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/dma-map-ops.h> 16 #include <linux/of.h> 17 #include <linux/of_device.h> 18 19 #include <asm/io.h> 20 #include <asm/mxcc.h> 21 #include <asm/mbus.h> 22 #include <asm/cacheflush.h> 23 #include <asm/tlbflush.h> 24 #include <asm/bitext.h> 25 #include <asm/iommu.h> 26 #include <asm/dma.h> 27 28 #include "mm_32.h" 29 30 /* 31 * This can be sized dynamically, but we will do this 32 * only when we have a guidance about actual I/O pressures. 33 */ 34 #define IOMMU_RNGE IOMMU_RNGE_256MB 35 #define IOMMU_START 0xF0000000 36 #define IOMMU_WINSIZE (256*1024*1024U) 37 #define IOMMU_NPTES (IOMMU_WINSIZE/PAGE_SIZE) /* 64K PTEs, 256KB */ 38 #define IOMMU_ORDER 6 /* 4096 * (1<<6) */ 39 40 static int viking_flush; 41 /* viking.S */ 42 extern void viking_flush_page(unsigned long page); 43 extern void viking_mxcc_flush_page(unsigned long page); 44 45 /* 46 * Values precomputed according to CPU type. 47 */ 48 static unsigned int ioperm_noc; /* Consistent mapping iopte flags */ 49 static pgprot_t dvma_prot; /* Consistent mapping pte flags */ 50 51 #define IOPERM (IOPTE_CACHE | IOPTE_WRITE | IOPTE_VALID) 52 #define MKIOPTE(pfn, perm) (((((pfn)<<8) & IOPTE_PAGE) | (perm)) & ~IOPTE_WAZ) 53 54 static const struct dma_map_ops sbus_iommu_dma_gflush_ops; 55 static const struct dma_map_ops sbus_iommu_dma_pflush_ops; 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 if (flush_page_for_dma_global) 134 op->dev.dma_ops = &sbus_iommu_dma_gflush_ops; 135 else 136 op->dev.dma_ops = &sbus_iommu_dma_pflush_ops; 137 } 138 139 static int __init iommu_init(void) 140 { 141 struct device_node *dp; 142 143 for_each_node_by_name(dp, "iommu") { 144 struct platform_device *op = of_find_device_by_node(dp); 145 146 sbus_iommu_init(op); 147 of_propagate_archdata(op); 148 } 149 150 return 0; 151 } 152 153 subsys_initcall(iommu_init); 154 155 /* Flush the iotlb entries to ram. */ 156 /* This could be better if we didn't have to flush whole pages. */ 157 static void iommu_flush_iotlb(iopte_t *iopte, unsigned int niopte) 158 { 159 unsigned long start; 160 unsigned long end; 161 162 start = (unsigned long)iopte; 163 end = PAGE_ALIGN(start + niopte*sizeof(iopte_t)); 164 start &= PAGE_MASK; 165 if (viking_mxcc_present) { 166 while(start < end) { 167 viking_mxcc_flush_page(start); 168 start += PAGE_SIZE; 169 } 170 } else if (viking_flush) { 171 while(start < end) { 172 viking_flush_page(start); 173 start += PAGE_SIZE; 174 } 175 } else { 176 while(start < end) { 177 __flush_page_to_ram(start); 178 start += PAGE_SIZE; 179 } 180 } 181 } 182 183 static dma_addr_t __sbus_iommu_map_page(struct device *dev, struct page *page, 184 unsigned long offset, size_t len, bool per_page_flush) 185 { 186 struct iommu_struct *iommu = dev->archdata.iommu; 187 phys_addr_t paddr = page_to_phys(page) + offset; 188 unsigned long off = paddr & ~PAGE_MASK; 189 unsigned long npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT; 190 unsigned long pfn = __phys_to_pfn(paddr); 191 unsigned int busa, busa0; 192 iopte_t *iopte, *iopte0; 193 int ioptex, i; 194 195 /* XXX So what is maxphys for us and how do drivers know it? */ 196 if (!len || len > 256 * 1024) 197 return DMA_MAPPING_ERROR; 198 199 /* 200 * We expect unmapped highmem pages to be not in the cache. 201 * XXX Is this a good assumption? 