Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | ** IA64 System Bus Adapter (SBA) I/O MMU manager |
| 3 | ** |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 4 | ** (c) Copyright 2002-2005 Alex Williamson |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5 | ** (c) Copyright 2002-2003 Grant Grundler |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 6 | ** (c) Copyright 2002-2005 Hewlett-Packard Company |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 7 | ** |
| 8 | ** Portions (c) 2000 Grant Grundler (from parisc I/O MMU code) |
| 9 | ** Portions (c) 1999 Dave S. Miller (from sparc64 I/O MMU code) |
| 10 | ** |
| 11 | ** This program is free software; you can redistribute it and/or modify |
| 12 | ** it under the terms of the GNU General Public License as published by |
| 13 | ** the Free Software Foundation; either version 2 of the License, or |
| 14 | ** (at your option) any later version. |
| 15 | ** |
| 16 | ** |
| 17 | ** This module initializes the IOC (I/O Controller) found on HP |
| 18 | ** McKinley machines and their successors. |
| 19 | ** |
| 20 | */ |
| 21 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 22 | #include <linux/types.h> |
| 23 | #include <linux/kernel.h> |
| 24 | #include <linux/module.h> |
| 25 | #include <linux/spinlock.h> |
| 26 | #include <linux/slab.h> |
| 27 | #include <linux/init.h> |
| 28 | #include <linux/mm.h> |
| 29 | #include <linux/string.h> |
| 30 | #include <linux/pci.h> |
| 31 | #include <linux/proc_fs.h> |
| 32 | #include <linux/seq_file.h> |
| 33 | #include <linux/acpi.h> |
| 34 | #include <linux/efi.h> |
| 35 | #include <linux/nodemask.h> |
| 36 | #include <linux/bitops.h> /* hweight64() */ |
| 37 | |
| 38 | #include <asm/delay.h> /* ia64_get_itc() */ |
| 39 | #include <asm/io.h> |
| 40 | #include <asm/page.h> /* PAGE_OFFSET */ |
| 41 | #include <asm/dma.h> |
| 42 | #include <asm/system.h> /* wmb() */ |
| 43 | |
| 44 | #include <asm/acpi-ext.h> |
| 45 | |
| 46 | #define PFX "IOC: " |
| 47 | |
| 48 | /* |
| 49 | ** Enabling timing search of the pdir resource map. Output in /proc. |
| 50 | ** Disabled by default to optimize performance. |
| 51 | */ |
| 52 | #undef PDIR_SEARCH_TIMING |
| 53 | |
| 54 | /* |
| 55 | ** This option allows cards capable of 64bit DMA to bypass the IOMMU. If |
| 56 | ** not defined, all DMA will be 32bit and go through the TLB. |
| 57 | ** There's potentially a conflict in the bio merge code with us |
| 58 | ** advertising an iommu, but then bypassing it. Since I/O MMU bypassing |
| 59 | ** appears to give more performance than bio-level virtual merging, we'll |
| 60 | ** do the former for now. NOTE: BYPASS_SG also needs to be undef'd to |
| 61 | ** completely restrict DMA to the IOMMU. |
| 62 | */ |
| 63 | #define ALLOW_IOV_BYPASS |
| 64 | |
| 65 | /* |
| 66 | ** This option specifically allows/disallows bypassing scatterlists with |
| 67 | ** multiple entries. Coalescing these entries can allow better DMA streaming |
| 68 | ** and in some cases shows better performance than entirely bypassing the |
| 69 | ** IOMMU. Performance increase on the order of 1-2% sequential output/input |
| 70 | ** using bonnie++ on a RAID0 MD device (sym2 & mpt). |
| 71 | */ |
| 72 | #undef ALLOW_IOV_BYPASS_SG |
| 73 | |
| 74 | /* |
| 75 | ** If a device prefetches beyond the end of a valid pdir entry, it will cause |
| 76 | ** a hard failure, ie. MCA. Version 3.0 and later of the zx1 LBA should |
| 77 | ** disconnect on 4k boundaries and prevent such issues. If the device is |
Matt LaPlante | 0779bf2 | 2006-11-30 05:24:39 +0100 | [diff] [blame^] | 78 | ** particularly aggressive, this option will keep the entire pdir valid such |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 79 | ** that prefetching will hit a valid address. This could severely impact |
| 80 | ** error containment, and is therefore off by default. The page that is |
| 81 | ** used for spill-over is poisoned, so that should help debugging somewhat. |
| 82 | */ |
| 83 | #undef FULL_VALID_PDIR |
| 84 | |
| 85 | #define ENABLE_MARK_CLEAN |
| 86 | |
| 87 | /* |
| 88 | ** The number of debug flags is a clue - this code is fragile. NOTE: since |
| 89 | ** tightening the use of res_lock the resource bitmap and actual pdir are no |
| 90 | ** longer guaranteed to stay in sync. The sanity checking code isn't going to |
| 91 | ** like that. |
| 92 | */ |
| 93 | #undef DEBUG_SBA_INIT |
| 94 | #undef DEBUG_SBA_RUN |
| 95 | #undef DEBUG_SBA_RUN_SG |
| 96 | #undef DEBUG_SBA_RESOURCE |
| 97 | #undef ASSERT_PDIR_SANITY |
| 98 | #undef DEBUG_LARGE_SG_ENTRIES |
| 99 | #undef DEBUG_BYPASS |
| 100 | |
| 101 | #if defined(FULL_VALID_PDIR) && defined(ASSERT_PDIR_SANITY) |
| 102 | #error FULL_VALID_PDIR and ASSERT_PDIR_SANITY are mutually exclusive |
| 103 | #endif |
| 104 | |
| 105 | #define SBA_INLINE __inline__ |
| 106 | /* #define SBA_INLINE */ |
| 107 | |
| 108 | #ifdef DEBUG_SBA_INIT |
| 109 | #define DBG_INIT(x...) printk(x) |
| 110 | #else |
| 111 | #define DBG_INIT(x...) |
| 112 | #endif |
| 113 | |
| 114 | #ifdef DEBUG_SBA_RUN |
| 115 | #define DBG_RUN(x...) printk(x) |
| 116 | #else |
| 117 | #define DBG_RUN(x...) |
| 118 | #endif |
| 119 | |
| 120 | #ifdef DEBUG_SBA_RUN_SG |
| 121 | #define DBG_RUN_SG(x...) printk(x) |
| 122 | #else |
| 123 | #define DBG_RUN_SG(x...) |
| 124 | #endif |
| 125 | |
| 126 | |
| 127 | #ifdef DEBUG_SBA_RESOURCE |
| 128 | #define DBG_RES(x...) printk(x) |
| 129 | #else |
| 130 | #define DBG_RES(x...) |
| 131 | #endif |
| 132 | |
| 133 | #ifdef DEBUG_BYPASS |
| 134 | #define DBG_BYPASS(x...) printk(x) |
| 135 | #else |
| 136 | #define DBG_BYPASS(x...) |
| 137 | #endif |
| 138 | |
| 139 | #ifdef ASSERT_PDIR_SANITY |
| 140 | #define ASSERT(expr) \ |
| 141 | if(!(expr)) { \ |
| 142 | printk( "\n" __FILE__ ":%d: Assertion " #expr " failed!\n",__LINE__); \ |
| 143 | panic(#expr); \ |
| 144 | } |
| 145 | #else |
| 146 | #define ASSERT(expr) |
| 147 | #endif |
| 148 | |
| 149 | /* |
| 150 | ** The number of pdir entries to "free" before issuing |
| 151 | ** a read to PCOM register to flush out PCOM writes. |
| 152 | ** Interacts with allocation granularity (ie 4 or 8 entries |
| 153 | ** allocated and free'd/purged at a time might make this |
| 154 | ** less interesting). |
| 155 | */ |
| 156 | #define DELAYED_RESOURCE_CNT 64 |
| 157 | |
Bjorn Helgaas | e15da40 | 2005-05-03 12:07:00 -0700 | [diff] [blame] | 158 | #define PCI_DEVICE_ID_HP_SX2000_IOC 0x12ec |
| 159 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 160 | #define ZX1_IOC_ID ((PCI_DEVICE_ID_HP_ZX1_IOC << 16) | PCI_VENDOR_ID_HP) |
| 161 | #define ZX2_IOC_ID ((PCI_DEVICE_ID_HP_ZX2_IOC << 16) | PCI_VENDOR_ID_HP) |
| 162 | #define REO_IOC_ID ((PCI_DEVICE_ID_HP_REO_IOC << 16) | PCI_VENDOR_ID_HP) |
| 163 | #define SX1000_IOC_ID ((PCI_DEVICE_ID_HP_SX1000_IOC << 16) | PCI_VENDOR_ID_HP) |
Bjorn Helgaas | e15da40 | 2005-05-03 12:07:00 -0700 | [diff] [blame] | 164 | #define SX2000_IOC_ID ((PCI_DEVICE_ID_HP_SX2000_IOC << 16) | PCI_VENDOR_ID_HP) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 165 | |
| 166 | #define ZX1_IOC_OFFSET 0x1000 /* ACPI reports SBA, we want IOC */ |
| 167 | |
| 168 | #define IOC_FUNC_ID 0x000 |
| 169 | #define IOC_FCLASS 0x008 /* function class, bist, header, rev... */ |
| 170 | #define IOC_IBASE 0x300 /* IO TLB */ |
| 171 | #define IOC_IMASK 0x308 |
| 172 | #define IOC_PCOM 0x310 |
| 173 | #define IOC_TCNFG 0x318 |
| 174 | #define IOC_PDIR_BASE 0x320 |
| 175 | |
| 176 | #define IOC_ROPE0_CFG 0x500 |
| 177 | #define IOC_ROPE_AO 0x10 /* Allow "Relaxed Ordering" */ |
| 178 | |
| 179 | |
| 180 | /* AGP GART driver looks for this */ |
| 181 | #define ZX1_SBA_IOMMU_COOKIE 0x0000badbadc0ffeeUL |
| 182 | |
| 183 | /* |
| 184 | ** The zx1 IOC supports 4/8/16/64KB page sizes (see TCNFG register) |
| 185 | ** |
| 186 | ** Some IOCs (sx1000) can run at the above pages sizes, but are |
| 187 | ** really only supported using the IOC at a 4k page size. |
| 188 | ** |
| 189 | ** iovp_size could only be greater than PAGE_SIZE if we are |
| 190 | ** confident the drivers really only touch the next physical |
| 191 | ** page iff that driver instance owns it. |
| 192 | */ |
| 193 | static unsigned long iovp_size; |
| 194 | static unsigned long iovp_shift; |
| 195 | static unsigned long iovp_mask; |
| 196 | |
| 197 | struct ioc { |
| 198 | void __iomem *ioc_hpa; /* I/O MMU base address */ |
| 199 | char *res_map; /* resource map, bit == pdir entry */ |
| 200 | u64 *pdir_base; /* physical base address */ |
| 201 | unsigned long ibase; /* pdir IOV Space base */ |
| 202 | unsigned long imask; /* pdir IOV Space mask */ |
| 203 | |
| 204 | unsigned long *res_hint; /* next avail IOVP - circular search */ |
| 205 | unsigned long dma_mask; |
| 206 | spinlock_t res_lock; /* protects the resource bitmap, but must be held when */ |
| 207 | /* clearing pdir to prevent races with allocations. */ |
| 208 | unsigned int res_bitshift; /* from the RIGHT! */ |
| 209 | unsigned int res_size; /* size of resource map in bytes */ |
| 210 | #ifdef CONFIG_NUMA |
| 211 | unsigned int node; /* node where this IOC lives */ |
| 212 | #endif |
| 213 | #if DELAYED_RESOURCE_CNT > 0 |
| 214 | spinlock_t saved_lock; /* may want to try to get this on a separate cacheline */ |
| 215 | /* than res_lock for bigger systems. */ |
| 216 | int saved_cnt; |
| 217 | struct sba_dma_pair { |
| 218 | dma_addr_t iova; |
| 219 | size_t size; |
| 220 | } saved[DELAYED_RESOURCE_CNT]; |
| 221 | #endif |
| 222 | |
| 223 | #ifdef PDIR_SEARCH_TIMING |
| 224 | #define SBA_SEARCH_SAMPLE 0x100 |
| 225 | unsigned long avg_search[SBA_SEARCH_SAMPLE]; |
| 226 | unsigned long avg_idx; /* current index into avg_search */ |
| 227 | #endif |
| 228 | |
| 229 | /* Stuff we don't need in performance path */ |
| 230 | struct ioc *next; /* list of IOC's in system */ |
| 231 | acpi_handle handle; /* for multiple IOC's */ |
| 232 | const char *name; |
| 233 | unsigned int func_id; |
| 234 | unsigned int rev; /* HW revision of chip */ |
| 235 | u32 iov_size; |
