of using Real Mode are: 24 ** o outbound DMA is slower - U2 won't prefetch data (GSC+ XQL signal). 25 ** o Inbound DMA less efficient - U2 can't use DMA_FAST attribute. 26 ** o Ability to do scatter/gather in HW is lost. 27 ** o Doesn't work under PCX-U/U+ machines since they didn't follow 28 ** the coherency design originally worked out. Only PCX-W does. 29 */ 30 31 #include <linux/types.h> 32 #include <linux/kernel.h> 33 #include <linux/init.h> 34 #include <linux/mm.h> 35 #include <linux/spinlock.h> 36 #include <linux/slab.h> 37 #include <linux/string.h> 38 #include <linux/pci.h> 39 #include <linux/reboot.h> 40 #include <linux/proc_fs.h> 41 #include <linux/seq_file.h> 42 #include <linux/scatterlist.h> 43 #include <linux/iommu-helper.h> 44 #include <linux/export.h> 45 46 #include <asm/byteorder.h> 47 #include <asm/cache.h> /* for L1_CACHE_BYTES */ 48 #include <linux/uaccess.h> 49 #include <asm/page.h> 50 #include <asm/dma.h> 51 #include <asm/io.h> 52 #include <asm/hardware.h> /* for register_module() */ 53 #include <asm/parisc-device.h> 54 55 #include "iommu.h" 56 57 /* 58 ** Choose "ccio" since that's what HP-UX calls it. 59 ** Make it easier for folks to migrate from one to the other :^) 60 */ 61 #define MODULE_NAME "ccio" 62 63 #undef DEBUG_CCIO_RES 64 #undef DEBUG_CCIO_RUN 65 #undef DEBUG_CCIO_INIT 66 #undef DEBUG_CCIO_RUN_SG 67 68 #ifdef CONFIG_PROC_FS 69 /* depends on proc fs support. But costs CPU performance. */ 70 #undef CCIO_COLLECT_STATS 71 #endif 72 73 #include <asm/runway.h> /* for proc_runway_root */ 74 75 #ifdef DEBUG_CCIO_INIT 76 #define DBG_INIT(x...) printk(x) 77 #else 78 #define DBG_INIT(x...) 79 #endif 80 81 #ifdef DEBUG_CCIO_RUN 82 #define DBG_RUN(x...) printk(x) 83 #else 84 #define DBG_RUN(x...) 85 #endif 86 87 #ifdef DEBUG_CCIO_RES 88 #define DBG_RES(x...) printk(x) 89 #else 90 #define DBG_RES(x...) 91 #endif 92 93 #ifdef DEBUG_CCIO_RUN_SG 94 #define DBG_RUN_SG(x...) printk(x) 95 #else 96 #define DBG_RUN_SG(x...) 97 #endif 98 99 #define CCIO_INLINE inline 100 #define WRITE_U32(value, addr) __raw_writel(value, addr) 101 #define READ_U32(addr) __raw_readl(addr) 102 103 #define U2_IOA_RUNWAY 0x580 104 #define U2_BC_GSC 0x501 105 #define UTURN_IOA_RUNWAY 0x581 106 #define UTURN_BC_GSC 0x502 107 108 #define IOA_NORMAL_MODE 0x00020080 /* IO_CONTROL to turn on CCIO */ 109 #define CMD_TLB_DIRECT_WRITE 35 /* IO_COMMAND for I/O TLB Writes */ 110 #define CMD_TLB_PURGE 33 /* IO_COMMAND to Purge I/O TLB entry */ 111 112 struct ioa_registers { 113 /* Runway Supervisory Set */ 114 int32_t unused1[12]; 115 uint32_t io_command; /* Offset 12 */ 116 uint32_t io_status; /* Offset 13 */ 117 uint32_t io_control; /* Offset 14 */ 118 int32_t unused2[1]; 119 120 /* Runway Auxiliary Register Set */ 121 uint32_t io_err_resp; /* Offset 0 */ 122 uint32_t io_err_info; /* Offset 1 */ 123 uint32_t io_err_req; /* Offset 2 */ 124 uint32_t io_err_resp_hi; /* Offset 3 */ 125 uint32_t io_tlb_entry_m; /* Offset 4 */ 126 uint32_t io_tlb_entry_l; /* Offset 5 */ 127 uint32_t unused3[1]; 128 uint32_t io_pdir_base; /* Offset 7 */ 129 uint32_t io_io_low_hv; /* Offset 8 */ 130 uint32_t io_io_high_hv; /* Offset 9 */ 131 uint32_t unused4[1]; 132 uint32_t io_chain_id_mask; /* Offset 11 */ 133 uint32_t unused5[2]; 134 uint32_t io_io_low; /* Offset 14 */ 135 uint32_t io_io_high; /* Offset 15 */ 136 }; 137 138 /* 139 ** IOA Registers 140 ** ------------- 141 ** 142 ** Runway IO_CONTROL Register (+0x38) 143 ** 144 ** The Runway IO_CONTROL register controls the forwarding of transactions. 145 ** 146 ** | 0 ... 13 | 14 15 | 16 ... 21 | 22 | 23 24 | 25 ... 31 | 147 ** | HV | TLB | reserved | HV | mode | reserved | 148 ** 149 ** o mode field indicates the address translation of transactions 150 ** forwarded from Runway to GSC+: 151 ** Mode Name Value Definition 152 ** Off (default) 0 Opaque to matching addresses. 153 ** Include 1 Transparent for matching addresses. 154 ** Peek 3 Map matching addresses. 155 ** 156 ** + "Off" mode: Runway transactions which match the I/O range 157 ** specified by the IO_IO_LOW/IO_IO_HIGH registers will be ignored. 158 ** + "Include" mode: all addresses within the I/O range specified 159 ** by the IO_IO_LOW and IO_IO_HIGH registers are transparently 160 ** forwarded. This is the I/O Adapter's normal operating mode. 161 ** + "Peek" mode: used during system configuration to initialize the 162 ** GSC+ bus. Runway Write_Shorts in the address range specified by 163 ** IO_IO_LOW and IO_IO_HIGH are forwarded through the I/O Adapter 164 ** *AND* the GSC+ address is remapped to the Broadcast Physical 165 ** Address space by setting the 14 high order address bits of the 166 ** 32 bit GSC+ address to ones. 167 ** 168 ** o TLB field affects transactions which are forwarded from GSC+ to Runway. 169 ** "Real" mode is the poweron default. 170 ** 171 ** TLB Mode Value Description 172 ** Real 0 No TLB translation. Address is directly mapped and the 173 ** virtual address is composed of selected physical bits. 174 ** Error 1 Software fills the TLB manually. 175 ** Normal 2 IOA fetches IO TLB misses from IO PDIR (in host memory). 176 ** 177 ** 178 ** IO_IO_LOW_HV +0x60 (HV dependent) 179 ** IO_IO_HIGH_HV +0x64 (HV dependent) 180 ** IO_IO_LOW +0x78 (Architected register) 181 ** IO_IO_HIGH +0x7c (Architected register) 182 ** 183 ** IO_IO_LOW and IO_IO_HIGH set the lower and upper bounds of the 184 ** I/O Adapter address space, respectively. 185 ** 186 ** 0 ... 7 | 8 ... 15 | 16 ... 31 | 187 ** 11111111 | 11111111 | address | 188 ** 189 ** Each LOW/HIGH pair describes a disjoint address space region. 190 ** (2 per GSC+ port). Each incoming Runway transaction address is compared 191 ** with both sets of LOW/HIGH registers. If the address is in the range 192 ** greater than or equal to IO_IO_LOW and less than IO_IO_HIGH the transaction 193 ** for forwarded to the respective GSC+ bus. 194 ** Specify IO_IO_LOW equal to or greater than IO_IO_HIGH to avoid specifying 195 ** an address space region. 196 ** 197 ** In order for a Runway address to reside within GSC+ extended address space: 198 ** Runway Address [0:7] must identically compare to 8'b11111111 199 ** Runway Address [8:11] must be equal to IO_IO_LOW(_HV)[16:19] 200 ** Runway Address [12:23] must be greater than or equal to 201 ** IO_IO_LOW(_HV)[20:31] and less than IO_IO_HIGH(_HV)[20:31]. 