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#include <linux/slab.h>
32#include <linux/memblock.h>
33
34#include <asm/prom.h>
35#include <asm/iommu.h>
36#include <asm/machdep.h>
37#include <asm/pci-bridge.h>
38#include <asm/udbg.h>
39#include <asm/firmware.h>
40#include <asm/cell-regs.h>
41
42#include "cell.h"
43#include "interrupt.h"
44
45/* Define CELL_IOMMU_REAL_UNMAP to actually unmap non-used pages
46 * instead of leaving them mapped to some dummy page. This can be
47 * enabled once the appropriate workarounds for spider bugs have
48 * been enabled
49 */
50#define CELL_IOMMU_REAL_UNMAP
51
52/* Define CELL_IOMMU_STRICT_PROTECTION to enforce protection of
53 * IO PTEs based on the transfer direction. That can be enabled
54 * once spider-net has been fixed to pass the correct direction
55 * to the DMA mapping functions
56 */
57#define CELL_IOMMU_STRICT_PROTECTION
58
59
60#define NR_IOMMUS			2
61
62/* IOC mmap registers */
63#define IOC_Reg_Size			0x2000
64
65#define IOC_IOPT_CacheInvd		0x908
66#define IOC_IOPT_CacheInvd_NE_Mask	0xffe0000000000000ul
67#define IOC_IOPT_CacheInvd_IOPTE_Mask	0x000003fffffffff8ul
68#define IOC_IOPT_CacheInvd_Busy		0x0000000000000001ul
69
70#define IOC_IOST_Origin			0x918
71#define IOC_IOST_Origin_E		0x8000000000000000ul
72#define IOC_IOST_Origin_HW		0x0000000000000800ul
73#define IOC_IOST_Origin_HL		0x0000000000000400ul
74
75#define IOC_IO_ExcpStat			0x920
76#define IOC_IO_ExcpStat_V		0x8000000000000000ul
77#define IOC_IO_ExcpStat_SPF_Mask	0x6000000000000000ul
78#define IOC_IO_ExcpStat_SPF_S		0x6000000000000000ul
79#define IOC_IO_ExcpStat_SPF_P		0x2000000000000000ul
80#define IOC_IO_ExcpStat_ADDR_Mask	0x00000007fffff000ul
81#define IOC_IO_ExcpStat_RW_Mask		0x0000000000000800ul
82#define IOC_IO_ExcpStat_IOID_Mask	0x00000000000007fful
83
84#define IOC_IO_ExcpMask			0x928
85#define IOC_IO_ExcpMask_SFE		0x4000000000000000ul
86#define IOC_IO_ExcpMask_PFE		0x2000000000000000ul
87
88#define IOC_IOCmd_Offset		0x1000
89
90#define IOC_IOCmd_Cfg			0xc00
91#define IOC_IOCmd_Cfg_TE		0x0000800000000000ul
92
93
94/* Segment table entries */
95#define IOSTE_V			0x8000000000000000ul /* valid */
96#define IOSTE_H			0x4000000000000000ul /* cache hint */
97#define IOSTE_PT_Base_RPN_Mask  0x3ffffffffffff000ul /* base RPN of IOPT */
98#define IOSTE_NPPT_Mask		0x0000000000000fe0ul /* no. pages in IOPT */
99#define IOSTE_PS_Mask		0x0000000000000007ul /* page size */
100#define IOSTE_PS_4K		0x0000000000000001ul /*   - 4kB  */
101#define IOSTE_PS_64K		0x0000000000000003ul /*   - 64kB */
102#define IOSTE_PS_1M		0x0000000000000005ul /*   - 1MB  */
103#define IOSTE_PS_16M		0x0000000000000007ul /*   - 16MB */
104
105
106/* IOMMU sizing */
107#define IO_SEGMENT_SHIFT	28
108#define IO_PAGENO_BITS(shift)	(IO_SEGMENT_SHIFT - (shift))
109
110/* The high bit needs to be set on every DMA address */
111#define SPIDER_DMA_OFFSET	0x80000000ul
112
113struct iommu_window {
114	struct list_head list;
115	struct cbe_iommu *iommu;
116	unsigned long offset;
117	unsigned long size;
118	unsigned int ioid;
119	struct iommu_table table;
120};
121
122#define NAMESIZE 8
123struct cbe_iommu {
124	int nid;
125	char name[NAMESIZE];
126	void __iomem *xlate_regs;
127	void __iomem *cmd_regs;
128	unsigned long *stab;
129	unsigned long *ptab;
130	void *pad_page;
131	struct list_head windows;
132};
133
134/* Static array of iommus, one per node
135 *   each contains a list of windows, keyed from dma_window property
136 *   - on bus setup, look for a matching window, or create one
137 *   - on dev setup, assign iommu_table ptr
138 */
139static struct cbe_iommu iommus[NR_IOMMUS];
140static int cbe_nr_iommus;
141
142static void invalidate_tce_cache(struct cbe_iommu *iommu, unsigned long *pte,
