Loading...
1/*
2 * ioport.c: Simple io mapping allocator.
3 *
4 * Copyright (C) 1995 David S. Miller (davem@caip.rutgers.edu)
5 * Copyright (C) 1995 Miguel de Icaza (miguel@nuclecu.unam.mx)
6 *
7 * 1996: sparc_free_io, 1999: ioremap()/iounmap() by Pete Zaitcev.
8 *
9 * 2000/01/29
10 * <rth> zait: as long as pci_alloc_consistent produces something addressable,
11 * things are ok.
12 * <zaitcev> rth: no, it is relevant, because get_free_pages returns you a
13 * pointer into the big page mapping
14 * <rth> zait: so what?
15 * <rth> zait: remap_it_my_way(virt_to_phys(get_free_page()))
16 * <zaitcev> Hmm
17 * <zaitcev> Suppose I did this remap_it_my_way(virt_to_phys(get_free_page())).
18 * So far so good.
19 * <zaitcev> Now, driver calls pci_free_consistent(with result of
20 * remap_it_my_way()).
21 * <zaitcev> How do you find the address to pass to free_pages()?
22 * <rth> zait: walk the page tables? It's only two or three level after all.
23 * <rth> zait: you have to walk them anyway to remove the mapping.
24 * <zaitcev> Hmm
25 * <zaitcev> Sounds reasonable
26 */
27
28#include <linux/module.h>
29#include <linux/sched.h>
30#include <linux/kernel.h>
31#include <linux/errno.h>
32#include <linux/types.h>
33#include <linux/ioport.h>
34#include <linux/mm.h>
35#include <linux/slab.h>
36#include <linux/pci.h> /* struct pci_dev */
37#include <linux/proc_fs.h>
38#include <linux/seq_file.h>
39#include <linux/scatterlist.h>
40#include <linux/of_device.h>
41
42#include <asm/io.h>
43#include <asm/vaddrs.h>
44#include <asm/oplib.h>
45#include <asm/prom.h>
46#include <asm/page.h>
47#include <asm/pgalloc.h>
48#include <asm/dma.h>
49#include <asm/iommu.h>
50#include <asm/io-unit.h>
51#include <asm/leon.h>
52
53const struct sparc32_dma_ops *sparc32_dma_ops;
54
55/* This function must make sure that caches and memory are coherent after DMA
56 * On LEON systems without cache snooping it flushes the entire D-CACHE.
57 */
58static inline void dma_make_coherent(unsigned long pa, unsigned long len)
59{
60 if (sparc_cpu_model == sparc_leon) {
61 if (!sparc_leon3_snooping_enabled())
62 leon_flush_dcache_all();
63 }
64}
65
66static void __iomem *_sparc_ioremap(struct resource *res, u32 bus, u32 pa, int sz);
67static void __iomem *_sparc_alloc_io(unsigned int busno, unsigned long phys,
68 unsigned long size, char *name);
69static void _sparc_free_io(struct resource *res);
70
71static void register_proc_sparc_ioport(void);
72
73/* This points to the next to use virtual memory for DVMA mappings */
74static struct resource _sparc_dvma = {
75 .name = "sparc_dvma", .start = DVMA_VADDR, .end = DVMA_END - 1
76};
77/* This points to the start of I/O mappings, cluable from outside. */
78/*ext*/ struct resource sparc_iomap = {
79 .name = "sparc_iomap", .start = IOBASE_VADDR, .end = IOBASE_END - 1
80};
81
82/*
83 * Our mini-allocator...
84 * Boy this is gross! We need it because we must map I/O for
85 * timers and interrupt controller before the kmalloc is available.
86 */
87
88#define XNMLN 15
89#define XNRES 10 /* SS-10 uses 8 */
90
91struct xresource {
92 struct resource xres; /* Must be first */
93 int xflag; /* 1 == used */
94 char xname[XNMLN+1];
95};
96
97static struct xresource xresv[XNRES];
98
99static struct xresource *xres_alloc(void) {
100 struct xresource *xrp;
101 int n;
102
103 xrp = xresv;
104 for (n = 0; n < XNRES; n++) {
105 if (xrp->xflag == 0) {
106 xrp->xflag = 1;
107 return xrp;
108 }
109 xrp++;
110 }
111 return NULL;
112}
113
114static void xres_free(struct xresource *xrp) {
115 xrp->xflag = 0;
116}
117
118/*
119 * These are typically used in PCI drivers
120 * which are trying to be cross-platform.
121 *
122 * Bus type is always zero on IIep.
123 */
124void __iomem *ioremap(unsigned long offset, unsigned long size)
125{
126 char name[14];
127
128 sprintf(name, "phys_%08x", (u32)offset);
129 return _sparc_alloc_io(0, offset, size, name);
130}
131EXPORT_SYMBOL(ioremap);
132
133/*
134 * Comlimentary to ioremap().
135 */
136void iounmap(volatile void __iomem *virtual)
137{
138 unsigned long vaddr = (unsigned long) virtual & PAGE_MASK;
139 struct resource *res;
140
141 /*
142 * XXX Too slow. Can have 8192 DVMA pages on sun4m in the worst case.
143 * This probably warrants some sort of hashing.
