2 * ioport.c: Simple io mapping allocator.
4 * Copyright (C) 1995 David S. Miller (davem@caip.rutgers.edu)
5 * Copyright (C) 1995 Miguel de Icaza (miguel@nuclecu.unam.mx)
7 * 1996: sparc_free_io, 1999: ioremap()/iounmap() by Pete Zaitcev.
10 * <rth> zait: as long as pci_alloc_consistent produces something addressable,
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()))
17 * <zaitcev> Suppose I did this remap_it_my_way(virt_to_phys(get_free_page())).
19 * <zaitcev> Now, driver calls pci_free_consistent(with result of
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.
25 * <zaitcev> Sounds reasonable
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>
35 #include <linux/slab.h>
36 #include <linux/pci.h> /* struct pci_dev */
37 #include <linux/proc_fs.h>
38 #include <linux/scatterlist.h>
39 #include <linux/of_device.h>
42 #include <asm/vaddrs.h>
43 #include <asm/oplib.h>
46 #include <asm/pgalloc.h>
48 #include <asm/iommu.h>
49 #include <asm/io-unit.h>
51 #define mmu_inval_dma_area(p, l) /* Anton pulled it out for 2.4.0-xx */
53 static struct resource *_sparc_find_resource(struct resource *r,
56 static void __iomem *_sparc_ioremap(struct resource *res, u32 bus, u32 pa, int sz);
57 static void __iomem *_sparc_alloc_io(unsigned int busno, unsigned long phys,
58 unsigned long size, char *name);
59 static void _sparc_free_io(struct resource *res);
61 static void register_proc_sparc_ioport(void);
63 /* This points to the next to use virtual memory for DVMA mappings */
64 static struct resource _sparc_dvma = {
65 .name = "sparc_dvma", .start = DVMA_VADDR, .end = DVMA_END - 1
67 /* This points to the start of I/O mappings, cluable from outside. */
68 /*ext*/ struct resource sparc_iomap = {
69 .name = "sparc_iomap", .start = IOBASE_VADDR, .end = IOBASE_END - 1
73 * Our mini-allocator...
74 * Boy this is gross! We need it because we must map I/O for
75 * timers and interrupt controller before the kmalloc is available.
79 #define XNRES 10 /* SS-10 uses 8 */
82 struct resource xres; /* Must be first */
83 int xflag; /* 1 == used */
87 static struct xresource xresv[XNRES];
89 static struct xresource *xres_alloc(void) {
90 struct xresource *xrp;
94 for (n = 0; n < XNRES; n++) {
95 if (xrp->xflag == 0) {
104 static void xres_free(struct xresource *xrp) {
109 * These are typically used in PCI drivers
110 * which are trying to be cross-platform.
112 * Bus type is always zero on IIep.
114 void __iomem *ioremap(unsigned long offset, unsigned long size)
118 sprintf(name, "phys_%08x", (u32)offset);
119 return _sparc_alloc_io(0, offset, size, name);
121 EXPORT_SYMBOL(ioremap);
124 * Comlimentary to ioremap().