143		long n_ptes)
144{
145	u64 __iomem *reg;
146	u64 val;
147	long n;
148
149	reg = iommu->xlate_regs + IOC_IOPT_CacheInvd;
150
151	while (n_ptes > 0) {
152		/* we can invalidate up to 1 << 11 PTEs at once */
153		n = min(n_ptes, 1l << 11);
154		val = (((n /*- 1*/) << 53) & IOC_IOPT_CacheInvd_NE_Mask)
155			| (__pa(pte) & IOC_IOPT_CacheInvd_IOPTE_Mask)
156		        | IOC_IOPT_CacheInvd_Busy;
157
158		out_be64(reg, val);
159		while (in_be64(reg) & IOC_IOPT_CacheInvd_Busy)
160			;
161
162		n_ptes -= n;
163		pte += n;
164	}
165}
166
167static int tce_build_cell(struct iommu_table *tbl, long index, long npages,
168		unsigned long uaddr, enum dma_data_direction direction,
169		struct dma_attrs *attrs)
170{
171	int i;
172	unsigned long *io_pte, base_pte;
173	struct iommu_window *window =
174		container_of(tbl, struct iommu_window, table);
175
176	/* implementing proper protection causes problems with the spidernet
177	 * driver - check mapping directions later, but allow read & write by
178	 * default for now.*/
179#ifdef CELL_IOMMU_STRICT_PROTECTION
180	/* to avoid referencing a global, we use a trick here to setup the
181	 * protection bit. "prot" is setup to be 3 fields of 4 bits apprended
182	 * together for each of the 3 supported direction values. It is then
183	 * shifted left so that the fields matching the desired direction
184	 * lands on the appropriate bits, and other bits are masked out.
185	 */
186	const unsigned long prot = 0xc48;
187	base_pte =
188		((prot << (52 + 4 * direction)) &
189		 (CBE_IOPTE_PP_W | CBE_IOPTE_PP_R)) |
190		CBE_IOPTE_M | CBE_IOPTE_SO_RW |
191		(window->ioid & CBE_IOPTE_IOID_Mask);
192#else
193	base_pte = CBE_IOPTE_PP_W | CBE_IOPTE_PP_R | CBE_IOPTE_M |
194		CBE_IOPTE_SO_RW | (window->ioid & CBE_IOPTE_IOID_Mask);
195#endif
196	if (unlikely(dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs)))
197		base_pte &= ~CBE_IOPTE_SO_RW;
198
199	io_pte = (unsigned long *)tbl->it_base + (index - tbl->it_offset);
200
201	for (i = 0; i < npages; i++, uaddr += (1 << tbl->it_page_shift))
202		io_pte[i] = base_pte | (__pa(uaddr) & CBE_IOPTE_RPN_Mask);
203
204	mb();
205
206	invalidate_tce_cache(window->iommu, io_pte, npages);
207
208	pr_debug("tce_build_cell(index=%lx,n=%lx,dir=%d,base_pte=%lx)\n",
209		 index, npages, direction, base_pte);
210	return 0;
211}
212
213static void tce_free_cell(struct iommu_table *tbl, long index, long npages)
214{
215
216	int i;
217	unsigned long *io_pte, pte;
218	struct iommu_window *window =
219		container_of(tbl, struct iommu_window, table);
220
221	pr_debug("tce_free_cell(index=%lx,n=%lx)\n", index, npages);
222
223#ifdef CELL_IOMMU_REAL_UNMAP
224	pte = 0;
225#else
226	/* spider bridge does PCI reads after freeing - insert a mapping
227	 * to a scratch page instead of an invalid entry */
228	pte = CBE_IOPTE_PP_R | CBE_IOPTE_M | CBE_IOPTE_SO_RW |
229		__pa(window->iommu->pad_page) |
230		(window->ioid & CBE_IOPTE_IOID_Mask);
231#endif
232
233	io_pte = (unsigned long *)tbl->it_base + (index - tbl->it_offset);
234
235	for (i = 0; i < npages; i++)
236		io_pte[i] = pte;
237
238	mb();
239
240	invalidate_tce_cache(window->iommu, io_pte, npages);
241}
242
243static irqreturn_t ioc_interrupt(int irq, void *data)
244{
245	unsigned long stat, spf;
246	struct cbe_iommu *iommu = data;
247
248	stat = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat);
249	spf = stat & IOC_IO_ExcpStat_SPF_Mask;
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	       (spf == IOC_IO_ExcpStat_SPF_S) ? 'S' : ' ',
256	       (spf == 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
269static 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
309static 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			__func__, 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
