2 * iommu.c: IOMMU specific routines for memory management.
4 * Copyright (C) 1995 David S. Miller (davem@caip.rutgers.edu)
5 * Copyright (C) 1995,2002 Pete Zaitcev (zaitcev@yahoo.com)
6 * Copyright (C) 1996 Eddie C. Dost (ecd@skynet.be)
7 * Copyright (C) 1997,1998 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
10 #include <linux/kernel.h>
11 #include <linux/init.h>
13 #include <linux/slab.h>
14 #include <linux/highmem.h> /* pte_offset_map => kmap_atomic */
16 #include <asm/scatterlist.h>
17 #include <asm/pgalloc.h>
18 #include <asm/pgtable.h>
23 #include <asm/cacheflush.h>
24 #include <asm/tlbflush.h>
25 #include <asm/bitext.h>
26 #include <asm/iommu.h>
30 * This can be sized dynamically, but we will do this
31 * only when we have a guidance about actual I/O pressures.
33 #define IOMMU_RNGE IOMMU_RNGE_256MB
34 #define IOMMU_START 0xF0000000
35 #define IOMMU_WINSIZE (256*1024*1024U)
36 #define IOMMU_NPTES (IOMMU_WINSIZE/PAGE_SIZE) /* 64K PTEs, 265KB */
37 #define IOMMU_ORDER 6 /* 4096 * (1<<6) */
40 extern int viking_mxcc_present;
41 BTFIXUPDEF_CALL(void, flush_page_for_dma, unsigned long)
42 #define flush_page_for_dma(page) BTFIXUP_CALL(flush_page_for_dma)(page)
43 extern int flush_page_for_dma_global;
44 static int viking_flush;
46 extern void viking_flush_page(unsigned long page);
47 extern void viking_mxcc_flush_page(unsigned long page);
50 * Values precomputed according to CPU type.
52 static unsigned int ioperm_noc; /* Consistent mapping iopte flags */
53 static pgprot_t dvma_prot; /* Consistent mapping pte flags */
55 #define IOPERM (IOPTE_CACHE | IOPTE_WRITE | IOPTE_VALID)
56 #define MKIOPTE(pfn, perm) (((((pfn)<<8) & IOPTE_PAGE) | (perm)) & ~IOPTE_WAZ)
59 iommu_init(int iommund, struct sbus_bus *sbus)
61 unsigned int impl, vers;
63 struct iommu_struct *iommu;
64 struct linux_prom_registers iommu_promregs[PROMREG_MAX];
66 unsigned long *bitmap;
68 iommu = kmalloc(sizeof(struct iommu_struct), GFP_ATOMIC);
70 prom_printf("Unable to allocate iommu structure\n");
74 if (prom_getproperty(iommund, "reg", (void *) iommu_promregs,
75 sizeof(iommu_promregs)) != -1) {
76 memset(&r, 0, sizeof(r));
77 r.flags = iommu_promregs[0].which_io;
78 r.start = iommu_promregs[0].phys_addr;
79 iommu->regs = (struct iommu_regs *)
80 sbus_ioremap(&r, 0, PAGE_SIZE * 3, "iommu_regs");
83 prom_printf("Cannot map IOMMU registers\n");
86 impl = (iommu->regs->control & IOMMU_CTRL_IMPL) >> 28;
87 vers = (iommu->regs->control & IOMMU_CTRL_VERS) >> 24;
88 tmp = iommu->regs->control;
89 tmp &= ~(IOMMU_CTRL_RNGE);
90 tmp |= (IOMMU_RNGE_256MB | IOMMU_CTRL_ENAB);
91 iommu->regs->control = tmp;
92 iommu_invalidate(iommu->regs);
93 iommu->start = IOMMU_START;
94 iommu->end = 0xffffffff;
96 /* Allocate IOMMU page table */
97 /* Stupid alignment constraints give me a headache.
