xref: /netbsd-src/sys/arch/xen/x86/xen_bus_dma.c (revision 274254cdae52594c1aa480a736aef78313d15c9c)
1 /*	$NetBSD: xen_bus_dma.c,v 1.14 2009/01/24 19:03:12 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  *
21  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
23  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
24  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
25  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
26  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
27  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
28  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
29  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
30  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
31  * POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #include <sys/cdefs.h>
35 __KERNEL_RCSID(0, "$NetBSD: xen_bus_dma.c,v 1.14 2009/01/24 19:03:12 bouyer Exp $");
36 
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/mbuf.h>
41 #include <sys/proc.h>
42 
43 #include <machine/bus.h>
44 #include <machine/bus_private.h>
45 
46 #include <uvm/uvm_extern.h>
47 
48 extern paddr_t avail_end;
49 
50 /* Pure 2^n version of get_order */
51 static inline int get_order(unsigned long size)
52 {
53 	int order = -1;
54 	size = (size - 1) >> (PAGE_SHIFT - 1);
55 	do {
56 		size >>= 1;
57 		order++;
58 	} while (size);
59 	return order;
60 }
61 
62 static int
63 _xen_alloc_contig(bus_size_t size, bus_size_t alignment, bus_size_t boundary,
64     struct pglist *mlistp, int flags, bus_addr_t low, bus_addr_t high)
65 {
66 	int order, i;
67 	unsigned long npagesreq, npages, mfn;
68 	bus_addr_t pa;
69 	struct vm_page *pg, *pgnext;
70 	int s, error;
71 #ifdef XEN3
72 	struct xen_memory_reservation res;
73 #endif
74 
75 	/*
76 	 * When requesting a contigous memory region, the hypervisor will
77 	 * return a memory range aligned on size. This will automagically
78 	 * handle "boundary", but the only way to enforce alignment
79 	 * is to request a memory region of size max(alignment, size).
80 	 */
81 	order = max(get_order(size), get_order(alignment));
82 	npages = (1 << order);
83 	npagesreq = (size >> PAGE_SHIFT);
84 	KASSERT(npages >= npagesreq);
85 
86 	/* get npages from UWM, and give them back to the hypervisor */
87 	error = uvm_pglistalloc(npages << PAGE_SHIFT, 0, avail_end, 0, 0,
88 	    mlistp, npages, (flags & BUS_DMA_NOWAIT) == 0);
89 	if (error)
90 		return (error);
91 
92 	for (pg = mlistp->tqh_first; pg != NULL; pg = pg->pageq.queue.tqe_next) {
93 		pa = VM_PAGE_TO_PHYS(pg);
94 		mfn = xpmap_ptom(pa) >> PAGE_SHIFT;
95 		xpmap_phys_to_machine_mapping[
96 		    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = INVALID_P2M_ENTRY;
97 #ifdef XEN3
98 		xenguest_handle(res.extent_start) = &mfn;
99 		res.nr_extents = 1;
100 		res.extent_order = 0;
101 		res.domid = DOMID_SELF;
102 		if (HYPERVISOR_memory_op(XENMEM_decrease_reservation, &res)
103 		    != 1) {
104 #ifdef DEBUG
105 			printf("xen_alloc_contig: XENMEM_decrease_reservation "
106 			    "failed!\n");
107 #endif
108 			xpmap_phys_to_machine_mapping[
109 			    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn;
110 
111 			error = ENOMEM;
112 			goto failed;
113 		}
114 #else
115 		if (HYPERVISOR_dom_mem_op(MEMOP_decrease_reservation,
116 		    &mfn, 1, 0) != 1) {
117 #ifdef DEBUG
118 			printf("xen_alloc_contig: MEMOP_decrease_reservation "
119 			    "failed!\n");
120 #endif
121 			xpmap_phys_to_machine_mapping[
122 			    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn;
123 			error = ENOMEM;
124 			goto failed;
125 		}
126 #endif
