Commit | Line | Data |
---|---|---|
ae209cf1 | 1 | /* |
f3f66f59 | 2 | * IOMMU implementation for Cell Broadband Processor Architecture |
ae209cf1 | 3 | * |
99e13912 | 4 | * (C) Copyright IBM Corporation 2006-2008 |
ae209cf1 | 5 | * |
165785e5 | 6 | * Author: Jeremy Kerr <jk@ozlabs.org> |
ae209cf1 | 7 | * |
165785e5 JK |
8 | * This program is free software; you can redistribute it and/or modify |
9 | * it under the terms of the GNU General Public License as published by | |
10 | * the Free Software Foundation; either version 2, or (at your option) | |
11 | * any later version. | |
12 | * | |
13 | * This program is distributed in the hope that it will be useful, | |
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
16 | * GNU General Public License for more details. | |
17 | * | |
18 | * You should have received a copy of the GNU General Public License | |
19 | * along with this program; if not, write to the Free Software | |
20 | * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. | |
ae209cf1 AB |
21 | */ |
22 | ||
23 | #undef DEBUG | |
24 | ||
25 | #include <linux/kernel.h> | |
ae209cf1 | 26 | #include <linux/init.h> |
165785e5 JK |
27 | #include <linux/interrupt.h> |
28 | #include <linux/notifier.h> | |
ccd05d08 | 29 | #include <linux/of.h> |
d8caf74f | 30 | #include <linux/of_platform.h> |
d9b2b2a2 | 31 | #include <linux/lmb.h> |
ae209cf1 | 32 | |
ae209cf1 | 33 | #include <asm/prom.h> |
165785e5 | 34 | #include <asm/iommu.h> |
ae209cf1 | 35 | #include <asm/machdep.h> |
165785e5 | 36 | #include <asm/pci-bridge.h> |
49d65b3a | 37 | #include <asm/udbg.h> |
9858ee8a | 38 | #include <asm/firmware.h> |
eef686a0 | 39 | #include <asm/cell-regs.h> |
ae209cf1 | 40 | |
165785e5 | 41 | #include "interrupt.h" |
ae209cf1 | 42 | |
165785e5 JK |
43 | /* Define CELL_IOMMU_REAL_UNMAP to actually unmap non-used pages |
44 | * instead of leaving them mapped to some dummy page. This can be | |
45 | * enabled once the appropriate workarounds for spider bugs have | |
46 | * been enabled | |
47 | */ | |
48 | #define CELL_IOMMU_REAL_UNMAP | |
49 | ||
50 | /* Define CELL_IOMMU_STRICT_PROTECTION to enforce protection of | |
51 | * IO PTEs based on the transfer direction. That can be enabled | |
52 | * once spider-net has been fixed to pass the correct direction | |
53 | * to the DMA mapping functions | |
54 | */ | |
55 | #define CELL_IOMMU_STRICT_PROTECTION | |
56 | ||
57 | ||
58 | #define NR_IOMMUS 2 | |
59 | ||
60 | /* IOC mmap registers */ | |
61 | #define IOC_Reg_Size 0x2000 | |
62 | ||
63 | #define IOC_IOPT_CacheInvd 0x908 | |
64 | #define IOC_IOPT_CacheInvd_NE_Mask 0xffe0000000000000ul | |
65 | #define IOC_IOPT_CacheInvd_IOPTE_Mask 0x000003fffffffff8ul | |
66 | #define IOC_IOPT_CacheInvd_Busy 0x0000000000000001ul | |
67 | ||
68 | #define IOC_IOST_Origin 0x918 | |
69 | #define IOC_IOST_Origin_E 0x8000000000000000ul | |
70 | #define IOC_IOST_Origin_HW 0x0000000000000800ul | |
71 | #define IOC_IOST_Origin_HL 0x0000000000000400ul | |
72 | ||
73 | #define IOC_IO_ExcpStat 0x920 | |
74 | #define IOC_IO_ExcpStat_V 0x8000000000000000ul | |
75 | #define IOC_IO_ExcpStat_SPF_Mask 0x6000000000000000ul | |
76 | #define IOC_IO_ExcpStat_SPF_S 0x6000000000000000ul | |
77 | #define IOC_IO_ExcpStat_SPF_P 0x4000000000000000ul | |
78 | #define IOC_IO_ExcpStat_ADDR_Mask 0x00000007fffff000ul | |
79 | #define IOC_IO_ExcpStat_RW_Mask 0x0000000000000800ul | |
80 | #define IOC_IO_ExcpStat_IOID_Mask 0x00000000000007fful | |
81 | ||
82 | #define IOC_IO_ExcpMask 0x928 | |
83 | #define IOC_IO_ExcpMask_SFE 0x4000000000000000ul | |
84 | #define IOC_IO_ExcpMask_PFE 0x2000000000000000ul | |
85 | ||
86 | #define IOC_IOCmd_Offset 0x1000 | |
87 | ||
88 | #define IOC_IOCmd_Cfg 0xc00 | |
89 | #define IOC_IOCmd_Cfg_TE 0x0000800000000000ul | |
90 | ||
91 | ||
92 | /* Segment table entries */ | |
93 | #define IOSTE_V 0x8000000000000000ul /* valid */ | |
94 | #define IOSTE_H 0x4000000000000000ul /* cache hint */ | |
95 | #define IOSTE_PT_Base_RPN_Mask 0x3ffffffffffff000ul /* base RPN of IOPT */ | |
96 | #define IOSTE_NPPT_Mask 0x0000000000000fe0ul /* no. pages in IOPT */ | |
97 | #define IOSTE_PS_Mask 0x0000000000000007ul /* page size */ | |
98 | #define IOSTE_PS_4K 0x0000000000000001ul /* - 4kB */ | |
99 | #define IOSTE_PS_64K 0x0000000000000003ul /* - 64kB */ | |
100 | #define IOSTE_PS_1M 0x0000000000000005ul /* - 1MB */ | |
101 | #define IOSTE_PS_16M 0x0000000000000007ul /* - 16MB */ | |
102 | ||
103 | /* Page table entries */ | |
104 | #define IOPTE_PP_W 0x8000000000000000ul /* protection: write */ | |
105 | #define IOPTE_PP_R 0x4000000000000000ul /* protection: read */ | |
106 | #define IOPTE_M 0x2000000000000000ul /* coherency required */ | |
107 | #define IOPTE_SO_R 0x1000000000000000ul /* ordering: writes */ | |
108 | #define IOPTE_SO_RW 0x1800000000000000ul /* ordering: r & w */ | |
109 | #define IOPTE_RPN_Mask 0x07fffffffffff000ul /* RPN */ | |
110 | #define IOPTE_H 0x0000000000000800ul /* cache hint */ | |
111 | #define IOPTE_IOID_Mask 0x00000000000007fful /* ioid */ | |
112 | ||
113 | ||
114 | /* IOMMU sizing */ | |
115 | #define IO_SEGMENT_SHIFT 28 | |
225d4905 | 116 | #define IO_PAGENO_BITS(shift) (IO_SEGMENT_SHIFT - (shift)) |
165785e5 JK |
117 | |
118 | /* The high bit needs to be set on every DMA address */ | |
119 | #define SPIDER_DMA_OFFSET 0x80000000ul | |
120 | ||
121 | struct iommu_window { | |
122 | struct list_head list; | |
123 | struct cbe_iommu *iommu; | |
124 | unsigned long offset; | |
125 | unsigned long size; | |
165785e5 JK |
126 | unsigned int ioid; |
127 | struct iommu_table table; | |
128 | }; | |
ae209cf1 | 129 | |
165785e5 JK |
130 | #define NAMESIZE 8 |
131 | struct cbe_iommu { | |
132 | int nid; | |
133 | char name[NAMESIZE]; | |
134 | void __iomem *xlate_regs; | |
135 | void __iomem *cmd_regs; | |
136 | unsigned long *stab; | |
137 | unsigned long *ptab; | |
138 | void *pad_page; | |
139 | struct list_head windows; | |
140 | }; | |
ae209cf1 | 141 | |
165785e5 JK |
142 | /* Static array of iommus, one per node |
143 | * each contains a list of windows, keyed from dma_window property | |
144 | * - on bus setup, look for a matching window, or create one | |
