1/*
2 * Copyright (C) 2001 Mike Corrigan & Dave Engebretsen, IBM Corporation
3 *
4 * Rewrite, cleanup:
5 *
6 * Copyright (C) 2004 Olof Johansson <olof@lixom.net>, IBM Corporation
7 * Copyright (C) 2006 Olof Johansson <olof@lixom.net>
8 *
9 * Dynamic DMA mapping support, pSeries-specific parts, both SMP and LPAR.
10 *
11 *
12 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU General Public License as published by
14 * the Free Software Foundation; either version 2 of the License, or
15 * (at your option) any later version.
16 *
17 * This program is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
20 * GNU General Public License for more details.
21 *
22 * You should have received a copy of the GNU General Public License
23 * along with this program; if not, write to the Free Software
24 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
25 */
26
27#include <linux/init.h>
28#include <linux/types.h>
29#include <linux/slab.h>
30#include <linux/mm.h>
31#include <linux/memblock.h>
32#include <linux/spinlock.h>
33#include <linux/string.h>
34#include <linux/pci.h>
35#include <linux/dma-mapping.h>
36#include <linux/crash_dump.h>
37#include <linux/memory.h>
38#include <linux/of.h>
39#include <asm/io.h>
40#include <asm/prom.h>
41#include <asm/rtas.h>
42#include <asm/iommu.h>
43#include <asm/pci-bridge.h>
44#include <asm/machdep.h>
45#include <asm/firmware.h>
46#include <asm/tce.h>
47#include <asm/ppc-pci.h>
48#include <asm/udbg.h>
49#include <asm/mmzone.h>
50#include <asm/plpar_wrappers.h>
51
52#include "pseries.h"
53
54static void tce_invalidate_pSeries_sw(struct iommu_table *tbl,
55				      __be64 *startp, __be64 *endp)
56{
57	u64 __iomem *invalidate = (u64 __iomem *)tbl->it_index;
58	unsigned long start, end, inc;
59
60	start = __pa(startp);
61	end = __pa(endp);
62	inc = L1_CACHE_BYTES; /* invalidate a cacheline of TCEs at a time */
63
64	/* If this is non-zero, change the format.  We shift the
65	 * address and or in the magic from the device tree. */
66	if (tbl->it_busno) {
67		start <<= 12;
68		end <<= 12;
69		inc <<= 12;
70		start |= tbl->it_busno;
71		end |= tbl->it_busno;
72	}
73
74	end |= inc - 1; /* round up end to be different than start */
75
76	mb(); /* Make sure TCEs in memory are written */
77	while (start <= end) {
78		out_be64(invalidate, start);
79		start += inc;
80	}
81}
82
83static int tce_build_pSeries(struct iommu_table *tbl, long index,
84			      long npages, unsigned long uaddr,
85			      enum dma_data_direction direction,
86			      struct dma_attrs *attrs)
87{
88	u64 proto_tce;
89	__be64 *tcep, *tces;
90	u64 rpn;
91
92	proto_tce = TCE_PCI_READ; // Read allowed
93
94	if (direction != DMA_TO_DEVICE)
95		proto_tce |= TCE_PCI_WRITE;
96
97	tces = tcep = ((__be64 *)tbl->it_base) + index;
98
99	while (npages--) {
100		/* can't move this out since we might cross MEMBLOCK boundary */
101		rpn = __pa(uaddr) >> TCE_SHIFT;
102		*tcep = cpu_to_be64(proto_tce | (rpn & TCE_RPN_MASK) << TCE_RPN_SHIFT);
103
104		uaddr += TCE_PAGE_SIZE;
105		tcep++;
106	}
107
108	if (tbl->it_type & TCE_PCI_SWINV_CREATE)
109		tce_invalidate_pSeries_sw(tbl, tces, tcep - 1);
110	return 0;
111}
112
113
114static void tce_free_pSeries(struct iommu_table *tbl, long index, long npages)
115{
116	__be64 *tcep, *tces;
117
118	tces = tcep = ((__be64 *)tbl->it_base) + index;
119
120	while (npages--)
121		*(tcep++) = 0;
122
123	if (tbl->it_type & TCE_PCI_SWINV_FREE)
124		tce_invalidate_pSeries_sw(tbl, tces, tcep - 1);
125}
126
127static unsigned long tce_get_pseries(struct iommu_table *tbl, long index)
128{
129	__be64 *tcep;
130
131	tcep = ((__be64 *)tbl->it_base) + index;
132
133	return be64_to_cpu(*tcep);
134}
135
136static void tce_free_pSeriesLP(struct iommu_table*, long, long);
137static void tce_freemulti_pSeriesLP(struct iommu_table*, long, long);
138
139static int tce_build_pSeriesLP(struct iommu_table *tbl, long tcenum,
140				long npages, unsigned long uaddr,
141				enum dma_data_direction direction,
142				struct dma_attrs *attrs)
143{
144	u64 rc = 0;
145	u64 proto_tce, tce;
146	u64 rpn;
147	int ret = 0;
148	long tcenum_start = tcenum, npages_start = npages;
149
150	rpn = __pa(uaddr) >> TCE_SHIFT;
151	proto_tce = TCE_PCI_READ;
152	if (direction != DMA_TO_DEVICE)
153		proto_tce |= TCE_PCI_WRITE;
154
155	while (npages--) {
156		tce = proto_tce | (rpn & TCE_RPN_MASK) << TCE_RPN_SHIFT;
157		rc = plpar_tce_put((u64)tbl->it_index, (u64)tcenum << 12, tce);
158
159		if (unlikely(rc == H_NOT_ENOUGH_RESOURCES)) {
160			ret = (int)rc;
161			tce_free_pSeriesLP(tbl, tcenum_start,
162			                   (npages_start - (npages + 1)));
163			break;
164		}
165
166		if (rc && printk_ratelimit()) {
167			printk("tce_build_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
168			printk("\tindex   = 0x%llx\n", (u64)tbl->it_index);
169			printk("\ttcenum  = 0x%llx\n", (u64)tcenum);
170			printk("\ttce val = 0x%llx\n", tce );
171			dump_stack();
172		}
173
174		tcenum++;
175		rpn++;
176	}
177	return ret;
178}
179
180static DEFINE_PER_CPU(__be64 *, tce_page);
181
182static int tce_buildmulti_pSeriesLP(struct iommu_table *tbl, long tcenum,
183				     long npages, unsigned long uaddr,
184				     enum dma_data_direction direction,
185				     struct dma_attrs *attrs)
186{
187	u64 rc = 0;
188	u64 proto_tce;
189	__be64 *tcep;
190	u64 rpn;
191	long l, limit;
192	long tcenum_start = tcenum, npages_start = npages;
193	int ret = 0;
194	unsigned long flags;
195
196	if (npages == 1) {
197		return tce_build_pSeriesLP(tbl, tcenum, npages, uaddr,
198		                           direction, attrs);
199	}
200
201	local_irq_save(flags);	/* to protect tcep and the page behind it */
202
203	tcep = __this_cpu_read(tce_page);
204
205	/* This is safe to do since interrupts are off when we're called
206	 * from iommu_alloc{,_sg}()
207	 */
208	if (!tcep) {
209		tcep = (__be64 *)__get_free_page(GFP_ATOMIC);
210		/* If allocation fails, fall back to the loop implementation */
211		if (!tcep) {
212			local_irq_restore(flags);
213			return tce_build_pSeriesLP(tbl, tcenum, npages, uaddr,
214					    direction, attrs);
215		}
216		__this_cpu_write(tce_page, tcep);
217	}
218
219	rpn = __pa(uaddr) >> TCE_SHIFT;
220	proto_tce = TCE_PCI_READ;
221	if (direction != DMA_TO_DEVICE)
222		proto_tce |= TCE_PCI_WRITE;
223
224	/* We can map max one pageful of TCEs at a time */
225	do {
226		/*
227		 * Set up the page with TCE data, looping through and setting
228		 * the values.
