Home | History | Annotate | Line # | Download | only in x86
xen_bus_dma.c revision 1.3.2.5
      1 /*	$NetBSD: xen_bus_dma.c,v 1.3.2.5 2006/09/16 11:18:59 ghen 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 1.3.2.5 2006/09/16 11:18:59 ghen 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)
     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 
     80 	/*
     81 	 * When requesting a contigous memory region, the hypervisor will
     82 	 * return a memory range aligned on size. This will automagically
     83 	 * handle "boundary", but the only way to enforce alignment
     84 	 * is to request a memory region of size max(alignment, size).
     85 	 */
     86 	order = max(get_order(size), get_order(alignment));
     87 	npages = (1 << order);
     88 	npagesreq = (size >> PAGE_SHIFT);
     89 	KASSERT(npages >= npagesreq);
     90 
     91 	/* get npages from UWM, and give them back to the hypervisor */
     92 	error = uvm_pglistalloc(npages << PAGE_SHIFT, 0, avail_end, 0, 0,
     93 	    mlistp, npages, (flags & BUS_DMA_NOWAIT) == 0);
     94 	if (error)
     95 		return (error);
     96 
     97 	for (pg = mlistp->tqh_first; pg != NULL; pg = pg->pageq.tqe_next) {
     98 		pa = VM_PAGE_TO_PHYS(pg);
     99 		mfn = xpmap_ptom(pa) >> PAGE_SHIFT;
    100 		xpmap_phys_to_machine_mapping[
    101 		    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = INVALID_P2M_ENTRY;
    102 		if (HYPERVISOR_dom_mem_op(MEMOP_decrease_reservation,
    103 		    &mfn, 1, 0) != 1) {
    104 #ifdef DEBUG
    105 			printf("xen_alloc_contig: MEMOP_decrease_reservation "
    106 			    "failed!\n");
    107 #endif
    108 			xpmap_phys_to_machine_mapping[
    109 			    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn;
    110 			error = ENOMEM;
    111 			goto failed;
    112 		}
    113 	}
    114 	/* Get the new contiguous memory extent */
    115 	if (HYPERVISOR_dom_mem_op(MEMOP_increase_reservation,
    116 	    &mfn, 1, order) != 1) {
    117 #ifdef DEBUG
    118 		printf("xen_alloc_contig: MEMOP_increase_reservation "
    119 		    "failed!\n");
    120 #endif
    121 		error = ENOMEM;
    122 		pg = NULL;
    123 		goto failed;
    124 	}
    125 	s = splvm();
    126 	/* Map the new extent in place of the old pages */
    127 	for (pg = mlistp->tqh_first, i = 0; pg != NULL; pg = pgnext, i++) {
    128 		pgnext = pg->pageq.tqe_next;
    129 		pa = VM_PAGE_TO_PHYS(pg);
    130 		xpmap_phys_to_machine_mapping[
    131 		    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn+i;
    132 		xpq_queue_machphys_update((mfn+i) << PAGE_SHIFT, pa);
    133 		/* while here, give extra pages back to UVM */
    134 		if (i >= npagesreq) {
    135 			TAILQ_REMOVE(mlistp, pg, pageq);
    136 			uvm_pagefree(pg);
    137 		}
    138 
    139 	}
    140 	/* Flush updates through and flush the TLB */
    141 	xpq_queue_tlb_flush();
    142 	xpq_flush_queue();
    143 	splx(s);
    144 	return 0;
    145 
    146 failed:
    147 	/*
    148 	 * Attempt to recover from a failed decrease or increase reservation:
    149 	 * if decrease_reservation failed, we don't have given all pages
    150 	 * back to Xen; give them back to UVM, and get the missing pages
    151 	 * from Xen.
    152 	 * if increase_reservation failed, we expect pg to be NULL and we just
    153 	 * get back the missing pages from Xen one by one.
    154 	 */
    155 	/* give back remaining pages to UVM */
    156 	for (; pg != NULL; pg = pgnext) {
    157 		pgnext = pg->pageq.tqe_next;
    158 		TAILQ_REMOVE(mlistp, pg, pageq);
    159 		uvm_pagefree(pg);
    160 	}
    161 	/* remplace the pages that we already gave to Xen */
    162 	s = splvm();
    163 	for (pg = mlistp->tqh_first; pg != NULL; pg = pgnext) {
    164 		pgnext = pg->pageq.tqe_next;
    165 		if (HYPERVISOR_dom_mem_op(MEMOP_increase_reservation,
    166 		    &mfn, 1, 0) != 1) {
    167 			printf("xen_alloc_contig: recovery "
    168 			    "MEMOP_increase_reservation failed!\n");
    169 			break;
    170 		}
    171 		pa = VM_PAGE_TO_PHYS(pg);
    172 		xpmap_phys_to_machine_mapping[
    173 		    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn;
    174 		xpq_queue_machphys_update((mfn) << PAGE_SHIFT, pa);
    175 		TAILQ_REMOVE(mlistp, pg, pageq);
    176 		uvm_pagefree(pg);
    177 	}
    178 	/* Flush updates through and flush the TLB */
    179 	xpq_queue_tlb_flush();
    180 	xpq_flush_queue();
    181 	splx(s);
    182 	return error;
    183 }
    184 
    185 
    186 /*
    187  * Allocate physical memory from the given physical address range.