144 */
145 if ((res = lookup_resource(&sparc_iomap, vaddr)) == NULL) {
146 printk("free_io/iounmap: cannot free %lx\n", vaddr);
147 return;
148 }
149 _sparc_free_io(res);
150
151 if ((char *)res >= (char*)xresv && (char *)res < (char *)&xresv[XNRES]) {
152 xres_free((struct xresource *)res);
153 } else {
154 kfree(res);
155 }
156}
157EXPORT_SYMBOL(iounmap);
158
159void __iomem *of_ioremap(struct resource *res, unsigned long offset,
160 unsigned long size, char *name)
161{
162 return _sparc_alloc_io(res->flags & 0xF,
163 res->start + offset,
164 size, name);
165}
166EXPORT_SYMBOL(of_ioremap);
167
168void of_iounmap(struct resource *res, void __iomem *base, unsigned long size)
169{
170 iounmap(base);
171}
172EXPORT_SYMBOL(of_iounmap);
173
174/*
175 * Meat of mapping
176 */
177static void __iomem *_sparc_alloc_io(unsigned int busno, unsigned long phys,
178 unsigned long size, char *name)
179{
180 static int printed_full;
181 struct xresource *xres;
182 struct resource *res;
183 char *tack;
184 int tlen;
185 void __iomem *va; /* P3 diag */
186
187 if (name == NULL) name = "???";
188
189 if ((xres = xres_alloc()) != 0) {
190 tack = xres->xname;
191 res = &xres->xres;
192 } else {
193 if (!printed_full) {
194 printk("ioremap: done with statics, switching to malloc\n");
195 printed_full = 1;
196 }
197 tlen = strlen(name);
198 tack = kmalloc(sizeof (struct resource) + tlen + 1, GFP_KERNEL);
199 if (tack == NULL) return NULL;
200 memset(tack, 0, sizeof(struct resource));
201 res = (struct resource *) tack;
202 tack += sizeof (struct resource);
203 }
204
205 strlcpy(tack, name, XNMLN+1);
206 res->name = tack;
207
208 va = _sparc_ioremap(res, busno, phys, size);
209 /* printk("ioremap(0x%x:%08lx[0x%lx])=%p\n", busno, phys, size, va); */ /* P3 diag */
210 return va;
211}
212
213/*
214 */
215static void __iomem *
216_sparc_ioremap(struct resource *res, u32 bus, u32 pa, int sz)
217{
218 unsigned long offset = ((unsigned long) pa) & (~PAGE_MASK);
219
220 if (allocate_resource(&sparc_iomap, res,
221 (offset + sz + PAGE_SIZE-1) & PAGE_MASK,
222 sparc_iomap.start, sparc_iomap.end, PAGE_SIZE, NULL, NULL) != 0) {
223 /* Usually we cannot see printks in this case. */
224 prom_printf("alloc_io_res(%s): cannot occupy\n",
225 (res->name != NULL)? res->name: "???");
226 prom_halt();
227 }
228
229 pa &= PAGE_MASK;
230 srmmu_mapiorange(bus, pa, res->start, resource_size(res));
231
232 return (void __iomem *)(unsigned long)(res->start + offset);
233}
234
235/*
236 * Comlimentary to _sparc_ioremap().
237 */
238static void _sparc_free_io(struct resource *res)
239{
240 unsigned long plen;
241
242 plen = resource_size(res);
243 BUG_ON((plen & (PAGE_SIZE-1)) != 0);
244 srmmu_unmapiorange(res->start, plen);
245 release_resource(res);
246}
247
248#ifdef CONFIG_SBUS
249
250void sbus_set_sbus64(struct device *dev, int x)
251{
252 printk("sbus_set_sbus64: unsupported\n");
253}
254EXPORT_SYMBOL(sbus_set_sbus64);
255
256/*
257 * Allocate a chunk of memory suitable for DMA.
258 * Typically devices use them for control blocks.
259 * CPU may access them without any explicit flushing.
260 */
261static void *sbus_alloc_coherent(struct device *dev, size_t len,
262 dma_addr_t *dma_addrp, gfp_t gfp,
263 struct dma_attrs *attrs)
264{
265 struct platform_device *op = to_platform_device(dev);
266 unsigned long len_total = PAGE_ALIGN(len);
267 unsigned long va;
268 struct resource *res;
269 int order;
270
271 /* XXX why are some lengths signed, others unsigned? */
272 if (len <= 0) {
273 return NULL;
274 }
275 /* XXX So what is maxphys for us and how do drivers know it? */
276 if (len > 256*1024) { /* __get_free_pages() limit */
277 return NULL;
278 }
279
280 order = get_order(len_total);
281 if ((va = __get_free_pages(GFP_KERNEL|__GFP_COMP, order)) == 0)
282 goto err_nopages;
283
284 if ((res = kzalloc(sizeof(struct resource), GFP_KERNEL)) == NULL)
285 goto err_nomem;
286
287 if (allocate_resource(&_sparc_dvma, res, len_total,
288 _sparc_dvma.start, _sparc_dvma.end, PAGE_SIZE, NULL, NULL) != 0) {
289 printk("sbus_alloc_consistent: cannot occupy 0x%lx", len_total);
290 goto err_nova;
291 }
292
293 // XXX The sbus_map_dma_area does this for us below, see comments.
294 // srmmu_mapiorange(0, virt_to_phys(va), res->start, len_total);
295 /*
296 * XXX That's where sdev would be used. Currently we load
297 * all iommu tables with the same translations.
298 */
299 if (sbus_map_dma_area(dev, dma_addrp, va, res->start, len_total) != 0)
300 goto err_noiommu;
301
302 res->name = op->dev.of_node->name;
303
304 return (void *)(unsigned long)res->start;
305
306err_noiommu:
307 release_resource(res);
308err_nova:
309 kfree(res);
310err_nomem:
311 free_pages(va, order);
312err_nopages:
313 return NULL;
314}
315
316static void sbus_free_coherent(struct device *dev, size_t n, void *p,
317 dma_addr_t ba, struct dma_attrs *attrs)
318{
319 struct resource *res;
320 struct page *pgv;
321
322 if ((res = lookup_resource(&_sparc_dvma,
323 (unsigned long)p)) == NULL) {
324 printk("sbus_free_consistent: cannot free %p\n", p);
325 return;
326 }
327
328 if (((unsigned long)p & (PAGE_SIZE-1)) != 0) {
329 printk("sbus_free_consistent: unaligned va %p\n", p);
330 return;
331 }
332
333 n = PAGE_ALIGN(n);
334 if (resource_size(res) != n) {
335 printk("sbus_free_consistent: region 0x%lx asked 0x%zx\n",
336 (long)resource_size(res), n);
337 return;
338 }
339
340 release_resource(res);
341 kfree(res);
342
343 pgv = virt_to_page(p);
344 sbus_unmap_dma_area(dev, ba, n);
345
346 __free_pages(pgv, get_order(n));
347}
348
349/*
350 * Map a chunk of memory so that devices can see it.
351 * CPU view of this memory may be inconsistent with
352 * a device view and explicit flushing is necessary.