| 236 | unsigned int pdir_size; /* in bytes, determined by IOV Space size */ |
| 237 | struct pci_dev *sac_only_dev; |
| 238 | }; |
| 239 | |
| 240 | static struct ioc *ioc_list; |
| 241 | static int reserve_sba_gart = 1; |
| 242 | |
| 243 | static SBA_INLINE void sba_mark_invalid(struct ioc *, dma_addr_t, size_t); |
| 244 | static SBA_INLINE void sba_free_range(struct ioc *, dma_addr_t, size_t); |
| 245 | |
| 246 | #define sba_sg_address(sg) (page_address((sg)->page) + (sg)->offset) |
| 247 | |
| 248 | #ifdef FULL_VALID_PDIR |
| 249 | static u64 prefetch_spill_page; |
| 250 | #endif |
| 251 | |
| 252 | #ifdef CONFIG_PCI |
| 253 | # define GET_IOC(dev) (((dev)->bus == &pci_bus_type) \ |
| 254 | ? ((struct ioc *) PCI_CONTROLLER(to_pci_dev(dev))->iommu) : NULL) |
| 255 | #else |
| 256 | # define GET_IOC(dev) NULL |
| 257 | #endif |
| 258 | |
| 259 | /* |
| 260 | ** DMA_CHUNK_SIZE is used by the SCSI mid-layer to break up |
Matt LaPlante | 0779bf2 | 2006-11-30 05:24:39 +0100 | [diff] [blame^] | 261 | ** (or rather not merge) DMAs into manageable chunks. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 262 | ** On parisc, this is more of the software/tuning constraint |
Matt LaPlante | 0779bf2 | 2006-11-30 05:24:39 +0100 | [diff] [blame^] | 263 | ** rather than the HW. I/O MMU allocation algorithms can be |
| 264 | ** faster with smaller sizes (to some degree). |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 265 | */ |
| 266 | #define DMA_CHUNK_SIZE (BITS_PER_LONG*iovp_size) |
| 267 | |
| 268 | #define ROUNDUP(x,y) ((x + ((y)-1)) & ~((y)-1)) |
| 269 | |
| 270 | /************************************ |
| 271 | ** SBA register read and write support |
| 272 | ** |
| 273 | ** BE WARNED: register writes are posted. |
| 274 | ** (ie follow writes which must reach HW with a read) |
| 275 | ** |
| 276 | */ |
| 277 | #define READ_REG(addr) __raw_readq(addr) |
| 278 | #define WRITE_REG(val, addr) __raw_writeq(val, addr) |
| 279 | |
| 280 | #ifdef DEBUG_SBA_INIT |
| 281 | |
| 282 | /** |
| 283 | * sba_dump_tlb - debugging only - print IOMMU operating parameters |
| 284 | * @hpa: base address of the IOMMU |
| 285 | * |
| 286 | * Print the size/location of the IO MMU PDIR. |
| 287 | */ |
| 288 | static void |
| 289 | sba_dump_tlb(char *hpa) |
| 290 | { |
| 291 | DBG_INIT("IO TLB at 0x%p\n", (void *)hpa); |
| 292 | DBG_INIT("IOC_IBASE : %016lx\n", READ_REG(hpa+IOC_IBASE)); |
| 293 | DBG_INIT("IOC_IMASK : %016lx\n", READ_REG(hpa+IOC_IMASK)); |
| 294 | DBG_INIT("IOC_TCNFG : %016lx\n", READ_REG(hpa+IOC_TCNFG)); |
| 295 | DBG_INIT("IOC_PDIR_BASE: %016lx\n", READ_REG(hpa+IOC_PDIR_BASE)); |
| 296 | DBG_INIT("\n"); |
| 297 | } |
| 298 | #endif |
| 299 | |
| 300 | |
| 301 | #ifdef ASSERT_PDIR_SANITY |
| 302 | |
| 303 | /** |
| 304 | * sba_dump_pdir_entry - debugging only - print one IOMMU PDIR entry |
| 305 | * @ioc: IO MMU structure which owns the pdir we are interested in. |
| 306 | * @msg: text to print ont the output line. |
| 307 | * @pide: pdir index. |
| 308 | * |
| 309 | * Print one entry of the IO MMU PDIR in human readable form. |
| 310 | */ |
| 311 | static void |
| 312 | sba_dump_pdir_entry(struct ioc *ioc, char *msg, uint pide) |
| 313 | { |
| 314 | /* start printing from lowest pde in rval */ |
| 315 | u64 *ptr = &ioc->pdir_base[pide & ~(BITS_PER_LONG - 1)]; |
| 316 | unsigned long *rptr = (unsigned long *) &ioc->res_map[(pide >>3) & -sizeof(unsigned long)]; |
| 317 | uint rcnt; |
| 318 | |
| 319 | printk(KERN_DEBUG "SBA: %s rp %p bit %d rval 0x%lx\n", |
| 320 | msg, rptr, pide & (BITS_PER_LONG - 1), *rptr); |
| 321 | |
| 322 | rcnt = 0; |
| 323 | while (rcnt < BITS_PER_LONG) { |
| 324 | printk(KERN_DEBUG "%s %2d %p %016Lx\n", |
| 325 | (rcnt == (pide & (BITS_PER_LONG - 1))) |
| 326 | ? " -->" : " ", |
| 327 | rcnt, ptr, (unsigned long long) *ptr ); |
| 328 | rcnt++; |
| 329 | ptr++; |
| 330 | } |
| 331 | printk(KERN_DEBUG "%s", msg); |
| 332 | } |
| 333 | |
| 334 | |
| 335 | /** |
| 336 | * sba_check_pdir - debugging only - consistency checker |
| 337 | * @ioc: IO MMU structure which owns the pdir we are interested in. |
| 338 | * @msg: text to print ont the output line. |
| 339 | * |
| 340 | * Verify the resource map and pdir state is consistent |
| 341 | */ |
| 342 | static int |
| 343 | sba_check_pdir(struct ioc *ioc, char *msg) |
| 344 | { |
| 345 | u64 *rptr_end = (u64 *) &(ioc->res_map[ioc->res_size]); |
| 346 | u64 *rptr = (u64 *) ioc->res_map; /* resource map ptr */ |
| 347 | u64 *pptr = ioc->pdir_base; /* pdir ptr */ |
| 348 | uint pide = 0; |
| 349 | |
| 350 | while (rptr < rptr_end) { |
| 351 | u64 rval; |
| 352 | int rcnt; /* number of bits we might check */ |
| 353 | |
| 354 | rval = *rptr; |
| 355 | rcnt = 64; |
| 356 | |
| 357 | while (rcnt) { |
| 358 | /* Get last byte and highest bit from that */ |
| 359 | u32 pde = ((u32)((*pptr >> (63)) & 0x1)); |
| 360 | if ((rval & 0x1) ^ pde) |
| 361 | { |
| 362 | /* |
| 363 | ** BUMMER! -- res_map != pdir -- |
| 364 | ** Dump rval and matching pdir entries |
| 365 | */ |
| 366 | sba_dump_pdir_entry(ioc, msg, pide); |
| 367 | return(1); |
| 368 | } |
| 369 | rcnt--; |
| 370 | rval >>= 1; /* try the next bit */ |
| 371 | pptr++; |
| 372 | pide++; |
| 373 | } |
| 374 | rptr++; /* look at next word of res_map */ |
| 375 | } |
| 376 | /* It'd be nice if we always got here :^) */ |
| 377 | return 0; |
| 378 | } |
| 379 | |
| 380 | |
| 381 | /** |
| 382 | * sba_dump_sg - debugging only - print Scatter-Gather list |
| 383 | * @ioc: IO MMU structure which owns the pdir we are interested in. |
| 384 | * @startsg: head of the SG list |
| 385 | * @nents: number of entries in SG list |
| 386 | * |
| 387 | * print the SG list so we can verify it's correct by hand. |
| 388 | */ |
| 389 | static void |
| 390 | sba_dump_sg( struct ioc *ioc, struct scatterlist *startsg, int nents) |
| 391 | { |
| 392 | while (nents-- > 0) { |
| 393 | printk(KERN_DEBUG " %d : DMA %08lx/%05x CPU %p\n", nents, |
| 394 | startsg->dma_address, startsg->dma_length, |
| 395 | sba_sg_address(startsg)); |
| 396 | startsg++; |
| 397 | } |
| 398 | } |
| 399 | |
| 400 | static void |
| 401 | sba_check_sg( struct ioc *ioc, struct scatterlist *startsg, int nents) |
| 402 | { |
| 403 | struct scatterlist *the_sg = startsg; |
| 404 | int the_nents = nents; |
| 405 | |
| 406 | while (the_nents-- > 0) { |
| 407 | if (sba_sg_address(the_sg) == 0x0UL) |
| 408 | sba_dump_sg(NULL, startsg, nents); |
| 409 | the_sg++; |
| 410 | } |
| 411 | } |
| 412 | |
| 413 | #endif /* ASSERT_PDIR_SANITY */ |
| 414 | |
| 415 | |
| 416 | |
| 417 | |
| 418 | /************************************************************** |
| 419 | * |
| 420 | * I/O Pdir Resource Management |
| 421 | * |
| 422 | * Bits set in the resource map are in use. |
| 423 | * Each bit can represent a number of pages. |
| 424 | * LSbs represent lower addresses (IOVA's). |
| 425 | * |
| 426 | ***************************************************************/ |
| 427 | #define PAGES_PER_RANGE 1 /* could increase this to 4 or 8 if needed */ |
| 428 | |
| 429 | /* Convert from IOVP to IOVA and vice versa. */ |
| 430 | #define SBA_IOVA(ioc,iovp,offset) ((ioc->ibase) | (iovp) | (offset)) |
| 431 | #define SBA_IOVP(ioc,iova) ((iova) & ~(ioc->ibase)) |
| 432 | |
| 433 | #define PDIR_ENTRY_SIZE sizeof(u64) |
| 434 | |
| 435 | #define PDIR_INDEX(iovp) ((iovp)>>iovp_shift) |
| 436 | |
| 437 | #define RESMAP_MASK(n) ~(~0UL << (n)) |
| 438 | #define RESMAP_IDX_MASK (sizeof(unsigned long) - 1) |
| 439 | |
| 440 | |
| 441 | /** |
| 442 | * For most cases the normal get_order is sufficient, however it limits us |
| 443 | * to PAGE_SIZE being the minimum mapping alignment and TC flush granularity. |
| 444 | * It only incurs about 1 clock cycle to use this one with the static variable |
| 445 | * and makes the code more intuitive. |
| 446 | */ |
| 447 | static SBA_INLINE int |
| 448 | get_iovp_order (unsigned long size) |
| 449 | { |
| 450 | long double d = size - 1; |
| 451 | long order; |
| 452 | |
| 453 | order = ia64_getf_exp(d); |
| 454 | order = order - iovp_shift - 0xffff + 1; |
| 455 | if (order < 0) |
| 456 | order = 0; |
| 457 | return order; |
| 458 | } |
| 459 | |
| 460 | /** |
| 461 | * sba_search_bitmap - find free space in IO PDIR resource bitmap |
| 462 | * @ioc: IO MMU structure which owns the pdir we are interested in. |
| 463 | * @bits_wanted: number of entries we need. |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 464 | * @use_hint: use res_hint to indicate where to start looking |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 465 | * |
| 466 | * Find consecutive free bits in resource bitmap. |
| 467 | * Each bit represents one entry in the IO Pdir. |
| 468 | * Cool perf optimization: search for log2(size) bits at a time. |
| 469 | */ |
| 470 | static SBA_INLINE unsigned long |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 471 | sba_search_bitmap(struct ioc *ioc, unsigned long bits_wanted, int use_hint) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 472 | { |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 473 | unsigned long *res_ptr; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 474 | unsigned long *res_end = (unsigned long *) &(ioc->res_map[ioc->res_size]); |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 475 | unsigned long flags, pide = ~0UL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 476 | |
| 477 | ASSERT(((unsigned long) ioc->res_hint & (sizeof(unsigned long) - 1UL)) == 0); |
| 478 | ASSERT(res_ptr < res_end); |
| 479 | |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 480 | spin_lock_irqsave(&ioc->res_lock, flags); |
| 481 | |
| 482 | /* Allow caller to force a search through the entire resource space */ |
| 483 | if (likely(use_hint)) { |
| 484 | res_ptr = ioc->res_hint; |
| 485 | } else { |
| 486 | res_ptr = (ulong *)ioc->res_map; |
| 487 | ioc->res_bitshift = 0; |
| 488 | } |
| 489 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 490 | /* |
| 491 | * N.B. REO/Grande defect AR2305 can cause TLB fetch timeouts |