229		 */
230		limit = min_t(long, npages, 4096/TCE_ENTRY_SIZE);
231
232		for (l = 0; l < limit; l++) {
233			tcep[l] = cpu_to_be64(proto_tce | (rpn & TCE_RPN_MASK) << TCE_RPN_SHIFT);
234			rpn++;
235		}
236
237		rc = plpar_tce_put_indirect((u64)tbl->it_index,
238					    (u64)tcenum << 12,
239					    (u64)__pa(tcep),
240					    limit);
241
242		npages -= limit;
243		tcenum += limit;
244	} while (npages > 0 && !rc);
245
246	local_irq_restore(flags);
247
248	if (unlikely(rc == H_NOT_ENOUGH_RESOURCES)) {
249		ret = (int)rc;
250		tce_freemulti_pSeriesLP(tbl, tcenum_start,
251		                        (npages_start - (npages + limit)));
252		return ret;
253	}
254
255	if (rc && printk_ratelimit()) {
256		printk("tce_buildmulti_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
257		printk("\tindex   = 0x%llx\n", (u64)tbl->it_index);
258		printk("\tnpages  = 0x%llx\n", (u64)npages);
259		printk("\ttce[0] val = 0x%llx\n", tcep[0]);
260		dump_stack();
261	}
262	return ret;
263}
264
265static void tce_free_pSeriesLP(struct iommu_table *tbl, long tcenum, long npages)
266{
267	u64 rc;
268
269	while (npages--) {
270		rc = plpar_tce_put((u64)tbl->it_index, (u64)tcenum << 12, 0);
271
272		if (rc && printk_ratelimit()) {
273			printk("tce_free_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
274			printk("\tindex   = 0x%llx\n", (u64)tbl->it_index);
275			printk("\ttcenum  = 0x%llx\n", (u64)tcenum);
276			dump_stack();
277		}
278
279		tcenum++;
280	}
281}
282
283
284static void tce_freemulti_pSeriesLP(struct iommu_table *tbl, long tcenum, long npages)
285{
286	u64 rc;
287
288	rc = plpar_tce_stuff((u64)tbl->it_index, (u64)tcenum << 12, 0, npages);
289
290	if (rc && printk_ratelimit()) {
291		printk("tce_freemulti_pSeriesLP: plpar_tce_stuff failed\n");
292		printk("\trc      = %lld\n", rc);
293		printk("\tindex   = 0x%llx\n", (u64)tbl->it_index);
294		printk("\tnpages  = 0x%llx\n", (u64)npages);
295		dump_stack();
296	}
297}
298
299static unsigned long tce_get_pSeriesLP(struct iommu_table *tbl, long tcenum)
300{
301	u64 rc;
302	unsigned long tce_ret;
303
304	rc = plpar_tce_get((u64)tbl->it_index, (u64)tcenum << 12, &tce_ret);
305
306	if (rc && printk_ratelimit()) {
307		printk("tce_get_pSeriesLP: plpar_tce_get failed. rc=%lld\n", rc);
308		printk("\tindex   = 0x%llx\n", (u64)tbl->it_index);
309		printk("\ttcenum  = 0x%llx\n", (u64)tcenum);
310		dump_stack();
311	}
312
313	return tce_ret;
314}
315
316/* this is compatible with cells for the device tree property */
317struct dynamic_dma_window_prop {
318	__be32	liobn;		/* tce table number */
319	__be64	dma_base;	/* address hi,lo */
320	__be32	tce_shift;	/* ilog2(tce_page_size) */
321	__be32	window_shift;	/* ilog2(tce_window_size) */
322};
323
324struct direct_window {
325	struct device_node *device;
326	const struct dynamic_dma_window_prop *prop;
327	struct list_head list;
328};
329
330/* Dynamic DMA Window support */
331struct ddw_query_response {
332	u32 windows_available;
333	u32 largest_available_block;
334	u32 page_size;
335	u32 migration_capable;
336};
337
338struct ddw_create_response {
339	u32 liobn;
340	u32 addr_hi;
341	u32 addr_lo;
342};
343
344static LIST_HEAD(direct_window_list);
345/* prevents races between memory on/offline and window creation */
346static DEFINE_SPINLOCK(direct_window_list_lock);
347/* protects initializing window twice for same device */
348static DEFINE_MUTEX(direct_window_init_mutex);
349#define DIRECT64_PROPNAME "linux,direct64-ddr-window-info"
350
351static int tce_clearrange_multi_pSeriesLP(unsigned long start_pfn,
352					unsigned long num_pfn, const void *arg)
353{
354	const struct dynamic_dma_window_prop *maprange = arg;
355	int rc;
356	u64 tce_size, num_tce, dma_offset, next;
357	u32 tce_shift;
358	long limit;
359
360	tce_shift = be32_to_cpu(maprange->tce_shift);
361	tce_size = 1ULL << tce_shift;
362	next = start_pfn << PAGE_SHIFT;
363	num_tce = num_pfn << PAGE_SHIFT;
364
365	/* round back to the beginning of the tce page size */
366	num_tce += next & (tce_size - 1);
367	next &= ~(tce_size - 1);
368
369	/* covert to number of tces */
370	num_tce |= tce_size - 1;
371	num_tce >>= tce_shift;
372
373	do {
374		/*
375		 * Set up the page with TCE data, looping through and setting
376		 * the values.