145 | * - on dev setup, assign iommu_table ptr | |
146 | */ | |
147 | static struct cbe_iommu iommus[NR_IOMMUS]; | |
148 | static int cbe_nr_iommus; | |
ae209cf1 | 149 | |
165785e5 JK |
150 | static void invalidate_tce_cache(struct cbe_iommu *iommu, unsigned long *pte, |
151 | long n_ptes) | |
ae209cf1 | 152 | { |
9340b0d3 AV |
153 | unsigned long __iomem *reg; |
154 | unsigned long val; | |
165785e5 | 155 | long n; |
ae209cf1 | 156 | |
165785e5 | 157 | reg = iommu->xlate_regs + IOC_IOPT_CacheInvd; |
ae209cf1 | 158 | |
165785e5 JK |
159 | while (n_ptes > 0) { |
160 | /* we can invalidate up to 1 << 11 PTEs at once */ | |
161 | n = min(n_ptes, 1l << 11); | |
162 | val = (((n /*- 1*/) << 53) & IOC_IOPT_CacheInvd_NE_Mask) | |
163 | | (__pa(pte) & IOC_IOPT_CacheInvd_IOPTE_Mask) | |
164 | | IOC_IOPT_CacheInvd_Busy; | |
ae209cf1 | 165 | |
165785e5 JK |
166 | out_be64(reg, val); |
167 | while (in_be64(reg) & IOC_IOPT_CacheInvd_Busy) | |
168 | ; | |
ae209cf1 | 169 | |
165785e5 JK |
170 | n_ptes -= n; |
171 | pte += n; | |
172 | } | |
ae209cf1 AB |
173 | } |
174 | ||
6490c490 | 175 | static int tce_build_cell(struct iommu_table *tbl, long index, long npages, |
4f3dd8a0 MN |
176 | unsigned long uaddr, enum dma_data_direction direction, |
177 | struct dma_attrs *attrs) | |
ae209cf1 | 178 | { |
165785e5 JK |
179 | int i; |
180 | unsigned long *io_pte, base_pte; | |
181 | struct iommu_window *window = | |
182 | container_of(tbl, struct iommu_window, table); | |
183 | ||
184 | /* implementing proper protection causes problems with the spidernet | |
185 | * driver - check mapping directions later, but allow read & write by | |
186 | * default for now.*/ | |
187 | #ifdef CELL_IOMMU_STRICT_PROTECTION | |
188 | /* to avoid referencing a global, we use a trick here to setup the | |
189 | * protection bit. "prot" is setup to be 3 fields of 4 bits apprended | |
190 | * together for each of the 3 supported direction values. It is then | |
191 | * shifted left so that the fields matching the desired direction | |
192 | * lands on the appropriate bits, and other bits are masked out. | |
193 | */ | |
194 | const unsigned long prot = 0xc48; | |
195 | base_pte = | |
196 | ((prot << (52 + 4 * direction)) & (IOPTE_PP_W | IOPTE_PP_R)) | |
197 | | IOPTE_M | IOPTE_SO_RW | (window->ioid & IOPTE_IOID_Mask); | |
198 | #else | |
199 | base_pte = IOPTE_PP_W | IOPTE_PP_R | IOPTE_M | IOPTE_SO_RW | | |
200 | (window->ioid & IOPTE_IOID_Mask); | |
201 | #endif | |
1ed6af73 MN |
202 | if (unlikely(dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs))) |
203 | base_pte &= ~IOPTE_SO_RW; | |
165785e5 | 204 | |
0d7386eb | 205 | io_pte = (unsigned long *)tbl->it_base + (index - tbl->it_offset); |
165785e5 JK |
206 | |
207 | for (i = 0; i < npages; i++, uaddr += IOMMU_PAGE_SIZE) | |
208 | io_pte[i] = base_pte | (__pa(uaddr) & IOPTE_RPN_Mask); | |
209 | ||
210 | mb(); | |
211 | ||
212 | invalidate_tce_cache(window->iommu, io_pte, npages); | |
213 | ||
214 | pr_debug("tce_build_cell(index=%lx,n=%lx,dir=%d,base_pte=%lx)\n", | |
215 | index, npages, direction, base_pte); | |
6490c490 | 216 | return 0; |
ae209cf1 AB |
217 | } |
218 | ||
165785e5 | 219 | static void tce_free_cell(struct iommu_table *tbl, long index, long npages) |
ae209cf1 | 220 | { |
ae209cf1 | 221 | |
165785e5 JK |
222 | int i; |
223 | unsigned long *io_pte, pte; | |
224 | struct iommu_window *window = | |
225 | container_of(tbl, struct iommu_window, table); | |
ae209cf1 | 226 | |
165785e5 | 227 | pr_debug("tce_free_cell(index=%lx,n=%lx)\n", index, npages); |
ae209cf1 | 228 | |
165785e5 JK |
229 | #ifdef CELL_IOMMU_REAL_UNMAP |
230 | pte = 0; | |
231 | #else | |
232 | /* spider bridge does PCI reads after freeing - insert a mapping | |
233 | * to a scratch page instead of an invalid entry */ | |
234 | pte = IOPTE_PP_R | IOPTE_M | IOPTE_SO_RW | __pa(window->iommu->pad_page) | |
235 | | (window->ioid & IOPTE_IOID_Mask); | |
236 | #endif | |
ae209cf1 | 237 | |
0d7386eb | 238 | io_pte = (unsigned long *)tbl->it_base + (index - tbl->it_offset); |
ae209cf1 | 239 | |
165785e5 JK |
240 | for (i = 0; i < npages; i++) |
241 | io_pte[i] = pte; | |
242 | ||
243 | mb(); | |
ae209cf1 | 244 | |
165785e5 | 245 | invalidate_tce_cache(window->iommu, io_pte, npages); |
ae209cf1 AB |
246 | } |
247 | ||
165785e5 | 248 | static irqreturn_t ioc_interrupt(int irq, void *data) |
ae209cf1 | 249 | { |
165785e5 JK |
250 | unsigned long stat; |
251 | struct cbe_iommu *iommu = data; | |
252 | ||
253 | stat = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat); | |
254 | ||
255 | /* Might want to rate limit it */ | |
256 | printk(KERN_ERR "iommu: DMA exception 0x%016lx\n", stat); | |
257 | printk(KERN_ERR " V=%d, SPF=[%c%c], RW=%s, IOID=0x%04x\n", | |
258 | !!(stat & IOC_IO_ExcpStat_V), | |
259 | (stat & IOC_IO_ExcpStat_SPF_S) ? 'S' : ' ', | |
260 | (stat & IOC_IO_ExcpStat_SPF_P) ? 'P' : ' ', | |
261 | (stat & IOC_IO_ExcpStat_RW_Mask) ? "Read" : "Write", | |
262 | (unsigned int)(stat & IOC_IO_ExcpStat_IOID_Mask)); | |
263 | printk(KERN_ERR " page=0x%016lx\n", | |
264 | stat & IOC_IO_ExcpStat_ADDR_Mask); | |
265 | ||
266 | /* clear interrupt */ | |
267 | stat &= ~IOC_IO_ExcpStat_V; | |
268 | out_be64(iommu->xlate_regs + IOC_IO_ExcpStat, stat); | |
269 | ||
270 | return IRQ_HANDLED; | |
ae209cf1 AB |
271 | } |
272 | ||
165785e5 | 273 | static int cell_iommu_find_ioc(int nid, unsigned long *base) |
ae209cf1 | 274 | { |
165785e5 JK |
275 | struct device_node *np; |
276 | struct resource r; | |
277 | ||
278 | *base = 0; | |
279 | ||
280 | /* First look for new style /be nodes */ | |
281 | for_each_node_by_name(np, "ioc") { | |
282 | if (of_node_to_nid(np) != nid) | |
283 | continue; | |
284 | if (of_address_to_resource(np, 0, &r)) { | |
285 | printk(KERN_ERR "iommu: can't get address for %s\n", | |
286 | np->full_name); | |
287 | continue; | |
288 | } | |
289 | *base = r.start; | |
290 | of_node_put(np); | |
291 | return 0; | |
292 | } | |
293 | ||
294 | /* Ok, let's try the old way */ | |
295 | for_each_node_by_type(np, "cpu") { | |
296 | const unsigned int *nidp; | |
297 | const unsigned long *tmp; | |
298 | ||
e2eb6392 | 299 | nidp = of_get_property(np, "node-id", NULL); |