126 void iounmap(volatile void __iomem *virtual)
128 unsigned long vaddr = (unsigned long) virtual & PAGE_MASK;
129 struct resource *res;
131 if ((res = _sparc_find_resource(&sparc_iomap, vaddr)) == NULL) {
132 printk("free_io/iounmap: cannot free %lx\n", vaddr);
137 if ((char *)res >= (char*)xresv && (char *)res < (char *)&xresv[XNRES]) {
138 xres_free((struct xresource *)res);
143 EXPORT_SYMBOL(iounmap);
145 void __iomem *of_ioremap(struct resource *res, unsigned long offset,
146 unsigned long size, char *name)
148 return _sparc_alloc_io(res->flags & 0xF,
152 EXPORT_SYMBOL(of_ioremap);
154 void of_iounmap(struct resource *res, void __iomem *base, unsigned long size)
158 EXPORT_SYMBOL(of_iounmap);
163 static void __iomem *_sparc_alloc_io(unsigned int busno, unsigned long phys,
164 unsigned long size, char *name)
166 static int printed_full;
167 struct xresource *xres;
168 struct resource *res;
171 void __iomem *va; /* P3 diag */
173 if (name == NULL) name = "???";
175 if ((xres = xres_alloc()) != 0) {
180 printk("ioremap: done with statics, switching to malloc\n");
184 tack = kmalloc(sizeof (struct resource) + tlen + 1, GFP_KERNEL);
185 if (tack == NULL) return NULL;
186 memset(tack, 0, sizeof(struct resource));
187 res = (struct resource *) tack;
188 tack += sizeof (struct resource);
191 strlcpy(tack, name, XNMLN+1);
194 va = _sparc_ioremap(res, busno, phys, size);
195 /* printk("ioremap(0x%x:%08lx[0x%lx])=%p\n", busno, phys, size, va); */ /* P3 diag */
201 static void __iomem *
202 _sparc_ioremap(struct resource *res, u32 bus, u32 pa, int sz)
204 unsigned long offset = ((unsigned long) pa) & (~PAGE_MASK);
206 if (allocate_resource(&sparc_iomap, res,
207 (offset + sz + PAGE_SIZE-1) & PAGE_MASK,
208 sparc_iomap.start, sparc_iomap.end, PAGE_SIZE, NULL, NULL) != 0) {
209 /* Usually we cannot see printks in this case. */
210 prom_printf("alloc_io_res(%s): cannot occupy\n",
211 (res->name != NULL)? res->name: "???");
216 sparc_mapiorange(bus, pa, res->start, res->end - res->start + 1);
218 return (void __iomem *)(unsigned long)(res->start + offset);
222 * Comlimentary to _sparc_ioremap().
224 static void _sparc_free_io(struct resource *res)
228 plen = res->end - res->start + 1;
229 BUG_ON((plen & (PAGE_SIZE-1)) != 0);
230 sparc_unmapiorange(res->start, plen);
231 release_resource(res);
236 void sbus_set_sbus64(struct device *dev, int x)
238 printk("sbus_set_sbus64: unsupported\n");
240 EXPORT_SYMBOL(sbus_set_sbus64);
243 * Allocate a chunk of memory suitable for DMA.
244 * Typically devices use them for control blocks.
245 * CPU may access them without any explicit flushing.
247 static void *sbus_alloc_coherent(struct device *dev, size_t len,
248 dma_addr_t *dma_addrp, gfp_t gfp)
250 struct of_device *op = to_of_device(dev);
251 unsigned long len_total = (len + PAGE_SIZE-1) & PAGE_MASK;
253 struct resource *res;
256 /* XXX why are some lengths signed, others unsigned? */
260 /* XXX So what is maxphys for us and how do drivers know it? */
261 if (len > 256*1024) { /* __get_free_pages() limit */
265 order = get_order(len_total);
266 if ((va = __get_free_pages(GFP_KERNEL|__GFP_COMP, order)) == 0)
269 if ((res = kzalloc(sizeof(struct resource), GFP_KERNEL)) == NULL)
272 if (allocate_resource(&_sparc_dvma, res, len_total,
273 _sparc_dvma.start, _sparc_dvma.end, PAGE_SIZE, NULL, NULL) != 0) {
274 printk("sbus_alloc_consistent: cannot occupy 0x%lx", len_total);
277 mmu_inval_dma_area(va, len_total);
278 // XXX The mmu_map_dma_area does this for us below, see comments.
279 // sparc_mapiorange(0, virt_to_phys(va), res->start, len_total);
281 * XXX That's where sdev would be used. Currently we load
282 * all iommu tables with the same translations.
284 if (mmu_map_dma_area(dev, dma_addrp, va, res->start, len_total) != 0)
287 res->name = op->node->name;
289 return (void *)(unsigned long)res->start;
292 release_resource(res);
294 free_pages(va, order);
301 static void sbus_free_coherent(struct device *dev, size_t n, void *p,
304 struct resource *res;
307 if ((res = _sparc_find_resource(&_sparc_dvma,
308 (unsigned long)p)) == NULL) {
309 printk("sbus_free_consistent: cannot free %p\n", p);
313 if (((unsigned long)p & (PAGE_SIZE-1)) != 0) {
314 printk("sbus_free_consistent: unaligned va %p\n", p);
318 n = (n + PAGE_SIZE-1) & PAGE_MASK;
319 if ((res->end-res->start)+1 != n) {
320 printk("sbus_free_consistent: region 0x%lx asked 0x%zx\n",
321 (long)((res->end-res->start)+1), n);
325 release_resource(res);
328 /* mmu_inval_dma_area(va, n); */ /* it's consistent, isn't it */
329 pgv = virt_to_page(p);
330 mmu_unmap_dma_area(dev, ba, n);
332 __free_pages(pgv, get_order(n));
336 * Map a chunk of memory so that devices can see it.
