]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - arch/powerpc/platforms/pseries/iommu.c
Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[karo-tx-linux.git] / arch / powerpc / platforms / pseries / iommu.c
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 <linux/iommu.h>
40 #include <linux/rculist.h>
41 #include <asm/io.h>
42 #include <asm/prom.h>
43 #include <asm/rtas.h>
44 #include <asm/iommu.h>
45 #include <asm/pci-bridge.h>
46 #include <asm/machdep.h>
47 #include <asm/firmware.h>
48 #include <asm/tce.h>
49 #include <asm/ppc-pci.h>
50 #include <asm/udbg.h>
51 #include <asm/mmzone.h>
52 #include <asm/plpar_wrappers.h>
53
54 #include "pseries.h"
55
56 static struct iommu_table_group *iommu_pseries_alloc_group(int node)
57 {
58         struct iommu_table_group *table_group = NULL;
59         struct iommu_table *tbl = NULL;
60         struct iommu_table_group_link *tgl = NULL;
61
62         table_group = kzalloc_node(sizeof(struct iommu_table_group), GFP_KERNEL,
63                            node);
64         if (!table_group)
65                 goto fail_exit;
66
67         tbl = kzalloc_node(sizeof(struct iommu_table), GFP_KERNEL, node);
68         if (!tbl)
69                 goto fail_exit;
70
71         tgl = kzalloc_node(sizeof(struct iommu_table_group_link), GFP_KERNEL,
72                         node);
73         if (!tgl)
74                 goto fail_exit;
75
76         INIT_LIST_HEAD_RCU(&tbl->it_group_list);
77         tgl->table_group = table_group;
78         list_add_rcu(&tgl->next, &tbl->it_group_list);
79
80         table_group->tables[0] = tbl;
81
82         return table_group;
83
84 fail_exit:
85         kfree(tgl);
86         kfree(table_group);
87         kfree(tbl);
88
89         return NULL;
90 }
91
92 static void iommu_pseries_free_group(struct iommu_table_group *table_group,
93                 const char *node_name)
94 {
95         struct iommu_table *tbl;
96 #ifdef CONFIG_IOMMU_API
97         struct iommu_table_group_link *tgl;
98 #endif
99
100         if (!table_group)
101                 return;
102
103         tbl = table_group->tables[0];
104 #ifdef CONFIG_IOMMU_API
105         tgl = list_first_entry_or_null(&tbl->it_group_list,
106                         struct iommu_table_group_link, next);
107
108         WARN_ON_ONCE(!tgl);
109         if (tgl) {
110                 list_del_rcu(&tgl->next);
111                 kfree(tgl);
112         }
113         if (table_group->group) {
114                 iommu_group_put(table_group->group);
115                 BUG_ON(table_group->group);
116         }
117 #endif
118         iommu_free_table(tbl, node_name);
119
120         kfree(table_group);
121 }
122
123 static void tce_invalidate_pSeries_sw(struct iommu_table *tbl,
124                                       __be64 *startp, __be64 *endp)
125 {
126         u64 __iomem *invalidate = (u64 __iomem *)tbl->it_index;
127         unsigned long start, end, inc;
128
129         start = __pa(startp);
130         end = __pa(endp);
131         inc = L1_CACHE_BYTES; /* invalidate a cacheline of TCEs at a time */
132
133         /* If this is non-zero, change the format.  We shift the
134          * address and or in the magic from the device tree. */
135         if (tbl->it_busno) {
136                 start <<= 12;
137                 end <<= 12;
138                 inc <<= 12;
139                 start |= tbl->it_busno;
140                 end |= tbl->it_busno;
141         }
142
143         end |= inc - 1; /* round up end to be different than start */
144
145         mb(); /* Make sure TCEs in memory are written */
146         while (start <= end) {
147                 out_be64(invalidate, start);
148                 start += inc;
149         }
150 }
151
152 static int tce_build_pSeries(struct iommu_table *tbl, long index,
153                               long npages, unsigned long uaddr,
154                               enum dma_data_direction direction,
155                               struct dma_attrs *attrs)
156 {
157         u64 proto_tce;
158         __be64 *tcep, *tces;
159         u64 rpn;
160
161         proto_tce = TCE_PCI_READ; // Read allowed
162
163         if (direction != DMA_TO_DEVICE)
164                 proto_tce |= TCE_PCI_WRITE;
165
166         tces = tcep = ((__be64 *)tbl->it_base) + index;
167
168         while (npages--) {
169                 /* can't move this out since we might cross MEMBLOCK boundary */
170                 rpn = __pa(uaddr) >> TCE_SHIFT;
171                 *tcep = cpu_to_be64(proto_tce | (rpn & TCE_RPN_MASK) << TCE_RPN_SHIFT);
172
173                 uaddr += TCE_PAGE_SIZE;
174                 tcep++;
175         }
176
177         if (tbl->it_type & TCE_PCI_SWINV_CREATE)
178                 tce_invalidate_pSeries_sw(tbl, tces, tcep - 1);
179         return 0;
180 }
181
182
183 static void tce_free_pSeries(struct iommu_table *tbl, long index, long npages)
184 {
185         __be64 *tcep, *tces;
186
187         tces = tcep = ((__be64 *)tbl->it_base) + index;
188
189         while (npages--)
190                 *(tcep++) = 0;
191
192         if (tbl->it_type & TCE_PCI_SWINV_FREE)
193                 tce_invalidate_pSeries_sw(tbl, tces, tcep - 1);
194 }
195
196 static unsigned long tce_get_pseries(struct iommu_table *tbl, long index)
197 {
198         __be64 *tcep;
199
200         tcep = ((__be64 *)tbl->it_base) + index;
201
202         return be64_to_cpu(*tcep);
203 }
204
205 static void tce_free_pSeriesLP(struct iommu_table*, long, long);
206 static void tce_freemulti_pSeriesLP(struct iommu_table*, long, long);
207
208 static int tce_build_pSeriesLP(struct iommu_table *tbl, long tcenum,
209                                 long npages, unsigned long uaddr,
210                                 enum dma_data_direction direction,
211                                 struct dma_attrs *attrs)
212 {
213         u64 rc = 0;
214         u64 proto_tce, tce;
215         u64 rpn;
216         int ret = 0;
217         long tcenum_start = tcenum, npages_start = npages;
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         while (npages--) {
225                 tce = proto_tce | (rpn & TCE_RPN_MASK) << TCE_RPN_SHIFT;
226                 rc = plpar_tce_put((u64)tbl->it_index, (u64)tcenum << 12, tce);
227
228                 if (unlikely(rc == H_NOT_ENOUGH_RESOURCES)) {
229                         ret = (int)rc;
230                         tce_free_pSeriesLP(tbl, tcenum_start,
231                                            (npages_start - (npages + 1)));
232                         break;
233                 }
234
235                 if (rc && printk_ratelimit()) {
236                         printk("tce_build_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
237                         printk("\tindex   = 0x%llx\n", (u64)tbl->it_index);
238                         printk("\ttcenum  = 0x%llx\n", (u64)tcenum);
239                         printk("\ttce val = 0x%llx\n", tce );
240                         dump_stack();
241                 }
242
243                 tcenum++;
244                 rpn++;
245         }
246         return ret;
247 }
248
249 static DEFINE_PER_CPU(__be64 *, tce_page);
250
251 static int tce_buildmulti_pSeriesLP(struct iommu_table *tbl, long tcenum,
252                                      long npages, unsigned long uaddr,
