xref: /openbmc/linux/arch/sparc/mm/iommu.c (revision e23feb16)
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 /*
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 /* srmmu.c */
41 extern int viking_mxcc_present;
42 extern int flush_page_for_dma_global;
43 static int viking_flush;
44 /* viking.S */
45 extern void viking_flush_page(unsigned long page);
46 extern void viking_mxcc_flush_page(unsigned long page);
47 
48 /*
49  * Values precomputed according to CPU type.
50  */
51 static unsigned int ioperm_noc;		/* Consistent mapping iopte flags */
52 static pgprot_t dvma_prot;		/* Consistent mapping pte flags */
53 
54 #define IOPERM        (IOPTE_CACHE | IOPTE_WRITE | IOPTE_VALID)
55 #define MKIOPTE(pfn, perm) (((((pfn)<<8) & IOPTE_PAGE) | (perm)) & ~IOPTE_WAZ)
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 tmp;
63 
64 	iommu = kmalloc(sizeof(struct iommu_struct), GFP_KERNEL);
65 	if (!iommu) {
66 		prom_printf("Unable to allocate iommu structure\n");
67 		prom_halt();
68 	}
69 
70 	iommu->regs = of_ioremap(&op->resource[0], 0, PAGE_SIZE * 3,
71 				 "iommu_regs");
72 	if (!iommu->regs) {
73 		prom_printf("Cannot map IOMMU registers\n");
74 		prom_halt();
75 	}
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;
85 
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
90            it to us. */
91         tmp = __get_free_pages(GFP_KERNEL, IOMMU_ORDER);
92 	if (!tmp) {
93 		prom_printf("Unable to allocate iommu table [0x%lx]\n",
94 			    IOMMU_NPTES * sizeof(iopte_t));
95 		prom_halt();
96 	}
97 	iommu->page_table = (iopte_t *)tmp;
98 
99 	/* Initialize new table. */
100 	memset(iommu->page_table, 0, IOMMU_NPTES*sizeof(iopte_t));
101 	flush_cache_all();
102 	flush_tlb_all();
103 	iommu->regs->base = __pa((unsigned long) iommu->page_table) >> 4;
104 	iommu_invalidate(iommu->regs);
105 
106 	bitmap = kmalloc(IOMMU_NPTES>>3, GFP_KERNEL);
107 	if (!bitmap) {
108 		prom_printf("Unable to allocate iommu bitmap [%d]\n",
109 			    (int)(IOMMU_NPTES>>3));
110 		prom_halt();
111 	}
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.
115 	 */
116 	if (srmmu_modtype == HyperSparc)
117 		iommu->usemap.num_colors = vac_cache_size >> PAGE_SHIFT;
118 	else
119 		iommu->usemap.num_colors = 1;
120 
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);
124 
125 	op->dev.archdata.iommu = iommu;
126 }
127 
128 static int __init iommu_init(void)
129 {
130 	struct device_node *dp;
131 
132 	for_each_node_by_name(dp, "iommu") {
133 		struct platform_device *op = of_find_device_by_node(dp);
134 
135 		sbus_iommu_init(op);
136 		of_propagate_archdata(op);
137 	}
138 
139 	return 0;
140 }
141 
142 subsys_initcall(iommu_init);
143 
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)
147 {
148 	unsigned long start;
149 	unsigned long end;
150 
151 	start = (unsigned long)iopte;
152 	end = PAGE_ALIGN(start + niopte*sizeof(iopte_t));
153 	start &= PAGE_MASK;
154 	if (viking_mxcc_present) {
155 		while(start < end) {
156 			viking_mxcc_flush_page(start);
157 			start += PAGE_SIZE;
158 		}
159 	} else if (viking_flush) {
160 		while(start < end) {
161 			viking_flush_page(start);
162 			start += PAGE_SIZE;
163 		}
164 	} else {
165 		while(start < end) {
166 			__flush_page_to_ram(start);