202 ** Runway Address [24:39] is not used in the comparison. 203 ** 204 ** When the Runway transaction is forwarded to GSC+, the GSC+ address is 205 ** as follows: 206 ** GSC+ Address[0:3] 4'b1111 207 ** GSC+ Address[4:29] Runway Address[12:37] 208 ** GSC+ Address[30:31] 2'b00 209 ** 210 ** All 4 Low/High registers must be initialized (by PDC) once the lower bus 211 ** is interrogated and address space is defined. The operating system will 212 ** modify the architectural IO_IO_LOW and IO_IO_HIGH registers following 213 ** the PDC initialization. However, the hardware version dependent IO_IO_LOW 214 ** and IO_IO_HIGH registers should not be subsequently altered by the OS. 215 ** 216 ** Writes to both sets of registers will take effect immediately, bypassing 217 ** the queues, which ensures that subsequent Runway transactions are checked 218 ** against the updated bounds values. However reads are queued, introducing 219 ** the possibility of a read being bypassed by a subsequent write to the same 220 ** register. This sequence can be avoided by having software wait for read 221 ** returns before issuing subsequent writes. 222 */ 223 224 struct ioc { 225 struct ioa_registers __iomem *ioc_regs; /* I/O MMU base address */ 226 u8 *res_map; /* resource map, bit == pdir entry */ 227 u64 *pdir_base; /* physical base address */ 228 u32 pdir_size; /* bytes, function of IOV Space size */ 229 u32 res_hint; /* next available IOVP - 230 circular search */ 231 u32 res_size; /* size of resource map in bytes */ 232 spinlock_t res_lock; 233 234 #ifdef CCIO_COLLECT_STATS 235 #define CCIO_SEARCH_SAMPLE 0x100 236 unsigned long avg_search[CCIO_SEARCH_SAMPLE]; 237 unsigned long avg_idx; /* current index into avg_search */ 238 unsigned long used_pages; 239 unsigned long msingle_calls; 240 unsigned long msingle_pages; 241 unsigned long msg_calls; 242 unsigned long msg_pages; 243 unsigned long usingle_calls; 244 unsigned long usingle_pages; 245 unsigned long usg_calls; 246 unsigned long usg_pages; 247 #endif 248 unsigned short cujo20_bug; 249 250 /* STUFF We don't need in performance path */ 251 u32 chainid_shift; /* specify bit location of chain_id */ 252 struct ioc *next; /* Linked list of discovered iocs */ 253 const char *name; /* device name from firmware */ 254 unsigned int hw_path; /* the hardware path this ioc is associatd with */ 255 struct pci_dev *fake_pci_dev; /* the fake pci_dev for non-pci devs */ 256 struct resource mmio_region[2]; /* The "routed" MMIO regions */ 257 }; 258 259 static struct ioc *ioc_list; 260 static int ioc_count; 261 262 /************************************************************** 263 * 264 * I/O Pdir Resource Management 265 * 266 * Bits set in the resource map are in use. 267 * Each bit can represent a number of pages. 268 * LSbs represent lower addresses (IOVA's). 269 * 270 * This was was copied from sba_iommu.c. Don't try to unify 271 * the two resource managers unless a way to have different 272 * allocation policies is also adjusted. We'd like to avoid 273 * I/O TLB thrashing by having resource allocation policy 274 * match the I/O TLB replacement policy. 275 * 276 ***************************************************************/ 277 #define IOVP_SIZE PAGE_SIZE 278 #define IOVP_SHIFT PAGE_SHIFT 279 #define IOVP_MASK PAGE_MASK 280 281 /* Convert from IOVP to IOVA and vice versa. */ 282 #define CCIO_IOVA(iovp,offset) ((iovp) | (offset)) 283 #define CCIO_IOVP(iova) ((iova) & IOVP_MASK) 284 285 #define PDIR_INDEX(iovp) ((iovp)>>IOVP_SHIFT) 286 #define MKIOVP(pdir_idx) ((long)(pdir_idx) << IOVP_SHIFT) 287 #define MKIOVA(iovp,offset) (dma_addr_t)((long)iovp | (long)offset) 288 289 /* 290 ** Don't worry about the 150% average search length on a miss. 291 ** If the search wraps around, and passes the res_hint, it will 292 ** cause the kernel to panic anyhow. 293 */ 294 #define CCIO_SEARCH_LOOP(ioc, res_idx, mask, size) \ 295 for(; res_ptr < res_end; ++res_ptr) { \ 296 int ret;\ 297 unsigned int idx;\ 298 idx = (unsigned int)((unsigned long)res_ptr - (unsigned long)ioc->res_map); \ 299 ret = iommu_is_span_boundary(idx << 3, pages_needed, 0, boundary_size);\ 300 if ((0 == (*res_ptr & mask)) && !ret) { \ 301 *res_ptr |= mask; \ 302 res_idx = idx;\ 303 ioc->res_hint = res_idx + (size >> 3); \ 304 goto resource_found; \ 305 } \ 306 } 307 308 #define CCIO_FIND_FREE_MAPPING(ioa, res_idx, mask, size) \ 309 u##size *res_ptr = (u##size *)&((ioc)->res_map[ioa->res_hint & ~((size >> 3) - 1)]); \ 310 u##size *res_end = (u##size *)&(ioc)->res_map[ioa->res_size]; \ 311 CCIO_SEARCH_LOOP(ioc, res_idx, mask, size); \ 312 res_ptr = (u##size *)&(ioc)->res_map[0]; \ 313 CCIO_SEARCH_LOOP(ioa, res_idx, mask, size); 314 315 /* 316 ** Find available bit in this ioa's resource map. 317 ** Use a "circular" search: 318 ** o Most IOVA's are "temporary" - avg search time should be small. 319 ** o keep a history of what happened for debugging 320 ** o KISS. 321 ** 322 ** Perf optimizations: 323 ** o search for log2(size) bits at a time. 324 ** o search for available resource bits using byte/word/whatever. 325 ** o use different search for "large" (eg > 4 pages) or "very large" 326 ** (eg > 16 pages) mappings. 327 */ 328 329 /** 330 * ccio_alloc_range - Allocate pages in the ioc's resource map. 331 * @ioc: The I/O Controller. 332 * @pages_needed: The requested number of pages to be mapped into the 333 * I/O Pdir... 334 * 335 * This function searches the resource map of the ioc to locate a range 336 * of available pages for the requested size. 337 */ 338 static int 339 ccio_alloc_range(struct ioc *ioc, struct device *dev, size_t size) 340 { 341 unsigned int pages_needed = size >> IOVP_SHIFT; 342 unsigned int res_idx; 343 unsigned long boundary_size; 344 #ifdef CCIO_COLLECT_STATS 345 unsigned long cr_start = mfctl(16); 346 #endif 347 348 BUG_ON(pages_needed == 0); 349 BUG_ON((pages_needed * IOVP_SIZE) > DMA_CHUNK_SIZE); 350 351 DBG_RES("%s() size: %d pages_needed %d\n", 352 __func__, size, pages_needed); 353 354 /* 355 ** "seek and ye shall find"...praying never hurts either... 356 ** ggg sacrifices another 710 to the computer gods. 357 */ 358 359 boundary_size = ALIGN((unsigned long long)dma_get_seg_boundary(dev) + 1, 360 1ULL << IOVP_SHIFT) >> IOVP_SHIFT; 361 362 if (pages_needed <= 8) { 363 /* 364 * LAN traffic will not thrash the TLB IFF the same NIC 365 * uses 8 adjacent pages to map separate payload data. 366 * ie the same byte in the resource bit map. 367 */ 368 #if 0 369 /* FIXME: bit search should shift it's way through 370 * an unsigned long - not byte at a time. As it is now, 371 * we effectively allocate this byte to this mapping. 