377		 */
378		limit = min_t(long, num_tce, 512);
379		dma_offset = next + be64_to_cpu(maprange->dma_base);
380
381		rc = plpar_tce_stuff((u64)be32_to_cpu(maprange->liobn),
382					     dma_offset,
383					     0, limit);
384		next += limit * tce_size;
385		num_tce -= limit;
386	} while (num_tce > 0 && !rc);
387
388	return rc;
389}
390
391static int tce_setrange_multi_pSeriesLP(unsigned long start_pfn,
392					unsigned long num_pfn, const void *arg)
393{
394	const struct dynamic_dma_window_prop *maprange = arg;
395	u64 tce_size, num_tce, dma_offset, next, proto_tce, liobn;
396	__be64 *tcep;
397	u32 tce_shift;
398	u64 rc = 0;
399	long l, limit;
400
401	local_irq_disable();	/* to protect tcep and the page behind it */
402	tcep = __this_cpu_read(tce_page);
403
404	if (!tcep) {
405		tcep = (__be64 *)__get_free_page(GFP_ATOMIC);
406		if (!tcep) {
407			local_irq_enable();
408			return -ENOMEM;
409		}
410		__this_cpu_write(tce_page, tcep);
411	}
412
413	proto_tce = TCE_PCI_READ | TCE_PCI_WRITE;
414
415	liobn = (u64)be32_to_cpu(maprange->liobn);
416	tce_shift = be32_to_cpu(maprange->tce_shift);
417	tce_size = 1ULL << tce_shift;
418	next = start_pfn << PAGE_SHIFT;
419	num_tce = num_pfn << PAGE_SHIFT;
420
421	/* round back to the beginning of the tce page size */
422	num_tce += next & (tce_size - 1);
423	next &= ~(tce_size - 1);
424
425	/* covert to number of tces */
426	num_tce |= tce_size - 1;
427	num_tce >>= tce_shift;
428
429	/* We can map max one pageful of TCEs at a time */
430	do {
431		/*
432		 * Set up the page with TCE data, looping through and setting
433		 * the values.
434		 */
435		limit = min_t(long, num_tce, 4096/TCE_ENTRY_SIZE);
436		dma_offset = next + be64_to_cpu(maprange->dma_base);
437
438		for (l = 0; l < limit; l++) {
439			tcep[l] = cpu_to_be64(proto_tce | next);
440			next += tce_size;
441		}
442
443		rc = plpar_tce_put_indirect(liobn,
444					    dma_offset,
445					    (u64)__pa(tcep),
446					    limit);
447
448		num_tce -= limit;
449	} while (num_tce > 0 && !rc);
450
451	/* error cleanup: caller will clear whole range */
452
453	local_irq_enable();
454	return rc;
455}
456
457static int tce_setrange_multi_pSeriesLP_walk(unsigned long start_pfn,
458		unsigned long num_pfn, void *arg)
459{
460	return tce_setrange_multi_pSeriesLP(start_pfn, num_pfn, arg);
461}
462
463
464#ifdef CONFIG_PCI
465static void iommu_table_setparms(struct pci_controller *phb,
466				 struct device_node *dn,
467				 struct iommu_table *tbl)
468{
469	struct device_node *node;
470	const unsigned long *basep, *sw_inval;
471	const u32 *sizep;
472
473	node = phb->dn;
474
475	basep = of_get_property(node, "linux,tce-base", NULL);
476	sizep = of_get_property(node, "linux,tce-size", NULL);
477	if (basep == NULL || sizep == NULL) {
478		printk(KERN_ERR "PCI_DMA: iommu_table_setparms: %s has "
479				"missing tce entries !\n", dn->full_name);
480		return;
481	}
482
483	tbl->it_base = (unsigned long)__va(*basep);
484
485	if (!is_kdump_kernel())
486		memset((void *)tbl->it_base, 0, *sizep);
487
488	tbl->it_busno = phb->bus->number;
489	tbl->it_page_shift = IOMMU_PAGE_SHIFT_4K;
490
491	/* Units of tce entries */
492	tbl->it_offset = phb->dma_window_base_cur >> tbl->it_page_shift;
493
494	/* Test if we are going over 2GB of DMA space */
495	if (phb->dma_window_base_cur + phb->dma_window_size > 0x80000000ul) {
496		udbg_printf("PCI_DMA: Unexpected number of IOAs under this PHB.\n");
497		panic("PCI_DMA: Unexpected number of IOAs under this PHB.\n");
498	}
499
500	phb->dma_window_base_cur += phb->dma_window_size;
501
502	/* Set the tce table size - measured in entries */
503	tbl->it_size = phb->dma_window_size >> tbl->it_page_shift;
504
505	tbl->it_index = 0;
506	tbl->it_blocksize = 16;
507	tbl->it_type = TCE_PCI;
508
509	sw_inval = of_get_property(node, "linux,tce-sw-invalidate-info", NULL);
510	if (sw_inval) {
511		/*
512		 * This property contains information on how to
513		 * invalidate the TCE entry.  The first property is
514		 * the base MMIO address used to invalidate entries.
515		 * The second property tells us the format of the TCE
516		 * invalidate (whether it needs to be shifted) and
517		 * some magic routing info to add to our invalidate
518		 * command.
519		 */
520		tbl->it_index = (unsigned long) ioremap(sw_inval[0], 8);
521		tbl->it_busno = sw_inval[1]; /* overload this with magic */
522		tbl->it_type = TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE;
523	}
524}
525
526/*
527 * iommu_table_setparms_lpar
528 *
529 * Function: On pSeries LPAR systems, return TCE table info, given a pci bus.