167 			start += PAGE_SIZE;
168 		}
169 	}
170 }
171 
172 static u32 iommu_get_one(struct device *dev, struct page *page, int npages)
173 {
174 	struct iommu_struct *iommu = dev->archdata.iommu;
175 	int ioptex;
176 	iopte_t *iopte, *iopte0;
177 	unsigned int busa, busa0;
178 	int i;
179 
180 	/* page color = pfn of page */
181 	ioptex = bit_map_string_get(&iommu->usemap, npages, page_to_pfn(page));
182 	if (ioptex < 0)
183 		panic("iommu out");
184 	busa0 = iommu->start + (ioptex << PAGE_SHIFT);
185 	iopte0 = &iommu->page_table[ioptex];
186 
187 	busa = busa0;
188 	iopte = iopte0;
189 	for (i = 0; i < npages; i++) {
190 		iopte_val(*iopte) = MKIOPTE(page_to_pfn(page), IOPERM);
191 		iommu_invalidate_page(iommu->regs, busa);
192 		busa += PAGE_SIZE;
193 		iopte++;
194 		page++;
195 	}
196 
197 	iommu_flush_iotlb(iopte0, npages);
198 
199 	return busa0;
200 }
201 
202 static u32 iommu_get_scsi_one(struct device *dev, char *vaddr, unsigned int len)
203 {
204 	unsigned long off;
205 	int npages;
206 	struct page *page;
207 	u32 busa;
208 
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);
213 	return busa + off;
214 }
215 
216 static __u32 iommu_get_scsi_one_gflush(struct device *dev, char *vaddr, unsigned long len)
217 {
218 	flush_page_for_dma(0);
219 	return iommu_get_scsi_one(dev, vaddr, len);
220 }
221 
222 static __u32 iommu_get_scsi_one_pflush(struct device *dev, char *vaddr, unsigned long len)
223 {
224 	unsigned long page = ((unsigned long) vaddr) & PAGE_MASK;
225 
226 	while(page < ((unsigned long)(vaddr + len))) {
227 		flush_page_for_dma(page);
228 		page += PAGE_SIZE;
229 	}
230 	return iommu_get_scsi_one(dev, vaddr, len);
231 }
232 
233 static void iommu_get_scsi_sgl_gflush(struct device *dev, struct scatterlist *sg, int sz)
234 {
235 	int n;
236 
237 	flush_page_for_dma(0);
238 	while (sz != 0) {
239 		--sz;
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;
243 		sg = sg_next(sg);
244 	}
245 }
246 
247 static void iommu_get_scsi_sgl_pflush(struct device *dev, struct scatterlist *sg, int sz)
248 {
249 	unsigned long page, oldpage = 0;
250 	int n, i;
251 
252 	while(sz != 0) {
253 		--sz;
254 
255 		n = (sg->length + sg->offset + PAGE_SIZE-1) >> PAGE_SHIFT;
256 
257 		/*
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?
261 		 */
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);
266 					oldpage = page;
267 				}
268 				page += PAGE_SIZE;
269 			}
270 		}
271 
272 		sg->dma_address = iommu_get_one(dev, sg_page(sg), n) + sg->offset;
273 		sg->dma_length = sg->length;
274 		sg = sg_next(sg);
275 	}
276 }
277 
278 static void iommu_release_one(struct device *dev, u32 busa, int npages)
279 {
280 	struct iommu_struct *iommu = dev->archdata.iommu;
281 	int ioptex;
282 	int i;
283 
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);
289 		busa += PAGE_SIZE;
290 	}
291 	bit_map_clear(&iommu->usemap, ioptex, npages);
292 }
293 
294 static void iommu_release_scsi_one(struct device *dev, __u32 vaddr, unsigned long len)
295 {
296 	unsigned long off;
297 	int npages;
298 
299 	off = vaddr & ~PAGE_MASK;
300 	npages = (off + len + PAGE_SIZE-1) >> PAGE_SHIFT;
301 	iommu_release_one(dev, vaddr & PAGE_MASK, npages);
302 }
303 
