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/config.h>
11 #include <linux/kernel.h>
12 #include <linux/init.h>
14 #include <linux/slab.h>
15 #include <linux/highmem.h> /* pte_offset_map => kmap_atomic */
17 #include <asm/scatterlist.h>
18 #include <asm/pgalloc.h>
19 #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, 265KB */
38 #define IOMMU_ORDER 6 /* 4096 * (1<<6) */
41 extern int viking_mxcc_present
;
42 BTFIXUPDEF_CALL(void, flush_page_for_dma
, unsigned long)
43 #define flush_page_for_dma(page) BTFIXUP_CALL(flush_page_for_dma)(page)
44 extern int flush_page_for_dma_global
;
45 static int viking_flush
;
47 extern void viking_flush_page(unsigned long page
);
48 extern void viking_mxcc_flush_page(unsigned long page
);
51 * Values precomputed according to CPU type.
53 static unsigned int ioperm_noc
; /* Consistent mapping iopte flags */
54 static pgprot_t dvma_prot
; /* Consistent mapping pte flags */
56 #define IOPERM (IOPTE_CACHE | IOPTE_WRITE | IOPTE_VALID)
57 #define MKIOPTE(pfn, perm) (((((pfn)<<8) & IOPTE_PAGE) | (perm)) & ~IOPTE_WAZ)
60 iommu_init(int iommund
, struct sbus_bus
*sbus
)
62 unsigned int impl
, vers
;
64 struct iommu_struct
*iommu
;
65 struct linux_prom_registers iommu_promregs
[PROMREG_MAX
];
67 unsigned long *bitmap
;
69 iommu
= kmalloc(sizeof(struct iommu_struct
), GFP_ATOMIC
);
71 prom_printf("Unable to allocate iommu structure\n");
75 if (prom_getproperty(iommund
, "reg", (void *) iommu_promregs
,
76 sizeof(iommu_promregs
)) != -1) {
77 memset(&r
, 0, sizeof(r
));
78 r
.flags
= iommu_promregs
[0].which_io
;
79 r
.start
= iommu_promregs
[0].phys_addr
;
80 iommu
->regs
= (struct iommu_regs
*)
81 sbus_ioremap(&r
, 0, PAGE_SIZE
* 3, "iommu_regs");
84 prom_printf("Cannot map IOMMU registers\n");
87 impl
= (iommu
->regs
->control
& IOMMU_CTRL_IMPL
) >> 28;
88 vers
= (iommu
->regs
->control
& IOMMU_CTRL_VERS
) >> 24;
89 tmp
= iommu
->regs
->control
;
90 tmp
&= ~(IOMMU_CTRL_RNGE
);
91 tmp
|= (IOMMU_RNGE_256MB
| IOMMU_CTRL_ENAB
);
92 iommu
->regs
->control
= tmp
;
93 iommu_invalidate(iommu
->regs
);
94 iommu
->start
= IOMMU_START
;
95 iommu
->end
= 0xffffffff;
97 /* Allocate IOMMU page table */
98 /* Stupid alignment constraints give me a headache.
99 We need 256K or 512K or 1M or 2M area aligned to
100 its size and current gfp will fortunately give
102 tmp
= __get_free_pages(GFP_KERNEL
, IOMMU_ORDER
);
104 prom_printf("Unable to allocate iommu table [0x%08x]\n",
105 IOMMU_NPTES
*sizeof(iopte_t
));
108 iommu
->page_table
= (iopte_t
*)tmp
;
110 /* Initialize new table. */
111 memset(iommu
->page_table
, 0, IOMMU_NPTES
*sizeof(iopte_t
));
114 iommu
->regs
->base
= __pa((unsigned long) iommu
->page_table
) >> 4;
115 iommu_invalidate(iommu
->regs
);
117 bitmap
= kmalloc(IOMMU_NPTES
>>3, GFP_KERNEL
);
119 prom_printf("Unable to allocate iommu bitmap [%d]\n",
120 (int)(IOMMU_NPTES
>>3));
123 bit_map_init(&iommu
->usemap
, bitmap
, IOMMU_NPTES
);
124 /* To be coherent on HyperSparc, the page color of DVMA
125 * and physical addresses must match.
