xref: /illumos-gate/usr/src/uts/sun4/io/px/px_fdvma.c (revision 0ea48847)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /*
26  * Internal PCI Fast DVMA implementation
27  */
28 #include <sys/types.h>
29 #include <sys/kmem.h>
30 #include <sys/async.h>
31 #include <sys/sysmacros.h>
32 #include <sys/sunddi.h>
33 #include <sys/ddi_impldefs.h>
34 #include <sys/dvma.h>
35 #include "px_obj.h"
36 
37 /*LINTLIBRARY*/
38 
39 static struct dvma_ops fdvma_ops;
40 typedef struct fast_dvma fdvma_t;
41 
42 /*
43  * The following routines are used to implement the sun4u fast dvma
44  * routines on this bus.
45  */
46 
47 /*ARGSUSED*/
48 static void
px_fdvma_load(ddi_dma_handle_t h,caddr_t a,uint_t len,uint_t index,ddi_dma_cookie_t * cp)49 px_fdvma_load(ddi_dma_handle_t h, caddr_t a, uint_t len, uint_t index,
50 	ddi_dma_cookie_t *cp)
51 {
52 	ddi_dma_impl_t *mp = (ddi_dma_impl_t *)h;
53 	fdvma_t *fdvma_p = (fdvma_t *)mp->dmai_fdvma;
54 	px_t *px_p = (px_t *)fdvma_p->softsp;
55 	px_mmu_t *mmu_p = px_p->px_mmu_p;
56 	dev_info_t *dip = px_p->px_dip;
57 	px_dvma_addr_t dvma_addr, dvma_pg;
58 	uint32_t offset;
59 	size_t npages, pg_index;
60 	io_attributes_t attr;
61 
62 	offset = (uint32_t)(uintptr_t)a & MMU_PAGE_OFFSET;
63 	npages = MMU_BTOPR(len + offset);
64 	if (!npages)
65 		return;
66 
67 	/* make sure we don't exceed reserved boundary */
68 	DBG(DBG_FAST_DVMA, dip, "load index=%x: %p+%x ", index, a, len);
69 	if (index + npages > mp->dmai_ndvmapages) {
70 		cmn_err(px_panic_on_fatal_errors ? CE_PANIC : CE_WARN,
71 		    "%s%d: kaddr_load index(%x)+pgs(%lx) exceeds limit\n",
72 		    ddi_driver_name(dip), ddi_get_instance(dip),
73 		    index, npages);
74 		return;
75 	}
76 	fdvma_p->pagecnt[index] = npages;
77 
78 	dvma_addr = mp->dmai_mapping + MMU_PTOB(index);
79 	dvma_pg = MMU_BTOP(dvma_addr);
80 	pg_index = dvma_pg - mmu_p->dvma_base_pg;
81 
82 	/* construct the dma cookie to be returned */
83 	MAKE_DMA_COOKIE(cp, dvma_addr | offset, len);
84 	DBG(DBG_FAST_DVMA | DBG_CONT, dip, "cookie: %x+%x\n",
85 	    cp->dmac_address, cp->dmac_size);
86 
87 	attr = PX_GET_TTE_ATTR(mp->dmai_rflags, mp->dmai_attr.dma_attr_flags);
88 
89 	if (px_lib_iommu_map(dip, PCI_TSBID(0, pg_index), npages,
90 	    PX_ADD_ATTR_EXTNS(attr, mp->dmai_bdf), (void *)a, 0,
91 	    MMU_MAP_BUF) != DDI_SUCCESS) {
92 		cmn_err(CE_WARN, "%s%d: kaddr_load can't get "
93 		    "page frame for vaddr %lx", ddi_driver_name(dip),
94 		    ddi_get_instance(dip), (uintptr_t)a);
95 	}
96 }
97 
98 /*ARGSUSED*/
99 static void
px_fdvma_unload(ddi_dma_handle_t h,uint_t index,uint_t sync_flag)100 px_fdvma_unload(ddi_dma_handle_t h, uint_t index, uint_t sync_flag)
101 {
102 	ddi_dma_impl_t *mp = (ddi_dma_impl_t *)h;
103 	fdvma_t *fdvma_p = (fdvma_t *)mp->dmai_fdvma;
104 	px_t *px_p = (px_t *)fdvma_p->softsp;
105 	size_t npages = fdvma_p->pagecnt[index];
