xref: /openbmc/linux/arch/sparc/mm/iommu.c (revision 23c2b932)
1 /*
2  * iommu.c:  IOMMU specific routines for memory management.
3  *
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)
8  */
9 
10 #include <linux/kernel.h>
11 #include <linux/init.h>
12 #include <linux/mm.h>
13 #include <linux/slab.h>
14 #include <linux/highmem.h>	/* pte_offset_map => kmap_atomic */
15 #include <linux/scatterlist.h>
16 #include <linux/of.h>
17 #include <linux/of_device.h>
18 
19 #include <asm/pgalloc.h>
20 #include <asm/pgtable.h>
21 #include <asm/io.h>
22 #include <asm/mxcc.h>
23 #include <asm/mbus.h>
24 #include <asm/cacheflush.h>
25 #include <asm/tlbflush.h>
26 #include <asm/bitext.h>
27 #include <asm/iommu.h>
28 #include <asm/dma.h>
29 
30 #include "mm_32.h"
31 
32 /*
33  * This can be sized dynamically, but we will do this
34  * only when we have a guidance about actual I/O pressures.
35  */
36 #define IOMMU_RNGE	IOMMU_RNGE_256MB
37 #define IOMMU_START	0xF0000000
38 #define IOMMU_WINSIZE	(256*1024*1024U)
39 #define IOMMU_NPTES	(IOMMU_WINSIZE/PAGE_SIZE)	/* 64K PTEs, 256KB */
40 #define IOMMU_ORDER	6				/* 4096 * (1<<6) */
41 
42 static int viking_flush;
43 /* viking.S */
44 extern void viking_flush_page(unsigned long page);
45 extern void viking_mxcc_flush_page(unsigned long page);
46 
47 /*
48  * Values precomputed according to CPU type.
49  */
50 static unsigned int ioperm_noc;		/* Consistent mapping iopte flags */
51 static pgprot_t dvma_prot;		/* Consistent mapping pte flags */
52 
53 #define IOPERM        (IOPTE_CACHE | IOPTE_WRITE | IOPTE_VALID)
54 #define MKIOPTE(pfn, perm) (((((pfn)<<8) & IOPTE_PAGE) | (perm)) & ~IOPTE_WAZ)
55 
56 static void __init sbus_iommu_init(struct platform_device *op)
57 {
58 	struct iommu_struct *iommu;
59 	unsigned int impl, vers;
60 	unsigned long *bitmap;
61 	unsigned long control;
62 	unsigned long base;
63 	unsigned long tmp;
64 
65 	iommu = kmalloc(sizeof(struct iommu_struct), GFP_KERNEL);
66 	if (!iommu) {
67 		prom_printf("Unable to allocate iommu structure\n");
68 		prom_halt();
69 	}
70 
71 	iommu->regs = of_ioremap(&op->resource[0], 0, PAGE_SIZE * 3,
72 				 "iommu_regs");
73 	if (!iommu->regs) {
74 		prom_printf("Cannot map IOMMU registers\n");
75 		prom_halt();
76 	}
77 
78 	control = sbus_readl(&iommu->regs->control);
79 	impl = (control & IOMMU_CTRL_IMPL) >> 28;
80 	vers = (control & IOMMU_CTRL_VERS) >> 24;
81 	control &= ~(IOMMU_CTRL_RNGE);
82 	control |= (IOMMU_RNGE_256MB | IOMMU_CTRL_ENAB);
83 	sbus_writel(control, &iommu->regs->control);
84 
85 	iommu_invalidate(iommu->regs);
86 	iommu->start = IOMMU_START;
87 	iommu->end = 0xffffffff;
88 
89 	/* Allocate IOMMU page table */
90 	/* Stupid alignment constraints give me a headache.
91 	   We need 256K or 512K or 1M or 2M area aligned to
92            its size and current gfp will fortunately give
93            it to us. */
94         tmp = __get_free_pages(GFP_KERNEL, IOMMU_ORDER);
95 	if (!tmp) {
96 		prom_printf("Unable to allocate iommu table [0x%lx]\n",
97 			    IOMMU_NPTES * sizeof(iopte_t));
98 		prom_halt();
99 	}
100 	iommu->page_table = (iopte_t *)tmp;
101 
102 	/* Initialize new table. */
103 	memset(iommu->page_table, 0, IOMMU_NPTES*sizeof(iopte_t));
104 	flush_cache_all();
105 	flush_tlb_all();
106 
107 	base = __pa((unsigned long)iommu->page_table) >> 4;
108 	sbus_writel(base, &iommu->regs->base);
109 	iommu_invalidate(iommu->regs);
110 
111 	bitmap = kmalloc(IOMMU_NPTES>>3, GFP_KERNEL);
112 	if (!bitmap) {
113 		prom_printf("Unable to allocate iommu bitmap [%d]\n",
114 			    (int)(IOMMU_NPTES>>3));
115 		prom_halt();
116 	}
117 	bit_map_init(&iommu->usemap, bitmap, IOMMU_NPTES);
118 	/* To be coherent on HyperSparc, the page color of DVMA
119 	 * and physical addresses must match.