253                                      enum dma_data_direction direction,
254                                      struct dma_attrs *attrs)
255 {
256         u64 rc = 0;
257         u64 proto_tce;
258         __be64 *tcep;
259         u64 rpn;
260         long l, limit;
261         long tcenum_start = tcenum, npages_start = npages;
262         int ret = 0;
263         unsigned long flags;
264
265         if ((npages == 1) || !firmware_has_feature(FW_FEATURE_MULTITCE)) {
266                 return tce_build_pSeriesLP(tbl, tcenum, npages, uaddr,
267                                            direction, attrs);
268         }
269
270         local_irq_save(flags);  /* to protect tcep and the page behind it */
271
272         tcep = __this_cpu_read(tce_page);
273
274         /* This is safe to do since interrupts are off when we're called
275          * from iommu_alloc{,_sg}()
276          */
277         if (!tcep) {
278                 tcep = (__be64 *)__get_free_page(GFP_ATOMIC);
279                 /* If allocation fails, fall back to the loop implementation */
280                 if (!tcep) {
281                         local_irq_restore(flags);
282                         return tce_build_pSeriesLP(tbl, tcenum, npages, uaddr,
283                                             direction, attrs);
284                 }
285                 __this_cpu_write(tce_page, tcep);
286         }
287
288         rpn = __pa(uaddr) >> TCE_SHIFT;
289         proto_tce = TCE_PCI_READ;
290         if (direction != DMA_TO_DEVICE)
291                 proto_tce |= TCE_PCI_WRITE;
292
293         /* We can map max one pageful of TCEs at a time */
294         do {
295                 /*
296                  * Set up the page with TCE data, looping through and setting
297                  * the values.
298                  */
299                 limit = min_t(long, npages, 4096/TCE_ENTRY_SIZE);
300
301                 for (l = 0; l < limit; l++) {
302                         tcep[l] = cpu_to_be64(proto_tce | (rpn & TCE_RPN_MASK) << TCE_RPN_SHIFT);
303                         rpn++;
304                 }
305
306                 rc = plpar_tce_put_indirect((u64)tbl->it_index,
307                                             (u64)tcenum << 12,
308                                             (u64)__pa(tcep),
309                                             limit);
310
311                 npages -= limit;
312                 tcenum += limit;
313         } while (npages > 0 && !rc);
314
315         local_irq_restore(flags);
316
317         if (unlikely(rc == H_NOT_ENOUGH_RESOURCES)) {
318                 ret = (int)rc;
319                 tce_freemulti_pSeriesLP(tbl, tcenum_start,
320                                         (npages_start - (npages + limit)));
321                 return ret;
322         }
323
324         if (rc && printk_ratelimit()) {
325                 printk("tce_buildmulti_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
326                 printk("\tindex   = 0x%llx\n", (u64)tbl->it_index);
327                 printk("\tnpages  = 0x%llx\n", (u64)npages);
328                 printk("\ttce[0] val = 0x%llx\n", tcep[0]);
329                 dump_stack();
330         }
331         return ret;
332 }
333
334 static void tce_free_pSeriesLP(struct iommu_table *tbl, long tcenum, long npages)
335 {
336         u64 rc;
337
338         while (npages--) {
339                 rc = plpar_tce_put((u64)tbl->it_index, (u64)tcenum << 12, 0);
340
341                 if (rc && printk_ratelimit()) {
342                         printk("tce_free_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
343                         printk("\tindex   = 0x%llx\n", (u64)tbl->it_index);
344                         printk("\ttcenum  = 0x%llx\n", (u64)tcenum);
345                         dump_stack();
346                 }
347
348                 tcenum++;
349         }
350 }
351
352
353 static void tce_freemulti_pSeriesLP(struct iommu_table *tbl, long tcenum, long npages)
354 {
355         u64 rc;
356
357         if (!firmware_has_feature(FW_FEATURE_MULTITCE))
358                 return tce_free_pSeriesLP(tbl, tcenum, npages);
359
360         rc = plpar_tce_stuff((u64)tbl->it_index, (u64)tcenum << 12, 0, npages);
361
362         if (rc && printk_ratelimit()) {
363                 printk("tce_freemulti_pSeriesLP: plpar_tce_stuff failed\n");
364                 printk("\trc      = %lld\n", rc);
365                 printk("\tindex   = 0x%llx\n", (u64)tbl->it_index);
366                 printk("\tnpages  = 0x%llx\n", (u64)npages);
367                 dump_stack();
368         }
369 }
370
371 static unsigned long tce_get_pSeriesLP(struct iommu_table *tbl, long tcenum)
372 {
373         u64 rc;
374         unsigned long tce_ret;
375
376         rc = plpar_tce_get((u64)tbl->it_index, (u64)tcenum << 12, &tce_ret);
377
378         if (rc && printk_ratelimit()) {
379                 printk("tce_get_pSeriesLP: plpar_tce_get failed. rc=%lld\n", rc);
380                 printk("\tindex   = 0x%llx\n", (u64)tbl->it_index);
381                 printk("\ttcenum  = 0x%llx\n", (u64)tcenum);
382                 dump_stack();
383         }
384
385         return tce_ret;
386 }
387
388 /* this is compatible with cells for the device tree property */
389 struct dynamic_dma_window_prop {
390         __be32  liobn;          /* tce table number */
391         __be64  dma_base;       /* address hi,lo */
392         __be32  tce_shift;      /* ilog2(tce_page_size) */
393         __be32  window_shift;   /* ilog2(tce_window_size) */
394 };
395
396 struct direct_window {
397         struct device_node *device;
398         const struct dynamic_dma_window_prop *prop;
399         struct list_head list;
400 };
401
402 /* Dynamic DMA Window support */
403 struct ddw_query_response {
404         u32 windows_available;
405         u32 largest_available_block;
406         u32 page_size;
407         u32 migration_capable;
408 };
409
410 struct ddw_create_response {
411         u32 liobn;
412         u32 addr_hi;
413         u32 addr_lo;
414 };
415
416 static LIST_HEAD(direct_window_list);
417 /* prevents races between memory on/offline and window creation */
418 static DEFINE_SPINLOCK(direct_window_list_lock);
419 /* protects initializing window twice for same device */
420 static DEFINE_MUTEX(direct_window_init_mutex);
421 #define DIRECT64_PROPNAME "linux,direct64-ddr-window-info"
422
423 static int tce_clearrange_multi_pSeriesLP(unsigned long start_pfn,
424                                         unsigned long num_pfn, const void *arg)
425 {
426         const struct dynamic_dma_window_prop *maprange = arg;
427         int rc;
428         u64 tce_size, num_tce, dma_offset, next;
429         u32 tce_shift;
430         long limit;
431
432         tce_shift = be32_to_cpu(maprange->tce_shift);
433         tce_size = 1ULL << tce_shift;
434         next = start_pfn << PAGE_SHIFT;
435         num_tce = num_pfn << PAGE_SHIFT;
436
437         /* round back to the beginning of the tce page size */
438         num_tce += next & (tce_size - 1);
439         next &= ~(tce_size - 1);
440
441         /* covert to number of tces */
442         num_tce |= tce_size - 1;
443         num_tce >>= tce_shift;
444
445         do {
446                 /*
447                  * Set up the page with TCE data, looping through and setting
448                  * the values.
