GNU Linux-libre 5.10.217-gnu1
[releases.git] / arch / powerpc / platforms / pseries / iommu.c
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Copyright (C) 2001 Mike Corrigan & Dave Engebretsen, IBM Corporation
4  *
5  * Rewrite, cleanup:
6  *
7  * Copyright (C) 2004 Olof Johansson <olof@lixom.net>, IBM Corporation
8  * Copyright (C) 2006 Olof Johansson <olof@lixom.net>
9  *
10  * Dynamic DMA mapping support, pSeries-specific parts, both SMP and LPAR.
11  */
12
13 #include <linux/init.h>
14 #include <linux/types.h>
15 #include <linux/slab.h>
16 #include <linux/mm.h>
17 #include <linux/memblock.h>
18 #include <linux/spinlock.h>
19 #include <linux/string.h>
20 #include <linux/pci.h>
21 #include <linux/dma-mapping.h>
22 #include <linux/crash_dump.h>
23 #include <linux/memory.h>
24 #include <linux/of.h>
25 #include <linux/iommu.h>
26 #include <linux/rculist.h>
27 #include <asm/io.h>
28 #include <asm/prom.h>
29 #include <asm/rtas.h>
30 #include <asm/iommu.h>
31 #include <asm/pci-bridge.h>
32 #include <asm/machdep.h>
33 #include <asm/firmware.h>
34 #include <asm/tce.h>
35 #include <asm/ppc-pci.h>
36 #include <asm/udbg.h>
37 #include <asm/mmzone.h>
38 #include <asm/plpar_wrappers.h>
39
40 #include "pseries.h"
41
42 enum {
43         DDW_QUERY_PE_DMA_WIN  = 0,
44         DDW_CREATE_PE_DMA_WIN = 1,
45         DDW_REMOVE_PE_DMA_WIN = 2,
46
47         DDW_APPLICABLE_SIZE
48 };
49
50 enum {
51         DDW_EXT_SIZE = 0,
52         DDW_EXT_RESET_DMA_WIN = 1,
53         DDW_EXT_QUERY_OUT_SIZE = 2
54 };
55
56 static struct iommu_table_group *iommu_pseries_alloc_group(int node)
57 {
58         struct iommu_table_group *table_group;
59         struct iommu_table *tbl;
60
61         table_group = kzalloc_node(sizeof(struct iommu_table_group), GFP_KERNEL,
62                            node);
63         if (!table_group)
64                 return NULL;
65
66         tbl = kzalloc_node(sizeof(struct iommu_table), GFP_KERNEL, node);
67         if (!tbl)
68                 goto free_group;
69
70         INIT_LIST_HEAD_RCU(&tbl->it_group_list);
71         kref_init(&tbl->it_kref);
72
73         table_group->tables[0] = tbl;
74
75         return table_group;
76
77 free_group:
78         kfree(table_group);
79         return NULL;
80 }
81
82 static void iommu_pseries_free_group(struct iommu_table_group *table_group,
83                 const char *node_name)
84 {
85         struct iommu_table *tbl;
86
87         if (!table_group)
88                 return;
89
90         tbl = table_group->tables[0];
91 #ifdef CONFIG_IOMMU_API
92         if (table_group->group) {
93                 iommu_group_put(table_group->group);
94                 BUG_ON(table_group->group);
95         }
96 #endif
97         iommu_tce_table_put(tbl);
98
99         kfree(table_group);
100 }
101
102 static int tce_build_pSeries(struct iommu_table *tbl, long index,
103                               long npages, unsigned long uaddr,
104                               enum dma_data_direction direction,
105                               unsigned long attrs)
106 {
107         u64 proto_tce;
108         __be64 *tcep;
109         u64 rpn;
110
111         proto_tce = TCE_PCI_READ; // Read allowed
112
113         if (direction != DMA_TO_DEVICE)
114                 proto_tce |= TCE_PCI_WRITE;
115
116         tcep = ((__be64 *)tbl->it_base) + index;
117
118         while (npages--) {
119                 /* can't move this out since we might cross MEMBLOCK boundary */
120                 rpn = __pa(uaddr) >> TCE_SHIFT;
121                 *tcep = cpu_to_be64(proto_tce | (rpn & TCE_RPN_MASK) << TCE_RPN_SHIFT);
122
123                 uaddr += TCE_PAGE_SIZE;
124                 tcep++;
125         }
126         return 0;
127 }
128
129
130 static void tce_free_pSeries(struct iommu_table *tbl, long index, long npages)
131 {
132         __be64 *tcep;
133
134         tcep = ((__be64 *)tbl->it_base) + index;
135
136         while (npages--)
137                 *(tcep++) = 0;
138 }
139
140 static unsigned long tce_get_pseries(struct iommu_table *tbl, long index)
141 {
142         __be64 *tcep;
143
144         tcep = ((__be64 *)tbl->it_base) + index;
145
146         return be64_to_cpu(*tcep);
147 }
148
149 static void tce_free_pSeriesLP(unsigned long liobn, long, long);
150 static void tce_freemulti_pSeriesLP(struct iommu_table*, long, long);
151
152 static int tce_build_pSeriesLP(unsigned long liobn, long tcenum, long tceshift,
153                                 long npages, unsigned long uaddr,
154                                 enum dma_data_direction direction,
155                                 unsigned long attrs)
156 {
157         u64 rc = 0;
158         u64 proto_tce, tce;
159         u64 rpn;
160         int ret = 0;
161         long tcenum_start = tcenum, npages_start = npages;
162
163         rpn = __pa(uaddr) >> tceshift;
164         proto_tce = TCE_PCI_READ;
165         if (direction != DMA_TO_DEVICE)
166                 proto_tce |= TCE_PCI_WRITE;
167
168         while (npages--) {
169                 tce = proto_tce | (rpn & TCE_RPN_MASK) << tceshift;
170                 rc = plpar_tce_put((u64)liobn, (u64)tcenum << tceshift, tce);
171
172                 if (unlikely(rc == H_NOT_ENOUGH_RESOURCES)) {
173                         ret = (int)rc;
174                         tce_free_pSeriesLP(liobn, tcenum_start,
175                                            (npages_start - (npages + 1)));
176                         break;
177                 }
178
179                 if (rc && printk_ratelimit()) {
180                         printk("tce_build_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
181                         printk("\tindex   = 0x%llx\n", (u64)liobn);
182                         printk("\ttcenum  = 0x%llx\n", (u64)tcenum);
183                         printk("\ttce val = 0x%llx\n", tce );
184                         dump_stack();
185                 }
186
187                 tcenum++;
188                 rpn++;
189         }
190         return ret;
191 }
192
193 static DEFINE_PER_CPU(__be64 *, tce_page);
194
195 static int tce_buildmulti_pSeriesLP(struct iommu_table *tbl, long tcenum,
196                                      long npages, unsigned long uaddr,
197                                      enum dma_data_direction direction,
198                                      unsigned long attrs)
199 {
200         u64 rc = 0;
201         u64 proto_tce;
202         __be64 *tcep;
203         u64 rpn;
204         long l, limit;
205         long tcenum_start = tcenum, npages_start = npages;
206         int ret = 0;
207         unsigned long flags;
208
209         if ((npages == 1) || !firmware_has_feature(FW_FEATURE_PUT_TCE_IND)) {
210                 return tce_build_pSeriesLP(tbl->it_index, tcenum,
211                                            tbl->it_page_shift, npages, uaddr,
212                                            direction, attrs);
213         }
214
215         local_irq_save(flags);  /* to protect tcep and the page behind it */
216
217         tcep = __this_cpu_read(tce_page);
218
219         /* This is safe to do since interrupts are off when we're called
220          * from iommu_alloc{,_sg}()
221          */
222         if (!tcep) {
223                 tcep = (__be64 *)__get_free_page(GFP_ATOMIC);
224                 /* If allocation fails, fall back to the loop implementation */
225                 if (!tcep) {
226                         local_irq_restore(flags);
227                         return tce_build_pSeriesLP(tbl->it_index, tcenum,
228                                         tbl->it_page_shift,
229                                         npages, uaddr, direction, attrs);
230                 }
231                 __this_cpu_write(tce_page, tcep);
232         }
233
234         rpn = __pa(uaddr) >> TCE_SHIFT;
235         proto_tce = TCE_PCI_READ;
236         if (direction != DMA_TO_DEVICE)
237                 proto_tce |= TCE_PCI_WRITE;
238
239         /* We can map max one pageful of TCEs at a time */
240         do {
241                 /*
242                  * Set up the page with TCE data, looping through and setting
243                  * the values.
