xref: /netbsd-src/sys/arch/xen/x86/xen_bus_dma.c (revision 7fa608457b817eca6e0977b37f758ae064f3c99c)
1 /*	$NetBSD: xen_bus_dma.c,v 1.9 2007/02/24 21:19:25 bouyer Exp $	*/
2 /*	NetBSD bus_dma.c,v 1.21 2005/04/16 07:53:35 yamt Exp */
3 
4 /*-
5  * Copyright (c) 1996, 1997, 1998 The NetBSD Foundation, Inc.
6  * All rights reserved.
7  *
8  * This code is derived from software contributed to The NetBSD Foundation
9  * by Charles M. Hannum and by Jason R. Thorpe of the Numerical Aerospace
10  * Simulation Facility, NASA Ames Research Center.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by the NetBSD
23  *	Foundation, Inc. and its contributors.
24  * 4. Neither the name of The NetBSD Foundation nor the names of its
25  *    contributors may be used to endorse or promote products derived
26  *    from this software without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
29  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
30  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
31  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
32  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
33  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
34  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
35  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
36  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
37  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
38  * POSSIBILITY OF SUCH DAMAGE.
39  */
40 
41 #include <sys/cdefs.h>
42 __KERNEL_RCSID(0, "$NetBSD: xen_bus_dma.c,v 1.9 2007/02/24 21:19:25 bouyer Exp $");
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/kernel.h>
47 #include <sys/malloc.h>
48 #include <sys/mbuf.h>
49 #include <sys/proc.h>
50 
51 #include <machine/bus.h>
52 #include <machine/bus_private.h>
53 
54 #include <uvm/uvm_extern.h>
55 
56 extern paddr_t avail_end;
57 
58 /* Pure 2^n version of get_order */
59 static inline int get_order(unsigned long size)
60 {
61 	int order = -1;
62 	size = (size - 1) >> (PAGE_SHIFT - 1);
63 	do {
64 		size >>= 1;
65 		order++;
66 	} while (size);
67 	return order;
68 }
69 
70 static int
71 _xen_alloc_contig(bus_size_t size, bus_size_t alignment, bus_size_t boundary,
72     struct pglist *mlistp, int flags, bus_addr_t low, bus_addr_t high)
73 {
74 	int order, i;
75 	unsigned long npagesreq, npages, mfn;
76 	bus_addr_t pa;
77 	struct vm_page *pg, *pgnext;
78 	int s, error;
79 #ifdef XEN3
80 	struct xen_memory_reservation res;
81 #endif
82 
83 	/*
84 	 * When requesting a contigous memory region, the hypervisor will
85 	 * return a memory range aligned on size. This will automagically
86 	 * handle "boundary", but the only way to enforce alignment
87 	 * is to request a memory region of size max(alignment, size).
88 	 */
89 	order = max(get_order(size), get_order(alignment));
90 	npages = (1 << order);
91 	npagesreq = (size >> PAGE_SHIFT);
92 	KASSERT(npages >= npagesreq);
93 
94 	/* get npages from UWM, and give them back to the hypervisor */
95 	error = uvm_pglistalloc(npages << PAGE_SHIFT, 0, avail_end, 0, 0,
96 	    mlistp, npages, (flags & BUS_DMA_NOWAIT) == 0);
97 	if (error)
98 		return (error);
99 
100 	for (pg = mlistp->tqh_first; pg != NULL; pg = pg->pageq.tqe_next) {
101 		pa = VM_PAGE_TO_PHYS(pg);
102 		mfn = xpmap_ptom(pa) >> PAGE_SHIFT;
103 		xpmap_phys_to_machine_mapping[
104 		    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = INVALID_P2M_ENTRY;
105 #ifdef XEN3
106 		res.extent_start = &mfn;
107 		res.nr_extents = 1;
108 		res.extent_order = 0;
109 		res.domid = DOMID_SELF;
110 		if (HYPERVISOR_memory_op(XENMEM_decrease_reservation, &res)
111 		    < 0) {
112 #ifdef DEBUG
113 			printf("xen_alloc_contig: XENMEM_decrease_reservation "
114 			    "failed!\n");
115 #endif
116 			xpmap_phys_to_machine_mapping[
117 			    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn;
118 
