1 /*
2  * IOMMU implementation for Cell Broadband Processor Architecture
3  *
4  * (C) Copyright IBM Corporation 2006-2008
5  *
6  * Author: Jeremy Kerr <jk@ozlabs.org>
7  *
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.
21  */
22 
23 #undef DEBUG
24 
25 #include <linux/kernel.h>
26 #include <linux/init.h>
27 #include <linux/interrupt.h>
28 #include <linux/notifier.h>
29 #include <linux/of.h>
30 #include <linux/of_platform.h>
31 
32 #include <asm/prom.h>
33 #include <asm/iommu.h>
34 #include <asm/machdep.h>
35 #include <asm/pci-bridge.h>
36 #include <asm/udbg.h>
37 #include <asm/lmb.h>
38 #include <asm/firmware.h>
39 #include <asm/cell-regs.h>
40 
41 #include "interrupt.h"
42 
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
116 #define IO_PAGENO_BITS(shift)	(IO_SEGMENT_SHIFT - (shift))
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;
126 	unsigned int ioid;
127 	struct iommu_table table;
128 };
129 
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 };
141 
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;
149 
150 static void invalidate_tce_cache(struct cbe_iommu *iommu, unsigned long *pte,
151 		long n_ptes)
152 {
153 	unsigned long __iomem *reg;
154 	unsigned long val;
155 	long n;
156 
157 	reg = iommu->xlate_regs + IOC_IOPT_CacheInvd;
158 
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;
165 
166 		out_be64(reg, val);
167 		while (in_be64(reg) & IOC_IOPT_CacheInvd_Busy)
168 			;
169 
170 		n_ptes -= n;
171 		pte += n;
172 	}
173 }
174 
175 static void tce_build_cell(struct iommu_table *tbl, long index, long npages,
176 		unsigned long uaddr, enum dma_data_direction direction)
177 {
178 	int i;
179 	unsigned long *io_pte, base_pte;
180 	struct iommu_window *window =
181 		container_of(tbl, struct iommu_window, table);
182 
183 	/* implementing proper protection causes problems with the spidernet
184 	 * driver - check mapping directions later, but allow read & write by
185 	 * default for now.*/
186 #ifdef CELL_IOMMU_STRICT_PROTECTION
187 	/* to avoid referencing a global, we use a trick here to setup the
188 	 * protection bit. "prot" is setup to be 3 fields of 4 bits apprended
189 	 * together for each of the 3 supported direction values. It is then
190 	 * shifted left so that the fields matching the desired direction
191 	 * lands on the appropriate bits, and other bits are masked out.
192 	 */
193 	const unsigned long prot = 0xc48;
194 	base_pte =
195 		((prot << (52 + 4 * direction)) & (IOPTE_PP_W | IOPTE_PP_R))
196 		| IOPTE_M | IOPTE_SO_RW | (window->ioid & IOPTE_IOID_Mask);
197 #else
198 	base_pte = IOPTE_PP_W | IOPTE_PP_R | IOPTE_M | IOPTE_SO_RW |
199 		(window->ioid & IOPTE_IOID_Mask);
200 #endif
201 
202 	io_pte = (unsigned long *)tbl->it_base + (index - tbl->it_offset);
203 
204 	for (i = 0; i < npages; i++, uaddr += IOMMU_PAGE_SIZE)
205 		io_pte[i] = base_pte | (__pa(uaddr) & IOPTE_RPN_Mask);
206 
207 	mb();
208 
209 	invalidate_tce_cache(window->iommu, io_pte, npages);
210 
211 	pr_debug("tce_build_cell(index=%lx,n=%lx,dir=%d,base_pte=%lx)\n",
212 		 index, npages, direction, base_pte);
213 }
214 
215 static void tce_free_cell(struct iommu_table *tbl, long index, long npages)
216 {
217 
218 	int i;
219 	unsigned long *io_pte, pte;
220 	struct iommu_window *window =
221 		container_of(tbl, struct iommu_window, table);
222 
223 	pr_debug("tce_free_cell(index=%lx,n=%lx)\n", index, npages);
224 
225 #ifdef CELL_IOMMU_REAL_UNMAP
226 	pte = 0;
227 #else
228 	/* spider bridge does PCI reads after freeing - insert a mapping
229 	 * to a scratch page instead of an invalid entry */
230 	pte = IOPTE_PP_R | IOPTE_M | IOPTE_SO_RW | __pa(window->iommu->pad_page)
231 		| (window->ioid & IOPTE_IOID_Mask);
232 #endif
233 
234 	io_pte = (unsigned long *)tbl->it_base + (index - tbl->it_offset);
235 
236 	for (i = 0; i < npages; i++)
237 		io_pte[i] = pte;
238 
239 	mb();
240 
241 	invalidate_tce_cache(window->iommu, io_pte, npages);
242 }
243 
244 static irqreturn_t ioc_interrupt(int irq, void *data)
245 {
246 	unsigned long stat;
247 	struct cbe_iommu *iommu = data;
248 
249 	stat = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat);
250 
251 	/* Might want to rate limit it */
252 	printk(KERN_ERR "iommu: DMA exception 0x%016lx\n", stat);
