Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * ioport.c: Simple io mapping allocator.
4 *
5 * Copyright (C) 1995 David S. Miller (davem@caip.rutgers.edu)
6 * Copyright (C) 1995 Miguel de Icaza (miguel@nuclecu.unam.mx)
7 *
8 * 1996: sparc_free_io, 1999: ioremap()/iounmap() by Pete Zaitcev.
9 *
10 * 2000/01/29
11 * <rth> zait: as long as pci_alloc_consistent produces something addressable,
12 * things are ok.
13 * <zaitcev> rth: no, it is relevant, because get_free_pages returns you a
14 * pointer into the big page mapping
15 * <rth> zait: so what?
16 * <rth> zait: remap_it_my_way(virt_to_phys(get_free_page()))
17 * <zaitcev> Hmm
18 * <zaitcev> Suppose I did this remap_it_my_way(virt_to_phys(get_free_page())).
19 * So far so good.
20 * <zaitcev> Now, driver calls pci_free_consistent(with result of
21 * remap_it_my_way()).
22 * <zaitcev> How do you find the address to pass to free_pages()?
23 * <rth> zait: walk the page tables? It's only two or three level after all.
24 * <rth> zait: you have to walk them anyway to remove the mapping.
25 * <zaitcev> Hmm
26 * <zaitcev> Sounds reasonable
27 */
28
29#include <linux/module.h>
30#include <linux/sched.h>
31#include <linux/kernel.h>
32#include <linux/errno.h>
33#include <linux/types.h>
34#include <linux/ioport.h>
35#include <linux/mm.h>
36#include <linux/slab.h>
37#include <linux/pci.h> /* struct pci_dev */
38#include <linux/proc_fs.h>
39#include <linux/seq_file.h>
40#include <linux/scatterlist.h>
41#include <linux/of_device.h>
42
43#include <asm/io.h>
44#include <asm/vaddrs.h>
45#include <asm/oplib.h>
46#include <asm/prom.h>
47#include <asm/page.h>
48#include <asm/pgalloc.h>
49#include <asm/dma.h>
50#include <asm/iommu.h>
51#include <asm/io-unit.h>
52#include <asm/leon.h>
53
54const struct sparc32_dma_ops *sparc32_dma_ops;
55
56/* This function must make sure that caches and memory are coherent after DMA
57 * On LEON systems without cache snooping it flushes the entire D-CACHE.
58 */
59static inline void dma_make_coherent(unsigned long pa, unsigned long len)
60{
61 if (sparc_cpu_model == sparc_leon) {
62 if (!sparc_leon3_snooping_enabled())
63 leon_flush_dcache_all();
64 }
65}
66
67static void __iomem *_sparc_ioremap(struct resource *res, u32 bus, u32 pa, int sz);
68static void __iomem *_sparc_alloc_io(unsigned int busno, unsigned long phys,
69 unsigned long size, char *name);
70static void _sparc_free_io(struct resource *res);
71
72static void register_proc_sparc_ioport(void);
73
74/* This points to the next to use virtual memory for DVMA mappings */
75static struct resource _sparc_dvma = {
76 .name = "sparc_dvma", .start = DVMA_VADDR, .end = DVMA_END - 1
77};
78/* This points to the start of I/O mappings, cluable from outside. */
79/*ext*/ struct resource sparc_iomap = {
80 .name = "sparc_iomap", .start = IOBASE_VADDR, .end = IOBASE_END - 1
81};
82
83/*
84 * Our mini-allocator...
85 * Boy this is gross! We need it because we must map I/O for
86 * timers and interrupt controller before the kmalloc is available.
87 */
88
89#define XNMLN 15
90#define XNRES 10 /* SS-10 uses 8 */
91
92struct xresource {
93 struct resource xres; /* Must be first */
94 int xflag; /* 1 == used */
95 char xname[XNMLN+1];
96};
97
98static struct xresource xresv[XNRES];
99
100static struct xresource *xres_alloc(void) {
101 struct xresource *xrp;
102 int n;
103
104 xrp = xresv;
105 for (n = 0; n < XNRES; n++) {
106 if (xrp->xflag == 0) {
107 xrp->xflag = 1;
108 return xrp;
109 }
110 xrp++;
111 }
112 return NULL;
113}
114
115static void xres_free(struct xresource *xrp) {
116 xrp->xflag = 0;
117}
118
119/*
120 * These are typically used in PCI drivers
121 * which are trying to be cross-platform.