165785e5 | 300 | if (nidp && *nidp == nid) { |
e2eb6392 | 301 | tmp = of_get_property(np, "ioc-translation", NULL); |
165785e5 JK |
302 | if (tmp) { |
303 | *base = *tmp; | |
304 | of_node_put(np); | |
305 | return 0; | |
306 | } | |
307 | } | |
308 | } | |
ae209cf1 | 309 | |
165785e5 | 310 | return -ENODEV; |
ae209cf1 AB |
311 | } |
312 | ||
7d432ff1 | 313 | static void cell_iommu_setup_stab(struct cbe_iommu *iommu, |
41347917 ME |
314 | unsigned long dbase, unsigned long dsize, |
315 | unsigned long fbase, unsigned long fsize) | |
ae209cf1 | 316 | { |
165785e5 | 317 | struct page *page; |
7d432ff1 | 318 | unsigned long segments, stab_size; |
41347917 ME |
319 | |
320 | segments = max(dbase + dsize, fbase + fsize) >> IO_SEGMENT_SHIFT; | |
165785e5 | 321 | |
7d432ff1 | 322 | pr_debug("%s: iommu[%d]: segments: %lu\n", |
e48b1b45 | 323 | __func__, iommu->nid, segments); |
165785e5 JK |
324 | |
325 | /* set up the segment table */ | |
3ca6644e ME |
326 | stab_size = segments * sizeof(unsigned long); |
327 | page = alloc_pages_node(iommu->nid, GFP_KERNEL, get_order(stab_size)); | |
165785e5 JK |
328 | BUG_ON(!page); |
329 | iommu->stab = page_address(page); | |
7d432ff1 ME |
330 | memset(iommu->stab, 0, stab_size); |
331 | } | |
332 | ||
333 | static unsigned long *cell_iommu_alloc_ptab(struct cbe_iommu *iommu, | |
334 | unsigned long base, unsigned long size, unsigned long gap_base, | |
225d4905 | 335 | unsigned long gap_size, unsigned long page_shift) |
7d432ff1 ME |
336 | { |
337 | struct page *page; | |
338 | int i; | |
339 | unsigned long reg, segments, pages_per_segment, ptab_size, | |
340 | n_pte_pages, start_seg, *ptab; | |
341 | ||
342 | start_seg = base >> IO_SEGMENT_SHIFT; | |
343 | segments = size >> IO_SEGMENT_SHIFT; | |
225d4905 ME |
344 | pages_per_segment = 1ull << IO_PAGENO_BITS(page_shift); |
345 | /* PTEs for each segment must start on a 4K bounday */ | |
346 | pages_per_segment = max(pages_per_segment, | |
347 | (1 << 12) / sizeof(unsigned long)); | |
165785e5 | 348 | |
165785e5 | 349 | ptab_size = segments * pages_per_segment * sizeof(unsigned long); |
e48b1b45 | 350 | pr_debug("%s: iommu[%d]: ptab_size: %lu, order: %d\n", __func__, |
165785e5 JK |
351 | iommu->nid, ptab_size, get_order(ptab_size)); |
352 | page = alloc_pages_node(iommu->nid, GFP_KERNEL, get_order(ptab_size)); | |
353 | BUG_ON(!page); | |
354 | ||
7d432ff1 ME |
355 | ptab = page_address(page); |
356 | memset(ptab, 0, ptab_size); | |
165785e5 | 357 | |
3d3e6da1 ME |
358 | /* number of 4K pages needed for a page table */ |
359 | n_pte_pages = (pages_per_segment * sizeof(unsigned long)) >> 12; | |
165785e5 JK |
360 | |
361 | pr_debug("%s: iommu[%d]: stab at %p, ptab at %p, n_pte_pages: %lu\n", | |
e48b1b45 | 362 | __func__, iommu->nid, iommu->stab, ptab, |
165785e5 JK |
363 | n_pte_pages); |
364 | ||
365 | /* initialise the STEs */ | |
366 | reg = IOSTE_V | ((n_pte_pages - 1) << 5); | |
367 | ||
225d4905 ME |
368 | switch (page_shift) { |
369 | case 12: reg |= IOSTE_PS_4K; break; | |
370 | case 16: reg |= IOSTE_PS_64K; break; | |
371 | case 20: reg |= IOSTE_PS_1M; break; | |
372 | case 24: reg |= IOSTE_PS_16M; break; | |
373 | default: BUG(); | |
165785e5 | 374 | } |
ae209cf1 | 375 | |
7d432ff1 ME |
376 | gap_base = gap_base >> IO_SEGMENT_SHIFT; |
377 | gap_size = gap_size >> IO_SEGMENT_SHIFT; | |
378 | ||
165785e5 | 379 | pr_debug("Setting up IOMMU stab:\n"); |
7d432ff1 ME |
380 | for (i = start_seg; i < (start_seg + segments); i++) { |
381 | if (i >= gap_base && i < (gap_base + gap_size)) { | |
382 | pr_debug("\toverlap at %d, skipping\n", i); | |
383 | continue; | |
384 | } | |
3d3e6da1 ME |
385 | iommu->stab[i] = reg | (__pa(ptab) + (n_pte_pages << 12) * |
386 | (i - start_seg)); | |
165785e5 JK |
387 | pr_debug("\t[%d] 0x%016lx\n", i, iommu->stab[i]); |
388 | } | |
7d432ff1 ME |
389 | |
390 | return ptab; | |
7fc67afc ME |
391 | } |
392 | ||
393 | static void cell_iommu_enable_hardware(struct cbe_iommu *iommu) | |
394 | { | |
395 | int ret; | |
396 | unsigned long reg, xlate_base; | |
397 | unsigned int virq; | |
398 | ||
399 | if (cell_iommu_find_ioc(iommu->nid, &xlate_base)) | |
400 | panic("%s: missing IOC register mappings for node %d\n", | |
e48b1b45 | 401 | __func__, iommu->nid); |
7fc67afc ME |
402 | |
403 | iommu->xlate_regs = ioremap(xlate_base, IOC_Reg_Size); | |
404 | iommu->cmd_regs = iommu->xlate_regs + IOC_IOCmd_Offset; | |
165785e5 JK |
405 | |
406 | /* ensure that the STEs have updated */ | |
407 | mb(); | |
408 | ||
409 | /* setup interrupts for the iommu. */ | |
410 | reg = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat); | |
411 | out_be64(iommu->xlate_regs + IOC_IO_ExcpStat, | |
412 | reg & ~IOC_IO_ExcpStat_V); | |
413 | out_be64(iommu->xlate_regs + IOC_IO_ExcpMask, | |
414 | IOC_IO_ExcpMask_PFE | IOC_IO_ExcpMask_SFE); | |
415 | ||
416 | virq = irq_create_mapping(NULL, | |
417 | IIC_IRQ_IOEX_ATI | (iommu->nid << IIC_IRQ_NODE_SHIFT)); | |
418 | BUG_ON(virq == NO_IRQ); | |
419 | ||
420 | ret = request_irq(virq, ioc_interrupt, IRQF_DISABLED, | |
421 | iommu->name, iommu); | |
422 | BUG_ON(ret); | |
49d65b3a | 423 | |
165785e5 JK |
424 | /* set the IOC segment table origin register (and turn on the iommu) */ |
425 | reg = IOC_IOST_Origin_E | __pa(iommu->stab) | IOC_IOST_Origin_HW; | |
426 | out_be64(iommu->xlate_regs + IOC_IOST_Origin, reg); | |
427 | in_be64(iommu->xlate_regs + IOC_IOST_Origin); | |
ae209cf1 | 428 | |
165785e5 JK |
429 | /* turn on IO translation */ |
430 | reg = in_be64(iommu->cmd_regs + IOC_IOCmd_Cfg) | IOC_IOCmd_Cfg_TE; | |
431 | out_be64(iommu->cmd_regs + IOC_IOCmd_Cfg, reg); | |
432 | } | |
433 | ||
7fc67afc ME |
434 | static void cell_iommu_setup_hardware(struct cbe_iommu *iommu, |
435 | unsigned long base, unsigned long size) | |
436 | { | |
7d432ff1 | 437 | cell_iommu_setup_stab(iommu, base, size, 0, 0); |
225d4905 ME |
438 | iommu->ptab = cell_iommu_alloc_ptab(iommu, base, size, 0, 0, |
439 | IOMMU_PAGE_SHIFT); | |
7fc67afc ME |
440 | cell_iommu_enable_hardware(iommu); |
441 | } | |
442 | ||
165785e5 JK |
443 | #if 0/* Unused for now */ |
444 | static struct iommu_window *find_window(struct cbe_iommu *iommu, | |