253 	printk(KERN_ERR "  V=%d, SPF=[%c%c], RW=%s, IOID=0x%04x\n",
254 	       !!(stat & IOC_IO_ExcpStat_V),
255 	       (stat & IOC_IO_ExcpStat_SPF_S) ? 'S' : ' ',
256 	       (stat & IOC_IO_ExcpStat_SPF_P) ? 'P' : ' ',
257 	       (stat & IOC_IO_ExcpStat_RW_Mask) ? "Read" : "Write",
258 	       (unsigned int)(stat & IOC_IO_ExcpStat_IOID_Mask));
259 	printk(KERN_ERR "  page=0x%016lx\n",
260 	       stat & IOC_IO_ExcpStat_ADDR_Mask);
261 
262 	/* clear interrupt */
263 	stat &= ~IOC_IO_ExcpStat_V;
264 	out_be64(iommu->xlate_regs + IOC_IO_ExcpStat, stat);
265 
266 	return IRQ_HANDLED;
267 }
268 
269 static int cell_iommu_find_ioc(int nid, unsigned long *base)
270 {
271 	struct device_node *np;
272 	struct resource r;
273 
274 	*base = 0;
275 
276 	/* First look for new style /be nodes */
277 	for_each_node_by_name(np, "ioc") {
278 		if (of_node_to_nid(np) != nid)
279 			continue;
280 		if (of_address_to_resource(np, 0, &r)) {
281 			printk(KERN_ERR "iommu: can't get address for %s\n",
282 			       np->full_name);
283 			continue;
284 		}
285 		*base = r.start;
286 		of_node_put(np);
287 		return 0;
288 	}
289 
290 	/* Ok, let's try the old way */
291 	for_each_node_by_type(np, "cpu") {
292 		const unsigned int *nidp;
293 		const unsigned long *tmp;
294 
295 		nidp = of_get_property(np, "node-id", NULL);
296 		if (nidp && *nidp == nid) {
297 			tmp = of_get_property(np, "ioc-translation", NULL);
298 			if (tmp) {
299 				*base = *tmp;
300 				of_node_put(np);
301 				return 0;
302 			}
303 		}
304 	}
305 
306 	return -ENODEV;
307 }
308 
309 static void cell_iommu_setup_stab(struct cbe_iommu *iommu,
310 				unsigned long dbase, unsigned long dsize,
311 				unsigned long fbase, unsigned long fsize)
312 {
313 	struct page *page;
314 	unsigned long segments, stab_size;
315 
316 	segments = max(dbase + dsize, fbase + fsize) >> IO_SEGMENT_SHIFT;
317 
318 	pr_debug("%s: iommu[%d]: segments: %lu\n",
319 			__FUNCTION__, iommu->nid, segments);
320 
321 	/* set up the segment table */
322 	stab_size = segments * sizeof(unsigned long);
323 	page = alloc_pages_node(iommu->nid, GFP_KERNEL, get_order(stab_size));
324 	BUG_ON(!page);
325 	iommu->stab = page_address(page);
326 	memset(iommu->stab, 0, stab_size);
327 }
328 
329 static unsigned long *cell_iommu_alloc_ptab(struct cbe_iommu *iommu,
330 		unsigned long base, unsigned long size, unsigned long gap_base,
331 		unsigned long gap_size, unsigned long page_shift)
332 {
333 	struct page *page;
334 	int i;
335 	unsigned long reg, segments, pages_per_segment, ptab_size,
336 		      n_pte_pages, start_seg, *ptab;
337 
338 	start_seg = base >> IO_SEGMENT_SHIFT;
339 	segments  = size >> IO_SEGMENT_SHIFT;
340 	pages_per_segment = 1ull << IO_PAGENO_BITS(page_shift);
341 	/* PTEs for each segment must start on a 4K bounday */
342 	pages_per_segment = max(pages_per_segment,
343 				(1 << 12) / sizeof(unsigned long));
344 
345 	ptab_size = segments * pages_per_segment * sizeof(unsigned long);
346 	pr_debug("%s: iommu[%d]: ptab_size: %lu, order: %d\n", __FUNCTION__,
347 			iommu->nid, ptab_size, get_order(ptab_size));
348 	page = alloc_pages_node(iommu->nid, GFP_KERNEL, get_order(ptab_size));
349 	BUG_ON(!page);
350 
351 	ptab = page_address(page);
352 	memset(ptab, 0, ptab_size);
353 
354 	/* number of 4K pages needed for a page table */
355 	n_pte_pages = (pages_per_segment * sizeof(unsigned long)) >> 12;
356 
357 	pr_debug("%s: iommu[%d]: stab at %p, ptab at %p, n_pte_pages: %lu\n",
358 			__FUNCTION__, iommu->nid, iommu->stab, ptab,
359 			n_pte_pages);
360 
361 	/* initialise the STEs */
362 	reg = IOSTE_V | ((n_pte_pages - 1) << 5);
363 
364 	switch (page_shift) {
365 	case 12: reg |= IOSTE_PS_4K;  break;
366 	case 16: reg |= IOSTE_PS_64K; break;
367 	case 20: reg |= IOSTE_PS_1M;  break;
368 	case 24: reg |= IOSTE_PS_16M; break;
369 	default: BUG();
370 	}
371 
372 	gap_base = gap_base >> IO_SEGMENT_SHIFT;
373 	gap_size = gap_size >> IO_SEGMENT_SHIFT;
374 
375 	pr_debug("Setting up IOMMU stab:\n");
376 	for (i = start_seg; i < (start_seg + segments); i++) {
377 		if (i >= gap_base && i < (gap_base + gap_size)) {
378 			pr_debug("\toverlap at %d, skipping\n", i);
379 			continue;
380 		}
