2 * iommu.c: IOMMU specific routines for memory management.
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 */
15 #include <linux/scatterlist.h>
17 #include <linux/of_device.h>
19 #include <asm/pgalloc.h>
20 #include <asm/pgtable.h>
24 #include <asm/cacheflush.h>
25 #include <asm/tlbflush.h>
26 #include <asm/bitext.h>
27 #include <asm/iommu.h>
31 * This can be sized dynamically, but we will do this
32 * only when we have a guidance about actual I/O pressures.
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) */
41 extern int viking_mxcc_present;
42 extern int flush_page_for_dma_global;
43 static int viking_flush;
45 extern void viking_flush_page(unsigned long page);
46 extern void viking_mxcc_flush_page(unsigned long page);
49 * Values precomputed according to CPU type.
51 static unsigned int ioperm_noc; /* Consistent mapping iopte flags */
52 static pgprot_t dvma_prot; /* Consistent mapping pte flags */
54 #define IOPERM (IOPTE_CACHE | IOPTE_WRITE | IOPTE_VALID)
55 #define MKIOPTE(pfn, perm) (((((pfn)<<8) & IOPTE_PAGE) | (perm)) & ~IOPTE_WAZ)
57 static void __init sbus_iommu_init(struct platform_device *op)
59 struct iommu_struct *iommu;
60 unsigned int impl, vers;
61 unsigned long *bitmap;
64 iommu = kmalloc(sizeof(struct iommu_struct), GFP_KERNEL);
66 prom_printf("Unable to allocate iommu structure\n");
70 iommu->regs = of_ioremap(&op->resource[0], 0, PAGE_SIZE * 3,
73 prom_printf("Cannot map IOMMU registers\n");
76 impl = (iommu->regs->control & IOMMU_CTRL_IMPL) >> 28;
77 vers = (iommu->regs->control & IOMMU_CTRL_VERS) >> 24;
78 tmp = iommu->regs->control;
79 tmp &= ~(IOMMU_CTRL_RNGE);
80 tmp |= (IOMMU_RNGE_256MB | IOMMU_CTRL_ENAB);
81 iommu->regs->control = tmp;
82 iommu_invalidate(iommu->regs);
83 iommu->start = IOMMU_START;
84 iommu->end = 0xffffffff;
86 /* Allocate IOMMU page table */
87 /* Stupid alignment constraints give me a headache.
88 We need 256K or 512K or 1M or 2M area aligned to
89 its size and current gfp will fortunately give
91 tmp = __get_free_pages(GFP_KERNEL, IOMMU_ORDER);
93 prom_printf("Unable to allocate iommu table [0x%lx]\n",
94 IOMMU_NPTES * sizeof(iopte_t));
97 iommu->page_table = (iopte_t *)tmp;
99 /* Initialize new table. */
100 memset(iommu->page_table, 0, IOMMU_NPTES*sizeof(iopte_t));
103 iommu->regs->base = __pa((unsigned long) iommu->page_table) >> 4;
104 iommu_invalidate(iommu->regs);
106 bitmap = kmalloc(IOMMU_NPTES>>3, GFP_KERNEL);
108 prom_printf("Unable to allocate iommu bitmap [%d]\n",
109 (int)(IOMMU_NPTES>>3));
112 bit_map_init(&iommu->usemap, bitmap, IOMMU_NPTES);
113 /* To be coherent on HyperSparc, the page color of DVMA
114 * and physical addresses must match.