244                  */
245                 limit = min_t(long, npages, 4096/TCE_ENTRY_SIZE);
246
247                 for (l = 0; l < limit; l++) {
248                         tcep[l] = cpu_to_be64(proto_tce | (rpn & TCE_RPN_MASK) << TCE_RPN_SHIFT);
249                         rpn++;
250                 }
251
252                 rc = plpar_tce_put_indirect((u64)tbl->it_index,
253                                             (u64)tcenum << 12,
254                                             (u64)__pa(tcep),
255                                             limit);
256
257                 npages -= limit;
258                 tcenum += limit;
259         } while (npages > 0 && !rc);
260
261         local_irq_restore(flags);
262
263         if (unlikely(rc == H_NOT_ENOUGH_RESOURCES)) {
264                 ret = (int)rc;
265                 tce_freemulti_pSeriesLP(tbl, tcenum_start,
266                                         (npages_start - (npages + limit)));
267                 return ret;
268         }
269
270         if (rc && printk_ratelimit()) {
271                 printk("tce_buildmulti_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
272                 printk("\tindex   = 0x%llx\n", (u64)tbl->it_index);
273                 printk("\tnpages  = 0x%llx\n", (u64)npages);
274                 printk("\ttce[0] val = 0x%llx\n", tcep[0]);
275                 dump_stack();
276         }
277         return ret;
278 }
279
280 static void tce_free_pSeriesLP(unsigned long liobn, long tcenum, long npages)
281 {
282         u64 rc;
283
284         while (npages--) {
285                 rc = plpar_tce_put((u64)liobn, (u64)tcenum << 12, 0);
286
287                 if (rc && printk_ratelimit()) {
288                         printk("tce_free_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
289                         printk("\tindex   = 0x%llx\n", (u64)liobn);
290                         printk("\ttcenum  = 0x%llx\n", (u64)tcenum);
291                         dump_stack();
292                 }
293
294                 tcenum++;
295         }
296 }
297
298
299 static void tce_freemulti_pSeriesLP(struct iommu_table *tbl, long tcenum, long npages)
300 {
301         u64 rc;
302
303         if (!firmware_has_feature(FW_FEATURE_STUFF_TCE))
304                 return tce_free_pSeriesLP(tbl->it_index, tcenum, npages);
305
306         rc = plpar_tce_stuff((u64)tbl->it_index, (u64)tcenum << 12, 0, npages);
307
308         if (rc && printk_ratelimit()) {
309                 printk("tce_freemulti_pSeriesLP: plpar_tce_stuff failed\n");
310                 printk("\trc      = %lld\n", rc);
311                 printk("\tindex   = 0x%llx\n", (u64)tbl->it_index);
312                 printk("\tnpages  = 0x%llx\n", (u64)npages);
313                 dump_stack();
314         }
315 }
316
317 static unsigned long tce_get_pSeriesLP(struct iommu_table *tbl, long tcenum)
318 {
319         u64 rc;
320         unsigned long tce_ret;
321
322         rc = plpar_tce_get((u64)tbl->it_index, (u64)tcenum << 12, &tce_ret);
323
324         if (rc && printk_ratelimit()) {
325                 printk("tce_get_pSeriesLP: plpar_tce_get failed. rc=%lld\n", rc);
326                 printk("\tindex   = 0x%llx\n", (u64)tbl->it_index);
327                 printk("\ttcenum  = 0x%llx\n", (u64)tcenum);
328                 dump_stack();
329         }
330
331         return tce_ret;
332 }
333
334 /* this is compatible with cells for the device tree property */
335 struct dynamic_dma_window_prop {
336         __be32  liobn;          /* tce table number */
337         __be64  dma_base;       /* address hi,lo */
338         __be32  tce_shift;      /* ilog2(tce_page_size) */
339         __be32  window_shift;   /* ilog2(tce_window_size) */
340 };
341
342 struct direct_window {
343         struct device_node *device;
344         const struct dynamic_dma_window_prop *prop;
345         struct list_head list;
346 };
347
348 /* Dynamic DMA Window support */
349 struct ddw_query_response {
350         u32 windows_available;
351         u64 largest_available_block;
352         u32 page_size;
353         u32 migration_capable;
354 };
355
356 struct ddw_create_response {
357         u32 liobn;
358         u32 addr_hi;
359         u32 addr_lo;
360 };
361
362 static LIST_HEAD(direct_window_list);
363 /* prevents races between memory on/offline and window creation */
364 static DEFINE_SPINLOCK(direct_window_list_lock);
365 /* protects initializing window twice for same device */
366 static DEFINE_MUTEX(direct_window_init_mutex);
367 #define DIRECT64_PROPNAME "linux,direct64-ddr-window-info"
368
369 static int tce_clearrange_multi_pSeriesLP(unsigned long start_pfn,
370                                         unsigned long num_pfn, const void *arg)
371 {
372         const struct dynamic_dma_window_prop *maprange = arg;
373         int rc;
374         u64 tce_size, num_tce, dma_offset, next;
375         u32 tce_shift;
376         long limit;
377
378         tce_shift = be32_to_cpu(maprange->tce_shift);
379         tce_size = 1ULL << tce_shift;
380         next = start_pfn << PAGE_SHIFT;
381         num_tce = num_pfn << PAGE_SHIFT;
382
383         /* round back to the beginning of the tce page size */
384         num_tce += next & (tce_size - 1);
385         next &= ~(tce_size - 1);
386
387         /* covert to number of tces */
388         num_tce |= tce_size - 1;
389         num_tce >>= tce_shift;
390
391         do {
392                 /*
393                  * Set up the page with TCE data, looping through and setting
394                  * the values.