329static 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", __func__,
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			__func__, 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
389static 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		      __func__, 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, 0, iommu->name, iommu);
417	BUG_ON(ret);
418
419	/* set the IOC segment table origin register (and turn on the iommu) */
420	reg = IOC_IOST_Origin_E | __pa(iommu->stab) | IOC_IOST_Origin_HW;
421	out_be64(iommu->xlate_regs + IOC_IOST_Origin, reg);
422	in_be64(iommu->xlate_regs + IOC_IOST_Origin);
423
424	/* turn on IO translation */
425	reg = in_be64(iommu->cmd_regs + IOC_IOCmd_Cfg) | IOC_IOCmd_Cfg_TE;
426	out_be64(iommu->cmd_regs + IOC_IOCmd_Cfg, reg);
427}
428
429static void cell_iommu_setup_hardware(struct cbe_iommu *iommu,
430	unsigned long base, unsigned long size)
431{
432	cell_iommu_setup_stab(iommu, base, size, 0, 0);
433	iommu->ptab = cell_iommu_alloc_ptab(iommu, base, size, 0, 0,
434					    IOMMU_PAGE_SHIFT_4K);
435	cell_iommu_enable_hardware(iommu);
436}
437
438#if 0/* Unused for now */
439static struct iommu_window *find_window(struct cbe_iommu *iommu,
440		unsigned long offset, unsigned long size)
441{
442	struct iommu_window *window;
443
444	/* todo: check for overlapping (but not equal) windows) */
445
446	list_for_each_entry(window, &(iommu->windows), list) {
447		if (window->offset == offset && window->size == size)
448			return window;
449	}
450
451	return NULL;
452}
453#endif
454
455static inline u32 cell_iommu_get_ioid(struct device_node *np)
456{
457	const u32 *ioid;
458
459	ioid = of_get_property(np, "ioid", NULL);
460	if (ioid == NULL) {
461		printk(KERN_WARNING "iommu: missing ioid for %s using 0\n",
462		       np->full_name);
463		return 0;
464	}
465
466	return *ioid;
467}
468
469static struct iommu_table_ops cell_iommu_ops = {
470	.set = tce_build_cell,
471	.clear = tce_free_cell
472};
473
474static struct iommu_window * __init
475cell_iommu_setup_window(struct cbe_iommu *iommu, struct device_node *np,
476			unsigned long offset, unsigned long size,
477			unsigned long pte_offset)
478{
479	struct iommu_window *window;
480	struct page *page;
481	u32 ioid;
482
483	ioid = cell_iommu_get_ioid(np);
484
485	window = kzalloc_node(sizeof(*window), GFP_KERNEL, iommu->nid);
486	BUG_ON(window == NULL);
487
488	window->offset = offset;
489	window->size = size;
490	window->ioid = ioid;
491	window->iommu = iommu;
492
493	window->table.it_blocksize = 16;
494	window->table.it_base = (unsigned long)iommu->ptab;
495	window->table.it_index = iommu->nid;
496	window->table.it_page_shift = IOMMU_PAGE_SHIFT_4K;
497	window->table.it_offset =
498		(offset >> window->table.it_page_shift) + pte_offset;
499	window->table.it_size = size >> window->table.it_page_shift;
500	window->table.it_ops = &cell_iommu_ops;
501
502	iommu_init_table(&window->table, iommu->nid);
503
504	pr_debug("\tioid      %d\n", window->ioid);
505	pr_debug("\tblocksize %ld\n", window->table.it_blocksize);
506	pr_debug("\tbase      0x%016lx\n", window->table.it_base);
507	pr_debug("\toffset    0x%lx\n", window->table.it_offset);
508	pr_debug("\tsize      %ld\n", window->table.it_size);
509
510	list_add(&window->list, &iommu->windows);
511
512	if (offset != 0)
513		return window;
514
515	/* We need to map and reserve the first IOMMU page since it's used
516	 * by the spider workaround. In theory, we only need to do that when
517	 * running on spider but it doesn't really matter.
518	 *
519	 * This code also assumes that we have a window that starts at 0,
520	 * which is the case on all spider based blades.