353 */
354static dma_addr_t sbus_map_page(struct device *dev, struct page *page,
355 unsigned long offset, size_t len,
356 enum dma_data_direction dir,
357 struct dma_attrs *attrs)
358{
359 void *va = page_address(page) + offset;
360
361 /* XXX why are some lengths signed, others unsigned? */
362 if (len <= 0) {
363 return 0;
364 }
365 /* XXX So what is maxphys for us and how do drivers know it? */
366 if (len > 256*1024) { /* __get_free_pages() limit */
367 return 0;
368 }
369 return mmu_get_scsi_one(dev, va, len);
370}
371
372static void sbus_unmap_page(struct device *dev, dma_addr_t ba, size_t n,
373 enum dma_data_direction dir, struct dma_attrs *attrs)
374{
375 mmu_release_scsi_one(dev, ba, n);
376}
377
378static int sbus_map_sg(struct device *dev, struct scatterlist *sg, int n,
379 enum dma_data_direction dir, struct dma_attrs *attrs)
380{
381 mmu_get_scsi_sgl(dev, sg, n);
382 return n;
383}
384
385static void sbus_unmap_sg(struct device *dev, struct scatterlist *sg, int n,
386 enum dma_data_direction dir, struct dma_attrs *attrs)
387{
388 mmu_release_scsi_sgl(dev, sg, n);
389}
390
391static void sbus_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
392 int n, enum dma_data_direction dir)
393{
394 BUG();
395}
396
397static void sbus_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
398 int n, enum dma_data_direction dir)
399{
400 BUG();
401}
402
403struct dma_map_ops sbus_dma_ops = {
404 .alloc = sbus_alloc_coherent,
405 .free = sbus_free_coherent,
406 .map_page = sbus_map_page,
407 .unmap_page = sbus_unmap_page,
408 .map_sg = sbus_map_sg,
409 .unmap_sg = sbus_unmap_sg,
410 .sync_sg_for_cpu = sbus_sync_sg_for_cpu,
411 .sync_sg_for_device = sbus_sync_sg_for_device,
412};
413
414static int __init sparc_register_ioport(void)
415{
416 register_proc_sparc_ioport();
417
418 return 0;
419}
420
421arch_initcall(sparc_register_ioport);
422
423#endif /* CONFIG_SBUS */
424
425
426/* Allocate and map kernel buffer using consistent mode DMA for a device.
427 * hwdev should be valid struct pci_dev pointer for PCI devices.
428 */
429static void *pci32_alloc_coherent(struct device *dev, size_t len,
430 dma_addr_t *pba, gfp_t gfp,
431 struct dma_attrs *attrs)
432{
433 unsigned long len_total = PAGE_ALIGN(len);
434 void *va;
435 struct resource *res;
436 int order;
437
438 if (len == 0) {
439 return NULL;
440 }
441 if (len > 256*1024) { /* __get_free_pages() limit */
442 return NULL;
443 }
444
445 order = get_order(len_total);
446 va = (void *) __get_free_pages(GFP_KERNEL, order);
447 if (va == NULL) {
448 printk("pci_alloc_consistent: no %ld pages\n", len_total>>PAGE_SHIFT);
449 goto err_nopages;
450 }
451
452 if ((res = kzalloc(sizeof(struct resource), GFP_KERNEL)) == NULL) {
453 printk("pci_alloc_consistent: no core\n");
454 goto err_nomem;
455 }
456
457 if (allocate_resource(&_sparc_dvma, res, len_total,
458 _sparc_dvma.start, _sparc_dvma.end, PAGE_SIZE, NULL, NULL) != 0) {
459 printk("pci_alloc_consistent: cannot occupy 0x%lx", len_total);
460 goto err_nova;
461 }
462 srmmu_mapiorange(0, virt_to_phys(va), res->start, len_total);
463
464 *pba = virt_to_phys(va); /* equals virt_to_bus (R.I.P.) for us. */
465 return (void *) res->start;
466
467err_nova:
468 kfree(res);
469err_nomem:
470 free_pages((unsigned long)va, order);
471err_nopages:
472 return NULL;
473}
474
475/* Free and unmap a consistent DMA buffer.
476 * cpu_addr is what was returned from pci_alloc_consistent,
477 * size must be the same as what as passed into pci_alloc_consistent,
478 * and likewise dma_addr must be the same as what *dma_addrp was set to.
479 *
480 * References to the memory and mappings associated with cpu_addr/dma_addr
481 * past this call are illegal.
482 */
483static void pci32_free_coherent(struct device *dev, size_t n, void *p,
484 dma_addr_t ba, struct dma_attrs *attrs)
485{
486 struct resource *res;
487
488 if ((res = lookup_resource(&_sparc_dvma,
489 (unsigned long)p)) == NULL) {
490 printk("pci_free_consistent: cannot free %p\n", p);
491 return;
492 }
493
494 if (((unsigned long)p & (PAGE_SIZE-1)) != 0) {
495 printk("pci_free_consistent: unaligned va %p\n", p);
496 return;
497 }
498
499 n = PAGE_ALIGN(n);
500 if (resource_size(res) != n) {
501 printk("pci_free_consistent: region 0x%lx asked 0x%lx\n",
502 (long)resource_size(res), (long)n);
503 return;
504 }
505
506 dma_make_coherent(ba, n);
507 srmmu_unmapiorange((unsigned long)p, n);
508
509 release_resource(res);
510 kfree(res);
511 free_pages((unsigned long)phys_to_virt(ba), get_order(n));
512}
513
514/*
515 * Same as pci_map_single, but with pages.
516 */
517static dma_addr_t pci32_map_page(struct device *dev, struct page *page,
518 unsigned long offset, size_t size,
519 enum dma_data_direction dir,
520 struct dma_attrs *attrs)
521{
522 /* IIep is write-through, not flushing. */
523 return page_to_phys(page) + offset;
524}
525
526static void pci32_unmap_page(struct device *dev, dma_addr_t ba, size_t size,
527 enum dma_data_direction dir, struct dma_attrs *attrs)
528{
529 if (dir != PCI_DMA_TODEVICE)
530 dma_make_coherent(ba, PAGE_ALIGN(size));
531}
532
533/* Map a set of buffers described by scatterlist in streaming
534 * mode for DMA. This is the scather-gather version of the
535 * above pci_map_single interface. Here the scatter gather list
536 * elements are each tagged with the appropriate dma address
537 * and length. They are obtained via sg_dma_{address,length}(SG).
538 *
539 * NOTE: An implementation may be able to use a smaller number of
540 * DMA address/length pairs than there are SG table elements.
