1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * IOMMU implementation for Cell Broadband Processor Architecture
4  *
5  * (C) Copyright IBM Corporation 2006-2008
6  *
7  * Author: Jeremy Kerr <jk@ozlabs.org>
8  */
9 
10 #undef DEBUG
11 
12 #include <linux/kernel.h>
13 #include <linux/init.h>
14 #include <linux/interrupt.h>
15 #include <linux/irqdomain.h>
16 #include <linux/notifier.h>
17 #include <linux/of.h>
18 #include <linux/of_address.h>
19 #include <linux/of_platform.h>
20 #include <linux/slab.h>
21 #include <linux/memblock.h>
22 
23 #include <asm/prom.h>
24 #include <asm/iommu.h>
25 #include <asm/machdep.h>
26 #include <asm/pci-bridge.h>
27 #include <asm/udbg.h>
28 #include <asm/firmware.h>
29 #include <asm/cell-regs.h>
30 
31 #include "cell.h"
32 #include "interrupt.h"
33 
34 /* Define CELL_IOMMU_REAL_UNMAP to actually unmap non-used pages
35  * instead of leaving them mapped to some dummy page. This can be
36  * enabled once the appropriate workarounds for spider bugs have
37  * been enabled
38  */
39 #define CELL_IOMMU_REAL_UNMAP
40 
41 /* Define CELL_IOMMU_STRICT_PROTECTION to enforce protection of
42  * IO PTEs based on the transfer direction. That can be enabled
43  * once spider-net has been fixed to pass the correct direction
44  * to the DMA mapping functions
45  */
46 #define CELL_IOMMU_STRICT_PROTECTION
47 
48 
49 #define NR_IOMMUS			2
50 
51 /* IOC mmap registers */
52 #define IOC_Reg_Size			0x2000
53 
54 #define IOC_IOPT_CacheInvd		0x908
55 #define IOC_IOPT_CacheInvd_NE_Mask	0xffe0000000000000ul
56 #define IOC_IOPT_CacheInvd_IOPTE_Mask	0x000003fffffffff8ul
57 #define IOC_IOPT_CacheInvd_Busy		0x0000000000000001ul
58 
59 #define IOC_IOST_Origin			0x918
60 #define IOC_IOST_Origin_E		0x8000000000000000ul
61 #define IOC_IOST_Origin_HW		0x0000000000000800ul
62 #define IOC_IOST_Origin_HL		0x0000000000000400ul
63 
64 #define IOC_IO_ExcpStat			0x920
65 #define IOC_IO_ExcpStat_V		0x8000000000000000ul
66 #define IOC_IO_ExcpStat_SPF_Mask	0x6000000000000000ul
67 #define IOC_IO_ExcpStat_SPF_S		0x6000000000000000ul
68 #define IOC_IO_ExcpStat_SPF_P		0x2000000000000000ul
69 #define IOC_IO_ExcpStat_ADDR_Mask	0x00000007fffff000ul
70 #define IOC_IO_ExcpStat_RW_Mask		0x0000000000000800ul
71 #define IOC_IO_ExcpStat_IOID_Mask	0x00000000000007fful
72 
73 #define IOC_IO_ExcpMask			0x928
74 #define IOC_IO_ExcpMask_SFE		0x4000000000000000ul
75 #define IOC_IO_ExcpMask_PFE		0x2000000000000000ul
76 
77 #define IOC_IOCmd_Offset		0x1000
78 
79 #define IOC_IOCmd_Cfg			0xc00
80 #define IOC_IOCmd_Cfg_TE		0x0000800000000000ul
81 
82 
83 /* Segment table entries */
84 #define IOSTE_V			0x8000000000000000ul /* valid */
85 #define IOSTE_H			0x4000000000000000ul /* cache hint */
86 #define IOSTE_PT_Base_RPN_Mask  0x3ffffffffffff000ul /* base RPN of IOPT */
87 #define IOSTE_NPPT_Mask		0x0000000000000fe0ul /* no. pages in IOPT */
88 #define IOSTE_PS_Mask		0x0000000000000007ul /* page size */
89 #define IOSTE_PS_4K		0x0000000000000001ul /*   - 4kB  */
90 #define IOSTE_PS_64K		0x0000000000000003ul /*   - 64kB */
91 #define IOSTE_PS_1M		0x0000000000000005ul /*   - 1MB  */
92 #define IOSTE_PS_16M		0x0000000000000007ul /*   - 16MB */
93 
94 
95 /* IOMMU sizing */
96 #define IO_SEGMENT_SHIFT	28
97 #define IO_PAGENO_BITS(shift)	(IO_SEGMENT_SHIFT - (shift))
98 
99 /* The high bit needs to be set on every DMA address */
100 #define SPIDER_DMA_OFFSET	0x80000000ul
101 
102 struct iommu_window {
103 	struct list_head list;
104 	struct cbe_iommu *iommu;
105 	unsigned long offset;
106 	unsigned long size;
107 	unsigned int ioid;
108 	struct iommu_table table;
109 };
110 
111 #define NAMESIZE 8
112 struct cbe_iommu {
113 	int nid;
114 	char name[NAMESIZE];
115 	void __iomem *xlate_regs;
116 	void __iomem *cmd_regs;
117 	unsigned long *stab;
118 	unsigned long *ptab;
119 	void *pad_page;
120 	struct list_head windows;
121 };
122 
123 /* Static array of iommus, one per node
124  *   each contains a list of windows, keyed from dma_window property
125  *   - on bus setup, look for a matching window, or create one
126  *   - on dev setup, assign iommu_table ptr
127  */
128 static struct cbe_iommu iommus[NR_IOMMUS];
129 static int cbe_nr_iommus;