521	 */
522	page = alloc_pages_node(iommu->nid, GFP_KERNEL, 0);
523	BUG_ON(!page);
524	iommu->pad_page = page_address(page);
525	clear_page(iommu->pad_page);
526
527	__set_bit(0, window->table.it_map);
528	tce_build_cell(&window->table, window->table.it_offset, 1,
529		       (unsigned long)iommu->pad_page, DMA_TO_DEVICE, NULL);
530
531	return window;
532}
533
534static struct cbe_iommu *cell_iommu_for_node(int nid)
535{
536	int i;
537
538	for (i = 0; i < cbe_nr_iommus; i++)
539		if (iommus[i].nid == nid)
540			return &iommus[i];
541	return NULL;
542}
543
544static unsigned long cell_dma_direct_offset;
545
546static unsigned long dma_iommu_fixed_base;
547
548/* iommu_fixed_is_weak is set if booted with iommu_fixed=weak */
549static int iommu_fixed_is_weak;
550
551static struct iommu_table *cell_get_iommu_table(struct device *dev)
552{
553	struct iommu_window *window;
554	struct cbe_iommu *iommu;
555
556	/* Current implementation uses the first window available in that
557	 * node's iommu. We -might- do something smarter later though it may
558	 * never be necessary
559	 */
560	iommu = cell_iommu_for_node(dev_to_node(dev));
561	if (iommu == NULL || list_empty(&iommu->windows)) {
562		dev_err(dev, "iommu: missing iommu for %s (node %d)\n",
563		       of_node_full_name(dev->of_node), dev_to_node(dev));
564		return NULL;
565	}
566	window = list_entry(iommu->windows.next, struct iommu_window, list);
567
568	return &window->table;
569}
570
571/* A coherent allocation implies strong ordering */
572
573static void *dma_fixed_alloc_coherent(struct device *dev, size_t size,
574				      dma_addr_t *dma_handle, gfp_t flag,
575				      struct dma_attrs *attrs)
576{
577	if (iommu_fixed_is_weak)
578		return iommu_alloc_coherent(dev, cell_get_iommu_table(dev),
579					    size, dma_handle,
580					    device_to_mask(dev), flag,
581					    dev_to_node(dev));
582	else
583		return dma_direct_ops.alloc(dev, size, dma_handle, flag,
584					    attrs);
585}
586
587static void dma_fixed_free_coherent(struct device *dev, size_t size,
588				    void *vaddr, dma_addr_t dma_handle,
589				    struct dma_attrs *attrs)
590{
591	if (iommu_fixed_is_weak)
592		iommu_free_coherent(cell_get_iommu_table(dev), size, vaddr,
593				    dma_handle);
594	else
595		dma_direct_ops.free(dev, size, vaddr, dma_handle, attrs);
596}
597
598static dma_addr_t dma_fixed_map_page(struct device *dev, struct page *page,
599				     unsigned long offset, size_t size,
600				     enum dma_data_direction direction,
601				     struct dma_attrs *attrs)
602{
603	if (iommu_fixed_is_weak == dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs))
604		return dma_direct_ops.map_page(dev, page, offset, size,
605					       direction, attrs);
606	else
607		return iommu_map_page(dev, cell_get_iommu_table(dev), page,
608				      offset, size, device_to_mask(dev),
609				      direction, attrs);
610}
611
612static void dma_fixed_unmap_page(struct device *dev, dma_addr_t dma_addr,
613				 size_t size, enum dma_data_direction direction,
614				 struct dma_attrs *attrs)
615{
616	if (iommu_fixed_is_weak == dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs))
617		dma_direct_ops.unmap_page(dev, dma_addr, size, direction,
618					  attrs);
619	else
620		iommu_unmap_page(cell_get_iommu_table(dev), dma_addr, size,
621				 direction, attrs);
622}
623
624static int dma_fixed_map_sg(struct device *dev, struct scatterlist *sg,
625			   int nents, enum dma_data_direction direction,
626			   struct dma_attrs *attrs)
627{
628	if (iommu_fixed_is_weak == dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs))
629		return dma_direct_ops.map_sg(dev, sg, nents, direction, attrs);
630	else
631		return ppc_iommu_map_sg(dev, cell_get_iommu_table(dev), sg,
632					nents, device_to_mask(dev),
633					direction, attrs);
634}
635
636static void dma_fixed_unmap_sg(struct device *dev, struct scatterlist *sg,
637			       int nents, enum dma_data_direction direction,
638			       struct dma_attrs *attrs)
639{
640	if (iommu_fixed_is_weak == dma_get_attr(DMA_ATTR_WEAK_ORDERING, attrs))
641		dma_direct_ops.unmap_sg(dev, sg, nents, direction, attrs);
642	else
643		ppc_iommu_unmap_sg(cell_get_iommu_table(dev), sg, nents,
644				   direction, attrs);
645}
646
647static int dma_fixed_dma_supported(struct device *dev, u64 mask)
648{
649	return mask == DMA_BIT_MASK(64);
650}
651
652static int dma_set_mask_and_switch(struct device *dev, u64 dma_mask);
653
654struct dma_map_ops dma_iommu_fixed_ops = {
655	.alloc          = dma_fixed_alloc_coherent,
656	.free           = dma_fixed_free_coherent,
657	.map_sg         = dma_fixed_map_sg,
658	.unmap_sg       = dma_fixed_unmap_sg,
659	.dma_supported  = dma_fixed_dma_supported,
660	.set_dma_mask   = dma_set_mask_and_switch,
661	.map_page       = dma_fixed_map_page,
662	.unmap_page     = dma_fixed_unmap_page,
663};
664
665static void cell_dma_dev_setup_fixed(struct device *dev);
666