120 	 */
121 	if (srmmu_modtype == HyperSparc)
122 		iommu->usemap.num_colors = vac_cache_size >> PAGE_SHIFT;
123 	else
124 		iommu->usemap.num_colors = 1;
125 
126 	printk(KERN_INFO "IOMMU: impl %d vers %d table 0x%p[%d B] map [%d b]\n",
127 	       impl, vers, iommu->page_table,
128 	       (int)(IOMMU_NPTES*sizeof(iopte_t)), (int)IOMMU_NPTES);
129 
130 	op->dev.archdata.iommu = iommu;
131 }
132 
133 static int __init iommu_init(void)
134 {
135 	struct device_node *dp;
136 
137 	for_each_node_by_name(dp, "iommu") {
138 		struct platform_device *op = of_find_device_by_node(dp);
139 
140 		sbus_iommu_init(op);
141 		of_propagate_archdata(op);
142 	}
143 
144 	return 0;
145 }
146 
147 subsys_initcall(iommu_init);
148 
149 /* Flush the iotlb entries to ram. */
150 /* This could be better if we didn't have to flush whole pages. */
151 static void iommu_flush_iotlb(iopte_t *iopte, unsigned int niopte)
152 {
153 	unsigned long start;
154 	unsigned long end;
155 
156 	start = (unsigned long)iopte;
157 	end = PAGE_ALIGN(start + niopte*sizeof(iopte_t));
158 	start &= PAGE_MASK;
159 	if (viking_mxcc_present) {
160 		while(start < end) {
161 			viking_mxcc_flush_page(start);
162 			start += PAGE_SIZE;
163 		}
164 	} else if (viking_flush) {
165 		while(start < end) {
166 			viking_flush_page(start);
167 			start += PAGE_SIZE;
168 		}
169 	} else {
170 		while(start < end) {
171 			__flush_page_to_ram(start);
172 			start += PAGE_SIZE;
173 		}
174 	}
175 }
176 
177 static u32 iommu_get_one(struct device *dev, struct page *page, int npages)
178 {
179 	struct iommu_struct *iommu = dev->archdata.iommu;
180 	int ioptex;
181 	iopte_t *iopte, *iopte0;
182 	unsigned int busa, busa0;
183 	int i;
184 
185 	/* page color = pfn of page */
186 	ioptex = bit_map_string_get(&iommu->usemap, npages, page_to_pfn(page));
187 	if (ioptex < 0)
188 		panic("iommu out");
189 	busa0 = iommu->start + (ioptex << PAGE_SHIFT);
190 	iopte0 = &iommu->page_table[ioptex];
191 
192 	busa = busa0;
193 	iopte = iopte0;
194 	for (i = 0; i < npages; i++) {
195 		iopte_val(*iopte) = MKIOPTE(page_to_pfn(page), IOPERM);
196 		iommu_invalidate_page(iommu->regs, busa);
197 		busa += PAGE_SIZE;
198 		iopte++;
199 		page++;
200 	}
201 
202 	iommu_flush_iotlb(iopte0, npages);
203 
204 	return busa0;
205 }
206 
207 static u32 iommu_get_scsi_one(struct device *dev, char *vaddr, unsigned int len)
208 {
209 	unsigned long off;
210 	int npages;
211 	struct page *page;
212 	u32 busa;
213 
214 	off = (unsigned long)vaddr & ~PAGE_MASK;
215 	npages = (off + len + PAGE_SIZE-1) >> PAGE_SHIFT;
216 	page = virt_to_page((unsigned long)vaddr & PAGE_MASK);
217 	busa = iommu_get_one(dev, page, npages);
218 	return busa + off;
219 }
220 
221 static __u32 iommu_get_scsi_one_gflush(struct device *dev, char *vaddr, unsigned long len)
222 {
223 	flush_page_for_dma(0);
224 	return iommu_get_scsi_one(dev, vaddr, len);
225 }
226 
227 static __u32 iommu_get_scsi_one_pflush(struct device *dev, char *vaddr, unsigned long len)
228 {
229 	unsigned long page = ((unsigned long) vaddr) & PAGE_MASK;
230 
231 	while(page < ((unsigned long)(vaddr + len))) {
232 		flush_page_for_dma(page);
233 		page += PAGE_SIZE;
234 	}
235 	return iommu_get_scsi_one(dev, vaddr, len);
236 }
237 
238 static void iommu_get_scsi_sgl_gflush(struct device *dev, struct scatterlist *sg, int sz)
239 {
240 	int n;
241 
242 	flush_page_for_dma(0);
243 	while (sz != 0) {
244 		--sz;