337 * CPU view of this memory may be inconsistent with
338 * a device view and explicit flushing is necessary.
340 static dma_addr_t sbus_map_page(struct device *dev, struct page *page,
341 unsigned long offset, size_t len,
342 enum dma_data_direction dir,
343 struct dma_attrs *attrs)
345 void *va = page_address(page) + offset;
347 /* XXX why are some lengths signed, others unsigned? */
351 /* XXX So what is maxphys for us and how do drivers know it? */
352 if (len > 256*1024) { /* __get_free_pages() limit */
355 return mmu_get_scsi_one(dev, va, len);
358 static void sbus_unmap_page(struct device *dev, dma_addr_t ba, size_t n,
359 enum dma_data_direction dir, struct dma_attrs *attrs)
361 mmu_release_scsi_one(dev, ba, n);
364 static int sbus_map_sg(struct device *dev, struct scatterlist *sg, int n,
365 enum dma_data_direction dir, struct dma_attrs *attrs)
367 mmu_get_scsi_sgl(dev, sg, n);
370 * XXX sparc64 can return a partial length here. sun4c should do this
371 * but it currently panics if it can't fulfill the request - Anton
376 static void sbus_unmap_sg(struct device *dev, struct scatterlist *sg, int n,
377 enum dma_data_direction dir, struct dma_attrs *attrs)
379 mmu_release_scsi_sgl(dev, sg, n);
382 static void sbus_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
383 int n, enum dma_data_direction dir)
388 static void sbus_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
389 int n, enum dma_data_direction dir)
394 struct dma_map_ops sbus_dma_ops = {
395 .alloc_coherent = sbus_alloc_coherent,
396 .free_coherent = sbus_free_coherent,
397 .map_page = sbus_map_page,
398 .unmap_page = sbus_unmap_page,
399 .map_sg = sbus_map_sg,
400 .unmap_sg = sbus_unmap_sg,
401 .sync_sg_for_cpu = sbus_sync_sg_for_cpu,
402 .sync_sg_for_device = sbus_sync_sg_for_device,
405 struct dma_map_ops *dma_ops = &sbus_dma_ops;
406 EXPORT_SYMBOL(dma_ops);
408 static int __init sparc_register_ioport(void)
410 register_proc_sparc_ioport();
415 arch_initcall(sparc_register_ioport);
417 #endif /* CONFIG_SBUS */
421 /* Allocate and map kernel buffer using consistent mode DMA for a device.
422 * hwdev should be valid struct pci_dev pointer for PCI devices.
424 static void *pci32_alloc_coherent(struct device *dev, size_t len,
425 dma_addr_t *pba, gfp_t gfp)
427 unsigned long len_total = (len + PAGE_SIZE-1) & PAGE_MASK;
429 struct resource *res;
435 if (len > 256*1024) { /* __get_free_pages() limit */
439 order = get_order(len_total);
440 va = __get_free_pages(GFP_KERNEL, order);
442 printk("pci_alloc_consistent: no %ld pages\n", len_total>>PAGE_SHIFT);
446 if ((res = kzalloc(sizeof(struct resource), GFP_KERNEL)) == NULL) {
447 free_pages(va, order);
448 printk("pci_alloc_consistent: no core\n");
452 if (allocate_resource(&_sparc_dvma, res, len_total,
453 _sparc_dvma.start, _sparc_dvma.end, PAGE_SIZE, NULL, NULL) != 0) {
454 printk("pci_alloc_consistent: cannot occupy 0x%lx", len_total);
455 free_pages(va, order);
459 mmu_inval_dma_area(va, len_total);
461 /* P3 */ printk("pci_alloc_consistent: kva %lx uncva %lx phys %lx size %lx\n",
462 (long)va, (long)res->start, (long)virt_to_phys(va), len_total);
464 sparc_mapiorange(0, virt_to_phys(va), res->start, len_total);
466 *pba = virt_to_phys(va); /* equals virt_to_bus (R.I.P.) for us. */
467 return (void *) res->start;
470 /* Free and unmap a consistent DMA buffer.