106 	px_dvma_addr_t dvma_pg = MMU_BTOP(mp->dmai_mapping + MMU_PTOB(index));
107 
108 	DBG(DBG_FAST_DVMA, px_p->px_dip,
109 	    "unload index=%x sync_flag=%x %x+%x+%x\n", index, sync_flag,
110 	    mp->dmai_mapping, MMU_PTOB(index), MMU_PTOB(npages));
111 
112 	px_mmu_unmap_pages(px_p->px_mmu_p, mp, dvma_pg, npages);
113 	fdvma_p->pagecnt[index] = 0;
114 }
115 
116 /*ARGSUSED*/
117 static void
px_fdvma_sync(ddi_dma_handle_t h,uint_t index,uint_t sync_flag)118 px_fdvma_sync(ddi_dma_handle_t h, uint_t index, uint_t sync_flag)
119 {
120 	ddi_dma_impl_t *mp = (ddi_dma_impl_t *)h;
121 	fdvma_t *fdvma_p = (fdvma_t *)mp->dmai_fdvma;
122 	px_t *px_p = (px_t *)fdvma_p->softsp;
123 	size_t npg = fdvma_p->pagecnt[index];
124 
125 	DBG(DBG_FAST_DVMA, px_p->px_dip,
126 	    "sync index=%x sync_flag=%x %x+%x+%x\n", index, sync_flag,
127 	    mp->dmai_mapping, MMU_PTOB(index), MMU_PTOB(npg));
128 }
129 
130 int
px_fdvma_reserve(dev_info_t * dip,dev_info_t * rdip,px_t * px_p,ddi_dma_req_t * dmareq,ddi_dma_handle_t * handlep)131 px_fdvma_reserve(dev_info_t *dip, dev_info_t *rdip, px_t *px_p,
132 	ddi_dma_req_t *dmareq, ddi_dma_handle_t *handlep)
133 {
134 	fdvma_t *fdvma_p;
135 	px_dvma_addr_t dvma_pg;
136 	px_mmu_t *mmu_p = px_p->px_mmu_p;
137 	size_t npages;
138 	ddi_dma_impl_t *mp;
139 	ddi_dma_lim_t *lim_p = dmareq->dmar_limits;
140 	ulong_t hi = lim_p->dlim_addr_hi;
141 	ulong_t lo = lim_p->dlim_addr_lo;
142 	size_t counter_max = (lim_p->dlim_cntr_max + 1) & MMU_PAGE_MASK;
143 
144 	if (px_disable_fdvma)
145 		return (DDI_FAILURE);
146 
147 	DBG(DBG_DMA_CTL, dip, "DDI_DMA_RESERVE: rdip=%s%d\n",
148 	    ddi_driver_name(rdip), ddi_get_instance(rdip));
149 
150 	/*
151 	 * Check the limit structure.
152 	 */
153 	if ((lo >= hi) || (hi < mmu_p->mmu_dvma_base))
154 		return (DDI_DMA_BADLIMITS);
155 
156 	/*
157 	 * Allocate DVMA space from reserve.
158 	 */
159 	npages = dmareq->dmar_object.dmao_size;
160 	if ((long)atomic_add_long_nv(&mmu_p->mmu_dvma_reserve, -npages) < 0) {
161 		atomic_add_long(&mmu_p->mmu_dvma_reserve, npages);
162 		return (DDI_DMA_NORESOURCES);
163 	}
164 
165 	/*
166 	 * Allocate the dma handle.
167 	 */
168 	mp = kmem_zalloc(sizeof (px_dma_hdl_t), KM_SLEEP);
169 
170 	/*
171 	 * Get entries from dvma space map.
172 	 * (vmem_t *vmp,
173 	 *	size_t size, size_t align, size_t phase,
174 	 *	size_t nocross, void *minaddr, void *maxaddr, int vmflag)
175 	 */
176 	dvma_pg = MMU_BTOP((ulong_t)vmem_xalloc(mmu_p->mmu_dvma_map,
177 	    MMU_PTOB(npages), MMU_PAGE_SIZE, 0,
178 	    counter_max, (void *)lo, (void *)(hi + 1),
179 	    dmareq->dmar_fp == DDI_DMA_SLEEP ? VM_SLEEP : VM_NOSLEEP));
180 	if (dvma_pg == 0) {
181 		atomic_add_long(&mmu_p->mmu_dvma_reserve, npages);
182 		kmem_free(mp, sizeof (px_dma_hdl_t));
183 		return (DDI_DMA_NOMAPPING);
184 	}
185 
186 	/*
187 	 * Create the fast dvma request structure.