372 */ 373 unsigned long mask = ~(~0UL >> pages_needed); 374 CCIO_FIND_FREE_MAPPING(ioc, res_idx, mask, 8); 375 #else 376 CCIO_FIND_FREE_MAPPING(ioc, res_idx, 0xff, 8); 377 #endif 378 } else if (pages_needed <= 16) { 379 CCIO_FIND_FREE_MAPPING(ioc, res_idx, 0xffff, 16); 380 } else if (pages_needed <= 32) { 381 CCIO_FIND_FREE_MAPPING(ioc, res_idx, ~(unsigned int)0, 32); 382 #ifdef __LP64__ 383 } else if (pages_needed <= 64) { 384 CCIO_FIND_FREE_MAPPING(ioc, res_idx, ~0UL, 64); 385 #endif 386 } else { 387 panic("%s: %s() Too many pages to map. pages_needed: %u\n", 388 __FILE__, __func__, pages_needed); 389 } 390 391 panic("%s: %s() I/O MMU is out of mapping resources.\n", __FILE__, 392 __func__); 393 394 resource_found: 395 396 DBG_RES("%s() res_idx %d res_hint: %d\n", 397 __func__, res_idx, ioc->res_hint); 398 399 #ifdef CCIO_COLLECT_STATS 400 { 401 unsigned long cr_end = mfctl(16); 402 unsigned long tmp = cr_end - cr_start; 403 /* check for roll over */ 404 cr_start = (cr_end < cr_start) ? -(tmp) : (tmp); 405 } 406 ioc->avg_search[ioc->avg_idx++] = cr_start; 407 ioc->avg_idx &= CCIO_SEARCH_SAMPLE - 1; 408 ioc->used_pages += pages_needed; 409 #endif 410 /* 411 ** return the bit address. 412 */ 413 return res_idx << 3; 414 } 415 416 #define CCIO_FREE_MAPPINGS(ioc, res_idx, mask, size) \ 417 u##size *res_ptr = (u##size *)&((ioc)->res_map[res_idx]); \ 418 BUG_ON((*res_ptr & mask) != mask); \ 419 *res_ptr &= ~(mask); 420 421 /** 422 * ccio_free_range - Free pages from the ioc's resource map. 423 * @ioc: The I/O Controller. 424 * @iova: The I/O Virtual Address. 425 * @pages_mapped: The requested number of pages to be freed from the 426 * I/O Pdir. 427 * 428 * This function frees the resouces allocated for the iova. 429 */ 430 static void 431 ccio_free_range(struct ioc *ioc, dma_addr_t iova, unsigned long pages_mapped) 432 { 433 unsigned long iovp = CCIO_IOVP(iova); 434 unsigned int res_idx = PDIR_INDEX(iovp) >> 3; 435 436 BUG_ON(pages_mapped == 0); 437 BUG_ON((pages_mapped * IOVP_SIZE) > DMA_CHUNK_SIZE); 438 BUG_ON(pages_mapped > BITS_PER_LONG); 439 440 DBG_RES("%s(): res_idx: %d pages_mapped %d\n", 441 __func__, res_idx, pages_mapped); 442 443 #ifdef CCIO_COLLECT_STATS 444 ioc->used_pages -= pages_mapped; 445 #endif 446 447 if(pages_mapped <= 8) { 448 #if 0 449 /* see matching comments in alloc_range */ 450 unsigned long mask = ~(~0UL >> pages_mapped); 451 CCIO_FREE_MAPPINGS(ioc, res_idx, mask, 8); 452 #else 453 CCIO_FREE_MAPPINGS(ioc, res_idx, 0xffUL, 8); 454 #endif 455 } else if(pages_mapped <= 16) { 456 CCIO_FREE_MAPPINGS(ioc, res_idx, 0xffffUL, 16); 457 } else if(pages_mapped <= 32) { 458 CCIO_FREE_MAPPINGS(ioc, res_idx, ~(unsigned int)0, 32); 459 #ifdef __LP64__ 460 } else if(pages_mapped <= 64) { 461 CCIO_FREE_MAPPINGS(ioc, res_idx, ~0UL, 64); 462 #endif 463 } else { 464 panic("%s:%s() Too many pages to unmap.\n", __FILE__, 465 __func__); 466 } 467 } 468 469 /**************************************************************** 470 ** 471 ** CCIO dma_ops support routines 472 ** 473 *****************************************************************/ 474 475 typedef unsigned long space_t; 476 #define KERNEL_SPACE 0 477 478 /* 479 ** DMA "Page Type" and Hints 480 ** o if SAFE_DMA isn't set, mapping is for FAST_DMA. SAFE_DMA should be 481 ** set for subcacheline DMA transfers since we don't want to damage the 482 ** other part of a cacheline. 483 ** o SAFE_DMA must be set for "memory" allocated via pci_alloc_consistent(). 484 ** This bit tells U2 to do R/M/W for partial cachelines. "Streaming" 485 ** data can avoid this if the mapping covers full cache lines. 486 ** o STOP_MOST is needed for atomicity across cachelines. 487 ** Apparently only "some EISA devices" need this. 488 ** Using CONFIG_ISA is hack. Only the IOA with EISA under it needs 489 ** to use this hint iff the EISA devices needs this feature. 490 ** According to the U2 ERS, STOP_MOST enabled pages hurt performance. 491 ** o PREFETCH should *not* be set for cases like Multiple PCI devices 492 ** behind GSCtoPCI (dino) bus converter. Only one cacheline per GSC 493 ** device can be fetched and multiply DMA streams will thrash the 494 ** prefetch buffer and burn memory bandwidth. See 6.7.3 "Prefetch Rules 495 ** and Invalidation of Prefetch Entries". 496 ** 497 ** FIXME: the default hints need to be per GSC device - not global. 498 ** 499 ** HP-UX dorks: linux device driver programming model is totally different 500 ** than HP-UX's. HP-UX always sets HINT_PREFETCH since it's drivers 501 ** do special things to work on non-coherent platforms...linux has to 502 ** be much more careful with this. 503 */ 504 #define IOPDIR_VALID 0x01UL 505 #define HINT_SAFE_DMA 0x02UL /* used for pci_alloc_consistent() pages */ 506 #ifdef CONFIG_EISA 507 #define HINT_STOP_MOST 0x04UL /* LSL support */ 508 #else 509 #define HINT_STOP_MOST 0x00UL /* only needed for "some EISA devices" */ 510 #endif 511 #define HINT_UDPATE_ENB 0x08UL /* not used/supported by U2 */ 512 #define HINT_PREFETCH 0x10UL /* for outbound pages which are not SAFE */ 513 514 515 /* 516 ** Use direction (ie PCI_DMA_TODEVICE) to pick hint. 517 ** ccio_alloc_consistent() depends on this to get SAFE_DMA 518 ** when it passes in BIDIRECTIONAL flag. 519 */ 520 static u32 hint_lookup[] = { 521 [PCI_DMA_BIDIRECTIONAL] = HINT_STOP_MOST | HINT_SAFE_DMA | IOPDIR_VALID, 522 [PCI_DMA_TODEVICE] = HINT_STOP_MOST | HINT_PREFETCH | IOPDIR_VALID, 523 [PCI_DMA_FROMDEVICE] = HINT_STOP_MOST | IOPDIR_VALID, 524 }; 525 526 /** 527 * ccio_io_pdir_entry - Initialize an I/O Pdir. 528 * @pdir_ptr: A pointer into I/O Pdir. 529 * @sid: The Space Identifier. 530 * @vba: The virtual address. 