541 * (for example via virtual mapping capabilities)
542 * The routine returns the number of addr/length pairs actually
543 * used, at most nents.
544 *
545 * Device ownership issues as mentioned above for pci_map_single are
546 * the same here.
547 */
548static int pci32_map_sg(struct device *device, struct scatterlist *sgl,
549 int nents, enum dma_data_direction dir,
550 struct dma_attrs *attrs)
551{
552 struct scatterlist *sg;
553 int n;
554
555 /* IIep is write-through, not flushing. */
556 for_each_sg(sgl, sg, nents, n) {
557 sg->dma_address = sg_phys(sg);
558 sg->dma_length = sg->length;
559 }
560 return nents;
561}
562
563/* Unmap a set of streaming mode DMA translations.
564 * Again, cpu read rules concerning calls here are the same as for
565 * pci_unmap_single() above.
566 */
567static void pci32_unmap_sg(struct device *dev, struct scatterlist *sgl,
568 int nents, enum dma_data_direction dir,
569 struct dma_attrs *attrs)
570{
571 struct scatterlist *sg;
572 int n;
573
574 if (dir != PCI_DMA_TODEVICE) {
575 for_each_sg(sgl, sg, nents, n) {
576 dma_make_coherent(sg_phys(sg), PAGE_ALIGN(sg->length));
577 }
578 }
579}
580
581/* Make physical memory consistent for a single
582 * streaming mode DMA translation before or after a transfer.
583 *
584 * If you perform a pci_map_single() but wish to interrogate the
585 * buffer using the cpu, yet do not wish to teardown the PCI dma
586 * mapping, you must call this function before doing so. At the
587 * next point you give the PCI dma address back to the card, you
588 * must first perform a pci_dma_sync_for_device, and then the
589 * device again owns the buffer.
590 */
591static void pci32_sync_single_for_cpu(struct device *dev, dma_addr_t ba,
592 size_t size, enum dma_data_direction dir)
593{
594 if (dir != PCI_DMA_TODEVICE) {
595 dma_make_coherent(ba, PAGE_ALIGN(size));
596 }
597}
598
599static void pci32_sync_single_for_device(struct device *dev, dma_addr_t ba,
600 size_t size, enum dma_data_direction dir)
601{
602 if (dir != PCI_DMA_TODEVICE) {
603 dma_make_coherent(ba, PAGE_ALIGN(size));
604 }
605}
606
607/* Make physical memory consistent for a set of streaming
608 * mode DMA translations after a transfer.
609 *
610 * The same as pci_dma_sync_single_* but for a scatter-gather list,
611 * same rules and usage.
612 */
613static void pci32_sync_sg_for_cpu(struct device *dev, struct scatterlist *sgl,
614 int nents, enum dma_data_direction dir)
615{
616 struct scatterlist *sg;
617 int n;
618
619 if (dir != PCI_DMA_TODEVICE) {
620 for_each_sg(sgl, sg, nents, n) {
621 dma_make_coherent(sg_phys(sg), PAGE_ALIGN(sg->length));
622 }
623 }
624}
625
626static void pci32_sync_sg_for_device(struct device *device, struct scatterlist *sgl,
627 int nents, enum dma_data_direction dir)
628{
629 struct scatterlist *sg;
630 int n;
631
632 if (dir != PCI_DMA_TODEVICE) {
633 for_each_sg(sgl, sg, nents, n) {
634 dma_make_coherent(sg_phys(sg), PAGE_ALIGN(sg->length));
635 }
636 }
637}
638
639struct dma_map_ops pci32_dma_ops = {
640 .alloc = pci32_alloc_coherent,
641 .free = pci32_free_coherent,
642 .map_page = pci32_map_page,
643 .unmap_page = pci32_unmap_page,
644 .map_sg = pci32_map_sg,
645 .unmap_sg = pci32_unmap_sg,
646 .sync_single_for_cpu = pci32_sync_single_for_cpu,
647 .sync_single_for_device = pci32_sync_single_for_device,
648 .sync_sg_for_cpu = pci32_sync_sg_for_cpu,
649 .sync_sg_for_device = pci32_sync_sg_for_device,
650};
651EXPORT_SYMBOL(pci32_dma_ops);
652
653/* leon re-uses pci32_dma_ops */
654struct dma_map_ops *leon_dma_ops = &pci32_dma_ops;
655EXPORT_SYMBOL(leon_dma_ops);
656
657struct dma_map_ops *dma_ops = &sbus_dma_ops;
658EXPORT_SYMBOL(dma_ops);
659
660
661/*
662 * Return whether the given PCI device DMA address mask can be
663 * supported properly. For example, if your device can only drive the
664 * low 24-bits during PCI bus mastering, then you would pass
665 * 0x00ffffff as the mask to this function.
666 */
667int dma_supported(struct device *dev, u64 mask)
668{
669#ifdef CONFIG_PCI
670 if (dev->bus == &pci_bus_type)
671 return 1;
672#endif
673 return 0;
674}
675EXPORT_SYMBOL(dma_supported);
676
677#ifdef CONFIG_PROC_FS
678
679static int sparc_io_proc_show(struct seq_file *m, void *v)
680{
681 struct resource *root = m->private, *r;
682 const char *nm;
683
684 for (r = root->child; r != NULL; r = r->sibling) {
685 if ((nm = r->name) == 0) nm = "???";
686 seq_printf(m, "%016llx-%016llx: %s\n",
687 (unsigned long long)r->start,
688 (unsigned long long)r->end, nm);
689 }
690
691 return 0;
692}
693
694static int sparc_io_proc_open(struct inode *inode, struct file *file)
695{
696 return single_open(file, sparc_io_proc_show, PDE(inode)->data);
697}
698
699static const struct file_operations sparc_io_proc_fops = {
700 .owner = THIS_MODULE,
701 .open = sparc_io_proc_open,
702 .read = seq_read,
703 .llseek = seq_lseek,
704 .release = single_release,
705};
706#endif /* CONFIG_PROC_FS */
707
708static void register_proc_sparc_ioport(void)
709{
710#ifdef CONFIG_PROC_FS
711 proc_create_data("io_map", 0, NULL, &sparc_io_proc_fops, &sparc_iomap);
712 proc_create_data("dvma_map", 0, NULL, &sparc_io_proc_fops, &_sparc_dvma);
713#endif
714}
1/*
2 * ioport.c: Simple io mapping allocator.