245 		n = (sg->length + sg->offset + PAGE_SIZE-1) >> PAGE_SHIFT;
246 		sg->dma_address = iommu_get_one(dev, sg_page(sg), n) + sg->offset;
247 		sg->dma_length = sg->length;
248 		sg = sg_next(sg);
249 	}
250 }
251 
252 static void iommu_get_scsi_sgl_pflush(struct device *dev, struct scatterlist *sg, int sz)
253 {
254 	unsigned long page, oldpage = 0;
255 	int n, i;
256 
257 	while(sz != 0) {
258 		--sz;
259 
260 		n = (sg->length + sg->offset + PAGE_SIZE-1) >> PAGE_SHIFT;
261 
262 		/*
263 		 * We expect unmapped highmem pages to be not in the cache.
264 		 * XXX Is this a good assumption?
265 		 * XXX What if someone else unmaps it here and races us?
266 		 */
267 		if ((page = (unsigned long) page_address(sg_page(sg))) != 0) {
268 			for (i = 0; i < n; i++) {
269 				if (page != oldpage) {	/* Already flushed? */
270 					flush_page_for_dma(page);
271 					oldpage = page;
272 				}
273 				page += PAGE_SIZE;
274 			}
275 		}
276 
277 		sg->dma_address = iommu_get_one(dev, sg_page(sg), n) + sg->offset;
278 		sg->dma_length = sg->length;
279 		sg = sg_next(sg);
280 	}
281 }
282 
283 static void iommu_release_one(struct device *dev, u32 busa, int npages)
284 {
285 	struct iommu_struct *iommu = dev->archdata.iommu;
286 	int ioptex;
287 	int i;
288 
289 	BUG_ON(busa < iommu->start);
290 	ioptex = (busa - iommu->start) >> PAGE_SHIFT;
291 	for (i = 0; i < npages; i++) {
292 		iopte_val(iommu->page_table[ioptex + i]) = 0;
293 		iommu_invalidate_page(iommu->regs, busa);
294 		busa += PAGE_SIZE;
295 	}
296 	bit_map_clear(&iommu->usemap, ioptex, npages);
297 }
298 
299 static void iommu_release_scsi_one(struct device *dev, __u32 vaddr, unsigned long len)
300 {
301 	unsigned long off;
302 	int npages;
303 
304 	off = vaddr & ~PAGE_MASK;
305 	npages = (off + len + PAGE_SIZE-1) >> PAGE_SHIFT;
306 	iommu_release_one(dev, vaddr & PAGE_MASK, npages);
307 }
308 
309 static void iommu_release_scsi_sgl(struct device *dev, struct scatterlist *sg, int sz)
310 {
311 	int n;
312 
313 	while(sz != 0) {
314 		--sz;
315 
316 		n = (sg->length + sg->offset + PAGE_SIZE-1) >> PAGE_SHIFT;
317 		iommu_release_one(dev, sg->dma_address & PAGE_MASK, n);
318 		sg->dma_address = 0x21212121;
319 		sg = sg_next(sg);
320 	}
321 }
322 
323 #ifdef CONFIG_SBUS
324 static int iommu_map_dma_area(struct device *dev, dma_addr_t *pba, unsigned long va,
325 			      unsigned long addr, int len)
326 {
327 	struct iommu_struct *iommu = dev->archdata.iommu;
328 	unsigned long page, end;
329 	iopte_t *iopte = iommu->page_table;
330 	iopte_t *first;
331 	int ioptex;
332 
333 	BUG_ON((va & ~PAGE_MASK) != 0);
334 	BUG_ON((addr & ~PAGE_MASK) != 0);
335 	BUG_ON((len & ~PAGE_MASK) != 0);
336 
337 	/* page color = physical address */
338 	ioptex = bit_map_string_get(&iommu->usemap, len >> PAGE_SHIFT,
339 		addr >> PAGE_SHIFT);
340 	if (ioptex < 0)
341 		panic("iommu out");
342 
343 	iopte += ioptex;
344 	first = iopte;
345 	end = addr + len;
346 	while(addr < end) {
347 		page = va;
348 		{
349 			pgd_t *pgdp;
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 			pgdp = pgd_offset(&init_mm, addr);
361 			pmdp = pmd_offset(pgdp, addr);
362 			ptep = pte_offset_map(pmdp, addr);
363 
364 			set_pte(ptep, mk_pte(virt_to_page(page), dvma_prot));
365 		}
366 		iopte_val(*iopte++) =
367 		    MKIOPTE(page_to_pfn(virt_to_page(page)), ioperm_noc);
368 		addr += PAGE_SIZE;
369 		va += PAGE_SIZE;
370 	}
371 	/* P3: why do we need this?