531 * @hints: The DMA Hint. 532 * 533 * Given a virtual address (vba, arg2) and space id, (sid, arg1), 534 * load the I/O PDIR entry pointed to by pdir_ptr (arg0). Each IO Pdir 535 * entry consists of 8 bytes as shown below (MSB == bit 0): 536 * 537 * 538 * WORD 0: 539 * +------+----------------+-----------------------------------------------+ 540 * | Phys | Virtual Index | Phys | 541 * | 0:3 | 0:11 | 4:19 | 542 * |4 bits| 12 bits | 16 bits | 543 * +------+----------------+-----------------------------------------------+ 544 * WORD 1: 545 * +-----------------------+-----------------------------------------------+ 546 * | Phys | Rsvd | Prefetch |Update |Rsvd |Lock |Safe |Valid | 547 * | 20:39 | | Enable |Enable | |Enable|DMA | | 548 * | 20 bits | 5 bits | 1 bit |1 bit |2 bits|1 bit |1 bit |1 bit | 549 * +-----------------------+-----------------------------------------------+ 550 * 551 * The virtual index field is filled with the results of the LCI 552 * (Load Coherence Index) instruction. The 8 bits used for the virtual 553 * index are bits 12:19 of the value returned by LCI. 554 */ 555 static void CCIO_INLINE 556 ccio_io_pdir_entry(u64 *pdir_ptr, space_t sid, unsigned long vba, 557 unsigned long hints) 558 { 559 register unsigned long pa; 560 register unsigned long ci; /* coherent index */ 561 562 /* We currently only support kernel addresses */ 563 BUG_ON(sid != KERNEL_SPACE); 564 565 /* 566 ** WORD 1 - low order word 567 ** "hints" parm includes the VALID bit! 568 ** "dep" clobbers the physical address offset bits as well. 569 */ 570 pa = lpa(vba); 571 asm volatile("depw %1,31,12,%0" : "+r" (pa) : "r" (hints)); 572 ((u32 *)pdir_ptr)[1] = (u32) pa; 573 574 /* 575 ** WORD 0 - high order word 576 */ 577 578 #ifdef __LP64__ 579 /* 580 ** get bits 12:15 of physical address 581 ** shift bits 16:31 of physical address 582 ** and deposit them 583 */ 584 asm volatile ("extrd,u %1,15,4,%0" : "=r" (ci) : "r" (pa)); 585 asm volatile ("extrd,u %1,31,16,%0" : "+r" (pa) : "r" (pa)); 586 asm volatile ("depd %1,35,4,%0" : "+r" (pa) : "r" (ci)); 587 #else 588 pa = 0; 589 #endif 590 /* 591 ** get CPU coherency index bits 592 ** Grab virtual index [0:11] 593 ** Deposit virt_idx bits into I/O PDIR word 594 */ 595 asm volatile ("lci %%r0(%1), %0" : "=r" (ci) : "r" (vba)); 596 asm volatile ("extru %1,19,12,%0" : "+r" (ci) : "r" (ci)); 597 asm volatile ("depw %1,15,12,%0" : "+r" (pa) : "r" (ci)); 598 599 ((u32 *)pdir_ptr)[0] = (u32) pa; 600 601 602 /* FIXME: PCX_W platforms don't need FDC/SYNC. (eg C360) 603 ** PCX-U/U+ do. (eg C200/C240) 604 ** PCX-T'? Don't know. (eg C110 or similar K-class) 605 ** 606 ** See PDC_MODEL/option 0/SW_CAP word for "Non-coherent IO-PDIR bit". 607 ** 608 ** "Since PCX-U employs an offset hash that is incompatible with 609 ** the real mode coherence index generation of U2, the PDIR entry 610 ** must be flushed to memory to retain coherence." 611 */ 612 asm_io_fdc(pdir_ptr); 613 asm_io_sync(); 614 } 615 616 /** 617 * ccio_clear_io_tlb - Remove stale entries from the I/O TLB. 618 * @ioc: The I/O Controller. 619 * @iovp: The I/O Virtual Page. 620 * @byte_cnt: The requested number of bytes to be freed from the I/O Pdir. 621 * 622 * Purge invalid I/O PDIR entries from the I/O TLB. 623 * 624 * FIXME: Can we change the byte_cnt to pages_mapped? 625 */ 626 static CCIO_INLINE void 627 ccio_clear_io_tlb(struct ioc *ioc, dma_addr_t iovp, size_t byte_cnt) 628 { 629 u32 chain_size = 1 << ioc->chainid_shift; 630 631 iovp &= IOVP_MASK; /* clear offset bits, just want pagenum */ 632 byte_cnt += chain_size; 633 634 while(byte_cnt > chain_size) { 635 WRITE_U32(CMD_TLB_PURGE | iovp, &ioc->ioc_regs->io_command); 636 iovp += chain_size; 637 byte_cnt -= chain_size; 638 } 639 } 640 641 /** 642 * ccio_mark_invalid - Mark the I/O Pdir entries invalid. 643 * @ioc: The I/O Controller. 644 * @iova: The I/O Virtual Address. 645 * @byte_cnt: The requested number of bytes to be freed from the I/O Pdir. 646 * 647 * Mark the I/O Pdir entries invalid and blow away the corresponding I/O 648 * TLB entries. 649 * 650 * FIXME: at some threshold it might be "cheaper" to just blow 651 * away the entire I/O TLB instead of individual entries. 652 * 653 * FIXME: Uturn has 256 TLB entries. We don't need to purge every 654 * PDIR entry - just once for each possible TLB entry. 655 * (We do need to maker I/O PDIR entries invalid regardless). 656 * 657 * FIXME: Can we change byte_cnt to pages_mapped? 658 */ 659 static CCIO_INLINE void 660 ccio_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt) 661 { 662 u32 iovp = (u32)CCIO_IOVP(iova); 663 size_t saved_byte_cnt; 664 665 /* round up to nearest page size */ 666 saved_byte_cnt = byte_cnt = ALIGN(byte_cnt, IOVP_SIZE); 667 668 while(byte_cnt > 0) { 669 /* invalidate one page at a time */ 670 unsigned int idx = PDIR_INDEX(iovp); 671 char *pdir_ptr = (char *) &(ioc->pdir_base[idx]); 672 673 BUG_ON(idx >= (ioc->pdir_size / sizeof(u64))); 674 pdir_ptr[7] = 0; /* clear only VALID bit */ 675 /* 676 ** FIXME: PCX_W platforms don't need FDC/SYNC. (eg C360) 677 ** PCX-U/U+ do. (eg C200/C240) 678 ** See PDC_MODEL/option 0/SW_CAP for "Non-coherent IO-PDIR bit". 679 */ 680 asm_io_fdc(pdir_ptr); 681 682 iovp += IOVP_SIZE; 683 byte_cnt -= IOVP_SIZE; 684 } 685 686 asm_io_sync(); 687 ccio_clear_io_tlb(ioc, CCIO_IOVP(iova), saved_byte_cnt); 688 } 689 690 /**************************************************************** 691 ** 692 ** CCIO dma_ops 693 ** 694 *****************************************************************/ 695 696 /** 697 * ccio_dma_supported - Verify the IOMMU supports the DMA address range. 698 * @dev: The PCI device. 699 * @mask: A bit mask describing the DMA address range of the device. 700 */ 701 static int 702 ccio_dma_supported(struct device *dev, u64 mask) 703 { 704 if(dev == NULL) { 705 printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n"); 706 BUG(); 707 return 0; 708 } 709 710 /* only support 32-bit or better devices (ie PCI/GSC) */ 711 return (int)(mask >= 0xffffffffUL); 712 } 713 714 /** 715 * ccio_map_single - Map an address range into the IOMMU. 716 * @dev: The PCI device. 717 * @addr: The start address of the DMA region. 718 * @size: The length of the DMA region. 719 * @direction: The direction of the DMA transaction (to/from device). 720 * 721 * This function implements the pci_map_single function. 