    188  * Called by DMA-safe memory allocation methods.
    189  * We need our own version to deal with physical vs machine addresses.
    190  */
    191 int
    192 _xen_bus_dmamem_alloc_range(bus_dma_tag_t t, bus_size_t size,
    193     bus_size_t alignment, bus_size_t boundary, bus_dma_segment_t *segs,
    194     int nsegs, int *rsegs, int flags, bus_addr_t low, bus_addr_t high)
    195 {
    196 	bus_addr_t curaddr, lastaddr;
    197 	struct vm_page *m;
    198 	struct pglist mlist;
    199 	int curseg, error;
    200 	int doingrealloc = 0;
    201 
    202 	/* Always round the size. */
    203 	size = round_page(size);
    204 
    205 	KASSERT((alignment & (alignment - 1)) == 0);
    206 	KASSERT((boundary & (boundary - 1)) == 0);
    207 	if (alignment < PAGE_SIZE)
    208 		alignment = PAGE_SIZE;
    209 	if (boundary != 0 && boundary < size)
    210 		return (EINVAL);
    211 
    212 	/*
    213 	 * Allocate pages from the VM system.
    214 	 */
    215 	error = uvm_pglistalloc(size, 0, avail_end, alignment, boundary,
    216 	    &mlist, nsegs, (flags & BUS_DMA_NOWAIT) == 0);
    217 	if (error)
    218 		return (error);
    219 again:
    220 
    221 	/*
    222 	 * Compute the location, size, and number of segments actually
    223 	 * returned by the VM code.
    224 	 */
    225 	m = mlist.tqh_first;
    226 	curseg = 0;
    227 	lastaddr = segs[curseg].ds_addr = _BUS_VM_PAGE_TO_BUS(m);
    228 	segs[curseg].ds_len = PAGE_SIZE;
    229 	m = m->pageq.tqe_next;
    230 	if ((segs[curseg].ds_addr & (alignment - 1)) != 0)
    231 		goto dorealloc;
    232 
    233 	for (; m != NULL; m = m->pageq.tqe_next) {
    234 		curaddr = _BUS_VM_PAGE_TO_BUS(m);
    235 		if ((lastaddr < low || lastaddr >= high) ||
    236 		    (curaddr < low || curaddr >= high)) {
    237 			/*
    238 			 * If machine addresses are outside the allowed
    239 			 * range we have to bail. Xen2 doesn't offer an
    240 			 * interface to get memory in a specific address
    241 			 * range.
    242 			 */
    243 			printf("_xen_bus_dmamem_alloc_range: no way to "
    244 			    "enforce address range\n");
    245 			uvm_pglistfree(&mlist);
    246 			return EINVAL;
    247 		}
    248 		if (curaddr == (lastaddr + PAGE_SIZE)) {
    249 			segs[curseg].ds_len += PAGE_SIZE;
    250 			if ((lastaddr & boundary) !=
    251 			    (curaddr & boundary))
    252 				goto dorealloc;
    253 		} else {
    254 			curseg++;
    255 			if (curseg >= nsegs ||
    256 			    (curaddr & (alignment - 1)) != 0) {
    257 dorealloc:
    258 				if (doingrealloc == 1)
    259 					panic("_xen_bus_dmamem_alloc_range: "
    260 					   "xen_alloc_contig returned "
    261 					   "too much segments");
    262 				doingrealloc = 1;
    263 				/*
    264 				 * Too much segments. Free this memory and
    265 				 * get a contigous segment from the hypervisor.
    266 				 */
    267 				uvm_pglistfree(&mlist);
    268 				for (curseg = 0; curseg < nsegs; curseg++) {
    269 					segs[curseg].ds_addr = 0;
    270 					segs[curseg].ds_len = 0;
    271 				}
    272 				error = _xen_alloc_contig(size, alignment,
    273 				    boundary, &mlist, flags);
    274 				if (error)
    275 					return error;
    276 				goto again;
    277 			}
    278 			segs[curseg].ds_addr = curaddr;
    279 			segs[curseg].ds_len = PAGE_SIZE;
    280 		}
    281 		lastaddr = curaddr;
    282 	}
    283 
    284 	*rsegs = curseg + 1;
    285 
    286 	return (0);
    287 }
    288