1 /*
2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
4 * for more details.
5 *
6 * Copyright (C) 2000,2002-2005 Silicon Graphics, Inc. All rights reserved.
7 *
8 * Routines for PCI DMA mapping. See Documentation/DMA-API.txt for
9 * a description of how these routines should be used.
10 */
11
12 #include <linux/gfp.h>
13 #include <linux/module.h>
14 #include <linux/dma-mapping.h>
15 #include <asm/dma.h>
16 #include <asm/sn/intr.h>
17 #include <asm/sn/pcibus_provider_defs.h>
18 #include <asm/sn/pcidev.h>
19 #include <asm/sn/sn_sal.h>
20
21 #define SG_ENT_VIRT_ADDRESS(sg) (sg_virt((sg)))
22 #define SG_ENT_PHYS_ADDRESS(SG) virt_to_phys(SG_ENT_VIRT_ADDRESS(SG))
23
24 /**
25 * sn_dma_supported - test a DMA mask
26 * @dev: device to test
27 * @mask: DMA mask to test
28 *
29 * Return whether the given PCI device DMA address mask can be supported
30 * properly. For example, if your device can only drive the low 24-bits
31 * during PCI bus mastering, then you would pass 0x00ffffff as the mask to
32 * this function. Of course, SN only supports devices that have 32 or more
33 * address bits when using the PMU.
34 */
sn_dma_supported(struct device * dev,u64 mask)35 static int sn_dma_supported(struct device *dev, u64 mask)
36 {
37 BUG_ON(!dev_is_pci(dev));
38
39 if (mask < 0x7fffffff)
40 return 0;
41 return 1;
42 }
43
44 /**
45 * sn_dma_set_mask - set the DMA mask
46 * @dev: device to set
47 * @dma_mask: new mask
48 *
49 * Set @dev's DMA mask if the hw supports it.
50 */
sn_dma_set_mask(struct device * dev,u64 dma_mask)51 int sn_dma_set_mask(struct device *dev, u64 dma_mask)
52 {
53 BUG_ON(!dev_is_pci(dev));
54
55 if (!sn_dma_supported(dev, dma_mask))
56 return 0;
57
58 *dev->dma_mask = dma_mask;
59 return 1;
60 }
61 EXPORT_SYMBOL(sn_dma_set_mask);
62
63 /**
64 * sn_dma_alloc_coherent - allocate memory for coherent DMA
65 * @dev: device to allocate for
66 * @size: size of the region
67 * @dma_handle: DMA (bus) address
68 * @flags: memory allocation flags
69 *
70 * dma_alloc_coherent() returns a pointer to a memory region suitable for
71 * coherent DMA traffic to/from a PCI device. On SN platforms, this means
72 * that @dma_handle will have the %PCIIO_DMA_CMD flag set.
73 *
74 * This interface is usually used for "command" streams (e.g. the command
75 * queue for a SCSI controller). See Documentation/DMA-API.txt for
76 * more information.
77 */
sn_dma_alloc_coherent(struct device * dev,size_t size,dma_addr_t * dma_handle,gfp_t flags,unsigned long attrs)78 static void *sn_dma_alloc_coherent(struct device *dev, size_t size,
79 dma_addr_t * dma_handle, gfp_t flags,
80 unsigned long attrs)
81 {
82 void *cpuaddr;
83 unsigned long phys_addr;
84 int node;
85 struct pci_dev *pdev = to_pci_dev(dev);
86 struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
87
88 BUG_ON(!dev_is_pci(dev));
89
90 /*
91 * Allocate the memory.
92 */
93 node = pcibus_to_node(pdev->bus);
94 if (likely(node >=0)) {
95 struct page *p = __alloc_pages_node(node,
96 flags, get_order(size));
97
98 if (likely(p))
99 cpuaddr = page_address(p);
100 else
101 return NULL;
102 } else
103 cpuaddr = (void *)__get_free_pages(flags, get_order(size));
104
105 if (unlikely(!cpuaddr))
106 return NULL;
107
108 memset(cpuaddr, 0x0, size);
109
110 /* physical addr. of the memory we just got */
111 phys_addr = __pa(cpuaddr);
112
113 /*
114 * 64 bit address translations should never fail.
115 * 32 bit translations can fail if there are insufficient mapping
116 * resources.
117 */
118
119 *dma_handle = provider->dma_map_consistent(pdev, phys_addr, size,
120 SN_DMA_ADDR_PHYS);
121 if (!*dma_handle) {
122 printk(KERN_ERR "%s: out of ATEs\n", __func__);
123 free_pages((unsigned long)cpuaddr, get_order(size));
124 return NULL;
125 }
126
127 return cpuaddr;
128 }
129
130 /**
131 * sn_pci_free_coherent - free memory associated with coherent DMAable region
132 * @dev: device to free for
133 * @size: size to free
134 * @cpu_addr: kernel virtual address to free
135 * @dma_handle: DMA address associated with this region
136 *
137 * Frees the memory allocated by dma_alloc_coherent(), potentially unmapping
138 * any associated IOMMU mappings.