395                  */
396                 limit = min_t(long, num_tce, 512);
397                 dma_offset = next + be64_to_cpu(maprange->dma_base);
398
399                 rc = plpar_tce_stuff((u64)be32_to_cpu(maprange->liobn),
400                                              dma_offset,
401                                              0, limit);
402                 next += limit * tce_size;
403                 num_tce -= limit;
404         } while (num_tce > 0 && !rc);
405
406         return rc;
407 }
408
409 static int tce_setrange_multi_pSeriesLP(unsigned long start_pfn,
410                                         unsigned long num_pfn, const void *arg)
411 {
412         const struct dynamic_dma_window_prop *maprange = arg;
413         u64 tce_size, num_tce, dma_offset, next, proto_tce, liobn;
414         __be64 *tcep;
415         u32 tce_shift;
416         u64 rc = 0;
417         long l, limit;
418
419         if (!firmware_has_feature(FW_FEATURE_PUT_TCE_IND)) {
420                 unsigned long tceshift = be32_to_cpu(maprange->tce_shift);
421                 unsigned long dmastart = (start_pfn << PAGE_SHIFT) +
422                                 be64_to_cpu(maprange->dma_base);
423                 unsigned long tcenum = dmastart >> tceshift;
424                 unsigned long npages = num_pfn << PAGE_SHIFT >> tceshift;
425                 void *uaddr = __va(start_pfn << PAGE_SHIFT);
426
427                 return tce_build_pSeriesLP(be32_to_cpu(maprange->liobn),
428                                 tcenum, tceshift, npages, (unsigned long) uaddr,
429                                 DMA_BIDIRECTIONAL, 0);
430         }
431
432         local_irq_disable();    /* to protect tcep and the page behind it */
433         tcep = __this_cpu_read(tce_page);
434
435         if (!tcep) {
436                 tcep = (__be64 *)__get_free_page(GFP_ATOMIC);
437                 if (!tcep) {
438                         local_irq_enable();
439                         return -ENOMEM;
440                 }
441                 __this_cpu_write(tce_page, tcep);
442         }
443
444         proto_tce = TCE_PCI_READ | TCE_PCI_WRITE;
445
446         liobn = (u64)be32_to_cpu(maprange->liobn);
447         tce_shift = be32_to_cpu(maprange->tce_shift);
448         tce_size = 1ULL << tce_shift;
449         next = start_pfn << PAGE_SHIFT;
450         num_tce = num_pfn << PAGE_SHIFT;
451
452         /* round back to the beginning of the tce page size */
453         num_tce += next & (tce_size - 1);
454         next &= ~(tce_size - 1);
455
456         /* covert to number of tces */
457         num_tce |= tce_size - 1;
458         num_tce >>= tce_shift;
459
460         /* We can map max one pageful of TCEs at a time */
461         do {
462                 /*
463                  * Set up the page with TCE data, looping through and setting
464                  * the values.
465                  */
466                 limit = min_t(long, num_tce, 4096/TCE_ENTRY_SIZE);
467                 dma_offset = next + be64_to_cpu(maprange->dma_base);
468
469                 for (l = 0; l < limit; l++) {
470                         tcep[l] = cpu_to_be64(proto_tce | next);
471                         next += tce_size;
472                 }
473
474                 rc = plpar_tce_put_indirect(liobn,
475                                             dma_offset,
476                                             (u64)__pa(tcep),
477                                             limit);
478
479                 num_tce -= limit;
480         } while (num_tce > 0 && !rc);
481
482         /* error cleanup: caller will clear whole range */
483
484         local_irq_enable();
485         return rc;
486 }
487
488 static int tce_setrange_multi_pSeriesLP_walk(unsigned long start_pfn,
489                 unsigned long num_pfn, void *arg)
490 {
491         return tce_setrange_multi_pSeriesLP(start_pfn, num_pfn, arg);
492 }
493
494 static void iommu_table_setparms(struct pci_controller *phb,
495                                  struct device_node *dn,
496                                  struct iommu_table *tbl)
497 {
498         struct device_node *node;
499         const unsigned long *basep;
500         const u32 *sizep;
501
502         node = phb->dn;
503
504         basep = of_get_property(node, "linux,tce-base", NULL);
505         sizep = of_get_property(node, "linux,tce-size", NULL);
506         if (basep == NULL || sizep == NULL) {
507                 printk(KERN_ERR "PCI_DMA: iommu_table_setparms: %pOF has "
508                                 "missing tce entries !\n", dn);
509                 return;
510         }
511
512         tbl->it_base = (unsigned long)__va(*basep);
513
514         if (!is_kdump_kernel())
515                 memset((void *)tbl->it_base, 0, *sizep);
516
517         tbl->it_busno = phb->bus->number;
518         tbl->it_page_shift = IOMMU_PAGE_SHIFT_4K;
519
520         /* Units of tce entries */
521         tbl->it_offset = phb->dma_window_base_cur >> tbl->it_page_shift;
522
523         /* Test if we are going over 2GB of DMA space */
524         if (phb->dma_window_base_cur + phb->dma_window_size > 0x80000000ul) {
525                 udbg_printf("PCI_DMA: Unexpected number of IOAs under this PHB.\n");
526                 panic("PCI_DMA: Unexpected number of IOAs under this PHB.\n");
527         }
528
529         phb->dma_window_base_cur += phb->dma_window_size;
530
531         /* Set the tce table size - measured in entries */
532         tbl->it_size = phb->dma_window_size >> tbl->it_page_shift;
533
534         tbl->it_index = 0;
535         tbl->it_blocksize = 16;
536         tbl->it_type = TCE_PCI;
537 }
538
539 /*
540  * iommu_table_setparms_lpar
541  *
542  * Function: On pSeries LPAR systems, return TCE table info, given a pci bus.
543  */
544 static void iommu_table_setparms_lpar(struct pci_controller *phb,
545                                       struct device_node *dn,
546                                       struct iommu_table *tbl,
547                                       struct iommu_table_group *table_group,
548                                       const __be32 *dma_window)
549 {
550         unsigned long offset, size;
551
552         of_parse_dma_window(dn, dma_window, &tbl->it_index, &offset, &size);
553
554         tbl->it_busno = phb->bus->number;
555         tbl->it_page_shift = IOMMU_PAGE_SHIFT_4K;
556         tbl->it_base   = 0;
557         tbl->it_blocksize  = 16;
558         tbl->it_type = TCE_PCI;
559         tbl->it_offset = offset >> tbl->it_page_shift;
560         tbl->it_size = size >> tbl->it_page_shift;
561
562         table_group->tce32_start = offset;
563         table_group->tce32_size = size;
564 }
565
566 struct iommu_table_ops iommu_table_pseries_ops = {
567         .set = tce_build_pSeries,
568         .clear = tce_free_pSeries,
569         .get = tce_get_pseries
570 };
571
572 static void pci_dma_bus_setup_pSeries(struct pci_bus *bus)
573 {
574         struct device_node *dn;
575         struct iommu_table *tbl;
576         struct device_node *isa_dn, *isa_dn_orig;
577         struct device_node *tmp;
578         struct pci_dn *pci;
579         int children;
580
581         dn = pci_bus_to_OF_node(bus);
582
583         pr_debug("pci_dma_bus_setup_pSeries: setting up bus %pOF\n", dn);
584
585         if (bus->self) {
586                 /* This is not a root bus, any setup will be done for the
587                  * device-side of the bridge in iommu_dev_setup_pSeries().
588                  */
589                 return;
590         }
591         pci = PCI_DN(dn);
592
593         /* Check if the ISA bus on the system is under
594          * this PHB.