116 if (srmmu_modtype == HyperSparc)
117 iommu->usemap.num_colors = vac_cache_size >> PAGE_SHIFT;
119 iommu->usemap.num_colors = 1;
121 printk(KERN_INFO "IOMMU: impl %d vers %d table 0x%p[%d B] map [%d b]\n",
122 impl, vers, iommu->page_table,
123 (int)(IOMMU_NPTES*sizeof(iopte_t)), (int)IOMMU_NPTES);
125 op->dev.archdata.iommu = iommu;
128 static int __init iommu_init(void)
130 struct device_node *dp;
132 for_each_node_by_name(dp, "iommu") {
133 struct platform_device *op = of_find_device_by_node(dp);
136 of_propagate_archdata(op);
142 subsys_initcall(iommu_init);
144 /* Flush the iotlb entries to ram. */
145 /* This could be better if we didn't have to flush whole pages. */
146 static void iommu_flush_iotlb(iopte_t *iopte, unsigned int niopte)
151 start = (unsigned long)iopte;
152 end = PAGE_ALIGN(start + niopte*sizeof(iopte_t));
154 if (viking_mxcc_present) {
156 viking_mxcc_flush_page(start);
159 } else if (viking_flush) {
161 viking_flush_page(start);
166 __flush_page_to_ram(start);
172 static u32 iommu_get_one(struct device *dev, struct page *page, int npages)
174 struct iommu_struct *iommu = dev->archdata.iommu;
176 iopte_t *iopte, *iopte0;
177 unsigned int busa, busa0;
180 /* page color = pfn of page */
181 ioptex = bit_map_string_get(&iommu->usemap, npages, page_to_pfn(page));
184 busa0 = iommu->start + (ioptex << PAGE_SHIFT);
185 iopte0 = &iommu->page_table[ioptex];
189 for (i = 0; i < npages; i++) {
190 iopte_val(*iopte) = MKIOPTE(page_to_pfn(page), IOPERM);
191 iommu_invalidate_page(iommu->regs, busa);
197 iommu_flush_iotlb(iopte0, npages);
202 static u32 iommu_get_scsi_one(struct device *dev, char *vaddr, unsigned int len)
209 off = (unsigned long)vaddr & ~PAGE_MASK;
210 npages = (off + len + PAGE_SIZE-1) >> PAGE_SHIFT;
211 page = virt_to_page((unsigned long)vaddr & PAGE_MASK);
212 busa = iommu_get_one(dev, page, npages);
216 static __u32 iommu_get_scsi_one_gflush(struct device *dev, char *vaddr, unsigned long len)
218 flush_page_for_dma(0);
219 return iommu_get_scsi_one(dev, vaddr, len);
222 static __u32 iommu_get_scsi_one_pflush(struct device *dev, char *vaddr, unsigned long len)
224 unsigned long page = ((unsigned long) vaddr) & PAGE_MASK;
226 while(page < ((unsigned long)(vaddr + len))) {
227 flush_page_for_dma(page);
230 return iommu_get_scsi_one(dev, vaddr, len);
233 static void iommu_get_scsi_sgl_gflush(struct device *dev, struct scatterlist *sg, int sz)
237 flush_page_for_dma(0);
240 n = (sg->length + sg->offset + PAGE_SIZE-1) >> PAGE_SHIFT;
241 sg->dma_address = iommu_get_one(dev, sg_page(sg), n) + sg->offset;
242 sg->dma_length = sg->length;
247 static void iommu_get_scsi_sgl_pflush(struct device *dev, struct scatterlist *sg, int sz)
249 unsigned long page, oldpage = 0;
255 n = (sg->length + sg->offset + PAGE_SIZE-1) >> PAGE_SHIFT;
258 * We expect unmapped highmem pages to be not in the cache.
259 * XXX Is this a good assumption?
260 * XXX What if someone else unmaps it here and races us?
262 if ((page = (unsigned long) page_address(sg_page(sg))) != 0) {
263 for (i = 0; i < n; i++) {
264 if (page != oldpage) { /* Already flushed? */
265 flush_page_for_dma(page);
272 sg->dma_address = iommu_get_one(dev, sg_page(sg), n) + sg->offset;
273 sg->dma_length = sg->length;
278 static void iommu_release_one(struct device *dev, u32 busa, int npages)
280 struct iommu_struct *iommu = dev->archdata.iommu;
284 BUG_ON(busa < iommu->start);
285 ioptex = (busa - iommu->start) >> PAGE_SHIFT;
286 for (i = 0; i < npages; i++) {
287 iopte_val(iommu->page_table[ioptex + i]) = 0;
288 iommu_invalidate_page(iommu->regs, busa);
291 bit_map_clear(&iommu->usemap, ioptex, npages);