445 | unsigned long offset, unsigned long size) | |
ae209cf1 | 446 | { |
165785e5 | 447 | struct iommu_window *window; |
ae209cf1 | 448 | |
165785e5 JK |
449 | /* todo: check for overlapping (but not equal) windows) */ |
450 | ||
451 | list_for_each_entry(window, &(iommu->windows), list) { | |
452 | if (window->offset == offset && window->size == size) | |
453 | return window; | |
49d65b3a | 454 | } |
165785e5 JK |
455 | |
456 | return NULL; | |
49d65b3a | 457 | } |
165785e5 | 458 | #endif |
ae209cf1 | 459 | |
c96b5126 ME |
460 | static inline u32 cell_iommu_get_ioid(struct device_node *np) |
461 | { | |
462 | const u32 *ioid; | |
463 | ||
464 | ioid = of_get_property(np, "ioid", NULL); | |
465 | if (ioid == NULL) { | |
466 | printk(KERN_WARNING "iommu: missing ioid for %s using 0\n", | |
467 | np->full_name); | |
468 | return 0; | |
469 | } | |
470 | ||
471 | return *ioid; | |
472 | } | |
473 | ||
165785e5 JK |
474 | static struct iommu_window * __init |
475 | cell_iommu_setup_window(struct cbe_iommu *iommu, struct device_node *np, | |
476 | unsigned long offset, unsigned long size, | |
477 | unsigned long pte_offset) | |
49d65b3a | 478 | { |
165785e5 | 479 | struct iommu_window *window; |
edf441fb | 480 | struct page *page; |
c96b5126 | 481 | u32 ioid; |
ae209cf1 | 482 | |
c96b5126 | 483 | ioid = cell_iommu_get_ioid(np); |
165785e5 JK |
484 | |
485 | window = kmalloc_node(sizeof(*window), GFP_KERNEL, iommu->nid); | |
486 | BUG_ON(window == NULL); | |
487 | ||
488 | window->offset = offset; | |
489 | window->size = size; | |
c96b5126 | 490 | window->ioid = ioid; |
165785e5 | 491 | window->iommu = iommu; |
165785e5 JK |
492 | |
493 | window->table.it_blocksize = 16; | |
494 | window->table.it_base = (unsigned long)iommu->ptab; | |
495 | window->table.it_index = iommu->nid; | |
08e02427 | 496 | window->table.it_offset = (offset >> IOMMU_PAGE_SHIFT) + pte_offset; |
165785e5 JK |
497 | window->table.it_size = size >> IOMMU_PAGE_SHIFT; |
498 | ||
499 | iommu_init_table(&window->table, iommu->nid); | |
500 | ||
501 | pr_debug("\tioid %d\n", window->ioid); | |
502 | pr_debug("\tblocksize %ld\n", window->table.it_blocksize); | |
503 | pr_debug("\tbase 0x%016lx\n", window->table.it_base); | |
504 | pr_debug("\toffset 0x%lx\n", window->table.it_offset); | |
505 | pr_debug("\tsize %ld\n", window->table.it_size); | |
506 | ||
507 | list_add(&window->list, &iommu->windows); | |
508 | ||
509 | if (offset != 0) | |
510 | return window; | |
511 | ||
512 | /* We need to map and reserve the first IOMMU page since it's used | |
513 | * by the spider workaround. In theory, we only need to do that when | |
514 | * running on spider but it doesn't really matter. | |
515 | * | |
516 | * This code also assumes that we have a window that starts at 0, | |
517 | * which is the case on all spider based blades. | |
518 | */ | |
edf441fb ME |
519 | page = alloc_pages_node(iommu->nid, GFP_KERNEL, 0); |
520 | BUG_ON(!page); | |
521 | iommu->pad_page = page_address(page); | |
522 | clear_page(iommu->pad_page); | |
523 | ||
165785e5 JK |
524 | __set_bit(0, window->table.it_map); |
525 | tce_build_cell(&window->table, window->table.it_offset, 1, | |
4f3dd8a0 | 526 | (unsigned long)iommu->pad_page, DMA_TO_DEVICE, NULL); |
165785e5 JK |
527 | window->table.it_hint = window->table.it_blocksize; |
528 | ||
529 | return window; | |
530 | } | |
ae209cf1 | 531 | |
165785e5 JK |
532 | static struct cbe_iommu *cell_iommu_for_node(int nid) |
533 | { | |
534 | int i; | |
49d65b3a | 535 | |
165785e5 JK |
536 | for (i = 0; i < cbe_nr_iommus; i++) |
537 | if (iommus[i].nid == nid) | |
538 | return &iommus[i]; | |
539 | return NULL; | |
540 | } | |
49d65b3a | 541 | |
f5d67bd5 ME |
542 | static unsigned long cell_dma_direct_offset; |
543 | ||
99e13912 | 544 | static unsigned long dma_iommu_fixed_base; |
1ed6af73 MN |
545 | |
546 | /* iommu_fixed_is_weak is set if booted with iommu_fixed=weak */ | |
547 | static int iommu_fixed_is_weak; | |
99e13912 | 548 | |
7e5f8105 | 549 | static struct iommu_table *cell_get_iommu_table(struct device *dev) |
165785e5 JK |
550 | { |
551 | struct iommu_window *window; | |
552 | struct cbe_iommu *iommu; | |
553 | struct dev_archdata *archdata = &dev->archdata; | |
554 | ||
165785e5 JK |
555 | /* Current implementation uses the first window available in that |
556 | * node's iommu. We -might- do something smarter later though it may | |
557 | * never be necessary | |
558 | */ | |
8fae0353 | 559 | iommu = cell_iommu_for_node(dev_to_node(dev)); |
165785e5 JK |
560 | if (iommu == NULL || list_empty(&iommu->windows)) { |
561 | printk(KERN_ERR "iommu: missing iommu for %s (node %d)\n", | |
562 | archdata->of_node ? archdata->of_node->full_name : "?", | |
8fae0353 | 563 | dev_to_node(dev)); |
7e5f8105 | 564 | return NULL; |
165785e5 JK |
565 | } |
566 | window = list_entry(iommu->windows.next, struct iommu_window, list); | |
49d65b3a | 567 | |
7e5f8105 | 568 | return &window->table; |
49d65b3a JO |
569 | } |
570 | ||
1ed6af73 MN |
571 | /* A coherent allocation implies strong ordering */ |
572 | ||
573 | static void *dma_fixed_alloc_coherent(struct device *dev, size_t size, | |
574 | dma_addr_t *dma_handle, gfp_t flag) | |
575 | { | |
576 | if (iommu_fixed_is_weak) | |
577 | return iommu_alloc_coherent(dev, cell_get_iommu_table(dev), | |
578 | size, dma_handle, | |
579 | device_to_mask(dev), flag, | |
8fae0353 | 580 | dev_to_node(dev)); |
1ed6af73 MN |
581 | else |
582 | return dma_direct_ops.alloc_coherent(dev, size, dma_handle, | |
583 | flag); | |
584 | } | |
585 | ||
586 | static void dma_fixed_free_coherent(struct device *dev, size_t size, | |
587 | void *vaddr, dma_addr_t dma_handle) | |
588 | { | |
589 | if (iommu_fixed_is_weak) | |
590 | iommu_free_coherent(cell_get_iommu_table(dev), size, vaddr, | |
591 | dma_handle); | |
592 | else | |
593 | dma_direct_ops.free_coherent(dev, size, vaddr, dma_handle); | |
594 | } | |
595 | ||
f9226d57 MN |
596 | static dma_addr_t dma_fixed_map_page(struct device *dev, struct page *page, |
597 | unsigned long offset, size_t size, | |
598 | enum dma_data_direction direction, | |
599 | struct dma_attrs *attrs) | |
1ed6af73 MN |
600 | { |