119 			error = ENOMEM;
120 			goto failed;
121 		}
122 #else
123 		if (HYPERVISOR_dom_mem_op(MEMOP_decrease_reservation,
124 		    &mfn, 1, 0) != 1) {
125 #ifdef DEBUG
126 			printf("xen_alloc_contig: MEMOP_decrease_reservation "
127 			    "failed!\n");
128 #endif
129 			xpmap_phys_to_machine_mapping[
130 			    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn;
131 			error = ENOMEM;
132 			goto failed;
133 		}
134 #endif
135 	}
136 	/* Get the new contiguous memory extent */
137 #ifdef XEN3
138 	res.extent_start = &mfn;
139 	res.nr_extents = 1;
140 	res.extent_order = order;
141 	res.address_bits = get_order(high) + PAGE_SHIFT;
142 	res.domid = DOMID_SELF;
143 	if (HYPERVISOR_memory_op(XENMEM_increase_reservation, &res) < 0) {
144 #ifdef DEBUG
145 		printf("xen_alloc_contig: XENMEM_increase_reservation "
146 		    "failed!\n");
147 #endif
148 		error = ENOMEM;
149 		pg = NULL;
150 		goto failed;
151 	}
152 #else
153 	if (HYPERVISOR_dom_mem_op(MEMOP_increase_reservation,
154 	    &mfn, 1, order) != 1) {
155 #ifdef DEBUG
156 		printf("xen_alloc_contig: MEMOP_increase_reservation "
157 		    "failed!\n");
158 #endif
159 		error = ENOMEM;
160 		pg = NULL;
161 		goto failed;
162 	}
163 #endif
164 	s = splvm();
165 	/* Map the new extent in place of the old pages */
166 	for (pg = mlistp->tqh_first, i = 0; pg != NULL; pg = pgnext, i++) {
167 		pgnext = pg->pageq.tqe_next;
168 		pa = VM_PAGE_TO_PHYS(pg);
169 		xpmap_phys_to_machine_mapping[
170 		    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn+i;
171 		xpq_queue_machphys_update((mfn+i) << PAGE_SHIFT, pa);
172 		/* while here, give extra pages back to UVM */
173 		if (i >= npagesreq) {
174 			TAILQ_REMOVE(mlistp, pg, pageq);
175 			uvm_pagefree(pg);
176 		}
177 
178 	}
179 	/* Flush updates through and flush the TLB */
180 	xpq_queue_tlb_flush();
181 	xpq_flush_queue();
182 	splx(s);
183 	return 0;
184 
185 failed:
186 	/*
187 	 * Attempt to recover from a failed decrease or increase reservation:
188 	 * if decrease_reservation failed, we don't have given all pages
189 	 * back to Xen; give them back to UVM, and get the missing pages
190 	 * from Xen.
191 	 * if increase_reservation failed, we expect pg to be NULL and we just
192 	 * get back the missing pages from Xen one by one.
193 	 */
194 	/* give back remaining pages to UVM */
195 	for (; pg != NULL; pg = pgnext) {
196 		pgnext = pg->pageq.tqe_next;
197 		TAILQ_REMOVE(mlistp, pg, pageq);
198 		uvm_pagefree(pg);
199 	}
200 	/* remplace the pages that we already gave to Xen */
201 	s = splvm();
202 	for (pg = mlistp->tqh_first; pg != NULL; pg = pgnext) {
203 		pgnext = pg->pageq.tqe_next;
204 #ifdef XEN3
205 		res.extent_start = &mfn;
206 		res.nr_extents = 1;
207 		res.extent_order = 0;
208 		res.address_bits = 32;
209 		res.domid = DOMID_SELF;
210 		if (HYPERVISOR_memory_op(XENMEM_increase_reservation, &res)
211 		    < 0) {
212 			printf("xen_alloc_contig: recovery "
213 			    "XENMEM_increase_reservation failed!\n");
214 			break;
215 		}
216 #else
217 		if (HYPERVISOR_dom_mem_op(MEMOP_increase_reservation,
218 		    &mfn, 1, 0) != 1) {
219 			printf("xen_alloc_contig: recovery "
220 			    "MEMOP_increase_reservation failed!\n");
221 			break;
222 		}
223 #endif
224 		pa = VM_PAGE_TO_PHYS(pg);
225 		xpmap_phys_to_machine_mapping[
226 		    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn;
227 		xpq_queue_machphys_update((mfn) << PAGE_SHIFT, pa);
228 		TAILQ_REMOVE(mlistp, pg, pageq);
229 		uvm_pagefree(pg);
230 	}
231 	/* Flush updates through and flush the TLB */
232 	xpq_queue_tlb_flush();
233 	xpq_flush_queue();
234 	splx(s);
235 	return error;
236 }
237 
238 
239 /*
240  * Allocate physical memory from the given physical address range.
241  * Called by DMA-safe memory allocation methods.
242  * We need our own version to deal with physical vs machine addresses.