381 		iommu->stab[i] = reg | (__pa(ptab) + (n_pte_pages << 12) *
382 					(i - start_seg));
383 		pr_debug("\t[%d] 0x%016lx\n", i, iommu->stab[i]);
384 	}
385 
386 	return ptab;
387 }
388 
389 static void cell_iommu_enable_hardware(struct cbe_iommu *iommu)
390 {
391 	int ret;
392 	unsigned long reg, xlate_base;
393 	unsigned int virq;
394 
395 	if (cell_iommu_find_ioc(iommu->nid, &xlate_base))
396 		panic("%s: missing IOC register mappings for node %d\n",
397 		      __FUNCTION__, iommu->nid);
398 
399 	iommu->xlate_regs = ioremap(xlate_base, IOC_Reg_Size);
400 	iommu->cmd_regs = iommu->xlate_regs + IOC_IOCmd_Offset;
401 
402 	/* ensure that the STEs have updated */
403 	mb();
404 
405 	/* setup interrupts for the iommu. */
406 	reg = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat);
407 	out_be64(iommu->xlate_regs + IOC_IO_ExcpStat,
408 			reg & ~IOC_IO_ExcpStat_V);
409 	out_be64(iommu->xlate_regs + IOC_IO_ExcpMask,
410 			IOC_IO_ExcpMask_PFE | IOC_IO_ExcpMask_SFE);
411 
412 	virq = irq_create_mapping(NULL,
413 			IIC_IRQ_IOEX_ATI | (iommu->nid << IIC_IRQ_NODE_SHIFT));
414 	BUG_ON(virq == NO_IRQ);
415 
416 	ret = request_irq(virq, ioc_interrupt, IRQF_DISABLED,
417 			iommu->name, iommu);
418 	BUG_ON(ret);
419 
420 	/* set the IOC segment table origin register (and turn on the iommu) */
421 	reg = IOC_IOST_Origin_E | __pa(iommu->stab) | IOC_IOST_Origin_HW;
422 	out_be64(iommu->xlate_regs + IOC_IOST_Origin, reg);
423 	in_be64(iommu->xlate_regs + IOC_IOST_Origin);
424 
425 	/* turn on IO translation */
426 	reg = in_be64(iommu->cmd_regs + IOC_IOCmd_Cfg) | IOC_IOCmd_Cfg_TE;
427 	out_be64(iommu->cmd_regs + IOC_IOCmd_Cfg, reg);
428 }
429 
430 static void cell_iommu_setup_hardware(struct cbe_iommu *iommu,
431 	unsigned long base, unsigned long size)
432 {
433 	cell_iommu_setup_stab(iommu, base, size, 0, 0);
434 	iommu->ptab = cell_iommu_alloc_ptab(iommu, base, size, 0, 0,
435 					    IOMMU_PAGE_SHIFT);
436 	cell_iommu_enable_hardware(iommu);
437 }
438 
439 #if 0/* Unused for now */
440 static struct iommu_window *find_window(struct cbe_iommu *iommu,
441 		unsigned long offset, unsigned long size)
442 {
443 	struct iommu_window *window;
444 
445 	/* todo: check for overlapping (but not equal) windows) */
446 
447 	list_for_each_entry(window, &(iommu->windows), list) {
448 		if (window->offset == offset && window->size == size)
449 			return window;
450 	}
451 
452 	return NULL;
453 }
454 #endif
455 
456 static inline u32 cell_iommu_get_ioid(struct device_node *np)
457 {
458 	const u32 *ioid;
459 
460 	ioid = of_get_property(np, "ioid", NULL);
461 	if (ioid == NULL) {
462 		printk(KERN_WARNING "iommu: missing ioid for %s using 0\n",
463 		       np->full_name);
464 		return 0;
465 	}
466 
467 	return *ioid;
468 }
469 
470 static struct iommu_window * __init
471 cell_iommu_setup_window(struct cbe_iommu *iommu, struct device_node *np,
472 			unsigned long offset, unsigned long size,
473 			unsigned long pte_offset)
474 {
475 	struct iommu_window *window;
476 	struct page *page;
477 	u32 ioid;
478 
479 	ioid = cell_iommu_get_ioid(np);
480 
481 	window = kmalloc_node(sizeof(*window), GFP_KERNEL, iommu->nid);
482 	BUG_ON(window == NULL);
483 
484 	window->offset = offset;
485 	window->size = size;
486 	window->ioid = ioid;
487 	window->iommu = iommu;
488 
489 	window->table.it_blocksize = 16;
490 	window->table.it_base = (unsigned long)iommu->ptab;
491 	window->table.it_index = iommu->nid;
492 	window->table.it_offset = (offset >> IOMMU_PAGE_SHIFT) + pte_offset;
493 	window->table.it_size = size >> IOMMU_PAGE_SHIFT;
494 
495 	iommu_init_table(&window->table, iommu->nid);
496 
497 	pr_debug("\tioid      %d\n", window->ioid);
498 	pr_debug("\tblocksize %ld\n", window->table.it_blocksize);
499 	pr_debug("\tbase      0x%016lx\n", window->table.it_base);
500 	pr_debug("\toffset    0x%lx\n", window->table.it_offset);
501 	pr_debug("\tsize      %ld\n", window->table.it_size);
502 
503 	list_add(&window->list, &iommu->windows);
504 
505 	if (offset != 0)
506 		return window;
507 
508 	/* We need to map and reserve the first IOMMU page since it's used
509 	 * by the spider workaround. In theory, we only need to do that when
510 	 * running on spider but it doesn't really matter.