595          */
596         isa_dn = isa_dn_orig = of_find_node_by_type(NULL, "isa");
597
598         while (isa_dn && isa_dn != dn)
599                 isa_dn = isa_dn->parent;
600
601         of_node_put(isa_dn_orig);
602
603         /* Count number of direct PCI children of the PHB. */
604         for (children = 0, tmp = dn->child; tmp; tmp = tmp->sibling)
605                 children++;
606
607         pr_debug("Children: %d\n", children);
608
609         /* Calculate amount of DMA window per slot. Each window must be
610          * a power of two (due to pci_alloc_consistent requirements).
611          *
612          * Keep 256MB aside for PHBs with ISA.
613          */
614
615         if (!isa_dn) {
616                 /* No ISA/IDE - just set window size and return */
617                 pci->phb->dma_window_size = 0x80000000ul; /* To be divided */
618
619                 while (pci->phb->dma_window_size * children > 0x80000000ul)
620                         pci->phb->dma_window_size >>= 1;
621                 pr_debug("No ISA/IDE, window size is 0x%llx\n",
622                          pci->phb->dma_window_size);
623                 pci->phb->dma_window_base_cur = 0;
624
625                 return;
626         }
627
628         /* If we have ISA, then we probably have an IDE
629          * controller too. Allocate a 128MB table but
630          * skip the first 128MB to avoid stepping on ISA
631          * space.
632          */
633         pci->phb->dma_window_size = 0x8000000ul;
634         pci->phb->dma_window_base_cur = 0x8000000ul;
635
636         pci->table_group = iommu_pseries_alloc_group(pci->phb->node);
637         tbl = pci->table_group->tables[0];
638
639         iommu_table_setparms(pci->phb, dn, tbl);
640         tbl->it_ops = &iommu_table_pseries_ops;
641         iommu_init_table(tbl, pci->phb->node, 0, 0);
642
643         /* Divide the rest (1.75GB) among the children */
644         pci->phb->dma_window_size = 0x80000000ul;
645         while (pci->phb->dma_window_size * children > 0x70000000ul)
646                 pci->phb->dma_window_size >>= 1;
647
648         pr_debug("ISA/IDE, window size is 0x%llx\n", pci->phb->dma_window_size);
649 }
650
651 #ifdef CONFIG_IOMMU_API
652 static int tce_exchange_pseries(struct iommu_table *tbl, long index, unsigned
653                                 long *tce, enum dma_data_direction *direction,
654                                 bool realmode)
655 {
656         long rc;
657         unsigned long ioba = (unsigned long) index << tbl->it_page_shift;
658         unsigned long flags, oldtce = 0;
659         u64 proto_tce = iommu_direction_to_tce_perm(*direction);
660         unsigned long newtce = *tce | proto_tce;
661
662         spin_lock_irqsave(&tbl->large_pool.lock, flags);
663
664         rc = plpar_tce_get((u64)tbl->it_index, ioba, &oldtce);
665         if (!rc)
666                 rc = plpar_tce_put((u64)tbl->it_index, ioba, newtce);
667
668         if (!rc) {
669                 *direction = iommu_tce_direction(oldtce);
670                 *tce = oldtce & ~(TCE_PCI_READ | TCE_PCI_WRITE);
671         }
672
673         spin_unlock_irqrestore(&tbl->large_pool.lock, flags);
674
675         return rc;
676 }
677 #endif
678
679 struct iommu_table_ops iommu_table_lpar_multi_ops = {
680         .set = tce_buildmulti_pSeriesLP,
681 #ifdef CONFIG_IOMMU_API
682         .xchg_no_kill = tce_exchange_pseries,
683 #endif
684         .clear = tce_freemulti_pSeriesLP,
685         .get = tce_get_pSeriesLP
686 };
687
688 static void pci_dma_bus_setup_pSeriesLP(struct pci_bus *bus)
689 {
690         struct iommu_table *tbl;
691         struct device_node *dn, *pdn;
692         struct pci_dn *ppci;
693         const __be32 *dma_window = NULL;
694
695         dn = pci_bus_to_OF_node(bus);
696
697         pr_debug("pci_dma_bus_setup_pSeriesLP: setting up bus %pOF\n",
698                  dn);
699
700         /* Find nearest ibm,dma-window, walking up the device tree */
701         for (pdn = dn; pdn != NULL; pdn = pdn->parent) {
702                 dma_window = of_get_property(pdn, "ibm,dma-window", NULL);
703                 if (dma_window != NULL)
704                         break;
705         }
706
707         if (dma_window == NULL) {
708                 pr_debug("  no ibm,dma-window property !\n");
709                 return;
710         }
711
712         ppci = PCI_DN(pdn);
713
714         pr_debug("  parent is %pOF, iommu_table: 0x%p\n",
715                  pdn, ppci->table_group);
716
717         if (!ppci->table_group) {
718                 ppci->table_group = iommu_pseries_alloc_group(ppci->phb->node);
719                 tbl = ppci->table_group->tables[0];
720                 iommu_table_setparms_lpar(ppci->phb, pdn, tbl,
721                                 ppci->table_group, dma_window);
722                 tbl->it_ops = &iommu_table_lpar_multi_ops;
723                 iommu_init_table(tbl, ppci->phb->node, 0, 0);
724                 iommu_register_group(ppci->table_group,
725                                 pci_domain_nr(bus), 0);
726                 pr_debug("  created table: %p\n", ppci->table_group);
727         }
728 }
729
730
731 static void pci_dma_dev_setup_pSeries(struct pci_dev *dev)
732 {
733         struct device_node *dn;
734         struct iommu_table *tbl;
735
736         pr_debug("pci_dma_dev_setup_pSeries: %s\n", pci_name(dev));
737
738         dn = dev->dev.of_node;
739
740         /* If we're the direct child of a root bus, then we need to allocate
741          * an iommu table ourselves. The bus setup code should have setup
742          * the window sizes already.
743          */
744         if (!dev->bus->self) {
745                 struct pci_controller *phb = PCI_DN(dn)->phb;
746
747                 pr_debug(" --> first child, no bridge. Allocating iommu table.\n");
748                 PCI_DN(dn)->table_group = iommu_pseries_alloc_group(phb->node);
749                 tbl = PCI_DN(dn)->table_group->tables[0];
750                 iommu_table_setparms(phb, dn, tbl);
751                 tbl->it_ops = &iommu_table_pseries_ops;
752                 iommu_init_table(tbl, phb->node, 0, 0);
753                 set_iommu_table_base(&dev->dev, tbl);
754                 return;
755         }
756
757         /* If this device is further down the bus tree, search upwards until
758          * an already allocated iommu table is found and use that.