667static void cell_dma_dev_setup(struct device *dev)
668{
669	/* Order is important here, these are not mutually exclusive */
670	if (get_dma_ops(dev) == &dma_iommu_fixed_ops)
671		cell_dma_dev_setup_fixed(dev);
672	else if (get_pci_dma_ops() == &dma_iommu_ops)
673		set_iommu_table_base(dev, cell_get_iommu_table(dev));
674	else if (get_pci_dma_ops() == &dma_direct_ops)
675		set_dma_offset(dev, cell_dma_direct_offset);
676	else
677		BUG();
678}
679
680static void cell_pci_dma_dev_setup(struct pci_dev *dev)
681{
682	cell_dma_dev_setup(&dev->dev);
683}
684
685static int cell_of_bus_notify(struct notifier_block *nb, unsigned long action,
686			      void *data)
687{
688	struct device *dev = data;
689
690	/* We are only intereted in device addition */
691	if (action != BUS_NOTIFY_ADD_DEVICE)
692		return 0;
693
694	/* We use the PCI DMA ops */
695	dev->archdata.dma_ops = get_pci_dma_ops();
696
697	cell_dma_dev_setup(dev);
698
699	return 0;
700}
701
702static struct notifier_block cell_of_bus_notifier = {
703	.notifier_call = cell_of_bus_notify
704};
705
706static int __init cell_iommu_get_window(struct device_node *np,
707					 unsigned long *base,
708					 unsigned long *size)
709{
710	const __be32 *dma_window;
711	unsigned long index;
712
713	/* Use ibm,dma-window if available, else, hard code ! */
714	dma_window = of_get_property(np, "ibm,dma-window", NULL);
715	if (dma_window == NULL) {
716		*base = 0;
717		*size = 0x80000000u;
718		return -ENODEV;
719	}
720
721	of_parse_dma_window(np, dma_window, &index, base, size);
722	return 0;
723}
724
725static struct cbe_iommu * __init cell_iommu_alloc(struct device_node *np)
726{
727	struct cbe_iommu *iommu;
728	int nid, i;
729
730	/* Get node ID */
731	nid = of_node_to_nid(np);
732	if (nid < 0) {
733		printk(KERN_ERR "iommu: failed to get node for %s\n",
734		       np->full_name);
735		return NULL;
736	}
737	pr_debug("iommu: setting up iommu for node %d (%s)\n",
738		 nid, np->full_name);
739
740	/* XXX todo: If we can have multiple windows on the same IOMMU, which
741	 * isn't the case today, we probably want here to check whether the
742	 * iommu for that node is already setup.
743	 * However, there might be issue with getting the size right so let's
744	 * ignore that for now. We might want to completely get rid of the
745	 * multiple window support since the cell iommu supports per-page ioids
746	 */
747
748	if (cbe_nr_iommus >= NR_IOMMUS) {
749		printk(KERN_ERR "iommu: too many IOMMUs detected ! (%s)\n",
750		       np->full_name);
751		return NULL;
752	}
753
754	/* Init base fields */
755	i = cbe_nr_iommus++;
756	iommu = &iommus[i];
757	iommu->stab = NULL;
758	iommu->nid = nid;
759	snprintf(iommu->name, sizeof(iommu->name), "iommu%d", i);
760	INIT_LIST_HEAD(&iommu->windows);
761
762	return iommu;
763}
764
765static void __init cell_iommu_init_one(struct device_node *np,
766				       unsigned long offset)
767{
768	struct cbe_iommu *iommu;
769	unsigned long base, size;
770
771	iommu = cell_iommu_alloc(np);
772	if (!iommu)
773		return;
774
775	/* Obtain a window for it */
776	cell_iommu_get_window(np, &base, &size);
777
778	pr_debug("\ttranslating window 0x%lx...0x%lx\n",
779		 base, base + size - 1);
780
781	/* Initialize the hardware */
782	cell_iommu_setup_hardware(iommu, base, size);
783
784	/* Setup the iommu_table */
785	cell_iommu_setup_window(iommu, np, base, size,
786				offset >> IOMMU_PAGE_SHIFT_4K);
787}
788
789static void __init cell_disable_iommus(void)
790{
791	int node;
792	unsigned long base, val;
793	void __iomem *xregs, *cregs;
794
795	/* Make sure IOC translation is disabled on all nodes */
796	for_each_online_node(node) {
797		if (cell_iommu_find_ioc(node, &base))
798			continue;
799		xregs = ioremap(base, IOC_Reg_Size);
800		if (xregs == NULL)
801			continue;
802		cregs = xregs + IOC_IOCmd_Offset;
803
804		pr_debug("iommu: cleaning up iommu on node %d\n", node);
805
806		out_be64(xregs + IOC_IOST_Origin, 0);
807		(void)in_be64(xregs + IOC_IOST_Origin);
808		val = in_be64(cregs + IOC_IOCmd_Cfg);
809		val &= ~IOC_IOCmd_Cfg_TE;
810		out_be64(cregs + IOC_IOCmd_Cfg, val);
811		(void)in_be64(cregs + IOC_IOCmd_Cfg);
812
813		iounmap(xregs);
814	}
815}
816
817static int __init cell_iommu_init_disabled(void)
818{
819	struct device_node *np = NULL;
820	unsigned long base = 0, size;
821
822	/* When no iommu is present, we use direct DMA ops */
823	set_pci_dma_ops(&dma_direct_ops);
824
825	/* First make sure all IOC translation is turned off */
826	cell_disable_iommus();
827
828	/* If we have no Axon, we set up the spider DMA magic offset */
829	if (of_find_node_by_name(NULL, "axon") == NULL)
830		cell_dma_direct_offset = SPIDER_DMA_OFFSET;
831
832	/* Now we need to check to see where the memory is mapped
833	 * in PCI space. We assume that all busses use the same dma
834	 * window which is always the case so far on Cell, thus we
835	 * pick up the first pci-internal node we can find and check
836	 * the DMA window from there.