449                  */
450                 limit = min_t(long, num_tce, 512);
451                 dma_offset = next + be64_to_cpu(maprange->dma_base);
452
453                 rc = plpar_tce_stuff((u64)be32_to_cpu(maprange->liobn),
454                                              dma_offset,
455                                              0, limit);
456                 next += limit * tce_size;
457                 num_tce -= limit;
458         } while (num_tce > 0 && !rc);
459
460         return rc;
461 }
462
463 static int tce_setrange_multi_pSeriesLP(unsigned long start_pfn,
464                                         unsigned long num_pfn, const void *arg)
465 {
466         const struct dynamic_dma_window_prop *maprange = arg;
467         u64 tce_size, num_tce, dma_offset, next, proto_tce, liobn;
468         __be64 *tcep;
469         u32 tce_shift;
470         u64 rc = 0;
471         long l, limit;
472
473         local_irq_disable();    /* to protect tcep and the page behind it */
474         tcep = __this_cpu_read(tce_page);
475
476         if (!tcep) {
477                 tcep = (__be64 *)__get_free_page(GFP_ATOMIC);
478                 if (!tcep) {
479                         local_irq_enable();
480                         return -ENOMEM;
481                 }
482                 __this_cpu_write(tce_page, tcep);
483         }
484
485         proto_tce = TCE_PCI_READ | TCE_PCI_WRITE;
486
487         liobn = (u64)be32_to_cpu(maprange->liobn);
488         tce_shift = be32_to_cpu(maprange->tce_shift);
489         tce_size = 1ULL << tce_shift;
490         next = start_pfn << PAGE_SHIFT;
491         num_tce = num_pfn << PAGE_SHIFT;
492
493         /* round back to the beginning of the tce page size */
494         num_tce += next & (tce_size - 1);
495         next &= ~(tce_size - 1);
496
497         /* covert to number of tces */
498         num_tce |= tce_size - 1;
499         num_tce >>= tce_shift;
500
501         /* We can map max one pageful of TCEs at a time */
502         do {
503                 /*
504                  * Set up the page with TCE data, looping through and setting
505                  * the values.
506                  */
507                 limit = min_t(long, num_tce, 4096/TCE_ENTRY_SIZE);
508                 dma_offset = next + be64_to_cpu(maprange->dma_base);
509
510                 for (l = 0; l < limit; l++) {
511                         tcep[l] = cpu_to_be64(proto_tce | next);
512                         next += tce_size;
513                 }
514
515                 rc = plpar_tce_put_indirect(liobn,
516                                             dma_offset,
517                                             (u64)__pa(tcep),
518                                             limit);
519
520                 num_tce -= limit;
521         } while (num_tce > 0 && !rc);
522
523         /* error cleanup: caller will clear whole range */
524
525         local_irq_enable();
526         return rc;
527 }
528
529 static int tce_setrange_multi_pSeriesLP_walk(unsigned long start_pfn,
530                 unsigned long num_pfn, void *arg)
531 {
532         return tce_setrange_multi_pSeriesLP(start_pfn, num_pfn, arg);
533 }
534
535 static void iommu_table_setparms(struct pci_controller *phb,
536                                  struct device_node *dn,
537                                  struct iommu_table *tbl)
538 {
539         struct device_node *node;
540         const unsigned long *basep, *sw_inval;
541         const u32 *sizep;
542
543         node = phb->dn;
544
545         basep = of_get_property(node, "linux,tce-base", NULL);
546         sizep = of_get_property(node, "linux,tce-size", NULL);
547         if (basep == NULL || sizep == NULL) {
548                 printk(KERN_ERR "PCI_DMA: iommu_table_setparms: %s has "
549                                 "missing tce entries !\n", dn->full_name);
550                 return;
551         }
552
553         tbl->it_base = (unsigned long)__va(*basep);
554
555         if (!is_kdump_kernel())
556                 memset((void *)tbl->it_base, 0, *sizep);
557
558         tbl->it_busno = phb->bus->number;
559         tbl->it_page_shift = IOMMU_PAGE_SHIFT_4K;
560
561         /* Units of tce entries */
562         tbl->it_offset = phb->dma_window_base_cur >> tbl->it_page_shift;
563
564         /* Test if we are going over 2GB of DMA space */
565         if (phb->dma_window_base_cur + phb->dma_window_size > 0x80000000ul) {
566                 udbg_printf("PCI_DMA: Unexpected number of IOAs under this PHB.\n");
567                 panic("PCI_DMA: Unexpected number of IOAs under this PHB.\n");
568         }
569
570         phb->dma_window_base_cur += phb->dma_window_size;
571
572         /* Set the tce table size - measured in entries */
573         tbl->it_size = phb->dma_window_size >> tbl->it_page_shift;
574
575         tbl->it_index = 0;
576         tbl->it_blocksize = 16;
577         tbl->it_type = TCE_PCI;
578
579         sw_inval = of_get_property(node, "linux,tce-sw-invalidate-info", NULL);
580         if (sw_inval) {
581                 /*
582                  * This property contains information on how to
583                  * invalidate the TCE entry.  The first property is
584                  * the base MMIO address used to invalidate entries.