759          */
760
761         while (dn && PCI_DN(dn) && PCI_DN(dn)->table_group == NULL)
762                 dn = dn->parent;
763
764         if (dn && PCI_DN(dn))
765                 set_iommu_table_base(&dev->dev,
766                                 PCI_DN(dn)->table_group->tables[0]);
767         else
768                 printk(KERN_WARNING "iommu: Device %s has no iommu table\n",
769                        pci_name(dev));
770 }
771
772 static int __read_mostly disable_ddw;
773
774 static int __init disable_ddw_setup(char *str)
775 {
776         disable_ddw = 1;
777         printk(KERN_INFO "ppc iommu: disabling ddw.\n");
778
779         return 0;
780 }
781
782 early_param("disable_ddw", disable_ddw_setup);
783
784 static void remove_dma_window(struct device_node *np, u32 *ddw_avail,
785                               struct property *win)
786 {
787         struct dynamic_dma_window_prop *dwp;
788         u64 liobn;
789         int ret;
790
791         dwp = win->value;
792         liobn = (u64)be32_to_cpu(dwp->liobn);
793
794         /* clear the whole window, note the arg is in kernel pages */
795         ret = tce_clearrange_multi_pSeriesLP(0,
796                 1ULL << (be32_to_cpu(dwp->window_shift) - PAGE_SHIFT), dwp);
797         if (ret)
798                 pr_warn("%pOF failed to clear tces in window.\n",
799                         np);
800         else
801                 pr_debug("%pOF successfully cleared tces in window.\n",
802                          np);
803
804         ret = rtas_call(ddw_avail[DDW_REMOVE_PE_DMA_WIN], 1, 1, NULL, liobn);
805         if (ret)
806                 pr_warn("%pOF: failed to remove direct window: rtas returned "
807                         "%d to ibm,remove-pe-dma-window(%x) %llx\n",
808                         np, ret, ddw_avail[DDW_REMOVE_PE_DMA_WIN], liobn);
809         else
810                 pr_debug("%pOF: successfully removed direct window: rtas returned "
811                         "%d to ibm,remove-pe-dma-window(%x) %llx\n",
812                         np, ret, ddw_avail[DDW_REMOVE_PE_DMA_WIN], liobn);
813 }
814
815 static void remove_ddw(struct device_node *np, bool remove_prop)
816 {
817         struct property *win;
818         u32 ddw_avail[DDW_APPLICABLE_SIZE];
819         int ret = 0;
820
821         ret = of_property_read_u32_array(np, "ibm,ddw-applicable",
822                                          &ddw_avail[0], DDW_APPLICABLE_SIZE);
823         if (ret)
824                 return;
825
826         win = of_find_property(np, DIRECT64_PROPNAME, NULL);
827         if (!win)
828                 return;
829
830         if (win->length >= sizeof(struct dynamic_dma_window_prop))
831                 remove_dma_window(np, ddw_avail, win);
832
833         if (!remove_prop)
834                 return;
835
836         ret = of_remove_property(np, win);
837         if (ret)
838                 pr_warn("%pOF: failed to remove direct window property: %d\n",
839                         np, ret);
840 }
841
842 static u64 find_existing_ddw(struct device_node *pdn)
843 {
844         struct direct_window *window;
845         const struct dynamic_dma_window_prop *direct64;
846         u64 dma_addr = 0;
847
848         spin_lock(&direct_window_list_lock);
849         /* check if we already created a window and dupe that config if so */
850         list_for_each_entry(window, &direct_window_list, list) {
851                 if (window->device == pdn) {
852                         direct64 = window->prop;
853                         dma_addr = be64_to_cpu(direct64->dma_base);
854                         break;
855                 }
856         }
857         spin_unlock(&direct_window_list_lock);
858
859         return dma_addr;
860 }
861
862 static int find_existing_ddw_windows(void)
863 {
864         int len;
865         struct device_node *pdn;
866         struct direct_window *window;
867         const struct dynamic_dma_window_prop *direct64;
868
869         if (!firmware_has_feature(FW_FEATURE_LPAR))
870                 return 0;
871
872         for_each_node_with_property(pdn, DIRECT64_PROPNAME) {
873                 direct64 = of_get_property(pdn, DIRECT64_PROPNAME, &len);
874                 if (!direct64)
875                         continue;
876
877                 window = kzalloc(sizeof(*window), GFP_KERNEL);
878                 if (!window || len < sizeof(struct dynamic_dma_window_prop)) {
879                         kfree(window);
880                         remove_ddw(pdn, true);
881                         continue;
882                 }
883
884                 window->device = pdn;
885                 window->prop = direct64;
886                 spin_lock(&direct_window_list_lock);
887                 list_add(&window->list, &direct_window_list);
888                 spin_unlock(&direct_window_list_lock);
889         }
890
891         return 0;
892 }
893 machine_arch_initcall(pseries, find_existing_ddw_windows);
894
895 /**
896  * ddw_read_ext - Get the value of an DDW extension
897  * @np:         device node from which the extension value is to be read.
898  * @extnum:     index number of the extension.
899  * @value:      pointer to return value, modified when extension is available.
900  *
901  * Checks if "ibm,ddw-extensions" exists for this node, and get the value
902  * on index 'extnum'.
903  * It can be used only to check if a property exists, passing value == NULL.
904  *
905  * Returns:
906  *      0 if extension successfully read
907  *      -EINVAL if the "ibm,ddw-extensions" does not exist,
908  *      -ENODATA if "ibm,ddw-extensions" does not have a value, and
909  *      -EOVERFLOW if "ibm,ddw-extensions" does not contain this extension.
910  */
911 static inline int ddw_read_ext(const struct device_node *np, int extnum,
912                                u32 *value)
913 {
914         static const char propname[] = "ibm,ddw-extensions";
915         u32 count;
916         int ret;
917
918         ret = of_property_read_u32_index(np, propname, DDW_EXT_SIZE, &count);
919         if (ret)
920                 return ret;
921
922         if (count < extnum)
923                 return -EOVERFLOW;
924
925         if (!value)
926                 value = &count;
927
928         return of_property_read_u32_index(np, propname, extnum, value);
929 }
930
931 static int query_ddw(struct pci_dev *dev, const u32 *ddw_avail,
932                      struct ddw_query_response *query,
933                      struct device_node *parent)
934 {
935         struct device_node *dn;
936         struct pci_dn *pdn;
937         u32 cfg_addr, ext_query, query_out[5];
938         u64 buid;
939         int ret, out_sz;
940
941         /*
942          * From LoPAR level 2.8, "ibm,ddw-extensions" index 3 can rule how many
943          * output parameters ibm,query-pe-dma-windows will have, ranging from
944          * 5 to 6.
945          */
946         ret = ddw_read_ext(parent, DDW_EXT_QUERY_OUT_SIZE, &ext_query);
947         if (!ret && ext_query == 1)
948                 out_sz = 6;
949         else
950                 out_sz = 5;
951
952         /*
953          * Get the config address and phb buid of the PE window.
954          * Rely on eeh to retrieve this for us.
955          * Retrieve them from the pci device, not the node with the
956          * dma-window property
957          */
958         dn = pci_device_to_OF_node(dev);
959         pdn = PCI_DN(dn);
960         buid = pdn->phb->buid;
961         cfg_addr = ((pdn->busno << 16) | (pdn->devfn << 8));
962
963         ret = rtas_call(ddw_avail[DDW_QUERY_PE_DMA_WIN], 3, out_sz, query_out,
964                         cfg_addr, BUID_HI(buid), BUID_LO(buid));
965         dev_info(&dev->dev, "ibm,query-pe-dma-windows(%x) %x %x %x returned %d\n",
966                  ddw_avail[DDW_QUERY_PE_DMA_WIN], cfg_addr, BUID_HI(buid),
967                  BUID_LO(buid), ret);
968
969         switch (out_sz) {
970         case 5:
971                 query->windows_available = query_out[0];
972                 query->largest_available_block = query_out[1];
973                 query->page_size = query_out[2];
974                 query->migration_capable = query_out[3];
975                 break;
976         case 6:
977                 query->windows_available = query_out[0];
978                 query->largest_available_block = ((u64)query_out[1] << 32) |
979                                                  query_out[2];
980                 query->page_size = query_out[3];
981                 query->migration_capable = query_out[4];
982                 break;
983         }
984
985         return ret;
986 }
987
988 static int create_ddw(struct pci_dev *dev, const u32 *ddw_avail,
989                         struct ddw_create_response *create, int page_shift,
990                         int window_shift)
991 {
992         struct device_node *dn;
993         struct pci_dn *pdn;
994         u32 cfg_addr;
995         u64 buid;
996         int ret;
997
998         /*
999          * Get the config address and phb buid of the PE window.
