2 * IOMMU implementation for Cell Broadband Processor Architecture
3 * We just establish a linear mapping at boot by setting all the
4 * IOPT cache entries in the CPU.
5 * The mapping functions should be identical to pci_direct_iommu,
6 * except for the handling of the high order bit that is required
7 * by the Spider bridge. These should be split into a separate
8 * file at the point where we get a different bridge chip.
10 * Copyright (C) 2005 IBM Deutschland Entwicklung GmbH,
11 * Arnd Bergmann <arndb@de.ibm.com>
13 * Based on linear mapping
14 * Copyright (C) 2003 Benjamin Herrenschmidt (benh@kernel.crashing.org)
16 * This program is free software; you can redistribute it and/or
17 * modify it under the terms of the GNU General Public License
18 * as published by the Free Software Foundation; either version
19 * 2 of the License, or (at your option) any later version.
24 #include <linux/kernel.h>
25 #include <linux/pci.h>
26 #include <linux/delay.h>
27 #include <linux/string.h>
28 #include <linux/init.h>
29 #include <linux/bootmem.h>
31 #include <linux/dma-mapping.h>
33 #include <asm/sections.h>
34 #include <asm/iommu.h>
37 #include <asm/pci-bridge.h>
38 #include <asm/machdep.h>
39 #include <asm/pmac_feature.h>
40 #include <asm/abs_addr.h>
41 #include <asm/system.h>
42 #include <asm/ppc-pci.h>
46 static inline unsigned long
47 get_iopt_entry(unsigned long real_address, unsigned long ioid,
50 return (prot & IOPT_PROT_MASK)
53 | (real_address & IOPT_RPN_MASK)
54 | (ioid & IOPT_IOID_MASK);
62 mk_ioste(unsigned long val)
64 ioste ioste = { .val = val, };
69 get_iost_entry(unsigned long iopt_base, unsigned long io_address, unsigned page_size)
79 nnpt = 0; /* one page per segment */
80 shift = 5; /* segment has 16 iopt entries */
85 nnpt = 0; /* one page per segment */
86 shift = 1; /* segment has 256 iopt entries */
91 nnpt = 0x07; /* 8 pages per io page table */
92 shift = 0; /* all entries are used */
97 nnpt = 0x7f; /* 128 pages per io page table */
98 shift = 0; /* all entries are used */
101 default: /* not a known compile time constant */
103 /* BUILD_BUG_ON() is not usable here */
104 extern void __get_iost_entry_bad_page_size(void);
105 __get_iost_entry_bad_page_size();
111 /* need 8 bytes per iopte */
112 (((io_address / page_size * 8)
113 /* align io page tables on 4k page boundaries */
115 /* nnpt+1 pages go into each iopt */
118 nnpt++; /* this seems to work, but the documentation is not clear
119 about wether we put nnpt or nnpt-1 into the ioste bits.