304 static void iommu_release_scsi_sgl(struct device *dev, struct scatterlist *sg, int sz)
305 {
306 	int n;
307 
308 	while(sz != 0) {
309 		--sz;
310 
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;
314 		sg = sg_next(sg);
315 	}
316 }
317 
318 #ifdef CONFIG_SBUS
319 static int iommu_map_dma_area(struct device *dev, dma_addr_t *pba, unsigned long va,
320 			      unsigned long addr, int len)
321 {
322 	struct iommu_struct *iommu = dev->archdata.iommu;
323 	unsigned long page, end;
324 	iopte_t *iopte = iommu->page_table;
325 	iopte_t *first;
326 	int ioptex;
327 
328 	BUG_ON((va & ~PAGE_MASK) != 0);
329 	BUG_ON((addr & ~PAGE_MASK) != 0);
330 	BUG_ON((len & ~PAGE_MASK) != 0);
331 
332 	/* page color = physical address */
333 	ioptex = bit_map_string_get(&iommu->usemap, len >> PAGE_SHIFT,
334 		addr >> PAGE_SHIFT);
335 	if (ioptex < 0)
336 		panic("iommu out");
337 
338 	iopte += ioptex;
339 	first = iopte;
340 	end = addr + len;
341 	while(addr < end) {
342 		page = va;
343 		{
344 			pgd_t *pgdp;
345 			pmd_t *pmdp;
346 			pte_t *ptep;
347 
348 			if (viking_mxcc_present)
349 				viking_mxcc_flush_page(page);
350 			else if (viking_flush)
351 				viking_flush_page(page);
352 			else
353 				__flush_page_to_ram(page);
354 
355 			pgdp = pgd_offset(&init_mm, addr);
356 			pmdp = pmd_offset(pgdp, addr);
357 			ptep = pte_offset_map(pmdp, addr);
358 
359 			set_pte(ptep, mk_pte(virt_to_page(page), dvma_prot));
360 		}
361 		iopte_val(*iopte++) =
362 		    MKIOPTE(page_to_pfn(virt_to_page(page)), ioperm_noc);
363 		addr += PAGE_SIZE;
364 		va += PAGE_SIZE;
365 	}
366 	/* P3: why do we need this?
367 	 *
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.
376 	 */
377 	flush_cache_all();
378 	iommu_flush_iotlb(first, len >> PAGE_SHIFT);
379 	flush_tlb_all();
380 	iommu_invalidate(iommu->regs);
381 
382 	*pba = iommu->start + (ioptex << PAGE_SHIFT);
383 	return 0;
384 }
385 
386 static void iommu_unmap_dma_area(struct device *dev, unsigned long busa, int len)
387 {
388 	struct iommu_struct *iommu = dev->archdata.iommu;
389 	iopte_t *iopte = iommu->page_table;
390 	unsigned long end;
391 	int ioptex = (busa - iommu->start) >> PAGE_SHIFT;
392 
393 	BUG_ON((busa & ~PAGE_MASK) != 0);
394 	BUG_ON((len & ~PAGE_MASK) != 0);
395 
396 	iopte += ioptex;
397 	end = busa + len;
398 	while (busa < end) {
399 		iopte_val(*iopte++) = 0;
400 		busa += PAGE_SIZE;
401 	}
402 	flush_tlb_all();
403 	iommu_invalidate(iommu->regs);
404 	bit_map_clear(&iommu->usemap, ioptex, len >> PAGE_SHIFT);
405 }
406 #endif
407 
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,
413 #ifdef CONFIG_SBUS
414 	.map_dma_area		= iommu_map_dma_area,
415 	.unmap_dma_area		= iommu_unmap_dma_area,
416 #endif
417 };
418 
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,
424 #ifdef CONFIG_SBUS
425 	.map_dma_area		= iommu_map_dma_area,
426 	.unmap_dma_area		= iommu_unmap_dma_area,
427 #endif
428 };
429 
430 void __init ld_mmu_iommu(void)
431 {
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;
435 	} else {
436 		sparc32_dma_ops = &iommu_dma_pflush_ops;
437 	}
438 
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;
442 	} else {
443 		dvma_prot = __pgprot(SRMMU_ET_PTE | SRMMU_PRIV);
444 		ioperm_noc = IOPTE_WRITE | IOPTE_VALID;
445 	}
446 }
447