98 We need 256K or 512K or 1M or 2M area aligned to
99 its size and current gfp will fortunately give
101 tmp = __get_free_pages(GFP_KERNEL, IOMMU_ORDER);
103 prom_printf("Unable to allocate iommu table [0x%08x]\n",
104 IOMMU_NPTES*sizeof(iopte_t));
107 iommu->page_table = (iopte_t *)tmp;
109 /* Initialize new table. */
110 memset(iommu->page_table, 0, IOMMU_NPTES*sizeof(iopte_t));
113 iommu->regs->base = __pa((unsigned long) iommu->page_table) >> 4;
114 iommu_invalidate(iommu->regs);
116 bitmap = kmalloc(IOMMU_NPTES>>3, GFP_KERNEL);
118 prom_printf("Unable to allocate iommu bitmap [%d]\n",
119 (int)(IOMMU_NPTES>>3));
122 bit_map_init(&iommu->usemap, bitmap, IOMMU_NPTES);
123 /* To be coherent on HyperSparc, the page color of DVMA
124 * and physical addresses must match.
126 if (srmmu_modtype == HyperSparc)
127 iommu->usemap.num_colors = vac_cache_size >> PAGE_SHIFT;
129 iommu->usemap.num_colors = 1;
131 printk("IOMMU: impl %d vers %d table 0x%p[%d B] map [%d b]\n",
132 impl, vers, iommu->page_table,
133 (int)(IOMMU_NPTES*sizeof(iopte_t)), (int)IOMMU_NPTES);
138 /* This begs to be btfixup-ed by srmmu. */
139 /* Flush the iotlb entries to ram. */
140 /* This could be better if we didn't have to flush whole pages. */
141 static void iommu_flush_iotlb(iopte_t *iopte, unsigned int niopte)
146 start = (unsigned long)iopte;
147 end = PAGE_ALIGN(start + niopte*sizeof(iopte_t));
149 if (viking_mxcc_present) {
151 viking_mxcc_flush_page(start);
154 } else if (viking_flush) {
156 viking_flush_page(start);
161 __flush_page_to_ram(start);
167 static u32 iommu_get_one(struct page *page, int npages, struct sbus_bus *sbus)
169 struct iommu_struct *iommu = sbus->iommu;
171 iopte_t *iopte, *iopte0;
172 unsigned int busa, busa0;
175 /* page color = pfn of page */
176 ioptex = bit_map_string_get(&iommu->usemap, npages, page_to_pfn(page));
179 busa0 = iommu->start + (ioptex << PAGE_SHIFT);
180 iopte0 = &iommu->page_table[ioptex];
184 for (i = 0; i < npages; i++) {
185 iopte_val(*iopte) = MKIOPTE(page_to_pfn(page), IOPERM);
186 iommu_invalidate_page(iommu->regs, busa);
192 iommu_flush_iotlb(iopte0, npages);
197 static u32 iommu_get_scsi_one(char *vaddr, unsigned int len,
198 struct sbus_bus *sbus)
205 off = (unsigned long)vaddr & ~PAGE_MASK;
206 npages = (off + len + PAGE_SIZE-1) >> PAGE_SHIFT;
207 page = virt_to_page((unsigned long)vaddr & PAGE_MASK);
208 busa = iommu_get_one(page, npages, sbus);
212 static __u32 iommu_get_scsi_one_noflush(char *vaddr, unsigned long len, struct sbus_bus *sbus)
214 return iommu_get_scsi_one(vaddr, len, sbus);
217 static __u32 iommu_get_scsi_one_gflush(char *vaddr, unsigned long len, struct sbus_bus *sbus)
219 flush_page_for_dma(0);
220 return iommu_get_scsi_one(vaddr, len, sbus);
223 static __u32 iommu_get_scsi_one_pflush(char *vaddr, unsigned long len, struct sbus_bus *sbus)
225 unsigned long page = ((unsigned long) vaddr) & PAGE_MASK;
227 while(page < ((unsigned long)(vaddr + len))) {
228 flush_page_for_dma(page);
231 return iommu_get_scsi_one(vaddr, len, sbus);
234 static void iommu_get_scsi_sgl_noflush(struct scatterlist *sg, int sz, struct sbus_bus *sbus)
240 n = (sg->length + sg->offset + PAGE_SIZE-1) >> PAGE_SHIFT;
241 sg->dvma_address = iommu_get_one(sg->page, n, sbus) + sg->offset;
242 sg->dvma_length = (__u32) sg->length;
247 static void iommu_get_scsi_sgl_gflush(struct scatterlist *sg, int sz, struct sbus_bus *sbus)
251 flush_page_for_dma(0);
254 n = (sg->length + sg->offset + PAGE_SIZE-1) >> PAGE_SHIFT;
255 sg->dvma_address = iommu_get_one(sg->page, n, sbus) + sg->offset;
256 sg->dvma_length = (__u32) sg->length;
261 static void iommu_get_scsi_sgl_pflush(struct scatterlist *sg, int sz, struct sbus_bus *sbus)
263 unsigned long page, oldpage = 0;
269 n = (sg->length + sg->offset + PAGE_SIZE-1) >> PAGE_SHIFT;
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?