585                  * The second property tells us the format of the TCE
586                  * invalidate (whether it needs to be shifted) and
587                  * some magic routing info to add to our invalidate
588                  * command.
589                  */
590                 tbl->it_index = (unsigned long) ioremap(sw_inval[0], 8);
591                 tbl->it_busno = sw_inval[1]; /* overload this with magic */
592                 tbl->it_type = TCE_PCI_SWINV_CREATE | TCE_PCI_SWINV_FREE;
593         }
594 }
595
596 /*
597  * iommu_table_setparms_lpar
598  *
599  * Function: On pSeries LPAR systems, return TCE table info, given a pci bus.
600  */
601 static void iommu_table_setparms_lpar(struct pci_controller *phb,
602                                       struct device_node *dn,
603                                       struct iommu_table *tbl,
604                                       const __be32 *dma_window)
605 {
606         unsigned long offset, size;
607
608         of_parse_dma_window(dn, dma_window, &tbl->it_index, &offset, &size);
609
610         tbl->it_busno = phb->bus->number;
611         tbl->it_page_shift = IOMMU_PAGE_SHIFT_4K;
612         tbl->it_base   = 0;
613         tbl->it_blocksize  = 16;
614         tbl->it_type = TCE_PCI;
615         tbl->it_offset = offset >> tbl->it_page_shift;
616         tbl->it_size = size >> tbl->it_page_shift;
617 }
618
619 struct iommu_table_ops iommu_table_pseries_ops = {
620         .set = tce_build_pSeries,
621         .clear = tce_free_pSeries,
622         .get = tce_get_pseries
623 };
624
625 static void pci_dma_bus_setup_pSeries(struct pci_bus *bus)
626 {
627         struct device_node *dn;
628         struct iommu_table *tbl;
629         struct device_node *isa_dn, *isa_dn_orig;
630         struct device_node *tmp;
631         struct pci_dn *pci;
632         int children;
633
634         dn = pci_bus_to_OF_node(bus);
635
636         pr_debug("pci_dma_bus_setup_pSeries: setting up bus %s\n", dn->full_name);
637
638         if (bus->self) {
639                 /* This is not a root bus, any setup will be done for the
640                  * device-side of the bridge in iommu_dev_setup_pSeries().
641                  */
642                 return;
643         }
644         pci = PCI_DN(dn);
645
646         /* Check if the ISA bus on the system is under
647          * this PHB.
648          */
649         isa_dn = isa_dn_orig = of_find_node_by_type(NULL, "isa");
650
651         while (isa_dn && isa_dn != dn)
652                 isa_dn = isa_dn->parent;
653
654         of_node_put(isa_dn_orig);
655
656         /* Count number of direct PCI children of the PHB. */
657         for (children = 0, tmp = dn->child; tmp; tmp = tmp->sibling)
658                 children++;
659
660         pr_debug("Children: %d\n", children);
661
662         /* Calculate amount of DMA window per slot. Each window must be
663          * a power of two (due to pci_alloc_consistent requirements).
664          *
665          * Keep 256MB aside for PHBs with ISA.
666          */
667
668         if (!isa_dn) {
669                 /* No ISA/IDE - just set window size and return */
670                 pci->phb->dma_window_size = 0x80000000ul; /* To be divided */
671
672                 while (pci->phb->dma_window_size * children > 0x80000000ul)
673                         pci->phb->dma_window_size >>= 1;
674                 pr_debug("No ISA/IDE, window size is 0x%llx\n",
675                          pci->phb->dma_window_size);
676                 pci->phb->dma_window_base_cur = 0;
677
678                 return;
679         }
680
681         /* If we have ISA, then we probably have an IDE
682          * controller too. Allocate a 128MB table but
683          * skip the first 128MB to avoid stepping on ISA
684          * space.
685          */
686         pci->phb->dma_window_size = 0x8000000ul;
687         pci->phb->dma_window_base_cur = 0x8000000ul;
688
689         pci->table_group = iommu_pseries_alloc_group(pci->phb->node);
690         tbl = pci->table_group->tables[0];
691
692         iommu_table_setparms(pci->phb, dn, tbl);
693         tbl->it_ops = &iommu_table_pseries_ops;
694         iommu_init_table(tbl, pci->phb->node);
695         iommu_register_group(pci->table_group, pci_domain_nr(bus), 0);
696
697         /* Divide the rest (1.75GB) among the children */
698         pci->phb->dma_window_size = 0x80000000ul;
699         while (pci->phb->dma_window_size * children > 0x70000000ul)
700                 pci->phb->dma_window_size >>= 1;
701
702         pr_debug("ISA/IDE, window size is 0x%llx\n", pci->phb->dma_window_size);
703 }
704
705 struct iommu_table_ops iommu_table_lpar_multi_ops = {
706         .set = tce_buildmulti_pSeriesLP,
707         .clear = tce_freemulti_pSeriesLP,
708         .get = tce_get_pSeriesLP
709 };
710
711 static void pci_dma_bus_setup_pSeriesLP(struct pci_bus *bus)
712 {
713         struct iommu_table *tbl;
714         struct device_node *dn, *pdn;
715         struct pci_dn *ppci;
716         const __be32 *dma_window = NULL;
717
718         dn = pci_bus_to_OF_node(bus);
719
720         pr_debug("pci_dma_bus_setup_pSeriesLP: setting up bus %s\n",
721                  dn->full_name);
722
723         /* Find nearest ibm,dma-window, walking up the device tree */
724         for (pdn = dn; pdn != NULL; pdn = pdn->parent) {
725                 dma_window = of_get_property(pdn, "ibm,dma-window", NULL);
726                 if (dma_window != NULL)
727                         break;
728         }
729
730         if (dma_window == NULL) {
731                 pr_debug("  no ibm,dma-window property !\n");
732                 return;
733         }
734
735         ppci = PCI_DN(pdn);
736
737         pr_debug("  parent is %s, iommu_table: 0x%p\n",
738                  pdn->full_name, ppci->table_group);
739
740         if (!ppci->table_group) {
741                 ppci->table_group = iommu_pseries_alloc_group(ppci->phb->node);
742                 tbl = ppci->table_group->tables[0];
743                 iommu_table_setparms_lpar(ppci->phb, pdn, tbl, dma_window);
744                 tbl->it_ops = &iommu_table_lpar_multi_ops;
745                 iommu_init_table(tbl, ppci->phb->node);
746                 iommu_register_group(ppci->table_group,
747                                 pci_domain_nr(bus), 0);
748                 pr_debug("  created table: %p\n", ppci->table_group);
749         }
750 }
751
752
753 static void pci_dma_dev_setup_pSeries(struct pci_dev *dev)
754 {
755         struct device_node *dn;
756         struct iommu_table *tbl;
757
758         pr_debug("pci_dma_dev_setup_pSeries: %s\n", pci_name(dev));
759
760         dn = dev->dev.of_node;
761
762         /* If we're the direct child of a root bus, then we need to allocate
763          * an iommu table ourselves. The bus setup code should have setup
764          * the window sizes already.