511 	 *
512 	 * This code also assumes that we have a window that starts at 0,
513 	 * which is the case on all spider based blades.
514 	 */
515 	page = alloc_pages_node(iommu->nid, GFP_KERNEL, 0);
516 	BUG_ON(!page);
517 	iommu->pad_page = page_address(page);
518 	clear_page(iommu->pad_page);
519 
520 	__set_bit(0, window->table.it_map);
521 	tce_build_cell(&window->table, window->table.it_offset, 1,
522 		       (unsigned long)iommu->pad_page, DMA_TO_DEVICE);
523 	window->table.it_hint = window->table.it_blocksize;
524 
525 	return window;
526 }
527 
528 static struct cbe_iommu *cell_iommu_for_node(int nid)
529 {
530 	int i;
531 
532 	for (i = 0; i < cbe_nr_iommus; i++)
533 		if (iommus[i].nid == nid)
534 			return &iommus[i];
535 	return NULL;
536 }
537 
538 static unsigned long cell_dma_direct_offset;
539 
540 static unsigned long dma_iommu_fixed_base;
541 struct dma_mapping_ops dma_iommu_fixed_ops;
542 
543 static void cell_dma_dev_setup_iommu(struct device *dev)
544 {
545 	struct iommu_window *window;
546 	struct cbe_iommu *iommu;
547 	struct dev_archdata *archdata = &dev->archdata;
548 
549 	/* Current implementation uses the first window available in that
550 	 * node's iommu. We -might- do something smarter later though it may
551 	 * never be necessary
552 	 */
553 	iommu = cell_iommu_for_node(archdata->numa_node);
554 	if (iommu == NULL || list_empty(&iommu->windows)) {
555 		printk(KERN_ERR "iommu: missing iommu for %s (node %d)\n",
556 		       archdata->of_node ? archdata->of_node->full_name : "?",
557 		       archdata->numa_node);
558 		return;
559 	}
560 	window = list_entry(iommu->windows.next, struct iommu_window, list);
561 
562 	archdata->dma_data = &window->table;
563 }
564 
565 static void cell_dma_dev_setup_fixed(struct device *dev);
566 
567 static void cell_dma_dev_setup(struct device *dev)
568 {
569 	struct dev_archdata *archdata = &dev->archdata;
570 
571 	/* Order is important here, these are not mutually exclusive */
572 	if (get_dma_ops(dev) == &dma_iommu_fixed_ops)
573 		cell_dma_dev_setup_fixed(dev);
574 	else if (get_pci_dma_ops() == &dma_iommu_ops)
575 		cell_dma_dev_setup_iommu(dev);
576 	else if (get_pci_dma_ops() == &dma_direct_ops)
577 		archdata->dma_data = (void *)cell_dma_direct_offset;
578 	else
579 		BUG();
580 }
581 
582 static void cell_pci_dma_dev_setup(struct pci_dev *dev)
583 {
584 	cell_dma_dev_setup(&dev->dev);
585 }
586 
587 static int cell_of_bus_notify(struct notifier_block *nb, unsigned long action,
588 			      void *data)
589 {
590 	struct device *dev = data;
591 
592 	/* We are only intereted in device addition */
593 	if (action != BUS_NOTIFY_ADD_DEVICE)
594 		return 0;
595 
596 	/* We use the PCI DMA ops */
597 	dev->archdata.dma_ops = get_pci_dma_ops();
598 
599 	cell_dma_dev_setup(dev);
600 
601 	return 0;
602 }
603 
604 static struct notifier_block cell_of_bus_notifier = {
605 	.notifier_call = cell_of_bus_notify
606 };
607 
608 static int __init cell_iommu_get_window(struct device_node *np,
609 					 unsigned long *base,
610 					 unsigned long *size)
611 {
612 	const void *dma_window;
613 	unsigned long index;
614 
615 	/* Use ibm,dma-window if available, else, hard code ! */
616 	dma_window = of_get_property(np, "ibm,dma-window", NULL);
617 	if (dma_window == NULL) {
618 		*base = 0;
619 		*size = 0x80000000u;
620 		return -ENODEV;
621 	}
622 
623 	of_parse_dma_window(np, dma_window, &index, base, size);
624 	return 0;
625 }
626 
627 static struct cbe_iommu * __init cell_iommu_alloc(struct device_node *np)
628 {
629 	struct cbe_iommu *iommu;
630 	int nid, i;
631 
632 	/* Get node ID */
633 	nid = of_node_to_nid(np);
634 	if (nid < 0) {
635 		printk(KERN_ERR "iommu: failed to get node for %s\n",
636 		       np->full_name);
637 		return NULL;
638 	}
639 	pr_debug("iommu: setting up iommu for node %d (%s)\n",
640 		 nid, np->full_name);
641 
642 	/* XXX todo: If we can have multiple windows on the same IOMMU, which
643 	 * isn't the case today, we probably want here to check wether the
644 	 * iommu for that node is already setup.