276 if ((page = (unsigned long) page_address(sg->page)) != 0) {
277 for (i = 0; i < n; i++) {
278 if (page != oldpage) { /* Already flushed? */
279 flush_page_for_dma(page);
286 sg->dvma_address = iommu_get_one(sg->page, n, sbus) + sg->offset;
287 sg->dvma_length = (__u32) sg->length;
292 static void iommu_release_one(u32 busa, int npages, struct sbus_bus *sbus)
294 struct iommu_struct *iommu = sbus->iommu;
298 BUG_ON(busa < iommu->start);
299 ioptex = (busa - iommu->start) >> PAGE_SHIFT;
300 for (i = 0; i < npages; i++) {
301 iopte_val(iommu->page_table[ioptex + i]) = 0;
302 iommu_invalidate_page(iommu->regs, busa);
305 bit_map_clear(&iommu->usemap, ioptex, npages);
308 static void iommu_release_scsi_one(__u32 vaddr, unsigned long len, struct sbus_bus *sbus)
313 off = vaddr & ~PAGE_MASK;
314 npages = (off + len + PAGE_SIZE-1) >> PAGE_SHIFT;
315 iommu_release_one(vaddr & PAGE_MASK, npages, sbus);
318 static void iommu_release_scsi_sgl(struct scatterlist *sg, int sz, struct sbus_bus *sbus)
325 n = (sg->length + sg->offset + PAGE_SIZE-1) >> PAGE_SHIFT;
326 iommu_release_one(sg->dvma_address & PAGE_MASK, n, sbus);
327 sg->dvma_address = 0x21212121;
333 static int iommu_map_dma_area(dma_addr_t *pba, unsigned long va,
334 unsigned long addr, int len)
336 unsigned long page, end;
337 struct iommu_struct *iommu = sbus_root->iommu;
338 iopte_t *iopte = iommu->page_table;
342 BUG_ON((va & ~PAGE_MASK) != 0);
343 BUG_ON((addr & ~PAGE_MASK) != 0);
344 BUG_ON((len & ~PAGE_MASK) != 0);
346 /* page color = physical address */
347 ioptex = bit_map_string_get(&iommu->usemap, len >> PAGE_SHIFT,
362 if (viking_mxcc_present)
363 viking_mxcc_flush_page(page);
364 else if (viking_flush)
365 viking_flush_page(page);
367 __flush_page_to_ram(page);
369 pgdp = pgd_offset(&init_mm, addr);
370 pmdp = pmd_offset(pgdp, addr);
371 ptep = pte_offset_map(pmdp, addr);
373 set_pte(ptep, mk_pte(virt_to_page(page), dvma_prot));
375 iopte_val(*iopte++) =
376 MKIOPTE(page_to_pfn(virt_to_page(page)), ioperm_noc);
380 /* P3: why do we need this?
382 * DAVEM: Because there are several aspects, none of which
383 * are handled by a single interface. Some cpus are
384 * completely not I/O DMA coherent, and some have
385 * virtually indexed caches. The driver DMA flushing
386 * methods handle the former case, but here during
387 * IOMMU page table modifications, and usage of non-cacheable
388 * cpu mappings of pages potentially in the cpu caches, we have
389 * to handle the latter case as well.