202 * XXX What if someone else unmaps it here and races us? 203 */ 204 if (per_page_flush && !PageHighMem(page)) { 205 unsigned long vaddr, p; 206 207 vaddr = (unsigned long)page_address(page) + offset; 208 for (p = vaddr & PAGE_MASK; p < vaddr + len; p += PAGE_SIZE) 209 flush_page_for_dma(p); 210 } 211 212 /* page color = pfn of page */ 213 ioptex = bit_map_string_get(&iommu->usemap, npages, pfn); 214 if (ioptex < 0) 215 panic("iommu out"); 216 busa0 = iommu->start + (ioptex << PAGE_SHIFT); 217 iopte0 = &iommu->page_table[ioptex]; 218 219 busa = busa0; 220 iopte = iopte0; 221 for (i = 0; i < npages; i++) { 222 iopte_val(*iopte) = MKIOPTE(pfn, IOPERM); 223 iommu_invalidate_page(iommu->regs, busa); 224 busa += PAGE_SIZE; 225 iopte++; 226 pfn++; 227 } 228 229 iommu_flush_iotlb(iopte0, npages); 230 return busa0 + off; 231 } 232 233 static dma_addr_t sbus_iommu_map_page_gflush(struct device *dev, 234 struct page *page, unsigned long offset, size_t len, 235 enum dma_data_direction dir, unsigned long attrs) 236 { 237 flush_page_for_dma(0); 238 return __sbus_iommu_map_page(dev, page, offset, len, false); 239 } 240 241 static dma_addr_t sbus_iommu_map_page_pflush(struct device *dev, 242 struct page *page, unsigned long offset, size_t len, 243 enum dma_data_direction dir, unsigned long attrs) 244 { 245 return __sbus_iommu_map_page(dev, page, offset, len, true); 246 } 247 248 static int __sbus_iommu_map_sg(struct device *dev, struct scatterlist *sgl, 249 int nents, enum dma_data_direction dir, unsigned long attrs, 250 bool per_page_flush) 251 { 252 struct scatterlist *sg; 253 int j; 254 255 for_each_sg(sgl, sg, nents, j) { 256 sg->dma_address =__sbus_iommu_map_page(dev, sg_page(sg), 257 sg->offset, sg->length, per_page_flush); 258 if (sg->dma_address == DMA_MAPPING_ERROR) 259 return 0; 260 sg->dma_length = sg->length; 261 } 262 263 return nents; 264 } 265 266 static int sbus_iommu_map_sg_gflush(struct device *dev, struct scatterlist *sgl, 267 int nents, enum dma_data_direction dir, unsigned long attrs) 268 { 269 flush_page_for_dma(0); 270 return __sbus_iommu_map_sg(dev, sgl, nents, dir, attrs, false); 271 } 272 273 static int sbus_iommu_map_sg_pflush(struct device *dev, struct scatterlist *sgl, 274 int nents, enum dma_data_direction dir, unsigned long attrs) 275 { 276 return __sbus_iommu_map_sg(dev, sgl, nents, dir, attrs, true); 277 } 278 279 static void sbus_iommu_unmap_page(struct device *dev, dma_addr_t dma_addr, 280 size_t len, enum dma_data_direction dir, unsigned long attrs) 281 { 282 struct iommu_struct *iommu = dev->archdata.iommu; 283 unsigned int busa = dma_addr & PAGE_MASK; 284 unsigned long off = dma_addr & ~PAGE_MASK; 285 unsigned int npages = (off + len + PAGE_SIZE-1) >> PAGE_SHIFT; 286 unsigned int ioptex = (busa - iommu->start) >> PAGE_SHIFT; 287 unsigned int i; 288 289 BUG_ON(busa < iommu->start); 290 for (i = 0; i < npages; i++) { 291 iopte_val(iommu->page_table[ioptex + i]) = 0; 292 iommu_invalidate_page(iommu->regs, busa); 293 busa += PAGE_SIZE; 294 } 295 bit_map_clear(&iommu->usemap, ioptex, npages); 296 } 297 298 static void sbus_iommu_unmap_sg(struct device *dev, struct scatterlist *sgl, 299 int nents, enum dma_data_direction dir, unsigned long attrs) 300 { 301 struct scatterlist *sg; 302 int i; 303 304 for_each_sg(sgl, sg, nents, i) { 305 sbus_iommu_unmap_page(dev, sg->dma_address, sg->length, dir, 306 attrs); 307 sg->dma_address = 0x21212121; 308 } 309 } 310 311 #ifdef CONFIG_SBUS 312 static void *sbus_iommu_alloc(struct device *dev, size_t len, 313 dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs) 314 { 315 struct iommu_struct *iommu = dev->archdata.iommu; 316 unsigned long va, addr, page, end, ret; 317 iopte_t *iopte = iommu->page_table; 318 iopte_t *first; 319 int ioptex; 320 321 /* XXX So what is maxphys for us and how do drivers know it? */ 322 if (!len || len > 256 * 1024) 323 return NULL; 324 325 len = PAGE_ALIGN(len); 326 va = __get_free_pages(gfp | __GFP_ZERO, get_order(len)); 327 if (va == 0) 328 return NULL; 329 330 addr = ret = sparc_dma_alloc_resource(dev, len); 331 if (!addr) 332 goto out_free_pages; 333 334 BUG_ON((va & ~PAGE_MASK) != 0); 335 BUG_ON((addr & ~PAGE_MASK) != 0); 336 BUG_ON((len & ~PAGE_MASK) != 0); 337 338 /* page color = physical address */ 339 ioptex = bit_map_string_get(&iommu->usemap, len >> PAGE_SHIFT, 340 addr >> PAGE_SHIFT); 341 if (ioptex < 0) 342 panic("iommu out"); 343 344 iopte += ioptex; 345 first = iopte; 346 end = addr + len; 347 while(addr < end) { 348 page = va; 349 { 350 pmd_t *pmdp; 351 pte_t *ptep; 352 353 if (viking_mxcc_present) 354 viking_mxcc_flush_page(page); 355 else if (viking_flush) 356 viking_flush_page(page); 357 else 358 __flush_page_to_ram(page); 359 360 pmdp = pmd_off_k(addr); 361 ptep = pte_offset_map(pmdp, addr); 362 363 set_pte(ptep, mk_pte(virt_to_page(page), dvma_prot)); 364 } 365 iopte_val(*iopte++) = 366 MKIOPTE(page_to_pfn(virt_to_page(page)), ioperm_noc); 367 addr += PAGE_SIZE; 368 va += PAGE_SIZE; 369 } 370 /* P3: why do we need this? 371 * 372 * DAVEM: Because there are several aspects, none of which 373 * are handled by a single interface. Some cpus are 374 * completely not I/O DMA coherent, and some have 375 * virtually indexed caches. The driver DMA flushing 376 * methods handle the former case, but here during 377 * IOMMU page table modifications, and usage of non-cacheable 378 * cpu mappings of pages potentially in the cpu caches, we have 379 * to handle the latter case as well. 380 */ 381 flush_cache_all(); 382 iommu_flush_iotlb(first, len >> PAGE_SHIFT); 383 flush_tlb_all(); 384 iommu_invalidate(iommu->regs); 385 386 *dma_handle = iommu->start + (ioptex << PAGE_SHIFT); 387 return (void *)ret; 388 389 out_free_pages: 390 free_pages(va, get_order(len)); 391 return NULL; 392 } 393 394 static void sbus_iommu_free(struct device *dev, size_t len, void *cpu_addr, 395 dma_addr_t busa, unsigned long attrs) 396 { 397 struct iommu_struct *iommu = dev->archdata.iommu; 398 iopte_t *iopte = iommu->page_table; 399 struct page *page = virt_to_page(cpu_addr); 400 int ioptex = (busa - iommu->start) >> PAGE_SHIFT; 401 unsigned long end; 402 403 if (!sparc_dma_free_resource(cpu_addr, len)) 404 return; 405 406 BUG_ON((busa & ~PAGE_MASK) != 0); 407 BUG_ON((len & ~PAGE_MASK) != 0); 408 409 iopte += ioptex; 410 end = busa + len; 411 while (busa < end) { 412 iopte_val(*iopte++) = 0; 413 busa += PAGE_SIZE; 414 } 415 flush_tlb_all(); 416 iommu_invalidate(iommu->regs); 417 bit_map_clear(&iommu->usemap, ioptex, len >> PAGE_SHIFT); 418 419 __free_pages(page, get_order(len)); 420 } 421 #endif 422 423 static const struct dma_map_ops sbus_iommu_dma_gflush_ops = { 424 #ifdef CONFIG_SBUS 425 .alloc = sbus_iommu_alloc, 426 .free = sbus_iommu_free, 427 #endif 428 .map_page = sbus_iommu_map_page_gflush, 429 .unmap_page = sbus_iommu_unmap_page, 430 .map_sg = sbus_iommu_map_sg_gflush, 431 .unmap_sg = sbus_iommu_unmap_sg, 432 }; 433 434 static const struct dma_map_ops sbus_iommu_dma_pflush_ops = { 435 #ifdef CONFIG_SBUS 436 .alloc = sbus_iommu_alloc, 437 .free = sbus_iommu_free, 438 #endif 439 .map_page = sbus_iommu_map_page_pflush, 440 .unmap_page = sbus_iommu_unmap_page, 441 .map_sg = sbus_iommu_map_sg_pflush, 442 .unmap_sg = sbus_iommu_unmap_sg, 443 }; 444 445 void __init ld_mmu_iommu(void) 446 { 447 if (viking_mxcc_present || srmmu_modtype == HyperSparc) { 448 dvma_prot = __pgprot(SRMMU_CACHE | SRMMU_ET_PTE | SRMMU_PRIV); 449 ioperm_noc = IOPTE_CACHE | IOPTE_WRITE | IOPTE_VALID; 450 } else { 451 dvma_prot = __pgprot(SRMMU_ET_PTE | SRMMU_PRIV); 452 ioperm_noc = IOPTE_WRITE | IOPTE_VALID; 453 } 454 } 455