3 *
4 * Copyright (C) 1995 David S. Miller (davem@caip.rutgers.edu)
5 * Copyright (C) 1995 Miguel de Icaza (miguel@nuclecu.unam.mx)
6 *
7 * 1996: sparc_free_io, 1999: ioremap()/iounmap() by Pete Zaitcev.
8 *
9 * 2000/01/29
10 * <rth> zait: as long as pci_alloc_consistent produces something addressable,
11 * things are ok.
12 * <zaitcev> rth: no, it is relevant, because get_free_pages returns you a
13 * pointer into the big page mapping
14 * <rth> zait: so what?
15 * <rth> zait: remap_it_my_way(virt_to_phys(get_free_page()))
16 * <zaitcev> Hmm
17 * <zaitcev> Suppose I did this remap_it_my_way(virt_to_phys(get_free_page())).
18 * So far so good.
19 * <zaitcev> Now, driver calls pci_free_consistent(with result of
20 * remap_it_my_way()).
21 * <zaitcev> How do you find the address to pass to free_pages()?
22 * <rth> zait: walk the page tables? It's only two or three level after all.
23 * <rth> zait: you have to walk them anyway to remove the mapping.
24 * <zaitcev> Hmm
25 * <zaitcev> Sounds reasonable
26 */
27
28#include <linux/module.h>
29#include <linux/sched.h>
30#include <linux/kernel.h>
31#include <linux/errno.h>
32#include <linux/types.h>
33#include <linux/ioport.h>
34#include <linux/mm.h>
35#include <linux/slab.h>
36#include <linux/pci.h> /* struct pci_dev */
37#include <linux/proc_fs.h>
38#include <linux/seq_file.h>
39#include <linux/scatterlist.h>
40#include <linux/of_device.h>
41
42#include <asm/io.h>
43#include <asm/vaddrs.h>
44#include <asm/oplib.h>
45#include <asm/prom.h>
46#include <asm/page.h>
47#include <asm/pgalloc.h>
48#include <asm/dma.h>
49#include <asm/iommu.h>
50#include <asm/io-unit.h>
51#include <asm/leon.h>
52
53const struct sparc32_dma_ops *sparc32_dma_ops;
54
55/* This function must make sure that caches and memory are coherent after DMA
56 * On LEON systems without cache snooping it flushes the entire D-CACHE.
57 */
58static inline void dma_make_coherent(unsigned long pa, unsigned long len)
59{
60 if (sparc_cpu_model == sparc_leon) {
61 if (!sparc_leon3_snooping_enabled())
62 leon_flush_dcache_all();
63 }
64}
65
66static void __iomem *_sparc_ioremap(struct resource *res, u32 bus, u32 pa, int sz);
67static void __iomem *_sparc_alloc_io(unsigned int busno, unsigned long phys,
68 unsigned long size, char *name);
69static void _sparc_free_io(struct resource *res);
70
71static void register_proc_sparc_ioport(void);
72
73/* This points to the next to use virtual memory for DVMA mappings */
74static struct resource _sparc_dvma = {
75 .name = "sparc_dvma", .start = DVMA_VADDR, .end = DVMA_END - 1
76};
77/* This points to the start of I/O mappings, cluable from outside. */
78/*ext*/ struct resource sparc_iomap = {
79 .name = "sparc_iomap", .start = IOBASE_VADDR, .end = IOBASE_END - 1
80};
81
82/*
83 * Our mini-allocator...
84 * Boy this is gross! We need it because we must map I/O for
85 * timers and interrupt controller before the kmalloc is available.
86 */
87
88#define XNMLN 15
89#define XNRES 10 /* SS-10 uses 8 */
90
91struct xresource {
92 struct resource xres; /* Must be first */
93 int xflag; /* 1 == used */
94 char xname[XNMLN+1];
95};
96
97static struct xresource xresv[XNRES];
98
99static struct xresource *xres_alloc(void) {
100 struct xresource *xrp;
101 int n;
102
103 xrp = xresv;
104 for (n = 0; n < XNRES; n++) {
105 if (xrp->xflag == 0) {
106 xrp->xflag = 1;
107 return xrp;
108 }
109 xrp++;
110 }
111 return NULL;
112}
113
114static void xres_free(struct xresource *xrp) {
115 xrp->xflag = 0;
116}
117
118/*
119 * These are typically used in PCI drivers
120 * which are trying to be cross-platform.
121 *
122 * Bus type is always zero on IIep.
123 */
124void __iomem *ioremap(unsigned long offset, unsigned long size)
125{
126 char name[14];
127
128 sprintf(name, "phys_%08x", (u32)offset);
129 return _sparc_alloc_io(0, offset, size, name);
130}
131EXPORT_SYMBOL(ioremap);
132
133/*
134 * Complementary to ioremap().
135 */
136void iounmap(volatile void __iomem *virtual)
137{
138 unsigned long vaddr = (unsigned long) virtual & PAGE_MASK;
139 struct resource *res;
140
141 /*
142 * XXX Too slow. Can have 8192 DVMA pages on sun4m in the worst case.
143 * This probably warrants some sort of hashing.