| 492 | * if a TLB entry is purged while in use. sba_mark_invalid() |
| 493 | * purges IOTLB entries in power-of-two sizes, so we also |
| 494 | * allocate IOVA space in power-of-two sizes. |
| 495 | */ |
| 496 | bits_wanted = 1UL << get_iovp_order(bits_wanted << iovp_shift); |
| 497 | |
| 498 | if (likely(bits_wanted == 1)) { |
| 499 | unsigned int bitshiftcnt; |
| 500 | for(; res_ptr < res_end ; res_ptr++) { |
| 501 | if (likely(*res_ptr != ~0UL)) { |
| 502 | bitshiftcnt = ffz(*res_ptr); |
| 503 | *res_ptr |= (1UL << bitshiftcnt); |
| 504 | pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map); |
| 505 | pide <<= 3; /* convert to bit address */ |
| 506 | pide += bitshiftcnt; |
| 507 | ioc->res_bitshift = bitshiftcnt + bits_wanted; |
| 508 | goto found_it; |
| 509 | } |
| 510 | } |
| 511 | goto not_found; |
| 512 | |
| 513 | } |
| 514 | |
| 515 | if (likely(bits_wanted <= BITS_PER_LONG/2)) { |
| 516 | /* |
| 517 | ** Search the resource bit map on well-aligned values. |
| 518 | ** "o" is the alignment. |
| 519 | ** We need the alignment to invalidate I/O TLB using |
| 520 | ** SBA HW features in the unmap path. |
| 521 | */ |
| 522 | unsigned long o = 1 << get_iovp_order(bits_wanted << iovp_shift); |
| 523 | uint bitshiftcnt = ROUNDUP(ioc->res_bitshift, o); |
| 524 | unsigned long mask, base_mask; |
| 525 | |
| 526 | base_mask = RESMAP_MASK(bits_wanted); |
| 527 | mask = base_mask << bitshiftcnt; |
| 528 | |
| 529 | DBG_RES("%s() o %ld %p", __FUNCTION__, o, res_ptr); |
| 530 | for(; res_ptr < res_end ; res_ptr++) |
| 531 | { |
| 532 | DBG_RES(" %p %lx %lx\n", res_ptr, mask, *res_ptr); |
| 533 | ASSERT(0 != mask); |
| 534 | for (; mask ; mask <<= o, bitshiftcnt += o) { |
| 535 | if(0 == ((*res_ptr) & mask)) { |
| 536 | *res_ptr |= mask; /* mark resources busy! */ |
| 537 | pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map); |
| 538 | pide <<= 3; /* convert to bit address */ |
| 539 | pide += bitshiftcnt; |
| 540 | ioc->res_bitshift = bitshiftcnt + bits_wanted; |
| 541 | goto found_it; |
| 542 | } |
| 543 | } |
| 544 | |
| 545 | bitshiftcnt = 0; |
| 546 | mask = base_mask; |
| 547 | |
| 548 | } |
| 549 | |
| 550 | } else { |
| 551 | int qwords, bits, i; |
| 552 | unsigned long *end; |
| 553 | |
| 554 | qwords = bits_wanted >> 6; /* /64 */ |
| 555 | bits = bits_wanted - (qwords * BITS_PER_LONG); |
| 556 | |
| 557 | end = res_end - qwords; |
| 558 | |
| 559 | for (; res_ptr < end; res_ptr++) { |
| 560 | for (i = 0 ; i < qwords ; i++) { |
| 561 | if (res_ptr[i] != 0) |
| 562 | goto next_ptr; |
| 563 | } |
| 564 | if (bits && res_ptr[i] && (__ffs(res_ptr[i]) < bits)) |
| 565 | continue; |
| 566 | |
| 567 | /* Found it, mark it */ |
| 568 | for (i = 0 ; i < qwords ; i++) |
| 569 | res_ptr[i] = ~0UL; |
| 570 | res_ptr[i] |= RESMAP_MASK(bits); |
| 571 | |
| 572 | pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map); |
| 573 | pide <<= 3; /* convert to bit address */ |
| 574 | res_ptr += qwords; |
| 575 | ioc->res_bitshift = bits; |
| 576 | goto found_it; |
| 577 | next_ptr: |
| 578 | ; |
| 579 | } |
| 580 | } |
| 581 | |
| 582 | not_found: |
| 583 | prefetch(ioc->res_map); |
| 584 | ioc->res_hint = (unsigned long *) ioc->res_map; |
| 585 | ioc->res_bitshift = 0; |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 586 | spin_unlock_irqrestore(&ioc->res_lock, flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 587 | return (pide); |
| 588 | |
| 589 | found_it: |
| 590 | ioc->res_hint = res_ptr; |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 591 | spin_unlock_irqrestore(&ioc->res_lock, flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 592 | return (pide); |
| 593 | } |
| 594 | |
| 595 | |
| 596 | /** |
| 597 | * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap |
| 598 | * @ioc: IO MMU structure which owns the pdir we are interested in. |
| 599 | * @size: number of bytes to create a mapping for |
| 600 | * |
| 601 | * Given a size, find consecutive unmarked and then mark those bits in the |
| 602 | * resource bit map. |
| 603 | */ |
| 604 | static int |
| 605 | sba_alloc_range(struct ioc *ioc, size_t size) |
| 606 | { |
| 607 | unsigned int pages_needed = size >> iovp_shift; |
| 608 | #ifdef PDIR_SEARCH_TIMING |
| 609 | unsigned long itc_start; |
| 610 | #endif |
| 611 | unsigned long pide; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 612 | |
| 613 | ASSERT(pages_needed); |
| 614 | ASSERT(0 == (size & ~iovp_mask)); |
| 615 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 616 | #ifdef PDIR_SEARCH_TIMING |
| 617 | itc_start = ia64_get_itc(); |
| 618 | #endif |
| 619 | /* |
| 620 | ** "seek and ye shall find"...praying never hurts either... |
| 621 | */ |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 622 | pide = sba_search_bitmap(ioc, pages_needed, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 623 | if (unlikely(pide >= (ioc->res_size << 3))) { |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 624 | pide = sba_search_bitmap(ioc, pages_needed, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 625 | if (unlikely(pide >= (ioc->res_size << 3))) { |
| 626 | #if DELAYED_RESOURCE_CNT > 0 |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 627 | unsigned long flags; |
| 628 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 629 | /* |
| 630 | ** With delayed resource freeing, we can give this one more shot. We're |
| 631 | ** getting close to being in trouble here, so do what we can to make this |
| 632 | ** one count. |
| 633 | */ |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 634 | spin_lock_irqsave(&ioc->saved_lock, flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 635 | if (ioc->saved_cnt > 0) { |
| 636 | struct sba_dma_pair *d; |
| 637 | int cnt = ioc->saved_cnt; |
| 638 | |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 639 | d = &(ioc->saved[ioc->saved_cnt - 1]); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 640 | |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 641 | spin_lock(&ioc->res_lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 642 | while (cnt--) { |
| 643 | sba_mark_invalid(ioc, d->iova, d->size); |
| 644 | sba_free_range(ioc, d->iova, d->size); |
| 645 | d--; |
| 646 | } |
| 647 | ioc->saved_cnt = 0; |
| 648 | READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */ |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 649 | spin_unlock(&ioc->res_lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 650 | } |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 651 | spin_unlock_irqrestore(&ioc->saved_lock, flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 652 | |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 653 | pide = sba_search_bitmap(ioc, pages_needed, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 654 | if (unlikely(pide >= (ioc->res_size << 3))) |
| 655 | panic(__FILE__ ": I/O MMU @ %p is out of mapping resources\n", |
| 656 | ioc->ioc_hpa); |
| 657 | #else |
| 658 | panic(__FILE__ ": I/O MMU @ %p is out of mapping resources\n", |
| 659 | ioc->ioc_hpa); |
| 660 | #endif |
| 661 | } |
| 662 | } |
| 663 | |
| 664 | #ifdef PDIR_SEARCH_TIMING |
| 665 | ioc->avg_search[ioc->avg_idx++] = (ia64_get_itc() - itc_start) / pages_needed; |
| 666 | ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1; |
| 667 | #endif |
| 668 | |
| 669 | prefetchw(&(ioc->pdir_base[pide])); |
| 670 | |
| 671 | #ifdef ASSERT_PDIR_SANITY |
| 672 | /* verify the first enable bit is clear */ |
| 673 | if(0x00 != ((u8 *) ioc->pdir_base)[pide*PDIR_ENTRY_SIZE + 7]) { |
| 674 | sba_dump_pdir_entry(ioc, "sba_search_bitmap() botched it?", pide); |
| 675 | } |
| 676 | #endif |
| 677 | |
| 678 | DBG_RES("%s(%x) %d -> %lx hint %x/%x\n", |
| 679 | __FUNCTION__, size, pages_needed, pide, |
| 680 | (uint) ((unsigned long) ioc->res_hint - (unsigned long) ioc->res_map), |
| 681 | ioc->res_bitshift ); |
| 682 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 683 | return (pide); |
| 684 | } |
| 685 | |
| 686 | |
| 687 | /** |
| 688 | * sba_free_range - unmark bits in IO PDIR resource bitmap |
| 689 | * @ioc: IO MMU structure which owns the pdir we are interested in. |
| 690 | * @iova: IO virtual address which was previously allocated. |
| 691 | * @size: number of bytes to create a mapping for |
| 692 | * |
| 693 | * clear bits in the ioc's resource map |
| 694 | */ |
| 695 | static SBA_INLINE void |
| 696 | sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size) |
| 697 | { |
| 698 | unsigned long iovp = SBA_IOVP(ioc, iova); |
| 699 | unsigned int pide = PDIR_INDEX(iovp); |
| 700 | unsigned int ridx = pide >> 3; /* convert bit to byte address */ |
| 701 | unsigned long *res_ptr = (unsigned long *) &((ioc)->res_map[ridx & ~RESMAP_IDX_MASK]); |
| 702 | int bits_not_wanted = size >> iovp_shift; |
| 703 | unsigned long m; |
| 704 | |
| 705 | /* Round up to power-of-two size: see AR2305 note above */ |
| 706 | bits_not_wanted = 1UL << get_iovp_order(bits_not_wanted << iovp_shift); |
| 707 | for (; bits_not_wanted > 0 ; res_ptr++) { |
| 708 | |
| 709 | if (unlikely(bits_not_wanted > BITS_PER_LONG)) { |
| 710 | |
| 711 | /* these mappings start 64bit aligned */ |
| 712 | *res_ptr = 0UL; |
| 713 | bits_not_wanted -= BITS_PER_LONG; |
| 714 | pide += BITS_PER_LONG; |
| 715 | |
| 716 | } else { |
| 717 | |
| 718 | /* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */ |
| 719 | m = RESMAP_MASK(bits_not_wanted) << (pide & (BITS_PER_LONG - 1)); |
| 720 | bits_not_wanted = 0; |
| 721 | |
| 722 | DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n", __FUNCTION__, (uint) iova, size, |
| 723 | bits_not_wanted, m, pide, res_ptr, *res_ptr); |
| 724 | |
| 725 | ASSERT(m != 0); |
| 726 | ASSERT(bits_not_wanted); |
| 727 | ASSERT((*res_ptr & m) == m); /* verify same bits are set */ |
| 728 | *res_ptr &= ~m; |
| 729 | } |
| 730 | } |
| 731 | } |
| 732 | |
| 733 | |
| 734 | /************************************************************** |
| 735 | * |
| 736 | * "Dynamic DMA Mapping" support (aka "Coherent I/O") |
| 737 | * |