530 */
531static void iommu_table_setparms_lpar(struct pci_controller *phb,
532				      struct device_node *dn,
533				      struct iommu_table *tbl,
534				      const __be32 *dma_window)
535{
536	unsigned long offset, size;
537
538	of_parse_dma_window(dn, dma_window, &tbl->it_index, &offset, &size);
539
540	tbl->it_busno = phb->bus->number;
541	tbl->it_page_shift = IOMMU_PAGE_SHIFT_4K;
542	tbl->it_base   = 0;
543	tbl->it_blocksize  = 16;
544	tbl->it_type = TCE_PCI;
545	tbl->it_offset = offset >> tbl->it_page_shift;
546	tbl->it_size = size >> tbl->it_page_shift;
547}
548
549static void pci_dma_bus_setup_pSeries(struct pci_bus *bus)
550{
551	struct device_node *dn;
552	struct iommu_table *tbl;
553	struct device_node *isa_dn, *isa_dn_orig;
554	struct device_node *tmp;
555	struct pci_dn *pci;
556	int children;
557
558	dn = pci_bus_to_OF_node(bus);
559
560	pr_debug("pci_dma_bus_setup_pSeries: setting up bus %s\n", dn->full_name);
561
562	if (bus->self) {
563		/* This is not a root bus, any setup will be done for the
564		 * device-side of the bridge in iommu_dev_setup_pSeries().
565		 */
566		return;
567	}
568	pci = PCI_DN(dn);
569
570	/* Check if the ISA bus on the system is under
571	 * this PHB.
572	 */
573	isa_dn = isa_dn_orig = of_find_node_by_type(NULL, "isa");
574
575	while (isa_dn && isa_dn != dn)
576		isa_dn = isa_dn->parent;
577
578	of_node_put(isa_dn_orig);
579
580	/* Count number of direct PCI children of the PHB. */
581	for (children = 0, tmp = dn->child; tmp; tmp = tmp->sibling)
582		children++;
583
584	pr_debug("Children: %d\n", children);
585
586	/* Calculate amount of DMA window per slot. Each window must be
587	 * a power of two (due to pci_alloc_consistent requirements).
588	 *
589	 * Keep 256MB aside for PHBs with ISA.
590	 */
591
592	if (!isa_dn) {
593		/* No ISA/IDE - just set window size and return */
594		pci->phb->dma_window_size = 0x80000000ul; /* To be divided */
595
596		while (pci->phb->dma_window_size * children > 0x80000000ul)
597			pci->phb->dma_window_size >>= 1;
598		pr_debug("No ISA/IDE, window size is 0x%llx\n",
599			 pci->phb->dma_window_size);
600		pci->phb->dma_window_base_cur = 0;
601
602		return;
603	}
604
605	/* If we have ISA, then we probably have an IDE
606	 * controller too. Allocate a 128MB table but
607	 * skip the first 128MB to avoid stepping on ISA
608	 * space.
609	 */
610	pci->phb->dma_window_size = 0x8000000ul;
611	pci->phb->dma_window_base_cur = 0x8000000ul;
612
613	tbl = kzalloc_node(sizeof(struct iommu_table), GFP_KERNEL,
614			   pci->phb->node);
615
616	iommu_table_setparms(pci->phb, dn, tbl);
617	pci->iommu_table = iommu_init_table(tbl, pci->phb->node);
618	iommu_register_group(tbl, pci_domain_nr(bus), 0);
619
620	/* Divide the rest (1.75GB) among the children */
621	pci->phb->dma_window_size = 0x80000000ul;
622	while (pci->phb->dma_window_size * children > 0x70000000ul)
623		pci->phb->dma_window_size >>= 1;
624
625	pr_debug("ISA/IDE, window size is 0x%llx\n", pci->phb->dma_window_size);
626}
627
628
629static void pci_dma_bus_setup_pSeriesLP(struct pci_bus *bus)
630{
631	struct iommu_table *tbl;
632	struct device_node *dn, *pdn;
633	struct pci_dn *ppci;
634	const __be32 *dma_window = NULL;
635
636	dn = pci_bus_to_OF_node(bus);
637
638	pr_debug("pci_dma_bus_setup_pSeriesLP: setting up bus %s\n",
639		 dn->full_name);
640
641	/* Find nearest ibm,dma-window, walking up the device tree */
642	for (pdn = dn; pdn != NULL; pdn = pdn->parent) {
643		dma_window = of_get_property(pdn, "ibm,dma-window", NULL);
644		if (dma_window != NULL)
645			break;
646	}
647
648	if (dma_window == NULL) {
649		pr_debug("  no ibm,dma-window property !\n");
650		return;
651	}
652
653	ppci = PCI_DN(pdn);
654
655	pr_debug("  parent is %s, iommu_table: 0x%p\n",
656		 pdn->full_name, ppci->iommu_table);
657
658	if (!ppci->iommu_table) {
659		tbl = kzalloc_node(sizeof(struct iommu_table), GFP_KERNEL,
660				   ppci->phb->node);
661		iommu_table_setparms_lpar(ppci->phb, pdn, tbl, dma_window);
662		ppci->iommu_table = iommu_init_table(tbl, ppci->phb->node);
663		iommu_register_group(tbl, pci_domain_nr(bus), 0);
664		pr_debug("  created table: %p\n", ppci->iommu_table);
665	}
666}
667
668
669static void pci_dma_dev_setup_pSeries(struct pci_dev *dev)
670{
671	struct device_node *dn;
672	struct iommu_table *tbl;
673
674	pr_debug("pci_dma_dev_setup_pSeries: %s\n", pci_name(dev));
675
676	dn = dev->dev.of_node;
677
678	/* If we're the direct child of a root bus, then we need to allocate
679	 * an iommu table ourselves. The bus setup code should have setup
680	 * the window sizes already.
681	 */
682	if (!dev->bus->self) {
683		struct pci_controller *phb = PCI_DN(dn)->phb;
684
685		pr_debug(" --> first child, no bridge. Allocating iommu table.\n");
686		tbl = kzalloc_node(sizeof(struct iommu_table), GFP_KERNEL,
687				   phb->node);
688		iommu_table_setparms(phb, dn, tbl);
689		PCI_DN(dn)->iommu_table = iommu_init_table(tbl, phb->node);
690		iommu_register_group(tbl, pci_domain_nr(phb->bus), 0);
691		set_iommu_table_base_and_group(&dev->dev,
692					       PCI_DN(dn)->iommu_table);
693		return;
694	}
695
696	/* If this device is further down the bus tree, search upwards until
697	 * an already allocated iommu table is found and use that.