188 	 */
189 	fdvma_p = kmem_alloc(sizeof (fdvma_t), KM_SLEEP);
190 	fdvma_p->pagecnt = kmem_alloc(npages * sizeof (uint_t), KM_SLEEP);
191 	fdvma_p->ops = &fdvma_ops;
192 	fdvma_p->softsp = (caddr_t)px_p;
193 	fdvma_p->sync_flag = NULL;
194 
195 	/*
196 	 * Initialize the handle.
197 	 */
198 	mp->dmai_rdip = rdip;
199 	mp->dmai_rflags = DMP_BYPASSNEXUS | DDI_DMA_READ | DMP_NOSYNC;
200 	mp->dmai_burstsizes = dmareq->dmar_limits->dlim_burstsizes;
201 	mp->dmai_mapping = MMU_PTOB(dvma_pg);
202 	mp->dmai_ndvmapages = npages;
203 	mp->dmai_size = npages * MMU_PAGE_SIZE;
204 	mp->dmai_nwin = 0;
205 	mp->dmai_fdvma = (caddr_t)fdvma_p;
206 
207 	/*
208 	 * The bdf protection value is set to immediate child
209 	 * at first. It gets modified by switch/bridge drivers
210 	 * as the code traverses down the fabric topology.
211 	 *
212 	 * XXX No IOMMU protection for broken devices.
213 	 */
214 	ASSERT((intptr_t)ddi_get_parent_data(rdip) >> 1 == 0);
215 	mp->dmai_bdf = ((intptr_t)ddi_get_parent_data(rdip) == 1) ?
216 	    PCIE_INVALID_BDF : pcie_get_bdf_for_dma_xfer(dip, rdip);
217 
218 	DBG(DBG_DMA_CTL, dip,
219 	    "DDI_DMA_RESERVE: mp=%p dvma=%x npages=%x private=%p\n",
220 	    mp, mp->dmai_mapping, npages, fdvma_p);
221 	*handlep = (ddi_dma_handle_t)mp;
222 	return (DDI_SUCCESS);
223 }
224 
225 int
px_fdvma_release(dev_info_t * dip,px_t * px_p,ddi_dma_impl_t * mp)226 px_fdvma_release(dev_info_t *dip, px_t *px_p, ddi_dma_impl_t *mp)
227 {
228 	px_mmu_t *mmu_p = px_p->px_mmu_p;
229 	size_t npages;
230 	fdvma_t *fdvma_p = (fdvma_t *)mp->dmai_fdvma;
231 
232 	if (px_disable_fdvma)
233 		return (DDI_FAILURE);
234 
235 	/* validate fdvma handle */
236 	if (!(mp->dmai_rflags & DMP_BYPASSNEXUS)) {
237 		DBG(DBG_DMA_CTL, dip, "DDI_DMA_RELEASE: not fast dma\n");
238 		return (DDI_FAILURE);
239 	}
240 
241 	/* flush all reserved dvma addresses from mmu */
242 	px_mmu_unmap_window(mmu_p, mp);
243 
244 	npages = mp->dmai_ndvmapages;
245 	vmem_xfree(mmu_p->mmu_dvma_map, (void *)mp->dmai_mapping,
246 	    MMU_PTOB(npages));
247 
248 	atomic_add_long(&mmu_p->mmu_dvma_reserve, npages);
249 	mp->dmai_ndvmapages = 0;
250 
251 	/* see if there is anyone waiting for dvma space */
252 	if (mmu_p->mmu_dvma_clid != 0) {
253 		DBG(DBG_DMA_CTL, dip, "run dvma callback\n");
254 		ddi_run_callback(&mmu_p->mmu_dvma_clid);
255 	}
256 
257 	/* free data structures */
258 	kmem_free(fdvma_p->pagecnt, npages * sizeof (uint_t));
259 	kmem_free(fdvma_p, sizeof (fdvma_t));
260 	kmem_free(mp, sizeof (px_dma_hdl_t));
261 
262 	/* see if there is anyone waiting for kmem */
263 	if (px_kmem_clid != 0) {
264 		DBG(DBG_DMA_CTL, dip, "run handle callback\n");
265 		ddi_run_callback(&px_kmem_clid);
266 	}
267 	return (DDI_SUCCESS);
268 }
269 
270 static struct dvma_ops fdvma_ops = {
271 	DVMAO_REV,
272 	px_fdvma_load,
273 	px_fdvma_unload,
274 	px_fdvma_sync
275 };
276