130 
131 static void invalidate_tce_cache(struct cbe_iommu *iommu, unsigned long *pte,
132 		long n_ptes)
133 {
134 	u64 __iomem *reg;
135 	u64 val;
136 	long n;
137 
138 	reg = iommu->xlate_regs + IOC_IOPT_CacheInvd;
139 
140 	while (n_ptes > 0) {
141 		/* we can invalidate up to 1 << 11 PTEs at once */
142 		n = min(n_ptes, 1l << 11);
143 		val = (((n /*- 1*/) << 53) & IOC_IOPT_CacheInvd_NE_Mask)
144 			| (__pa(pte) & IOC_IOPT_CacheInvd_IOPTE_Mask)
145 		        | IOC_IOPT_CacheInvd_Busy;
146 
147 		out_be64(reg, val);
148 		while (in_be64(reg) & IOC_IOPT_CacheInvd_Busy)
149 			;
150 
151 		n_ptes -= n;
152 		pte += n;
153 	}
154 }
155 
156 static int tce_build_cell(struct iommu_table *tbl, long index, long npages,
157 		unsigned long uaddr, enum dma_data_direction direction,
158 		unsigned long attrs)
159 {
160 	int i;
161 	unsigned long *io_pte, base_pte;
162 	struct iommu_window *window =
163 		container_of(tbl, struct iommu_window, table);
164 
165 	/* implementing proper protection causes problems with the spidernet
166 	 * driver - check mapping directions later, but allow read & write by
167 	 * default for now.*/
168 #ifdef CELL_IOMMU_STRICT_PROTECTION
169 	/* to avoid referencing a global, we use a trick here to setup the
170 	 * protection bit. "prot" is setup to be 3 fields of 4 bits appended
171 	 * together for each of the 3 supported direction values. It is then
172 	 * shifted left so that the fields matching the desired direction
173 	 * lands on the appropriate bits, and other bits are masked out.
174 	 */
175 	const unsigned long prot = 0xc48;
176 	base_pte =
177 		((prot << (52 + 4 * direction)) &
178 		 (CBE_IOPTE_PP_W | CBE_IOPTE_PP_R)) |
179 		CBE_IOPTE_M | CBE_IOPTE_SO_RW |
180 		(window->ioid & CBE_IOPTE_IOID_Mask);
181 #else
182 	base_pte = CBE_IOPTE_PP_W | CBE_IOPTE_PP_R | CBE_IOPTE_M |
183 		CBE_IOPTE_SO_RW | (window->ioid & CBE_IOPTE_IOID_Mask);
184 #endif
185 	if (unlikely(attrs & DMA_ATTR_WEAK_ORDERING))
186 		base_pte &= ~CBE_IOPTE_SO_RW;
187 
188 	io_pte = (unsigned long *)tbl->it_base + (index - tbl->it_offset);
189 
190 	for (i = 0; i < npages; i++, uaddr += (1 << tbl->it_page_shift))
191 		io_pte[i] = base_pte | (__pa(uaddr) & CBE_IOPTE_RPN_Mask);
192 
193 	mb();
194 
195 	invalidate_tce_cache(window->iommu, io_pte, npages);
196 
197 	pr_debug("tce_build_cell(index=%lx,n=%lx,dir=%d,base_pte=%lx)\n",
198 		 index, npages, direction, base_pte);
199 	return 0;
200 }
201 
202 static void tce_free_cell(struct iommu_table *tbl, long index, long npages)
203 {
204 
205 	int i;
206 	unsigned long *io_pte, pte;
207 	struct iommu_window *window =
208 		container_of(tbl, struct iommu_window, table);
209 
210 	pr_debug("tce_free_cell(index=%lx,n=%lx)\n", index, npages);
211 
212 #ifdef CELL_IOMMU_REAL_UNMAP
213 	pte = 0;
214 #else
215 	/* spider bridge does PCI reads after freeing - insert a mapping
216 	 * to a scratch page instead of an invalid entry */
217 	pte = CBE_IOPTE_PP_R | CBE_IOPTE_M | CBE_IOPTE_SO_RW |
218 		__pa(window->iommu->pad_page) |
219 		(window->ioid & CBE_IOPTE_IOID_Mask);
220 #endif
221 
222 	io_pte = (unsigned long *)tbl->it_base + (index - tbl->it_offset);
223 
224 	for (i = 0; i < npages; i++)
225 		io_pte[i] = pte;
226 
227 	mb();
228 
229 	invalidate_tce_cache(window->iommu, io_pte, npages);
230 }
231 
232 static irqreturn_t ioc_interrupt(int irq, void *data)
233 {
234 	unsigned long stat, spf;
235 	struct cbe_iommu *iommu = data;
236 
237 	stat = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat);
238 	spf = stat & IOC_IO_ExcpStat_SPF_Mask;
239 
240 	/* Might want to rate limit it */
241 	printk(KERN_ERR "iommu: DMA exception 0x%016lx\n", stat);
242 	printk(KERN_ERR "  V=%d, SPF=[%c%c], RW=%s, IOID=0x%04x\n",
243 	       !!(stat & IOC_IO_ExcpStat_V),
244 	       (spf == IOC_IO_ExcpStat_SPF_S) ? 'S' : ' ',
245 	       (spf == IOC_IO_ExcpStat_SPF_P) ? 'P' : ' ',
246 	       (stat & IOC_IO_ExcpStat_RW_Mask) ? "Read" : "Write",
247 	       (unsigned int)(stat & IOC_IO_ExcpStat_IOID_Mask));
248 	printk(KERN_ERR "  page=0x%016lx\n",
249 	       stat & IOC_IO_ExcpStat_ADDR_Mask);
250 
251 	/* clear interrupt */
252 	stat &= ~IOC_IO_ExcpStat_V;
