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 <linux/kmemleak.h> 42 #include <asm/io.h> 43 #include <asm/prom.h> 44 #include <asm/iommu.h> 45 #include <asm/pci-bridge.h> 46 #include <asm/machdep.h> 47 #include <asm/cacheflush.h> 48 #include <asm/ppc-pci.h> 49 50 #include "dart.h" 51 52 /* DART table address and size */ 53 static u32 *dart_tablebase; 54 static unsigned long dart_tablesize; 55 56 /* Mapped base address for the dart */ 57 static unsigned int __iomem *dart; 58 59 /* Dummy val that entries are set to when unused */ 60 static unsigned int dart_emptyval; 61 62 static struct iommu_table iommu_table_dart; 63 static int iommu_table_dart_inited; 64 static int dart_dirty; 65 static int dart_is_u4; 66 67 #define DART_U4_BYPASS_BASE 0x8000000000ull 68 69 #define DBG(...) 70 71 static DEFINE_SPINLOCK(invalidate_lock); 72 73 static inline void dart_tlb_invalidate_all(void) 74 { 75 unsigned long l = 0; 76 unsigned int reg, inv_bit; 77 unsigned long limit; 78 unsigned long flags; 79 80 spin_lock_irqsave(&invalidate_lock, flags); 81 82 DBG("dart: flush\n"); 83 84 /* To invalidate the DART, set the DARTCNTL_FLUSHTLB bit in the 85 * control register and wait for it to clear. 86 * 87 * Gotcha: Sometimes, the DART won't detect that the bit gets 88 * set. If so, clear it and set it again. 89 */ 90 91 limit = 0; 92 93 inv_bit = dart_is_u4 ? DART_CNTL_U4_FLUSHTLB : DART_CNTL_U3_FLUSHTLB; 94 retry: 95 l = 0; 96 reg = DART_IN(DART_CNTL); 97 reg |= inv_bit; 98 DART_OUT(DART_CNTL, reg); 99 100 while ((DART_IN(DART_CNTL) & inv_bit) && l < (1L << limit)) 101 l++; 102 if (l == (1L << limit)) { 103 if (limit < 4) { 104 limit++; 105 reg = DART_IN(DART_CNTL); 106 reg &= ~inv_bit; 107 DART_OUT(DART_CNTL, reg); 108 goto retry; 109 } else 110 panic("DART: TLB did not flush after waiting a long " 111 "time. Buggy U3 ?"); 112 } 113 114 spin_unlock_irqrestore(&invalidate_lock, flags); 115 } 116 117 static inline void dart_tlb_invalidate_one(unsigned long bus_rpn) 118 { 119 unsigned int reg; 120 unsigned int l, limit; 121 unsigned long flags; 122 123 spin_lock_irqsave(&invalidate_lock, flags); 124 125 reg = DART_CNTL_U4_ENABLE | DART_CNTL_U4_IONE | 126 (bus_rpn & DART_CNTL_U4_IONE_MASK); 127 DART_OUT(DART_CNTL, reg); 128 129 limit = 0; 130 wait_more: 131 l = 0; 132 while ((DART_IN(DART_CNTL) & DART_CNTL_U4_IONE) && l < (1L << limit)) { 133 rmb(); 134 l++; 135 } 136 137 if (l == (1L << limit)) { 138 if (limit < 4) { 139 limit++; 140 goto wait_more; 141 } else 142 panic("DART: TLB did not flush after waiting a long " 143 "time. Buggy U4 ?"); 144 } 145 146 spin_unlock_irqrestore(&invalidate_lock, flags); 147 } 148 149 static void dart_cache_sync(unsigned int *base, unsigned int count) 150 { 151 /* 152 * We add 1 to the number of entries to flush, following a 153 * comment in Darwin indicating that the memory controller 154 * can prefetch unmapped memory under some circumstances. 155 */ 156 unsigned long start = (unsigned long)base; 157 unsigned long end = start + (count + 1) * sizeof(unsigned int); 158 unsigned int tmp; 159 160 /* Perform a standard cache flush */ 161 flush_inval_dcache_range(start, end); 162 163 /* 164 * Perform the sequence described in the CPC925 manual to 165 * ensure all the data gets to a point the cache incoherent 166 * DART hardware will see. 167 */ 168 asm volatile(" sync;" 169 " isync;" 170 " dcbf 0,%1;" 171 " sync;" 172 " isync;" 173 " lwz %0,0(%1);" 174 " isync" : "=r" (tmp) : "r" (end) : "memory"); 175 } 176 177 static void dart_flush(struct iommu_table *tbl) 178 { 179 mb(); 180 if (dart_dirty) { 181 dart_tlb_invalidate_all(); 182 dart_dirty = 0; 183 } 184 } 185 186 static int dart_build(struct iommu_table *tbl, long index, 187 long npages, unsigned long uaddr, 188 enum dma_data_direction direction, 189 unsigned long attrs) 190 { 191 unsigned int *dp, *orig_dp; 192 unsigned int rpn; 193 long l; 194 195 DBG("dart: build at: %lx, %lx, addr: %x\n", index, npages, uaddr); 196 197 orig_dp = dp = ((unsigned int*)tbl->it_base) + index; 198 199 /* On U3, all memory is contiguous, so we can move this 200 * out of the loop. 201 */ 202 l = npages; 203 while (l--) { 204 rpn = __pa(uaddr) >> DART_PAGE_SHIFT; 205 206 *(dp++) = DARTMAP_VALID | (rpn & DARTMAP_RPNMASK); 207 208 uaddr += DART_PAGE_SIZE; 209 } 210 dart_cache_sync(orig_dp, npages); 211 212 if (dart_is_u4) { 213 rpn = index; 214 while (npages--) 215 dart_tlb_invalidate_one(rpn++); 216 } else { 217 dart_dirty = 1; 218 } 219 return 0; 220 } 221 222 223 static void dart_free(struct iommu_table *tbl, long index, long npages) 224 { 225 unsigned int *dp, *orig_dp; 226 long orig_npages = npages; 227 228 /* We don't worry about flushing the TLB cache. The only drawback of 229 * not doing it is that we won't catch buggy device drivers doing 230 * bad DMAs, but then no 32-bit architecture ever does either. 231 */ 232 233 DBG("dart: free at: %lx, %lx\n", index, npages); 234 235 orig_dp = dp = ((unsigned int *)tbl->it_base) + index; 236 237 while (npages--) 238 *(dp++) = dart_emptyval; 239 240 dart_cache_sync(orig_dp, orig_npages); 241 } 242 243 static void allocate_dart(void) 244 { 245 unsigned long tmp; 246 247 /* 512 pages (2MB) is max DART tablesize. */ 248 dart_tablesize = 1UL << 21; 249 250 /* 251 * 16MB (1 << 24) alignment. We allocate a full 16Mb chuck since we 252 * will blow up an entire large page anyway in the kernel mapping. 253 */ 254 dart_tablebase = __va(memblock_alloc_base(1UL<<24, 255 1UL<<24, 0x80000000L)); 256 257 /* There is no point scanning the DART space for leaks*/ 258 kmemleak_no_scan((void *)dart_tablebase); 259 260 /* Allocate a spare page to map all invalid DART pages. We need to do 261 * that to work around what looks like a problem with the HT bridge 262 * prefetching into invalid pages and corrupting data 263 */ 264 tmp = memblock_alloc(DART_PAGE_SIZE, DART_PAGE_SIZE); 265 dart_emptyval = DARTMAP_VALID | ((tmp >> DART_PAGE_SHIFT) & 266 DARTMAP_RPNMASK); 267 268 printk(KERN_INFO "DART table allocated at: %p\n", dart_tablebase); 269 } 270 271 static int __init dart_init(struct device_node *dart_node) 272 { 273 unsigned int i; 274 unsigned long base, size; 275 struct resource r; 276 277 /* IOMMU disabled by the user ? bail out */ 278 if (iommu_is_off) 279 return -ENODEV; 280 281 /* 282 * Only use the DART if the machine has more than 1GB of RAM 283 * or if requested with iommu=on on cmdline. 284 * 285 * 1GB of RAM is picked as limit because some default devices 286 * (i.e. Airport Extreme) have 30 bit address range limits. 287 */ 288 289 if (!iommu_force_on && memblock_end_of_DRAM() <= 0x40000000ull) 290 return -ENODEV; 291 292 /* Get DART registers */ 293 if (of_address_to_resource(dart_node, 0, &r)) 294 panic("DART: can't get register base ! "); 295 296 /* Map in DART registers */ 297 dart = ioremap(r.start, resource_size(&r)); 298 if (dart == NULL) 299 panic("DART: Cannot map registers!"); 300 301 /* Allocate the DART and dummy page */ 302 allocate_dart(); 303 304 /* Fill initial table */ 305 for (i = 0; i < dart_tablesize/4; i++) 306 dart_tablebase[i] = dart_emptyval; 307 308 /* Push to memory */ 309 dart_cache_sync(dart_tablebase, dart_tablesize / sizeof(u32)); 310 311 /* Initialize DART with table base and enable it. */ 312 base = ((unsigned long)dart_tablebase) >> DART_PAGE_SHIFT; 313 size = dart_tablesize >> DART_PAGE_SHIFT; 314 if (dart_is_u4) { 315 size &= DART_SIZE_U4_SIZE_MASK; 316 DART_OUT(DART_BASE_U4, base); 317 DART_OUT(DART_SIZE_U4, size); 318 DART_OUT(DART_CNTL, DART_CNTL_U4_ENABLE); 319 } else { 320 size &= DART_CNTL_U3_SIZE_MASK; 321 DART_OUT(DART_CNTL, 322 DART_CNTL_U3_ENABLE | 323 (base << DART_CNTL_U3_BASE_SHIFT) | 324 (size << DART_CNTL_U3_SIZE_SHIFT)); 325 } 326 327 /* Invalidate DART to get rid of possible stale TLBs */ 328 dart_tlb_invalidate_all(); 329 330 printk(KERN_INFO "DART IOMMU initialized