| 738 | ***************************************************************/ |
| 739 | |
| 740 | /** |
| 741 | * sba_io_pdir_entry - fill in one IO PDIR entry |
| 742 | * @pdir_ptr: pointer to IO PDIR entry |
| 743 | * @vba: Virtual CPU address of buffer to map |
| 744 | * |
| 745 | * SBA Mapping Routine |
| 746 | * |
| 747 | * Given a virtual address (vba, arg1) sba_io_pdir_entry() |
| 748 | * loads the I/O PDIR entry pointed to by pdir_ptr (arg0). |
| 749 | * Each IO Pdir entry consists of 8 bytes as shown below |
| 750 | * (LSB == bit 0): |
| 751 | * |
| 752 | * 63 40 11 7 0 |
| 753 | * +-+---------------------+----------------------------------+----+--------+ |
| 754 | * |V| U | PPN[39:12] | U | FF | |
| 755 | * +-+---------------------+----------------------------------+----+--------+ |
| 756 | * |
| 757 | * V == Valid Bit |
| 758 | * U == Unused |
| 759 | * PPN == Physical Page Number |
| 760 | * |
| 761 | * The physical address fields are filled with the results of virt_to_phys() |
| 762 | * on the vba. |
| 763 | */ |
| 764 | |
| 765 | #if 1 |
| 766 | #define sba_io_pdir_entry(pdir_ptr, vba) *pdir_ptr = ((vba & ~0xE000000000000FFFULL) \ |
| 767 | | 0x8000000000000000ULL) |
| 768 | #else |
| 769 | void SBA_INLINE |
| 770 | sba_io_pdir_entry(u64 *pdir_ptr, unsigned long vba) |
| 771 | { |
| 772 | *pdir_ptr = ((vba & ~0xE000000000000FFFULL) | 0x80000000000000FFULL); |
| 773 | } |
| 774 | #endif |
| 775 | |
| 776 | #ifdef ENABLE_MARK_CLEAN |
| 777 | /** |
| 778 | * Since DMA is i-cache coherent, any (complete) pages that were written via |
| 779 | * DMA can be marked as "clean" so that lazy_mmu_prot_update() doesn't have to |
| 780 | * flush them when they get mapped into an executable vm-area. |
| 781 | */ |
| 782 | static void |
| 783 | mark_clean (void *addr, size_t size) |
| 784 | { |
| 785 | unsigned long pg_addr, end; |
| 786 | |
| 787 | pg_addr = PAGE_ALIGN((unsigned long) addr); |
| 788 | end = (unsigned long) addr + size; |
| 789 | while (pg_addr + PAGE_SIZE <= end) { |
| 790 | struct page *page = virt_to_page((void *)pg_addr); |
| 791 | set_bit(PG_arch_1, &page->flags); |
| 792 | pg_addr += PAGE_SIZE; |
| 793 | } |
| 794 | } |
| 795 | #endif |
| 796 | |
| 797 | /** |
| 798 | * sba_mark_invalid - invalidate one or more IO PDIR entries |
| 799 | * @ioc: IO MMU structure which owns the pdir we are interested in. |
| 800 | * @iova: IO Virtual Address mapped earlier |
| 801 | * @byte_cnt: number of bytes this mapping covers. |
| 802 | * |
| 803 | * Marking the IO PDIR entry(ies) as Invalid and invalidate |
| 804 | * corresponding IO TLB entry. The PCOM (Purge Command Register) |
| 805 | * is to purge stale entries in the IO TLB when unmapping entries. |
| 806 | * |
| 807 | * The PCOM register supports purging of multiple pages, with a minium |
| 808 | * of 1 page and a maximum of 2GB. Hardware requires the address be |
| 809 | * aligned to the size of the range being purged. The size of the range |
| 810 | * must be a power of 2. The "Cool perf optimization" in the |
| 811 | * allocation routine helps keep that true. |
| 812 | */ |
| 813 | static SBA_INLINE void |
| 814 | sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt) |
| 815 | { |
| 816 | u32 iovp = (u32) SBA_IOVP(ioc,iova); |
| 817 | |
| 818 | int off = PDIR_INDEX(iovp); |
| 819 | |
| 820 | /* Must be non-zero and rounded up */ |
| 821 | ASSERT(byte_cnt > 0); |
| 822 | ASSERT(0 == (byte_cnt & ~iovp_mask)); |
| 823 | |
| 824 | #ifdef ASSERT_PDIR_SANITY |
| 825 | /* Assert first pdir entry is set */ |
| 826 | if (!(ioc->pdir_base[off] >> 60)) { |
| 827 | sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp)); |
| 828 | } |
| 829 | #endif |
| 830 | |
| 831 | if (byte_cnt <= iovp_size) |
| 832 | { |
| 833 | ASSERT(off < ioc->pdir_size); |
| 834 | |
| 835 | iovp |= iovp_shift; /* set "size" field for PCOM */ |
| 836 | |
| 837 | #ifndef FULL_VALID_PDIR |
| 838 | /* |
| 839 | ** clear I/O PDIR entry "valid" bit |
| 840 | ** Do NOT clear the rest - save it for debugging. |
| 841 | ** We should only clear bits that have previously |
| 842 | ** been enabled. |
| 843 | */ |
| 844 | ioc->pdir_base[off] &= ~(0x80000000000000FFULL); |
| 845 | #else |
| 846 | /* |
| 847 | ** If we want to maintain the PDIR as valid, put in |
| 848 | ** the spill page so devices prefetching won't |
| 849 | ** cause a hard fail. |
| 850 | */ |
| 851 | ioc->pdir_base[off] = (0x80000000000000FFULL | prefetch_spill_page); |
| 852 | #endif |
| 853 | } else { |
| 854 | u32 t = get_iovp_order(byte_cnt) + iovp_shift; |
| 855 | |
| 856 | iovp |= t; |
| 857 | ASSERT(t <= 31); /* 2GB! Max value of "size" field */ |
| 858 | |
| 859 | do { |
| 860 | /* verify this pdir entry is enabled */ |
| 861 | ASSERT(ioc->pdir_base[off] >> 63); |
| 862 | #ifndef FULL_VALID_PDIR |
| 863 | /* clear I/O Pdir entry "valid" bit first */ |
| 864 | ioc->pdir_base[off] &= ~(0x80000000000000FFULL); |
| 865 | #else |
| 866 | ioc->pdir_base[off] = (0x80000000000000FFULL | prefetch_spill_page); |
| 867 | #endif |
| 868 | off++; |
| 869 | byte_cnt -= iovp_size; |
| 870 | } while (byte_cnt > 0); |
| 871 | } |
| 872 | |
| 873 | WRITE_REG(iovp | ioc->ibase, ioc->ioc_hpa+IOC_PCOM); |
| 874 | } |
| 875 | |
| 876 | /** |
| 877 | * sba_map_single - map one buffer and return IOVA for DMA |
| 878 | * @dev: instance of PCI owned by the driver that's asking. |
| 879 | * @addr: driver buffer to map. |
| 880 | * @size: number of bytes to map in driver buffer. |
| 881 | * @dir: R/W or both. |
| 882 | * |
| 883 | * See Documentation/DMA-mapping.txt |
| 884 | */ |
| 885 | dma_addr_t |
| 886 | sba_map_single(struct device *dev, void *addr, size_t size, int dir) |
| 887 | { |
| 888 | struct ioc *ioc; |
| 889 | dma_addr_t iovp; |
| 890 | dma_addr_t offset; |
| 891 | u64 *pdir_start; |
| 892 | int pide; |
| 893 | #ifdef ASSERT_PDIR_SANITY |
| 894 | unsigned long flags; |
| 895 | #endif |
| 896 | #ifdef ALLOW_IOV_BYPASS |
| 897 | unsigned long pci_addr = virt_to_phys(addr); |
| 898 | #endif |
| 899 | |
| 900 | #ifdef ALLOW_IOV_BYPASS |
| 901 | ASSERT(to_pci_dev(dev)->dma_mask); |
| 902 | /* |
| 903 | ** Check if the PCI device can DMA to ptr... if so, just return ptr |
| 904 | */ |
| 905 | if (likely((pci_addr & ~to_pci_dev(dev)->dma_mask) == 0)) { |
| 906 | /* |
| 907 | ** Device is bit capable of DMA'ing to the buffer... |
| 908 | ** just return the PCI address of ptr |
| 909 | */ |
| 910 | DBG_BYPASS("sba_map_single() bypass mask/addr: 0x%lx/0x%lx\n", |
| 911 | to_pci_dev(dev)->dma_mask, pci_addr); |
| 912 | return pci_addr; |
| 913 | } |
| 914 | #endif |
| 915 | ioc = GET_IOC(dev); |
| 916 | ASSERT(ioc); |
| 917 | |
| 918 | prefetch(ioc->res_hint); |
| 919 | |
| 920 | ASSERT(size > 0); |
| 921 | ASSERT(size <= DMA_CHUNK_SIZE); |
| 922 | |
| 923 | /* save offset bits */ |
| 924 | offset = ((dma_addr_t) (long) addr) & ~iovp_mask; |
| 925 | |
| 926 | /* round up to nearest iovp_size */ |
| 927 | size = (size + offset + ~iovp_mask) & iovp_mask; |
| 928 | |
| 929 | #ifdef ASSERT_PDIR_SANITY |
| 930 | spin_lock_irqsave(&ioc->res_lock, flags); |
| 931 | if (sba_check_pdir(ioc,"Check before sba_map_single()")) |
| 932 | panic("Sanity check failed"); |
| 933 | spin_unlock_irqrestore(&ioc->res_lock, flags); |
| 934 | #endif |
| 935 | |
| 936 | pide = sba_alloc_range(ioc, size); |
| 937 | |
| 938 | iovp = (dma_addr_t) pide << iovp_shift; |
| 939 | |
| 940 | DBG_RUN("%s() 0x%p -> 0x%lx\n", |
| 941 | __FUNCTION__, addr, (long) iovp | offset); |
| 942 | |
| 943 | pdir_start = &(ioc->pdir_base[pide]); |
| 944 | |
| 945 | while (size > 0) { |
| 946 | ASSERT(((u8 *)pdir_start)[7] == 0); /* verify availability */ |
| 947 | sba_io_pdir_entry(pdir_start, (unsigned long) addr); |
| 948 | |
| 949 | DBG_RUN(" pdir 0x%p %lx\n", pdir_start, *pdir_start); |
| 950 | |
| 951 | addr += iovp_size; |
| 952 | size -= iovp_size; |
| 953 | pdir_start++; |
| 954 | } |
| 955 | /* force pdir update */ |
| 956 | wmb(); |
| 957 | |
| 958 | /* form complete address */ |
| 959 | #ifdef ASSERT_PDIR_SANITY |
| 960 | spin_lock_irqsave(&ioc->res_lock, flags); |
| 961 | sba_check_pdir(ioc,"Check after sba_map_single()"); |
| 962 | spin_unlock_irqrestore(&ioc->res_lock, flags); |
| 963 | #endif |
| 964 | return SBA_IOVA(ioc, iovp, offset); |
| 965 | } |
| 966 | |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 967 | #ifdef ENABLE_MARK_CLEAN |
| 968 | static SBA_INLINE void |
| 969 | sba_mark_clean(struct ioc *ioc, dma_addr_t iova, size_t size) |
| 970 | { |
| 971 | u32 iovp = (u32) SBA_IOVP(ioc,iova); |
| 972 | int off = PDIR_INDEX(iovp); |
| 973 | void *addr; |
| 974 | |
| 975 | if (size <= iovp_size) { |
| 976 | addr = phys_to_virt(ioc->pdir_base[off] & |
| 977 | ~0xE000000000000FFFULL); |
| 978 | mark_clean(addr, size); |
| 979 | } else { |
| 980 | do { |
| 981 | addr = phys_to_virt(ioc->pdir_base[off] & |
| 982 | ~0xE000000000000FFFULL); |
| 983 | mark_clean(addr, min(size, iovp_size)); |
| 984 | off++; |
| 985 | size -= iovp_size; |
| 986 | } while (size > 0); |
| 987 | } |
| 988 | } |
| 989 | #endif |
| 990 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 991 | /** |
| 992 | * sba_unmap_single - unmap one IOVA and free resources |
| 993 | * @dev: instance of PCI owned by the driver that's asking. |
| 994 | * @iova: IOVA of driver buffer previously mapped. |
| 995 | * @size: number of bytes mapped in driver buffer. |
| 996 | * @dir: R/W or both. |
| 997 | * |
| 998 | * See Documentation/DMA-mapping.txt |
| 999 | */ |
| 1000 | void sba_unmap_single(struct device *dev, dma_addr_t iova, size_t size, int dir) |
| 1001 | { |
| 1002 | struct ioc *ioc; |
| 1003 | #if DELAYED_RESOURCE_CNT > 0 |
| 1004 | struct sba_dma_pair *d; |
| 1005 | #endif |
| 1006 | unsigned long flags; |
| 1007 | dma_addr_t offset; |
| 1008 | |
| 1009 | ioc = GET_IOC(dev); |
| 1010 | ASSERT(ioc); |
| 1011 | |
| 1012 | #ifdef ALLOW_IOV_BYPASS |
| 1013 | if (likely((iova & ioc->imask) != ioc->ibase)) { |
| 1014 | /* |
| 1015 | ** Address does not fall w/in IOVA, must be bypassing |
| 1016 | */ |