120 In theory, this can't work for 4k pages. */
121 return mk_ioste(IOST_VALID_MASK
122 | (iostep & IOST_PT_BASE_MASK)
123 | ((nnpt << 5) & IOST_NNPT_MASK)
124 | (ps & IOST_PS_MASK));
127 /* compute the address of an io pte */
128 static inline unsigned long
129 get_ioptep(ioste iost_entry, unsigned long io_address)
131 unsigned long iopt_base;
132 unsigned long page_size;
133 unsigned long page_number;
134 unsigned long iopt_offset;
136 iopt_base = iost_entry.val & IOST_PT_BASE_MASK;
137 page_size = iost_entry.val & IOST_PS_MASK;
139 /* decode page size to compute page number */
140 page_number = (io_address & 0x0fffffff) >> (10 + 2 * page_size);
141 /* page number is an offset into the io page table */
142 iopt_offset = (page_number << 3) & 0x7fff8ul;
143 return iopt_base + iopt_offset;
146 /* compute the tag field of the iopt cache entry */
147 static inline unsigned long
148 get_ioc_tag(ioste iost_entry, unsigned long io_address)
150 unsigned long iopte = get_ioptep(iost_entry, io_address);
152 return IOPT_VALID_MASK
153 | ((iopte & 0x00000000000000ff8ul) >> 3)
154 | ((iopte & 0x0000003fffffc0000ul) >> 9);
157 /* compute the hashed 6 bit index for the 4-way associative pte cache */
158 static inline unsigned long
159 get_ioc_hash(ioste iost_entry, unsigned long io_address)
161 unsigned long iopte = get_ioptep(iost_entry, io_address);
163 return ((iopte & 0x000000000000001f8ul) >> 3)
164 ^ ((iopte & 0x00000000000020000ul) >> 17)
165 ^ ((iopte & 0x00000000000010000ul) >> 15)
166 ^ ((iopte & 0x00000000000008000ul) >> 13)
167 ^ ((iopte & 0x00000000000004000ul) >> 11)
168 ^ ((iopte & 0x00000000000002000ul) >> 9)
169 ^ ((iopte & 0x00000000000001000ul) >> 7);
172 /* same as above, but pretend that we have a simpler 1-way associative
173 pte cache with an 8 bit index */
174 static inline unsigned long
175 get_ioc_hash_1way(ioste iost_entry, unsigned long io_address)
177 unsigned long iopte = get_ioptep(iost_entry, io_address);
179 return ((iopte & 0x000000000000001f8ul) >> 3)
180 ^ ((iopte & 0x00000000000020000ul) >> 17)
181 ^ ((iopte & 0x00000000000010000ul) >> 15)
182 ^ ((iopte & 0x00000000000008000ul) >> 13)
183 ^ ((iopte & 0x00000000000004000ul) >> 11)
184 ^ ((iopte & 0x00000000000002000ul) >> 9)
185 ^ ((iopte & 0x00000000000001000ul) >> 7)
186 ^ ((iopte & 0x0000000000000c000ul) >> 8);
190 get_iost_cache(void __iomem *base, unsigned long index)
192 unsigned long __iomem *p = (base + IOC_ST_CACHE_DIR);
193 return mk_ioste(in_be64(&p[index]));
197 set_iost_cache(void __iomem *base, unsigned long index, ioste ste)
199 unsigned long __iomem *p = (base + IOC_ST_CACHE_DIR);
200 pr_debug("ioste %02lx was %016lx, store %016lx", index,
201 get_iost_cache(base, index).val, ste.val);
202 out_be64(&p[index], ste.val);
203 pr_debug(" now %016lx\n", get_iost_cache(base, index).val);
206 static inline unsigned long
207 get_iopt_cache(void __iomem *base, unsigned long index, unsigned long *tag)
209 unsigned long __iomem *tags = (void *)(base + IOC_PT_CACHE_DIR);
210 unsigned long __iomem *p = (void *)(base + IOC_PT_CACHE_REG);
218 set_iopt_cache(void __iomem *base, unsigned long index,
219 unsigned long tag, unsigned long val)
221 unsigned long __iomem *tags = base + IOC_PT_CACHE_DIR;
222 unsigned long __iomem *p = base + IOC_PT_CACHE_REG;
223 pr_debug("iopt %02lx was v%016lx/t%016lx, store v%016lx/t%016lx\n",
224 index, get_iopt_cache(base, index, &oldtag), oldtag, val, tag);
227 out_be64(&tags[index], tag);
231 set_iost_origin(void __iomem *base)
233 unsigned long __iomem *p = base + IOC_ST_ORIGIN;
234 unsigned long origin = IOSTO_ENABLE | IOSTO_SW;