253 	out_be64(iommu->xlate_regs + IOC_IO_ExcpStat, stat);
254 
255 	return IRQ_HANDLED;
256 }
257 
258 static int __init cell_iommu_find_ioc(int nid, unsigned long *base)
259 {
260 	struct device_node *np;
261 	struct resource r;
262 
263 	*base = 0;
264 
265 	/* First look for new style /be nodes */
266 	for_each_node_by_name(np, "ioc") {
267 		if (of_node_to_nid(np) != nid)
268 			continue;
269 		if (of_address_to_resource(np, 0, &r)) {
270 			printk(KERN_ERR "iommu: can't get address for %pOF\n",
271 			       np);
272 			continue;
273 		}
274 		*base = r.start;
275 		of_node_put(np);
276 		return 0;
277 	}
278 
279 	/* Ok, let's try the old way */
280 	for_each_node_by_type(np, "cpu") {
281 		const unsigned int *nidp;
282 		const unsigned long *tmp;
283 
284 		nidp = of_get_property(np, "node-id", NULL);
285 		if (nidp && *nidp == nid) {
286 			tmp = of_get_property(np, "ioc-translation", NULL);
287 			if (tmp) {
288 				*base = *tmp;
289 				of_node_put(np);
290 				return 0;
291 			}
292 		}
293 	}
294 
295 	return -ENODEV;
296 }
297 
298 static void __init cell_iommu_setup_stab(struct cbe_iommu *iommu,
299 				unsigned long dbase, unsigned long dsize,
300 				unsigned long fbase, unsigned long fsize)
301 {
302 	struct page *page;
303 	unsigned long segments, stab_size;
304 
305 	segments = max(dbase + dsize, fbase + fsize) >> IO_SEGMENT_SHIFT;
306 
307 	pr_debug("%s: iommu[%d]: segments: %lu\n",
308 			__func__, iommu->nid, segments);
309 
310 	/* set up the segment table */
311 	stab_size = segments * sizeof(unsigned long);
312 	page = alloc_pages_node(iommu->nid, GFP_KERNEL, get_order(stab_size));
313 	BUG_ON(!page);
314 	iommu->stab = page_address(page);
315 	memset(iommu->stab, 0, stab_size);
316 }
317 
318 static unsigned long *__init cell_iommu_alloc_ptab(struct cbe_iommu *iommu,
319 		unsigned long base, unsigned long size, unsigned long gap_base,
320 		unsigned long gap_size, unsigned long page_shift)
321 {
322 	struct page *page;
323 	int i;
324 	unsigned long reg, segments, pages_per_segment, ptab_size,
325 		      n_pte_pages, start_seg, *ptab;
326 
327 	start_seg = base >> IO_SEGMENT_SHIFT;
328 	segments  = size >> IO_SEGMENT_SHIFT;
329 	pages_per_segment = 1ull << IO_PAGENO_BITS(page_shift);
330 	/* PTEs for each segment must start on a 4K boundary */
331 	pages_per_segment = max(pages_per_segment,
332 				(1 << 12) / sizeof(unsigned long));
333 
334 	ptab_size = segments * pages_per_segment * sizeof(unsigned long);
335 	pr_debug("%s: iommu[%d]: ptab_size: %lu, order: %d\n", __func__,
336 			iommu->nid, ptab_size, get_order(ptab_size));
337 	page = alloc_pages_node(iommu->nid, GFP_KERNEL, get_order(ptab_size));
338 	BUG_ON(!page);
339 
340 	ptab = page_address(page);
341 	memset(ptab, 0, ptab_size);
342 
343 	/* number of 4K pages needed for a page table */
344 	n_pte_pages = (pages_per_segment * sizeof(unsigned long)) >> 12;
345 
346 	pr_debug("%s: iommu[%d]: stab at %p, ptab at %p, n_pte_pages: %lu\n",
347 			__func__, iommu->nid, iommu->stab, ptab,
348 			n_pte_pages);
349 
350 	/* initialise the STEs */
351 	reg = IOSTE_V | ((n_pte_pages - 1) << 5);
352 
353 	switch (page_shift) {
354 	case 12: reg |= IOSTE_PS_4K;  break;
355 	case 16: reg |= IOSTE_PS_64K; break;
356 	case 20: reg |= IOSTE_PS_1M;  break;
357 	case 24: reg |= IOSTE_PS_16M; break;
358 	default: BUG();
359 	}
360 
361 	gap_base = gap_base >> IO_SEGMENT_SHIFT;
362 	gap_size = gap_size >> IO_SEGMENT_SHIFT;
363 
364 	pr_debug("Setting up IOMMU stab:\n");
365 	for (i = start_seg; i < (start_seg + segments); i++) {
366 		if (i >= gap_base && i < (gap_base + gap_size)) {
367 			pr_debug("\toverlap at %d, skipping\n", i);
368 			continue;
369 		}
370 		iommu->stab[i] = reg | (__pa(ptab) + (n_pte_pages << 12) *
371 					(i - start_seg));
372 		pr_debug("\t[%d] 0x%016lx\n", i, iommu->stab[i]);
373 	}
374 
375 	return ptab;
376 }
377 
378 static void __init cell_iommu_enable_hardware(struct cbe_iommu *iommu)
379 {
380 	int ret;
381 	unsigned long reg, xlate_base;
382 	unsigned int virq;
383 
384 	if (cell_iommu_find_ioc(iommu->nid, &xlate_base))
385 		panic("%s: missing IOC register mappings for node %d\n",