122 *
123 * Bus type is always zero on IIep.
124 */
125void __iomem *ioremap(phys_addr_t offset, size_t size)
126{
127 char name[14];
128
129 sprintf(name, "phys_%08x", (u32)offset);
130 return _sparc_alloc_io(0, (unsigned long)offset, size, name);
131}
132EXPORT_SYMBOL(ioremap);
133
134/*
135 * Complementary to ioremap().
136 */
137void iounmap(volatile void __iomem *virtual)
138{
139 unsigned long vaddr = (unsigned long) virtual & PAGE_MASK;
140 struct resource *res;
141
142 /*
143 * XXX Too slow. Can have 8192 DVMA pages on sun4m in the worst case.
144 * This probably warrants some sort of hashing.
145 */
146 if ((res = lookup_resource(&sparc_iomap, vaddr)) == NULL) {
147 printk("free_io/iounmap: cannot free %lx\n", vaddr);
148 return;
149 }
150 _sparc_free_io(res);
151
152 if ((char *)res >= (char*)xresv && (char *)res < (char *)&xresv[XNRES]) {
153 xres_free((struct xresource *)res);
154 } else {
155 kfree(res);
156 }
157}
158EXPORT_SYMBOL(iounmap);
159
160void __iomem *of_ioremap(struct resource *res, unsigned long offset,
161 unsigned long size, char *name)
162{
163 return _sparc_alloc_io(res->flags & 0xF,
164 res->start + offset,
165 size, name);
166}
167EXPORT_SYMBOL(of_ioremap);
168
169void of_iounmap(struct resource *res, void __iomem *base, unsigned long size)
170{
171 iounmap(base);
172}
173EXPORT_SYMBOL(of_iounmap);
174
175/*
176 * Meat of mapping
177 */
178static void __iomem *_sparc_alloc_io(unsigned int busno, unsigned long phys,
179 unsigned long size, char *name)
180{
181 static int printed_full;
182 struct xresource *xres;
183 struct resource *res;
184 char *tack;
185 int tlen;
186 void __iomem *va; /* P3 diag */
187
188 if (name == NULL) name = "???";
189
190 if ((xres = xres_alloc()) != NULL) {
191 tack = xres->xname;
192 res = &xres->xres;
193 } else {
194 if (!printed_full) {
195 printk("ioremap: done with statics, switching to malloc\n");
196 printed_full = 1;
197 }
198 tlen = strlen(name);
199 tack = kmalloc(sizeof (struct resource) + tlen + 1, GFP_KERNEL);
200 if (tack == NULL) return NULL;
201 memset(tack, 0, sizeof(struct resource));
202 res = (struct resource *) tack;
203 tack += sizeof (struct resource);
204 }
205
206 strlcpy(tack, name, XNMLN+1);
207 res->name = tack;
208
209 va = _sparc_ioremap(res, busno, phys, size);
210 /* printk("ioremap(0x%x:%08lx[0x%lx])=%p\n", busno, phys, size, va); */ /* P3 diag */
211 return va;
212}
213
214/*
215 */
216static void __iomem *
217_sparc_ioremap(struct resource *res, u32 bus, u32 pa, int sz)
218{
219 unsigned long offset = ((unsigned long) pa) & (~PAGE_MASK);
220
221 if (allocate_resource(&sparc_iomap, res,
222 (offset + sz + PAGE_SIZE-1) & PAGE_MASK,
223 sparc_iomap.start, sparc_iomap.end, PAGE_SIZE, NULL, NULL) != 0) {
224 /* Usually we cannot see printks in this case. */
225 prom_printf("alloc_io_res(%s): cannot occupy\n",
226 (res->name != NULL)? res->name: "???");
227 prom_halt();
228 }
229
230 pa &= PAGE_MASK;
231 srmmu_mapiorange(bus, pa, res->start, resource_size(res));
232
233 return (void __iomem *)(unsigned long)(res->start + offset);
234}
235
236/*
237 * Complementary to _sparc_ioremap().