294 static void iommu_release_scsi_one(struct device *dev, __u32 vaddr, unsigned long len)
299 off = vaddr & ~PAGE_MASK;
300 npages = (off + len + PAGE_SIZE-1) >> PAGE_SHIFT;
301 iommu_release_one(dev, vaddr & PAGE_MASK, npages);
304 static void iommu_release_scsi_sgl(struct device *dev, struct scatterlist *sg, int sz)
311 n = (sg->length + sg->offset + PAGE_SIZE-1) >> PAGE_SHIFT;
312 iommu_release_one(dev, sg->dma_address & PAGE_MASK, n);
313 sg->dma_address = 0x21212121;
319 static int iommu_map_dma_area(struct device *dev, dma_addr_t *pba, unsigned long va,
320 unsigned long addr, int len)
322 struct iommu_struct *iommu = dev->archdata.iommu;
323 unsigned long page, end;
324 iopte_t *iopte = iommu->page_table;
328 BUG_ON((va & ~PAGE_MASK) != 0);
329 BUG_ON((addr & ~PAGE_MASK) != 0);
330 BUG_ON((len & ~PAGE_MASK) != 0);
332 /* page color = physical address */
333 ioptex = bit_map_string_get(&iommu->usemap, len >> PAGE_SHIFT,
348 if (viking_mxcc_present)
349 viking_mxcc_flush_page(page);
350 else if (viking_flush)
351 viking_flush_page(page);
353 __flush_page_to_ram(page);
355 pgdp = pgd_offset(&init_mm, addr);
356 pmdp = pmd_offset(pgdp, addr);
357 ptep = pte_offset_map(pmdp, addr);
359 set_pte(ptep, mk_pte(virt_to_page(page), dvma_prot));
361 iopte_val(*iopte++) =
362 MKIOPTE(page_to_pfn(virt_to_page(page)), ioperm_noc);
366 /* P3: why do we need this?
368 * DAVEM: Because there are several aspects, none of which
369 * are handled by a single interface. Some cpus are
370 * completely not I/O DMA coherent, and some have
371 * virtually indexed caches. The driver DMA flushing
372 * methods handle the former case, but here during
373 * IOMMU page table modifications, and usage of non-cacheable
374 * cpu mappings of pages potentially in the cpu caches, we have
375 * to handle the latter case as well.
378 iommu_flush_iotlb(first, len >> PAGE_SHIFT);
380 iommu_invalidate(iommu->regs);
382 *pba = iommu->start + (ioptex << PAGE_SHIFT);
386 static void iommu_unmap_dma_area(struct device *dev, unsigned long busa, int len)
388 struct iommu_struct *iommu = dev->archdata.iommu;
389 iopte_t *iopte = iommu->page_table;
391 int ioptex = (busa - iommu->start) >> PAGE_SHIFT;
393 BUG_ON((busa & ~PAGE_MASK) != 0);
394 BUG_ON((len & ~PAGE_MASK) != 0);
399 iopte_val(*iopte++) = 0;
403 iommu_invalidate(iommu->regs);
404 bit_map_clear(&iommu->usemap, ioptex, len >> PAGE_SHIFT);
408 static const struct sparc32_dma_ops iommu_dma_gflush_ops = {
409 .get_scsi_one = iommu_get_scsi_one_gflush,
410 .get_scsi_sgl = iommu_get_scsi_sgl_gflush,
411 .release_scsi_one = iommu_release_scsi_one,
412 .release_scsi_sgl = iommu_release_scsi_sgl,
414 .map_dma_area = iommu_map_dma_area,
415 .unmap_dma_area = iommu_unmap_dma_area,
419 static const struct sparc32_dma_ops iommu_dma_pflush_ops = {
420 .get_scsi_one = iommu_get_scsi_one_pflush,
421 .get_scsi_sgl = iommu_get_scsi_sgl_pflush,
422 .release_scsi_one = iommu_release_scsi_one,
423 .release_scsi_sgl = iommu_release_scsi_sgl,
425 .map_dma_area = iommu_map_dma_area,
426 .unmap_dma_area = iommu_unmap_dma_area,
430 void __init ld_mmu_iommu(void)
432 if (flush_page_for_dma_global) {
433 /* flush_page_for_dma flushes everything, no matter of what page is it */
434 sparc32_dma_ops = &iommu_dma_gflush_ops;
436 sparc32_dma_ops = &iommu_dma_pflush_ops;
439 if (viking_mxcc_present || srmmu_modtype == HyperSparc) {
440 dvma_prot = __pgprot(SRMMU_CACHE | SRMMU_ET_PTE | SRMMU_PRIV);
441 ioperm_noc = IOPTE_CACHE | IOPTE_WRITE | IOPTE_VALID;
443 dvma_prot = __pgprot(SRMMU_ET_PTE | SRMMU_PRIV);
444 ioperm_noc = IOPTE_WRITE | IOPTE_VALID;