386 		      __func__, iommu->nid);
387 
388 	iommu->xlate_regs = ioremap(xlate_base, IOC_Reg_Size);
389 	iommu->cmd_regs = iommu->xlate_regs + IOC_IOCmd_Offset;
390 
391 	/* ensure that the STEs have updated */
392 	mb();
393 
394 	/* setup interrupts for the iommu. */
395 	reg = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat);
396 	out_be64(iommu->xlate_regs + IOC_IO_ExcpStat,
397 			reg & ~IOC_IO_ExcpStat_V);
398 	out_be64(iommu->xlate_regs + IOC_IO_ExcpMask,
399 			IOC_IO_ExcpMask_PFE | IOC_IO_ExcpMask_SFE);
400 
401 	virq = irq_create_mapping(NULL,
402 			IIC_IRQ_IOEX_ATI | (iommu->nid << IIC_IRQ_NODE_SHIFT));
403 	BUG_ON(!virq);
404 
405 	ret = request_irq(virq, ioc_interrupt, 0, iommu->name, iommu);
406 	BUG_ON(ret);
407 
408 	/* set the IOC segment table origin register (and turn on the iommu) */
409 	reg = IOC_IOST_Origin_E | __pa(iommu->stab) | IOC_IOST_Origin_HW;
410 	out_be64(iommu->xlate_regs + IOC_IOST_Origin, reg);
411 	in_be64(iommu->xlate_regs + IOC_IOST_Origin);
412 
413 	/* turn on IO translation */
414 	reg = in_be64(iommu->cmd_regs + IOC_IOCmd_Cfg) | IOC_IOCmd_Cfg_TE;
415 	out_be64(iommu->cmd_regs + IOC_IOCmd_Cfg, reg);
416 }
417 
418 static void __init cell_iommu_setup_hardware(struct cbe_iommu *iommu,
419 	unsigned long base, unsigned long size)
420 {
421 	cell_iommu_setup_stab(iommu, base, size, 0, 0);
422 	iommu->ptab = cell_iommu_alloc_ptab(iommu, base, size, 0, 0,
423 					    IOMMU_PAGE_SHIFT_4K);
424 	cell_iommu_enable_hardware(iommu);
425 }
426 
427 #if 0/* Unused for now */
428 static struct iommu_window *find_window(struct cbe_iommu *iommu,
429 		unsigned long offset, unsigned long size)
430 {
431 	struct iommu_window *window;
432 
433 	/* todo: check for overlapping (but not equal) windows) */
434 
435 	list_for_each_entry(window, &(iommu->windows), list) {
436 		if (window->offset == offset && window->size == size)
437 			return window;
438 	}
439 
440 	return NULL;
441 }
442 #endif
443 
444 static inline u32 cell_iommu_get_ioid(struct device_node *np)
445 {
446 	const u32 *ioid;
447 
448 	ioid = of_get_property(np, "ioid", NULL);
449 	if (ioid == NULL) {
450 		printk(KERN_WARNING "iommu: missing ioid for %pOF using 0\n",
451 		       np);
452 		return 0;
453 	}
454 
455 	return *ioid;
456 }
457 
458 static struct iommu_table_ops cell_iommu_ops = {
459 	.set = tce_build_cell,
460 	.clear = tce_free_cell
461 };
462 
463 static struct iommu_window * __init
464 cell_iommu_setup_window(struct cbe_iommu *iommu, struct device_node *np,
465 			unsigned long offset, unsigned long size,
466 			unsigned long pte_offset)
467 {
468 	struct iommu_window *window;
469 	struct page *page;
470 	u32 ioid;
471 
472 	ioid = cell_iommu_get_ioid(np);
473 
474 	window = kzalloc_node(sizeof(*window), GFP_KERNEL, iommu->nid);
475 	BUG_ON(window == NULL);
476 
477 	window->offset = offset;
478 	window->size = size;
479 	window->ioid = ioid;
480 	window->iommu = iommu;
481 
482 	window->table.it_blocksize = 16;
483 	window->table.it_base = (unsigned long)iommu->ptab;
484 	window->table.it_index = iommu->nid;
485 	window->table.it_page_shift = IOMMU_PAGE_SHIFT_4K;
486 	window->table.it_offset =
487 		(offset >> window->table.it_page_shift) + pte_offset;
488 	window->table.it_size = size >> window->table.it_page_shift;
489 	window->table.it_ops = &cell_iommu_ops;
490 
491 	if (!iommu_init_table(&window->table, iommu->nid, 0, 0))
492 		panic("Failed to initialize iommu table");
493 
494 	pr_debug("\tioid      %d\n", window->ioid);
495 	pr_debug("\tblocksize %ld\n", window->table.it_blocksize);
496 	pr_debug("\tbase      0x%016lx\n", window->table.it_base);
497 	pr_debug("\toffset    0x%lx\n", window->table.it_offset);
498 	pr_debug("\tsize      %ld\n", window->table.it_size);
499 
500 	list_add(&window->list, &iommu->windows);
501 
502 	if (offset != 0)
503 		return window;
504 
505 	/* We need to map and reserve the first IOMMU page since it's used
506 	 * by the spider workaround. In theory, we only need to do that when
507 	 * running on spider but it doesn't really matter.
508 	 *
509 	 * This code also assumes that we have a window that starts at 0,
510 	 * which is the case on all spider based blades.