238 */
239static void _sparc_free_io(struct resource *res)
240{
241 unsigned long plen;
242
243 plen = resource_size(res);
244 BUG_ON((plen & (PAGE_SIZE-1)) != 0);
245 srmmu_unmapiorange(res->start, plen);
246 release_resource(res);
247}
248
249#ifdef CONFIG_SBUS
250
251void sbus_set_sbus64(struct device *dev, int x)
252{
253 printk("sbus_set_sbus64: unsupported\n");
254}
255EXPORT_SYMBOL(sbus_set_sbus64);
256
257/*
258 * Allocate a chunk of memory suitable for DMA.
259 * Typically devices use them for control blocks.
260 * CPU may access them without any explicit flushing.
261 */
262static void *sbus_alloc_coherent(struct device *dev, size_t len,
263 dma_addr_t *dma_addrp, gfp_t gfp,
264 unsigned long attrs)
265{
266 struct platform_device *op = to_platform_device(dev);
267 unsigned long len_total = PAGE_ALIGN(len);
268 unsigned long va;
269 struct resource *res;
270 int order;
271
272 /* XXX why are some lengths signed, others unsigned? */
273 if (len <= 0) {
274 return NULL;
275 }
276 /* XXX So what is maxphys for us and how do drivers know it? */
277 if (len > 256*1024) { /* __get_free_pages() limit */
278 return NULL;
279 }
280
281 order = get_order(len_total);
282 va = __get_free_pages(gfp, order);
283 if (va == 0)
284 goto err_nopages;
285
286 if ((res = kzalloc(sizeof(struct resource), GFP_KERNEL)) == NULL)
287 goto err_nomem;
288
289 if (allocate_resource(&_sparc_dvma, res, len_total,
290 _sparc_dvma.start, _sparc_dvma.end, PAGE_SIZE, NULL, NULL) != 0) {
291 printk("sbus_alloc_consistent: cannot occupy 0x%lx", len_total);
292 goto err_nova;
293 }
294
295 // XXX The sbus_map_dma_area does this for us below, see comments.
296 // srmmu_mapiorange(0, virt_to_phys(va), res->start, len_total);
297 /*
298 * XXX That's where sdev would be used. Currently we load
299 * all iommu tables with the same translations.
300 */
301 if (sbus_map_dma_area(dev, dma_addrp, va, res->start, len_total) != 0)
302 goto err_noiommu;
303
304 res->name = op->dev.of_node->name;
305
306 return (void *)(unsigned long)res->start;
307
308err_noiommu:
309 release_resource(res);
310err_nova:
311 kfree(res);
312err_nomem:
313 free_pages(va, order);
314err_nopages:
315 return NULL;
316}
317
318static void sbus_free_coherent(struct device *dev, size_t n, void *p,
319 dma_addr_t ba, unsigned long attrs)
320{
321 struct resource *res;
322 struct page *pgv;
323
324 if ((res = lookup_resource(&_sparc_dvma,
325 (unsigned long)p)) == NULL) {
326 printk("sbus_free_consistent: cannot free %p\n", p);
327 return;
328 }
329
330 if (((unsigned long)p & (PAGE_SIZE-1)) != 0) {
331 printk("sbus_free_consistent: unaligned va %p\n", p);
332 return;
333 }
334
335 n = PAGE_ALIGN(n);
336 if (resource_size(res) != n) {
337 printk("sbus_free_consistent: region 0x%lx asked 0x%zx\n",
338 (long)resource_size(res), n);
339 return;
340 }
341
342 release_resource(res);
343 kfree(res);
344
345 pgv = virt_to_page(p);
346 sbus_unmap_dma_area(dev, ba, n);
347
348 __free_pages(pgv, get_order(n));
349}
350
351/*
352 * Map a chunk of memory so that devices can see it.