471 * cpu_addr is what was returned from pci_alloc_consistent,
472 * size must be the same as what as passed into pci_alloc_consistent,
473 * and likewise dma_addr must be the same as what *dma_addrp was set to.
475 * References to the memory and mappings associated with cpu_addr/dma_addr
476 * past this call are illegal.
478 static void pci32_free_coherent(struct device *dev, size_t n, void *p,
481 struct resource *res;
484 if ((res = _sparc_find_resource(&_sparc_dvma,
485 (unsigned long)p)) == NULL) {
486 printk("pci_free_consistent: cannot free %p\n", p);
490 if (((unsigned long)p & (PAGE_SIZE-1)) != 0) {
491 printk("pci_free_consistent: unaligned va %p\n", p);
495 n = (n + PAGE_SIZE-1) & PAGE_MASK;
496 if ((res->end-res->start)+1 != n) {
497 printk("pci_free_consistent: region 0x%lx asked 0x%lx\n",
498 (long)((res->end-res->start)+1), (long)n);
502 pgp = (unsigned long) phys_to_virt(ba); /* bus_to_virt actually */
503 mmu_inval_dma_area(pgp, n);
504 sparc_unmapiorange((unsigned long)p, n);
506 release_resource(res);
509 free_pages(pgp, get_order(n));
513 * Same as pci_map_single, but with pages.
515 static dma_addr_t pci32_map_page(struct device *dev, struct page *page,
516 unsigned long offset, size_t size,
517 enum dma_data_direction dir,
518 struct dma_attrs *attrs)
520 /* IIep is write-through, not flushing. */
521 return page_to_phys(page) + offset;
524 /* Map a set of buffers described by scatterlist in streaming
525 * mode for DMA. This is the scather-gather version of the
526 * above pci_map_single interface. Here the scatter gather list
527 * elements are each tagged with the appropriate dma address
528 * and length. They are obtained via sg_dma_{address,length}(SG).
530 * NOTE: An implementation may be able to use a smaller number of
531 * DMA address/length pairs than there are SG table elements.
532 * (for example via virtual mapping capabilities)
533 * The routine returns the number of addr/length pairs actually
534 * used, at most nents.
536 * Device ownership issues as mentioned above for pci_map_single are
539 static int pci32_map_sg(struct device *device, struct scatterlist *sgl,
540 int nents, enum dma_data_direction dir,
541 struct dma_attrs *attrs)
543 struct scatterlist *sg;
546 /* IIep is write-through, not flushing. */
547 for_each_sg(sgl, sg, nents, n) {
548 BUG_ON(page_address(sg_page(sg)) == NULL);
549 sg->dma_address = virt_to_phys(sg_virt(sg));
550 sg->dma_length = sg->length;
555 /* Unmap a set of streaming mode DMA translations.
556 * Again, cpu read rules concerning calls here are the same as for
557 * pci_unmap_single() above.
559 static void pci32_unmap_sg(struct device *dev, struct scatterlist *sgl,
560 int nents, enum dma_data_direction dir,
561 struct dma_attrs *attrs)
563 struct scatterlist *sg;
566 if (dir != PCI_DMA_TODEVICE) {
567 for_each_sg(sgl, sg, nents, n) {
568 BUG_ON(page_address(sg_page(sg)) == NULL);
570 (unsigned long) page_address(sg_page(sg)),
571 (sg->length + PAGE_SIZE-1) & PAGE_MASK);
576 /* Make physical memory consistent for a single
577 * streaming mode DMA translation before or after a transfer.
579 * If you perform a pci_map_single() but wish to interrogate the
580 * buffer using the cpu, yet do not wish to teardown the PCI dma
581 * mapping, you must call this function before doing so. At the
582 * next point you give the PCI dma address back to the card, you
583 * must first perform a pci_dma_sync_for_device, and then the
584 * device again owns the buffer.
586 static void pci32_sync_single_for_cpu(struct device *dev, dma_addr_t ba,
587 size_t size, enum dma_data_direction dir)
589 if (dir != PCI_DMA_TODEVICE) {
590 mmu_inval_dma_area((unsigned long)phys_to_virt(ba),
591 (size + PAGE_SIZE-1) & PAGE_MASK);
595 static void pci32_sync_single_for_device(struct device *dev, dma_addr_t ba,
596 size_t size, enum dma_data_direction dir)
598 if (dir != PCI_DMA_TODEVICE) {
599 mmu_inval_dma_area((unsigned long)phys_to_virt(ba),
600 (size + PAGE_SIZE-1) & PAGE_MASK);
604 /* Make physical memory consistent for a set of streaming
605 * mode DMA translations after a transfer.
