1 /* 2 * arch/powerpc/sysdev/dart_iommu.c 3 * 4 * Copyright (C) 2004 Olof Johansson <olof@lixom.net>, IBM Corporation 5 * Copyright (C) 2005 Benjamin Herrenschmidt <benh@kernel.crashing.org>, 6 * IBM Corporation 7 * 8 * Based on pSeries_iommu.c: 9 * Copyright (C) 2001 Mike Corrigan & Dave Engebretsen, IBM Corporation 10 * Copyright (C) 2004 Olof Johansson <olof@lixom.net>, IBM Corporation 11 * 12 * Dynamic DMA mapping support, Apple U3, U4 & IBM CPC925 "DART" iommu. 13 * 14 * 15 * This program is free software; you can redistribute it and/or modify 16 * it under the terms of the GNU General Public License as published by 17 * the Free Software Foundation; either version 2 of the License, or 18 * (at your option) any later version. 19 * 20 * This program is distributed in the hope that it will be useful, 21 * but WITHOUT ANY WARRANTY; without even the implied warranty of 22 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 23 * GNU General Public License for more details. 24 * 25 * You should have received a copy of the GNU General Public License 26 * along with this program; if not, write to the Free Software 27 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA 28 */ 29 30 #include <linux/init.h> 31 #include <linux/types.h> 32 #include <linux/mm.h> 33 #include <linux/spinlock.h> 34 #include <linux/string.h> 35 #include <linux/pci.h> 36 #include <linux/dma-mapping.h> 37 #include <linux/vmalloc.h> 38 #include <linux/suspend.h> 39 #include <linux/memblock.h> 40 #include <linux/gfp.h> 41 #include <asm/io.h> 42 #include <asm/prom.h> 43 #include <asm/iommu.h> 44 #include <asm/pci-bridge.h> 45 #include <asm/machdep.h> 46 #include <asm/cacheflush.h> 47 #include <asm/ppc-pci.h> 48 49 #include "dart.h" 50 51 /* DART table address and size */ 52 static u32 *dart_tablebase; 53 static unsigned long dart_tablesize; 54 55 /* Mapped base address for the dart */ 56 static unsigned int __iomem *dart; 57 58 /* Dummy val that entries are set to when unused */ 59 static unsigned int dart_emptyval; 60 61 static struct iommu_table iommu_table_dart; 62 static int iommu_table_dart_inited; 63 static int dart_dirty; 64 static int dart_is_u4; 65 66 #define DART_U4_BYPASS_BASE 0x8000000000ull 67 68 #define DBG(...) 69 70 static DEFINE_SPINLOCK(invalidate_lock); 71 72 static inline void dart_tlb_invalidate_all(void) 73 { 74 unsigned long l = 0; 75 unsigned int reg, inv_bit; 76 unsigned long limit; 77 unsigned long flags; 78 79 spin_lock_irqsave(&invalidate_lock, flags); 80 81 DBG("dart: flush\n"); 82 83 /* To invalidate the DART, set the DARTCNTL_FLUSHTLB bit in the 84 * control register and wait for it to clear. 85 * 86 * Gotcha: Sometimes, the DART won't detect that the bit gets 87 * set. If so, clear it and set it again. 88 */ 89 90 limit = 0; 91 92 inv_bit = dart_is_u4 ? DART_CNTL_U4_FLUSHTLB : DART_CNTL_U3_FLUSHTLB; 93 retry: 94 l = 0; 95 reg = DART_IN(DART_CNTL); 96 reg |= inv_bit; 97 DART_OUT(DART_CNTL, reg); 98 99 while ((DART_IN(DART_CNTL) & inv_bit) && l < (1L << limit)) 100 l++; 101 if (l == (1L << limit)) { 102 if (limit < 4) { 103 limit++; 104 reg = DART_IN(DART_CNTL); 105 reg &= ~inv_bit; 106 DART_OUT(DART_CNTL, reg); 107 goto retry; 108 } else 109 panic("DART: TLB did not flush after waiting a long " 110 "time. Buggy U3 ?"); 111 } 112 113 