127 	}
128 	/* Get the new contiguous memory extent */
129 #ifdef XEN3
130 	xenguest_handle(res.extent_start) = &mfn;
131 	res.nr_extents = 1;
132 	res.extent_order = order;
133 	res.address_bits = get_order(high) + PAGE_SHIFT;
134 	res.domid = DOMID_SELF;
135 	error = HYPERVISOR_memory_op(XENMEM_increase_reservation, &res);
136 	if (error != 1) {
137 #ifdef DEBUG
138 		printf("xen_alloc_contig: XENMEM_increase_reservation "
139 		    "failed: %d (order %d address_bits %d)\n",
140 		    error, order, res.address_bits);
141 #endif
142 		error = ENOMEM;
143 		pg = NULL;
144 		goto failed;
145 	}
146 #else
147 	if (HYPERVISOR_dom_mem_op(MEMOP_increase_reservation,
148 	    &mfn, 1, order) != 1) {
149 #ifdef DEBUG
150 		printf("xen_alloc_contig: MEMOP_increase_reservation "
151 		    "failed!\n");
152 #endif
153 		error = ENOMEM;
154 		pg = NULL;
155 		goto failed;
156 	}
157 #endif
158 	s = splvm();
159 	/* Map the new extent in place of the old pages */
160 	for (pg = mlistp->tqh_first, i = 0; pg != NULL; pg = pgnext, i++) {
161 		pgnext = pg->pageq.queue.tqe_next;
162 		pa = VM_PAGE_TO_PHYS(pg);
163 		xpmap_phys_to_machine_mapping[
164 		    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn+i;
165 		xpq_queue_machphys_update((mfn+i) << PAGE_SHIFT, pa);
166 		/* while here, give extra pages back to UVM */
167 		if (i >= npagesreq) {
168 			TAILQ_REMOVE(mlistp, pg, pageq.queue);
169 			uvm_pagefree(pg);
170 		}
171 	}
172 	/* Flush updates through and flush the TLB */
173 	xpq_queue_tlb_flush();
174 	xpq_flush_queue();
175 	splx(s);
176 	return 0;
177 
178 failed:
179 	/*
180 	 * Attempt to recover from a failed decrease or increase reservation:
181 	 * if decrease_reservation failed, we don't have given all pages
182 	 * back to Xen; give them back to UVM, and get the missing pages
183 	 * from Xen.
184 	 * if increase_reservation failed, we expect pg to be NULL and we just
185 	 * get back the missing pages from Xen one by one.
186 	 */
187 	/* give back remaining pages to UVM */
188 	for (; pg != NULL; pg = pgnext) {
189 		pgnext = pg->pageq.queue.tqe_next;
190 		TAILQ_REMOVE(mlistp, pg, pageq.queue);
191 		uvm_pagefree(pg);
192 	}
193 	/* remplace the pages that we already gave to Xen */
194 	s = splvm();
195 	for (pg = mlistp->tqh_first; pg != NULL; pg = pgnext) {
196 		pgnext = pg->pageq.queue.tqe_next;
197 #ifdef XEN3
198 		xenguest_handle(res.extent_start) = &mfn;
199 		res.nr_extents = 1;
200 		res.extent_order = 0;
201 		res.address_bits = 32;
202 		res.domid = DOMID_SELF;
203 		if (HYPERVISOR_memory_op(XENMEM_increase_reservation, &res)
204 		    < 0) {
205 			printf("xen_alloc_contig: recovery "
206 			    "XENMEM_increase_reservation failed!\n");
207 			break;
208 		}
209 #else
210 		if (HYPERVISOR_dom_mem_op(MEMOP_increase_reservation,
211 		    &mfn, 1, 0) != 1) {
212 			printf("xen_alloc_contig: recovery "
213 			    "MEMOP_increase_reservation failed!\n");
214 			break;
215 		}
216 #endif
217 		pa = VM_PAGE_TO_PHYS(pg);
218 		xpmap_phys_to_machine_mapping[
219 		    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn;
220 		xpq_queue_machphys_update((mfn) << PAGE_SHIFT, pa);
221 		TAILQ_REMOVE(mlistp, pg, pageq.queue);
222 		uvm_pagefree(pg);
223 	}
224 	/* Flush updates through and flush the TLB */
225 	xpq_queue_tlb_flush();
226 	xpq_flush_queue();
227 	splx(s);
228 	return error;
229 }
230 
231 
232 /*
233  * Allocate physical memory from the given physical address range.
234  * Called by DMA-safe memory allocation methods.
235  * We need our own version to deal with physical vs machine addresses.