722 */ 723 static dma_addr_t 724 ccio_map_single(struct device *dev, void *addr, size_t size, 725 enum dma_data_direction direction) 726 { 727 int idx; 728 struct ioc *ioc; 729 unsigned long flags; 730 dma_addr_t iovp; 731 dma_addr_t offset; 732 u64 *pdir_start; 733 unsigned long hint = hint_lookup[(int)direction]; 734 735 BUG_ON(!dev); 736 ioc = GET_IOC(dev); 737 if (!ioc) 738 return DMA_MAPPING_ERROR; 739 740 BUG_ON(size <= 0); 741 742 /* save offset bits */ 743 offset = ((unsigned long) addr) & ~IOVP_MASK; 744 745 /* round up to nearest IOVP_SIZE */ 746 size = ALIGN(size + offset, IOVP_SIZE); 747 spin_lock_irqsave(&ioc->res_lock, flags); 748 749 #ifdef CCIO_COLLECT_STATS 750 ioc->msingle_calls++; 751 ioc->msingle_pages += size >> IOVP_SHIFT; 752 #endif 753 754 idx = ccio_alloc_range(ioc, dev, size); 755 iovp = (dma_addr_t)MKIOVP(idx); 756 757 pdir_start = &(ioc->pdir_base[idx]); 758 759 DBG_RUN("%s() 0x%p -> 0x%lx size: %0x%x\n", 760 __func__, addr, (long)iovp | offset, size); 761 762 /* If not cacheline aligned, force SAFE_DMA on the whole mess */ 763 if((size % L1_CACHE_BYTES) || ((unsigned long)addr % L1_CACHE_BYTES)) 764 hint |= HINT_SAFE_DMA; 765 766 while(size > 0) { 767 ccio_io_pdir_entry(pdir_start, KERNEL_SPACE, (unsigned long)addr, hint); 768 769 DBG_RUN(" pdir %p %08x%08x\n", 770 pdir_start, 771 (u32) (((u32 *) pdir_start)[0]), 772 (u32) (((u32 *) pdir_start)[1])); 773 ++pdir_start; 774 addr += IOVP_SIZE; 775 size -= IOVP_SIZE; 776 } 777 778 spin_unlock_irqrestore(&ioc->res_lock, flags); 779 780 /* form complete address */ 781 return CCIO_IOVA(iovp, offset); 782 } 783 784 785 static dma_addr_t 786 ccio_map_page(struct device *dev, struct page *page, unsigned long offset, 787 size_t size, enum dma_data_direction direction, 788 unsigned long attrs) 789 { 790 return ccio_map_single(dev, page_address(page) + offset, size, 791 direction); 792 } 793 794 795 /** 796 * ccio_unmap_page - Unmap an address range from the IOMMU. 797 * @dev: The PCI device. 798 * @addr: The start address of the DMA region. 799 * @size: The length of the DMA region. 800 * @direction: The direction of the DMA transaction (to/from device). 801 */ 802 static void 803 ccio_unmap_page(struct device *dev, dma_addr_t iova, size_t size, 804 enum dma_data_direction direction, unsigned long attrs) 805 { 806 struct ioc *ioc; 807 unsigned long flags; 808 dma_addr_t offset = iova & ~IOVP_MASK; 809 810 BUG_ON(!dev); 811 ioc = GET_IOC(dev); 812 if (!ioc) { 813 WARN_ON(!ioc); 814 return; 815 } 816 817 DBG_RUN("%s() iovp 0x%lx/%x\n", 818 __func__, (long)iova, size); 819 820 iova ^= offset; /* clear offset bits */ 821 size += offset; 822 size = ALIGN(size, IOVP_SIZE); 823 824 spin_lock_irqsave(&ioc->res_lock, flags); 825 826 #ifdef CCIO_COLLECT_STATS 827 ioc->usingle_calls++; 828 ioc->usingle_pages += size >> IOVP_SHIFT; 829 #endif 830 831 ccio_mark_invalid(ioc, iova, size); 832 ccio_free_range(ioc, iova, (size >> IOVP_SHIFT)); 833 spin_unlock_irqrestore(&ioc->res_lock, flags); 834 } 835 836 /** 837 * ccio_alloc - Allocate a consistent DMA mapping. 838 * @dev: The PCI device. 839 * @size: The length of the DMA region. 840 * @dma_handle: The DMA address handed back to the device (not the cpu). 841 * 842 * This function implements the pci_alloc_consistent function. 843 */ 844 static void * 845 ccio_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle, gfp_t flag, 846 unsigned long attrs) 847 { 848 void *ret; 849 #if 0 850 /* GRANT Need to establish hierarchy for non-PCI devs as well 851 ** and then provide matching gsc_map_xxx() functions for them as well. 852 */ 853 if(!hwdev) { 854 /* only support PCI */ 855 *dma_handle = 0; 856 return 0; 857 } 858 #endif 859 ret = (void *) __get_free_pages(flag, get_order(size)); 860 861 if (ret) { 862 memset(ret, 0, size); 863 *dma_handle = ccio_map_single(dev, ret, size, PCI_DMA_BIDIRECTIONAL); 864 } 865 866 return ret; 867 } 868 869 /** 870 * ccio_free - Free a consistent DMA mapping. 871 * @dev: The PCI device. 872 * @size: The length of the DMA region. 873 * @cpu_addr: The cpu address returned from the ccio_alloc_consistent. 874 * @dma_handle: The device address returned from the ccio_alloc_consistent. 875 * 876 * This function implements the pci_free_consistent function. 877 */ 878 static void 879 ccio_free(struct device *dev, size_t size, void *cpu_addr, 880 dma_addr_t dma_handle, unsigned long attrs) 881 { 882 ccio_unmap_page(dev, dma_handle, size, 0, 0); 883 free_pages((unsigned long)cpu_addr, get_order(size)); 884 } 885 886 /* 887 ** Since 0 is a valid pdir_base index value, can't use that 888 ** to determine if a value is valid or not. Use a flag to indicate 889 ** the SG list entry contains a valid pdir index. 890 */ 891 #define PIDE_FLAG 0x80000000UL 892 893 #ifdef CCIO_COLLECT_STATS 894 #define IOMMU_MAP_STATS 895 #endif 896 #include "iommu-helpers.h" 897 898 /** 899 * ccio_map_sg - Map the scatter/gather list into the IOMMU. 900 * @dev: The PCI device. 901 * @sglist: The scatter/gather list to be mapped in the IOMMU. 902 * @nents: The number of entries in the scatter/gather list. 903 * @direction: The direction of the DMA transaction (to/from device). 904 * 905 * This function implements the pci_map_sg function. 906 */ 907 static int 908 ccio_map_sg(struct device *dev, struct scatterlist *sglist, int nents, 909 enum dma_data_direction direction, unsigned long attrs) 910 { 911 struct ioc *ioc; 912 int coalesced, filled = 0; 913 unsigned long flags; 914 unsigned long hint = hint_lookup[(int)direction]; 915 unsigned long prev_len = 0, current_len = 0; 916 int i; 917 918 BUG_ON(!dev); 919 ioc = GET_IOC(dev); 920 if (!ioc) 921 return 0; 922 923 DBG_RUN_SG("%s() START %d entries\n", __func__, nents); 924 925 /* Fast path single entry scatterlists. */ 926 if (nents == 1) { 927 sg_dma_address(sglist) = ccio_map_single(dev, 928 sg_virt(sglist), sglist->length, 929 direction); 930 sg_dma_len(sglist) = sglist->length; 931 return 1; 932 } 933 934 for(i = 0; i < nents; i++) 935 prev_len += sglist[i].length; 936 937 spin_lock_irqsave(&ioc->res_lock, flags); 938 939 #ifdef CCIO_COLLECT_STATS 940 ioc->msg_calls++; 941 #endif 942 943 /* 944 ** First coalesce the chunks and allocate I/O pdir space 945 ** 946 ** If this is one DMA stream, we can properly map using the 947 ** correct virtual address associated with each DMA page. 948 ** w/o this association, we wouldn't have coherent DMA! 949 ** Access to the virtual address is what forces a two pass algorithm. 950 */ 951 coalesced = iommu_coalesce_chunks(ioc, dev, sglist, nents, ccio_alloc_range); 952 953 /* 954 ** Program the I/O Pdir 955 ** 956 ** map the virtual addresses to the I/O Pdir 957 ** o dma_address will contain the pdir index 958 ** o dma_len will contain the number of bytes to map 959 ** o page/offset contain the virtual address. 