| 1017 | DBG_BYPASS("sba_unmap_single() bypass addr: 0x%lx\n", iova); |
| 1018 | |
| 1019 | #ifdef ENABLE_MARK_CLEAN |
| 1020 | if (dir == DMA_FROM_DEVICE) { |
| 1021 | mark_clean(phys_to_virt(iova), size); |
| 1022 | } |
| 1023 | #endif |
| 1024 | return; |
| 1025 | } |
| 1026 | #endif |
| 1027 | offset = iova & ~iovp_mask; |
| 1028 | |
| 1029 | DBG_RUN("%s() iovp 0x%lx/%x\n", |
| 1030 | __FUNCTION__, (long) iova, size); |
| 1031 | |
| 1032 | iova ^= offset; /* clear offset bits */ |
| 1033 | size += offset; |
| 1034 | size = ROUNDUP(size, iovp_size); |
| 1035 | |
Alex Williamson | 5f6602a | 2005-04-25 13:14:36 -0700 | [diff] [blame] | 1036 | #ifdef ENABLE_MARK_CLEAN |
| 1037 | if (dir == DMA_FROM_DEVICE) |
| 1038 | sba_mark_clean(ioc, iova, size); |
| 1039 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1040 | |
| 1041 | #if DELAYED_RESOURCE_CNT > 0 |
| 1042 | spin_lock_irqsave(&ioc->saved_lock, flags); |
| 1043 | d = &(ioc->saved[ioc->saved_cnt]); |
| 1044 | d->iova = iova; |
| 1045 | d->size = size; |
| 1046 | if (unlikely(++(ioc->saved_cnt) >= DELAYED_RESOURCE_CNT)) { |
| 1047 | int cnt = ioc->saved_cnt; |
| 1048 | spin_lock(&ioc->res_lock); |
| 1049 | while (cnt--) { |
| 1050 | sba_mark_invalid(ioc, d->iova, d->size); |
| 1051 | sba_free_range(ioc, d->iova, d->size); |
| 1052 | d--; |
| 1053 | } |
| 1054 | ioc->saved_cnt = 0; |
| 1055 | READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */ |
| 1056 | spin_unlock(&ioc->res_lock); |
| 1057 | } |
| 1058 | spin_unlock_irqrestore(&ioc->saved_lock, flags); |
| 1059 | #else /* DELAYED_RESOURCE_CNT == 0 */ |
| 1060 | spin_lock_irqsave(&ioc->res_lock, flags); |
| 1061 | sba_mark_invalid(ioc, iova, size); |
| 1062 | sba_free_range(ioc, iova, size); |
| 1063 | READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */ |
| 1064 | spin_unlock_irqrestore(&ioc->res_lock, flags); |
| 1065 | #endif /* DELAYED_RESOURCE_CNT == 0 */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1066 | } |
| 1067 | |
| 1068 | |
| 1069 | /** |
| 1070 | * sba_alloc_coherent - allocate/map shared mem for DMA |
| 1071 | * @dev: instance of PCI owned by the driver that's asking. |
| 1072 | * @size: number of bytes mapped in driver buffer. |
| 1073 | * @dma_handle: IOVA of new buffer. |
| 1074 | * |
| 1075 | * See Documentation/DMA-mapping.txt |
| 1076 | */ |
| 1077 | void * |
Al Viro | 06a5449 | 2005-10-21 03:21:03 -0400 | [diff] [blame] | 1078 | sba_alloc_coherent (struct device *dev, size_t size, dma_addr_t *dma_handle, gfp_t flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1079 | { |
| 1080 | struct ioc *ioc; |
| 1081 | void *addr; |
| 1082 | |
| 1083 | ioc = GET_IOC(dev); |
| 1084 | ASSERT(ioc); |
| 1085 | |
| 1086 | #ifdef CONFIG_NUMA |
| 1087 | { |
| 1088 | struct page *page; |
| 1089 | page = alloc_pages_node(ioc->node == MAX_NUMNODES ? |
| 1090 | numa_node_id() : ioc->node, flags, |
| 1091 | get_order(size)); |
| 1092 | |
| 1093 | if (unlikely(!page)) |
| 1094 | return NULL; |
| 1095 | |
| 1096 | addr = page_address(page); |
| 1097 | } |
| 1098 | #else |
| 1099 | addr = (void *) __get_free_pages(flags, get_order(size)); |
| 1100 | #endif |
| 1101 | if (unlikely(!addr)) |
| 1102 | return NULL; |
| 1103 | |
| 1104 | memset(addr, 0, size); |
| 1105 | *dma_handle = virt_to_phys(addr); |
| 1106 | |
| 1107 | #ifdef ALLOW_IOV_BYPASS |
| 1108 | ASSERT(dev->coherent_dma_mask); |
| 1109 | /* |
| 1110 | ** Check if the PCI device can DMA to ptr... if so, just return ptr |
| 1111 | */ |
| 1112 | if (likely((*dma_handle & ~dev->coherent_dma_mask) == 0)) { |
| 1113 | DBG_BYPASS("sba_alloc_coherent() bypass mask/addr: 0x%lx/0x%lx\n", |
| 1114 | dev->coherent_dma_mask, *dma_handle); |
| 1115 | |
| 1116 | return addr; |
| 1117 | } |
| 1118 | #endif |
| 1119 | |
| 1120 | /* |
| 1121 | * If device can't bypass or bypass is disabled, pass the 32bit fake |
| 1122 | * device to map single to get an iova mapping. |
| 1123 | */ |
| 1124 | *dma_handle = sba_map_single(&ioc->sac_only_dev->dev, addr, size, 0); |
| 1125 | |
| 1126 | return addr; |
| 1127 | } |
| 1128 | |
| 1129 | |
| 1130 | /** |
| 1131 | * sba_free_coherent - free/unmap shared mem for DMA |
| 1132 | * @dev: instance of PCI owned by the driver that's asking. |
| 1133 | * @size: number of bytes mapped in driver buffer. |
| 1134 | * @vaddr: virtual address IOVA of "consistent" buffer. |
| 1135 | * @dma_handler: IO virtual address of "consistent" buffer. |
| 1136 | * |
| 1137 | * See Documentation/DMA-mapping.txt |
| 1138 | */ |
| 1139 | void sba_free_coherent (struct device *dev, size_t size, void *vaddr, dma_addr_t dma_handle) |
| 1140 | { |
| 1141 | sba_unmap_single(dev, dma_handle, size, 0); |
| 1142 | free_pages((unsigned long) vaddr, get_order(size)); |
| 1143 | } |
| 1144 | |
| 1145 | |
| 1146 | /* |
| 1147 | ** Since 0 is a valid pdir_base index value, can't use that |
| 1148 | ** to determine if a value is valid or not. Use a flag to indicate |
| 1149 | ** the SG list entry contains a valid pdir index. |
| 1150 | */ |
| 1151 | #define PIDE_FLAG 0x1UL |
| 1152 | |
| 1153 | #ifdef DEBUG_LARGE_SG_ENTRIES |
| 1154 | int dump_run_sg = 0; |
| 1155 | #endif |
| 1156 | |
| 1157 | |
| 1158 | /** |
| 1159 | * sba_fill_pdir - write allocated SG entries into IO PDIR |
| 1160 | * @ioc: IO MMU structure which owns the pdir we are interested in. |
| 1161 | * @startsg: list of IOVA/size pairs |
| 1162 | * @nents: number of entries in startsg list |
| 1163 | * |
| 1164 | * Take preprocessed SG list and write corresponding entries |
| 1165 | * in the IO PDIR. |
| 1166 | */ |
| 1167 | |
| 1168 | static SBA_INLINE int |
| 1169 | sba_fill_pdir( |
| 1170 | struct ioc *ioc, |
| 1171 | struct scatterlist *startsg, |
| 1172 | int nents) |
| 1173 | { |
| 1174 | struct scatterlist *dma_sg = startsg; /* pointer to current DMA */ |
| 1175 | int n_mappings = 0; |
| 1176 | u64 *pdirp = NULL; |
| 1177 | unsigned long dma_offset = 0; |
| 1178 | |
| 1179 | dma_sg--; |
| 1180 | while (nents-- > 0) { |
| 1181 | int cnt = startsg->dma_length; |
| 1182 | startsg->dma_length = 0; |
| 1183 | |
| 1184 | #ifdef DEBUG_LARGE_SG_ENTRIES |
| 1185 | if (dump_run_sg) |
| 1186 | printk(" %2d : %08lx/%05x %p\n", |
| 1187 | nents, startsg->dma_address, cnt, |
| 1188 | sba_sg_address(startsg)); |
| 1189 | #else |
| 1190 | DBG_RUN_SG(" %d : %08lx/%05x %p\n", |
| 1191 | nents, startsg->dma_address, cnt, |
| 1192 | sba_sg_address(startsg)); |
| 1193 | #endif |
| 1194 | /* |
| 1195 | ** Look for the start of a new DMA stream |
| 1196 | */ |
| 1197 | if (startsg->dma_address & PIDE_FLAG) { |
| 1198 | u32 pide = startsg->dma_address & ~PIDE_FLAG; |
| 1199 | dma_offset = (unsigned long) pide & ~iovp_mask; |
| 1200 | startsg->dma_address = 0; |
| 1201 | dma_sg++; |
| 1202 | dma_sg->dma_address = pide | ioc->ibase; |
| 1203 | pdirp = &(ioc->pdir_base[pide >> iovp_shift]); |
| 1204 | n_mappings++; |
| 1205 | } |
| 1206 | |
| 1207 | /* |
| 1208 | ** Look for a VCONTIG chunk |
| 1209 | */ |
| 1210 | if (cnt) { |
| 1211 | unsigned long vaddr = (unsigned long) sba_sg_address(startsg); |
| 1212 | ASSERT(pdirp); |
| 1213 | |
| 1214 | /* Since multiple Vcontig blocks could make up |
| 1215 | ** one DMA stream, *add* cnt to dma_len. |
| 1216 | */ |
| 1217 | dma_sg->dma_length += cnt; |
| 1218 | cnt += dma_offset; |
| 1219 | dma_offset=0; /* only want offset on first chunk */ |
| 1220 | cnt = ROUNDUP(cnt, iovp_size); |
| 1221 | do { |
| 1222 | sba_io_pdir_entry(pdirp, vaddr); |
| 1223 | vaddr += iovp_size; |
| 1224 | cnt -= iovp_size; |
| 1225 | pdirp++; |
| 1226 | } while (cnt > 0); |
| 1227 | } |
| 1228 | startsg++; |
| 1229 | } |
| 1230 | /* force pdir update */ |
| 1231 | wmb(); |
| 1232 | |
| 1233 | #ifdef DEBUG_LARGE_SG_ENTRIES |
| 1234 | dump_run_sg = 0; |
| 1235 | #endif |
| 1236 | return(n_mappings); |
| 1237 | } |
| 1238 | |
| 1239 | |
| 1240 | /* |
| 1241 | ** Two address ranges are DMA contiguous *iff* "end of prev" and |
| 1242 | ** "start of next" are both on an IOV page boundary. |
| 1243 | ** |
| 1244 | ** (shift left is a quick trick to mask off upper bits) |
| 1245 | */ |
| 1246 | #define DMA_CONTIG(__X, __Y) \ |
| 1247 | (((((unsigned long) __X) | ((unsigned long) __Y)) << (BITS_PER_LONG - iovp_shift)) == 0UL) |
| 1248 | |
| 1249 | |
| 1250 | /** |
| 1251 | * sba_coalesce_chunks - preprocess the SG list |
| 1252 | * @ioc: IO MMU structure which owns the pdir we are interested in. |
| 1253 | * @startsg: list of IOVA/size pairs |
| 1254 | * @nents: number of entries in startsg list |
| 1255 | * |
| 1256 | * First pass is to walk the SG list and determine where the breaks are |
| 1257 | * in the DMA stream. Allocates PDIR entries but does not fill them. |
| 1258 | * Returns the number of DMA chunks. |
| 1259 | * |
| 1260 | * Doing the fill separate from the coalescing/allocation keeps the |
| 1261 | * code simpler. Future enhancement could make one pass through |
| 1262 | * the sglist do both. |
| 1263 | */ |
| 1264 | static SBA_INLINE int |
| 1265 | sba_coalesce_chunks( struct ioc *ioc, |
| 1266 | struct scatterlist *startsg, |
| 1267 | int nents) |
| 1268 | { |
| 1269 | struct scatterlist *vcontig_sg; /* VCONTIG chunk head */ |
| 1270 | unsigned long vcontig_len; /* len of VCONTIG chunk */ |
| 1271 | unsigned long vcontig_end; |
| 1272 | struct scatterlist *dma_sg; /* next DMA stream head */ |
| 1273 | unsigned long dma_offset, dma_len; /* start/len of DMA stream */ |
| 1274 | int n_mappings = 0; |
| 1275 | |
| 1276 | while (nents > 0) { |
| 1277 | unsigned long vaddr = (unsigned long) sba_sg_address(startsg); |
| 1278 | |
| 1279 | /* |
| 1280 | ** Prepare for first/next DMA stream |
| 1281 | */ |
| 1282 | dma_sg = vcontig_sg = startsg; |
| 1283 | dma_len = vcontig_len = vcontig_end = startsg->length; |
| 1284 | vcontig_end += vaddr; |
| 1285 | dma_offset = vaddr & ~iovp_mask; |
| 1286 | |
| 1287 | /* PARANOID: clear entries */ |
| 1288 | startsg->dma_address = startsg->dma_length = 0; |
| 1289 | |
| 1290 | /* |
| 1291 | ** This loop terminates one iteration "early" since |
| 1292 | ** it's always looking one "ahead". |