1000          * Rely on eeh to retrieve this for us.
1001          * Retrieve them from the pci device, not the node with the
1002          * dma-window property
1003          */
1004         dn = pci_device_to_OF_node(dev);
1005         pdn = PCI_DN(dn);
1006         buid = pdn->phb->buid;
1007         cfg_addr = ((pdn->busno << 16) | (pdn->devfn << 8));
1008
1009         do {
1010                 /* extra outputs are LIOBN and dma-addr (hi, lo) */
1011                 ret = rtas_call(ddw_avail[DDW_CREATE_PE_DMA_WIN], 5, 4,
1012                                 (u32 *)create, cfg_addr, BUID_HI(buid),
1013                                 BUID_LO(buid), page_shift, window_shift);
1014         } while (rtas_busy_delay(ret));
1015         dev_info(&dev->dev,
1016                 "ibm,create-pe-dma-window(%x) %x %x %x %x %x returned %d "
1017                 "(liobn = 0x%x starting addr = %x %x)\n",
1018                  ddw_avail[DDW_CREATE_PE_DMA_WIN], cfg_addr, BUID_HI(buid),
1019                  BUID_LO(buid), page_shift, window_shift, ret, create->liobn,
1020                  create->addr_hi, create->addr_lo);
1021
1022         return ret;
1023 }
1024
1025 struct failed_ddw_pdn {
1026         struct device_node *pdn;
1027         struct list_head list;
1028 };
1029
1030 static LIST_HEAD(failed_ddw_pdn_list);
1031
1032 static phys_addr_t ddw_memory_hotplug_max(void)
1033 {
1034         phys_addr_t max_addr = memory_hotplug_max();
1035         struct device_node *memory;
1036
1037         for_each_node_by_type(memory, "memory") {
1038                 unsigned long start, size;
1039                 int n_mem_addr_cells, n_mem_size_cells, len;
1040                 const __be32 *memcell_buf;
1041
1042                 memcell_buf = of_get_property(memory, "reg", &len);
1043                 if (!memcell_buf || len <= 0)
1044                         continue;
1045
1046                 n_mem_addr_cells = of_n_addr_cells(memory);
1047                 n_mem_size_cells = of_n_size_cells(memory);
1048
1049                 start = of_read_number(memcell_buf, n_mem_addr_cells);
1050                 memcell_buf += n_mem_addr_cells;
1051                 size = of_read_number(memcell_buf, n_mem_size_cells);
1052                 memcell_buf += n_mem_size_cells;
1053
1054                 max_addr = max_t(phys_addr_t, max_addr, start + size);
1055         }
1056
1057         return max_addr;
1058 }
1059
1060 /*
1061  * Platforms supporting the DDW option starting with LoPAR level 2.7 implement
1062  * ibm,ddw-extensions, which carries the rtas token for
1063  * ibm,reset-pe-dma-windows.
1064  * That rtas-call can be used to restore the default DMA window for the device.
1065  */
1066 static void reset_dma_window(struct pci_dev *dev, struct device_node *par_dn)
1067 {
1068         int ret;
1069         u32 cfg_addr, reset_dma_win;
1070         u64 buid;
1071         struct device_node *dn;
1072         struct pci_dn *pdn;
1073
1074         ret = ddw_read_ext(par_dn, DDW_EXT_RESET_DMA_WIN, &reset_dma_win);
1075         if (ret)
1076                 return;
1077
1078         dn = pci_device_to_OF_node(dev);
1079         pdn = PCI_DN(dn);
1080         buid = pdn->phb->buid;
1081         cfg_addr = (pdn->busno << 16) | (pdn->devfn << 8);
1082
1083         ret = rtas_call(reset_dma_win, 3, 1, NULL, cfg_addr, BUID_HI(buid),
1084                         BUID_LO(buid));
1085         if (ret)
1086                 dev_info(&dev->dev,
1087                          "ibm,reset-pe-dma-windows(%x) %x %x %x returned %d ",
1088                          reset_dma_win, cfg_addr, BUID_HI(buid), BUID_LO(buid),
1089                          ret);
1090 }
1091
1092 /*
1093  * If the PE supports dynamic dma windows, and there is space for a table
1094  * that can map all pages in a linear offset, then setup such a table,
1095  * and record the dma-offset in the struct device.
1096  *
1097  * dev: the pci device we are checking
1098  * pdn: the parent pe node with the ibm,dma_window property
1099  * Future: also check if we can remap the base window for our base page size
1100  *
1101  * returns the dma offset for use by the direct mapped DMA code.
1102  */
1103 static u64 enable_ddw(struct pci_dev *dev, struct device_node *pdn)
1104 {
1105         int len, ret;
1106         struct ddw_query_response query;
1107         struct ddw_create_response create;
1108         int page_shift;
1109         u64 dma_addr, max_addr;
1110         struct device_node *dn;
1111         u32 ddw_avail[DDW_APPLICABLE_SIZE];
1112         struct direct_window *window;
1113         struct property *win64;
1114         struct dynamic_dma_window_prop *ddwprop;
1115         struct failed_ddw_pdn *fpdn;
1116         bool default_win_removed = false;
1117
1118         mutex_lock(&direct_window_init_mutex);
1119
1120         dma_addr = find_existing_ddw(pdn);
1121         if (dma_addr != 0)
1122                 goto out_unlock;
1123
1124         /*
1125          * If we already went through this for a previous function of
1126          * the same device and failed, we don't want to muck with the
1127          * DMA window again, as it will race with in-flight operations
1128          * and can lead to EEHs. The above mutex protects access to the
1129          * list.
1130          */
1131         list_for_each_entry(fpdn, &failed_ddw_pdn_list, list) {
1132                 if (fpdn->pdn == pdn)
1133                         goto out_unlock;
1134         }
1135
1136         /*
1137          * the ibm,ddw-applicable property holds the tokens for:
1138          * ibm,query-pe-dma-window
1139          * ibm,create-pe-dma-window
1140          * ibm,remove-pe-dma-window
1141          * for the given node in that order.
1142          * the property is actually in the parent, not the PE
1143          */
1144         ret = of_property_read_u32_array(pdn, "ibm,ddw-applicable",
1145                                          &ddw_avail[0], DDW_APPLICABLE_SIZE);
1146         if (ret)
1147                 goto out_failed;
1148
1149        /*
1150          * Query if there is a second window of size to map the
1151          * whole partition.  Query returns number of windows, largest
1152          * block assigned to PE (partition endpoint), and two bitmasks
1153          * of page sizes: supported and supported for migrate-dma.
1154          */
1155         dn = pci_device_to_OF_node(dev);
1156         ret = query_ddw(dev, ddw_avail, &query, pdn);
1157         if (ret != 0)
1158                 goto out_failed;
1159
1160         /*
1161          * If there is no window available, remove the default DMA window,
1162          * if it's present. This will make all the resources available to the
1163          * new DDW window.