144 */
145 if ((res = lookup_resource(&sparc_iomap, vaddr)) == NULL) {
146 printk("free_io/iounmap: cannot free %lx\n", vaddr);
147 return;
148 }
149 _sparc_free_io(res);
150
151 if ((char *)res >= (char*)xresv && (char *)res < (char *)&xresv[XNRES]) {
152 xres_free((struct xresource *)res);
153 } else {
154 kfree(res);
155 }
156}
157EXPORT_SYMBOL(iounmap);
158
159void __iomem *of_ioremap(struct resource *res, unsigned long offset,
160 unsigned long size, char *name)
161{
162 return _sparc_alloc_io(res->flags & 0xF,
163 res->start + offset,
164 size, name);
165}
166EXPORT_SYMBOL(of_ioremap);
167
168void of_iounmap(struct resource *res, void __iomem *base, unsigned long size)
169{
170 iounmap(base);
171}
172EXPORT_SYMBOL(of_iounmap);
173
174/*
175 * Meat of mapping
176 */
177static void __iomem *_sparc_alloc_io(unsigned int busno, unsigned long phys,
178 unsigned long size, char *name)
179{
180 static int printed_full;
181 struct xresource *xres;
182 struct resource *res;
183 char *tack;
184 int tlen;
185 void __iomem *va; /* P3 diag */
186
187 if (name == NULL) name = "???";
188
189 if ((xres = xres_alloc()) != NULL) {
190 tack = xres->xname;
191 res = &xres->xres;
192 } else {
193 if (!printed_full) {
194 printk("ioremap: done with statics, switching to malloc\n");
195 printed_full = 1;
196 }
197 tlen = strlen(name);
198 tack = kmalloc(sizeof (struct resource) + tlen + 1, GFP_KERNEL);
199 if (tack == NULL) return NULL;
200 memset(tack, 0, sizeof(struct resource));
201 res = (struct resource *) tack;
202 tack += sizeof (struct resource);
203 }
204
205 strlcpy(tack, name, XNMLN+1);
206 res->name = tack;
207
208 va = _sparc_ioremap(res, busno, phys, size);
209 /* printk("ioremap(0x%x:%08lx[0x%lx])=%p\n", busno, phys, size, va); */ /* P3 diag */
210 return va;
211}
212
213/*
214 */
215static void __iomem *
216_sparc_ioremap(struct resource *res, u32 bus, u32 pa, int sz)
217{
218 unsigned long offset = ((unsigned long) pa) & (~PAGE_MASK);
219
220 if (allocate_resource(&sparc_iomap, res,
221 (offset + sz + PAGE_SIZE-1) & PAGE_MASK,
222 sparc_iomap.start, sparc_iomap.end, PAGE_SIZE, NULL, NULL) != 0) {
223 /* Usually we cannot see printks in this case. */
224 prom_printf("alloc_io_res(%s): cannot occupy\n",
225 (res->name != NULL)? res->name: "???");
226 prom_halt();
227 }
228
229 pa &= PAGE_MASK;
230 srmmu_mapiorange(bus, pa, res->start, resource_size(res));
231
232 return (void __iomem *)(unsigned long)(res->start + offset);
233}
234
235/*
236 * Complementary to _sparc_ioremap().
237 */
238static void _sparc_free_io(struct resource *res)
239{
240 unsigned long plen;
241
242 plen = resource_size(res);
243 BUG_ON((plen & (PAGE_SIZE-1)) != 0);
244 srmmu_unmapiorange(res->start, plen);
245 release_resource(res);
246}
247
248#ifdef CONFIG_SBUS
249
250void sbus_set_sbus64(struct device *dev, int x)
251{
252 printk("sbus_set_sbus64: unsupported\n");
253}
254EXPORT_SYMBOL(sbus_set_sbus64);
255
256/*
257 * Allocate a chunk of memory suitable for DMA.
258 * Typically devices use them for control blocks.
259 * CPU may access them without any explicit flushing.
260 */
261static void *sbus_alloc_coherent(struct device *dev, size_t len,
262 dma_addr_t *dma_addrp, gfp_t gfp,
263 struct dma_attrs *attrs)
264{
265 struct platform_device *op = to_platform_device(dev);
266 unsigned long len_total = PAGE_ALIGN(len);
267 unsigned long va;
268 struct resource *res;
269 int order;
270
271 /* XXX why are some lengths signed, others unsigned? */
272 if (len <= 0) {
273 return NULL;
274 }
275 /* XXX So what is maxphys for us and how do drivers know it? */
276 if (len > 256*1024) { /* __get_free_pages() limit */
277 return NULL;
278 }
279
280 order = get_order(len_total);
281 va = __get_free_pages(gfp, order);
282 if (va == 0)
283 goto err_nopages;
284
285 if ((res = kzalloc(sizeof(struct resource), GFP_KERNEL)) == NULL)
286 goto err_nomem;
287
288 if (allocate_resource(&_sparc_dvma, res, len_total,
289 _sparc_dvma.start, _sparc_dvma.end, PAGE_SIZE, NULL, NULL) != 0) {
290 printk("sbus_alloc_consistent: cannot occupy 0x%lx", len_total);
291 goto err_nova;
292 }
293
294 // XXX The sbus_map_dma_area does this for us below, see comments.
295 // srmmu_mapiorange(0, virt_to_phys(va), res->start, len_total);
296 /*
297 * XXX That's where sdev would be used. Currently we load
298 * all iommu tables with the same translations.
299 */
300 if (sbus_map_dma_area(dev, dma_addrp, va, res->start, len_total) != 0)
301 goto err_noiommu;
302
303 res->name = op->dev.of_node->name;
304
305 return (void *)(unsigned long)res->start;
306
307err_noiommu:
308 release_resource(res);
309err_nova:
310 kfree(res);
311err_nomem:
312 free_pages(va, order);
313err_nopages:
314 return NULL;
315}
316
317static void sbus_free_coherent(struct device *dev, size_t n, void *p,
318 dma_addr_t ba, struct dma_attrs *attrs)
319{
320 struct resource *res;
321 struct page *pgv;
322
323 if ((res = lookup_resource(&_sparc_dvma,
324 (unsigned long)p)) == NULL) {
325 printk("sbus_free_consistent: cannot free %p\n", p);
326 return;
327 }
328
329 if (((unsigned long)p & (PAGE_SIZE-1)) != 0) {
330 printk("sbus_free_consistent: unaligned va %p\n", p);
331 return;
332 }
333
334 n = PAGE_ALIGN(n);
335 if (resource_size(res) != n) {
336 printk("sbus_free_consistent: region 0x%lx asked 0x%zx\n",
337 (long)resource_size(res), n);
338 return;
339 }
340
341 release_resource(res);
342 kfree(res);
343
344 pgv = virt_to_page(p);
345 sbus_unmap_dma_area(dev, ba, n);
346
347 __free_pages(pgv, get_order(n));
348}
349
350/*
351 * Map a chunk of memory so that devices can see it.
352 * CPU view of this memory may be inconsistent with
353 * a device view and explicit flushing is necessary.