511 	 */
512 	page = alloc_pages_node(iommu->nid, GFP_KERNEL, 0);
513 	BUG_ON(!page);
514 	iommu->pad_page = page_address(page);
515 	clear_page(iommu->pad_page);
516 
517 	__set_bit(0, window->table.it_map);
518 	tce_build_cell(&window->table, window->table.it_offset, 1,
519 		       (unsigned long)iommu->pad_page, DMA_TO_DEVICE, 0);
520 
521 	return window;
522 }
523 
524 static struct cbe_iommu *cell_iommu_for_node(int nid)
525 {
526 	int i;
527 
528 	for (i = 0; i < cbe_nr_iommus; i++)
529 		if (iommus[i].nid == nid)
530 			return &iommus[i];
531 	return NULL;
532 }
533 
534 static unsigned long cell_dma_nommu_offset;
535 
536 static unsigned long dma_iommu_fixed_base;
537 static bool cell_iommu_enabled;
538 
539 /* iommu_fixed_is_weak is set if booted with iommu_fixed=weak */
540 bool iommu_fixed_is_weak;
541 
542 static struct iommu_table *cell_get_iommu_table(struct device *dev)
543 {
544 	struct iommu_window *window;
545 	struct cbe_iommu *iommu;
546 
547 	/* Current implementation uses the first window available in that
548 	 * node's iommu. We -might- do something smarter later though it may
549 	 * never be necessary
550 	 */
551 	iommu = cell_iommu_for_node(dev_to_node(dev));
552 	if (iommu == NULL || list_empty(&iommu->windows)) {
553 		dev_err(dev, "iommu: missing iommu for %pOF (node %d)\n",
554 		       dev->of_node, dev_to_node(dev));
555 		return NULL;
556 	}
557 	window = list_entry(iommu->windows.next, struct iommu_window, list);
558 
559 	return &window->table;
560 }
561 
562 static u64 cell_iommu_get_fixed_address(struct device *dev);
563 
564 static void cell_dma_dev_setup(struct device *dev)
565 {
566 	if (cell_iommu_enabled) {
567 		u64 addr = cell_iommu_get_fixed_address(dev);
568 
569 		if (addr != OF_BAD_ADDR)
570 			dev->archdata.dma_offset = addr + dma_iommu_fixed_base;
571 		set_iommu_table_base(dev, cell_get_iommu_table(dev));
572 	} else {
573 		dev->archdata.dma_offset = cell_dma_nommu_offset;
574 	}
575 }
576 
577 static void cell_pci_dma_dev_setup(struct pci_dev *dev)
578 {
579 	cell_dma_dev_setup(&dev->dev);
580 }
581 
582 static int cell_of_bus_notify(struct notifier_block *nb, unsigned long action,
583 			      void *data)
584 {
585 	struct device *dev = data;
586 
587 	/* We are only interested in device addition */
588 	if (action != BUS_NOTIFY_ADD_DEVICE)
589 		return 0;
590 
591 	if (cell_iommu_enabled)
592 		dev->dma_ops = &dma_iommu_ops;
593 	cell_dma_dev_setup(dev);
594 	return 0;
595 }
596 
597 static struct notifier_block cell_of_bus_notifier = {
598 	.notifier_call = cell_of_bus_notify
599 };
600 
601 static int __init cell_iommu_get_window(struct device_node *np,
602 					 unsigned long *base,
603 					 unsigned long *size)
604 {
605 	const __be32 *dma_window;
606 	unsigned long index;
607 
608 	/* Use ibm,dma-window if available, else, hard code ! */
609 	dma_window = of_get_property(np, "ibm,dma-window", NULL);
610 	if (dma_window == NULL) {
611 		*base = 0;
612 		*size = 0x80000000u;
613 		return -ENODEV;
614 	}
615 
616 	of_parse_dma_window(np, dma_window, &index, base, size);
617 	return 0;
618 }
619 
620 static struct cbe_iommu * __init cell_iommu_alloc(struct device_node *np)
621 {
622 	struct cbe_iommu *iommu;
623 	int nid, i;
624 
625 	/* Get node ID */
626 	nid = of_node_to_nid(np);
627 	if (nid < 0) {
628 		printk(KERN_ERR "iommu: failed to get node for %pOF\n",
629 		       np);
630 		return NULL;
631 	}
632 	pr_debug("iommu: setting up iommu for node %d (%pOF)\n",
633 		 nid, np);
634 
635 	/* XXX todo: If we can have multiple windows on the same IOMMU, which
636 	 * isn't the case today, we probably want here to check whether the
637 	 * iommu for that node is already setup.