for %s type chipset\n", 331 dart_is_u4 ? "U4" : "U3"); 332 333 return 0; 334 } 335 336 static struct iommu_table_ops iommu_dart_ops = { 337 .set = dart_build, 338 .clear = dart_free, 339 .flush = dart_flush, 340 }; 341 342 static void iommu_table_dart_setup(void) 343 { 344 iommu_table_dart.it_busno = 0; 345 iommu_table_dart.it_offset = 0; 346 /* it_size is in number of entries */ 347 iommu_table_dart.it_size = dart_tablesize / sizeof(u32); 348 iommu_table_dart.it_page_shift = IOMMU_PAGE_SHIFT_4K; 349 350 /* Initialize the common IOMMU code */ 351 iommu_table_dart.it_base = (unsigned long)dart_tablebase; 352 iommu_table_dart.it_index = 0; 353 iommu_table_dart.it_blocksize = 1; 354 iommu_table_dart.it_ops = &iommu_dart_ops; 355 iommu_init_table(&iommu_table_dart, -1); 356 357 /* Reserve the last page of the DART to avoid possible prefetch 358 * past the DART mapped area 359 */ 360 set_bit(iommu_table_dart.it_size - 1, iommu_table_dart.it_map); 361 } 362 363 static void pci_dma_dev_setup_dart(struct pci_dev *dev) 364 { 365 if (dart_is_u4) 366 set_dma_offset(&dev->dev, DART_U4_BYPASS_BASE); 367 set_iommu_table_base(&dev->dev, &iommu_table_dart); 368 } 369 370 static void pci_dma_bus_setup_dart(struct pci_bus *bus) 371 { 372 if (!iommu_table_dart_inited) { 373 iommu_table_dart_inited = 1; 374 iommu_table_dart_setup(); 375 } 376 } 377 378 static bool dart_device_on_pcie(struct device *dev) 379 { 380 struct device_node *np = of_node_get(dev->of_node); 381 382 while(np) { 383 if (of_device_is_compatible(np, "U4-pcie") || 384 of_device_is_compatible(np, "u4-pcie")) { 385 of_node_put(np); 386 return true; 387 } 388 np = of_get_next_parent(np); 389 } 390 return false; 391 } 392 393 static int dart_dma_set_mask(struct device *dev, u64 dma_mask) 394 { 395 if (!dev->dma_mask || !dma_supported(dev, dma_mask)) 396 return -EIO; 397 398 /* U4 supports a DART bypass, we use it for 64-bit capable 399 * devices to improve performances. However, that only works 400 * for devices connected to U4 own PCIe interface, not bridged 401 * through hypertransport. We need the device to support at 402 * least 40 bits of addresses. 403 */ 404 if (dart_device_on_pcie(dev) && dma_mask >= DMA_BIT_MASK(40)) { 405 dev_info(dev, "Using 64-bit DMA iommu bypass\n"); 406 set_dma_ops(dev, &dma_nommu_ops); 407 } else { 408 dev_info(dev, "Using 32-bit DMA via iommu\n"); 409 set_dma_ops(dev, &dma_iommu_ops); 410 } 411 412 *dev->dma_mask = dma_mask; 413 return 0; 414 } 415 416 void __init iommu_init_early_dart(struct pci_controller_ops *controller_ops) 417 { 418 struct device_node *dn; 419 420 /* Find the DART in the device-tree */ 421 dn = of_find_compatible_node(NULL, "dart", "u3-dart"); 422 if (dn == NULL) { 423 dn = of_find_compatible_node(NULL, "dart", "u4-dart"); 424 if (dn == NULL) 425 return; /* use default direct_dma_ops */ 426 dart_is_u4 = 1; 427 } 428 429 /* Initialize the DART HW */ 430 if (dart_init(dn) != 0) 431 goto bail; 432 433 /* Setup bypass if supported */ 434 if (dart_is_u4) 435 ppc_md.dma_set_mask = dart_dma_set_mask; 436 437 controller_ops->dma_dev_setup = pci_dma_dev_setup_dart; 438 controller_ops->dma_bus_setup = pci_dma_bus_setup_dart; 439 440 /* Setup pci_dma ops */ 441 set_pci_dma_ops(&dma_iommu_ops); 442 return; 443 444 bail: 445 /* If init failed, use direct iommu and null setup functions */ 446 controller_ops->dma_dev_setup = NULL; 447 controller_ops->dma_bus_setup = NULL; 448 449 /* Setup pci_dma ops */ 450 set_pci_dma_ops(&dma_nommu_ops); 451 } 452 453 #ifdef CONFIG_PM 454 static void iommu_dart_restore(void) 455 { 456 dart_cache_sync(dart_tablebase, dart_tablesize / sizeof(u32)); 457 dart_tlb_invalidate_all(); 458 } 459 460 static int __init iommu_init_late_dart(void) 461 { 462 if (!dart_tablebase) 463 return 0; 464 465 ppc_md.iommu_restore = iommu_dart_restore; 466 467 return 0; 468 } 469 470 late_initcall(iommu_init_late_dart); 471 #endif /* CONFIG_PM */ 472