372 	 *
373 	 * DAVEM: Because there are several aspects, none of which
374 	 *        are handled by a single interface.  Some cpus are
375 	 *        completely not I/O DMA coherent, and some have
376 	 *        virtually indexed caches.  The driver DMA flushing
377 	 *        methods handle the former case, but here during
378 	 *        IOMMU page table modifications, and usage of non-cacheable
379 	 *        cpu mappings of pages potentially in the cpu caches, we have
380 	 *        to handle the latter case as well.
381 	 */
382 	flush_cache_all();
383 	iommu_flush_iotlb(first, len >> PAGE_SHIFT);
384 	flush_tlb_all();
385 	iommu_invalidate(iommu->regs);
386 
387 	*pba = iommu->start + (ioptex << PAGE_SHIFT);
388 	return 0;
389 }
390 
391 static void iommu_unmap_dma_area(struct device *dev, unsigned long busa, int len)
392 {
393 	struct iommu_struct *iommu = dev->archdata.iommu;
394 	iopte_t *iopte = iommu->page_table;
395 	unsigned long end;
396 	int ioptex = (busa - iommu->start) >> PAGE_SHIFT;
397 
398 	BUG_ON((busa & ~PAGE_MASK) != 0);
399 	BUG_ON((len & ~PAGE_MASK) != 0);
400 
401 	iopte += ioptex;
402 	end = busa + len;
403 	while (busa < end) {
404 		iopte_val(*iopte++) = 0;
405 		busa += PAGE_SIZE;
406 	}
407 	flush_tlb_all();
408 	iommu_invalidate(iommu->regs);
409 	bit_map_clear(&iommu->usemap, ioptex, len >> PAGE_SHIFT);
410 }
411 #endif
412 
413 static const struct sparc32_dma_ops iommu_dma_gflush_ops = {
414 	.get_scsi_one		= iommu_get_scsi_one_gflush,
415 	.get_scsi_sgl		= iommu_get_scsi_sgl_gflush,
416 	.release_scsi_one	= iommu_release_scsi_one,
417 	.release_scsi_sgl	= iommu_release_scsi_sgl,
418 #ifdef CONFIG_SBUS
419 	.map_dma_area		= iommu_map_dma_area,
420 	.unmap_dma_area		= iommu_unmap_dma_area,
421 #endif
422 };
423 
424 static const struct sparc32_dma_ops iommu_dma_pflush_ops = {
425 	.get_scsi_one		= iommu_get_scsi_one_pflush,
426 	.get_scsi_sgl		= iommu_get_scsi_sgl_pflush,
427 	.release_scsi_one	= iommu_release_scsi_one,
428 	.release_scsi_sgl	= iommu_release_scsi_sgl,
429 #ifdef CONFIG_SBUS
430 	.map_dma_area		= iommu_map_dma_area,
431 	.unmap_dma_area		= iommu_unmap_dma_area,
432 #endif
433 };
434 
435 void __init ld_mmu_iommu(void)
436 {
437 	if (flush_page_for_dma_global) {
438 		/* flush_page_for_dma flushes everything, no matter of what page is it */
439 		sparc32_dma_ops = &iommu_dma_gflush_ops;
440 	} else {
441 		sparc32_dma_ops = &iommu_dma_pflush_ops;
442 	}
443 
444 	if (viking_mxcc_present || srmmu_modtype == HyperSparc) {
445 		dvma_prot = __pgprot(SRMMU_CACHE | SRMMU_ET_PTE | SRMMU_PRIV);
446 		ioperm_noc = IOPTE_CACHE | IOPTE_WRITE | IOPTE_VALID;
447 	} else {
448 		dvma_prot = __pgprot(SRMMU_ET_PTE | SRMMU_PRIV);
449 		ioperm_noc = IOPTE_WRITE | IOPTE_VALID;
450 	}
451 }
452