765          */
766         if (!dev->bus->self) {
767                 struct pci_controller *phb = PCI_DN(dn)->phb;
768
769                 pr_debug(" --> first child, no bridge. Allocating iommu table.\n");
770                 PCI_DN(dn)->table_group = iommu_pseries_alloc_group(phb->node);
771                 tbl = PCI_DN(dn)->table_group->tables[0];
772                 iommu_table_setparms(phb, dn, tbl);
773                 tbl->it_ops = &iommu_table_pseries_ops;
774                 iommu_init_table(tbl, phb->node);
775                 iommu_register_group(PCI_DN(dn)->table_group,
776                                 pci_domain_nr(phb->bus), 0);
777                 set_iommu_table_base(&dev->dev, tbl);
778                 iommu_add_device(&dev->dev);
779                 return;
780         }
781
782         /* If this device is further down the bus tree, search upwards until
783          * an already allocated iommu table is found and use that.
784          */
785
786         while (dn && PCI_DN(dn) && PCI_DN(dn)->table_group == NULL)
787                 dn = dn->parent;
788
789         if (dn && PCI_DN(dn)) {
790                 set_iommu_table_base(&dev->dev,
791                                 PCI_DN(dn)->table_group->tables[0]);
792                 iommu_add_device(&dev->dev);
793         } else
794                 printk(KERN_WARNING "iommu: Device %s has no iommu table\n",
795                        pci_name(dev));
796 }
797
798 static int __read_mostly disable_ddw;
799
800 static int __init disable_ddw_setup(char *str)
801 {
802         disable_ddw = 1;
803         printk(KERN_INFO "ppc iommu: disabling ddw.\n");
804
805         return 0;
806 }
807
808 early_param("disable_ddw", disable_ddw_setup);
809
810 static void remove_ddw(struct device_node *np, bool remove_prop)
811 {
812         struct dynamic_dma_window_prop *dwp;
813         struct property *win64;
814         u32 ddw_avail[3];
815         u64 liobn;
816         int ret = 0;
817
818         ret = of_property_read_u32_array(np, "ibm,ddw-applicable",
819                                          &ddw_avail[0], 3);
820
821         win64 = of_find_property(np, DIRECT64_PROPNAME, NULL);
822         if (!win64)
823                 return;
824
825         if (ret || win64->length < sizeof(*dwp))
826                 goto delprop;
827
828         dwp = win64->value;
829         liobn = (u64)be32_to_cpu(dwp->liobn);
830
831         /* clear the whole window, note the arg is in kernel pages */
832         ret = tce_clearrange_multi_pSeriesLP(0,
833                 1ULL << (be32_to_cpu(dwp->window_shift) - PAGE_SHIFT), dwp);
834         if (ret)
835                 pr_warning("%s failed to clear tces in window.\n",
836                          np->full_name);
837         else
838                 pr_debug("%s successfully cleared tces in window.\n",
839                          np->full_name);
840
841         ret = rtas_call(ddw_avail[2], 1, 1, NULL, liobn);
842         if (ret)
843                 pr_warning("%s: failed to remove direct window: rtas returned "
844                         "%d to ibm,remove-pe-dma-window(%x) %llx\n",
845                         np->full_name, ret, ddw_avail[2], liobn);
846         else
847                 pr_debug("%s: successfully removed direct window: rtas returned "
848                         "%d to ibm,remove-pe-dma-window(%x) %llx\n",
849                         np->full_name, ret, ddw_avail[2], liobn);
850
851 delprop:
852         if (remove_prop)
853                 ret = of_remove_property(np, win64);
854         if (ret)
855                 pr_warning("%s: failed to remove direct window property: %d\n",
856                         np->full_name, ret);
857 }
858
859 static u64 find_existing_ddw(struct device_node *pdn)
860 {
861         struct direct_window *window;
862         const struct dynamic_dma_window_prop *direct64;
863         u64 dma_addr = 0;
864
865         spin_lock(&direct_window_list_lock);
866         /* check if we already created a window and dupe that config if so */
867         list_for_each_entry(window, &direct_window_list, list) {
868                 if (window->device == pdn) {
869                         direct64 = window->prop;
870                         dma_addr = be64_to_cpu(direct64->dma_base);
871                         break;
872                 }
873         }
874         spin_unlock(&direct_window_list_lock);
875
876         return dma_addr;
877 }
878
879 static int find_existing_ddw_windows(void)
880 {
881         int len;
882         struct device_node *pdn;
883         struct direct_window *window;
884         const struct dynamic_dma_window_prop *direct64;
885
886         if (!firmware_has_feature(FW_FEATURE_LPAR))
887                 return 0;
888
889         for_each_node_with_property(pdn, DIRECT64_PROPNAME) {
890                 direct64 = of_get_property(pdn, DIRECT64_PROPNAME, &len);
891                 if (!direct64)
892                         continue;
893
894                 window = kzalloc(sizeof(*window), GFP_KERNEL);
895                 if (!window || len < sizeof(struct dynamic_dma_window_prop)) {
896                         kfree(window);
897                         remove_ddw(pdn, true);
898                         continue;
899                 }
900
901                 window->device = pdn;
902                 window->prop = direct64;
903                 spin_lock(&direct_window_list_lock);
904                 list_add(&window->list, &direct_window_list);
905                 spin_unlock(&direct_window_list_lock);
906         }
907
908         return 0;
909 }
910 machine_arch_initcall(pseries, find_existing_ddw_windows);
911
912 static int query_ddw(struct pci_dev *dev, const u32 *ddw_avail,
913                         struct ddw_query_response *query)
914 {
915         struct device_node *dn;
916         struct pci_dn *pdn;
917         u32 cfg_addr;
918         u64 buid;
919         int ret;
920
921         /*
922          * Get the config address and phb buid of the PE window.
923          * Rely on eeh to retrieve this for us.