139 */
sn_dma_free_coherent(struct device * dev,size_t size,void * cpu_addr,dma_addr_t dma_handle,unsigned long attrs)140 static void sn_dma_free_coherent(struct device *dev, size_t size, void *cpu_addr,
141 dma_addr_t dma_handle, unsigned long attrs)
142 {
143 struct pci_dev *pdev = to_pci_dev(dev);
144 struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
145
146 BUG_ON(!dev_is_pci(dev));
147
148 provider->dma_unmap(pdev, dma_handle, 0);
149 free_pages((unsigned long)cpu_addr, get_order(size));
150 }
151
152 /**
153 * sn_dma_map_single_attrs - map a single page for DMA
154 * @dev: device to map for
155 * @cpu_addr: kernel virtual address of the region to map
156 * @size: size of the region
157 * @direction: DMA direction
158 * @attrs: optional dma attributes
159 *
160 * Map the region pointed to by @cpu_addr for DMA and return the
161 * DMA address.
162 *
163 * We map this to the one step pcibr_dmamap_trans interface rather than
164 * the two step pcibr_dmamap_alloc/pcibr_dmamap_addr because we have
165 * no way of saving the dmamap handle from the alloc to later free
166 * (which is pretty much unacceptable).
167 *
168 * mappings with the DMA_ATTR_WRITE_BARRIER get mapped with
169 * dma_map_consistent() so that writes force a flush of pending DMA.
170 * (See "SGI Altix Architecture Considerations for Linux Device Drivers",
171 * Document Number: 007-4763-001)
172 *
173 * TODO: simplify our interface;
174 * figure out how to save dmamap handle so can use two step.
175 */
sn_dma_map_page(struct device * dev,struct page * page,unsigned long offset,size_t size,enum dma_data_direction dir,unsigned long attrs)176 static dma_addr_t sn_dma_map_page(struct device *dev, struct page *page,
177 unsigned long offset, size_t size,
178 enum dma_data_direction dir,
179 unsigned long attrs)
180 {
181 void *cpu_addr = page_address(page) + offset;
182 dma_addr_t dma_addr;
183 unsigned long phys_addr;
184 struct pci_dev *pdev = to_pci_dev(dev);
185 struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
186
187 BUG_ON(!dev_is_pci(dev));
188
189 phys_addr = __pa(cpu_addr);
190 if (attrs & DMA_ATTR_WRITE_BARRIER)
191 dma_addr = provider->dma_map_consistent(pdev, phys_addr,
192 size, SN_DMA_ADDR_PHYS);
193 else
194 dma_addr = provider->dma_map(pdev, phys_addr, size,
195 SN_DMA_ADDR_PHYS);
196
197 if (!dma_addr) {
198 printk(KERN_ERR "%s: out of ATEs\n", __func__);
199 return 0;
200 }
201 return dma_addr;
202 }
203
204 /**
205 * sn_dma_unmap_single_attrs - unamp a DMA mapped page
206 * @dev: device to sync
207 * @dma_addr: DMA address to sync
208 * @size: size of region
209 * @direction: DMA direction
210 * @attrs: optional dma attributes
211 *
212 * This routine is supposed to sync the DMA region specified
213 * by @dma_handle into the coherence domain. On SN, we're always cache
214 * coherent, so we just need to free any ATEs associated with this mapping.
215 */
sn_dma_unmap_page(struct device * dev,dma_addr_t dma_addr,size_t size,enum dma_data_direction dir,unsigned long attrs)216 static void sn_dma_unmap_page(struct device *dev, dma_addr_t dma_addr,
217 size_t size, enum dma_data_direction dir,
218 unsigned long attrs)
219 {
220 struct pci_dev *pdev = to_pci_dev(dev);
221 struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
222
223 BUG_ON(!dev_is_pci(dev));
224
225 provider->dma_unmap(pdev, dma_addr, dir);
226 }
227
228 /**
229 * sn_dma_unmap_sg - unmap a DMA scatterlist
230 * @dev: device to unmap
231 * @sg: scatterlist to unmap
232 * @nhwentries: number of scatterlist entries
233 * @direction: DMA direction
234 * @attrs: optional dma attributes
235 *
236 * Unmap a set of streaming mode DMA translations.