| 1293 | */ |
| 1294 | while (--nents > 0) { |
| 1295 | unsigned long vaddr; /* tmp */ |
| 1296 | |
| 1297 | startsg++; |
| 1298 | |
| 1299 | /* PARANOID */ |
| 1300 | startsg->dma_address = startsg->dma_length = 0; |
| 1301 | |
| 1302 | /* catch brokenness in SCSI layer */ |
| 1303 | ASSERT(startsg->length <= DMA_CHUNK_SIZE); |
| 1304 | |
| 1305 | /* |
| 1306 | ** First make sure current dma stream won't |
| 1307 | ** exceed DMA_CHUNK_SIZE if we coalesce the |
| 1308 | ** next entry. |
| 1309 | */ |
| 1310 | if (((dma_len + dma_offset + startsg->length + ~iovp_mask) & iovp_mask) |
| 1311 | > DMA_CHUNK_SIZE) |
| 1312 | break; |
| 1313 | |
| 1314 | /* |
| 1315 | ** Then look for virtually contiguous blocks. |
| 1316 | ** |
| 1317 | ** append the next transaction? |
| 1318 | */ |
| 1319 | vaddr = (unsigned long) sba_sg_address(startsg); |
| 1320 | if (vcontig_end == vaddr) |
| 1321 | { |
| 1322 | vcontig_len += startsg->length; |
| 1323 | vcontig_end += startsg->length; |
| 1324 | dma_len += startsg->length; |
| 1325 | continue; |
| 1326 | } |
| 1327 | |
| 1328 | #ifdef DEBUG_LARGE_SG_ENTRIES |
| 1329 | dump_run_sg = (vcontig_len > iovp_size); |
| 1330 | #endif |
| 1331 | |
| 1332 | /* |
| 1333 | ** Not virtually contigous. |
| 1334 | ** Terminate prev chunk. |
| 1335 | ** Start a new chunk. |
| 1336 | ** |
| 1337 | ** Once we start a new VCONTIG chunk, dma_offset |
| 1338 | ** can't change. And we need the offset from the first |
| 1339 | ** chunk - not the last one. Ergo Successive chunks |
| 1340 | ** must start on page boundaries and dove tail |
| 1341 | ** with it's predecessor. |
| 1342 | */ |
| 1343 | vcontig_sg->dma_length = vcontig_len; |
| 1344 | |
| 1345 | vcontig_sg = startsg; |
| 1346 | vcontig_len = startsg->length; |
| 1347 | |
| 1348 | /* |
| 1349 | ** 3) do the entries end/start on page boundaries? |
| 1350 | ** Don't update vcontig_end until we've checked. |
| 1351 | */ |
| 1352 | if (DMA_CONTIG(vcontig_end, vaddr)) |
| 1353 | { |
| 1354 | vcontig_end = vcontig_len + vaddr; |
| 1355 | dma_len += vcontig_len; |
| 1356 | continue; |
| 1357 | } else { |
| 1358 | break; |
| 1359 | } |
| 1360 | } |
| 1361 | |
| 1362 | /* |
| 1363 | ** End of DMA Stream |
| 1364 | ** Terminate last VCONTIG block. |
| 1365 | ** Allocate space for DMA stream. |
| 1366 | */ |
| 1367 | vcontig_sg->dma_length = vcontig_len; |
| 1368 | dma_len = (dma_len + dma_offset + ~iovp_mask) & iovp_mask; |
| 1369 | ASSERT(dma_len <= DMA_CHUNK_SIZE); |
| 1370 | dma_sg->dma_address = (dma_addr_t) (PIDE_FLAG |
| 1371 | | (sba_alloc_range(ioc, dma_len) << iovp_shift) |
| 1372 | | dma_offset); |
| 1373 | n_mappings++; |
| 1374 | } |
| 1375 | |
| 1376 | return n_mappings; |
| 1377 | } |
| 1378 | |
| 1379 | |
| 1380 | /** |
| 1381 | * sba_map_sg - map Scatter/Gather list |
| 1382 | * @dev: instance of PCI owned by the driver that's asking. |
| 1383 | * @sglist: array of buffer/length pairs |
| 1384 | * @nents: number of entries in list |
| 1385 | * @dir: R/W or both. |
| 1386 | * |
| 1387 | * See Documentation/DMA-mapping.txt |
| 1388 | */ |
| 1389 | int sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents, int dir) |
| 1390 | { |
| 1391 | struct ioc *ioc; |
| 1392 | int coalesced, filled = 0; |
| 1393 | #ifdef ASSERT_PDIR_SANITY |
| 1394 | unsigned long flags; |
| 1395 | #endif |
| 1396 | #ifdef ALLOW_IOV_BYPASS_SG |
| 1397 | struct scatterlist *sg; |
| 1398 | #endif |
| 1399 | |
| 1400 | DBG_RUN_SG("%s() START %d entries\n", __FUNCTION__, nents); |
| 1401 | ioc = GET_IOC(dev); |
| 1402 | ASSERT(ioc); |
| 1403 | |
| 1404 | #ifdef ALLOW_IOV_BYPASS_SG |
| 1405 | ASSERT(to_pci_dev(dev)->dma_mask); |
| 1406 | if (likely((ioc->dma_mask & ~to_pci_dev(dev)->dma_mask) == 0)) { |
| 1407 | for (sg = sglist ; filled < nents ; filled++, sg++){ |
| 1408 | sg->dma_length = sg->length; |
| 1409 | sg->dma_address = virt_to_phys(sba_sg_address(sg)); |
| 1410 | } |
| 1411 | return filled; |
| 1412 | } |
| 1413 | #endif |
| 1414 | /* Fast path single entry scatterlists. */ |
| 1415 | if (nents == 1) { |
| 1416 | sglist->dma_length = sglist->length; |
| 1417 | sglist->dma_address = sba_map_single(dev, sba_sg_address(sglist), sglist->length, dir); |
| 1418 | return 1; |
| 1419 | } |
| 1420 | |
| 1421 | #ifdef ASSERT_PDIR_SANITY |
| 1422 | spin_lock_irqsave(&ioc->res_lock, flags); |
| 1423 | if (sba_check_pdir(ioc,"Check before sba_map_sg()")) |
| 1424 | { |
| 1425 | sba_dump_sg(ioc, sglist, nents); |
| 1426 | panic("Check before sba_map_sg()"); |
| 1427 | } |
| 1428 | spin_unlock_irqrestore(&ioc->res_lock, flags); |
| 1429 | #endif |
| 1430 | |
| 1431 | prefetch(ioc->res_hint); |
| 1432 | |
| 1433 | /* |
| 1434 | ** First coalesce the chunks and allocate I/O pdir space |
| 1435 | ** |
| 1436 | ** If this is one DMA stream, we can properly map using the |
| 1437 | ** correct virtual address associated with each DMA page. |
| 1438 | ** w/o this association, we wouldn't have coherent DMA! |
| 1439 | ** Access to the virtual address is what forces a two pass algorithm. |
| 1440 | */ |
| 1441 | coalesced = sba_coalesce_chunks(ioc, sglist, nents); |
| 1442 | |
| 1443 | /* |
| 1444 | ** Program the I/O Pdir |
| 1445 | ** |
| 1446 | ** map the virtual addresses to the I/O Pdir |
| 1447 | ** o dma_address will contain the pdir index |
| 1448 | ** o dma_len will contain the number of bytes to map |
| 1449 | ** o address contains the virtual address. |
| 1450 | */ |
| 1451 | filled = sba_fill_pdir(ioc, sglist, nents); |
| 1452 | |
| 1453 | #ifdef ASSERT_PDIR_SANITY |
| 1454 | spin_lock_irqsave(&ioc->res_lock, flags); |
| 1455 | if (sba_check_pdir(ioc,"Check after sba_map_sg()")) |
| 1456 | { |
| 1457 | sba_dump_sg(ioc, sglist, nents); |
| 1458 | panic("Check after sba_map_sg()\n"); |
| 1459 | } |
| 1460 | spin_unlock_irqrestore(&ioc->res_lock, flags); |
| 1461 | #endif |
| 1462 | |
| 1463 | ASSERT(coalesced == filled); |
| 1464 | DBG_RUN_SG("%s() DONE %d mappings\n", __FUNCTION__, filled); |
| 1465 | |
| 1466 | return filled; |
| 1467 | } |
| 1468 | |
| 1469 | |
| 1470 | /** |
| 1471 | * sba_unmap_sg - unmap Scatter/Gather list |
| 1472 | * @dev: instance of PCI owned by the driver that's asking. |
| 1473 | * @sglist: array of buffer/length pairs |
| 1474 | * @nents: number of entries in list |
| 1475 | * @dir: R/W or both. |
| 1476 | * |
| 1477 | * See Documentation/DMA-mapping.txt |
| 1478 | */ |
| 1479 | void sba_unmap_sg (struct device *dev, struct scatterlist *sglist, int nents, int dir) |
| 1480 | { |
| 1481 | #ifdef ASSERT_PDIR_SANITY |
| 1482 | struct ioc *ioc; |
| 1483 | unsigned long flags; |
| 1484 | #endif |
| 1485 | |
| 1486 | DBG_RUN_SG("%s() START %d entries, %p,%x\n", |
| 1487 | __FUNCTION__, nents, sba_sg_address(sglist), sglist->length); |
| 1488 | |
| 1489 | #ifdef ASSERT_PDIR_SANITY |
| 1490 | ioc = GET_IOC(dev); |
| 1491 | ASSERT(ioc); |
| 1492 | |
| 1493 | spin_lock_irqsave(&ioc->res_lock, flags); |
| 1494 | sba_check_pdir(ioc,"Check before sba_unmap_sg()"); |
| 1495 | spin_unlock_irqrestore(&ioc->res_lock, flags); |
| 1496 | #endif |
| 1497 | |
| 1498 | while (nents && sglist->dma_length) { |
| 1499 | |
| 1500 | sba_unmap_single(dev, sglist->dma_address, sglist->dma_length, dir); |
| 1501 | sglist++; |
| 1502 | nents--; |
| 1503 | } |
| 1504 | |
| 1505 | DBG_RUN_SG("%s() DONE (nents %d)\n", __FUNCTION__, nents); |
| 1506 | |
| 1507 | #ifdef ASSERT_PDIR_SANITY |
| 1508 | spin_lock_irqsave(&ioc->res_lock, flags); |
| 1509 | sba_check_pdir(ioc,"Check after sba_unmap_sg()"); |
| 1510 | spin_unlock_irqrestore(&ioc->res_lock, flags); |
| 1511 | #endif |
| 1512 | |
| 1513 | } |
| 1514 | |
| 1515 | /************************************************************** |
| 1516 | * |
| 1517 | * Initialization and claim |
| 1518 | * |
| 1519 | ***************************************************************/ |
| 1520 | |
| 1521 | static void __init |
| 1522 | ioc_iova_init(struct ioc *ioc) |
| 1523 | { |
| 1524 | int tcnfg; |
| 1525 | int agp_found = 0; |
| 1526 | struct pci_dev *device = NULL; |
| 1527 | #ifdef FULL_VALID_PDIR |
| 1528 | unsigned long index; |
| 1529 | #endif |
| 1530 | |
| 1531 | /* |
| 1532 | ** Firmware programs the base and size of a "safe IOVA space" |
| 1533 | ** (one that doesn't overlap memory or LMMIO space) in the |
| 1534 | ** IBASE and IMASK registers. |
| 1535 | */ |
| 1536 | ioc->ibase = READ_REG(ioc->ioc_hpa + IOC_IBASE) & ~0x1UL; |
| 1537 | ioc->imask = READ_REG(ioc->ioc_hpa + IOC_IMASK) | 0xFFFFFFFF00000000UL; |
| 1538 | |
| 1539 | ioc->iov_size = ~ioc->imask + 1; |
| 1540 | |
| 1541 | DBG_INIT("%s() hpa %p IOV base 0x%lx mask 0x%lx (%dMB)\n", |
| 1542 | __FUNCTION__, ioc->ioc_hpa, ioc->ibase, ioc->imask, |
| 1543 | ioc->iov_size >> 20); |
| 1544 | |
| 1545 | switch (iovp_size) { |
| 1546 | case 4*1024: tcnfg = 0; break; |
| 1547 | case 8*1024: tcnfg = 1; break; |
| 1548 | case 16*1024: tcnfg = 2; break; |
| 1549 | case 64*1024: tcnfg = 3; break; |
| 1550 | default: |
| 1551 | panic(PFX "Unsupported IOTLB page size %ldK", |
| 1552 | iovp_size >> 10); |
| 1553 | break; |
| 1554 | } |
| 1555 | WRITE_REG(tcnfg, ioc->ioc_hpa + IOC_TCNFG); |
| 1556 | |
| 1557 | ioc->pdir_size = (ioc->iov_size / iovp_size) * PDIR_ENTRY_SIZE; |
| 1558 | ioc->pdir_base = (void *) __get_free_pages(GFP_KERNEL, |
| 1559 | get_order(ioc->pdir_size)); |
| 1560 | if (!ioc->pdir_base) |
| 1561 | panic(PFX "Couldn't allocate I/O Page Table\n"); |
| 1562 | |
| 1563 | memset(ioc->pdir_base, 0, ioc->pdir_size); |
| 1564 | |
| 1565 | DBG_INIT("%s() IOV page size %ldK pdir %p size %x\n", __FUNCTION__, |
| 1566 | iovp_size >> 10, ioc->pdir_base, ioc->pdir_size); |
| 1567 | |
| 1568 | ASSERT(ALIGN((unsigned long) ioc->pdir_base, 4*1024) == (unsigned long) ioc->pdir_base); |
| 1569 | WRITE_REG(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE); |
| 1570 | |
| 1571 | /* |
| 1572 | ** If an AGP device is present, only use half of the IOV space |
| 1573 | ** for PCI DMA. Unfortunately we can't know ahead of time |
| 1574 | ** whether GART support will actually be used, for now we |