354 */
355static dma_addr_t sbus_map_page(struct device *dev, struct page *page,
356 unsigned long offset, size_t len,
357 enum dma_data_direction dir,
358 struct dma_attrs *attrs)
359{
360 void *va = page_address(page) + offset;
361
362 /* XXX why are some lengths signed, others unsigned? */
363 if (len <= 0) {
364 return 0;
365 }
366 /* XXX So what is maxphys for us and how do drivers know it? */
367 if (len > 256*1024) { /* __get_free_pages() limit */
368 return 0;
369 }
370 return mmu_get_scsi_one(dev, va, len);
371}
372
373static void sbus_unmap_page(struct device *dev, dma_addr_t ba, size_t n,
374 enum dma_data_direction dir, struct dma_attrs *attrs)
375{
376 mmu_release_scsi_one(dev, ba, n);
377}
378
379static int sbus_map_sg(struct device *dev, struct scatterlist *sg, int n,
380 enum dma_data_direction dir, struct dma_attrs *attrs)
381{
382 mmu_get_scsi_sgl(dev, sg, n);
383 return n;
384}
385
386static void sbus_unmap_sg(struct device *dev, struct scatterlist *sg, int n,
387 enum dma_data_direction dir, struct dma_attrs *attrs)
388{
389 mmu_release_scsi_sgl(dev, sg, n);
390}
391
392static void sbus_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
393 int n, enum dma_data_direction dir)
394{
395 BUG();
396}
397
398static void sbus_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
399 int n, enum dma_data_direction dir)
400{
401 BUG();
402}
403
404static struct dma_map_ops sbus_dma_ops = {
405 .alloc = sbus_alloc_coherent,
406 .free = sbus_free_coherent,
407 .map_page = sbus_map_page,
408 .unmap_page = sbus_unmap_page,
409 .map_sg = sbus_map_sg,
410 .unmap_sg = sbus_unmap_sg,
411 .sync_sg_for_cpu = sbus_sync_sg_for_cpu,
412 .sync_sg_for_device = sbus_sync_sg_for_device,
413};
414
415static int __init sparc_register_ioport(void)
416{
417 register_proc_sparc_ioport();
418
419 return 0;
420}
421
422arch_initcall(sparc_register_ioport);
423
424#endif /* CONFIG_SBUS */
425
426
427/* Allocate and map kernel buffer using consistent mode DMA for a device.
428 * hwdev should be valid struct pci_dev pointer for PCI devices.
429 */
430static void *pci32_alloc_coherent(struct device *dev, size_t len,
431 dma_addr_t *pba, gfp_t gfp,
432 struct dma_attrs *attrs)
433{
434 unsigned long len_total = PAGE_ALIGN(len);
435 void *va;
436 struct resource *res;
437 int order;
438
439 if (len == 0) {
440 return NULL;
441 }
442 if (len > 256*1024) { /* __get_free_pages() limit */
443 return NULL;
444 }
445
446 order = get_order(len_total);
447 va = (void *) __get_free_pages(gfp, order);
448 if (va == NULL) {
449 printk("pci_alloc_consistent: no %ld pages\n", len_total>>PAGE_SHIFT);
450 goto err_nopages;
451 }
452
453 if ((res = kzalloc(sizeof(struct resource), GFP_KERNEL)) == NULL) {
454 printk("pci_alloc_consistent: no core\n");
455 goto err_nomem;
456 }
457
458 if (allocate_resource(&_sparc_dvma, res, len_total,
459 _sparc_dvma.start, _sparc_dvma.end, PAGE_SIZE, NULL, NULL) != 0) {
460 printk("pci_alloc_consistent: cannot occupy 0x%lx", len_total);
461 goto err_nova;
462 }
463 srmmu_mapiorange(0, virt_to_phys(va), res->start, len_total);
464
465 *pba = virt_to_phys(va); /* equals virt_to_bus (R.I.P.) for us. */
466 return (void *) res->start;
467
468err_nova:
469 kfree(res);
470err_nomem:
471 free_pages((unsigned long)va, order);
472err_nopages:
473 return NULL;
474}
475
476/* Free and unmap a consistent DMA buffer.
477 * cpu_addr is what was returned from pci_alloc_consistent,
478 * size must be the same as what as passed into pci_alloc_consistent,
479 * and likewise dma_addr must be the same as what *dma_addrp was set to.
480 *
481 * References to the memory and mappings associated with cpu_addr/dma_addr
482 * past this call are illegal.
483 */
484static void pci32_free_coherent(struct device *dev, size_t n, void *p,
485 dma_addr_t ba, struct dma_attrs *attrs)
486{
487 struct resource *res;
488
489 if ((res = lookup_resource(&_sparc_dvma,
490 (unsigned long)p)) == NULL) {
491 printk("pci_free_consistent: cannot free %p\n", p);
492 return;
493 }
494
495 if (((unsigned long)p & (PAGE_SIZE-1)) != 0) {
496 printk("pci_free_consistent: unaligned va %p\n", p);
497 return;
498 }
499
500 n = PAGE_ALIGN(n);
501 if (resource_size(res) != n) {
502 printk("pci_free_consistent: region 0x%lx asked 0x%lx\n",
503 (long)resource_size(res), (long)n);
504 return;
505 }
506
507 dma_make_coherent(ba, n);
508 srmmu_unmapiorange((unsigned long)p, n);
509
510 release_resource(res);
511 kfree(res);
512 free_pages((unsigned long)phys_to_virt(ba), get_order(n));
513}
514
515/*
516 * Same as pci_map_single, but with pages.
517 */
518static dma_addr_t pci32_map_page(struct device *dev, struct page *page,
519 unsigned long offset, size_t size,
520 enum dma_data_direction dir,
521 struct dma_attrs *attrs)
522{
523 /* IIep is write-through, not flushing. */
524 return page_to_phys(page) + offset;
525}
526
527static void pci32_unmap_page(struct device *dev, dma_addr_t ba, size_t size,
528 enum dma_data_direction dir, struct dma_attrs *attrs)
529{
530 if (dir != PCI_DMA_TODEVICE)
531 dma_make_coherent(ba, PAGE_ALIGN(size));
532}
533
534/* Map a set of buffers described by scatterlist in streaming
535 * mode for DMA. This is the scatter-gather version of the
536 * above pci_map_single interface. Here the scatter gather list
537 * elements are each tagged with the appropriate dma address
538 * and length. They are obtained via sg_dma_{address,length}(SG).
539 *
540 * NOTE: An implementation may be able to use a smaller number of
541 * DMA address/length pairs than there are SG table elements.