960 */ 961 filled = iommu_fill_pdir(ioc, sglist, nents, hint, ccio_io_pdir_entry); 962 963 spin_unlock_irqrestore(&ioc->res_lock, flags); 964 965 BUG_ON(coalesced != filled); 966 967 DBG_RUN_SG("%s() DONE %d mappings\n", __func__, filled); 968 969 for (i = 0; i < filled; i++) 970 current_len += sg_dma_len(sglist + i); 971 972 BUG_ON(current_len != prev_len); 973 974 return filled; 975 } 976 977 /** 978 * ccio_unmap_sg - Unmap the scatter/gather list from the IOMMU. 979 * @dev: The PCI device. 980 * @sglist: The scatter/gather list to be unmapped from the IOMMU. 981 * @nents: The number of entries in the scatter/gather list. 982 * @direction: The direction of the DMA transaction (to/from device). 983 * 984 * This function implements the pci_unmap_sg function. 985 */ 986 static void 987 ccio_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents, 988 enum dma_data_direction direction, unsigned long attrs) 989 { 990 struct ioc *ioc; 991 992 BUG_ON(!dev); 993 ioc = GET_IOC(dev); 994 if (!ioc) { 995 WARN_ON(!ioc); 996 return; 997 } 998 999 DBG_RUN_SG("%s() START %d entries, %p,%x\n", 1000 __func__, nents, sg_virt(sglist), sglist->length); 1001 1002 #ifdef CCIO_COLLECT_STATS 1003 ioc->usg_calls++; 1004 #endif 1005 1006 while(sg_dma_len(sglist) && nents--) { 1007 1008 #ifdef CCIO_COLLECT_STATS 1009 ioc->usg_pages += sg_dma_len(sglist) >> PAGE_SHIFT; 1010 #endif 1011 ccio_unmap_page(dev, sg_dma_address(sglist), 1012 sg_dma_len(sglist), direction, 0); 1013 ++sglist; 1014 } 1015 1016 DBG_RUN_SG("%s() DONE (nents %d)\n", __func__, nents); 1017 } 1018 1019 static const struct dma_map_ops ccio_ops = { 1020 .dma_supported = ccio_dma_supported, 1021 .alloc = ccio_alloc, 1022 .free = ccio_free, 1023 .map_page = ccio_map_page, 1024 .unmap_page = ccio_unmap_page, 1025 .map_sg = ccio_map_sg, 1026 .unmap_sg = ccio_unmap_sg, 1027 .get_sgtable = dma_common_get_sgtable, 1028 }; 1029 1030 #ifdef CONFIG_PROC_FS 1031 static int ccio_proc_info(struct seq_file *m, void *p) 1032 { 1033 struct ioc *ioc = ioc_list; 1034 1035 while (ioc != NULL) { 1036 unsigned int total_pages = ioc->res_size << 3; 1037 #ifdef CCIO_COLLECT_STATS 1038 unsigned long avg = 0, min, max; 1039 int j; 1040 #endif 1041 1042 seq_printf(m, "%s\n", ioc->name); 1043 1044 seq_printf(m, "Cujo 2.0 bug : %s\n", 1045 (ioc->cujo20_bug ? "yes" : "no")); 1046 1047 seq_printf(m, "IO PDIR size : %d bytes (%d entries)\n", 1048 total_pages * 8, total_pages); 1049 1050 #ifdef CCIO_COLLECT_STATS 1051 seq_printf(m, "IO PDIR entries : %ld free %ld used (%d%%)\n", 1052 total_pages - ioc->used_pages, ioc->used_pages, 1053 (int)(ioc->used_pages * 100 / total_pages)); 1054 #endif 1055 1056 seq_printf(m, "Resource bitmap : %d bytes (%d pages)\n", 1057 ioc->res_size, total_pages); 1058 1059 #ifdef CCIO_COLLECT_STATS 1060 min = max = ioc->avg_search[0]; 1061 for(j = 0; j < CCIO_SEARCH_SAMPLE; ++j) { 1062 avg += ioc->avg_search[j]; 1063 if(ioc->avg_search[j] > max) 1064 max = ioc->avg_search[j]; 1065 if(ioc->avg_search[j] < min) 1066 min = ioc->avg_search[j]; 1067 } 1068 avg /= CCIO_SEARCH_SAMPLE; 1069 seq_printf(m, " Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles)\n", 1070 min, avg, max); 1071 1072 seq_printf(m, "pci_map_single(): %8ld calls %8ld pages (avg %d/1000)\n", 1073 ioc->msingle_calls, ioc->msingle_pages, 1074 (int)((ioc->msingle_pages * 1000)/ioc->msingle_calls)); 1075 1076 /* KLUGE - unmap_sg calls unmap_page for each mapped page */ 1077 min = ioc->usingle_calls - ioc->usg_calls; 1078 max = ioc->usingle_pages - ioc->usg_pages; 1079 seq_printf(m, "pci_unmap_single: %8ld calls %8ld pages (avg %d/1000)\n", 1080 min, max, (int)((max * 1000)/min)); 1081 1082 seq_printf(m, "pci_map_sg() : %8ld calls %8ld pages (avg %d/1000)\n", 1083 ioc->msg_calls, ioc->msg_pages, 1084 (int)((ioc->msg_pages * 1000)/ioc->msg_calls)); 1085 1086 seq_printf(m, "pci_unmap_sg() : %8ld calls %8ld pages (avg %d/1000)\n\n\n", 1087 ioc->usg_calls, ioc->usg_pages, 1088 (int)((ioc->usg_pages * 1000)/ioc->usg_calls)); 1089 #endif /* CCIO_COLLECT_STATS */ 1090 1091 ioc = ioc->next; 1092 } 1093 1094 return 0; 1095 } 1096 1097 static int ccio_proc_bitmap_info(struct seq_file *m, void *p) 1098 { 1099 struct ioc *ioc = ioc_list; 1100 1101 while (ioc != NULL) { 1102 seq_hex_dump(m, " ", DUMP_PREFIX_NONE, 32, 4, ioc->res_map, 1103 ioc->res_size, false); 1104 seq_putc(m, '\n'); 1105 ioc = ioc->next; 1106 break; /* XXX - remove me */ 1107 } 1108 1109 return 0; 1110 } 1111 #endif /* CONFIG_PROC_FS */ 1112 1113 /** 1114 * ccio_find_ioc - Find the ioc in the ioc_list 1115 * @hw_path: The hardware path of the ioc. 1116 * 1117 * This function searches the ioc_list for an ioc that matches 1118 * the provide hardware path. 1119 */ 1120 static struct ioc * ccio_find_ioc(int hw_path) 1121 { 1122 int i; 1123 struct ioc *ioc; 1124 1125 ioc = ioc_list; 1126 for (i = 0; i < ioc_count; i++) { 1127 if (ioc->hw_path == hw_path) 1128 return ioc; 1129 1130 ioc = ioc->next; 1131 } 1132 1133 return NULL; 1134 } 1135 1136 /** 1137 * ccio_get_iommu - Find the iommu which controls this device 1138 * @dev: The parisc device. 1139 * 1140 * This function searches through the registered IOMMU's and returns 1141 * the appropriate IOMMU for the device based on its hardware path. 1142 */ 1143 void * ccio_get_iommu(const struct parisc_device *dev) 1144 { 1145 dev = find_pa_parent_type(dev, HPHW_IOA); 1146 if (!dev) 1147 return NULL; 1148 1149 return ccio_find_ioc(dev->hw_path); 1150 } 1151 1152 #define CUJO_20_STEP 0x10000000 /* inc upper nibble */ 1153 1154 /* Cujo 2.0 has a bug which will silently corrupt data being transferred 1155 * to/from certain pages. To avoid this happening, we mark these pages 1156 * as `used', and ensure that nothing will try to allocate from them. 1157 */ 1158 void __init ccio_cujo20_fixup(struct parisc_device *cujo, u32 iovp) 1159 { 1160 unsigned int idx; 1161 struct parisc_device *dev = parisc_parent(cujo); 1162 struct ioc *ioc = ccio_get_iommu(dev); 1163 u8 *res_ptr; 1164 1165 ioc->cujo20_bug = 1; 1166 res_ptr = ioc->res_map; 1167 idx = PDIR_INDEX(iovp) >> 3; 1168 1169 while (idx < ioc->res_size) { 1170 res_ptr[idx] |= 0xff; 1171 idx += PDIR_INDEX(CUJO_20_STEP) >> 3; 1172 } 1173 } 1174 1175 #if 0 1176 /* GRANT - is this needed for U2 or not? */ 1177 1178 /* 1179 ** Get the size of the I/O TLB for this I/O MMU. 1180 ** 1181 ** If spa_shift is non-zero (ie probably U2), 1182 ** then calculate the I/O TLB size using spa_shift. 