601 | if (iommu_fixed_is_weak == dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs)) | |
f9226d57 MN |
602 | return dma_direct_ops.map_page(dev, page, offset, size, |
603 | direction, attrs); | |
1ed6af73 | 604 | else |
f9226d57 MN |
605 | return iommu_map_page(dev, cell_get_iommu_table(dev), page, |
606 | offset, size, device_to_mask(dev), | |
607 | direction, attrs); | |
1ed6af73 MN |
608 | } |
609 | ||
f9226d57 MN |
610 | static void dma_fixed_unmap_page(struct device *dev, dma_addr_t dma_addr, |
611 | size_t size, enum dma_data_direction direction, | |
612 | struct dma_attrs *attrs) | |
1ed6af73 MN |
613 | { |
614 | if (iommu_fixed_is_weak == dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs)) | |
f9226d57 MN |
615 | dma_direct_ops.unmap_page(dev, dma_addr, size, direction, |
616 | attrs); | |
1ed6af73 | 617 | else |
f9226d57 MN |
618 | iommu_unmap_page(cell_get_iommu_table(dev), dma_addr, size, |
619 | direction, attrs); | |
1ed6af73 MN |
620 | } |
621 | ||
622 | static int dma_fixed_map_sg(struct device *dev, struct scatterlist *sg, | |
623 | int nents, enum dma_data_direction direction, | |
624 | struct dma_attrs *attrs) | |
625 | { | |
626 | if (iommu_fixed_is_weak == dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs)) | |
627 | return dma_direct_ops.map_sg(dev, sg, nents, direction, attrs); | |
628 | else | |
629 | return iommu_map_sg(dev, cell_get_iommu_table(dev), sg, nents, | |
630 | device_to_mask(dev), direction, attrs); | |
631 | } | |
632 | ||
633 | static void dma_fixed_unmap_sg(struct device *dev, struct scatterlist *sg, | |
634 | int nents, enum dma_data_direction direction, | |
635 | struct dma_attrs *attrs) | |
636 | { | |
637 | if (iommu_fixed_is_weak == dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs)) | |
638 | dma_direct_ops.unmap_sg(dev, sg, nents, direction, attrs); | |
639 | else | |
640 | iommu_unmap_sg(cell_get_iommu_table(dev), sg, nents, direction, | |
641 | attrs); | |
642 | } | |
643 | ||
644 | static int dma_fixed_dma_supported(struct device *dev, u64 mask) | |
645 | { | |
646 | return mask == DMA_64BIT_MASK; | |
647 | } | |
648 | ||
649 | static int dma_set_mask_and_switch(struct device *dev, u64 dma_mask); | |
650 | ||
651 | struct dma_mapping_ops dma_iommu_fixed_ops = { | |
652 | .alloc_coherent = dma_fixed_alloc_coherent, | |
653 | .free_coherent = dma_fixed_free_coherent, | |
1ed6af73 MN |
654 | .map_sg = dma_fixed_map_sg, |
655 | .unmap_sg = dma_fixed_unmap_sg, | |
656 | .dma_supported = dma_fixed_dma_supported, | |
657 | .set_dma_mask = dma_set_mask_and_switch, | |
f9226d57 MN |
658 | .map_page = dma_fixed_map_page, |
659 | .unmap_page = dma_fixed_unmap_page, | |
1ed6af73 MN |
660 | }; |
661 | ||
f9660e8a | 662 | static void cell_dma_dev_setup_fixed(struct device *dev); |
99e13912 | 663 | |
86865771 ME |
664 | static void cell_dma_dev_setup(struct device *dev) |
665 | { | |
666 | struct dev_archdata *archdata = &dev->archdata; | |
667 | ||
99e13912 ME |
668 | /* Order is important here, these are not mutually exclusive */ |
669 | if (get_dma_ops(dev) == &dma_iommu_fixed_ops) | |
f9660e8a | 670 | cell_dma_dev_setup_fixed(dev); |
99e13912 | 671 | else if (get_pci_dma_ops() == &dma_iommu_ops) |
7e5f8105 | 672 | archdata->dma_data = cell_get_iommu_table(dev); |
86865771 ME |
673 | else if (get_pci_dma_ops() == &dma_direct_ops) |
674 | archdata->dma_data = (void *)cell_dma_direct_offset; | |
675 | else | |
676 | BUG(); | |
677 | } | |
678 | ||
165785e5 | 679 | static void cell_pci_dma_dev_setup(struct pci_dev *dev) |
49d65b3a | 680 | { |
165785e5 JK |
681 | cell_dma_dev_setup(&dev->dev); |
682 | } | |
49d65b3a | 683 | |
165785e5 JK |
684 | static int cell_of_bus_notify(struct notifier_block *nb, unsigned long action, |
685 | void *data) | |
686 | { | |
687 | struct device *dev = data; | |
49d65b3a | 688 | |
165785e5 JK |
689 | /* We are only intereted in device addition */ |
690 | if (action != BUS_NOTIFY_ADD_DEVICE) | |
691 | return 0; | |
49d65b3a | 692 | |
165785e5 | 693 | /* We use the PCI DMA ops */ |
57190708 | 694 | dev->archdata.dma_ops = get_pci_dma_ops(); |
49d65b3a | 695 | |
165785e5 | 696 | cell_dma_dev_setup(dev); |
49d65b3a | 697 | |
165785e5 JK |
698 | return 0; |
699 | } | |
49d65b3a | 700 | |
165785e5 JK |
701 | static struct notifier_block cell_of_bus_notifier = { |
702 | .notifier_call = cell_of_bus_notify | |
703 | }; | |
49d65b3a | 704 | |
165785e5 JK |
705 | static int __init cell_iommu_get_window(struct device_node *np, |
706 | unsigned long *base, | |
707 | unsigned long *size) | |
708 | { | |
709 | const void *dma_window; | |
710 | unsigned long index; | |
49d65b3a | 711 | |
165785e5 | 712 | /* Use ibm,dma-window if available, else, hard code ! */ |
e2eb6392 | 713 | dma_window = of_get_property(np, "ibm,dma-window", NULL); |
165785e5 JK |
714 | if (dma_window == NULL) { |
715 | *base = 0; | |
716 | *size = 0x80000000u; | |
717 | return -ENODEV; | |
718 | } | |
49d65b3a | 719 | |
165785e5 | 720 | of_parse_dma_window(np, dma_window, &index, base, size); |
49d65b3a | 721 | return 0; |
ae209cf1 AB |
722 | } |
723 | ||
209bfbb4 | 724 | static struct cbe_iommu * __init cell_iommu_alloc(struct device_node *np) |
49d65b3a | 725 | { |
165785e5 | 726 | struct cbe_iommu *iommu; |
165785e5 JK |
727 | int nid, i; |
728 | ||
729 | /* Get node ID */ | |
730 | nid = of_node_to_nid(np); | |
731 | if (nid < 0) { | |
732 | printk(KERN_ERR "iommu: failed to get node for %s\n", | |
733 | np->full_name); | |
209bfbb4 | 734 | return NULL; |
165785e5 JK |
735 | } |
736 | pr_debug("iommu: setting up iommu for node %d (%s)\n", | |
737 | nid, np->full_name); | |
738 | ||
739 | /* XXX todo: If we can have multiple windows on the same IOMMU, which | |
740 | * isn't the case today, we probably want here to check wether the | |
741 | * iommu for that node is already setup. | |
742 | * However, there might be issue with getting the size right so let's | |
743 | * ignore that for now. We might want to completely get rid of the | |
744 | * multiple window support since the cell iommu supports per-page ioids | |
745 | */ | |
746 | ||
747 | if (cbe_nr_iommus >= NR_IOMMUS) { | |
748 | printk(KERN_ERR "iommu: too many IOMMUs detected ! (%s)\n", | |