| 1575 | ** can just key on an AGP device found in the system. |
| 1576 | ** We program the next pdir index after we stop w/ a key for |
| 1577 | ** the GART code to handshake on. |
| 1578 | */ |
| 1579 | for_each_pci_dev(device) |
| 1580 | agp_found |= pci_find_capability(device, PCI_CAP_ID_AGP); |
| 1581 | |
| 1582 | if (agp_found && reserve_sba_gart) { |
| 1583 | printk(KERN_INFO PFX "reserving %dMb of IOVA space at 0x%lx for agpgart\n", |
| 1584 | ioc->iov_size/2 >> 20, ioc->ibase + ioc->iov_size/2); |
| 1585 | ioc->pdir_size /= 2; |
| 1586 | ((u64 *)ioc->pdir_base)[PDIR_INDEX(ioc->iov_size/2)] = ZX1_SBA_IOMMU_COOKIE; |
| 1587 | } |
| 1588 | #ifdef FULL_VALID_PDIR |
| 1589 | /* |
| 1590 | ** Check to see if the spill page has been allocated, we don't need more than |
| 1591 | ** one across multiple SBAs. |
| 1592 | */ |
| 1593 | if (!prefetch_spill_page) { |
| 1594 | char *spill_poison = "SBAIOMMU POISON"; |
| 1595 | int poison_size = 16; |
| 1596 | void *poison_addr, *addr; |
| 1597 | |
| 1598 | addr = (void *)__get_free_pages(GFP_KERNEL, get_order(iovp_size)); |
| 1599 | if (!addr) |
| 1600 | panic(PFX "Couldn't allocate PDIR spill page\n"); |
| 1601 | |
| 1602 | poison_addr = addr; |
| 1603 | for ( ; (u64) poison_addr < addr + iovp_size; poison_addr += poison_size) |
| 1604 | memcpy(poison_addr, spill_poison, poison_size); |
| 1605 | |
| 1606 | prefetch_spill_page = virt_to_phys(addr); |
| 1607 | |
| 1608 | DBG_INIT("%s() prefetch spill addr: 0x%lx\n", __FUNCTION__, prefetch_spill_page); |
| 1609 | } |
| 1610 | /* |
| 1611 | ** Set all the PDIR entries valid w/ the spill page as the target |
| 1612 | */ |
| 1613 | for (index = 0 ; index < (ioc->pdir_size / PDIR_ENTRY_SIZE) ; index++) |
| 1614 | ((u64 *)ioc->pdir_base)[index] = (0x80000000000000FF | prefetch_spill_page); |
| 1615 | #endif |
| 1616 | |
| 1617 | /* Clear I/O TLB of any possible entries */ |
| 1618 | WRITE_REG(ioc->ibase | (get_iovp_order(ioc->iov_size) + iovp_shift), ioc->ioc_hpa + IOC_PCOM); |
| 1619 | READ_REG(ioc->ioc_hpa + IOC_PCOM); |
| 1620 | |
| 1621 | /* Enable IOVA translation */ |
| 1622 | WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa + IOC_IBASE); |
| 1623 | READ_REG(ioc->ioc_hpa + IOC_IBASE); |
| 1624 | } |
| 1625 | |
| 1626 | static void __init |
| 1627 | ioc_resource_init(struct ioc *ioc) |
| 1628 | { |
| 1629 | spin_lock_init(&ioc->res_lock); |
| 1630 | #if DELAYED_RESOURCE_CNT > 0 |
| 1631 | spin_lock_init(&ioc->saved_lock); |
| 1632 | #endif |
| 1633 | |
| 1634 | /* resource map size dictated by pdir_size */ |
| 1635 | ioc->res_size = ioc->pdir_size / PDIR_ENTRY_SIZE; /* entries */ |
| 1636 | ioc->res_size >>= 3; /* convert bit count to byte count */ |
| 1637 | DBG_INIT("%s() res_size 0x%x\n", __FUNCTION__, ioc->res_size); |
| 1638 | |
| 1639 | ioc->res_map = (char *) __get_free_pages(GFP_KERNEL, |
| 1640 | get_order(ioc->res_size)); |
| 1641 | if (!ioc->res_map) |
| 1642 | panic(PFX "Couldn't allocate resource map\n"); |
| 1643 | |
| 1644 | memset(ioc->res_map, 0, ioc->res_size); |
| 1645 | /* next available IOVP - circular search */ |
| 1646 | ioc->res_hint = (unsigned long *) ioc->res_map; |
| 1647 | |
| 1648 | #ifdef ASSERT_PDIR_SANITY |
| 1649 | /* Mark first bit busy - ie no IOVA 0 */ |
| 1650 | ioc->res_map[0] = 0x1; |
| 1651 | ioc->pdir_base[0] = 0x8000000000000000ULL | ZX1_SBA_IOMMU_COOKIE; |
| 1652 | #endif |
| 1653 | #ifdef FULL_VALID_PDIR |
| 1654 | /* Mark the last resource used so we don't prefetch beyond IOVA space */ |
| 1655 | ioc->res_map[ioc->res_size - 1] |= 0x80UL; /* res_map is chars */ |
| 1656 | ioc->pdir_base[(ioc->pdir_size / PDIR_ENTRY_SIZE) - 1] = (0x80000000000000FF |
| 1657 | | prefetch_spill_page); |
| 1658 | #endif |
| 1659 | |
| 1660 | DBG_INIT("%s() res_map %x %p\n", __FUNCTION__, |
| 1661 | ioc->res_size, (void *) ioc->res_map); |
| 1662 | } |
| 1663 | |
| 1664 | static void __init |
| 1665 | ioc_sac_init(struct ioc *ioc) |
| 1666 | { |
| 1667 | struct pci_dev *sac = NULL; |
| 1668 | struct pci_controller *controller = NULL; |
| 1669 | |
| 1670 | /* |
| 1671 | * pci_alloc_coherent() must return a DMA address which is |
| 1672 | * SAC (single address cycle) addressable, so allocate a |
| 1673 | * pseudo-device to enforce that. |
| 1674 | */ |
| 1675 | sac = kmalloc(sizeof(*sac), GFP_KERNEL); |
| 1676 | if (!sac) |
| 1677 | panic(PFX "Couldn't allocate struct pci_dev"); |
| 1678 | memset(sac, 0, sizeof(*sac)); |
| 1679 | |
| 1680 | controller = kmalloc(sizeof(*controller), GFP_KERNEL); |
| 1681 | if (!controller) |
| 1682 | panic(PFX "Couldn't allocate struct pci_controller"); |
| 1683 | memset(controller, 0, sizeof(*controller)); |
| 1684 | |
| 1685 | controller->iommu = ioc; |
| 1686 | sac->sysdata = controller; |
| 1687 | sac->dma_mask = 0xFFFFFFFFUL; |
| 1688 | #ifdef CONFIG_PCI |
| 1689 | sac->dev.bus = &pci_bus_type; |
| 1690 | #endif |
| 1691 | ioc->sac_only_dev = sac; |
| 1692 | } |
| 1693 | |
| 1694 | static void __init |
| 1695 | ioc_zx1_init(struct ioc *ioc) |
| 1696 | { |
| 1697 | unsigned long rope_config; |
| 1698 | unsigned int i; |
| 1699 | |
| 1700 | if (ioc->rev < 0x20) |
| 1701 | panic(PFX "IOC 2.0 or later required for IOMMU support\n"); |
| 1702 | |
| 1703 | /* 38 bit memory controller + extra bit for range displaced by MMIO */ |
| 1704 | ioc->dma_mask = (0x1UL << 39) - 1; |
| 1705 | |
| 1706 | /* |
| 1707 | ** Clear ROPE(N)_CONFIG AO bit. |
| 1708 | ** Disables "NT Ordering" (~= !"Relaxed Ordering") |
| 1709 | ** Overrides bit 1 in DMA Hint Sets. |
| 1710 | ** Improves netperf UDP_STREAM by ~10% for tg3 on bcm5701. |
| 1711 | */ |
| 1712 | for (i=0; i<(8*8); i+=8) { |
| 1713 | rope_config = READ_REG(ioc->ioc_hpa + IOC_ROPE0_CFG + i); |
| 1714 | rope_config &= ~IOC_ROPE_AO; |
| 1715 | WRITE_REG(rope_config, ioc->ioc_hpa + IOC_ROPE0_CFG + i); |
| 1716 | } |
| 1717 | } |
| 1718 | |
| 1719 | typedef void (initfunc)(struct ioc *); |
| 1720 | |
| 1721 | struct ioc_iommu { |
| 1722 | u32 func_id; |
| 1723 | char *name; |
| 1724 | initfunc *init; |
| 1725 | }; |
| 1726 | |
| 1727 | static struct ioc_iommu ioc_iommu_info[] __initdata = { |
| 1728 | { ZX1_IOC_ID, "zx1", ioc_zx1_init }, |
| 1729 | { ZX2_IOC_ID, "zx2", NULL }, |
| 1730 | { SX1000_IOC_ID, "sx1000", NULL }, |
Bjorn Helgaas | e15da40 | 2005-05-03 12:07:00 -0700 | [diff] [blame] | 1731 | { SX2000_IOC_ID, "sx2000", NULL }, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1732 | }; |
| 1733 | |
| 1734 | static struct ioc * __init |
| 1735 | ioc_init(u64 hpa, void *handle) |
| 1736 | { |
| 1737 | struct ioc *ioc; |
| 1738 | struct ioc_iommu *info; |
| 1739 | |
| 1740 | ioc = kmalloc(sizeof(*ioc), GFP_KERNEL); |
| 1741 | if (!ioc) |
| 1742 | return NULL; |
| 1743 | |
| 1744 | memset(ioc, 0, sizeof(*ioc)); |
| 1745 | |
| 1746 | ioc->next = ioc_list; |
| 1747 | ioc_list = ioc; |
| 1748 | |
| 1749 | ioc->handle = handle; |
| 1750 | ioc->ioc_hpa = ioremap(hpa, 0x1000); |
| 1751 | |
| 1752 | ioc->func_id = READ_REG(ioc->ioc_hpa + IOC_FUNC_ID); |
| 1753 | ioc->rev = READ_REG(ioc->ioc_hpa + IOC_FCLASS) & 0xFFUL; |
| 1754 | ioc->dma_mask = 0xFFFFFFFFFFFFFFFFUL; /* conservative */ |
| 1755 | |
| 1756 | for (info = ioc_iommu_info; info < ioc_iommu_info + ARRAY_SIZE(ioc_iommu_info); info++) { |
| 1757 | if (ioc->func_id == info->func_id) { |
| 1758 | ioc->name = info->name; |
| 1759 | if (info->init) |
| 1760 | (info->init)(ioc); |
| 1761 | } |
| 1762 | } |
| 1763 | |
| 1764 | iovp_size = (1 << iovp_shift); |
| 1765 | iovp_mask = ~(iovp_size - 1); |
| 1766 | |
| 1767 | DBG_INIT("%s: PAGE_SIZE %ldK, iovp_size %ldK\n", __FUNCTION__, |
| 1768 | PAGE_SIZE >> 10, iovp_size >> 10); |
| 1769 | |
| 1770 | if (!ioc->name) { |
| 1771 | ioc->name = kmalloc(24, GFP_KERNEL); |
| 1772 | if (ioc->name) |
| 1773 | sprintf((char *) ioc->name, "Unknown (%04x:%04x)", |
| 1774 | ioc->func_id & 0xFFFF, (ioc->func_id >> 16) & 0xFFFF); |
| 1775 | else |
| 1776 | ioc->name = "Unknown"; |
| 1777 | } |
| 1778 | |
| 1779 | ioc_iova_init(ioc); |
| 1780 | ioc_resource_init(ioc); |
| 1781 | ioc_sac_init(ioc); |
| 1782 | |
| 1783 | if ((long) ~iovp_mask > (long) ia64_max_iommu_merge_mask) |
| 1784 | ia64_max_iommu_merge_mask = ~iovp_mask; |
| 1785 | |
| 1786 | printk(KERN_INFO PFX |
| 1787 | "%s %d.%d HPA 0x%lx IOVA space %dMb at 0x%lx\n", |
| 1788 | ioc->name, (ioc->rev >> 4) & 0xF, ioc->rev & 0xF, |
| 1789 | hpa, ioc->iov_size >> 20, ioc->ibase); |
| 1790 | |
| 1791 | return ioc; |
| 1792 | } |
| 1793 | |
| 1794 | |
| 1795 | |
| 1796 | /************************************************************************** |
| 1797 | ** |
| 1798 | ** SBA initialization code (HW and SW) |
| 1799 | ** |
| 1800 | ** o identify SBA chip itself |
| 1801 | ** o FIXME: initialize DMA hints for reasonable defaults |
| 1802 | ** |
| 1803 | **************************************************************************/ |
| 1804 | |
| 1805 | #ifdef CONFIG_PROC_FS |
| 1806 | static void * |
| 1807 | ioc_start(struct seq_file *s, loff_t *pos) |
| 1808 | { |
| 1809 | struct ioc *ioc; |
| 1810 | loff_t n = *pos; |
| 1811 | |
| 1812 | for (ioc = ioc_list; ioc; ioc = ioc->next) |
| 1813 | if (!n--) |
| 1814 | return ioc; |
| 1815 | |
| 1816 | return NULL; |
| 1817 | } |
| 1818 | |
| 1819 | static void * |
| 1820 | ioc_next(struct seq_file *s, void *v, loff_t *pos) |
| 1821 | { |
| 1822 | struct ioc *ioc = v; |
| 1823 | |
| 1824 | ++*pos; |
| 1825 | return ioc->next; |
| 1826 | } |
| 1827 | |
| 1828 | static void |
| 1829 | ioc_stop(struct seq_file *s, void *v) |
| 1830 | { |
| 1831 | } |
| 1832 | |
| 1833 | static int |
| 1834 | ioc_show(struct seq_file *s, void *v) |
| 1835 | { |
| 1836 | struct ioc *ioc = v; |
| 1837 | unsigned long *res_ptr = (unsigned long *)ioc->res_map; |
| 1838 | int i, used = 0; |
| 1839 | |
| 1840 | seq_printf(s, "Hewlett Packard %s IOC rev %d.%d\n", |
| 1841 | ioc->name, ((ioc->rev >> 4) & 0xF), (ioc->rev & 0xF)); |
| 1842 | #ifdef CONFIG_NUMA |
| 1843 | if (ioc->node != MAX_NUMNODES) |
| 1844 | seq_printf(s, "NUMA node : %d\n", ioc->node); |
| 1845 | #endif |