236  */
237 int
238 _xen_bus_dmamem_alloc_range(bus_dma_tag_t t, bus_size_t size,
239     bus_size_t alignment, bus_size_t boundary, bus_dma_segment_t *segs,
240     int nsegs, int *rsegs, int flags, bus_addr_t low, bus_addr_t high)
241 {
242 	bus_addr_t curaddr, lastaddr;
243 	struct vm_page *m;
244 	struct pglist mlist;
245 	int curseg, error;
246 	int doingrealloc = 0;
247 
248 	/* Always round the size. */
249 	size = round_page(size);
250 
251 	KASSERT((alignment & (alignment - 1)) == 0);
252 	KASSERT((boundary & (boundary - 1)) == 0);
253 	if (alignment < PAGE_SIZE)
254 		alignment = PAGE_SIZE;
255 	if (boundary != 0 && boundary < size)
256 		return (EINVAL);
257 
258 	/*
259 	 * Allocate pages from the VM system.
260 	 */
261 	error = uvm_pglistalloc(size, 0, avail_end, alignment, boundary,
262 	    &mlist, nsegs, (flags & BUS_DMA_NOWAIT) == 0);
263 	if (error)
264 		return (error);
265 again:
266 
267 	/*
268 	 * Compute the location, size, and number of segments actually
269 	 * returned by the VM code.
270 	 */
271 	m = mlist.tqh_first;
272 	curseg = 0;
273 	curaddr = lastaddr = segs[curseg].ds_addr = _BUS_VM_PAGE_TO_BUS(m);
274 	if (curaddr < low || curaddr >= high)
275 		goto badaddr;
276 	segs[curseg].ds_len = PAGE_SIZE;
277 	m = m->pageq.queue.tqe_next;
278 	if ((segs[curseg].ds_addr & (alignment - 1)) != 0)
279 		goto dorealloc;
280 
281 	for (; m != NULL; m = m->pageq.queue.tqe_next) {
282 		curaddr = _BUS_VM_PAGE_TO_BUS(m);
283 		if (curaddr < low || curaddr >= high)
284 			goto badaddr;
285 		if (curaddr == (lastaddr + PAGE_SIZE)) {
286 			segs[curseg].ds_len += PAGE_SIZE;
287 			if ((lastaddr & boundary) != (curaddr & boundary))
288 				goto dorealloc;
289 		} else {
290 			curseg++;
291 			if (curseg >= nsegs || (curaddr & (alignment - 1)) != 0)
292 				goto dorealloc;
293 			segs[curseg].ds_addr = curaddr;
294 			segs[curseg].ds_len = PAGE_SIZE;
295 		}
296 		lastaddr = curaddr;
297 	}
298 
299 	*rsegs = curseg + 1;
300 	return (0);
301 
302 badaddr:
303 #ifdef XEN3
304 	if (doingrealloc == 0)
305 		goto dorealloc;
306 	if (curaddr < low) {
307 		/* no way to enforce this */
308 		printf("_xen_bus_dmamem_alloc_range: no way to "
309 		    "enforce address range (0x%" PRIx64 " - 0x%" PRIx64 ")\n",
310 		    (uint64_t)low, (uint64_t)high);
311 		uvm_pglistfree(&mlist);
312 		return EINVAL;
313 	}
314 	printf("xen_bus_dmamem_alloc_range: "
315 	    "curraddr=0x%lx > high=0x%lx\n",
316 	    (u_long)curaddr, (u_long)high);
317 	panic("xen_bus_dmamem_alloc_range 1");
318 #else /* !XEN3 */
319 	/*
320 	 * If machine addresses are outside the allowed
321 	 * range we have to bail. Xen2 doesn't offer an
322 	 * interface to get memory in a specific address
323 	 * range.
324 	 */
325 	printf("_xen_bus_dmamem_alloc_range: no way to "
326 	    "enforce address range\n");
327 	uvm_pglistfree(&mlist);
328 	return EINVAL;
329 #endif /* XEN3 */
330 dorealloc:
331 	if (doingrealloc == 1)
332 		panic("_xen_bus_dmamem_alloc_range: "
333 		   "xen_alloc_contig returned "
334 		   "too much segments");
335 	doingrealloc = 1;
336 	/*
337 	 * Too much segments, or memory doesn't fit
338 	 * constraints. Free this memory and
339 	 * get a contigous segment from the hypervisor.
340 	 */
341 	uvm_pglistfree(&mlist);
342 	for (curseg = 0; curseg < nsegs; curseg++) {
343 		segs[curseg].ds_addr = 0;
344 		segs[curseg].ds_len = 0;
345 	}
346 	error = _xen_alloc_contig(size, alignment,
347 	    boundary, &mlist, flags, low, high);
348 	if (error)
349 		return error;
350 	goto again;
351 }
352