749 | np->full_name); | |
209bfbb4 | 750 | return NULL; |
165785e5 JK |
751 | } |
752 | ||
753 | /* Init base fields */ | |
754 | i = cbe_nr_iommus++; | |
755 | iommu = &iommus[i]; | |
9340b0d3 | 756 | iommu->stab = NULL; |
165785e5 JK |
757 | iommu->nid = nid; |
758 | snprintf(iommu->name, sizeof(iommu->name), "iommu%d", i); | |
759 | INIT_LIST_HEAD(&iommu->windows); | |
49d65b3a | 760 | |
209bfbb4 ME |
761 | return iommu; |
762 | } | |
763 | ||
764 | static void __init cell_iommu_init_one(struct device_node *np, | |
765 | unsigned long offset) | |
766 | { | |
767 | struct cbe_iommu *iommu; | |
768 | unsigned long base, size; | |
769 | ||
770 | iommu = cell_iommu_alloc(np); | |
771 | if (!iommu) | |
772 | return; | |
773 | ||
165785e5 JK |
774 | /* Obtain a window for it */ |
775 | cell_iommu_get_window(np, &base, &size); | |
49d65b3a | 776 | |
165785e5 JK |
777 | pr_debug("\ttranslating window 0x%lx...0x%lx\n", |
778 | base, base + size - 1); | |
49d65b3a | 779 | |
165785e5 | 780 | /* Initialize the hardware */ |
7fc67afc | 781 | cell_iommu_setup_hardware(iommu, base, size); |
49d65b3a | 782 | |
165785e5 JK |
783 | /* Setup the iommu_table */ |
784 | cell_iommu_setup_window(iommu, np, base, size, | |
785 | offset >> IOMMU_PAGE_SHIFT); | |
786 | } | |
49d65b3a | 787 | |
165785e5 JK |
788 | static void __init cell_disable_iommus(void) |
789 | { | |
790 | int node; | |
791 | unsigned long base, val; | |
792 | void __iomem *xregs, *cregs; | |
793 | ||
794 | /* Make sure IOC translation is disabled on all nodes */ | |
795 | for_each_online_node(node) { | |
796 | if (cell_iommu_find_ioc(node, &base)) | |
797 | continue; | |
798 | xregs = ioremap(base, IOC_Reg_Size); | |
799 | if (xregs == NULL) | |
800 | continue; | |
801 | cregs = xregs + IOC_IOCmd_Offset; | |
802 | ||
803 | pr_debug("iommu: cleaning up iommu on node %d\n", node); | |
804 | ||
805 | out_be64(xregs + IOC_IOST_Origin, 0); | |
806 | (void)in_be64(xregs + IOC_IOST_Origin); | |
807 | val = in_be64(cregs + IOC_IOCmd_Cfg); | |
808 | val &= ~IOC_IOCmd_Cfg_TE; | |
809 | out_be64(cregs + IOC_IOCmd_Cfg, val); | |
810 | (void)in_be64(cregs + IOC_IOCmd_Cfg); | |
811 | ||
812 | iounmap(xregs); | |
813 | } | |
814 | } | |
49d65b3a | 815 | |
165785e5 JK |
816 | static int __init cell_iommu_init_disabled(void) |
817 | { | |
818 | struct device_node *np = NULL; | |
819 | unsigned long base = 0, size; | |
49d65b3a | 820 | |
165785e5 | 821 | /* When no iommu is present, we use direct DMA ops */ |
98747770 | 822 | set_pci_dma_ops(&dma_direct_ops); |
49d65b3a | 823 | |
165785e5 JK |
824 | /* First make sure all IOC translation is turned off */ |
825 | cell_disable_iommus(); | |
826 | ||
827 | /* If we have no Axon, we set up the spider DMA magic offset */ | |
828 | if (of_find_node_by_name(NULL, "axon") == NULL) | |
f5d67bd5 | 829 | cell_dma_direct_offset = SPIDER_DMA_OFFSET; |
165785e5 JK |
830 | |
831 | /* Now we need to check to see where the memory is mapped | |
832 | * in PCI space. We assume that all busses use the same dma | |
833 | * window which is always the case so far on Cell, thus we | |
834 | * pick up the first pci-internal node we can find and check | |
835 | * the DMA window from there. | |
836 | */ | |
837 | for_each_node_by_name(np, "axon") { | |
838 | if (np->parent == NULL || np->parent->parent != NULL) | |
839 | continue; | |
840 | if (cell_iommu_get_window(np, &base, &size) == 0) | |
841 | break; | |
842 | } | |
843 | if (np == NULL) { | |
844 | for_each_node_by_name(np, "pci-internal") { | |
845 | if (np->parent == NULL || np->parent->parent != NULL) | |
846 | continue; | |
847 | if (cell_iommu_get_window(np, &base, &size) == 0) | |
848 | break; | |
849 | } | |
850 | } | |
851 | of_node_put(np); | |
852 | ||
853 | /* If we found a DMA window, we check if it's big enough to enclose | |
854 | * all of physical memory. If not, we force enable IOMMU | |
855 | */ | |
856 | if (np && size < lmb_end_of_DRAM()) { | |
857 | printk(KERN_WARNING "iommu: force-enabled, dma window" | |
858 | " (%ldMB) smaller than total memory (%ldMB)\n", | |
859 | size >> 20, lmb_end_of_DRAM() >> 20); | |
860 | return -ENODEV; | |
49d65b3a JO |
861 | } |
862 | ||
f5d67bd5 | 863 | cell_dma_direct_offset += base; |
165785e5 | 864 | |
f5d67bd5 | 865 | if (cell_dma_direct_offset != 0) |
110f95c9 ME |
866 | ppc_md.pci_dma_dev_setup = cell_pci_dma_dev_setup; |
867 | ||
165785e5 | 868 | printk("iommu: disabled, direct DMA offset is 0x%lx\n", |
f5d67bd5 | 869 | cell_dma_direct_offset); |
165785e5 JK |
870 | |
871 | return 0; | |
49d65b3a JO |
872 | } |
873 | ||
99e13912 ME |
874 | /* |
875 | * Fixed IOMMU mapping support | |
876 | * | |
877 | * This code adds support for setting up a fixed IOMMU mapping on certain | |
878 | * cell machines. For 64-bit devices this avoids the performance overhead of | |
879 | * mapping and unmapping pages at runtime. 32-bit devices are unable to use | |
880 | * the fixed mapping. | |
881 | * | |
882 | * The fixed mapping is established at boot, and maps all of physical memory | |
883 | * 1:1 into device space at some offset. On machines with < 30 GB of memory | |
884 | * we setup the fixed mapping immediately above the normal IOMMU window. | |
885 | * | |
886 | * For example a machine with 4GB of memory would end up with the normal | |
887 | * IOMMU window from 0-2GB and the fixed mapping window from 2GB to 6GB. In | |
888 | * this case a 64-bit device wishing to DMA to 1GB would be told to DMA to | |
889 | * 3GB, plus any offset required by firmware. The firmware offset is encoded | |
890 | * in the "dma-ranges" property. | |
891 | * | |
892 | * On machines with 30GB or more of memory, we are unable to place the fixed | |
893 | * mapping above the normal IOMMU window as we would run out of address space. | |
894 | * Instead we move the normal IOMMU window to coincide with the hash page | |
895 | * table, this region does not need to be part of the fixed mapping as no | |
896 | * device should ever be DMA'ing to it. We then setup the fixed mapping | |
897 | * from 0 to 32GB. | |
898 | */ | |
899 | ||
900 | static u64 cell_iommu_get_fixed_address(struct device *dev) | |