698	 */
699
700	while (dn && PCI_DN(dn) && PCI_DN(dn)->iommu_table == NULL)
701		dn = dn->parent;
702
703	if (dn && PCI_DN(dn))
704		set_iommu_table_base_and_group(&dev->dev,
705					       PCI_DN(dn)->iommu_table);
706	else
707		printk(KERN_WARNING "iommu: Device %s has no iommu table\n",
708		       pci_name(dev));
709}
710
711static int __read_mostly disable_ddw;
712
713static int __init disable_ddw_setup(char *str)
714{
715	disable_ddw = 1;
716	printk(KERN_INFO "ppc iommu: disabling ddw.\n");
717
718	return 0;
719}
720
721early_param("disable_ddw", disable_ddw_setup);
722
723static void remove_ddw(struct device_node *np, bool remove_prop)
724{
725	struct dynamic_dma_window_prop *dwp;
726	struct property *win64;
727	u32 ddw_avail[3];
728	u64 liobn;
729	int ret = 0;
730
731	ret = of_property_read_u32_array(np, "ibm,ddw-applicable",
732					 &ddw_avail[0], 3);
733
734	win64 = of_find_property(np, DIRECT64_PROPNAME, NULL);
735	if (!win64)
736		return;
737
738	if (ret || win64->length < sizeof(*dwp))
739		goto delprop;
740
741	dwp = win64->value;
742	liobn = (u64)be32_to_cpu(dwp->liobn);
743
744	/* clear the whole window, note the arg is in kernel pages */
745	ret = tce_clearrange_multi_pSeriesLP(0,
746		1ULL << (be32_to_cpu(dwp->window_shift) - PAGE_SHIFT), dwp);
747	if (ret)
748		pr_warning("%s failed to clear tces in window.\n",
749			 np->full_name);
750	else
751		pr_debug("%s successfully cleared tces in window.\n",
752			 np->full_name);
753
754	ret = rtas_call(ddw_avail[2], 1, 1, NULL, liobn);
755	if (ret)
756		pr_warning("%s: failed to remove direct window: rtas returned "
757			"%d to ibm,remove-pe-dma-window(%x) %llx\n",
758			np->full_name, ret, ddw_avail[2], liobn);
759	else
760		pr_debug("%s: successfully removed direct window: rtas returned "
761			"%d to ibm,remove-pe-dma-window(%x) %llx\n",
762			np->full_name, ret, ddw_avail[2], liobn);
763
764delprop:
765	if (remove_prop)
766		ret = of_remove_property(np, win64);
767	if (ret)
768		pr_warning("%s: failed to remove direct window property: %d\n",
769			np->full_name, ret);
770}
771
772static u64 find_existing_ddw(struct device_node *pdn)
773{
774	struct direct_window *window;
775	const struct dynamic_dma_window_prop *direct64;
776	u64 dma_addr = 0;
777
778	spin_lock(&direct_window_list_lock);
779	/* check if we already created a window and dupe that config if so */
780	list_for_each_entry(window, &direct_window_list, list) {
781		if (window->device == pdn) {
782			direct64 = window->prop;
783			dma_addr = be64_to_cpu(direct64->dma_base);
784			break;
785		}
786	}
787	spin_unlock(&direct_window_list_lock);
788
789	return dma_addr;
790}
791
792static int find_existing_ddw_windows(void)
793{
794	int len;
795	struct device_node *pdn;
796	struct direct_window *window;
797	const struct dynamic_dma_window_prop *direct64;
798
799	if (!firmware_has_feature(FW_FEATURE_LPAR))
800		return 0;
801
802	for_each_node_with_property(pdn, DIRECT64_PROPNAME) {
803		direct64 = of_get_property(pdn, DIRECT64_PROPNAME, &len);
804		if (!direct64)
805			continue;
806
807		window = kzalloc(sizeof(*window), GFP_KERNEL);
808		if (!window || len < sizeof(struct dynamic_dma_window_prop)) {
809			kfree(window);
810			remove_ddw(pdn, true);
811			continue;
812		}
813
814		window->device = pdn;
815		window->prop = direct64;
816		spin_lock(&direct_window_list_lock);
817		list_add(&window->list, &direct_window_list);
818		spin_unlock(&direct_window_list_lock);
819	}
820
821	return 0;
822}
823machine_arch_initcall(pseries, find_existing_ddw_windows);
824
825static int query_ddw(struct pci_dev *dev, const u32 *ddw_avail,
826			struct ddw_query_response *query)
827{
828	struct eeh_dev *edev;
829	u32 cfg_addr;
830	u64 buid;
831	int ret;
832
833	/*
834	 * Get the config address and phb buid of the PE window.
835	 * Rely on eeh to retrieve this for us.
836	 * Retrieve them from the pci device, not the node with the
837	 * dma-window property
838	 */
839	edev = pci_dev_to_eeh_dev(dev);
840	cfg_addr = edev->config_addr;
841	if (edev->pe_config_addr)
842		cfg_addr = edev->pe_config_addr;
843	buid = edev->phb->buid;
844
845	ret = rtas_call(ddw_avail[0], 3, 5, (u32 *)query,
846		  cfg_addr, BUID_HI(buid), BUID_LO(buid));
847	dev_info(&dev->dev, "ibm,query-pe-dma-windows(%x) %x %x %x"
848		" returned %d\n", ddw_avail[0], cfg_addr, BUID_HI(buid),
849		BUID_LO(buid), ret);
850	return ret;
851}
852
853static int create_ddw(struct pci_dev *dev, const u32 *ddw_avail,
854			struct ddw_create_response *create, int page_shift,
855			int window_shift)
856{
857	struct eeh_dev *edev;
858	u32 cfg_addr;
859	u64 buid;
860	int ret;
861
862	/*
863	 * Get the config address and phb buid of the PE window.
864	 * Rely on eeh to retrieve this for us.