| 1846 | seq_printf(s, "IOVA size : %ld MB\n", ((ioc->pdir_size >> 3) * iovp_size)/(1024*1024)); |
| 1847 | seq_printf(s, "IOVA page size : %ld kb\n", iovp_size/1024); |
| 1848 | |
| 1849 | for (i = 0; i < (ioc->res_size / sizeof(unsigned long)); ++i, ++res_ptr) |
| 1850 | used += hweight64(*res_ptr); |
| 1851 | |
| 1852 | seq_printf(s, "PDIR size : %d entries\n", ioc->pdir_size >> 3); |
| 1853 | seq_printf(s, "PDIR used : %d entries\n", used); |
| 1854 | |
| 1855 | #ifdef PDIR_SEARCH_TIMING |
| 1856 | { |
| 1857 | unsigned long i = 0, avg = 0, min, max; |
| 1858 | min = max = ioc->avg_search[0]; |
| 1859 | for (i = 0; i < SBA_SEARCH_SAMPLE; i++) { |
| 1860 | avg += ioc->avg_search[i]; |
| 1861 | if (ioc->avg_search[i] > max) max = ioc->avg_search[i]; |
| 1862 | if (ioc->avg_search[i] < min) min = ioc->avg_search[i]; |
| 1863 | } |
| 1864 | avg /= SBA_SEARCH_SAMPLE; |
| 1865 | seq_printf(s, "Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles/IOVA page)\n", |
| 1866 | min, avg, max); |
| 1867 | } |
| 1868 | #endif |
| 1869 | #ifndef ALLOW_IOV_BYPASS |
| 1870 | seq_printf(s, "IOVA bypass disabled\n"); |
| 1871 | #endif |
| 1872 | return 0; |
| 1873 | } |
| 1874 | |
| 1875 | static struct seq_operations ioc_seq_ops = { |
| 1876 | .start = ioc_start, |
| 1877 | .next = ioc_next, |
| 1878 | .stop = ioc_stop, |
| 1879 | .show = ioc_show |
| 1880 | }; |
| 1881 | |
| 1882 | static int |
| 1883 | ioc_open(struct inode *inode, struct file *file) |
| 1884 | { |
| 1885 | return seq_open(file, &ioc_seq_ops); |
| 1886 | } |
| 1887 | |
| 1888 | static struct file_operations ioc_fops = { |
| 1889 | .open = ioc_open, |
| 1890 | .read = seq_read, |
| 1891 | .llseek = seq_lseek, |
| 1892 | .release = seq_release |
| 1893 | }; |
| 1894 | |
| 1895 | static void __init |
| 1896 | ioc_proc_init(void) |
| 1897 | { |
| 1898 | struct proc_dir_entry *dir, *entry; |
| 1899 | |
| 1900 | dir = proc_mkdir("bus/mckinley", NULL); |
| 1901 | if (!dir) |
| 1902 | return; |
| 1903 | |
| 1904 | entry = create_proc_entry(ioc_list->name, 0, dir); |
| 1905 | if (entry) |
| 1906 | entry->proc_fops = &ioc_fops; |
| 1907 | } |
| 1908 | #endif |
| 1909 | |
| 1910 | static void |
| 1911 | sba_connect_bus(struct pci_bus *bus) |
| 1912 | { |
| 1913 | acpi_handle handle, parent; |
| 1914 | acpi_status status; |
| 1915 | struct ioc *ioc; |
| 1916 | |
| 1917 | if (!PCI_CONTROLLER(bus)) |
| 1918 | panic(PFX "no sysdata on bus %d!\n", bus->number); |
| 1919 | |
| 1920 | if (PCI_CONTROLLER(bus)->iommu) |
| 1921 | return; |
| 1922 | |
| 1923 | handle = PCI_CONTROLLER(bus)->acpi_handle; |
| 1924 | if (!handle) |
| 1925 | return; |
| 1926 | |
| 1927 | /* |
| 1928 | * The IOC scope encloses PCI root bridges in the ACPI |
| 1929 | * namespace, so work our way out until we find an IOC we |
| 1930 | * claimed previously. |
| 1931 | */ |
| 1932 | do { |
| 1933 | for (ioc = ioc_list; ioc; ioc = ioc->next) |
| 1934 | if (ioc->handle == handle) { |
| 1935 | PCI_CONTROLLER(bus)->iommu = ioc; |
| 1936 | return; |
| 1937 | } |
| 1938 | |
| 1939 | status = acpi_get_parent(handle, &parent); |
| 1940 | handle = parent; |
| 1941 | } while (ACPI_SUCCESS(status)); |
| 1942 | |
| 1943 | printk(KERN_WARNING "No IOC for PCI Bus %04x:%02x in ACPI\n", pci_domain_nr(bus), bus->number); |
| 1944 | } |
| 1945 | |
| 1946 | #ifdef CONFIG_NUMA |
| 1947 | static void __init |
| 1948 | sba_map_ioc_to_node(struct ioc *ioc, acpi_handle handle) |
| 1949 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1950 | unsigned int node; |
Alex Williamson | bb0fc08 | 2005-03-24 22:58:00 -0700 | [diff] [blame] | 1951 | int pxm; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1952 | |
| 1953 | ioc->node = MAX_NUMNODES; |
| 1954 | |
Alex Williamson | bb0fc08 | 2005-03-24 22:58:00 -0700 | [diff] [blame] | 1955 | pxm = acpi_get_pxm(handle); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1956 | |
Alex Williamson | bb0fc08 | 2005-03-24 22:58:00 -0700 | [diff] [blame] | 1957 | if (pxm < 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1958 | return; |
| 1959 | |
Yasunori Goto | 762834e | 2006-06-23 02:03:19 -0700 | [diff] [blame] | 1960 | node = pxm_to_node(pxm); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1961 | |
| 1962 | if (node >= MAX_NUMNODES || !node_online(node)) |
| 1963 | return; |
| 1964 | |
| 1965 | ioc->node = node; |
| 1966 | return; |
| 1967 | } |
| 1968 | #else |
| 1969 | #define sba_map_ioc_to_node(ioc, handle) |
| 1970 | #endif |
| 1971 | |
| 1972 | static int __init |
| 1973 | acpi_sba_ioc_add(struct acpi_device *device) |
| 1974 | { |
| 1975 | struct ioc *ioc; |
| 1976 | acpi_status status; |
| 1977 | u64 hpa, length; |
| 1978 | struct acpi_buffer buffer; |
| 1979 | struct acpi_device_info *dev_info; |
| 1980 | |
| 1981 | status = hp_acpi_csr_space(device->handle, &hpa, &length); |
| 1982 | if (ACPI_FAILURE(status)) |
| 1983 | return 1; |
| 1984 | |
| 1985 | buffer.length = ACPI_ALLOCATE_LOCAL_BUFFER; |
| 1986 | status = acpi_get_object_info(device->handle, &buffer); |
| 1987 | if (ACPI_FAILURE(status)) |
| 1988 | return 1; |
| 1989 | dev_info = buffer.pointer; |
| 1990 | |
| 1991 | /* |
| 1992 | * For HWP0001, only SBA appears in ACPI namespace. It encloses the PCI |
| 1993 | * root bridges, and its CSR space includes the IOC function. |
| 1994 | */ |
| 1995 | if (strncmp("HWP0001", dev_info->hardware_id.value, 7) == 0) { |
| 1996 | hpa += ZX1_IOC_OFFSET; |
| 1997 | /* zx1 based systems default to kernel page size iommu pages */ |
| 1998 | if (!iovp_shift) |
| 1999 | iovp_shift = min(PAGE_SHIFT, 16); |
| 2000 | } |
Len Brown | 144c87b | 2006-04-02 00:15:39 -0500 | [diff] [blame] | 2001 | kfree(dev_info); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2002 | |
| 2003 | /* |
| 2004 | * default anything not caught above or specified on cmdline to 4k |
| 2005 | * iommu page size |
| 2006 | */ |
| 2007 | if (!iovp_shift) |
| 2008 | iovp_shift = 12; |
| 2009 | |
| 2010 | ioc = ioc_init(hpa, device->handle); |
| 2011 | if (!ioc) |
| 2012 | return 1; |
| 2013 | |
| 2014 | /* setup NUMA node association */ |
| 2015 | sba_map_ioc_to_node(ioc, device->handle); |
| 2016 | return 0; |
| 2017 | } |
| 2018 | |
| 2019 | static struct acpi_driver acpi_sba_ioc_driver = { |
| 2020 | .name = "IOC IOMMU Driver", |
| 2021 | .ids = "HWP0001,HWP0004", |
| 2022 | .ops = { |
| 2023 | .add = acpi_sba_ioc_add, |
| 2024 | }, |
| 2025 | }; |
| 2026 | |
| 2027 | static int __init |
| 2028 | sba_init(void) |
| 2029 | { |
Alex Williamson | 0b9afed | 2005-09-06 11:20:49 -0600 | [diff] [blame] | 2030 | if (!ia64_platform_is("hpzx1") && !ia64_platform_is("hpzx1_swiotlb")) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2031 | return 0; |
| 2032 | |
Alex Williamson | 0b9afed | 2005-09-06 11:20:49 -0600 | [diff] [blame] | 2033 | acpi_bus_register_driver(&acpi_sba_ioc_driver); |
| 2034 | if (!ioc_list) { |
| 2035 | #ifdef CONFIG_IA64_GENERIC |
| 2036 | extern int swiotlb_late_init_with_default_size (size_t size); |
| 2037 | |
| 2038 | /* |
| 2039 | * If we didn't find something sba_iommu can claim, we |
| 2040 | * need to setup the swiotlb and switch to the dig machvec. |
| 2041 | */ |
| 2042 | if (swiotlb_late_init_with_default_size(64 * (1<<20)) != 0) |
| 2043 | panic("Unable to find SBA IOMMU or initialize " |
| 2044 | "software I/O TLB: Try machvec=dig boot option"); |
| 2045 | machvec_init("dig"); |
| 2046 | #else |
| 2047 | panic("Unable to find SBA IOMMU: Try a generic or DIG kernel"); |
| 2048 | #endif |
| 2049 | return 0; |
| 2050 | } |
| 2051 | |
| 2052 | #if defined(CONFIG_IA64_GENERIC) || defined(CONFIG_IA64_HP_ZX1_SWIOTLB) |
| 2053 | /* |
| 2054 | * hpzx1_swiotlb needs to have a fairly small swiotlb bounce |
| 2055 | * buffer setup to support devices with smaller DMA masks than |
| 2056 | * sba_iommu can handle. |
| 2057 | */ |
| 2058 | if (ia64_platform_is("hpzx1_swiotlb")) { |
| 2059 | extern void hwsw_init(void); |
| 2060 | |
| 2061 | hwsw_init(); |
| 2062 | } |
| 2063 | #endif |
| 2064 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2065 | #ifdef CONFIG_PCI |
| 2066 | { |
| 2067 | struct pci_bus *b = NULL; |
| 2068 | while ((b = pci_find_next_bus(b)) != NULL) |
| 2069 | sba_connect_bus(b); |
| 2070 | } |
| 2071 | #endif |
| 2072 | |
| 2073 | #ifdef CONFIG_PROC_FS |
| 2074 | ioc_proc_init(); |
| 2075 | #endif |
| 2076 | return 0; |
| 2077 | } |
| 2078 | |
| 2079 | subsys_initcall(sba_init); /* must be initialized after ACPI etc., but before any drivers... */ |
| 2080 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2081 | static int __init |
| 2082 | nosbagart(char *str) |
| 2083 | { |
| 2084 | reserve_sba_gart = 0; |
| 2085 | return 1; |
| 2086 | } |
| 2087 | |
| 2088 | int |
| 2089 | sba_dma_supported (struct device *dev, u64 mask) |
| 2090 | { |
| 2091 | /* make sure it's at least 32bit capable */ |
| 2092 | return ((mask & 0xFFFFFFFFUL) == 0xFFFFFFFFUL); |
| 2093 | } |
| 2094 | |
| 2095 | int |
| 2096 | sba_dma_mapping_error (dma_addr_t dma_addr) |
| 2097 | { |
| 2098 | return 0; |
| 2099 | } |
| 2100 | |
| 2101 | __setup("nosbagart", nosbagart); |
| 2102 | |
| 2103 | static int __init |
| 2104 | sba_page_override(char *str) |
| 2105 | { |
| 2106 | unsigned long page_size; |
| 2107 | |
| 2108 | page_size = memparse(str, &str); |
| 2109 | switch (page_size) { |
| 2110 | case 4096: |
| 2111 | case 8192: |
| 2112 | case 16384: |
| 2113 | case 65536: |
| 2114 | iovp_shift = ffs(page_size) - 1; |
| 2115 | break; |
| 2116 | default: |
| 2117 | printk("%s: unknown/unsupported iommu page size %ld\n", |
| 2118 | __FUNCTION__, page_size); |
| 2119 | } |
| 2120 | |
| 2121 | return 1; |
| 2122 | } |
| 2123 | |
| 2124 | __setup("sbapagesize=",sba_page_override); |
| 2125 | |
| 2126 | EXPORT_SYMBOL(sba_dma_mapping_error); |
| 2127 | EXPORT_SYMBOL(sba_map_single); |
| 2128 | EXPORT_SYMBOL(sba_unmap_single); |
| 2129 | EXPORT_SYMBOL(sba_map_sg); |
| 2130 | EXPORT_SYMBOL(sba_unmap_sg); |
| 2131 | EXPORT_SYMBOL(sba_dma_supported); |
| 2132 | EXPORT_SYMBOL(sba_alloc_coherent); |
| 2133 | EXPORT_SYMBOL(sba_free_coherent); |