607 * The same as pci_dma_sync_single_* but for a scatter-gather list,
608 * same rules and usage.
610 static void pci32_sync_sg_for_cpu(struct device *dev, struct scatterlist *sgl,
611 int nents, enum dma_data_direction dir)
613 struct scatterlist *sg;
616 if (dir != PCI_DMA_TODEVICE) {
617 for_each_sg(sgl, sg, nents, n) {
618 BUG_ON(page_address(sg_page(sg)) == NULL);
620 (unsigned long) page_address(sg_page(sg)),
621 (sg->length + PAGE_SIZE-1) & PAGE_MASK);
626 static void pci32_sync_sg_for_device(struct device *device, struct scatterlist *sgl,
627 int nents, enum dma_data_direction dir)
629 struct scatterlist *sg;
632 if (dir != PCI_DMA_TODEVICE) {
633 for_each_sg(sgl, sg, nents, n) {
634 BUG_ON(page_address(sg_page(sg)) == NULL);
636 (unsigned long) page_address(sg_page(sg)),
637 (sg->length + PAGE_SIZE-1) & PAGE_MASK);
642 struct dma_map_ops pci32_dma_ops = {
643 .alloc_coherent = pci32_alloc_coherent,
644 .free_coherent = pci32_free_coherent,
645 .map_page = pci32_map_page,
646 .map_sg = pci32_map_sg,
647 .unmap_sg = pci32_unmap_sg,
648 .sync_single_for_cpu = pci32_sync_single_for_cpu,
649 .sync_single_for_device = pci32_sync_single_for_device,
650 .sync_sg_for_cpu = pci32_sync_sg_for_cpu,
651 .sync_sg_for_device = pci32_sync_sg_for_device,
653 EXPORT_SYMBOL(pci32_dma_ops);
655 #endif /* CONFIG_PCI */
658 * Return whether the given PCI device DMA address mask can be
659 * supported properly. For example, if your device can only drive the
660 * low 24-bits during PCI bus mastering, then you would pass
661 * 0x00ffffff as the mask to this function.
663 int dma_supported(struct device *dev, u64 mask)
666 if (dev->bus == &pci_bus_type)
671 EXPORT_SYMBOL(dma_supported);
673 int dma_set_mask(struct device *dev, u64 dma_mask)
676 if (dev->bus == &pci_bus_type)
677 return pci_set_dma_mask(to_pci_dev(dev), dma_mask);
681 EXPORT_SYMBOL(dma_set_mask);
684 #ifdef CONFIG_PROC_FS
687 _sparc_io_get_info(char *buf, char **start, off_t fpos, int length, int *eof,
690 char *p = buf, *e = buf + length;
694 for (r = ((struct resource *)data)->child; r != NULL; r = r->sibling) {
695 if (p + 32 >= e) /* Better than nothing */
697 if ((nm = r->name) == 0) nm = "???";
698 p += sprintf(p, "%016llx-%016llx: %s\n",
699 (unsigned long long)r->start,
700 (unsigned long long)r->end, nm);
706 #endif /* CONFIG_PROC_FS */
709 * This is a version of find_resource and it belongs to kernel/resource.c.
710 * Until we have agreement with Linus and Martin, it lingers here.
712 * XXX Too slow. Can have 8192 DVMA pages on sun4m in the worst case.
713 * This probably warrants some sort of hashing.
715 static struct resource *_sparc_find_resource(struct resource *root,
718 struct resource *tmp;
720 for (tmp = root->child; tmp != 0; tmp = tmp->sibling) {
721 if (tmp->start <= hit && tmp->end >= hit)
727 static void register_proc_sparc_ioport(void)
729 #ifdef CONFIG_PROC_FS
730 create_proc_read_entry("io_map",0,NULL,_sparc_io_get_info,&sparc_iomap);
731 create_proc_read_entry("dvma_map",0,NULL,_sparc_io_get_info,&_sparc_dvma);