spin_unlock_irqrestore(&invalidate_lock, flags); 114 } 115 116 static inline void dart_tlb_invalidate_one(unsigned long bus_rpn) 117 { 118 unsigned int reg; 119 unsigned int l, limit; 120 unsigned long flags; 121 122 spin_lock_irqsave(&invalidate_lock, flags); 123 124 reg = DART_CNTL_U4_ENABLE | DART_CNTL_U4_IONE | 125 (bus_rpn & DART_CNTL_U4_IONE_MASK); 126 DART_OUT(DART_CNTL, reg); 127 128 limit = 0; 129 wait_more: 130 l = 0; 131 while ((DART_IN(DART_CNTL) & DART_CNTL_U4_IONE) && l < (1L << limit)) { 132 rmb(); 133 l++; 134 } 135 136 if (l == (1L << limit)) { 137 if (limit < 4) { 138 limit++; 139 goto wait_more; 140 } else 141 panic("DART: TLB did not flush after waiting a long " 142 "time. Buggy U4 ?"); 143 } 144 145 spin_unlock_irqrestore(&invalidate_lock, flags); 146 } 147 148 static void dart_cache_sync(unsigned int *base, unsigned int count) 149 { 150 /* 151 * We add 1 to the number of entries to flush, following a 152 * comment in Darwin indicating that the memory controller 153 * can prefetch unmapped memory under some circumstances. 154 */ 155 unsigned long start = (unsigned long)base; 156 unsigned long end = start + (count + 1) * sizeof(unsigned int); 157 unsigned int tmp; 158 159 /* Perform a standard cache flush */ 160 flush_inval_dcache_range(start, end); 161 162 /* 163 * Perform the sequence described in the CPC925 manual to 164 * ensure all the data gets to a point the cache incoherent 165 * DART hardware will see. 166 */ 167 asm volatile(" sync;" 168 " isync;" 169 " dcbf 0,%1;" 170 " sync;" 171 " isync;" 172 " lwz %0,0(%1);" 173 " isync" : "=r" (tmp) : "r" (end) : "memory"); 174 } 175 176 static void dart_flush(struct iommu_table *tbl) 177 { 178 mb(); 179 if (dart_dirty) { 180 dart_tlb_invalidate_all(); 181 dart_dirty = 0; 182 } 183 } 184 185 static int dart_build(struct iommu_table *tbl, long index, 186 long npages, unsigned long uaddr, 187 enum dma_data_direction direction, 188 unsigned long attrs) 189 { 190 unsigned int *dp, *orig_dp; 191 unsigned int rpn; 192 long l; 193 194 DBG("dart: build at: %lx, %lx, addr: %x\n", index, npages, uaddr); 195 196 orig_dp = dp = ((unsigned int*)tbl->it_base) + index; 197 198 /* On U3, all memory is contiguous, so we can move this 199 * out of the loop. 200 */ 201 l = npages; 202 while (l--) { 203 rpn = __pa(uaddr) >> DART_PAGE_SHIFT; 204 205 *(dp++) = DARTMAP_VALID | (rpn & DARTMAP_RPNMASK); 206 207 uaddr += DART_PAGE_SIZE; 208 } 209 dart_cache_sync(orig_dp, npages); 210 211 if (dart_is_u4) { 212 rpn = index; 213 while (npages--) 214 dart_tlb_invalidate_one(rpn++); 215 } else { 216 dart_dirty = 1; 217 } 218 return 0; 219 } 220 221 222 static void dart_free(struct iommu_table *tbl, long index, long npages) 223 { 224 unsigned int *dp, *orig_dp; 225 long orig_npages = npages; 226 227 /* We don't worry about flushing the TLB cache. The only drawback of 228 * not doing it is that we won't catch buggy device drivers doing 229 * bad DMAs, but then no 32-bit architecture ever does either. 230 */ 231 232 DBG("dart: free at: %lx, %lx\n", index, npages); 233 234 orig_dp = dp = ((unsigned int *)tbl->it_base) + index; 235 236 while (npages--) 237 *(dp++) = dart_emptyval; 238 239 dart_cache_sync(orig_dp, orig_npages); 240 } 241 242 static void allocate_dart(void) 243 { 244 unsigned long tmp; 245 246 /* 512 pages (2MB) is max DART tablesize. */ 247 dart_tablesize = 1UL << 21; 248 249 /* 250 * 16MB (1 << 24) alignment. We allocate a full 16Mb chuck since we 251 * will blow up an entire large page anyway in the kernel mapping. 