901 | { | |
3a4295d1 | 902 | u64 cpu_addr, size, best_size, dev_addr = OF_BAD_ADDR; |
ccd05d08 | 903 | struct device_node *np; |
99e13912 | 904 | const u32 *ranges = NULL; |
3a4295d1 | 905 | int i, len, best, naddr, nsize, pna, range_size; |
99e13912 | 906 | |
ccd05d08 | 907 | np = of_node_get(dev->archdata.of_node); |
3a4295d1 ME |
908 | while (1) { |
909 | naddr = of_n_addr_cells(np); | |
910 | nsize = of_n_size_cells(np); | |
911 | np = of_get_next_parent(np); | |
912 | if (!np) | |
913 | break; | |
914 | ||
99e13912 | 915 | ranges = of_get_property(np, "dma-ranges", &len); |
3a4295d1 ME |
916 | |
917 | /* Ignore empty ranges, they imply no translation required */ | |
918 | if (ranges && len > 0) | |
ccd05d08 | 919 | break; |
99e13912 ME |
920 | } |
921 | ||
922 | if (!ranges) { | |
923 | dev_dbg(dev, "iommu: no dma-ranges found\n"); | |
924 | goto out; | |
925 | } | |
926 | ||
927 | len /= sizeof(u32); | |
928 | ||
3a4295d1 ME |
929 | pna = of_n_addr_cells(np); |
930 | range_size = naddr + nsize + pna; | |
931 | ||
99e13912 | 932 | /* dma-ranges format: |
3a4295d1 ME |
933 | * child addr : naddr cells |
934 | * parent addr : pna cells | |
935 | * size : nsize cells | |
99e13912 | 936 | */ |
3a4295d1 ME |
937 | for (i = 0, best = -1, best_size = 0; i < len; i += range_size) { |
938 | cpu_addr = of_translate_dma_address(np, ranges + i + naddr); | |
939 | size = of_read_number(ranges + i + naddr + pna, nsize); | |
99e13912 ME |
940 | |
941 | if (cpu_addr == 0 && size > best_size) { | |
942 | best = i; | |
943 | best_size = size; | |
944 | } | |
945 | } | |
946 | ||
3a4295d1 ME |
947 | if (best >= 0) { |
948 | dev_addr = of_read_number(ranges + best, naddr); | |
949 | } else | |
99e13912 ME |
950 | dev_dbg(dev, "iommu: no suitable range found!\n"); |
951 | ||
952 | out: | |
953 | of_node_put(np); | |
954 | ||
3a4295d1 | 955 | return dev_addr; |
99e13912 ME |
956 | } |
957 | ||
958 | static int dma_set_mask_and_switch(struct device *dev, u64 dma_mask) | |
959 | { | |
960 | if (!dev->dma_mask || !dma_supported(dev, dma_mask)) | |
961 | return -EIO; | |
962 | ||
4a8df150 ME |
963 | if (dma_mask == DMA_BIT_MASK(64) && |
964 | cell_iommu_get_fixed_address(dev) != OF_BAD_ADDR) | |
965 | { | |
966 | dev_dbg(dev, "iommu: 64-bit OK, using fixed ops\n"); | |
967 | set_dma_ops(dev, &dma_iommu_fixed_ops); | |
99e13912 ME |
968 | } else { |
969 | dev_dbg(dev, "iommu: not 64-bit, using default ops\n"); | |
970 | set_dma_ops(dev, get_pci_dma_ops()); | |
971 | } | |
972 | ||
4a8df150 ME |
973 | cell_dma_dev_setup(dev); |
974 | ||
99e13912 ME |
975 | *dev->dma_mask = dma_mask; |
976 | ||
977 | return 0; | |
978 | } | |
979 | ||
f9660e8a | 980 | static void cell_dma_dev_setup_fixed(struct device *dev) |
99e13912 ME |
981 | { |
982 | struct dev_archdata *archdata = &dev->archdata; | |
983 | u64 addr; | |
984 | ||
985 | addr = cell_iommu_get_fixed_address(dev) + dma_iommu_fixed_base; | |
986 | archdata->dma_data = (void *)addr; | |
987 | ||
988 | dev_dbg(dev, "iommu: fixed addr = %lx\n", addr); | |
989 | } | |
990 | ||
da40451b ME |
991 | static void insert_16M_pte(unsigned long addr, unsigned long *ptab, |
992 | unsigned long base_pte) | |
993 | { | |
994 | unsigned long segment, offset; | |
995 | ||
996 | segment = addr >> IO_SEGMENT_SHIFT; | |
997 | offset = (addr >> 24) - (segment << IO_PAGENO_BITS(24)); | |
998 | ptab = ptab + (segment * (1 << 12) / sizeof(unsigned long)); | |
999 | ||
1000 | pr_debug("iommu: addr %lx ptab %p segment %lx offset %lx\n", | |
1001 | addr, ptab, segment, offset); | |
1002 | ||
1003 | ptab[offset] = base_pte | (__pa(addr) & IOPTE_RPN_Mask); | |
1004 | } | |
1005 | ||
99e13912 ME |
1006 | static void cell_iommu_setup_fixed_ptab(struct cbe_iommu *iommu, |
1007 | struct device_node *np, unsigned long dbase, unsigned long dsize, | |
1008 | unsigned long fbase, unsigned long fsize) | |
1009 | { | |
da40451b | 1010 | unsigned long base_pte, uaddr, ioaddr, *ptab; |
99e13912 | 1011 | |
da40451b | 1012 | ptab = cell_iommu_alloc_ptab(iommu, fbase, fsize, dbase, dsize, 24); |
99e13912 | 1013 | |
99e13912 | 1014 | dma_iommu_fixed_base = fbase; |
99e13912 ME |
1015 | |
1016 | pr_debug("iommu: mapping 0x%lx pages from 0x%lx\n", fsize, fbase); | |
1017 | ||
1ed6af73 | 1018 | base_pte = IOPTE_PP_W | IOPTE_PP_R | IOPTE_M |
99e13912 ME |
1019 | | (cell_iommu_get_ioid(np) & IOPTE_IOID_Mask); |
1020 | ||
1ed6af73 MN |
1021 | if (iommu_fixed_is_weak) |
1022 | pr_info("IOMMU: Using weak ordering for fixed mapping\n"); | |
1023 | else { | |
1024 | pr_info("IOMMU: Using strong ordering for fixed mapping\n"); | |
1025 | base_pte |= IOPTE_SO_RW; | |
1026 | } | |
1027 | ||
da40451b | 1028 | for (uaddr = 0; uaddr < fsize; uaddr += (1 << 24)) { |
99e13912 | 1029 | /* Don't touch the dynamic region */ |
da40451b ME |
1030 | ioaddr = uaddr + fbase; |
1031 | if (ioaddr >= dbase && ioaddr < (dbase + dsize)) { | |
f9660e8a | 1032 | pr_debug("iommu: fixed/dynamic overlap, skipping\n"); |
99e13912 ME |
1033 | continue; |
1034 | } | |
da40451b ME |
1035 | |
1036 | insert_16M_pte(uaddr, ptab, base_pte); | |
99e13912 ME |
1037 | } |
1038 | ||
1039 | mb(); | |
1040 | } | |
1041 | ||
1042 | static int __init cell_iommu_fixed_mapping_init(void) | |
1043 | { | |
1044 | unsigned long dbase, dsize, fbase, fsize, hbase, hend; | |
1045 | struct cbe_iommu *iommu; | |
1046 | struct device_node *np; | |
1047 | ||
1048 | /* The fixed mapping is only supported on axon machines */ | |
1049 | np = of_find_node_by_name(NULL, "axon"); | |
1050 | if (!np) { | |
1051 | pr_debug("iommu: fixed mapping disabled, no axons found\n"); | |
1052 | return -1; | |
1053 | } | |
1054 | ||
0e0b47ab ME |
1055 | /* We must have dma-ranges properties for fixed mapping to work */ |
1056 | for (np = NULL; (np = of_find_all_nodes(np));) { | |
1057 | if (of_find_property(np, "dma-ranges", NULL)) | |
1058 | break; | |
1059 | } | |
1060 | of_node_put(np); | |
1061 | ||
1062 | if (!np) { | |
1063 | pr_debug("iommu: no dma-ranges found, no fixed mapping\n"); | |
1064 | return -1; | |
1065 | } | |
1066 | ||
99e13912 ME |
1067 | /* The default setup is to have the fixed mapping sit after the |