865	 * Retrieve them from the pci device, not the node with the
866	 * dma-window property
867	 */
868	edev = pci_dev_to_eeh_dev(dev);
869	cfg_addr = edev->config_addr;
870	if (edev->pe_config_addr)
871		cfg_addr = edev->pe_config_addr;
872	buid = edev->phb->buid;
873
874	do {
875		/* extra outputs are LIOBN and dma-addr (hi, lo) */
876		ret = rtas_call(ddw_avail[1], 5, 4, (u32 *)create,
877				cfg_addr, BUID_HI(buid), BUID_LO(buid),
878				page_shift, window_shift);
879	} while (rtas_busy_delay(ret));
880	dev_info(&dev->dev,
881		"ibm,create-pe-dma-window(%x) %x %x %x %x %x returned %d "
882		"(liobn = 0x%x starting addr = %x %x)\n", ddw_avail[1],
883		 cfg_addr, BUID_HI(buid), BUID_LO(buid), page_shift,
884		 window_shift, ret, create->liobn, create->addr_hi, create->addr_lo);
885
886	return ret;
887}
888
889struct failed_ddw_pdn {
890	struct device_node *pdn;
891	struct list_head list;
892};
893
894static LIST_HEAD(failed_ddw_pdn_list);
895
896/*
897 * If the PE supports dynamic dma windows, and there is space for a table
898 * that can map all pages in a linear offset, then setup such a table,
899 * and record the dma-offset in the struct device.
900 *
901 * dev: the pci device we are checking
902 * pdn: the parent pe node with the ibm,dma_window property
903 * Future: also check if we can remap the base window for our base page size
904 *
905 * returns the dma offset for use by dma_set_mask
906 */
907static u64 enable_ddw(struct pci_dev *dev, struct device_node *pdn)
908{
909	int len, ret;
910	struct ddw_query_response query;
911	struct ddw_create_response create;
912	int page_shift;
913	u64 dma_addr, max_addr;
914	struct device_node *dn;
915	u32 ddw_avail[3];
916	struct direct_window *window;
917	struct property *win64;
918	struct dynamic_dma_window_prop *ddwprop;
919	struct failed_ddw_pdn *fpdn;
920
921	mutex_lock(&direct_window_init_mutex);
922
923	dma_addr = find_existing_ddw(pdn);
924	if (dma_addr != 0)
925		goto out_unlock;
926
927	/*
928	 * If we already went through this for a previous function of
929	 * the same device and failed, we don't want to muck with the
930	 * DMA window again, as it will race with in-flight operations
931	 * and can lead to EEHs. The above mutex protects access to the
932	 * list.
933	 */
934	list_for_each_entry(fpdn, &failed_ddw_pdn_list, list) {
935		if (!strcmp(fpdn->pdn->full_name, pdn->full_name))
936			goto out_unlock;
937	}
938
939	/*
940	 * the ibm,ddw-applicable property holds the tokens for:
941	 * ibm,query-pe-dma-window
942	 * ibm,create-pe-dma-window
943	 * ibm,remove-pe-dma-window
944	 * for the given node in that order.
945	 * the property is actually in the parent, not the PE
946	 */
947	ret = of_property_read_u32_array(pdn, "ibm,ddw-applicable",
948					 &ddw_avail[0], 3);
949	if (ret)
950		goto out_failed;
951
952       /*
953	 * Query if there is a second window of size to map the
954	 * whole partition.  Query returns number of windows, largest
955	 * block assigned to PE (partition endpoint), and two bitmasks
956	 * of page sizes: supported and supported for migrate-dma.
957	 */
958	dn = pci_device_to_OF_node(dev);
959	ret = query_ddw(dev, ddw_avail, &query);
960	if (ret != 0)
961		goto out_failed;
962
963	if (query.windows_available == 0) {
964		/*
965		 * no additional windows are available for this device.
966		 * We might be able to reallocate the existing window,
967		 * trading in for a larger page size.
968		 */
969		dev_dbg(&dev->dev, "no free dynamic windows");
970		goto out_failed;
971	}
972	if (query.page_size & 4) {
973		page_shift = 24; /* 16MB */
974	} else if (query.page_size & 2) {
975		page_shift = 16; /* 64kB */
976	} else if (query.page_size & 1) {
977		page_shift = 12; /* 4kB */
978	} else {
979		dev_dbg(&dev->dev, "no supported direct page size in mask %x",
980			  query.page_size);
981		goto out_failed;
982	}
983	/* verify the window * number of ptes will map the partition */
984	/* check largest block * page size > max memory hotplug addr */
985	max_addr = memory_hotplug_max();
986	if (query.largest_available_block < (max_addr >> page_shift)) {
987		dev_dbg(&dev->dev, "can't map partiton max 0x%llx with %u "
988			  "%llu-sized pages\n", max_addr,  query.largest_available_block,
989			  1ULL << page_shift);
990		goto out_failed;
991	}
992	len = order_base_2(max_addr);
993	win64 = kzalloc(sizeof(struct property), GFP_KERNEL);
994	if (!win64) {
995		dev_info(&dev->dev,
996			"couldn't allocate property for 64bit dma window\n");
997		goto out_failed;
998	}
999	win64->name = kstrdup(DIRECT64_PROPNAME, GFP_KERNEL);
1000	win64->value = ddwprop = kmalloc(sizeof(*ddwprop), GFP_KERNEL);
1001	win64->length = sizeof(*ddwprop);
1002	if (!win64->name || !win64->value) {
1003		dev_info(&dev->dev,
1004			"couldn't allocate property name and value\n");
1005		goto out_free_prop;
1006	}
1007
1008	ret = create_ddw(dev, ddw_avail, &create, page_shift, len);
1009	if (ret != 0)
1010		goto out_free_prop;
1011
1012	ddwprop->liobn = cpu_to_be32(create.liobn);
1013	ddwprop->dma_base = cpu_to_be64(((u64)create.addr_hi << 32) |
1014			create.addr_lo);
1015	ddwprop->tce_shift = cpu_to_be32(page_shift);
1016	ddwprop->window_shift = cpu_to_be32(len);
1017
1018	dev_dbg(&dev->dev, "created tce table LIOBN 0x%x for %s\n",
1019		  create.liobn, dn->full_name);
1020
1021	window = kzalloc(sizeof(*window), GFP_KERNEL);
1022	if (!window)
1023		goto out_clear_window;
1024
1025	ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> PAGE_SHIFT,
1026			win64->value, tce_setrange_multi_pSeriesLP_walk);
1027	if (ret) {
1028		dev_info(&dev->dev, "failed to map direct window for %s: %d\n",
1029			 dn->full_name, ret);
1030		goto out_free_window;
1031	}
1032
1033	ret = of_add_property(pdn, win64);
1034	if (ret) {
1035		dev_err(&dev->dev, "unable to add dma window property for %s: %d",
1036			 pdn->full_name, ret);
1037		goto out_free_window;
1038	}
1039
1040	window->device = pdn;
1041	window->prop = ddwprop;
1042	spin_lock(&direct_window_list_lock);
1043	list_add(&window->list, &direct_window_list);
1044	spin_unlock(&direct_window_list_lock);
1045
1046	dma_addr = be64_to_cpu(ddwprop->dma_base);
1047	goto out_unlock;
1048
1049out_free_window:
1050	kfree(window);
1051
1052out_clear_window:
1053	remove_ddw(pdn, true);
1054
1055out_free_prop:
1056	kfree(win64->name);
1057	kfree(win64->value);
1058	kfree(win64);
1059
1060out_failed:
1061
1062	fpdn = kzalloc(sizeof(*fpdn), GFP_KERNEL);
1063	if (!fpdn)
1064		goto out_unlock;
1065	fpdn->pdn = pdn;
1066	list_add(&fpdn->list, &failed_ddw_pdn_list);
1067
1068out_unlock:
1069	mutex_unlock(&direct_window_init_mutex);
1070	return dma_addr;
1071}
1072
1073static void pci_dma_dev_setup_pSeriesLP(struct pci_dev *dev)
1074{
1075	struct device_node *pdn, *dn;
1076	struct iommu_table *tbl;
1077	const __be32 *dma_window = NULL;
1078	struct pci_dn *pci;
1079
1080	pr_debug("pci_dma_dev_setup_pSeriesLP: %s\n", pci_name(dev));
1081
1082	/* dev setup for LPAR is a little tricky, since the device tree might
1083	 * contain the dma-window properties per-device and not necessarily
1084	 * for the bus. So we need to search upwards in the tree until we
1085	 * either hit a dma-window property, OR find a parent with a table
1086	 * already allocated.