353 * CPU view of this memory may be inconsistent with
354 * a device view and explicit flushing is necessary.
355 */
356static dma_addr_t sbus_map_page(struct device *dev, struct page *page,
357 unsigned long offset, size_t len,
358 enum dma_data_direction dir,
359 unsigned long attrs)
360{
361 void *va = page_address(page) + offset;
362
363 /* XXX why are some lengths signed, others unsigned? */
364 if (len <= 0) {
365 return 0;
366 }
367 /* XXX So what is maxphys for us and how do drivers know it? */
368 if (len > 256*1024) { /* __get_free_pages() limit */
369 return 0;
370 }
371 return mmu_get_scsi_one(dev, va, len);
372}
373
374static void sbus_unmap_page(struct device *dev, dma_addr_t ba, size_t n,
375 enum dma_data_direction dir, unsigned long attrs)
376{
377 mmu_release_scsi_one(dev, ba, n);
378}
379
380static int sbus_map_sg(struct device *dev, struct scatterlist *sg, int n,
381 enum dma_data_direction dir, unsigned long attrs)
382{
383 mmu_get_scsi_sgl(dev, sg, n);
384 return n;
385}
386
387static void sbus_unmap_sg(struct device *dev, struct scatterlist *sg, int n,
388 enum dma_data_direction dir, unsigned long attrs)
389{
390 mmu_release_scsi_sgl(dev, sg, n);
391}
392
393static void sbus_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
394 int n, enum dma_data_direction dir)
395{
396 BUG();
397}
398
399static void sbus_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
400 int n, enum dma_data_direction dir)
401{
402 BUG();
403}
404
405static int sbus_dma_supported(struct device *dev, u64 mask)
406{
407 return 0;
408}
409
410static const struct dma_map_ops sbus_dma_ops = {
411 .alloc = sbus_alloc_coherent,
412 .free = sbus_free_coherent,
413 .map_page = sbus_map_page,
414 .unmap_page = sbus_unmap_page,
415 .map_sg = sbus_map_sg,
416 .unmap_sg = sbus_unmap_sg,
417 .sync_sg_for_cpu = sbus_sync_sg_for_cpu,
418 .sync_sg_for_device = sbus_sync_sg_for_device,
419 .dma_supported = sbus_dma_supported,
420};
421
422static int __init sparc_register_ioport(void)
423{
424 register_proc_sparc_ioport();
425
426 return 0;
427}
428
429arch_initcall(sparc_register_ioport);
430
431#endif /* CONFIG_SBUS */
432
433
434/* Allocate and map kernel buffer using consistent mode DMA for a device.
435 * hwdev should be valid struct pci_dev pointer for PCI devices.
436 */
437static void *pci32_alloc_coherent(struct device *dev, size_t len,
438 dma_addr_t *pba, gfp_t gfp,
439 unsigned long attrs)
440{
441 unsigned long len_total = PAGE_ALIGN(len);
442 void *va;
443 struct resource *res;
444 int order;
445
446 if (len == 0) {
447 return NULL;
448 }
449 if (len > 256*1024) { /* __get_free_pages() limit */
450 return NULL;
451 }
452
453 order = get_order(len_total);
454 va = (void *) __get_free_pages(gfp, order);
455 if (va == NULL) {
456 printk("pci_alloc_consistent: no %ld pages\n", len_total>>PAGE_SHIFT);
457 goto err_nopages;
458 }
459
460 if ((res = kzalloc(sizeof(struct resource), GFP_KERNEL)) == NULL) {
461 printk("pci_alloc_consistent: no core\n");
462 goto err_nomem;
463 }
464
465 if (allocate_resource(&_sparc_dvma, res, len_total,
466 _sparc_dvma.start, _sparc_dvma.end, PAGE_SIZE, NULL, NULL) != 0) {
467 printk("pci_alloc_consistent: cannot occupy 0x%lx", len_total);
468 goto err_nova;
469 }
470 srmmu_mapiorange(0, virt_to_phys(va), res->start, len_total);
471
472 *pba = virt_to_phys(va); /* equals virt_to_bus (R.I.P.) for us. */
473 return (void *) res->start;
474
475err_nova:
476 kfree(res);
477err_nomem:
478 free_pages((unsigned long)va, order);
479err_nopages:
480 return NULL;
481}
482
483/* Free and unmap a consistent DMA buffer.