252 */ 253 dart_tablebase = __va(memblock_alloc_base(1UL<<24, 254 1UL<<24, 0x80000000L)); 255 256 /* There is no point scanning the DART space for leaks*/ 257 kmemleak_no_scan((void *)dart_tablebase); 258 259 /* Allocate a spare page to map all invalid DART pages. We need to do 260 * that to work around what looks like a problem with the HT bridge 261 * prefetching into invalid pages and corrupting data 262 */ 263 tmp = memblock_alloc(DART_PAGE_SIZE, DART_PAGE_SIZE); 264 dart_emptyval = DARTMAP_VALID | ((tmp >> DART_PAGE_SHIFT) & 265 DARTMAP_RPNMASK); 266 267 printk(KERN_INFO "DART table allocated at: %p\n", dart_tablebase); 268 } 269 270 static int __init dart_init(struct device_node *dart_node) 271 { 272 unsigned int i; 273 unsigned long base, size; 274 struct resource r; 275 276 /* IOMMU disabled by the user ? bail out */ 277 if (iommu_is_off) 278 return -ENODEV; 279 280 /* 281 * Only use the DART if the machine has more than 1GB of RAM 282 * or if requested with iommu=on on cmdline. 283 * 284 * 1GB of RAM is picked as limit because some default devices 285 * (i.e. Airport Extreme) have 30 bit address range limits. 286 */ 287 288 if (!iommu_force_on && memblock_end_of_DRAM() <= 0x40000000ull) 289 return -ENODEV; 290 291 /* Get DART registers */ 292 if (of_address_to_resource(dart_node, 0, &r)) 293 panic("DART: can't get register base ! "); 294 295 /* Map in DART registers */ 296 dart = ioremap(r.start, resource_size(&r)); 297 if (dart == NULL) 298 panic("DART: Cannot map registers!"); 299 300 /* Allocate the DART and dummy page */ 301 allocate_dart(); 302 303 /* Fill initial table */ 304 for (i = 0; i < dart_tablesize/4; i++) 305 dart_tablebase[i] = dart_emptyval; 306 307 /* Push to memory */ 308 dart_cache_sync(dart_tablebase, dart_tablesize / sizeof(u32)); 309 310 /* Initialize DART with table base and enable it. */ 311 base = ((unsigned long)dart_tablebase) >> DART_PAGE_SHIFT; 312 size = dart_tablesize >> DART_PAGE_SHIFT; 313 if (dart_is_u4) { 314 size &= DART_SIZE_U4_SIZE_MASK; 315 DART_OUT(DART_BASE_U4, base); 316 DART_OUT(DART_SIZE_U4, size); 317 DART_OUT(DART_CNTL, DART_CNTL_U4_ENABLE); 318 } else { 319 size &= DART_CNTL_U3_SIZE_MASK; 320 DART_OUT(DART_CNTL, 321 DART_CNTL_U3_ENABLE | 322 (base << DART_CNTL_U3_BASE_SHIFT) | 323 (size << DART_CNTL_U3_SIZE_SHIFT)); 324 } 325 326 /* Invalidate DART to get rid of possible stale TLBs */ 327 dart_tlb_invalidate_all(); 328 329 printk(KERN_INFO "DART IOMMU initialized for %s type chipset\n", 330 dart_is_u4 ? "U4" : "U3"); 331 332 return 0; 333 } 334 335 static struct iommu_table_ops iommu_dart_ops = { 336 .set = dart_build, 337 .clear = dart_free, 338 .flush = dart_flush, 339 }; 340 341 static void iommu_table_dart_setup(void) 342 { 343 iommu_table_dart.it_busno = 0; 344 iommu_table_dart.it_offset = 0; 345 /* it_size is in number of entries */ 346 iommu_table_dart.it_size = dart_tablesize / sizeof(u32); 347 iommu_table_dart.it_page_shift = IOMMU_PAGE_SHIFT_4K; 348 349 /* Initialize the common IOMMU code */ 350 iommu_table_dart.it_base = (unsigned long)dart_tablebase; 351 