1183 ** 1184 ** Otherwise we are supposed to get the IODC entry point ENTRY TLB 1185 ** and execute it. However, both U2 and Uturn firmware supplies spa_shift. 1186 ** I think only Java (K/D/R-class too?) systems don't do this. 1187 */ 1188 static int 1189 ccio_get_iotlb_size(struct parisc_device *dev) 1190 { 1191 if (dev->spa_shift == 0) { 1192 panic("%s() : Can't determine I/O TLB size.\n", __func__); 1193 } 1194 return (1 << dev->spa_shift); 1195 } 1196 #else 1197 1198 /* Uturn supports 256 TLB entries */ 1199 #define CCIO_CHAINID_SHIFT 8 1200 #define CCIO_CHAINID_MASK 0xff 1201 #endif /* 0 */ 1202 1203 /* We *can't* support JAVA (T600). Venture there at your own risk. */ 1204 static const struct parisc_device_id ccio_tbl[] __initconst = { 1205 { HPHW_IOA, HVERSION_REV_ANY_ID, U2_IOA_RUNWAY, 0xb }, /* U2 */ 1206 { HPHW_IOA, HVERSION_REV_ANY_ID, UTURN_IOA_RUNWAY, 0xb }, /* UTurn */ 1207 { 0, } 1208 }; 1209 1210 static int ccio_probe(struct parisc_device *dev); 1211 1212 static struct parisc_driver ccio_driver __refdata = { 1213 .name = "ccio", 1214 .id_table = ccio_tbl, 1215 .probe = ccio_probe, 1216 }; 1217 1218 /** 1219 * ccio_ioc_init - Initialize the I/O Controller 1220 * @ioc: The I/O Controller. 1221 * 1222 * Initialize the I/O Controller which includes setting up the 1223 * I/O Page Directory, the resource map, and initalizing the 1224 * U2/Uturn chip into virtual mode. 1225 */ 1226 static void __init 1227 ccio_ioc_init(struct ioc *ioc) 1228 { 1229 int i; 1230 unsigned int iov_order; 1231 u32 iova_space_size; 1232 1233 /* 1234 ** Determine IOVA Space size from memory size. 1235 ** 1236 ** Ideally, PCI drivers would register the maximum number 1237 ** of DMA they can have outstanding for each device they 1238 ** own. Next best thing would be to guess how much DMA 1239 ** can be outstanding based on PCI Class/sub-class. Both 1240 ** methods still require some "extra" to support PCI 1241 ** Hot-Plug/Removal of PCI cards. (aka PCI OLARD). 1242 */ 1243 1244 iova_space_size = (u32) (totalram_pages() / count_parisc_driver(&ccio_driver)); 1245 1246 /* limit IOVA space size to 1MB-1GB */ 1247 1248 if (iova_space_size < (1 << (20 - PAGE_SHIFT))) { 1249 iova_space_size = 1 << (20 - PAGE_SHIFT); 1250 #ifdef __LP64__ 1251 } else if (iova_space_size > (1 << (30 - PAGE_SHIFT))) { 1252 iova_space_size = 1 << (30 - PAGE_SHIFT); 1253 #endif 1254 } 1255 1256 /* 1257 ** iova space must be log2() in size. 1258 ** thus, pdir/res_map will also be log2(). 1259 */ 1260 1261 /* We could use larger page sizes in order to *decrease* the number 1262 ** of mappings needed. (ie 8k pages means 1/2 the mappings). 1263 ** 1264 ** Note: Grant Grunder says "Using 8k I/O pages isn't trivial either 1265 ** since the pages must also be physically contiguous - typically 1266 ** this is the case under linux." 1267 */ 1268 1269 iov_order = get_order(iova_space_size << PAGE_SHIFT); 1270 1271 /* iova_space_size is now bytes, not pages */ 1272 iova_space_size = 1 << (iov_order + PAGE_SHIFT); 1273 1274 ioc->pdir_size = (iova_space_size / IOVP_SIZE) * sizeof(u64); 1275 1276 BUG_ON(ioc->pdir_size > 8 * 1024 * 1024); /* max pdir size <= 8MB */ 1277 1278 /* Verify it's a power of two */ 1279 BUG_ON((1 << get_order(ioc->pdir_size)) != (ioc->pdir_size >> PAGE_SHIFT)); 1280 1281 DBG_INIT("%s() hpa 0x%p mem %luMB IOV %dMB (%d bits)\n", 1282 __func__, ioc->ioc_regs, 1283 (unsigned long) totalram_pages() >> (20 - PAGE_SHIFT), 1284 iova_space_size>>20, 1285 iov_order + PAGE_SHIFT); 1286 1287 ioc->pdir_base = (u64 *)__get_free_pages(GFP_KERNEL, 1288 get_order(ioc->pdir_size)); 1289 if(NULL == ioc->pdir_base) { 1290 panic("%s() could not allocate I/O Page Table\n", __func__); 1291 } 1292 memset(ioc->pdir_base, 0, ioc->pdir_size); 1293 1294 BUG_ON((((unsigned long)ioc->pdir_base) & PAGE_MASK) != (unsigned long)ioc->pdir_base); 1295 DBG_INIT(" base %p\n", ioc->pdir_base); 1296 1297 /* resource map size dictated by pdir_size */ 1298 ioc->res_size = (ioc->pdir_size / sizeof(u64)) >> 3; 1299 DBG_INIT("%s() res_size 0x%x\n", __func__, ioc->res_size); 1300 1301 ioc->res_map = (u8 *)__get_free_pages(GFP_KERNEL, 1302 get_order(ioc->res_size)); 1303 if(NULL == ioc->res_map) { 1304 panic("%s() could not allocate resource map\n", __func__); 1305 } 1306 memset(ioc->res_map, 0, ioc->res_size); 1307 1308 /* Initialize the res_hint to 16 */ 1309 ioc->res_hint = 16; 1310 1311 /* Initialize the spinlock */ 1312 spin_lock_init(&ioc->res_lock); 1313 1314 /* 1315 ** Chainid is the upper most bits of an IOVP used to determine 1316 ** which TLB entry an IOVP will use. 1317 */ 1318 ioc->chainid_shift = get_order(iova_space_size) + PAGE_SHIFT - CCIO_CHAINID_SHIFT; 1319 DBG_INIT(" chainid_shift 0x%x\n", ioc->chainid_shift); 1320 1321 /* 1322 ** Initialize IOA hardware 1323 */ 1324 WRITE_U32(CCIO_CHAINID_MASK << ioc->chainid_shift, 1325 &ioc->ioc_regs->io_chain_id_mask); 1326 1327 WRITE_U32(virt_to_phys(ioc->pdir_base), 1328 &ioc->ioc_regs->io_pdir_base); 1329 1330 /* 1331 ** Go to "Virtual Mode" 1332 */ 1333 WRITE_U32(IOA_NORMAL_MODE, &ioc->ioc_regs->io_control); 1334 1335 /* 1336 ** Initialize all I/O TLB entries to 0 (Valid bit off). 1337 */ 1338 WRITE_U32(0, &ioc->ioc_regs->io_tlb_entry_m); 1339 WRITE_U32(0, &ioc->ioc_regs->io_tlb_entry_l); 1340 1341 for(i = 1 << CCIO_CHAINID_SHIFT; i ; i--) { 1342 WRITE_U32((CMD_TLB_DIRECT_WRITE | (i << ioc->chainid_shift)), 1343 &ioc->ioc_regs->io_command); 1344 } 1345 } 1346 1347 static void __init 1348 ccio_init_resource(struct resource *res, char *name, void __iomem *ioaddr) 1349 { 1350 int result; 1351 1352 res->parent = NULL; 1353 res->flags = IORESOURCE_MEM; 1354 /* 1355 * bracing ((signed) ...) are required for 64bit kernel because 1356 * we only want to sign extend the lower 16 bits of the register. 1357 * The upper 16-bits of range registers are hardcoded to 0xffff. 1358 */ 1359 res->start = (unsigned long)((signed) READ_U32(ioaddr) << 16); 1360 res->end = (unsigned long)((signed) (READ_U32(ioaddr + 4) << 16) - 1); 1361 res->name = name; 1362 /* 1363 * Check if this MMIO range is disable 1364 */ 1365 if (res->end + 1 == res->start) 1366 return; 1367 1368 /* On some platforms (e.g. K-Class), we have already registered 1369 * resources for devices reported by firmware. Some are children 1370 * of ccio. 1371 * "insert" ccio ranges in the mmio hierarchy (/proc/iomem). 