542 * (for example via virtual mapping capabilities)
543 * The routine returns the number of addr/length pairs actually
544 * used, at most nents.
545 *
546 * Device ownership issues as mentioned above for pci_map_single are
547 * the same here.
548 */
549static int pci32_map_sg(struct device *device, struct scatterlist *sgl,
550 int nents, enum dma_data_direction dir,
551 struct dma_attrs *attrs)
552{
553 struct scatterlist *sg;
554 int n;
555
556 /* IIep is write-through, not flushing. */
557 for_each_sg(sgl, sg, nents, n) {
558 sg->dma_address = sg_phys(sg);
559 sg->dma_length = sg->length;
560 }
561 return nents;
562}
563
564/* Unmap a set of streaming mode DMA translations.
565 * Again, cpu read rules concerning calls here are the same as for
566 * pci_unmap_single() above.
567 */
568static void pci32_unmap_sg(struct device *dev, struct scatterlist *sgl,
569 int nents, enum dma_data_direction dir,
570 struct dma_attrs *attrs)
571{
572 struct scatterlist *sg;
573 int n;
574
575 if (dir != PCI_DMA_TODEVICE) {
576 for_each_sg(sgl, sg, nents, n) {
577 dma_make_coherent(sg_phys(sg), PAGE_ALIGN(sg->length));
578 }
579 }
580}
581
582/* Make physical memory consistent for a single
583 * streaming mode DMA translation before or after a transfer.
584 *
585 * If you perform a pci_map_single() but wish to interrogate the
586 * buffer using the cpu, yet do not wish to teardown the PCI dma
587 * mapping, you must call this function before doing so. At the
588 * next point you give the PCI dma address back to the card, you
589 * must first perform a pci_dma_sync_for_device, and then the
590 * device again owns the buffer.
591 */
592static void pci32_sync_single_for_cpu(struct device *dev, dma_addr_t ba,
593 size_t size, enum dma_data_direction dir)
594{
595 if (dir != PCI_DMA_TODEVICE) {
596 dma_make_coherent(ba, PAGE_ALIGN(size));
597 }
598}
599
600static void pci32_sync_single_for_device(struct device *dev, dma_addr_t ba,
601 size_t size, enum dma_data_direction dir)
602{
603 if (dir != PCI_DMA_TODEVICE) {
604 dma_make_coherent(ba, PAGE_ALIGN(size));
605 }
606}
607
608/* Make physical memory consistent for a set of streaming
609 * mode DMA translations after a transfer.
610 *
611 * The same as pci_dma_sync_single_* but for a scatter-gather list,
612 * same rules and usage.
613 */
614static void pci32_sync_sg_for_cpu(struct device *dev, struct scatterlist *sgl,
615 int nents, enum dma_data_direction dir)
616{
617 struct scatterlist *sg;
618 int n;
619
620 if (dir != PCI_DMA_TODEVICE) {
621 for_each_sg(sgl, sg, nents, n) {
622 dma_make_coherent(sg_phys(sg), PAGE_ALIGN(sg->length));
623 }
624 }
625}
626
627static void pci32_sync_sg_for_device(struct device *device, struct scatterlist *sgl,
628 int nents, enum dma_data_direction dir)
629{
630 struct scatterlist *sg;
631 int n;
632
633 if (dir != PCI_DMA_TODEVICE) {
634 for_each_sg(sgl, sg, nents, n) {
635 dma_make_coherent(sg_phys(sg), PAGE_ALIGN(sg->length));
636 }
637 }
638}
639
640struct dma_map_ops pci32_dma_ops = {
641 .alloc = pci32_alloc_coherent,
642 .free = pci32_free_coherent,
643 .map_page = pci32_map_page,
644 .unmap_page = pci32_unmap_page,
645 .map_sg = pci32_map_sg,
646 .unmap_sg = pci32_unmap_sg,
647 .sync_single_for_cpu = pci32_sync_single_for_cpu,
648 .sync_single_for_device = pci32_sync_single_for_device,
649 .sync_sg_for_cpu = pci32_sync_sg_for_cpu,
650 .sync_sg_for_device = pci32_sync_sg_for_device,
651};
652EXPORT_SYMBOL(pci32_dma_ops);
653
654/* leon re-uses pci32_dma_ops */
655struct dma_map_ops *leon_dma_ops = &pci32_dma_ops;
656EXPORT_SYMBOL(leon_dma_ops);
657
658struct dma_map_ops *dma_ops = &sbus_dma_ops;
659EXPORT_SYMBOL(dma_ops);
660
661
662/*
663 * Return whether the given PCI device DMA address mask can be
664 * supported properly. For example, if your device can only drive the
665 * low 24-bits during PCI bus mastering, then you would pass
666 * 0x00ffffff as the mask to this function.
667 */
668int dma_supported(struct device *dev, u64 mask)
669{
670 if (dev_is_pci(dev))
671 return 1;
672
673 return 0;
674}
675EXPORT_SYMBOL(dma_supported);
676
677#ifdef CONFIG_PROC_FS
678
679static int sparc_io_proc_show(struct seq_file *m, void *v)
680{
681 struct resource *root = m->private, *r;
682 const char *nm;
683
684 for (r = root->child; r != NULL; r = r->sibling) {
685 if ((nm = r->name) == NULL) nm = "???";
686 seq_printf(m, "%016llx-%016llx: %s\n",
687 (unsigned long long)r->start,
688 (unsigned long long)r->end, nm);
689 }
690
691 return 0;
692}
693
694static int sparc_io_proc_open(struct inode *inode, struct file *file)
695{
696 return single_open(file, sparc_io_proc_show, PDE_DATA(inode));
697}
698
699static const struct file_operations sparc_io_proc_fops = {
700 .owner = THIS_MODULE,
701 .open = sparc_io_proc_open,
702 .read = seq_read,
703 .llseek = seq_lseek,
704 .release = single_release,
705};
706#endif /* CONFIG_PROC_FS */
707
708static void register_proc_sparc_ioport(void)
709{
710#ifdef CONFIG_PROC_FS
711 proc_create_data("io_map", 0, NULL, &sparc_io_proc_fops, &sparc_iomap);
712 proc_create_data("dvma_map", 0, NULL, &sparc_io_proc_fops, &_sparc_dvma);
713#endif
714}