127 if (srmmu_modtype
== HyperSparc
)
128 iommu
->usemap
.num_colors
= vac_cache_size
>> PAGE_SHIFT
;
130 iommu
->usemap
.num_colors
= 1;
132 printk("IOMMU: impl %d vers %d table 0x%p[%d B] map [%d b]\n",
133 impl
, vers
, iommu
->page_table
,
134 (int)(IOMMU_NPTES
*sizeof(iopte_t
)), (int)IOMMU_NPTES
);
139 /* This begs to be btfixup-ed by srmmu. */
140 /* Flush the iotlb entries to ram. */
141 /* This could be better if we didn't have to flush whole pages. */
142 static void iommu_flush_iotlb(iopte_t
*iopte
, unsigned int niopte
)
147 start
= (unsigned long)iopte
& PAGE_MASK
;
148 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 if (busa
< iommu
->start
)
300 ioptex
= (busa
- iommu
->start
) >> PAGE_SHIFT
;
301 for (i
= 0; i
< npages
; i
++) {
302 iopte_val(iommu
->page_table
[ioptex
+ i
]) = 0;
303 iommu_invalidate_page(iommu
->regs
, busa
);
306 bit_map_clear(&iommu
->usemap
, ioptex
, npages
);
309 static void iommu_release_scsi_one(__u32 vaddr
, unsigned long len
, struct sbus_bus
*sbus
)
314 off
= vaddr
& ~PAGE_MASK
;
315 npages
= (off
+ len
+ PAGE_SIZE
-1) >> PAGE_SHIFT
;
316 iommu_release_one(vaddr
& PAGE_MASK
, npages
, sbus
);
319 static void iommu_release_scsi_sgl(struct scatterlist
*sg
, int sz
, struct sbus_bus
*sbus
)
326 n
= (sg
->length
+ sg
->offset
+ PAGE_SIZE
-1) >> PAGE_SHIFT
;
327 iommu_release_one(sg
->dvma_address
& PAGE_MASK
, n
, sbus
);
328 sg
->dvma_address
= 0x21212121;
334 static int iommu_map_dma_area(dma_addr_t
*pba
, unsigned long va
,
335 unsigned long addr
, int len
)
337 unsigned long page
, end
;
338 struct iommu_struct
*iommu
= sbus_root
->iommu
;
339 iopte_t
*iopte
= iommu
->page_table
;
343 if ((va
& ~PAGE_MASK
) != 0) BUG();
344 if ((addr
& ~PAGE_MASK
) != 0) BUG();
345 if ((len
& ~PAGE_MASK
) != 0) BUG();
347 /* page color = physical address */
348 ioptex
= bit_map_string_get(&iommu
->usemap
, len
>> PAGE_SHIFT
,
363 if (viking_mxcc_present
)
364 viking_mxcc_flush_page(page
);
365 else if (viking_flush
)
366 viking_flush_page(page
);
368 __flush_page_to_ram(page
);
370 pgdp
= pgd_offset(&init_mm
, addr
);
371 pmdp
= pmd_offset(pgdp
, addr
);
372 ptep
= pte_offset_map(pmdp
, addr
);
374 set_pte(ptep
, mk_pte(virt_to_page(page
), dvma_prot
));
376 iopte_val(*iopte
++) =
377 MKIOPTE(page_to_pfn(virt_to_page(page
)), ioperm_noc
);
381 /* P3: why do we need this?
383 * DAVEM: Because there are several aspects, none of which
384 * are handled by a single interface. Some cpus are
385 * completely not I/O DMA coherent, and some have
386 * virtually indexed caches. The driver DMA flushing
387 * methods handle the former case, but here during
388 * IOMMU page table modifications, and usage of non-cacheable
389 * cpu mappings of pages potentially in the cpu caches, we have
390 * to handle the latter case as well.