837	 */
838	for_each_node_by_name(np, "axon") {
839		if (np->parent == NULL || np->parent->parent != NULL)
840			continue;
841		if (cell_iommu_get_window(np, &base, &size) == 0)
842			break;
843	}
844	if (np == NULL) {
845		for_each_node_by_name(np, "pci-internal") {
846			if (np->parent == NULL || np->parent->parent != NULL)
847				continue;
848			if (cell_iommu_get_window(np, &base, &size) == 0)
849				break;
850		}
851	}
852	of_node_put(np);
853
854	/* If we found a DMA window, we check if it's big enough to enclose
855	 * all of physical memory. If not, we force enable IOMMU
856	 */
857	if (np && size < memblock_end_of_DRAM()) {
858		printk(KERN_WARNING "iommu: force-enabled, dma window"
859		       " (%ldMB) smaller than total memory (%lldMB)\n",
860		       size >> 20, memblock_end_of_DRAM() >> 20);
861		return -ENODEV;
862	}
863
864	cell_dma_direct_offset += base;
865
866	if (cell_dma_direct_offset != 0)
867		cell_pci_controller_ops.dma_dev_setup = cell_pci_dma_dev_setup;
868
869	printk("iommu: disabled, direct DMA offset is 0x%lx\n",
870	       cell_dma_direct_offset);
871
872	return 0;
873}
874
875/*
876 *  Fixed IOMMU mapping support
877 *
878 *  This code adds support for setting up a fixed IOMMU mapping on certain
879 *  cell machines. For 64-bit devices this avoids the performance overhead of
880 *  mapping and unmapping pages at runtime. 32-bit devices are unable to use
881 *  the fixed mapping.
882 *
883 *  The fixed mapping is established at boot, and maps all of physical memory
884 *  1:1 into device space at some offset. On machines with < 30 GB of memory
885 *  we setup the fixed mapping immediately above the normal IOMMU window.
886 *
887 *  For example a machine with 4GB of memory would end up with the normal
888 *  IOMMU window from 0-2GB and the fixed mapping window from 2GB to 6GB. In
889 *  this case a 64-bit device wishing to DMA to 1GB would be told to DMA to
890 *  3GB, plus any offset required by firmware. The firmware offset is encoded
891 *  in the "dma-ranges" property.
892 *
893 *  On machines with 30GB or more of memory, we are unable to place the fixed
894 *  mapping above the normal IOMMU window as we would run out of address space.
895 *  Instead we move the normal IOMMU window to coincide with the hash page
896 *  table, this region does not need to be part of the fixed mapping as no
897 *  device should ever be DMA'ing to it. We then setup the fixed mapping
898 *  from 0 to 32GB.
899 */
900
901static u64 cell_iommu_get_fixed_address(struct device *dev)
902{
903	u64 cpu_addr, size, best_size, dev_addr = OF_BAD_ADDR;
904	struct device_node *np;
905	const u32 *ranges = NULL;
906	int i, len, best, naddr, nsize, pna, range_size;
907
908	np = of_node_get(dev->of_node);
909	while (1) {
910		naddr = of_n_addr_cells(np);
911		nsize = of_n_size_cells(np);
912		np = of_get_next_parent(np);
913		if (!np)
914			break;
915
916		ranges = of_get_property(np, "dma-ranges", &len);
917
918		/* Ignore empty ranges, they imply no translation required */
919		if (ranges && len > 0)
920			break;
921	}
922
923	if (!ranges) {
924		dev_dbg(dev, "iommu: no dma-ranges found\n");
925		goto out;
926	}
927
928	len /= sizeof(u32);
929
930	pna = of_n_addr_cells(np);
931	range_size = naddr + nsize + pna;
932
933	/* dma-ranges format:
934	 * child addr	: naddr cells
935	 * parent addr	: pna cells
936	 * size		: nsize cells
937	 */
938	for (i = 0, best = -1, best_size = 0; i < len; i += range_size) {
939		cpu_addr = of_translate_dma_address(np, ranges + i + naddr);
940		size = of_read_number(ranges + i + naddr + pna, nsize);