237 */
sn_dma_unmap_sg(struct device * dev,struct scatterlist * sgl,int nhwentries,enum dma_data_direction dir,unsigned long attrs)238 static void sn_dma_unmap_sg(struct device *dev, struct scatterlist *sgl,
239 int nhwentries, enum dma_data_direction dir,
240 unsigned long attrs)
241 {
242 int i;
243 struct pci_dev *pdev = to_pci_dev(dev);
244 struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
245 struct scatterlist *sg;
246
247 BUG_ON(!dev_is_pci(dev));
248
249 for_each_sg(sgl, sg, nhwentries, i) {
250 provider->dma_unmap(pdev, sg->dma_address, dir);
251 sg->dma_address = (dma_addr_t) NULL;
252 sg->dma_length = 0;
253 }
254 }
255
256 /**
257 * sn_dma_map_sg - map a scatterlist for DMA
258 * @dev: device to map for
259 * @sg: scatterlist to map
260 * @nhwentries: number of entries
261 * @direction: direction of the DMA transaction
262 * @attrs: optional dma attributes
263 *
264 * mappings with the DMA_ATTR_WRITE_BARRIER get mapped with
265 * dma_map_consistent() so that writes force a flush of pending DMA.
266 * (See "SGI Altix Architecture Considerations for Linux Device Drivers",
267 * Document Number: 007-4763-001)
268 *
269 * Maps each entry of @sg for DMA.
270 */
sn_dma_map_sg(struct device * dev,struct scatterlist * sgl,int nhwentries,enum dma_data_direction dir,unsigned long attrs)271 static int sn_dma_map_sg(struct device *dev, struct scatterlist *sgl,
272 int nhwentries, enum dma_data_direction dir,
273 unsigned long attrs)
274 {
275 unsigned long phys_addr;
276 struct scatterlist *saved_sg = sgl, *sg;
277 struct pci_dev *pdev = to_pci_dev(dev);
278 struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
279 int i;
280
281 BUG_ON(!dev_is_pci(dev));
282
283 /*
284 * Setup a DMA address for each entry in the scatterlist.
285 */
286 for_each_sg(sgl, sg, nhwentries, i) {
287 dma_addr_t dma_addr;
288 phys_addr = SG_ENT_PHYS_ADDRESS(sg);
289 if (attrs & DMA_ATTR_WRITE_BARRIER)
290 dma_addr = provider->dma_map_consistent(pdev,
291 phys_addr,
292 sg->length,
293 SN_DMA_ADDR_PHYS);
294 else
295 dma_addr = provider->dma_map(pdev, phys_addr,
296 sg->length,
297 SN_DMA_ADDR_PHYS);
298
299 sg->dma_address = dma_addr;
300 if (!sg->dma_address) {
301 printk(KERN_ERR "%s: out of ATEs\n", __func__);
302
303 /*
304 * Free any successfully allocated entries.
305 */
306 if (i > 0)
307 sn_dma_unmap_sg(dev, saved_sg, i, dir, attrs);
308 return 0;
309 }
310
311 sg->dma_length = sg->length;
312 }
313
314 return nhwentries;
315 }
316
sn_dma_sync_single_for_cpu(struct device * dev,dma_addr_t dma_handle,size_t size,enum dma_data_direction dir)317 static void sn_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
318 size_t size, enum dma_data_direction dir)
319 {
320 BUG_ON(!dev_is_pci(dev));
321 }
322
sn_dma_sync_single_for_device(struct device * dev,dma_addr_t dma_handle,size_t size,enum dma_data_direction dir)323 static void sn_dma_sync_single_for_device(struct device *dev, dma_addr_t dma_handle,
324 size_t size,
325 enum dma_data_direction dir)
326 {
327 BUG_ON(!dev_is_pci(dev));
328 }
329
sn_dma_sync_sg_for_cpu(struct device * dev,struct scatterlist * sg,int nelems,enum dma_data_direction dir)330 static void sn_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
331 int nelems, enum dma_data_direction dir)
332 {
333 BUG_ON(!dev_is_pci(dev));
334 }
335
sn_dma_sync_sg_for_device(struct device * dev,struct scatterlist * sg,int nelems,enum dma_data_direction dir)336 static void sn_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
337 int nelems, enum dma_data_direction dir)
338 {
339 BUG_ON(!dev_is_pci(dev));
340 }
341
sn_dma_mapping_error(struct device * dev,dma_addr_t dma_addr)342 static int sn_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
343 {
344 return 0;
345 }
346
sn_dma_get_required_mask(struct device * dev)347 u64 sn_dma_get_required_mask(struct device *dev)
348 {
349 return DMA_BIT_MASK(64);
350 }
351 EXPORT_SYMBOL_GPL(sn_dma_get_required_mask);
352
sn_pci_get_legacy_mem(struct pci_bus * bus)353 char *sn_pci_get_legacy_mem(struct pci_bus *bus)
354 {
355 if (!SN_PCIBUS_BUSSOFT(bus))
356 return ERR_PTR(-ENODEV);
357
358 return (char *)(SN_PCIBUS_BUSSOFT(bus)->bs_legacy_mem | __IA64_UNCACHED_OFFSET);
359 }
360
sn_pci_legacy_read(struct pci_bus * bus,u16 port,u32 * val,u8 size)361 int sn_pci_legacy_read(struct pci_bus *bus, u16 port, u32 *val, u8 size)
362 {
363 unsigned long addr;
364 int ret;
365 struct ia64_sal_retval isrv;
366
367 /*
368 * First, try the SN_SAL_IOIF_PCI_SAFE SAL call which can work
369 * around hw issues at the pci bus level. SGI proms older than
370 * 4.10 don't implement this.