1068 | * dynamic region, so find the top of the largest IOMMU window | |
1069 | * on any axon, then add the size of RAM and that's our max value. | |
1070 | * If that is > 32GB we have to do other shennanigans. | |
1071 | */ | |
1072 | fbase = 0; | |
1073 | for_each_node_by_name(np, "axon") { | |
1074 | cell_iommu_get_window(np, &dbase, &dsize); | |
1075 | fbase = max(fbase, dbase + dsize); | |
1076 | } | |
1077 | ||
1078 | fbase = _ALIGN_UP(fbase, 1 << IO_SEGMENT_SHIFT); | |
1079 | fsize = lmb_phys_mem_size(); | |
1080 | ||
1081 | if ((fbase + fsize) <= 0x800000000) | |
1082 | hbase = 0; /* use the device tree window */ | |
1083 | else { | |
1084 | /* If we're over 32 GB we need to cheat. We can't map all of | |
1085 | * RAM with the fixed mapping, and also fit the dynamic | |
1086 | * region. So try to place the dynamic region where the hash | |
1087 | * table sits, drivers never need to DMA to it, we don't | |
1088 | * need a fixed mapping for that area. | |
1089 | */ | |
1090 | if (!htab_address) { | |
1091 | pr_debug("iommu: htab is NULL, on LPAR? Huh?\n"); | |
1092 | return -1; | |
1093 | } | |
1094 | hbase = __pa(htab_address); | |
1095 | hend = hbase + htab_size_bytes; | |
1096 | ||
1097 | /* The window must start and end on a segment boundary */ | |
1098 | if ((hbase != _ALIGN_UP(hbase, 1 << IO_SEGMENT_SHIFT)) || | |
1099 | (hend != _ALIGN_UP(hend, 1 << IO_SEGMENT_SHIFT))) { | |
1100 | pr_debug("iommu: hash window not segment aligned\n"); | |
1101 | return -1; | |
1102 | } | |
1103 | ||
1104 | /* Check the hash window fits inside the real DMA window */ | |
1105 | for_each_node_by_name(np, "axon") { | |
1106 | cell_iommu_get_window(np, &dbase, &dsize); | |
1107 | ||
1108 | if (hbase < dbase || (hend > (dbase + dsize))) { | |
1109 | pr_debug("iommu: hash window doesn't fit in" | |
1110 | "real DMA window\n"); | |
1111 | return -1; | |
1112 | } | |
1113 | } | |
1114 | ||
1115 | fbase = 0; | |
1116 | } | |
1117 | ||
1118 | /* Setup the dynamic regions */ | |
1119 | for_each_node_by_name(np, "axon") { | |
1120 | iommu = cell_iommu_alloc(np); | |
1121 | BUG_ON(!iommu); | |
1122 | ||
1123 | if (hbase == 0) | |
1124 | cell_iommu_get_window(np, &dbase, &dsize); | |
1125 | else { | |
1126 | dbase = hbase; | |
1127 | dsize = htab_size_bytes; | |
1128 | } | |
1129 | ||
44621be4 ME |
1130 | printk(KERN_DEBUG "iommu: node %d, dynamic window 0x%lx-0x%lx " |
1131 | "fixed window 0x%lx-0x%lx\n", iommu->nid, dbase, | |
99e13912 ME |
1132 | dbase + dsize, fbase, fbase + fsize); |
1133 | ||
7d432ff1 | 1134 | cell_iommu_setup_stab(iommu, dbase, dsize, fbase, fsize); |
225d4905 ME |
1135 | iommu->ptab = cell_iommu_alloc_ptab(iommu, dbase, dsize, 0, 0, |
1136 | IOMMU_PAGE_SHIFT); | |
99e13912 ME |
1137 | cell_iommu_setup_fixed_ptab(iommu, np, dbase, dsize, |
1138 | fbase, fsize); | |
1139 | cell_iommu_enable_hardware(iommu); | |
1140 | cell_iommu_setup_window(iommu, np, dbase, dsize, 0); | |
1141 | } | |
1142 | ||
99e13912 ME |
1143 | dma_iommu_ops.set_dma_mask = dma_set_mask_and_switch; |
1144 | set_pci_dma_ops(&dma_iommu_ops); | |
1145 | ||
99e13912 ME |
1146 | return 0; |
1147 | } | |
1148 | ||
1149 | static int iommu_fixed_disabled; | |
1150 | ||
1151 | static int __init setup_iommu_fixed(char *str) | |
1152 | { | |
7886250e MN |
1153 | struct device_node *pciep; |
1154 | ||
99e13912 ME |
1155 | if (strcmp(str, "off") == 0) |
1156 | iommu_fixed_disabled = 1; | |
1157 | ||
7886250e MN |
1158 | /* If we can find a pcie-endpoint in the device tree assume that |
1159 | * we're on a triblade or a CAB so by default the fixed mapping | |
1160 | * should be set to be weakly ordered; but only if the boot | |
1161 | * option WASN'T set for strong ordering | |
1162 | */ | |
1163 | pciep = of_find_node_by_type(NULL, "pcie-endpoint"); | |
1164 | ||
1165 | if (strcmp(str, "weak") == 0 || (pciep && strcmp(str, "strong") != 0)) | |
1ed6af73 MN |
1166 | iommu_fixed_is_weak = 1; |
1167 | ||
7886250e MN |
1168 | of_node_put(pciep); |
1169 | ||
99e13912 ME |
1170 | return 1; |
1171 | } | |
1172 | __setup("iommu_fixed=", setup_iommu_fixed); | |
1173 | ||
165785e5 | 1174 | static int __init cell_iommu_init(void) |
ae209cf1 | 1175 | { |
165785e5 JK |
1176 | struct device_node *np; |
1177 | ||
165785e5 JK |
1178 | /* If IOMMU is disabled or we have little enough RAM to not need |
1179 | * to enable it, we setup a direct mapping. | |
1180 | * | |
1181 | * Note: should we make sure we have the IOMMU actually disabled ? | |
1182 | */ | |
1183 | if (iommu_is_off || | |
1184 | (!iommu_force_on && lmb_end_of_DRAM() <= 0x80000000ull)) | |
1185 | if (cell_iommu_init_disabled() == 0) | |
1186 | goto bail; | |
1187 | ||
1188 | /* Setup various ppc_md. callbacks */ | |
1189 | ppc_md.pci_dma_dev_setup = cell_pci_dma_dev_setup; | |
1190 | ppc_md.tce_build = tce_build_cell; | |
1191 | ppc_md.tce_free = tce_free_cell; | |
1192 | ||
99e13912 ME |
1193 | if (!iommu_fixed_disabled && cell_iommu_fixed_mapping_init() == 0) |
1194 | goto bail; | |
1195 | ||
165785e5 JK |
1196 | /* Create an iommu for each /axon node. */ |
1197 | for_each_node_by_name(np, "axon") { | |
1198 | if (np->parent == NULL || np->parent->parent != NULL) | |
1199 | continue; | |
1200 | cell_iommu_init_one(np, 0); | |
49d65b3a | 1201 | } |
ae209cf1 | 1202 | |
165785e5 JK |
1203 | /* Create an iommu for each toplevel /pci-internal node for |
1204 | * old hardware/firmware | |
1205 | */ | |
1206 | for_each_node_by_name(np, "pci-internal") { | |
1207 | if (np->parent == NULL || np->parent->parent != NULL) | |
1208 | continue; | |
1209 | cell_iommu_init_one(np, SPIDER_DMA_OFFSET); | |
1210 | } | |
1211 | ||
1212 | /* Setup default PCI iommu ops */ | |
98747770 | 1213 | set_pci_dma_ops(&dma_iommu_ops); |
165785e5 JK |
1214 | |
1215 | bail: | |
1216 | /* Register callbacks on OF platform device addition/removal | |
1217 | * to handle linking them to the right DMA operations | |
1218 | */ | |
1219 | bus_register_notifier(&of_platform_bus_type, &cell_of_bus_notifier); | |
1220 | ||
1221 | return 0; | |
ae209cf1 | 1222 | } |
e25c47ff GL |
1223 | machine_arch_initcall(cell, cell_iommu_init); |
1224 | machine_arch_initcall(celleb_native, cell_iommu_init); | |
165785e5 | 1225 |