1164          * If anything fails after this, we need to restore it, so also check
1165          * for extensions presence.
1166          */
1167         if (query.windows_available == 0) {
1168                 struct property *default_win;
1169                 int reset_win_ext;
1170
1171                 default_win = of_find_property(pdn, "ibm,dma-window", NULL);
1172                 if (!default_win)
1173                         goto out_failed;
1174
1175                 reset_win_ext = ddw_read_ext(pdn, DDW_EXT_RESET_DMA_WIN, NULL);
1176                 if (reset_win_ext)
1177                         goto out_failed;
1178
1179                 remove_dma_window(pdn, ddw_avail, default_win);
1180                 default_win_removed = true;
1181
1182                 /* Query again, to check if the window is available */
1183                 ret = query_ddw(dev, ddw_avail, &query, pdn);
1184                 if (ret != 0)
1185                         goto out_failed;
1186
1187                 if (query.windows_available == 0) {
1188                         /* no windows are available for this device. */
1189                         dev_dbg(&dev->dev, "no free dynamic windows");
1190                         goto out_failed;
1191                 }
1192         }
1193         if (query.page_size & 4) {
1194                 page_shift = 24; /* 16MB */
1195         } else if (query.page_size & 2) {
1196                 page_shift = 16; /* 64kB */
1197         } else if (query.page_size & 1) {
1198                 page_shift = 12; /* 4kB */
1199         } else {
1200                 dev_dbg(&dev->dev, "no supported direct page size in mask %x",
1201                           query.page_size);
1202                 goto out_failed;
1203         }
1204         /* verify the window * number of ptes will map the partition */
1205         /* check largest block * page size > max memory hotplug addr */
1206         max_addr = ddw_memory_hotplug_max();
1207         if (query.largest_available_block < (max_addr >> page_shift)) {
1208                 dev_dbg(&dev->dev, "can't map partition max 0x%llx with %llu "
1209                           "%llu-sized pages\n", max_addr,  query.largest_available_block,
1210                           1ULL << page_shift);
1211                 goto out_failed;
1212         }
1213         len = order_base_2(max_addr);
1214         win64 = kzalloc(sizeof(struct property), GFP_KERNEL);
1215         if (!win64) {
1216                 dev_info(&dev->dev,
1217                         "couldn't allocate property for 64bit dma window\n");
1218                 goto out_failed;
1219         }
1220         win64->name = kstrdup(DIRECT64_PROPNAME, GFP_KERNEL);
1221         win64->value = ddwprop = kmalloc(sizeof(*ddwprop), GFP_KERNEL);
1222         win64->length = sizeof(*ddwprop);
1223         if (!win64->name || !win64->value) {
1224                 dev_info(&dev->dev,
1225                         "couldn't allocate property name and value\n");
1226                 goto out_free_prop;
1227         }
1228
1229         ret = create_ddw(dev, ddw_avail, &create, page_shift, len);
1230         if (ret != 0)
1231                 goto out_free_prop;
1232
1233         ddwprop->liobn = cpu_to_be32(create.liobn);
1234         ddwprop->dma_base = cpu_to_be64(((u64)create.addr_hi << 32) |
1235                         create.addr_lo);
1236         ddwprop->tce_shift = cpu_to_be32(page_shift);
1237         ddwprop->window_shift = cpu_to_be32(len);
1238
1239         dev_dbg(&dev->dev, "created tce table LIOBN 0x%x for %pOF\n",
1240                   create.liobn, dn);
1241
1242         window = kzalloc(sizeof(*window), GFP_KERNEL);
1243         if (!window)
1244                 goto out_clear_window;
1245
1246         ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> PAGE_SHIFT,
1247                         win64->value, tce_setrange_multi_pSeriesLP_walk);
1248         if (ret) {
1249                 dev_info(&dev->dev, "failed to map direct window for %pOF: %d\n",
1250                          dn, ret);
1251                 goto out_free_window;
1252         }
1253
1254         ret = of_add_property(pdn, win64);
1255         if (ret) {
1256                 dev_err(&dev->dev, "unable to add dma window property for %pOF: %d",
1257                          pdn, ret);
1258                 goto out_free_window;
1259         }
1260
1261         window->device = pdn;
1262         window->prop = ddwprop;
1263         spin_lock(&direct_window_list_lock);
1264         list_add(&window->list, &direct_window_list);
1265         spin_unlock(&direct_window_list_lock);
1266
1267         dma_addr = be64_to_cpu(ddwprop->dma_base);
1268         goto out_unlock;
1269
1270 out_free_window:
1271         kfree(window);
1272
1273 out_clear_window:
1274         remove_ddw(pdn, true);
1275
1276 out_free_prop:
1277         kfree(win64->name);
1278         kfree(win64->value);
1279         kfree(win64);
1280
1281 out_failed:
1282         if (default_win_removed)
1283                 reset_dma_window(dev, pdn);
1284
1285         fpdn = kzalloc(sizeof(*fpdn), GFP_KERNEL);
1286         if (!fpdn)
1287                 goto out_unlock;
1288         fpdn->pdn = pdn;
1289         list_add(&fpdn->list, &failed_ddw_pdn_list);
1290
1291 out_unlock:
1292         mutex_unlock(&direct_window_init_mutex);
1293         return dma_addr;
1294 }
1295
1296 static void pci_dma_dev_setup_pSeriesLP(struct pci_dev *dev)
1297 {
1298         struct device_node *pdn, *dn;
1299         struct iommu_table *tbl;
1300         const __be32 *dma_window = NULL;
1301         struct pci_dn *pci;
1302
1303         pr_debug("pci_dma_dev_setup_pSeriesLP: %s\n", pci_name(dev));
1304
1305         /* dev setup for LPAR is a little tricky, since the device tree might
1306          * contain the dma-window properties per-device and not necessarily
1307          * for the bus. So we need to search upwards in the tree until we
1308          * either hit a dma-window property, OR find a parent with a table
1309          * already allocated.
1310          */
1311         dn = pci_device_to_OF_node(dev);
1312         pr_debug("  node is %pOF\n", dn);
1313
1314         for (pdn = dn; pdn && PCI_DN(pdn) && !PCI_DN(pdn)->table_group;
1315              pdn = pdn->parent) {
1316                 dma_window = of_get_property(pdn, "ibm,dma-window", NULL);
1317                 if (dma_window)
1318                         break;
1319         }
1320
1321         if (!pdn || !PCI_DN(pdn)) {
1322                 printk(KERN_WARNING "pci_dma_dev_setup_pSeriesLP: "
1323                        "no DMA window found for pci dev=%s dn=%pOF\n",
1324                                  pci_name(dev), dn);
1325                 return;
1326         }
1327         pr_debug("  parent is %pOF\n", pdn);
1328
1329         pci = PCI_DN(pdn);
1330         if (!pci->table_group) {
1331                 pci->table_group = iommu_pseries_alloc_group(pci->phb->node);
1332                 tbl = pci->table_group->tables[0];
1333                 iommu_table_setparms_lpar(pci->phb, pdn, tbl,
1334                                 pci->table_group, dma_window);
1335                 tbl->it_ops = &iommu_table_lpar_multi_ops;
1336                 iommu_init_table(tbl, pci->phb->node, 0, 0);
1337                 iommu_register_group(pci->table_group,
1338                                 pci_domain_nr(pci->phb->bus), 0);
1339                 pr_debug("  created table: %p\n", pci->table_group);
1340         } else {
1341                 pr_debug("  found DMA window, table: %p\n", pci->table_group);
1342         }
1343
1344         set_iommu_table_base(&dev->dev, pci->table_group->tables[0]);
1345         iommu_add_device(pci->table_group, &dev->dev);
1346 }
1347
1348 static bool iommu_bypass_supported_pSeriesLP(struct pci_dev *pdev, u64 dma_mask)
1349 {
1350         struct device_node *dn = pci_device_to_OF_node(pdev), *pdn;
1351         const __be32 *dma_window = NULL;
1352
1353         /* only attempt to use a new window if 64-bit DMA is requested */
1354         if (dma_mask < DMA_BIT_MASK(64))
1355                 return false;
1356
1357         dev_dbg(&pdev->dev, "node is %pOF\n", dn);
1358
1359         /*
1360          * the device tree might contain the dma-window properties
1361          * per-device and not necessarily for the bus. So we need to
1362          * search upwards in the tree until we either hit a dma-window
1363          * property, OR find a parent with a table already allocated.