392 iommu_flush_iotlb(first, len >> PAGE_SHIFT);
394 iommu_invalidate(iommu->regs);
396 *pba = iommu->start + (ioptex << PAGE_SHIFT);
400 static void iommu_unmap_dma_area(unsigned long busa, int len)
402 struct iommu_struct *iommu = sbus_root->iommu;
403 iopte_t *iopte = iommu->page_table;
405 int ioptex = (busa - iommu->start) >> PAGE_SHIFT;
407 BUG_ON((busa & ~PAGE_MASK) != 0);
408 BUG_ON((len & ~PAGE_MASK) != 0);
413 iopte_val(*iopte++) = 0;
417 iommu_invalidate(iommu->regs);
418 bit_map_clear(&iommu->usemap, ioptex, len >> PAGE_SHIFT);
421 static struct page *iommu_translate_dvma(unsigned long busa)
423 struct iommu_struct *iommu = sbus_root->iommu;
424 iopte_t *iopte = iommu->page_table;
426 iopte += ((busa - iommu->start) >> PAGE_SHIFT);
427 return pfn_to_page((iopte_val(*iopte) & IOPTE_PAGE) >> (PAGE_SHIFT-4));
431 static char *iommu_lockarea(char *vaddr, unsigned long len)
436 static void iommu_unlockarea(char *vaddr, unsigned long len)
440 void __init ld_mmu_iommu(void)
442 viking_flush = (BTFIXUPVAL_CALL(flush_page_for_dma) == (unsigned long)viking_flush_page);
443 BTFIXUPSET_CALL(mmu_lockarea, iommu_lockarea, BTFIXUPCALL_RETO0);
444 BTFIXUPSET_CALL(mmu_unlockarea, iommu_unlockarea, BTFIXUPCALL_NOP);
446 if (!BTFIXUPVAL_CALL(flush_page_for_dma)) {
447 /* IO coherent chip */
448 BTFIXUPSET_CALL(mmu_get_scsi_one, iommu_get_scsi_one_noflush, BTFIXUPCALL_RETO0);
449 BTFIXUPSET_CALL(mmu_get_scsi_sgl, iommu_get_scsi_sgl_noflush, BTFIXUPCALL_NORM);
450 } else if (flush_page_for_dma_global) {
451 /* flush_page_for_dma flushes everything, no matter of what page is it */
452 BTFIXUPSET_CALL(mmu_get_scsi_one, iommu_get_scsi_one_gflush, BTFIXUPCALL_NORM);
453 BTFIXUPSET_CALL(mmu_get_scsi_sgl, iommu_get_scsi_sgl_gflush, BTFIXUPCALL_NORM);
455 BTFIXUPSET_CALL(mmu_get_scsi_one, iommu_get_scsi_one_pflush, BTFIXUPCALL_NORM);
456 BTFIXUPSET_CALL(mmu_get_scsi_sgl, iommu_get_scsi_sgl_pflush, BTFIXUPCALL_NORM);
458 BTFIXUPSET_CALL(mmu_release_scsi_one, iommu_release_scsi_one, BTFIXUPCALL_NORM);
459 BTFIXUPSET_CALL(mmu_release_scsi_sgl, iommu_release_scsi_sgl, BTFIXUPCALL_NORM);
462 BTFIXUPSET_CALL(mmu_map_dma_area, iommu_map_dma_area, BTFIXUPCALL_NORM);
463 BTFIXUPSET_CALL(mmu_unmap_dma_area, iommu_unmap_dma_area, BTFIXUPCALL_NORM);
464 BTFIXUPSET_CALL(mmu_translate_dvma, iommu_translate_dvma, BTFIXUPCALL_NORM);
467 if (viking_mxcc_present || srmmu_modtype == HyperSparc) {
468 dvma_prot = __pgprot(SRMMU_CACHE | SRMMU_ET_PTE | SRMMU_PRIV);
469 ioperm_noc = IOPTE_CACHE | IOPTE_WRITE | IOPTE_VALID;
471 dvma_prot = __pgprot(SRMMU_ET_PTE | SRMMU_PRIV);
472 ioperm_noc = IOPTE_WRITE | IOPTE_VALID;