393 iommu_flush_iotlb(first
, len
>> PAGE_SHIFT
);
395 iommu_invalidate(iommu
->regs
);
397 *pba
= iommu
->start
+ (ioptex
<< PAGE_SHIFT
);
401 static void iommu_unmap_dma_area(unsigned long busa
, int len
)
403 struct iommu_struct
*iommu
= sbus_root
->iommu
;
404 iopte_t
*iopte
= iommu
->page_table
;
406 int ioptex
= (busa
- iommu
->start
) >> PAGE_SHIFT
;
408 if ((busa
& ~PAGE_MASK
) != 0) BUG();
409 if ((len
& ~PAGE_MASK
) != 0) BUG();
414 iopte_val(*iopte
++) = 0;
418 iommu_invalidate(iommu
->regs
);
419 bit_map_clear(&iommu
->usemap
, ioptex
, len
>> PAGE_SHIFT
);
422 static struct page
*iommu_translate_dvma(unsigned long busa
)
424 struct iommu_struct
*iommu
= sbus_root
->iommu
;
425 iopte_t
*iopte
= iommu
->page_table
;
427 iopte
+= ((busa
- iommu
->start
) >> PAGE_SHIFT
);
428 return pfn_to_page((iopte_val(*iopte
) & IOPTE_PAGE
) >> (PAGE_SHIFT
-4));
432 static char *iommu_lockarea(char *vaddr
, unsigned long len
)
437 static void iommu_unlockarea(char *vaddr
, unsigned long len
)
441 void __init
ld_mmu_iommu(void)
443 viking_flush
= (BTFIXUPVAL_CALL(flush_page_for_dma
) == (unsigned long)viking_flush_page
);
444 BTFIXUPSET_CALL(mmu_lockarea
, iommu_lockarea
, BTFIXUPCALL_RETO0
);
445 BTFIXUPSET_CALL(mmu_unlockarea
, iommu_unlockarea
, BTFIXUPCALL_NOP
);
447 if (!BTFIXUPVAL_CALL(flush_page_for_dma
)) {
448 /* IO coherent chip */
449 BTFIXUPSET_CALL(mmu_get_scsi_one
, iommu_get_scsi_one_noflush
, BTFIXUPCALL_RETO0
);
450 BTFIXUPSET_CALL(mmu_get_scsi_sgl
, iommu_get_scsi_sgl_noflush
, BTFIXUPCALL_NORM
);
451 } else if (flush_page_for_dma_global
) {
452 /* flush_page_for_dma flushes everything, no matter of what page is it */
453 BTFIXUPSET_CALL(mmu_get_scsi_one
, iommu_get_scsi_one_gflush
, BTFIXUPCALL_NORM
);
454 BTFIXUPSET_CALL(mmu_get_scsi_sgl
, iommu_get_scsi_sgl_gflush
, BTFIXUPCALL_NORM
);
456 BTFIXUPSET_CALL(mmu_get_scsi_one
, iommu_get_scsi_one_pflush
, BTFIXUPCALL_NORM
);
457 BTFIXUPSET_CALL(mmu_get_scsi_sgl
, iommu_get_scsi_sgl_pflush
, BTFIXUPCALL_NORM
);
459 BTFIXUPSET_CALL(mmu_release_scsi_one
, iommu_release_scsi_one
, BTFIXUPCALL_NORM
);
460 BTFIXUPSET_CALL(mmu_release_scsi_sgl
, iommu_release_scsi_sgl
, BTFIXUPCALL_NORM
);
463 BTFIXUPSET_CALL(mmu_map_dma_area
, iommu_map_dma_area
, BTFIXUPCALL_NORM
);
464 BTFIXUPSET_CALL(mmu_unmap_dma_area
, iommu_unmap_dma_area
, BTFIXUPCALL_NORM
);
465 BTFIXUPSET_CALL(mmu_translate_dvma
, iommu_translate_dvma
, BTFIXUPCALL_NORM
);
468 if (viking_mxcc_present
|| srmmu_modtype
== HyperSparc
) {
469 dvma_prot
= __pgprot(SRMMU_CACHE
| SRMMU_ET_PTE
| SRMMU_PRIV
);
470 ioperm_noc
= IOPTE_CACHE
| IOPTE_WRITE
| IOPTE_VALID
;
472 dvma_prot
= __pgprot(SRMMU_ET_PTE
| SRMMU_PRIV
);
473 ioperm_noc
= IOPTE_WRITE
| IOPTE_VALID
;