645 	 * However, there might be issue with getting the size right so let's
646 	 * ignore that for now. We might want to completely get rid of the
647 	 * multiple window support since the cell iommu supports per-page ioids
648 	 */
649 
650 	if (cbe_nr_iommus >= NR_IOMMUS) {
651 		printk(KERN_ERR "iommu: too many IOMMUs detected ! (%s)\n",
652 		       np->full_name);
653 		return NULL;
654 	}
655 
656 	/* Init base fields */
657 	i = cbe_nr_iommus++;
658 	iommu = &iommus[i];
659 	iommu->stab = NULL;
660 	iommu->nid = nid;
661 	snprintf(iommu->name, sizeof(iommu->name), "iommu%d", i);
662 	INIT_LIST_HEAD(&iommu->windows);
663 
664 	return iommu;
665 }
666 
667 static void __init cell_iommu_init_one(struct device_node *np,
668 				       unsigned long offset)
669 {
670 	struct cbe_iommu *iommu;
671 	unsigned long base, size;
672 
673 	iommu = cell_iommu_alloc(np);
674 	if (!iommu)
675 		return;
676 
677 	/* Obtain a window for it */
678 	cell_iommu_get_window(np, &base, &size);
679 
680 	pr_debug("\ttranslating window 0x%lx...0x%lx\n",
681 		 base, base + size - 1);
682 
683 	/* Initialize the hardware */
684 	cell_iommu_setup_hardware(iommu, base, size);
685 
686 	/* Setup the iommu_table */
687 	cell_iommu_setup_window(iommu, np, base, size,
688 				offset >> IOMMU_PAGE_SHIFT);
689 }
690 
691 static void __init cell_disable_iommus(void)
692 {
693 	int node;
694 	unsigned long base, val;
695 	void __iomem *xregs, *cregs;
696 
697 	/* Make sure IOC translation is disabled on all nodes */
698 	for_each_online_node(node) {
699 		if (cell_iommu_find_ioc(node, &base))
700 			continue;
701 		xregs = ioremap(base, IOC_Reg_Size);
702 		if (xregs == NULL)
703 			continue;
704 		cregs = xregs + IOC_IOCmd_Offset;
705 
706 		pr_debug("iommu: cleaning up iommu on node %d\n", node);
707 
708 		out_be64(xregs + IOC_IOST_Origin, 0);
709 		(void)in_be64(xregs + IOC_IOST_Origin);
710 		val = in_be64(cregs + IOC_IOCmd_Cfg);
711 		val &= ~IOC_IOCmd_Cfg_TE;
712 		out_be64(cregs + IOC_IOCmd_Cfg, val);
713 		(void)in_be64(cregs + IOC_IOCmd_Cfg);
714 
715 		iounmap(xregs);
716 	}
717 }
718 
719 static int __init cell_iommu_init_disabled(void)
720 {
721 	struct device_node *np = NULL;
722 	unsigned long base = 0, size;
723 
724 	/* When no iommu is present, we use direct DMA ops */
725 	set_pci_dma_ops(&dma_direct_ops);
726 
727 	/* First make sure all IOC translation is turned off */
728 	cell_disable_iommus();
729 
730 	/* If we have no Axon, we set up the spider DMA magic offset */
731 	if (of_find_node_by_name(NULL, "axon") == NULL)
732 		cell_dma_direct_offset = SPIDER_DMA_OFFSET;
733 
734 	/* Now we need to check to see where the memory is mapped
735 	 * in PCI space. We assume that all busses use the same dma
736 	 * window which is always the case so far on Cell, thus we
737 	 * pick up the first pci-internal node we can find and check
738 	 * the DMA window from there.