371 */
372
373 SAL_CALL(isrv, SN_SAL_IOIF_PCI_SAFE,
374 pci_domain_nr(bus), bus->number,
375 0, /* io */
376 0, /* read */
377 port, size, __pa(val));
378
379 if (isrv.status == 0)
380 return size;
381
382 /*
383 * If the above failed, retry using the SAL_PROBE call which should
384 * be present in all proms (but which cannot work round PCI chipset
385 * bugs). This code is retained for compatibility with old
386 * pre-4.10 proms, and should be removed at some point in the future.
387 */
388
389 if (!SN_PCIBUS_BUSSOFT(bus))
390 return -ENODEV;
391
392 addr = SN_PCIBUS_BUSSOFT(bus)->bs_legacy_io | __IA64_UNCACHED_OFFSET;
393 addr += port;
394
395 ret = ia64_sn_probe_mem(addr, (long)size, (void *)val);
396
397 if (ret == 2)
398 return -EINVAL;
399
400 if (ret == 1)
401 *val = -1;
402
403 return size;
404 }
405
sn_pci_legacy_write(struct pci_bus * bus,u16 port,u32 val,u8 size)406 int sn_pci_legacy_write(struct pci_bus *bus, u16 port, u32 val, u8 size)
407 {
408 int ret = size;
409 unsigned long paddr;
410 unsigned long *addr;
411 struct ia64_sal_retval isrv;
412
413 /*
414 * First, try the SN_SAL_IOIF_PCI_SAFE SAL call which can work
415 * around hw issues at the pci bus level. SGI proms older than
416 * 4.10 don't implement this.
417 */
418
419 SAL_CALL(isrv, SN_SAL_IOIF_PCI_SAFE,
420 pci_domain_nr(bus), bus->number,
421 0, /* io */
422 1, /* write */
423 port, size, __pa(&val));
424
425 if (isrv.status == 0)
426 return size;
427
428 /*
429 * If the above failed, retry using the SAL_PROBE call which should
430 * be present in all proms (but which cannot work round PCI chipset
431 * bugs). This code is retained for compatibility with old
432 * pre-4.10 proms, and should be removed at some point in the future.
433 */
434
435 if (!SN_PCIBUS_BUSSOFT(bus)) {
436 ret = -ENODEV;
437 goto out;
438 }
439
440 /* Put the phys addr in uncached space */
441 paddr = SN_PCIBUS_BUSSOFT(bus)->bs_legacy_io | __IA64_UNCACHED_OFFSET;
442 paddr += port;
443 addr = (unsigned long *)paddr;
444
445 switch (size) {
446 case 1:
447 *(volatile u8 *)(addr) = (u8)(val);
448 break;
449 case 2:
450 *(volatile u16 *)(addr) = (u16)(val);
451 break;
452 case 4:
453 *(volatile u32 *)(addr) = (u32)(val);
454 break;
455 default:
456 ret = -EINVAL;
457 break;
458 }
459 out:
460 return ret;
461 }
462
463 static struct dma_map_ops sn_dma_ops = {
464 .alloc = sn_dma_alloc_coherent,
465 .free = sn_dma_free_coherent,
466 .map_page = sn_dma_map_page,
467 .unmap_page = sn_dma_unmap_page,
468 .map_sg = sn_dma_map_sg,
469 .unmap_sg = sn_dma_unmap_sg,
470 .sync_single_for_cpu = sn_dma_sync_single_for_cpu,
471 .sync_sg_for_cpu = sn_dma_sync_sg_for_cpu,
472 .sync_single_for_device = sn_dma_sync_single_for_device,
473 .sync_sg_for_device = sn_dma_sync_sg_for_device,
474 .mapping_error = sn_dma_mapping_error,
475 .dma_supported = sn_dma_supported,
476 };
477
sn_dma_init(void)478 void sn_dma_init(void)
479 {
480 dma_ops = &sn_dma_ops;
481 }
482