638 	 * However, there might be issue with getting the size right so let's
639 	 * ignore that for now. We might want to completely get rid of the
640 	 * multiple window support since the cell iommu supports per-page ioids
641 	 */
642 
643 	if (cbe_nr_iommus >= NR_IOMMUS) {
644 		printk(KERN_ERR "iommu: too many IOMMUs detected ! (%pOF)\n",
645 		       np);
646 		return NULL;
647 	}
648 
649 	/* Init base fields */
650 	i = cbe_nr_iommus++;
651 	iommu = &iommus[i];
652 	iommu->stab = NULL;
653 	iommu->nid = nid;
654 	snprintf(iommu->name, sizeof(iommu->name), "iommu%d", i);
655 	INIT_LIST_HEAD(&iommu->windows);
656 
657 	return iommu;
658 }
659 
660 static void __init cell_iommu_init_one(struct device_node *np,
661 				       unsigned long offset)
662 {
663 	struct cbe_iommu *iommu;
664 	unsigned long base, size;
665 
666 	iommu = cell_iommu_alloc(np);
667 	if (!iommu)
668 		return;
669 
670 	/* Obtain a window for it */
671 	cell_iommu_get_window(np, &base, &size);
672 
673 	pr_debug("\ttranslating window 0x%lx...0x%lx\n",
674 		 base, base + size - 1);
675 
676 	/* Initialize the hardware */
677 	cell_iommu_setup_hardware(iommu, base, size);
678 
679 	/* Setup the iommu_table */
680 	cell_iommu_setup_window(iommu, np, base, size,
681 				offset >> IOMMU_PAGE_SHIFT_4K);
682 }
683 
684 static void __init cell_disable_iommus(void)
685 {
686 	int node;
687 	unsigned long base, val;
688 	void __iomem *xregs, *cregs;
689 
690 	/* Make sure IOC translation is disabled on all nodes */
691 	for_each_online_node(node) {
692 		if (cell_iommu_find_ioc(node, &base))
693 			continue;
694 		xregs = ioremap(base, IOC_Reg_Size);
695 		if (xregs == NULL)
696 			continue;
697 		cregs = xregs + IOC_IOCmd_Offset;
698 
699 		pr_debug("iommu: cleaning up iommu on node %d\n", node);
700 
701 		out_be64(xregs + IOC_IOST_Origin, 0);
702 		(void)in_be64(xregs + IOC_IOST_Origin);
703 		val = in_be64(cregs + IOC_IOCmd_Cfg);
704 		val &= ~IOC_IOCmd_Cfg_TE;
705 		out_be64(cregs + IOC_IOCmd_Cfg, val);
706 		(void)in_be64(cregs + IOC_IOCmd_Cfg);
707 
708 		iounmap(xregs);
709 	}
710 }
711 
712 static int __init cell_iommu_init_disabled(void)
713 {
714 	struct device_node *np = NULL;
715 	unsigned long base = 0, size;
716 
717 	/* When no iommu is present, we use direct DMA ops */
718 
719 	/* First make sure all IOC translation is turned off */
720 	cell_disable_iommus();
721 
722 	/* If we have no Axon, we set up the spider DMA magic offset */
723 	np = of_find_node_by_name(NULL, "axon");
724 	if (!np)
725 		cell_dma_nommu_offset = SPIDER_DMA_OFFSET;
726 	of_node_put(np);
727 
728 	/* Now we need to check to see where the memory is mapped
729 	 * in PCI space. We assume that all busses use the same dma
730 	 * window which is always the case so far on Cell, thus we
731 	 * pick up the first pci-internal node we can find and check
732 	 * the DMA window from there.
733 	 */
734 	for_each_node_by_name(np, "axon") {
735 		if (np->parent == NULL || np->parent->parent != NULL)
736 			continue;
737 		if (cell_iommu_get_window(np, &base, &size) == 0)
738 			break;
739 	}
740 	if (np == NULL) {
741 		for_each_node_by_name(np, "pci-internal") {
742 			if (np->parent == NULL || np->parent->parent != NULL)
743 				continue;
744 			if (cell_iommu_get_window(np, &base, &size) == 0)
745 				break;
746 		}
747 	}
748 	of_node_put(np);
749 
750 	/* If we found a DMA window, we check if it's big enough to enclose
751 	 * all of physical memory. If not, we force enable IOMMU
752 	 */
753 	if (np && size < memblock_end_of_DRAM()) {
754 		printk(KERN_WARNING "iommu: force-enabled, dma window"
755 		       " (%ldMB) smaller than total memory (%lldMB)\n",
756 		       size >> 20, memblock_end_of_DRAM() >> 20);
757 		return -ENODEV;
758 	}
759 
760 	cell_dma_nommu_offset += base;
761 
762 	if (cell_dma_nommu_offset != 0)
763 		cell_pci_controller_ops.dma_dev_setup = cell_pci_dma_dev_setup;
764 
765 	printk("iommu: disabled, direct DMA offset is 0x%lx\n",
766 	       cell_dma_nommu_offset);
767 
768 	return 0;
769 }
770 
771 /*
772  *  Fixed IOMMU mapping support
773  *
774  *  This code adds support for setting up a fixed IOMMU mapping on certain
775  *  cell machines. For 64-bit devices this avoids the performance overhead of
776  *  mapping and unmapping pages at runtime. 32-bit devices are unable to use
777  *  the fixed mapping.
778  *
779  *  The fixed mapping is established at boot, and maps all of physical memory
780  *  1:1 into device space at some offset. On machines with < 30 GB of memory
781  *  we setup the fixed mapping immediately above the normal IOMMU window.
782  *
783  *  For example a machine with 4GB of memory would end up with the normal
784  *  IOMMU window from 0-2GB and the fixed mapping window from 2GB to 6GB. In
785  *  this case a 64-bit device wishing to DMA to 1GB would be told to DMA to
786  *  3GB, plus any offset required by firmware. The firmware offset is encoded
787  *  in the "dma-ranges" property.
788  *
789  *  On machines with 30GB or more of memory, we are unable to place the fixed
790  *  mapping above the normal IOMMU window as we would run out of address space.
791  *  Instead we move the normal IOMMU window to coincide with the hash page
792  *  table, this region does not need to be part of the fixed mapping as no
793  *  device should ever be DMA'ing to it. We then setup the fixed mapping
794  *  from 0 to 32GB.