739 	 */
740 	for_each_node_by_name(np, "axon") {
741 		if (np->parent == NULL || np->parent->parent != NULL)
742 			continue;
743 		if (cell_iommu_get_window(np, &base, &size) == 0)
744 			break;
745 	}
746 	if (np == NULL) {
747 		for_each_node_by_name(np, "pci-internal") {
748 			if (np->parent == NULL || np->parent->parent != NULL)
749 				continue;
750 			if (cell_iommu_get_window(np, &base, &size) == 0)
751 				break;
752 		}
753 	}
754 	of_node_put(np);
755 
756 	/* If we found a DMA window, we check if it's big enough to enclose
757 	 * all of physical memory. If not, we force enable IOMMU
758 	 */
759 	if (np && size < lmb_end_of_DRAM()) {
760 		printk(KERN_WARNING "iommu: force-enabled, dma window"
761 		       " (%ldMB) smaller than total memory (%ldMB)\n",
762 		       size >> 20, lmb_end_of_DRAM() >> 20);
763 		return -ENODEV;
764 	}
765 
766 	cell_dma_direct_offset += base;
767 
768 	if (cell_dma_direct_offset != 0)
769 		ppc_md.pci_dma_dev_setup = cell_pci_dma_dev_setup;
770 
771 	printk("iommu: disabled, direct DMA offset is 0x%lx\n",
772 	       cell_dma_direct_offset);
773 
774 	return 0;
775 }
776 
777 /*
778  *  Fixed IOMMU mapping support
779  *
780  *  This code adds support for setting up a fixed IOMMU mapping on certain
781  *  cell machines. For 64-bit devices this avoids the performance overhead of
782  *  mapping and unmapping pages at runtime. 32-bit devices are unable to use
783  *  the fixed mapping.
784  *
785  *  The fixed mapping is established at boot, and maps all of physical memory
786  *  1:1 into device space at some offset. On machines with < 30 GB of memory
787  *  we setup the fixed mapping immediately above the normal IOMMU window.
788  *
789  *  For example a machine with 4GB of memory would end up with the normal
790  *  IOMMU window from 0-2GB and the fixed mapping window from 2GB to 6GB. In
791  *  this case a 64-bit device wishing to DMA to 1GB would be told to DMA to
792  *  3GB, plus any offset required by firmware. The firmware offset is encoded
793  *  in the "dma-ranges" property.
794  *
795  *  On machines with 30GB or more of memory, we are unable to place the fixed
796  *  mapping above the normal IOMMU window as we would run out of address space.
797  *  Instead we move the normal IOMMU window to coincide with the hash page
798  *  table, this region does not need to be part of the fixed mapping as no
799  *  device should ever be DMA'ing to it. We then setup the fixed mapping
800  *  from 0 to 32GB.
801  */
802 
803 static u64 cell_iommu_get_fixed_address(struct device *dev)
804 {
805 	u64 cpu_addr, size, best_size, dev_addr = OF_BAD_ADDR;
806 	struct device_node *np;
807 	const u32 *ranges = NULL;
808 	int i, len, best, naddr, nsize, pna, range_size;
809 
810 	np = of_node_get(dev->archdata.of_node);
811 	while (1) {
812 		naddr = of_n_addr_cells(np);
813 		nsize = of_n_size_cells(np);
814 		np = of_get_next_parent(np);
815 		if (!np)
816 			break;
817 
818 		ranges = of_get_property(np, "dma-ranges", &len);
819 
820 		/* Ignore empty ranges, they imply no translation required */
821 		if (ranges && len > 0)
822 			break;
823 	}
824 
825 	if (!ranges) {
826 		dev_dbg(dev, "iommu: no dma-ranges found\n");
827 		goto out;
828 	}
829 
830 	len /= sizeof(u32);
831 
832 	pna = of_n_addr_cells(np);
833 	range_size = naddr + nsize + pna;
834 
835 	/* dma-ranges format:
836 	 * child addr	: naddr cells
837 	 * parent addr	: pna cells
838 	 * size		: nsize cells
839 	 */
840 	for (i = 0, best = -1, best_size = 0; i < len; i += range_size) {
841 		cpu_addr = of_translate_dma_address(np, ranges + i + naddr);
842 		size = of_read_number(ranges + i + naddr + pna, nsize);
843 
844 		if (cpu_addr == 0 && size > best_size) {
845 			best = i;
846 			best_size = size;
847 		}
848 	}
849 
850 	if (best >= 0) {
851 		dev_addr = of_read_number(ranges + best, naddr);
852 	} else
853 		dev_dbg(dev, "iommu: no suitable range found!\n");
854 
855 out:
856 	of_node_put(np);
857 
858 	return dev_addr;
859 }
860 