1087	 */
1088	dn = pci_device_to_OF_node(dev);
1089	pr_debug("  node is %s\n", dn->full_name);
1090
1091	for (pdn = dn; pdn && PCI_DN(pdn) && !PCI_DN(pdn)->iommu_table;
1092	     pdn = pdn->parent) {
1093		dma_window = of_get_property(pdn, "ibm,dma-window", NULL);
1094		if (dma_window)
1095			break;
1096	}
1097
1098	if (!pdn || !PCI_DN(pdn)) {
1099		printk(KERN_WARNING "pci_dma_dev_setup_pSeriesLP: "
1100		       "no DMA window found for pci dev=%s dn=%s\n",
1101				 pci_name(dev), of_node_full_name(dn));
1102		return;
1103	}
1104	pr_debug("  parent is %s\n", pdn->full_name);
1105
1106	pci = PCI_DN(pdn);
1107	if (!pci->iommu_table) {
1108		tbl = kzalloc_node(sizeof(struct iommu_table), GFP_KERNEL,
1109				   pci->phb->node);
1110		iommu_table_setparms_lpar(pci->phb, pdn, tbl, dma_window);
1111		pci->iommu_table = iommu_init_table(tbl, pci->phb->node);
1112		iommu_register_group(tbl, pci_domain_nr(pci->phb->bus), 0);
1113		pr_debug("  created table: %p\n", pci->iommu_table);
1114	} else {
1115		pr_debug("  found DMA window, table: %p\n", pci->iommu_table);
1116	}
1117
1118	set_iommu_table_base_and_group(&dev->dev, pci->iommu_table);
1119}
1120
1121static int dma_set_mask_pSeriesLP(struct device *dev, u64 dma_mask)
1122{
1123	bool ddw_enabled = false;
1124	struct device_node *pdn, *dn;
1125	struct pci_dev *pdev;
1126	const __be32 *dma_window = NULL;
1127	u64 dma_offset;
1128
1129	if (!dev->dma_mask)
1130		return -EIO;
1131
1132	if (!dev_is_pci(dev))
1133		goto check_mask;
1134
1135	pdev = to_pci_dev(dev);
1136
1137	/* only attempt to use a new window if 64-bit DMA is requested */
1138	if (!disable_ddw && dma_mask == DMA_BIT_MASK(64)) {
1139		dn = pci_device_to_OF_node(pdev);
1140		dev_dbg(dev, "node is %s\n", dn->full_name);
1141
1142		/*
1143		 * the device tree might contain the dma-window properties
1144		 * per-device and not necessarily for the bus. So we need to
1145		 * search upwards in the tree until we either hit a dma-window
1146		 * property, OR find a parent with a table already allocated.