924          * Retrieve them from the pci device, not the node with the
925          * dma-window property
926          */
927         dn = pci_device_to_OF_node(dev);
928         pdn = PCI_DN(dn);
929         buid = pdn->phb->buid;
930         cfg_addr = ((pdn->busno << 16) | (pdn->devfn << 8));
931
932         ret = rtas_call(ddw_avail[0], 3, 5, (u32 *)query,
933                   cfg_addr, BUID_HI(buid), BUID_LO(buid));
934         dev_info(&dev->dev, "ibm,query-pe-dma-windows(%x) %x %x %x"
935                 " returned %d\n", ddw_avail[0], cfg_addr, BUID_HI(buid),
936                 BUID_LO(buid), ret);
937         return ret;
938 }
939
940 static int create_ddw(struct pci_dev *dev, const u32 *ddw_avail,
941                         struct ddw_create_response *create, int page_shift,
942                         int window_shift)
943 {
944         struct device_node *dn;
945         struct pci_dn *pdn;
946         u32 cfg_addr;
947         u64 buid;
948         int ret;
949
950         /*
951          * Get the config address and phb buid of the PE window.
952          * Rely on eeh to retrieve this for us.
953          * Retrieve them from the pci device, not the node with the
954          * dma-window property
955          */
956         dn = pci_device_to_OF_node(dev);
957         pdn = PCI_DN(dn);
958         buid = pdn->phb->buid;
959         cfg_addr = ((pdn->busno << 16) | (pdn->devfn << 8));
960
961         do {
962                 /* extra outputs are LIOBN and dma-addr (hi, lo) */
963                 ret = rtas_call(ddw_avail[1], 5, 4, (u32 *)create,
964                                 cfg_addr, BUID_HI(buid), BUID_LO(buid),
965                                 page_shift, window_shift);
966         } while (rtas_busy_delay(ret));
967         dev_info(&dev->dev,
968                 "ibm,create-pe-dma-window(%x) %x %x %x %x %x returned %d "
969                 "(liobn = 0x%x starting addr = %x %x)\n", ddw_avail[1],
970                  cfg_addr, BUID_HI(buid), BUID_LO(buid), page_shift,
971                  window_shift, ret, create->liobn, create->addr_hi, create->addr_lo);
972
973         return ret;
974 }
975
976 struct failed_ddw_pdn {
977         struct device_node *pdn;
978         struct list_head list;
979 };
980
981 static LIST_HEAD(failed_ddw_pdn_list);
982
983 /*
984  * If the PE supports dynamic dma windows, and there is space for a table
985  * that can map all pages in a linear offset, then setup such a table,
986  * and record the dma-offset in the struct device.
987  *
988  * dev: the pci device we are checking
989  * pdn: the parent pe node with the ibm,dma_window property
990  * Future: also check if we can remap the base window for our base page size
991  *
992  * returns the dma offset for use by dma_set_mask
993  */
994 static u64 enable_ddw(struct pci_dev *dev, struct device_node *pdn)
995 {
996         int len, ret;
997         struct ddw_query_response query;
998         struct ddw_create_response create;
999         int page_shift;
1000         u64 dma_addr, max_addr;
1001         struct device_node *dn;
1002         u32 ddw_avail[3];
1003         struct direct_window *window;
1004         struct property *win64;
1005         struct dynamic_dma_window_prop *ddwprop;
1006         struct failed_ddw_pdn *fpdn;
1007
1008         mutex_lock(&direct_window_init_mutex);
1009
1010         dma_addr = find_existing_ddw(pdn);
1011         if (dma_addr != 0)
1012                 goto out_unlock;
1013
1014         /*
1015          * If we already went through this for a previous function of
1016          * the same device and failed, we don't want to muck with the
1017          * DMA window again, as it will race with in-flight operations
1018          * and can lead to EEHs. The above mutex protects access to the
1019          * list.
1020          */
1021         list_for_each_entry(fpdn, &failed_ddw_pdn_list, list) {
1022                 if (!strcmp(fpdn->pdn->full_name, pdn->full_name))
1023                         goto out_unlock;
1024         }
1025
1026         /*
1027          * the ibm,ddw-applicable property holds the tokens for:
1028          * ibm,query-pe-dma-window
1029          * ibm,create-pe-dma-window
1030          * ibm,remove-pe-dma-window
1031          * for the given node in that order.
1032          * the property is actually in the parent, not the PE
1033          */
1034         ret = of_property_read_u32_array(pdn, "ibm,ddw-applicable",
1035                                          &ddw_avail[0], 3);
1036         if (ret)
1037                 goto out_failed;
1038
1039        /*
1040          * Query if there is a second window of size to map the
1041          * whole partition.  Query returns number of windows, largest
1042          * block assigned to PE (partition endpoint), and two bitmasks
1043          * of page sizes: supported and supported for migrate-dma.
1044          */
1045         dn = pci_device_to_OF_node(dev);
1046         ret = query_ddw(dev, ddw_avail, &query);
1047         if (ret != 0)
1048                 goto out_failed;
1049
1050         if (query.windows_available == 0) {
1051                 /*
1052                  * no additional windows are available for this device.
1053                  * We might be able to reallocate the existing window,
1054                  * trading in for a larger page size.