941
942		if (cpu_addr == 0 && size > best_size) {
943			best = i;
944			best_size = size;
945		}
946	}
947
948	if (best >= 0) {
949		dev_addr = of_read_number(ranges + best, naddr);
950	} else
951		dev_dbg(dev, "iommu: no suitable range found!\n");
952
953out:
954	of_node_put(np);
955
956	return dev_addr;
957}
958
959static int dma_set_mask_and_switch(struct device *dev, u64 dma_mask)
960{
961	if (!dev->dma_mask || !dma_supported(dev, dma_mask))
962		return -EIO;
963
964	if (dma_mask == DMA_BIT_MASK(64) &&
965		cell_iommu_get_fixed_address(dev) != OF_BAD_ADDR)
966	{
967		dev_dbg(dev, "iommu: 64-bit OK, using fixed ops\n");
968		set_dma_ops(dev, &dma_iommu_fixed_ops);
969	} else {
970		dev_dbg(dev, "iommu: not 64-bit, using default ops\n");
971		set_dma_ops(dev, get_pci_dma_ops());
972	}
973
974	cell_dma_dev_setup(dev);
975
976	*dev->dma_mask = dma_mask;
977
978	return 0;
979}
980
981static void cell_dma_dev_setup_fixed(struct device *dev)
982{
983	u64 addr;
984
985	addr = cell_iommu_get_fixed_address(dev) + dma_iommu_fixed_base;
986	set_dma_offset(dev, addr);
987
988	dev_dbg(dev, "iommu: fixed addr = %llx\n", addr);
989}
990
991static 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) & CBE_IOPTE_RPN_Mask);
1004}
1005
1006static 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{
1010	unsigned long base_pte, uaddr, ioaddr, *ptab;
1011
1012	ptab = cell_iommu_alloc_ptab(iommu, fbase, fsize, dbase, dsize, 24);
1013
1014	dma_iommu_fixed_base = fbase;
1015
1016	pr_debug("iommu: mapping 0x%lx pages from 0x%lx\n", fsize, fbase);
1017
1018	base_pte = CBE_IOPTE_PP_W | CBE_IOPTE_PP_R | CBE_IOPTE_M |
1019		(cell_iommu_get_ioid(np) & CBE_IOPTE_IOID_Mask);
1020
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 |= CBE_IOPTE_SO_RW;
1026	}
1027
1028	for (uaddr = 0; uaddr < fsize; uaddr += (1 << 24)) {
1029		/* Don't touch the dynamic region */
1030		ioaddr = uaddr + fbase;
1031		if (ioaddr >= dbase && ioaddr < (dbase + dsize)) {
1032			pr_debug("iommu: fixed/dynamic overlap, skipping\n");
1033			continue;
1034		}
1035
1036		insert_16M_pte(uaddr, ptab, base_pte);
1037	}
1038
1039	mb();
1040}
1041
1042static 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	of_node_put(np);
1051
1052	if (!np) {
1053		pr_debug("iommu: fixed mapping disabled, no axons found\n");
1054		return -1;
1055	}
1056
1057	/* We must have dma-ranges properties for fixed mapping to work */
1058	np = of_find_node_with_property(NULL, "dma-ranges");
1059	of_node_put(np);
1060
1061	if (!np) {
1062		pr_debug("iommu: no dma-ranges found, no fixed mapping\n");
1063		return -1;
1064	}
1065
1066	/* The default setup is to have the fixed mapping sit after the
1067	 * dynamic region, so find the top of the largest IOMMU window
1068	 * on any axon, then add the size of RAM and that's our max value.
1069	 * If that is > 32GB we have to do other shennanigans.
1070	 */
1071	fbase = 0;
1072	for_each_node_by_name(np, "axon") {
1073		cell_iommu_get_window(np, &dbase, &dsize);
1074		fbase = max(fbase, dbase + dsize);
1075	}
1076
1077	fbase = _ALIGN_UP(fbase, 1 << IO_SEGMENT_SHIFT);
1078	fsize = memblock_phys_mem_size();
1079
1080	if ((fbase + fsize) <= 0x800000000ul)
1081		hbase = 0; /* use the device tree window */
1082	else {
1083		/* If we're over 32 GB we need to cheat. We can't map all of
1084		 * RAM with the fixed mapping, and also fit the dynamic
1085		 * region. So try to place the dynamic region where the hash
1086		 * table sits, drivers never need to DMA to it, we don't
1087		 * need a fixed mapping for that area.