1147		 */
1148		for (pdn = dn; pdn && PCI_DN(pdn) && !PCI_DN(pdn)->iommu_table;
1149				pdn = pdn->parent) {
1150			dma_window = of_get_property(pdn, "ibm,dma-window", NULL);
1151			if (dma_window)
1152				break;
1153		}
1154		if (pdn && PCI_DN(pdn)) {
1155			dma_offset = enable_ddw(pdev, pdn);
1156			if (dma_offset != 0) {
1157				dev_info(dev, "Using 64-bit direct DMA at offset %llx\n", dma_offset);
1158				set_dma_offset(dev, dma_offset);
1159				set_dma_ops(dev, &dma_direct_ops);
1160				ddw_enabled = true;
1161			}
1162		}
1163	}
1164
1165	/* fall back on iommu ops, restore table pointer with ops */
1166	if (!ddw_enabled && get_dma_ops(dev) != &dma_iommu_ops) {
1167		dev_info(dev, "Restoring 32-bit DMA via iommu\n");
1168		set_dma_ops(dev, &dma_iommu_ops);
1169		pci_dma_dev_setup_pSeriesLP(pdev);
1170	}
1171
1172check_mask:
1173	if (!dma_supported(dev, dma_mask))
1174		return -EIO;
1175
1176	*dev->dma_mask = dma_mask;
1177	return 0;
1178}
1179
1180static u64 dma_get_required_mask_pSeriesLP(struct device *dev)
1181{
1182	if (!dev->dma_mask)
1183		return 0;
1184
1185	if (!disable_ddw && dev_is_pci(dev)) {
1186		struct pci_dev *pdev = to_pci_dev(dev);
1187		struct device_node *dn;
1188
1189		dn = pci_device_to_OF_node(pdev);
1190
1191		/* search upwards for ibm,dma-window */
1192		for (; dn && PCI_DN(dn) && !PCI_DN(dn)->iommu_table;
1193				dn = dn->parent)
1194			if (of_get_property(dn, "ibm,dma-window", NULL))
1195				break;
1196		/* if there is a ibm,ddw-applicable property require 64 bits */
1197		if (dn && PCI_DN(dn) &&
1198				of_get_property(dn, "ibm,ddw-applicable", NULL))
1199			return DMA_BIT_MASK(64);
1200	}
1201
1202	return dma_iommu_ops.get_required_mask(dev);
1203}
1204
1205#else  /* CONFIG_PCI */
1206#define pci_dma_bus_setup_pSeries	NULL
1207#define pci_dma_dev_setup_pSeries	NULL
1208#define pci_dma_bus_setup_pSeriesLP	NULL
1209#define pci_dma_dev_setup_pSeriesLP	NULL
1210#define dma_set_mask_pSeriesLP		NULL
1211#define dma_get_required_mask_pSeriesLP	NULL
1212#endif /* !CONFIG_PCI */
1213
1214static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
1215		void *data)
1216{
1217	struct direct_window *window;
1218	struct memory_notify *arg = data;
1219	int ret = 0;
1220
1221	switch (action) {
1222	case MEM_GOING_ONLINE:
1223		spin_lock(&direct_window_list_lock);
1224		list_for_each_entry(window, &direct_window_list, list) {
1225			ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn,
1226					arg->nr_pages, window->prop);
1227			/* XXX log error */
1228		}
1229		spin_unlock(&direct_window_list_lock);
1230		break;
1231	case MEM_CANCEL_ONLINE:
1232	case MEM_OFFLINE:
1233		spin_lock(&direct_window_list_lock);
1234		list_for_each_entry(window, &direct_window_list, list) {
1235			ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn,
1236					arg->nr_pages, window->prop);
1237			/* XXX log error */
1238		}
1239		spin_unlock(&direct_window_list_lock);
1240		break;
1241	default:
1242		break;
1243	}
1244	if (ret && action != MEM_CANCEL_ONLINE)
1245		return NOTIFY_BAD;
1246
1247	return NOTIFY_OK;
1248}
1249
1250static struct notifier_block iommu_mem_nb = {
1251	.notifier_call = iommu_mem_notifier,
1252};
1253
1254static int iommu_reconfig_notifier(struct notifier_block *nb, unsigned long action, void *data)
1255{
1256	int err = NOTIFY_OK;
1257	struct of_reconfig_data *rd = data;
1258	struct device_node *np = rd->dn;
1259	struct pci_dn *pci = PCI_DN(np);
1260	struct direct_window *window;
1261
1262	switch (action) {
1263	case OF_RECONFIG_DETACH_NODE:
1264		/*
1265		 * Removing the property will invoke the reconfig
1266		 * notifier again, which causes dead-lock on the
1267		 * read-write semaphore of the notifier chain. So
1268		 * we have to remove the property when releasing
1269		 * the device node.
1270		 */
1271		remove_ddw(np, false);
1272		if (pci && pci->iommu_table)
1273			iommu_free_table(pci->iommu_table, np->full_name);
1274
1275		spin_lock(&direct_window_list_lock);
1276		list_for_each_entry(window, &direct_window_list, list) {
1277			if (window->device == np) {
1278				list_del(&window->list);
1279				kfree(window);
1280				break;
1281			}
1282		}
1283		spin_unlock(&direct_window_list_lock);
1284		break;
1285	default:
1286		err = NOTIFY_DONE;
1287		break;
1288	}
1289	return err;
1290}
1291
1292static struct notifier_block iommu_reconfig_nb = {
1293	.notifier_call = iommu_reconfig_notifier,
1294};
1295
1296/* These are called very early. */
1297void iommu_init_early_pSeries(void)
1298{
1299	if (of_chosen && of_get_property(of_chosen, "linux,iommu-off", NULL))
1300		return;
1301
1302	if (firmware_has_feature(FW_FEATURE_LPAR)) {
1303		if (firmware_has_feature(FW_FEATURE_MULTITCE)) {
1304			ppc_md.tce_build = tce_buildmulti_pSeriesLP;
1305			ppc_md.tce_free	 = tce_freemulti_pSeriesLP;
1306		} else {
1307			ppc_md.tce_build = tce_build_pSeriesLP;
1308			ppc_md.tce_free	 = tce_free_pSeriesLP;
1309		}
1310		ppc_md.tce_get   = tce_get_pSeriesLP;
1311		pseries_pci_controller_ops.dma_bus_setup = pci_dma_bus_setup_pSeriesLP;
1312		pseries_pci_controller_ops.dma_dev_setup = pci_dma_dev_setup_pSeriesLP;
1313		ppc_md.dma_set_mask = dma_set_mask_pSeriesLP;
1314		ppc_md.dma_get_required_mask = dma_get_required_mask_pSeriesLP;
1315	} else {
1316		ppc_md.tce_build = tce_build_pSeries;
1317		ppc_md.tce_free  = tce_free_pSeries;
1318		ppc_md.tce_get   = tce_get_pseries;
1319		pseries_pci_controller_ops.dma_bus_setup = pci_dma_bus_setup_pSeries;
1320		pseries_pci_controller_ops.dma_dev_setup = pci_dma_dev_setup_pSeries;
1321	}
1322
1323
1324	of_reconfig_notifier_register(&iommu_reconfig_nb);
1325	register_memory_notifier(&iommu_mem_nb);
1326
1327	set_pci_dma_ops(&dma_iommu_ops);
1328}
1329
1330static int __init disable_multitce(char *str)
1331{
1332	if (strcmp(str, "off") == 0 &&
1333	    firmware_has_feature(FW_FEATURE_LPAR) &&
1334	    firmware_has_feature(FW_FEATURE_MULTITCE)) {
1335		printk(KERN_INFO "Disabling MULTITCE firmware feature\n");
1336		ppc_md.tce_build = tce_build_pSeriesLP;
1337		ppc_md.tce_free	 = tce_free_pSeriesLP;
1338		powerpc_firmware_features &= ~FW_FEATURE_MULTITCE;
1339	}
1340	return 1;
1341}
1342
1343__setup("multitce=", disable_multitce);
1344
1345machine_subsys_initcall_sync(pseries, tce_iommu_bus_notifier_init);
1346