1372 */ 1373 result = insert_resource(&iomem_resource, res); 1374 if (result < 0) { 1375 printk(KERN_ERR "%s() failed to claim CCIO bus address space (%08lx,%08lx)\n", 1376 __func__, (unsigned long)res->start, (unsigned long)res->end); 1377 } 1378 } 1379 1380 static void __init ccio_init_resources(struct ioc *ioc) 1381 { 1382 struct resource *res = ioc->mmio_region; 1383 char *name = kmalloc(14, GFP_KERNEL); 1384 1385 snprintf(name, 14, "GSC Bus [%d/]", ioc->hw_path); 1386 1387 ccio_init_resource(res, name, &ioc->ioc_regs->io_io_low); 1388 ccio_init_resource(res + 1, name, &ioc->ioc_regs->io_io_low_hv); 1389 } 1390 1391 static int new_ioc_area(struct resource *res, unsigned long size, 1392 unsigned long min, unsigned long max, unsigned long align) 1393 { 1394 if (max <= min) 1395 return -EBUSY; 1396 1397 res->start = (max - size + 1) &~ (align - 1); 1398 res->end = res->start + size; 1399 1400 /* We might be trying to expand the MMIO range to include 1401 * a child device that has already registered it's MMIO space. 1402 * Use "insert" instead of request_resource(). 1403 */ 1404 if (!insert_resource(&iomem_resource, res)) 1405 return 0; 1406 1407 return new_ioc_area(res, size, min, max - size, align); 1408 } 1409 1410 static int expand_ioc_area(struct resource *res, unsigned long size, 1411 unsigned long min, unsigned long max, unsigned long align) 1412 { 1413 unsigned long start, len; 1414 1415 if (!res->parent) 1416 return new_ioc_area(res, size, min, max, align); 1417 1418 start = (res->start - size) &~ (align - 1); 1419 len = res->end - start + 1; 1420 if (start >= min) { 1421 if (!adjust_resource(res, start, len)) 1422 return 0; 1423 } 1424 1425 start = res->start; 1426 len = ((size + res->end + align) &~ (align - 1)) - start; 1427 if (start + len <= max) { 1428 if (!adjust_resource(res, start, len)) 1429 return 0; 1430 } 1431 1432 return -EBUSY; 1433 } 1434 1435 /* 1436 * Dino calls this function. Beware that we may get called on systems 1437 * which have no IOC (725, B180, C160L, etc) but do have a Dino. 1438 * So it's legal to find no parent IOC. 1439 * 1440 * Some other issues: one of the resources in the ioc may be unassigned. 1441 */ 1442 int ccio_allocate_resource(const struct parisc_device *dev, 1443 struct resource *res, unsigned long size, 1444 unsigned long min, unsigned long max, unsigned long align) 1445 { 1446 struct resource *parent = &iomem_resource; 1447 struct ioc *ioc = ccio_get_iommu(dev); 1448 if (!ioc) 1449 goto out; 1450 1451 parent = ioc->mmio_region; 1452 if (parent->parent && 1453 !allocate_resource(parent, res, size, min, max, align, NULL, NULL)) 1454 return 0; 1455 1456 if ((parent + 1)->parent && 1457 !allocate_resource(parent + 1, res, size, min, max, align, 1458 NULL, NULL)) 1459 return 0; 1460 1461 if (!expand_ioc_area(parent, size, min, max, align)) { 1462 __raw_writel(((parent->start)>>16) | 0xffff0000, 1463 &ioc->ioc_regs->io_io_low); 1464 __raw_writel(((parent->end)>>16) | 0xffff0000, 1465 &ioc->ioc_regs->io_io_high); 1466 } else if (!expand_ioc_area(parent + 1, size, min, max, align)) { 1467 parent++; 1468 __raw_writel(((parent->start)>>16) | 0xffff0000, 1469 &ioc->ioc_regs->io_io_low_hv); 1470 __raw_writel(((parent->end)>>16) | 0xffff0000, 1471 &ioc->ioc_regs->io_io_high_hv); 1472 } else { 1473 return -EBUSY; 1474 } 1475 1476 out: 1477 return allocate_resource(parent, res, size, min, max, align, NULL,NULL); 1478 } 1479 1480 int ccio_request_resource(const struct parisc_device *dev, 1481 struct resource *res) 1482 { 1483 struct resource *parent; 1484 struct ioc *ioc = ccio_get_iommu(dev); 1485 1486 if (!ioc) { 1487 parent = &iomem_resource; 1488 } else if ((ioc->mmio_region->start <= res->start) && 1489 (res->end <= ioc->mmio_region->end)) { 1490 parent = ioc->mmio_region; 1491 } else if (((ioc->mmio_region + 1)->start <= res->start) && 1492 (res->end <= (ioc->mmio_region + 1)->end)) { 1493 parent = ioc->mmio_region + 1; 1494 } else { 1495 return -EBUSY; 1496 } 1497 1498 /* "transparent" bus bridges need to register MMIO resources 1499 * firmware assigned them. e.g. children of hppb.c (e.g. K-class) 1500 * registered their resources in the PDC "bus walk" (See 1501 * arch/parisc/kernel/inventory.c). 1502 */ 1503 return insert_resource(parent, res); 1504 } 1505 1506 /** 1507 * ccio_probe - Determine if ccio should claim this device. 1508 * @dev: The device which has been found 1509 * 1510 * Determine if ccio should claim this chip (return 0) or not (return 1). 1511 * If so, initialize the chip and tell other partners in crime they 1512 * have work to do. 1513 */ 1514 static int __init ccio_probe(struct parisc_device *dev) 1515 { 1516 int i; 1517 struct ioc *ioc, **ioc_p = &ioc_list; 1518 struct pci_hba_data *hba; 1519 1520 ioc = kzalloc(sizeof(struct ioc), GFP_KERNEL); 1521 if (ioc == NULL) { 1522 printk(KERN_ERR MODULE_NAME ": memory allocation failure\n"); 1523 return -ENOMEM; 1524 } 1525 1526 ioc->name = dev->id.hversion == U2_IOA_RUNWAY ? "U2" : "UTurn"; 1527 1528 printk(KERN_INFO "Found %s at 0x%lx\n", ioc->name, 1529 (unsigned long)dev->hpa.start); 1530 1531 for (i = 0; i < ioc_count; i++) { 1532 ioc_p = &(*ioc_p)->next; 1533 } 1534 *ioc_p = ioc; 1535 1536 ioc->hw_path = dev->hw_path; 1537 ioc->ioc_regs = ioremap_nocache(dev->hpa.start, 4096); 1538 if (!ioc->ioc_regs) { 1539 kfree(ioc); 1540 return -ENOMEM; 1541 } 1542 ccio_ioc_init(ioc); 1543 ccio_init_resources(ioc); 1544 hppa_dma_ops = &ccio_ops; 1545 1546 hba = kzalloc(sizeof(*hba), GFP_KERNEL); 1547 /* if this fails, no I/O cards will work, so may as well bug */ 1548 BUG_ON(hba == NULL); 1549 1550 hba->iommu = ioc; 1551 dev->dev.platform_data = hba; 1552 1553 #ifdef CONFIG_PROC_FS 1554 if (ioc_count == 0) { 1555 proc_create_single(MODULE_NAME, 0, proc_runway_root, 1556 ccio_proc_info); 1557 proc_create_single(MODULE_NAME"-bitmap", 0, proc_runway_root, 1558 ccio_proc_bitmap_info); 1559 } 1560 #endif 1561 ioc_count++; 1562 return 0; 1563 } 1564 1565 /** 1566 * ccio_init - ccio initialization procedure. 1567 * 1568 * Register this driver. 1569 */ 1570 void __init ccio_init(void) 1571 { 1572 register_parisc_driver(&ccio_driver); 1573 } 1574 1575