861 static int dma_set_mask_and_switch(struct device *dev, u64 dma_mask)
862 {
863 	if (!dev->dma_mask || !dma_supported(dev, dma_mask))
864 		return -EIO;
865 
866 	if (dma_mask == DMA_BIT_MASK(64) &&
867 		cell_iommu_get_fixed_address(dev) != OF_BAD_ADDR)
868 	{
869 		dev_dbg(dev, "iommu: 64-bit OK, using fixed ops\n");
870 		set_dma_ops(dev, &dma_iommu_fixed_ops);
871 	} else {
872 		dev_dbg(dev, "iommu: not 64-bit, using default ops\n");
873 		set_dma_ops(dev, get_pci_dma_ops());
874 	}
875 
876 	cell_dma_dev_setup(dev);
877 
878 	*dev->dma_mask = dma_mask;
879 
880 	return 0;
881 }
882 
883 static void cell_dma_dev_setup_fixed(struct device *dev)
884 {
885 	struct dev_archdata *archdata = &dev->archdata;
886 	u64 addr;
887 
888 	addr = cell_iommu_get_fixed_address(dev) + dma_iommu_fixed_base;
889 	archdata->dma_data = (void *)addr;
890 
891 	dev_dbg(dev, "iommu: fixed addr = %lx\n", addr);
892 }
893 
894 static void insert_16M_pte(unsigned long addr, unsigned long *ptab,
895 			   unsigned long base_pte)
896 {
897 	unsigned long segment, offset;
898 
899 	segment = addr >> IO_SEGMENT_SHIFT;
900 	offset = (addr >> 24) - (segment << IO_PAGENO_BITS(24));
901 	ptab = ptab + (segment * (1 << 12) / sizeof(unsigned long));
902 
903 	pr_debug("iommu: addr %lx ptab %p segment %lx offset %lx\n",
904 		  addr, ptab, segment, offset);
905 
906 	ptab[offset] = base_pte | (__pa(addr) & IOPTE_RPN_Mask);
907 }
908 
909 static void cell_iommu_setup_fixed_ptab(struct cbe_iommu *iommu,
910 	struct device_node *np, unsigned long dbase, unsigned long dsize,
911 	unsigned long fbase, unsigned long fsize)
912 {
913 	unsigned long base_pte, uaddr, ioaddr, *ptab;
914 
915 	ptab = cell_iommu_alloc_ptab(iommu, fbase, fsize, dbase, dsize, 24);
916 
917 	dma_iommu_fixed_base = fbase;
918 
919 	pr_debug("iommu: mapping 0x%lx pages from 0x%lx\n", fsize, fbase);
920 
921 	base_pte = IOPTE_PP_W | IOPTE_PP_R | IOPTE_M | IOPTE_SO_RW
922 		    | (cell_iommu_get_ioid(np) & IOPTE_IOID_Mask);
923 
924 	for (uaddr = 0; uaddr < fsize; uaddr += (1 << 24)) {
925 		/* Don't touch the dynamic region */
926 		ioaddr = uaddr + fbase;
927 		if (ioaddr >= dbase && ioaddr < (dbase + dsize)) {
928 			pr_debug("iommu: fixed/dynamic overlap, skipping\n");
929 			continue;
930 		}
931 
932 		insert_16M_pte(uaddr, ptab, base_pte);
933 	}
934 
935 	mb();
936 }
937 
938 static int __init cell_iommu_fixed_mapping_init(void)
939 {
940 	unsigned long dbase, dsize, fbase, fsize, hbase, hend;
941 	struct cbe_iommu *iommu;
942 	struct device_node *np;
943 
944 	/* The fixed mapping is only supported on axon machines */
945 	np = of_find_node_by_name(NULL, "axon");
946 	if (!np) {
947 		pr_debug("iommu: fixed mapping disabled, no axons found\n");
948 		return -1;
949 	}
950 
951 	/* We must have dma-ranges properties for fixed mapping to work */
952 	for (np = NULL; (np = of_find_all_nodes(np));) {
953 		if (of_find_property(np, "dma-ranges", NULL))
954 			break;
955 	}
956 	of_node_put(np);
957 
958 	if (!np) {
959 		pr_debug("iommu: no dma-ranges found, no fixed mapping\n");
960 		return -1;
961 	}
962 
963 	/* The default setup is to have the fixed mapping sit after the
964 	 * dynamic region, so find the top of the largest IOMMU window
965 	 * on any axon, then add the size of RAM and that's our max value.
966 	 * If that is > 32GB we have to do other shennanigans.
967 	 */
968 	fbase = 0;
969 	for_each_node_by_name(np, "axon") {
970 		cell_iommu_get_window(np, &dbase, &dsize);
971 		fbase = max(fbase, dbase + dsize);
972 	}
973 
974 	fbase = _ALIGN_UP(fbase, 1 << IO_SEGMENT_SHIFT);
975 	fsize = lmb_phys_mem_size();
976 
977 	if ((fbase + fsize) <= 0x800000000)
978 		hbase = 0; /* use the device tree window */
979 	else {
980 		/* If we're over 32 GB we need to cheat. We can't map all of
981 		 * RAM with the fixed mapping, and also fit the dynamic
982 		 * region. So try to place the dynamic region where the hash
983 		 * table sits, drivers never need to DMA to it, we don't
984 		 * need a fixed mapping for that area.