1088		 */
1089		if (!htab_address) {
1090			pr_debug("iommu: htab is NULL, on LPAR? Huh?\n");
1091			return -1;
1092		}
1093		hbase = __pa(htab_address);
1094		hend  = hbase + htab_size_bytes;
1095
1096		/* The window must start and end on a segment boundary */
1097		if ((hbase != _ALIGN_UP(hbase, 1 << IO_SEGMENT_SHIFT)) ||
1098		    (hend != _ALIGN_UP(hend, 1 << IO_SEGMENT_SHIFT))) {
1099			pr_debug("iommu: hash window not segment aligned\n");
1100			return -1;
1101		}
1102
1103		/* Check the hash window fits inside the real DMA window */
1104		for_each_node_by_name(np, "axon") {
1105			cell_iommu_get_window(np, &dbase, &dsize);
1106
1107			if (hbase < dbase || (hend > (dbase + dsize))) {
1108				pr_debug("iommu: hash window doesn't fit in"
1109					 "real DMA window\n");
1110				return -1;
1111			}
1112		}
1113
1114		fbase = 0;
1115	}
1116
1117	/* Setup the dynamic regions */
1118	for_each_node_by_name(np, "axon") {
1119		iommu = cell_iommu_alloc(np);
1120		BUG_ON(!iommu);
1121
1122		if (hbase == 0)
1123			cell_iommu_get_window(np, &dbase, &dsize);
1124		else {
1125			dbase = hbase;
1126			dsize = htab_size_bytes;
1127		}
1128
1129		printk(KERN_DEBUG "iommu: node %d, dynamic window 0x%lx-0x%lx "
1130			"fixed window 0x%lx-0x%lx\n", iommu->nid, dbase,
1131			 dbase + dsize, fbase, fbase + fsize);
1132
1133		cell_iommu_setup_stab(iommu, dbase, dsize, fbase, fsize);
1134		iommu->ptab = cell_iommu_alloc_ptab(iommu, dbase, dsize, 0, 0,
1135						    IOMMU_PAGE_SHIFT_4K);
1136		cell_iommu_setup_fixed_ptab(iommu, np, dbase, dsize,
1137					     fbase, fsize);
1138		cell_iommu_enable_hardware(iommu);
1139		cell_iommu_setup_window(iommu, np, dbase, dsize, 0);
1140	}
1141
1142	dma_iommu_ops.set_dma_mask = dma_set_mask_and_switch;
1143	set_pci_dma_ops(&dma_iommu_ops);
1144
1145	return 0;
1146}
1147
1148static int iommu_fixed_disabled;
1149
1150static int __init setup_iommu_fixed(char *str)
1151{
1152	struct device_node *pciep;
1153
1154	if (strcmp(str, "off") == 0)
1155		iommu_fixed_disabled = 1;
1156
1157	/* If we can find a pcie-endpoint in the device tree assume that
1158	 * we're on a triblade or a CAB so by default the fixed mapping
1159	 * should be set to be weakly ordered; but only if the boot
1160	 * option WASN'T set for strong ordering
1161	 */
1162	pciep = of_find_node_by_type(NULL, "pcie-endpoint");
1163
1164	if (strcmp(str, "weak") == 0 || (pciep && strcmp(str, "strong") != 0))
1165		iommu_fixed_is_weak = 1;
1166
1167	of_node_put(pciep);
1168
1169	return 1;
1170}
1171__setup("iommu_fixed=", setup_iommu_fixed);
1172
1173static u64 cell_dma_get_required_mask(struct device *dev)
1174{
1175	struct dma_map_ops *dma_ops;
1176
1177	if (!dev->dma_mask)
1178		return 0;
1179
1180	if (!iommu_fixed_disabled &&
1181			cell_iommu_get_fixed_address(dev) != OF_BAD_ADDR)
1182		return DMA_BIT_MASK(64);
1183
1184	dma_ops = get_dma_ops(dev);
1185	if (dma_ops->get_required_mask)
1186		return dma_ops->get_required_mask(dev);
1187
1188	WARN_ONCE(1, "no get_required_mask in %p ops", dma_ops);
1189
1190	return DMA_BIT_MASK(64);
1191}
1192
1193static int __init cell_iommu_init(void)
1194{
1195	struct device_node *np;
1196
1197	/* If IOMMU is disabled or we have little enough RAM to not need
1198	 * to enable it, we setup a direct mapping.
1199	 *
1200	 * Note: should we make sure we have the IOMMU actually disabled ?
1201	 */
1202	if (iommu_is_off ||
1203	    (!iommu_force_on && memblock_end_of_DRAM() <= 0x80000000ull))
1204		if (cell_iommu_init_disabled() == 0)
1205			goto bail;
1206
1207	/* Setup various callbacks */
1208	cell_pci_controller_ops.dma_dev_setup = cell_pci_dma_dev_setup;
1209	ppc_md.dma_get_required_mask = cell_dma_get_required_mask;
1210
1211	if (!iommu_fixed_disabled && cell_iommu_fixed_mapping_init() == 0)
1212		goto bail;
1213
1214	/* Create an iommu for each /axon node.  */
1215	for_each_node_by_name(np, "axon") {
1216		if (np->parent == NULL || np->parent->parent != NULL)
1217			continue;
1218		cell_iommu_init_one(np, 0);
1219	}
1220
1221	/* Create an iommu for each toplevel /pci-internal node for
1222	 * old hardware/firmware
1223	 */
1224	for_each_node_by_name(np, "pci-internal") {
1225		if (np->parent == NULL || np->parent->parent != NULL)
1226			continue;
1227		cell_iommu_init_one(np, SPIDER_DMA_OFFSET);
1228	}
1229
1230	/* Setup default PCI iommu ops */
1231	set_pci_dma_ops(&dma_iommu_ops);
1232
1233 bail:
1234	/* Register callbacks on OF platform device addition/removal
1235	 * to handle linking them to the right DMA operations
1236	 */
1237	bus_register_notifier(&platform_bus_type, &cell_of_bus_notifier);
1238
1239	return 0;
1240}
1241machine_arch_initcall(cell, cell_iommu_init);
1242