1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * iommu.c: IOMMU specific routines for memory management. 4 * 5 * Copyright (C) 1995 David S. Miller (davem@caip.rutgers.edu) 6 * Copyright (C) 1995,2002 Pete Zaitcev (zaitcev@yahoo.com) 7 * Copyright (C) 1996 Eddie C. Dost (ecd@skynet.be) 8 * Copyright (C) 1997,1998 Jakub Jelinek (jj@sunsite.mff.cuni.cz) 9 */ 10 11 #include <linux/kernel.h> 12 #include <linux/init.h> 13 #include <linux/mm.h> 14 #include <linux/slab.h> 15 #include <linux/highmem.h> /* pte_offset_map => kmap_atomic */ 16 #include <linux/dma-mapping.h> 17 #include <linux/of.h> 18 #include <linux/of_device.h> 19 20 #include <asm/pgalloc.h> 21 #include <asm/pgtable.h> 22 #include <asm/io.h> 23 #include <asm/mxcc.h> 24 #include <asm/mbus.h> 25 #include <asm/cacheflush.h> 26 #include <asm/tlbflush.h> 27 #include <asm/bitext.h> 28 #include <asm/iommu.h> 29 #include <asm/dma.h> 30 31 #include "mm_32.h" 32 33 /* 34 * This can be sized dynamically, but we will do this 35 * only when we have a guidance about actual I/O pressures. 36 */ 37 #define IOMMU_RNGE IOMMU_RNGE_256MB 38 #define IOMMU_START 0xF0000000 39 #define IOMMU_WINSIZE (256*1024*1024U) 40 #define IOMMU_NPTES (IOMMU_WINSIZE/PAGE_SIZE) /* 64K PTEs, 256KB */ 41 #define IOMMU_ORDER 6 /* 4096 * (1<<6) */ 42 43 static int viking_flush; 44 /* viking.S */ 45 extern void viking_flush_page(unsigned long page); 46 extern void viking_mxcc_flush_page(unsigned long page); 47 48 /* 49 * Values precomputed according to CPU type. 50 */ 51 static unsigned int ioperm_noc; /* Consistent mapping iopte flags */ 52 static pgprot_t dvma_prot; /* Consistent mapping pte flags */ 53 54 #define IOPERM (IOPTE_CACHE | IOPTE_WRITE | IOPTE_VALID) 55 #define MKIOPTE(pfn, perm) (((((pfn)<<8) & IOPTE_PAGE) | (perm)) & ~IOPTE_WAZ) 56 57 static void __init sbus_iommu_init(struct platform_device *op) 58 { 59 struct iommu_struct *iommu; 60 unsigned int impl, vers; 61 unsigned long *bitmap; 62 unsigned long control; 63 unsigned long base; 64 unsigned long tmp; 65 66 iommu = kmalloc(sizeof(struct iommu_struct), GFP_KERNEL); 67 if (!iommu) { 68 prom_printf("Unable to allocate iommu structure\n"); 69 prom_halt(); 70 } 71 72 iommu->regs = of_ioremap(&op->resource[0], 0, PAGE_SIZE * 3, 73 "iommu_regs"); 74 if (!iommu->regs) { 75 prom_printf("Cannot map IOMMU registers\n"); 76 prom_halt(); 77 } 78 79 control = sbus_readl(&iommu->regs->control); 80 impl = (control & IOMMU_CTRL_IMPL) >> 28; 81 vers = (control & IOMMU_CTRL_VERS) >> 24; 82 control &= ~(IOMMU_CTRL_RNGE); 83 control |= (IOMMU_RNGE_256MB | IOMMU_CTRL_ENAB); 84 sbus_writel(control, &iommu->regs->control); 85 86 iommu_invalidate(iommu->regs); 87 iommu->start = IOMMU_START; 88 iommu->end = 0xffffffff; 89 90 /* Allocate IOMMU page table */ 91 /* Stupid alignment constraints give me a headache. 92 We need 256K or 512K or 1M or 2M area aligned to 93 its size and current gfp will fortunately give 94 it to us. */ 95 tmp = __get_free_pages(GFP_KERNEL, IOMMU_ORDER); 96 if (!tmp) { 97 prom_printf("Unable to allocate iommu table [0x%lx]\n", 98 IOMMU_NPTES * sizeof(iopte_t)); 99 prom_halt(); 100 } 101 iommu->page_table = (iopte_t *)tmp; 102 103 /* Initialize new table. */ 104 memset(iommu->page_table, 0, IOMMU_NPTES*sizeof(iopte_t)); 105 flush_cache_all(); 106 flush_tlb_all(); 107 108 base = __pa((unsigned long)iommu->page_table) >> 4; 109 sbus_writel(base, &iommu->regs->base); 110 iommu_invalidate(iommu->regs); 111 112 bitmap = kmalloc(IOMMU_NPTES>>3, GFP_KERNEL); 113 if (!bitmap) { 114 prom_printf("Unable to allocate iommu bitmap [%d]\n", 115 (int)(IOMMU_NPTES>>3)); 116 prom_halt(); 117 } 118 bit_map_init(&iommu->usemap, bitmap, IOMMU_NPTES); 119 /* To be coherent on HyperSparc, the page color of DVMA 120 * and physical addresses must match. 