985 		 */
986 		if (!htab_address) {
987 			pr_debug("iommu: htab is NULL, on LPAR? Huh?\n");
988 			return -1;
989 		}
990 		hbase = __pa(htab_address);
991 		hend  = hbase + htab_size_bytes;
992 
993 		/* The window must start and end on a segment boundary */
994 		if ((hbase != _ALIGN_UP(hbase, 1 << IO_SEGMENT_SHIFT)) ||
995 		    (hend != _ALIGN_UP(hend, 1 << IO_SEGMENT_SHIFT))) {
996 			pr_debug("iommu: hash window not segment aligned\n");
997 			return -1;
998 		}
999 
1000 		/* Check the hash window fits inside the real DMA window */
1001 		for_each_node_by_name(np, "axon") {
1002 			cell_iommu_get_window(np, &dbase, &dsize);
1003 
1004 			if (hbase < dbase || (hend > (dbase + dsize))) {
1005 				pr_debug("iommu: hash window doesn't fit in"
1006 					 "real DMA window\n");
1007 				return -1;
1008 			}
1009 		}
1010 
1011 		fbase = 0;
1012 	}
1013 
1014 	/* Setup the dynamic regions */
1015 	for_each_node_by_name(np, "axon") {
1016 		iommu = cell_iommu_alloc(np);
1017 		BUG_ON(!iommu);
1018 
1019 		if (hbase == 0)
1020 			cell_iommu_get_window(np, &dbase, &dsize);
1021 		else {
1022 			dbase = hbase;
1023 			dsize = htab_size_bytes;
1024 		}
1025 
1026 		printk(KERN_DEBUG "iommu: node %d, dynamic window 0x%lx-0x%lx "
1027 			"fixed window 0x%lx-0x%lx\n", iommu->nid, dbase,
1028 			 dbase + dsize, fbase, fbase + fsize);
1029 
1030 		cell_iommu_setup_stab(iommu, dbase, dsize, fbase, fsize);
1031 		iommu->ptab = cell_iommu_alloc_ptab(iommu, dbase, dsize, 0, 0,
1032 						    IOMMU_PAGE_SHIFT);
1033 		cell_iommu_setup_fixed_ptab(iommu, np, dbase, dsize,
1034 					     fbase, fsize);
1035 		cell_iommu_enable_hardware(iommu);
1036 		cell_iommu_setup_window(iommu, np, dbase, dsize, 0);
1037 	}
1038 
1039 	dma_iommu_fixed_ops = dma_direct_ops;
1040 	dma_iommu_fixed_ops.set_dma_mask = dma_set_mask_and_switch;
1041 
1042 	dma_iommu_ops.set_dma_mask = dma_set_mask_and_switch;
1043 	set_pci_dma_ops(&dma_iommu_ops);
1044 
1045 	return 0;
1046 }
1047 
1048 static int iommu_fixed_disabled;
1049 
1050 static int __init setup_iommu_fixed(char *str)
1051 {
1052 	if (strcmp(str, "off") == 0)
1053 		iommu_fixed_disabled = 1;
1054 
1055 	return 1;
1056 }
1057 __setup("iommu_fixed=", setup_iommu_fixed);
1058 
1059 static int __init cell_iommu_init(void)
1060 {
1061 	struct device_node *np;
1062 
1063 	/* If IOMMU is disabled or we have little enough RAM to not need
1064 	 * to enable it, we setup a direct mapping.
1065 	 *
1066 	 * Note: should we make sure we have the IOMMU actually disabled ?
1067 	 */
1068 	if (iommu_is_off ||
1069 	    (!iommu_force_on && lmb_end_of_DRAM() <= 0x80000000ull))
1070 		if (cell_iommu_init_disabled() == 0)
1071 			goto bail;
1072 
1073 	/* Setup various ppc_md. callbacks */
1074 	ppc_md.pci_dma_dev_setup = cell_pci_dma_dev_setup;
1075 	ppc_md.tce_build = tce_build_cell;
1076 	ppc_md.tce_free = tce_free_cell;
1077 
1078 	if (!iommu_fixed_disabled && cell_iommu_fixed_mapping_init() == 0)
1079 		goto bail;
1080 
1081 	/* Create an iommu for each /axon node.  */
1082 	for_each_node_by_name(np, "axon") {
1083 		if (np->parent == NULL || np->parent->parent != NULL)
1084 			continue;
1085 		cell_iommu_init_one(np, 0);
1086 	}
1087 
1088 	/* Create an iommu for each toplevel /pci-internal node for
1089 	 * old hardware/firmware
1090 	 */
1091 	for_each_node_by_name(np, "pci-internal") {
1092 		if (np->parent == NULL || np->parent->parent != NULL)
1093 			continue;
1094 		cell_iommu_init_one(np, SPIDER_DMA_OFFSET);
1095 	}
1096 
1097 	/* Setup default PCI iommu ops */
1098 	set_pci_dma_ops(&dma_iommu_ops);
1099 
1100  bail:
1101 	/* Register callbacks on OF platform device addition/removal
1102 	 * to handle linking them to the right DMA operations
1103 	 */
1104 	bus_register_notifier(&of_platform_bus_type, &cell_of_bus_notifier);
1105 
1106 	return 0;
1107 }
1108 machine_arch_initcall(cell, cell_iommu_init);
1109 machine_arch_initcall(celleb_native, cell_iommu_init);
1110 
1111