484 * cpu_addr is what was returned from pci_alloc_consistent,
485 * size must be the same as what as passed into pci_alloc_consistent,
486 * and likewise dma_addr must be the same as what *dma_addrp was set to.
487 *
488 * References to the memory and mappings associated with cpu_addr/dma_addr
489 * past this call are illegal.
490 */
491static void pci32_free_coherent(struct device *dev, size_t n, void *p,
492 dma_addr_t ba, unsigned long attrs)
493{
494 struct resource *res;
495
496 if ((res = lookup_resource(&_sparc_dvma,
497 (unsigned long)p)) == NULL) {
498 printk("pci_free_consistent: cannot free %p\n", p);
499 return;
500 }
501
502 if (((unsigned long)p & (PAGE_SIZE-1)) != 0) {
503 printk("pci_free_consistent: unaligned va %p\n", p);
504 return;
505 }
506
507 n = PAGE_ALIGN(n);
508 if (resource_size(res) != n) {
509 printk("pci_free_consistent: region 0x%lx asked 0x%lx\n",
510 (long)resource_size(res), (long)n);
511 return;
512 }
513
514 dma_make_coherent(ba, n);
515 srmmu_unmapiorange((unsigned long)p, n);
516
517 release_resource(res);
518 kfree(res);
519 free_pages((unsigned long)phys_to_virt(ba), get_order(n));
520}
521
522/*
523 * Same as pci_map_single, but with pages.
524 */
525static dma_addr_t pci32_map_page(struct device *dev, struct page *page,
526 unsigned long offset, size_t size,
527 enum dma_data_direction dir,
528 unsigned long attrs)
529{
530 /* IIep is write-through, not flushing. */
531 return page_to_phys(page) + offset;
532}
533
534static void pci32_unmap_page(struct device *dev, dma_addr_t ba, size_t size,
535 enum dma_data_direction dir, unsigned long attrs)
536{
537 if (dir != PCI_DMA_TODEVICE && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
538 dma_make_coherent(ba, PAGE_ALIGN(size));
539}
540
541/* Map a set of buffers described by scatterlist in streaming
542 * mode for DMA. This is the scatter-gather version of the
543 * above pci_map_single interface. Here the scatter gather list
544 * elements are each tagged with the appropriate dma address
545 * and length. They are obtained via sg_dma_{address,length}(SG).
546 *
547 * NOTE: An implementation may be able to use a smaller number of
548 * DMA address/length pairs than there are SG table elements.
549 * (for example via virtual mapping capabilities)
550 * The routine returns the number of addr/length pairs actually
551 * used, at most nents.
552 *
553 * Device ownership issues as mentioned above for pci_map_single are
554 * the same here.
555 */
556static int pci32_map_sg(struct device *device, struct scatterlist *sgl,
557 int nents, enum dma_data_direction dir,
558 unsigned long attrs)
559{
560 struct scatterlist *sg;
561 int n;
562
563 /* IIep is write-through, not flushing. */
564 for_each_sg(sgl, sg, nents, n) {
565 sg->dma_address = sg_phys(sg);
566 sg->dma_length = sg->length;
567 }
568 return nents;
569}
570
571/* Unmap a set of streaming mode DMA translations.
572 * Again, cpu read rules concerning calls here are the same as for
573 * pci_unmap_single() above.
574 */
575static void pci32_unmap_sg(struct device *dev, struct scatterlist *sgl,
576 int nents, enum dma_data_direction dir,
577 unsigned long attrs)
578{
579 struct scatterlist *sg;
580 int n;
581
582 if (dir != PCI_DMA_TODEVICE && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
583 for_each_sg(sgl, sg, nents, n) {
584 dma_make_coherent(sg_phys(sg), PAGE_ALIGN(sg->length));
585 }
586 }
587}
588
589/* Make physical memory consistent for a single
590 * streaming mode DMA translation before or after a transfer.