795  */
796 
797 static u64 cell_iommu_get_fixed_address(struct device *dev)
798 {
799 	u64 cpu_addr, size, best_size, dev_addr = OF_BAD_ADDR;
800 	struct device_node *np;
801 	const u32 *ranges = NULL;
802 	int i, len, best, naddr, nsize, pna, range_size;
803 
804 	/* We can be called for platform devices that have no of_node */
805 	np = of_node_get(dev->of_node);
806 	if (!np)
807 		goto out;
808 
809 	while (1) {
810 		naddr = of_n_addr_cells(np);
811 		nsize = of_n_size_cells(np);
812 		np = of_get_next_parent(np);
813 		if (!np)
814 			break;
815 
816 		ranges = of_get_property(np, "dma-ranges", &len);
817 
818 		/* Ignore empty ranges, they imply no translation required */
819 		if (ranges && len > 0)
820 			break;
821 	}
822 
823 	if (!ranges) {
824 		dev_dbg(dev, "iommu: no dma-ranges found\n");
825 		goto out;
826 	}
827 
828 	len /= sizeof(u32);
829 
830 	pna = of_n_addr_cells(np);
831 	range_size = naddr + nsize + pna;
832 
833 	/* dma-ranges format:
834 	 * child addr	: naddr cells
835 	 * parent addr	: pna cells
836 	 * size		: nsize cells
837 	 */
838 	for (i = 0, best = -1, best_size = 0; i < len; i += range_size) {
839 		cpu_addr = of_translate_dma_address(np, ranges + i + naddr);
840 		size = of_read_number(ranges + i + naddr + pna, nsize);
841 
842 		if (cpu_addr == 0 && size > best_size) {
843 			best = i;
844 			best_size = size;
845 		}
846 	}
847 
848 	if (best >= 0) {
849 		dev_addr = of_read_number(ranges + best, naddr);
850 	} else
851 		dev_dbg(dev, "iommu: no suitable range found!\n");
852 
853 out:
854 	of_node_put(np);
855 
856 	return dev_addr;
857 }
858 
859 static bool cell_pci_iommu_bypass_supported(struct pci_dev *pdev, u64 mask)
860 {
861 	return mask == DMA_BIT_MASK(64) &&
862 		cell_iommu_get_fixed_address(&pdev->dev) != OF_BAD_ADDR;
863 }
864 
865 static void __init insert_16M_pte(unsigned long addr, unsigned long *ptab,
866 			   unsigned long base_pte)
867 {
868 	unsigned long segment, offset;
869 
870 	segment = addr >> IO_SEGMENT_SHIFT;
871 	offset = (addr >> 24) - (segment << IO_PAGENO_BITS(24));
872 	ptab = ptab + (segment * (1 << 12) / sizeof(unsigned long));
873 
874 	pr_debug("iommu: addr %lx ptab %p segment %lx offset %lx\n",
875 		  addr, ptab, segment, offset);
876 
877 	ptab[offset] = base_pte | (__pa(addr) & CBE_IOPTE_RPN_Mask);
878 }
879 
880 static void __init cell_iommu_setup_fixed_ptab(struct cbe_iommu *iommu,
881 	struct device_node *np, unsigned long dbase, unsigned long dsize,
882 	unsigned long fbase, unsigned long fsize)
883 {
884 	unsigned long base_pte, uaddr, ioaddr, *ptab;
885 
886 	ptab = cell_iommu_alloc_ptab(iommu, fbase, fsize, dbase, dsize, 24);
887 
888 	dma_iommu_fixed_base = fbase;
889 
890 	pr_debug("iommu: mapping 0x%lx pages from 0x%lx\n", fsize, fbase);
891 
892 	base_pte = CBE_IOPTE_PP_W | CBE_IOPTE_PP_R | CBE_IOPTE_M |
893 		(cell_iommu_get_ioid(np) & CBE_IOPTE_IOID_Mask);
894 
895 	if (iommu_fixed_is_weak)
896 		pr_info("IOMMU: Using weak ordering for fixed mapping\n");
897 	else {
898 		pr_info("IOMMU: Using strong ordering for fixed mapping\n");
899 		base_pte |= CBE_IOPTE_SO_RW;
900 	}
901 
902 	for (uaddr = 0; uaddr < fsize; uaddr += (1 << 24)) {
903 		/* Don't touch the dynamic region */
904 		ioaddr = uaddr + fbase;
905 		if (ioaddr >= dbase && ioaddr < (dbase + dsize)) {
906 			pr_debug("iommu: fixed/dynamic overlap, skipping\n");
907 			continue;
908 		}
909 
910 		insert_16M_pte(uaddr, ptab, base_pte);
911 	}
912 
913 	mb();
914 }
915 
916 static int __init cell_iommu_fixed_mapping_init(void)
917 {
918 	unsigned long dbase, dsize, fbase, fsize, hbase, hend;
919 	struct cbe_iommu *iommu;
920 	struct device_node *np;
921 
922 	/* The fixed mapping is only supported on axon machines */
923 	np = of_find_node_by_name(NULL, "axon");
924 	of_node_put(np);
925 
926 	if (!np) {
927 		pr_debug("iommu: fixed mapping disabled, no axons found\n");
928 		return -1;
929 	}
930 
931 	/* We must have dma-ranges properties for fixed mapping to work */
932 	np = of_find_node_with_property(NULL, "dma-ranges");
933 	of_node_put(np);
934 
935 	if (!np) {
936 		pr_debug("iommu: no dma-ranges found, no fixed mapping\n");
937 		return -1;
938 	}
939 
940 	/* The default setup is to have the fixed mapping sit after the
941 	 * dynamic region, so find the top of the largest IOMMU window
942 	 * on any axon, then add the size of RAM and that's our max value.
943 	 * If that is > 32GB we have to do other shennanigans.
944 	 */
945 	fbase = 0;
946 	for_each_node_by_name(np, "axon") {
947 		cell_iommu_get_window(np, &dbase, &dsize);
948 		fbase = max(fbase, dbase + dsize);
949 	}
950 
951 	fbase = ALIGN(fbase, 1 << IO_SEGMENT_SHIFT);
952 	fsize = memblock_phys_mem_size();
953 
954 	if ((fbase + fsize) <= 0x800000000ul)
955 		hbase = 0; /* use the device tree window */
956 	else {
957 		/* If we're over 32 GB we need to cheat. We can't map all of
958 		 * RAM with the fixed mapping, and also fit the dynamic
959 		 * region. So try to place the dynamic region where the hash
960 		 * table sits, drivers never need to DMA to it, we don't
961 		 * need a fixed mapping for that area.