1055                  */
1056                 dev_dbg(&dev->dev, "no free dynamic windows");
1057                 goto out_failed;
1058         }
1059         if (query.page_size & 4) {
1060                 page_shift = 24; /* 16MB */
1061         } else if (query.page_size & 2) {
1062                 page_shift = 16; /* 64kB */
1063         } else if (query.page_size & 1) {
1064                 page_shift = 12; /* 4kB */
1065         } else {
1066                 dev_dbg(&dev->dev, "no supported direct page size in mask %x",
1067                           query.page_size);
1068                 goto out_failed;
1069         }
1070         /* verify the window * number of ptes will map the partition */
1071         /* check largest block * page size > max memory hotplug addr */
1072         max_addr = memory_hotplug_max();
1073         if (query.largest_available_block < (max_addr >> page_shift)) {
1074                 dev_dbg(&dev->dev, "can't map partiton max 0x%llx with %u "
1075                           "%llu-sized pages\n", max_addr,  query.largest_available_block,
1076                           1ULL << page_shift);
1077                 goto out_failed;
1078         }
1079         len = order_base_2(max_addr);
1080         win64 = kzalloc(sizeof(struct property), GFP_KERNEL);
1081         if (!win64) {
1082                 dev_info(&dev->dev,
1083                         "couldn't allocate property for 64bit dma window\n");
1084                 goto out_failed;
1085         }
1086         win64->name = kstrdup(DIRECT64_PROPNAME, GFP_KERNEL);
1087         win64->value = ddwprop = kmalloc(sizeof(*ddwprop), GFP_KERNEL);
1088         win64->length = sizeof(*ddwprop);
1089         if (!win64->name || !win64->value) {
1090                 dev_info(&dev->dev,
1091                         "couldn't allocate property name and value\n");
1092                 goto out_free_prop;
1093         }
1094
1095         ret = create_ddw(dev, ddw_avail, &create, page_shift, len);
1096         if (ret != 0)
1097                 goto out_free_prop;
1098
1099         ddwprop->liobn = cpu_to_be32(create.liobn);
1100         ddwprop->dma_base = cpu_to_be64(((u64)create.addr_hi << 32) |
1101                         create.addr_lo);
1102         ddwprop->tce_shift = cpu_to_be32(page_shift);
1103         ddwprop->window_shift = cpu_to_be32(len);
1104
1105         dev_dbg(&dev->dev, "created tce table LIOBN 0x%x for %s\n",
1106                   create.liobn, dn->full_name);
1107
1108         window = kzalloc(sizeof(*window), GFP_KERNEL);
1109         if (!window)
1110                 goto out_clear_window;
1111
1112         ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> PAGE_SHIFT,
1113                         win64->value, tce_setrange_multi_pSeriesLP_walk);
1114         if (ret) {
1115                 dev_info(&dev->dev, "failed to map direct window for %s: %d\n",
1116                          dn->full_name, ret);
1117                 goto out_free_window;
1118         }
1119
1120         ret = of_add_property(pdn, win64);
1121         if (ret) {
1122                 dev_err(&dev->dev, "unable to add dma window property for %s: %d",
1123                          pdn->full_name, ret);
1124                 goto out_free_window;
1125         }
1126
1127         window->device = pdn;
1128         window->prop = ddwprop;
1129         spin_lock(&direct_window_list_lock);
1130         list_add(&window->list, &direct_window_list);
1131         spin_unlock(&direct_window_list_lock);
1132
1133         dma_addr = be64_to_cpu(ddwprop->dma_base);
1134         goto out_unlock;
1135
1136 out_free_window:
1137         kfree(window);
1138
1139 out_clear_window:
1140         remove_ddw(pdn, true);
1141
1142 out_free_prop:
1143         kfree(win64->name);
1144         kfree(win64->value);
1145         kfree(win64);
1146
1147 out_failed:
1148
1149         fpdn = kzalloc(sizeof(*fpdn), GFP_KERNEL);
1150         if (!fpdn)
1151                 goto out_unlock;
1152         fpdn->pdn = pdn;
1153         list_add(&fpdn->list, &failed_ddw_pdn_list);
1154
1155 out_unlock:
1156         mutex_unlock(&direct_window_init_mutex);
1157         return dma_addr;
1158 }
1159
1160 static void pci_dma_dev_setup_pSeriesLP(struct pci_dev *dev)
1161 {
1162         struct device_node *pdn, *dn;
1163         struct iommu_table *tbl;
1164         const __be32 *dma_window = NULL;
1165         struct pci_dn *pci;
1166
1167         pr_debug("pci_dma_dev_setup_pSeriesLP: %s\n", pci_name(dev));
1168
1169         /* dev setup for LPAR is a little tricky, since the device tree might
1170          * contain the dma-window properties per-device and not necessarily
1171          * for the bus. So we need to search upwards in the tree until we
1172          * either hit a dma-window property, OR find a parent with a table
1173          * already allocated.
1174          */
1175         dn = pci_device_to_OF_node(dev);
1176         pr_debug("  node is %s\n", dn->full_name);
1177
1178         for (pdn = dn; pdn && PCI_DN(pdn) && !PCI_DN(pdn)->table_group;
1179              pdn = pdn->parent) {
1180                 dma_window = of_get_property(pdn, "ibm,dma-window", NULL);
1181                 if (dma_window)
1182                         break;
1183         }
1184
1185         if (!pdn || !PCI_DN(pdn)) {
1186                 printk(KERN_WARNING "pci_dma_dev_setup_pSeriesLP: "
1187                        "no DMA window found for pci dev=%s dn=%s\n",
1188                                  pci_name(dev), of_node_full_name(dn));
1189                 return;
1190         }
1191         pr_debug("  parent is %s\n", pdn->full_name);
1192
1193         pci = PCI_DN(pdn);
1194         if (!pci->table_group) {
1195                 pci->table_group = iommu_pseries_alloc_group(pci->phb->node);
1196                 tbl = pci->table_group->tables[0];
1197                 iommu_table_setparms_lpar(pci->phb, pdn, tbl, dma_window);
1198                 tbl->it_ops = &iommu_table_lpar_multi_ops;
1199                 iommu_init_table(tbl, pci->phb->node);
1200                 iommu_register_group(pci->table_group,
1201                                 pci_domain_nr(pci->phb->bus), 0);
1202                 pr_debug("  created table: %p\n", pci->table_group);
1203         } else {
1204                 pr_debug("  found DMA window, table: %p\n", pci->table_group);
1205         }
1206
1207         set_iommu_table_base(&dev->dev, pci->table_group->tables[0]);
1208         iommu_add_device(&dev->dev);
1209 }
1210
1211 static int dma_set_mask_pSeriesLP(struct device *dev, u64 dma_mask)
1212 {
1213         bool ddw_enabled = false;
1214         struct device_node *pdn, *dn;
1215         struct pci_dev *pdev;
1216         const __be32 *dma_window = NULL;
1217         u64 dma_offset;
1218
1219         if (!dev->dma_mask)
1220                 return -EIO;
1221
1222         if (!dev_is_pci(dev))
1223                 goto check_mask;
1224
1225         pdev = to_pci_dev(dev);
1226
1227         /* only attempt to use a new window if 64-bit DMA is requested */
1228         if (!disable_ddw && dma_mask == DMA_BIT_MASK(64)) {
1229                 dn = pci_device_to_OF_node(pdev);
1230                 dev_dbg(dev, "node is %s\n", dn->full_name);
1231
1232                 /*
1233                  * the device tree might contain the dma-window properties
1234                  * per-device and not necessarily for the bus. So we need to
1235                  * search upwards in the tree until we either hit a dma-window
1236                  * property, OR find a parent with a table already allocated.