236 pr_debug("iost_origin %016lx, now %016lx\n", in_be64(p), origin);
241 set_iocmd_config(void __iomem *base)
243 unsigned long __iomem *p = base + 0xc00;
247 pr_debug("iost_conf %016lx, now %016lx\n", conf, conf | IOCMD_CONF_TE);
248 out_be64(p, conf | IOCMD_CONF_TE);
251 /* FIXME: get these from the device tree */
252 #define ioc_base 0x20000511000ull
253 #define ioc_mmio_base 0x20000510000ull
255 #define iopt_phys_offset (- 0x20000000) /* We have a 512MB offset from the SB */
256 #define io_page_size 0x1000000
258 static unsigned long map_iopt_entry(unsigned long address)
260 switch (address >> 20) {
262 address = 0x24020000000ull; /* spider i/o */
265 address += iopt_phys_offset;
269 return get_iopt_entry(address, ioid, IOPT_PROT_RW);
272 static void iommu_bus_setup_null(struct pci_bus *b) { }
273 static void iommu_dev_setup_null(struct pci_dev *d) { }
275 /* initialize the iommu to support a simple linear mapping
276 * for each DMA window used by any device. For now, we
277 * happen to know that there is only one DMA window in use,
278 * starting at iopt_phys_offset. */
279 static void cell_map_iommu(void)
281 unsigned long address;
286 base = __ioremap(ioc_base, 0x1000, _PAGE_NO_CACHE);
287 pr_debug("%lx mapped to %p\n", ioc_base, base);
288 set_iocmd_config(base);
291 base = __ioremap(ioc_mmio_base, 0x1000, _PAGE_NO_CACHE);
292 pr_debug("%lx mapped to %p\n", ioc_mmio_base, base);
294 set_iost_origin(base);
296 for (address = 0; address < 0x100000000ul; address += io_page_size) {
297 ioste = get_iost_entry(0x10000000000ul, address, io_page_size);
298 if ((address & 0xfffffff) == 0) /* segment start */
299 set_iost_cache(base, address >> 28, ioste);
300 index = get_ioc_hash_1way(ioste, address);
301 pr_debug("addr %08lx, index %02lx, ioste %016lx\n",
302 address, index, ioste.val);
304 get_ioc_hash_1way(ioste, address),
305 get_ioc_tag(ioste, address),
306 map_iopt_entry(address));
312 static void *cell_alloc_coherent(struct device *hwdev, size_t size,
313 dma_addr_t *dma_handle, gfp_t flag)
317 ret = (void *)__get_free_pages(flag, get_order(size));
319 memset(ret, 0, size);
320 *dma_handle = virt_to_abs(ret) | CELL_DMA_VALID;
325 static void cell_free_coherent(struct device *hwdev, size_t size,
326 void *vaddr, dma_addr_t dma_handle)
328 free_pages((unsigned long)vaddr, get_order(size));
331 static dma_addr_t cell_map_single(struct device *hwdev, void *ptr,
332 size_t size, enum dma_data_direction direction)
334 return virt_to_abs(ptr) | CELL_DMA_VALID;
337 static void cell_unmap_single(struct device *hwdev, dma_addr_t dma_addr,
338 size_t size, enum dma_data_direction direction)
342 static int cell_map_sg(struct device *hwdev, struct scatterlist *sg,
343 int nents, enum dma_data_direction direction)
347 for (i = 0; i < nents; i++, sg++) {
348 sg->dma_address = (page_to_phys(sg->page) + sg->offset)
350 sg->dma_length = sg->length;
356 static void cell_unmap_sg(struct device *hwdev, struct scatterlist *sg,
357 int nents, enum dma_data_direction direction)
361 static int cell_dma_supported(struct device *dev, u64 mask)
363 return mask < 0x100000000ull;
366 void cell_init_iommu(void)
370 /* Direct I/O, IOMMU off */
371 ppc_md.iommu_dev_setup = iommu_dev_setup_null;
372 ppc_md.iommu_bus_setup = iommu_bus_setup_null;
374 pci_dma_ops.alloc_coherent = cell_alloc_coherent;
375 pci_dma_ops.free_coherent = cell_free_coherent;
376 pci_dma_ops.map_single = cell_map_single;
377 pci_dma_ops.unmap_single = cell_unmap_single;
378 pci_dma_ops.map_sg = cell_map_sg;
379 pci_dma_ops.unmap_sg = cell_unmap_sg;
380 pci_dma_ops.dma_supported = cell_dma_supported;