243  */
244 int
245 _xen_bus_dmamem_alloc_range(bus_dma_tag_t t, bus_size_t size,
246     bus_size_t alignment, bus_size_t boundary, bus_dma_segment_t *segs,
247     int nsegs, int *rsegs, int flags, bus_addr_t low, bus_addr_t high)
248 {
249 	bus_addr_t curaddr, lastaddr;
250 	struct vm_page *m;
251 	struct pglist mlist;
252 	int curseg, error;
253 	int doingrealloc = 0;
254 
255 	/* Always round the size. */
256 	size = round_page(size);
257 
258 	KASSERT((alignment & (alignment - 1)) == 0);
259 	KASSERT((boundary & (boundary - 1)) == 0);
260 	if (alignment < PAGE_SIZE)
261 		alignment = PAGE_SIZE;
262 	if (boundary != 0 && boundary < size)
263 		return (EINVAL);
264 
265 	/*
266 	 * Allocate pages from the VM system.
267 	 */
268 	error = uvm_pglistalloc(size, 0, avail_end, alignment, boundary,
269 	    &mlist, nsegs, (flags & BUS_DMA_NOWAIT) == 0);
270 	if (error)
271 		return (error);
272 again:
273 
274 	/*
275 	 * Compute the location, size, and number of segments actually
276 	 * returned by the VM code.
277 	 */
278 	m = mlist.tqh_first;
279 	curseg = 0;
280 	curaddr = lastaddr = segs[curseg].ds_addr = _BUS_VM_PAGE_TO_BUS(m);
281 	if (curaddr < low || curaddr >= high)
282 		goto badaddr;
283 	segs[curseg].ds_len = PAGE_SIZE;
284 	m = m->pageq.tqe_next;
285 	if ((segs[curseg].ds_addr & (alignment - 1)) != 0)
286 		goto dorealloc;
287 
288 	for (; m != NULL; m = m->pageq.tqe_next) {
289 		curaddr = _BUS_VM_PAGE_TO_BUS(m);
290 		if (curaddr < low || curaddr >= high)
291 			goto badaddr;
292 		if (curaddr == (lastaddr + PAGE_SIZE)) {
293 			segs[curseg].ds_len += PAGE_SIZE;
294 			if ((lastaddr & boundary) != (curaddr & boundary))
295 				goto dorealloc;
296 		} else {
297 			curseg++;
298 			if (curseg >= nsegs || (curaddr & (alignment - 1)) != 0)
299 				goto dorealloc;
300 			segs[curseg].ds_addr = curaddr;
301 			segs[curseg].ds_len = PAGE_SIZE;
302 		}
303 		lastaddr = curaddr;
304 	}
305 
306 	*rsegs = curseg + 1;
307 	return (0);
308 
309 badaddr:
310 #ifdef XEN3
311 	if (doingrealloc == 0)
312 		goto dorealloc;
313 	if (curaddr < low) {
314 		/* no way to enforce this */
315 		printf("_xen_bus_dmamem_alloc_range: no way to "
316 		    "enforce address range\n");
317 		uvm_pglistfree(&mlist);
318 		return EINVAL;
319 	}
320 	printf("xen_bus_dmamem_alloc_range: "
321 	    "curraddr=0x%lx > high=0x%lx\n",
322 	    (u_long)curaddr, (u_long)high);
323 	panic("xen_bus_dmamem_alloc_range 1");
324 #else /* !XEN3 */
325 	/*
326 	 * If machine addresses are outside the allowed
327 	 * range we have to bail. Xen2 doesn't offer an
328 	 * interface to get memory in a specific address
329 	 * range.
330 	 */
331 	printf("_xen_bus_dmamem_alloc_range: no way to "
332 	    "enforce address range\n");
333 	uvm_pglistfree(&mlist);
334 	return EINVAL;
335 #endif /* XEN3 */
336 dorealloc:
337 	if (doingrealloc == 1)
338 		panic("_xen_bus_dmamem_alloc_range: "
339 		   "xen_alloc_contig returned "
340 		   "too much segments");
341 	doingrealloc = 1;
342 	/*
343 	 * Too much segments, or memory doesn't fit
344 	 * constraints. Free this memory and
345 	 * get a contigous segment from the hypervisor.
346 	 */
347 	uvm_pglistfree(&mlist);
348 	for (curseg = 0; curseg < nsegs; curseg++) {
349 		segs[curseg].ds_addr = 0;
350 		segs[curseg].ds_len = 0;
351 	}
352 	error = _xen_alloc_contig(size, alignment,
353 	    boundary, &mlist, flags, low, high);
354 	if (error)
355 		return error;
356 	goto again;
357 }
358