121 */ 122 if (srmmu_modtype == HyperSparc) 123 iommu->usemap.num_colors = vac_cache_size >> PAGE_SHIFT; 124 else 125 iommu->usemap.num_colors = 1; 126 127 printk(KERN_INFO "IOMMU: impl %d vers %d table 0x%p[%d B] map [%d b]\n", 128 impl, vers, iommu->page_table, 129 (int)(IOMMU_NPTES*sizeof(iopte_t)), (int)IOMMU_NPTES); 130 131 op->dev.archdata.iommu = iommu; 132 } 133 134 static int __init iommu_init(void) 135 { 136 struct device_node *dp; 137 138 for_each_node_by_name(dp, "iommu") { 139 struct platform_device *op = of_find_device_by_node(dp); 140 141 sbus_iommu_init(op); 142 of_propagate_archdata(op); 143 } 144 145 return 0; 146 } 147 148 subsys_initcall(iommu_init); 149 150 /* Flush the iotlb entries to ram. */ 151 /* This could be better if we didn't have to flush whole pages. */ 152 static void iommu_flush_iotlb(iopte_t *iopte, unsigned int niopte) 153 { 154 unsigned long start; 155 unsigned long end; 156 157 start = (unsigned long)iopte; 158 end = PAGE_ALIGN(start + niopte*sizeof(iopte_t)); 159 start &= PAGE_MASK; 160 if (viking_mxcc_present) { 161 while(start < end) { 162 viking_mxcc_flush_page(start); 163 start += PAGE_SIZE; 164 } 165 } else if (viking_flush) { 166 while(start < end) { 167 viking_flush_page(start); 168 start += PAGE_SIZE; 169 } 170 } else { 171 while(start < end) { 172 __flush_page_to_ram(start); 173 start += PAGE_SIZE; 174 } 175 } 176 } 177 178 static u32 iommu_get_one(struct device *dev, struct page *page, int npages) 179 { 180 struct iommu_struct *iommu = dev->archdata.iommu; 181 int ioptex; 182 iopte_t *iopte, *iopte0; 183 unsigned int busa, busa0; 184 int i; 185 186 /* page color = pfn of page */ 187 ioptex = bit_map_string_get(&iommu->usemap, npages, page_to_pfn(page)); 188 if (ioptex < 0) 189 panic("iommu out"); 190 busa0 = iommu->start + (ioptex << PAGE_SHIFT); 191 iopte0 = &iommu->page_table[ioptex]; 192 193 busa = busa0; 194 iopte = iopte0; 195 for (i = 0; i < npages; i++) { 196 iopte_val(*iopte) = MKIOPTE(page_to_pfn(page), IOPERM); 197 iommu_invalidate_page(iommu->regs, busa); 198 busa += PAGE_SIZE; 199 iopte++; 200 page++; 201 } 202 203 iommu_flush_iotlb(iopte0, npages); 204 205 return busa0; 206 } 207 208 static dma_addr_t __sbus_iommu_map_page(struct device *dev, struct page *page, 209 unsigned long offset, size_t len) 210 { 211 void *vaddr = page_address(page) + offset; 212 unsigned long off = (unsigned long)vaddr & ~PAGE_MASK; 213 unsigned long npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT; 214 215 /* XXX So what is maxphys for us and how do drivers know it? */ 216 if (!len || len > 256 * 1024) 217 return DMA_MAPPING_ERROR; 218 return iommu_get_one(dev, virt_to_page(vaddr), npages) + off; 219 } 220 221 static dma_addr_t sbus_iommu_map_page_gflush(struct device *dev, 222 struct page *page, unsigned long offset, size_t len, 223 enum dma_data_direction dir, unsigned long attrs) 224 { 225 flush_page_for_dma(0); 226 return __sbus_iommu_map_page(dev, page, offset, len); 227 } 228 229 static dma_addr_t sbus_iommu_map_page_pflush(struct device *dev, 230 struct page *page, unsigned long offset, size_t len, 231 enum dma_data_direction dir, unsigned long attrs) 232 { 233 void *vaddr = page_address(page) + offset; 234 unsigned long p = ((unsigned long)vaddr) & PAGE_MASK; 235 236 while (p < (unsigned long)vaddr + len) { 237 flush_page_for_dma(p); 238 p += PAGE_SIZE; 239 } 240 241 return __sbus_iommu_map_page(dev, page, offset, len); 242 } 243 244 static int sbus_iommu_map_sg_gflush(struct device *dev, struct scatterlist *sgl, 245 int nents, enum dma_data_direction dir, unsigned long attrs) 246 { 247 struct scatterlist *sg; 248 int i, n; 249 250 flush_page_for_dma(0); 251 252 for_each_sg(sgl, sg, nents, i) { 253 n = (sg->length + sg->offset + PAGE_SIZE-1) >> PAGE_SHIFT; 254 sg->dma_address = iommu_get_one(dev, sg_page(sg), n) + sg->offset; 255 sg->dma_length = sg->length; 256 } 257 258 return nents; 259 } 260 261 static int sbus_iommu_map_sg_pflush(struct device *dev, struct scatterlist *sgl, 262 int nents, enum dma_data_direction dir, unsigned long attrs) 263 { 264 unsigned long page, oldpage = 0; 265 struct scatterlist *sg; 266 int i, j, n; 267 268 for_each_sg(sgl, sg, nents, j) { 269 n = (sg->length + sg->offset + PAGE_SIZE-1) >> PAGE_SHIFT; 270 271 /* 272 * We expect unmapped highmem pages to be not in the cache. 273 * XXX Is this a good assumption? 274 * XXX What if someone else unmaps it here and races us? 275 */ 276 if (!PageHighMem(sg_page(sg))) { 277 page = (unsigned long)page_address(sg_page(sg)); 278 for (i = 0; i < n; i++) { 279 if (page != oldpage) { /* Already flushed? */ 280 flush_page_for_dma(page); 281 oldpage = page; 282 } 283 page += PAGE_SIZE; 284 } 285 } 286 287 sg->dma_address = iommu_get_one(dev, sg_page(sg), n) + sg->offset; 288 sg->dma_length = sg->length; 289 } 290 291 return nents; 292 } 293 294 static void sbus_iommu_unmap_page(struct device *dev, dma_addr_t dma_addr, 295 size_t len, enum dma_data_direction dir, unsigned long attrs) 296 { 297 struct iommu_struct *iommu = dev->archdata.iommu; 298 unsigned int busa = dma_addr & PAGE_MASK; 299 unsigned long off = dma_addr & ~PAGE_MASK; 300 unsigned int npages = (off + len + PAGE_SIZE-1) >> PAGE_SHIFT; 301 unsigned int ioptex = (busa - iommu->start) >> PAGE_SHIFT; 302 unsigned int i; 303 304 BUG_ON(busa < iommu->start); 305 for (i = 0; i < npages; i++) { 306 iopte_val(iommu->page_table[ioptex + i]) = 0; 307 iommu_invalidate_page(iommu->regs, busa); 308 busa += PAGE_SIZE; 309 } 310 bit_map_clear(&iommu->usemap, ioptex, npages); 311 } 312 313 static void sbus_iommu_unmap_sg(struct device *dev, struct scatterlist *sgl, 314 int nents, enum dma_data_direction dir, unsigned long attrs) 315 { 316 struct scatterlist *sg; 317 int i; 318 319 for_each_sg(sgl, sg, nents, i) { 320 sbus_iommu_unmap_page(dev, sg->dma_address, sg->length, dir, 321 attrs); 322 sg->dma_address = 0x21212121; 323 } 324 } 325 326 #ifdef CONFIG_SBUS 327 static void *sbus_iommu_alloc(struct device *dev, size_t len, 328 dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs) 329 { 330 struct iommu_struct *iommu = dev->archdata.iommu; 331 unsigned long va, addr, page, end, ret; 332 iopte_t *iopte = iommu->page_table; 333 iopte_t *first; 334 int ioptex; 335 336 /* XXX So what is maxphys for us and how do drivers know it? */ 337 if (!len || len > 256 * 1024) 338 return NULL; 339 340 len = PAGE_ALIGN(len); 341 va = __get_free_pages(gfp | __GFP_ZERO, get_order(len)); 342 if (va == 0) 343 return NULL; 344 345 addr = ret = sparc_dma_alloc_resource(dev, len); 346 if (!addr) 347 goto out_free_pages; 348 349 BUG_ON((va & ~PAGE_MASK) != 0); 350 BUG_ON((addr & ~PAGE_MASK) != 0); 351 BUG_ON((len & ~PAGE_MASK) != 0); 352 353 /* page color = physical address */ 354 ioptex = bit_map_string_get(&iommu->usemap, len >> PAGE_SHIFT, 355 addr >> PAGE_SHIFT); 356 if (ioptex < 0) 357 panic("iommu out"); 358 359 iopte += ioptex; 360 first = iopte; 361 end = addr + len; 362 while(addr < end) { 363 page = va; 364 { 365 pgd_t *pgdp; 366 pmd_t *pmdp; 367 pte_t *ptep; 368 369 if (viking_mxcc_present) 370 viking_mxcc_flush_page(page); 371 else if (viking_flush) 372 viking_flush_page(page); 373 else 374 __flush_page_to_ram(page); 375 376 pgdp = pgd_offset(&init_mm, addr); 377 pmdp = pmd_offset(pgdp, addr); 378 ptep = pte_offset_map(pmdp, addr); 379 380 set_pte(ptep, mk_pte(virt_to_page(page), dvma_prot)); 381 } 382 iopte_val(*iopte++) = 383 MKIOPTE(page_to_pfn(virt_to_page(page)), ioperm_noc); 384 addr += PAGE_SIZE; 385 va += PAGE_SIZE; 386 } 387 /* P3: why do we need this? 388 * 389 * DAVEM: Because there are several aspects, none of which 390 * are handled by a single interface. Some cpus are 391 * completely not I/O DMA coherent, and some have 392 * virtually indexed caches. The driver DMA flushing 393 * methods handle the former case, but here during 394 * IOMMU page table modifications, and usage of non-cacheable 395 * cpu mappings of pages potentially in the cpu caches, we have 396 * to handle the latter case as well. 397 */ 398 flush_cache_all(); 399 iommu_flush_iotlb(first, len >> PAGE_SHIFT); 400 flush_tlb_all(); 401 iommu_invalidate(iommu->regs); 402 403 *dma_handle = iommu->start + (ioptex << PAGE_SHIFT); 404 return (void *)ret; 405 406 out_free_pages: 407 free_pages(va, get_order(len)); 408 return NULL; 409 } 410 411 static void sbus_iommu_free(struct device *dev, size_t len, void *cpu_addr, 412 dma_addr_t busa, unsigned long attrs) 413 { 414 struct iommu_struct *iommu = dev->archdata.iommu; 415 iopte_t *iopte = iommu->page_table; 416 struct page *page = virt_to_page(cpu_addr); 417 int ioptex = (busa - iommu->start) >> PAGE_SHIFT; 418 unsigned long end; 419 420 if (!sparc_dma_free_resource(cpu_addr, len)) 421 return; 422 423 BUG_ON((busa & ~PAGE_MASK) != 0); 424 BUG_ON((len & ~PAGE_MASK) != 0); 425 426 iopte += ioptex; 427 end = busa + len; 428 while (busa < end) { 429 iopte_val(*iopte++) = 0; 430 busa += PAGE_SIZE; 431 } 432 flush_tlb_all(); 433 iommu_invalidate(iommu->regs); 434 bit_map_clear(&iommu->usemap, ioptex, len >> PAGE_SHIFT); 435 436 __free_pages(page, get_order(len)); 437 } 438 #endif 439 440 static const struct dma_map_ops sbus_iommu_dma_gflush_ops = { 441 #ifdef CONFIG_SBUS 442 .alloc = sbus_iommu_alloc, 443 .free = sbus_iommu_free, 444 #endif 445 .map_page = sbus_iommu_map_page_gflush, 446 .unmap_page = sbus_iommu_unmap_page, 447 .map_sg = sbus_iommu_map_sg_gflush, 448 .unmap_sg = sbus_iommu_unmap_sg, 449 }; 450 451 static const struct dma_map_ops sbus_iommu_dma_pflush_ops = { 452 #ifdef CONFIG_SBUS 453 .alloc = sbus_iommu_alloc, 454 .free = sbus_iommu_free, 455 #endif 456 .map_page = sbus_iommu_map_page_pflush, 457 .unmap_page = sbus_iommu_unmap_page, 458 .map_sg = sbus_iommu_map_sg_pflush, 459 .unmap_sg = sbus_iommu_unmap_sg, 460 }; 461 462 void __init ld_mmu_iommu(void) 463 { 464 if (flush_page_for_dma_global) { 465 /* flush_page_for_dma flushes everything, no matter of what page is it */ 466 dma_ops = &sbus_iommu_dma_gflush_ops; 467 } else { 468 dma_ops = &sbus_iommu_dma_pflush_ops; 469 } 470 471 if (viking_mxcc_present || srmmu_modtype == HyperSparc) { 472 dvma_prot = __pgprot(SRMMU_CACHE | SRMMU_ET_PTE | SRMMU_PRIV); 473 ioperm_noc = IOPTE_CACHE | IOPTE_WRITE | IOPTE_VALID; 474 } else { 475 dvma_prot = __pgprot(SRMMU_ET_PTE | SRMMU_PRIV); 476 ioperm_noc = IOPTE_WRITE | IOPTE_VALID; 477 } 478 } 479