xref: /netbsd-src/sys/arch/xen/x86/xen_bus_dma.c (revision 4e6df137e8e14049b5a701d249962c480449c141)
1 /*	$NetBSD: xen_bus_dma.c,v 1.19 2010/03/02 00:13:50 jym 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 1.19 2010/03/02 00:13:50 jym 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 	struct xen_memory_reservation res;
72 
73 	/*
74 	 * When requesting a contigous memory region, the hypervisor will
75 	 * return a memory range aligned on size. This will automagically
76 	 * handle "boundary", but the only way to enforce alignment
77 	 * is to request a memory region of size max(alignment, size).
78 	 */
79 	order = max(get_order(size), get_order(alignment));
80 	npages = (1 << order);
81 	npagesreq = (size >> PAGE_SHIFT);
82 	KASSERT(npages >= npagesreq);
83 
84 	/* get npages from UVM, and give them back to the hypervisor */
85 	error = uvm_pglistalloc(((psize_t)npages) << PAGE_SHIFT,
86             0, avail_end, 0, 0, mlistp, npages, (flags & BUS_DMA_NOWAIT) == 0);
87 	if (error)
88 		return (error);
89 
90 	for (pg = mlistp->tqh_first; pg != NULL; pg = pg->pageq.queue.tqe_next) {
91 		pa = VM_PAGE_TO_PHYS(pg);
92 		mfn = xpmap_ptom(pa) >> PAGE_SHIFT;
93 		xpmap_phys_to_machine_mapping[
94 		    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = INVALID_P2M_ENTRY;
95 		xenguest_handle(res.extent_start) = &mfn;
96 		res.nr_extents = 1;
97 		res.extent_order = 0;
98 		res.domid = DOMID_SELF;
99 		error = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &res);
100 		if (error != 1) {
101 #ifdef DEBUG
102 			printf("xen_alloc_contig: XENMEM_decrease_reservation "
103 			    "failed: err %d (pa %#" PRIxPADDR " mfn %#lx)\n",
104 			    error, pa, mfn);
105 #endif
106 			xpmap_phys_to_machine_mapping[
107 			    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn;
108 
109 			error = ENOMEM;
110 			goto failed;
111 		}
112 	}
113 	/* Get the new contiguous memory extent */
114 	xenguest_handle(res.extent_start) = &mfn;
115 	res.nr_extents = 1;
116 	res.extent_order = order;
117 	res.address_bits = get_order(high) + PAGE_SHIFT;
118 	res.domid = DOMID_SELF;
119 	error = HYPERVISOR_memory_op(XENMEM_increase_reservation, &res);
120 	if (error != 1) {
121 #ifdef DEBUG
122 		printf("xen_alloc_contig: XENMEM_increase_reservation "
123 		    "failed: %d (order %d address_bits %d)\n",
124 		    error, order, res.address_bits);
125 #endif
126 		error = ENOMEM;
127 		pg = NULL;
128 		goto failed;
129 	}
130 	s = splvm();
131 	/* Map the new extent in place of the old pages */
132 	for (pg = mlistp->tqh_first, i = 0; pg != NULL; pg = pgnext, i++) {
133 		pgnext = pg->pageq.queue.tqe_next;
134 		pa = VM_PAGE_TO_PHYS(pg);
135 		xpmap_phys_to_machine_mapping[
136 		    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn+i;
137 		xpq_queue_machphys_update(((paddr_t)(mfn+i)) << PAGE_SHIFT, pa);
138 		/* while here, give extra pages back to UVM */
139 		if (i >= npagesreq) {
140 			TAILQ_REMOVE(mlistp, pg, pageq.queue);
141 			uvm_pagefree(pg);
142 		}
143 	}
144 	/* Flush updates through and flush the TLB */
145 	xpq_queue_tlb_flush();
146 	splx(s);
147 	return 0;
148 
149 failed:
150 	/*
151 	 * Attempt to recover from a failed decrease or increase reservation:
152 	 * if decrease_reservation failed, we don't have given all pages
153 	 * back to Xen; give them back to UVM, and get the missing pages
154 	 * from Xen.
155 	 * if increase_reservation failed, we expect pg to be NULL and we just
156 	 * get back the missing pages from Xen one by one.
157 	 */
158 	/* give back remaining pages to UVM */
159 	for (; pg != NULL; pg = pgnext) {
160 		pgnext = pg->pageq.queue.tqe_next;
161 		TAILQ_REMOVE(mlistp, pg, pageq.queue);
162 		uvm_pagefree(pg);
163 	}
164 	/* remplace the pages that we already gave to Xen */
165 	s = splvm();
166 	for (pg = mlistp->tqh_first; pg != NULL; pg = pgnext) {
167 		pgnext = pg->pageq.queue.tqe_next;
168 		xenguest_handle(res.extent_start) = &mfn;
169 		res.nr_extents = 1;
170 		res.extent_order = 0;
171 		res.address_bits = 32;
172 		res.domid = DOMID_SELF;
173 		if (HYPERVISOR_memory_op(XENMEM_increase_reservation, &res)
174 		    < 0) {
175 			printf("xen_alloc_contig: recovery "
176 			    "XENMEM_increase_reservation failed!\n");
177 			break;
178 		}
179 		pa = VM_PAGE_TO_PHYS(pg);
180 		xpmap_phys_to_machine_mapping[
181 		    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn;
182 		xpq_queue_machphys_update(((paddr_t)mfn) << PAGE_SHIFT, pa);
183 		TAILQ_REMOVE(mlistp, pg, pageq.queue);
184 		uvm_pagefree(pg);
185 	}
186 	/* Flush updates through and flush the TLB */
187 	xpq_queue_tlb_flush();
188 	splx(s);
189 	return error;
190 }
191 
192 
193 /*
194  * Allocate physical memory from the given physical address range.