591 *
592 * If you perform a pci_map_single() but wish to interrogate the
593 * buffer using the cpu, yet do not wish to teardown the PCI dma
594 * mapping, you must call this function before doing so. At the
595 * next point you give the PCI dma address back to the card, you
596 * must first perform a pci_dma_sync_for_device, and then the
597 * device again owns the buffer.
598 */
599static void pci32_sync_single_for_cpu(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
607static void pci32_sync_single_for_device(struct device *dev, dma_addr_t ba,
608 size_t size, enum dma_data_direction dir)
609{
610 if (dir != PCI_DMA_TODEVICE) {
611 dma_make_coherent(ba, PAGE_ALIGN(size));
612 }
613}
614
615/* Make physical memory consistent for a set of streaming
616 * mode DMA translations after a transfer.
617 *
618 * The same as pci_dma_sync_single_* but for a scatter-gather list,
619 * same rules and usage.
620 */
621static void pci32_sync_sg_for_cpu(struct device *dev, struct scatterlist *sgl,
622 int nents, enum dma_data_direction dir)
623{
624 struct scatterlist *sg;
625 int n;
626
627 if (dir != PCI_DMA_TODEVICE) {
628 for_each_sg(sgl, sg, nents, n) {
629 dma_make_coherent(sg_phys(sg), PAGE_ALIGN(sg->length));
630 }
631 }
632}
633
634static void pci32_sync_sg_for_device(struct device *device, struct scatterlist *sgl,
635 int nents, enum dma_data_direction dir)
636{
637 struct scatterlist *sg;
638 int n;
639
640 if (dir != PCI_DMA_TODEVICE) {
641 for_each_sg(sgl, sg, nents, n) {
642 dma_make_coherent(sg_phys(sg), PAGE_ALIGN(sg->length));
643 }
644 }
645}
646
647/* note: leon re-uses pci32_dma_ops */
648const struct dma_map_ops pci32_dma_ops = {
649 .alloc = pci32_alloc_coherent,
650 .free = pci32_free_coherent,
651 .map_page = pci32_map_page,
652 .unmap_page = pci32_unmap_page,
653 .map_sg = pci32_map_sg,
654 .unmap_sg = pci32_unmap_sg,
655 .sync_single_for_cpu = pci32_sync_single_for_cpu,
656 .sync_single_for_device = pci32_sync_single_for_device,
657 .sync_sg_for_cpu = pci32_sync_sg_for_cpu,
658 .sync_sg_for_device = pci32_sync_sg_for_device,
659};
660EXPORT_SYMBOL(pci32_dma_ops);
661
662const struct dma_map_ops *dma_ops = &sbus_dma_ops;
663EXPORT_SYMBOL(dma_ops);
664
665#ifdef CONFIG_PROC_FS
666
667static int sparc_io_proc_show(struct seq_file *m, void *v)
668{
669 struct resource *root = m->private, *r;
670 const char *nm;
671
672 for (r = root->child; r != NULL; r = r->sibling) {
673 if ((nm = r->name) == NULL) nm = "???";
674 seq_printf(m, "%016llx-%016llx: %s\n",
675 (unsigned long long)r->start,
676 (unsigned long long)r->end, nm);
677 }
678
679 return 0;
680}
681
682static int sparc_io_proc_open(struct inode *inode, struct file *file)
683{
684 return single_open(file, sparc_io_proc_show, PDE_DATA(inode));
685}
686
687static const struct file_operations sparc_io_proc_fops = {
688 .owner = THIS_MODULE,
689 .open = sparc_io_proc_open,
690 .read = seq_read,
691 .llseek = seq_lseek,
692 .release = single_release,
693};
694#endif /* CONFIG_PROC_FS */
695
696static void register_proc_sparc_ioport(void)
697{
698#ifdef CONFIG_PROC_FS
699 proc_create_data("io_map", 0, NULL, &sparc_io_proc_fops, &sparc_iomap);
700 proc_create_data("dvma_map", 0, NULL, &sparc_io_proc_fops, &_sparc_dvma);
701#endif
702}
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}