iommu_table_dart.it_index = 0; 352 iommu_table_dart.it_blocksize = 1; 353 iommu_table_dart.it_ops = &iommu_dart_ops; 354 iommu_init_table(&iommu_table_dart, -1); 355 356 /* Reserve the last page of the DART to avoid possible prefetch 357 * past the DART mapped area 358 */ 359 set_bit(iommu_table_dart.it_size - 1, iommu_table_dart.it_map); 360 } 361 362 static void pci_dma_dev_setup_dart(struct pci_dev *dev) 363 { 364 if (dart_is_u4) 365 set_dma_offset(&dev->dev, DART_U4_BYPASS_BASE); 366 set_iommu_table_base(&dev->dev, &iommu_table_dart); 367 } 368 369 static void pci_dma_bus_setup_dart(struct pci_bus *bus) 370 { 371 if (!iommu_table_dart_inited) { 372 iommu_table_dart_inited = 1; 373 iommu_table_dart_setup(); 374 } 375 } 376 377 static bool dart_device_on_pcie(struct device *dev) 378 { 379 struct device_node *np = of_node_get(dev->of_node); 380 381 while(np) { 382 if (of_device_is_compatible(np, "U4-pcie") || 383 of_device_is_compatible(np, "u4-pcie")) { 384 of_node_put(np); 385 return true; 386 } 387 np = of_get_next_parent(np); 388 } 389 return false; 390 } 391 392 static int dart_dma_set_mask(struct device *dev, u64 dma_mask) 393 { 394 if (!dev->dma_mask || !dma_supported(dev, dma_mask)) 395 return -EIO; 396 397 /* U4 supports a DART bypass, we use it for 64-bit capable 398 * devices to improve performances. However, that only works 399 * for devices connected to U4 own PCIe interface, not bridged 400 * through hypertransport. We need the device to support at 401 * least 40 bits of addresses. 402 */ 403 if (dart_device_on_pcie(dev) && dma_mask >= DMA_BIT_MASK(40)) { 404 dev_info(dev, "Using 64-bit DMA iommu bypass\n"); 405 set_dma_ops(dev, &dma_nommu_ops); 406 } else { 407 dev_info(dev, "Using 32-bit DMA via iommu\n"); 408 set_dma_ops(dev, &dma_iommu_ops); 409 } 410 411 *dev->dma_mask = dma_mask; 412 return 0; 413 } 414 415 void __init iommu_init_early_dart(struct pci_controller_ops *controller_ops) 416 { 417 struct device_node *dn; 418 419 /* Find the DART in the device-tree */ 420 dn = of_find_compatible_node(NULL, "dart", "u3-dart"); 421 if (dn == NULL) { 422 dn = of_find_compatible_node(NULL, "dart", "u4-dart"); 423 if (dn == NULL) 424 return; /* use default direct_dma_ops */ 425 dart_is_u4 = 1; 426 } 427 428 /* Initialize the DART HW */ 429 if (dart_init(dn) != 0) 430 goto bail; 431 432 /* Setup bypass if supported */ 433 if (dart_is_u4) 434 ppc_md.dma_set_mask = dart_dma_set_mask; 435 436 controller_ops->dma_dev_setup = pci_dma_dev_setup_dart; 437 controller_ops->dma_bus_setup = pci_dma_bus_setup_dart; 438 439 /* Setup pci_dma ops */ 440 set_pci_dma_ops(&dma_iommu_ops); 441 return; 442 443 bail: 444 /* If init failed, use direct iommu and null setup functions */ 445 controller_ops->dma_dev_setup = NULL; 446 controller_ops->dma_bus_setup = NULL; 447 448 /* Setup pci_dma ops */ 449 set_pci_dma_ops(&dma_nommu_ops); 450 } 451 452 #ifdef CONFIG_PM 453 static void iommu_dart_restore(void) 454 { 455 dart_cache_sync(dart_tablebase, dart_tablesize / sizeof(u32)); 456 dart_tlb_invalidate_all(); 457 } 458 459 static int __init iommu_init_late_dart(void) 460 { 461 if (!dart_tablebase) 462 return 0; 463 464 ppc_md.iommu_restore = iommu_dart_restore; 465 466 return 0; 467 } 468 469 late_initcall(iommu_init_late_dart); 470 #endif /* CONFIG_PM */ 471