195  * Called by DMA-safe memory allocation methods.
196  * We need our own version to deal with physical vs machine addresses.
197  */
198 int
199 _xen_bus_dmamem_alloc_range(bus_dma_tag_t t, bus_size_t size,
200     bus_size_t alignment, bus_size_t boundary, bus_dma_segment_t *segs,
201     int nsegs, int *rsegs, int flags, bus_addr_t low, bus_addr_t high)
202 {
203 	bus_addr_t curaddr, lastaddr;
204 	struct vm_page *m;
205 	struct pglist mlist;
206 	int curseg, error;
207 	int doingrealloc = 0;
208 
209 	/* Always round the size. */
210 	size = round_page(size);
211 
212 	KASSERT((alignment & (alignment - 1)) == 0);
213 	KASSERT((boundary & (boundary - 1)) == 0);
214 	if (alignment < PAGE_SIZE)
215 		alignment = PAGE_SIZE;
216 	if (boundary != 0 && boundary < size)
217 		return (EINVAL);
218 
219 	/*
220 	 * Allocate pages from the VM system.
221 	 */
222 	error = uvm_pglistalloc(size, 0, avail_end, alignment, boundary,
223 	    &mlist, nsegs, (flags & BUS_DMA_NOWAIT) == 0);
224 	if (error)
225 		return (error);
226 again:
227 
228 	/*
229 	 * Compute the location, size, and number of segments actually
230 	 * returned by the VM code.
231 	 */
232 	m = mlist.tqh_first;
233 	curseg = 0;
234 	curaddr = lastaddr = segs[curseg].ds_addr = _BUS_VM_PAGE_TO_BUS(m);
235 	if (curaddr < low || curaddr >= high)
236 		goto badaddr;
237 	segs[curseg].ds_len = PAGE_SIZE;
238 	m = m->pageq.queue.tqe_next;
239 	if ((segs[curseg].ds_addr & (alignment - 1)) != 0)
240 		goto dorealloc;
241 
242 	for (; m != NULL; m = m->pageq.queue.tqe_next) {
243 		curaddr = _BUS_VM_PAGE_TO_BUS(m);
244 		if (curaddr < low || curaddr >= high)
245 			goto badaddr;
246 		if (curaddr == (lastaddr + PAGE_SIZE)) {
247 			segs[curseg].ds_len += PAGE_SIZE;
248 			if ((lastaddr & boundary) != (curaddr & boundary))
249 				goto dorealloc;
250 		} else {
251 			curseg++;
252 			if (curseg >= nsegs || (curaddr & (alignment - 1)) != 0)
253 				goto dorealloc;
254 			segs[curseg].ds_addr = curaddr;
255 			segs[curseg].ds_len = PAGE_SIZE;
256 		}
257 		lastaddr = curaddr;
258 	}
259 
260 	*rsegs = curseg + 1;
261 	return (0);
262 
263 badaddr:
264 	if (doingrealloc == 0)
265 		goto dorealloc;
266 	if (curaddr < low) {
267 		/* no way to enforce this */
268 		printf("_xen_bus_dmamem_alloc_range: no way to "
269 		    "enforce address range (0x%" PRIx64 " - 0x%" PRIx64 ")\n",
270 		    (uint64_t)low, (uint64_t)high);
271 		uvm_pglistfree(&mlist);
272 		return EINVAL;
273 	}
274 	printf("xen_bus_dmamem_alloc_range: "
275 	    "curraddr=0x%lx > high=0x%lx\n",
276 	    (u_long)curaddr, (u_long)high);
277 	panic("xen_bus_dmamem_alloc_range 1");
278 dorealloc:
279 	if (doingrealloc == 1)
280 		panic("_xen_bus_dmamem_alloc_range: "
281 		   "xen_alloc_contig returned "
282 		   "too much segments");
283 	doingrealloc = 1;
284 	/*
285 	 * Too much segments, or memory doesn't fit
286 	 * constraints. Free this memory and
287 	 * get a contigous segment from the hypervisor.
288 	 */
289 	uvm_pglistfree(&mlist);
290 	for (curseg = 0; curseg < nsegs; curseg++) {
291 		segs[curseg].ds_addr = 0;
292 		segs[curseg].ds_len = 0;
293 	}
294 	error = _xen_alloc_contig(size, alignment,
295 	    boundary, &mlist, flags, low, high);
296 	if (error)
297 		return error;
298 	goto again;
299 }
300