1364          */
1365         for (pdn = dn; pdn && PCI_DN(pdn) && !PCI_DN(pdn)->table_group;
1366                         pdn = pdn->parent) {
1367                 dma_window = of_get_property(pdn, "ibm,dma-window", NULL);
1368                 if (dma_window)
1369                         break;
1370         }
1371
1372         if (pdn && PCI_DN(pdn)) {
1373                 pdev->dev.archdata.dma_offset = enable_ddw(pdev, pdn);
1374                 if (pdev->dev.archdata.dma_offset)
1375                         return true;
1376         }
1377
1378         return false;
1379 }
1380
1381 static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
1382                 void *data)
1383 {
1384         struct direct_window *window;
1385         struct memory_notify *arg = data;
1386         int ret = 0;
1387
1388         switch (action) {
1389         case MEM_GOING_ONLINE:
1390                 spin_lock(&direct_window_list_lock);
1391                 list_for_each_entry(window, &direct_window_list, list) {
1392                         ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn,
1393                                         arg->nr_pages, window->prop);
1394                         /* XXX log error */
1395                 }
1396                 spin_unlock(&direct_window_list_lock);
1397                 break;
1398         case MEM_CANCEL_ONLINE:
1399         case MEM_OFFLINE:
1400                 spin_lock(&direct_window_list_lock);
1401                 list_for_each_entry(window, &direct_window_list, list) {
1402                         ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn,
1403                                         arg->nr_pages, window->prop);
1404                         /* XXX log error */
1405                 }
1406                 spin_unlock(&direct_window_list_lock);
1407                 break;
1408         default:
1409                 break;
1410         }
1411         if (ret && action != MEM_CANCEL_ONLINE)
1412                 return NOTIFY_BAD;
1413
1414         return NOTIFY_OK;
1415 }
1416
1417 static struct notifier_block iommu_mem_nb = {
1418         .notifier_call = iommu_mem_notifier,
1419 };
1420
1421 static int iommu_reconfig_notifier(struct notifier_block *nb, unsigned long action, void *data)
1422 {
1423         int err = NOTIFY_OK;
1424         struct of_reconfig_data *rd = data;
1425         struct device_node *np = rd->dn;
1426         struct pci_dn *pci = PCI_DN(np);
1427         struct direct_window *window;
1428
1429         switch (action) {
1430         case OF_RECONFIG_DETACH_NODE:
1431                 /*
1432                  * Removing the property will invoke the reconfig
1433                  * notifier again, which causes dead-lock on the
1434                  * read-write semaphore of the notifier chain. So
1435                  * we have to remove the property when releasing
1436                  * the device node.
1437                  */
1438                 remove_ddw(np, false);
1439                 if (pci && pci->table_group)
1440                         iommu_pseries_free_group(pci->table_group,
1441                                         np->full_name);
1442
1443                 spin_lock(&direct_window_list_lock);
1444                 list_for_each_entry(window, &direct_window_list, list) {
1445                         if (window->device == np) {
1446                                 list_del(&window->list);
1447                                 kfree(window);
1448                                 break;
1449                         }
1450                 }
1451                 spin_unlock(&direct_window_list_lock);
1452                 break;
1453         default:
1454                 err = NOTIFY_DONE;
1455                 break;
1456         }
1457         return err;
1458 }
1459
1460 static struct notifier_block iommu_reconfig_nb = {
1461         .notifier_call = iommu_reconfig_notifier,
1462 };
1463
1464 /* These are called very early. */
1465 void iommu_init_early_pSeries(void)
1466 {
1467         if (of_chosen && of_get_property(of_chosen, "linux,iommu-off", NULL))
1468                 return;
1469
1470         if (firmware_has_feature(FW_FEATURE_LPAR)) {
1471                 pseries_pci_controller_ops.dma_bus_setup = pci_dma_bus_setup_pSeriesLP;
1472                 pseries_pci_controller_ops.dma_dev_setup = pci_dma_dev_setup_pSeriesLP;
1473                 if (!disable_ddw)
1474                         pseries_pci_controller_ops.iommu_bypass_supported =
1475                                 iommu_bypass_supported_pSeriesLP;
1476         } else {
1477                 pseries_pci_controller_ops.dma_bus_setup = pci_dma_bus_setup_pSeries;
1478                 pseries_pci_controller_ops.dma_dev_setup = pci_dma_dev_setup_pSeries;
1479         }
1480
1481
1482         of_reconfig_notifier_register(&iommu_reconfig_nb);
1483         register_memory_notifier(&iommu_mem_nb);
1484
1485         set_pci_dma_ops(&dma_iommu_ops);
1486 }
1487
1488 static int __init disable_multitce(char *str)
1489 {
1490         if (strcmp(str, "off") == 0 &&
1491             firmware_has_feature(FW_FEATURE_LPAR) &&
1492             (firmware_has_feature(FW_FEATURE_PUT_TCE_IND) ||
1493              firmware_has_feature(FW_FEATURE_STUFF_TCE))) {
1494                 printk(KERN_INFO "Disabling MULTITCE firmware feature\n");
1495                 powerpc_firmware_features &=
1496                         ~(FW_FEATURE_PUT_TCE_IND | FW_FEATURE_STUFF_TCE);
1497         }
1498         return 1;
1499 }
1500
1501 __setup("multitce=", disable_multitce);
1502
1503 static int tce_iommu_bus_notifier(struct notifier_block *nb,
1504                 unsigned long action, void *data)
1505 {
1506         struct device *dev = data;
1507
1508         switch (action) {
1509         case BUS_NOTIFY_DEL_DEVICE:
1510                 iommu_del_device(dev);
1511                 return 0;
1512         default:
1513                 return 0;
1514         }
1515 }
1516
1517 static struct notifier_block tce_iommu_bus_nb = {
1518         .notifier_call = tce_iommu_bus_notifier,
1519 };
1520
1521 static int __init tce_iommu_bus_notifier_init(void)
1522 {
1523         bus_register_notifier(&pci_bus_type, &tce_iommu_bus_nb);
1524         return 0;
1525 }
1526 machine_subsys_initcall_sync(pseries, tce_iommu_bus_notifier_init);