1237                  */
1238                 for (pdn = dn; pdn && PCI_DN(pdn) && !PCI_DN(pdn)->table_group;
1239                                 pdn = pdn->parent) {
1240                         dma_window = of_get_property(pdn, "ibm,dma-window", NULL);
1241                         if (dma_window)
1242                                 break;
1243                 }
1244                 if (pdn && PCI_DN(pdn)) {
1245                         dma_offset = enable_ddw(pdev, pdn);
1246                         if (dma_offset != 0) {
1247                                 dev_info(dev, "Using 64-bit direct DMA at offset %llx\n", dma_offset);
1248                                 set_dma_offset(dev, dma_offset);
1249                                 set_dma_ops(dev, &dma_direct_ops);
1250                                 ddw_enabled = true;
1251                         }
1252                 }
1253         }
1254
1255         /* fall back on iommu ops */
1256         if (!ddw_enabled && get_dma_ops(dev) != &dma_iommu_ops) {
1257                 dev_info(dev, "Restoring 32-bit DMA via iommu\n");
1258                 set_dma_ops(dev, &dma_iommu_ops);
1259         }
1260
1261 check_mask:
1262         if (!dma_supported(dev, dma_mask))
1263                 return -EIO;
1264
1265         *dev->dma_mask = dma_mask;
1266         return 0;
1267 }
1268
1269 static u64 dma_get_required_mask_pSeriesLP(struct device *dev)
1270 {
1271         if (!dev->dma_mask)
1272                 return 0;
1273
1274         if (!disable_ddw && dev_is_pci(dev)) {
1275                 struct pci_dev *pdev = to_pci_dev(dev);
1276                 struct device_node *dn;
1277
1278                 dn = pci_device_to_OF_node(pdev);
1279
1280                 /* search upwards for ibm,dma-window */
1281                 for (; dn && PCI_DN(dn) && !PCI_DN(dn)->table_group;
1282                                 dn = dn->parent)
1283                         if (of_get_property(dn, "ibm,dma-window", NULL))
1284                                 break;
1285                 /* if there is a ibm,ddw-applicable property require 64 bits */
1286                 if (dn && PCI_DN(dn) &&
1287                                 of_get_property(dn, "ibm,ddw-applicable", NULL))
1288                         return DMA_BIT_MASK(64);
1289         }
1290
1291         return dma_iommu_ops.get_required_mask(dev);
1292 }
1293
1294 static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
1295                 void *data)
1296 {
1297         struct direct_window *window;
1298         struct memory_notify *arg = data;
1299         int ret = 0;
1300
1301         switch (action) {
1302         case MEM_GOING_ONLINE:
1303                 spin_lock(&direct_window_list_lock);
1304                 list_for_each_entry(window, &direct_window_list, list) {
1305                         ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn,
1306                                         arg->nr_pages, window->prop);
1307                         /* XXX log error */
1308                 }
1309                 spin_unlock(&direct_window_list_lock);
1310                 break;
1311         case MEM_CANCEL_ONLINE:
1312         case MEM_OFFLINE:
1313                 spin_lock(&direct_window_list_lock);
1314                 list_for_each_entry(window, &direct_window_list, list) {
1315                         ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn,
1316                                         arg->nr_pages, window->prop);
1317                         /* XXX log error */
1318                 }
1319                 spin_unlock(&direct_window_list_lock);
1320                 break;
1321         default:
1322                 break;
1323         }
1324         if (ret && action != MEM_CANCEL_ONLINE)
1325                 return NOTIFY_BAD;
1326
1327         return NOTIFY_OK;
1328 }
1329
1330 static struct notifier_block iommu_mem_nb = {
1331         .notifier_call = iommu_mem_notifier,
1332 };
1333
1334 static int iommu_reconfig_notifier(struct notifier_block *nb, unsigned long action, void *data)
1335 {
1336         int err = NOTIFY_OK;
1337         struct of_reconfig_data *rd = data;
1338         struct device_node *np = rd->dn;
1339         struct pci_dn *pci = PCI_DN(np);
1340         struct direct_window *window;
1341
1342         switch (action) {
1343         case OF_RECONFIG_DETACH_NODE:
1344                 /*
1345                  * Removing the property will invoke the reconfig
1346                  * notifier again, which causes dead-lock on the
1347                  * read-write semaphore of the notifier chain. So
1348                  * we have to remove the property when releasing
1349                  * the device node.
1350                  */
1351                 remove_ddw(np, false);
1352                 if (pci && pci->table_group)
1353                         iommu_pseries_free_group(pci->table_group,
1354                                         np->full_name);
1355
1356                 spin_lock(&direct_window_list_lock);
1357                 list_for_each_entry(window, &direct_window_list, list) {
1358                         if (window->device == np) {
1359                                 list_del(&window->list);
1360                                 kfree(window);
1361                                 break;
1362                         }
1363                 }
1364                 spin_unlock(&direct_window_list_lock);
1365                 break;
1366         default:
1367                 err = NOTIFY_DONE;
1368                 break;
1369         }
1370         return err;
1371 }
1372
1373 static struct notifier_block iommu_reconfig_nb = {
1374         .notifier_call = iommu_reconfig_notifier,
1375 };
1376
1377 /* These are called very early. */
1378 void iommu_init_early_pSeries(void)
1379 {
1380         if (of_chosen && of_get_property(of_chosen, "linux,iommu-off", NULL))
1381                 return;
1382
1383         if (firmware_has_feature(FW_FEATURE_LPAR)) {
1384                 pseries_pci_controller_ops.dma_bus_setup = pci_dma_bus_setup_pSeriesLP;
1385                 pseries_pci_controller_ops.dma_dev_setup = pci_dma_dev_setup_pSeriesLP;
1386                 ppc_md.dma_set_mask = dma_set_mask_pSeriesLP;
1387                 ppc_md.dma_get_required_mask = dma_get_required_mask_pSeriesLP;
1388         } else {
1389                 pseries_pci_controller_ops.dma_bus_setup = pci_dma_bus_setup_pSeries;
1390                 pseries_pci_controller_ops.dma_dev_setup = pci_dma_dev_setup_pSeries;
1391         }
1392
1393
1394         of_reconfig_notifier_register(&iommu_reconfig_nb);
1395         register_memory_notifier(&iommu_mem_nb);
1396
1397         set_pci_dma_ops(&dma_iommu_ops);
1398 }
1399
1400 static int __init disable_multitce(char *str)
1401 {
1402         if (strcmp(str, "off") == 0 &&
1403             firmware_has_feature(FW_FEATURE_LPAR) &&
1404             firmware_has_feature(FW_FEATURE_MULTITCE)) {
1405                 printk(KERN_INFO "Disabling MULTITCE firmware feature\n");
1406                 powerpc_firmware_features &= ~FW_FEATURE_MULTITCE;
1407         }
1408         return 1;
1409 }
1410
1411 __setup("multitce=", disable_multitce);
1412
1413 machine_subsys_initcall_sync(pseries, tce_iommu_bus_notifier_init);