962 		 */
963 		if (!htab_address) {
964 			pr_debug("iommu: htab is NULL, on LPAR? Huh?\n");
965 			return -1;
966 		}
967 		hbase = __pa(htab_address);
968 		hend  = hbase + htab_size_bytes;
969 
970 		/* The window must start and end on a segment boundary */
971 		if ((hbase != ALIGN(hbase, 1 << IO_SEGMENT_SHIFT)) ||
972 		    (hend != ALIGN(hend, 1 << IO_SEGMENT_SHIFT))) {
973 			pr_debug("iommu: hash window not segment aligned\n");
974 			return -1;
975 		}
976 
977 		/* Check the hash window fits inside the real DMA window */
978 		for_each_node_by_name(np, "axon") {
979 			cell_iommu_get_window(np, &dbase, &dsize);
980 
981 			if (hbase < dbase || (hend > (dbase + dsize))) {
982 				pr_debug("iommu: hash window doesn't fit in"
983 					 "real DMA window\n");
984 				of_node_put(np);
985 				return -1;
986 			}
987 		}
988 
989 		fbase = 0;
990 	}
991 
992 	/* Setup the dynamic regions */
993 	for_each_node_by_name(np, "axon") {
994 		iommu = cell_iommu_alloc(np);
995 		BUG_ON(!iommu);
996 
997 		if (hbase == 0)
998 			cell_iommu_get_window(np, &dbase, &dsize);
999 		else {
1000 			dbase = hbase;
1001 			dsize = htab_size_bytes;
1002 		}
1003 
1004 		printk(KERN_DEBUG "iommu: node %d, dynamic window 0x%lx-0x%lx "
1005 			"fixed window 0x%lx-0x%lx\n", iommu->nid, dbase,
1006 			 dbase + dsize, fbase, fbase + fsize);
1007 
1008 		cell_iommu_setup_stab(iommu, dbase, dsize, fbase, fsize);
1009 		iommu->ptab = cell_iommu_alloc_ptab(iommu, dbase, dsize, 0, 0,
1010 						    IOMMU_PAGE_SHIFT_4K);
1011 		cell_iommu_setup_fixed_ptab(iommu, np, dbase, dsize,
1012 					     fbase, fsize);
1013 		cell_iommu_enable_hardware(iommu);
1014 		cell_iommu_setup_window(iommu, np, dbase, dsize, 0);
1015 	}
1016 
1017 	cell_pci_controller_ops.iommu_bypass_supported =
1018 		cell_pci_iommu_bypass_supported;
1019 	return 0;
1020 }
1021 
1022 static int iommu_fixed_disabled;
1023 
1024 static int __init setup_iommu_fixed(char *str)
1025 {
1026 	struct device_node *pciep;
1027 
1028 	if (strcmp(str, "off") == 0)
1029 		iommu_fixed_disabled = 1;
1030 
1031 	/* If we can find a pcie-endpoint in the device tree assume that
1032 	 * we're on a triblade or a CAB so by default the fixed mapping
1033 	 * should be set to be weakly ordered; but only if the boot
1034 	 * option WASN'T set for strong ordering
1035 	 */
1036 	pciep = of_find_node_by_type(NULL, "pcie-endpoint");
1037 
1038 	if (strcmp(str, "weak") == 0 || (pciep && strcmp(str, "strong") != 0))
1039 		iommu_fixed_is_weak = true;
1040 
1041 	of_node_put(pciep);
1042 
1043 	return 1;
1044 }
1045 __setup("iommu_fixed=", setup_iommu_fixed);
1046 
1047 static int __init cell_iommu_init(void)
1048 {
1049 	struct device_node *np;
1050 
1051 	/* If IOMMU is disabled or we have little enough RAM to not need
1052 	 * to enable it, we setup a direct mapping.
1053 	 *
1054 	 * Note: should we make sure we have the IOMMU actually disabled ?
1055 	 */
1056 	if (iommu_is_off ||
1057 	    (!iommu_force_on && memblock_end_of_DRAM() <= 0x80000000ull))
1058 		if (cell_iommu_init_disabled() == 0)
1059 			goto bail;
1060 
1061 	/* Setup various callbacks */
1062 	cell_pci_controller_ops.dma_dev_setup = cell_pci_dma_dev_setup;
1063 
1064 	if (!iommu_fixed_disabled && cell_iommu_fixed_mapping_init() == 0)
1065 		goto done;
1066 
1067 	/* Create an iommu for each /axon node.  */
1068 	for_each_node_by_name(np, "axon") {
1069 		if (np->parent == NULL || np->parent->parent != NULL)
1070 			continue;
1071 		cell_iommu_init_one(np, 0);
1072 	}
1073 
1074 	/* Create an iommu for each toplevel /pci-internal node for
1075 	 * old hardware/firmware
1076 	 */
1077 	for_each_node_by_name(np, "pci-internal") {
1078 		if (np->parent == NULL || np->parent->parent != NULL)
1079 			continue;
1080 		cell_iommu_init_one(np, SPIDER_DMA_OFFSET);
1081 	}
1082  done:
1083 	/* Setup default PCI iommu ops */
1084 	set_pci_dma_ops(&dma_iommu_ops);
1085 	cell_iommu_enabled = true;
1086  bail:
1087 	/* Register callbacks on OF platform device addition/removal
1088 	 * to handle linking them to the right DMA operations
1089 	 */
1090 	bus_register_notifier(&platform_bus_type, &cell_of_bus_notifier);
1091 
1092 	return 0;
1093 }
1094 machine_arch_initcall(cell, cell_iommu_init);
1095