forked from ~ljy/RK356X_SDK_RELEASE

blame | history | patch | commit | commitdiff
hc
2024-05-16 8d2a02b24d66aa359e83eebc1ed3c0f85367a1cb 
 kernel/drivers/gpu/host1x/job.c
....@@ -1,24 +1,14 @@
1
+// SPDX-License-Identifier: GPL-2.0-only
12 /*
23  * Tegra host1x Job
34  *
45  * Copyright (c) 2010-2015, NVIDIA Corporation.
5
- *
6
- * This program is free software; you can redistribute it and/or modify it
7
- * under the terms and conditions of the GNU General Public License,
8
- * version 2, as published by the Free Software Foundation.
9
- *
10
- * This program is distributed in the hope it will be useful, but WITHOUT
11
- * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12
- * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
13
- * more details.
14
- *
15
- * You should have received a copy of the GNU General Public License
16
- * along with this program.  If not, see <http://www.gnu.org/licenses/>.
176  */
187 
198 #include <linux/dma-mapping.h>
209 #include <linux/err.h>
2110 #include <linux/host1x.h>
11
+#include <linux/iommu.h>
2212 #include <linux/kref.h>
2313 #include <linux/module.h>
2414 #include <linux/scatterlist.h>
....@@ -37,9 +27,12 @@
3727 				    u32 num_cmdbufs, u32 num_relocs)
3828 {
3929 	struct host1x_job *job = NULL;
40
-	unsigned int num_unpins = num_cmdbufs + num_relocs;
30
+	unsigned int num_unpins = num_relocs;
4131 	u64 total;
4232 	void *mem;
33
+
34
+	if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL))
35
+		num_unpins += num_cmdbufs;
4336 
4437 	/* Check that we're not going to overflow */
4538 	total = sizeof(struct host1x_job) +
....@@ -110,15 +103,20 @@
110103 
111104 static unsigned int pin_job(struct host1x *host, struct host1x_job *job)
112105 {
106
+	struct host1x_client *client = job->client;
107
+	struct device *dev = client->dev;
108
+	struct host1x_job_gather *g;
109
+	struct iommu_domain *domain;
113110 	unsigned int i;
114111 	int err;
115112 
113
+	domain = iommu_get_domain_for_dev(dev);
116114 	job->num_unpins = 0;
117115 
118116 	for (i = 0; i < job->num_relocs; i++) {
119117 		struct host1x_reloc *reloc = &job->relocs[i];
118
+		dma_addr_t phys_addr, *phys;
120119 		struct sg_table *sgt;
121
-		dma_addr_t phys_addr;
122120 
123121 		reloc->target.bo = host1x_bo_get(reloc->target.bo);
124122 		if (!reloc->target.bo) {
....@@ -126,7 +124,60 @@
126124 			goto unpin;
127125 		}
128126 
129
-		phys_addr = host1x_bo_pin(reloc->target.bo, &sgt);
127
+		/*
128
+		 * If the client device is not attached to an IOMMU, the
129
+		 * physical address of the buffer object can be used.
130
+		 *
131
+		 * Similarly, when an IOMMU domain is shared between all
132
+		 * host1x clients, the IOVA is already available, so no
133
+		 * need to map the buffer object again.
134
+		 *
135
+		 * XXX Note that this isn't always safe to do because it
136
+		 * relies on an assumption that no cache maintenance is
137
+		 * needed on the buffer objects.
138
+		 */
139
+		if (!domain || client->group)
140
+			phys = &phys_addr;
141
+		else
142
+			phys = NULL;
143
+
144
+		sgt = host1x_bo_pin(dev, reloc->target.bo, phys);
145
+		if (IS_ERR(sgt)) {
146
+			err = PTR_ERR(sgt);
147
+			goto unpin;
148
+		}
149
+
150
+		if (sgt) {
151
+			unsigned long mask = HOST1X_RELOC_READ |
152
+					     HOST1X_RELOC_WRITE;
153
+			enum dma_data_direction dir;
154
+
155
+			switch (reloc->flags & mask) {
156
+			case HOST1X_RELOC_READ:
157
+				dir = DMA_TO_DEVICE;
158
+				break;
159
+
160
+			case HOST1X_RELOC_WRITE:
161
+				dir = DMA_FROM_DEVICE;
162
+				break;
163
+
164
+			case HOST1X_RELOC_READ | HOST1X_RELOC_WRITE:
165
+				dir = DMA_BIDIRECTIONAL;
166
+				break;
167
+
168
+			default:
169
+				err = -EINVAL;
170
+				goto unpin;
171
+			}
172
+
173
+			err = dma_map_sgtable(dev, sgt, dir, 0);
174
+			if (err)
175
+				goto unpin;
176
+
177
+			job->unpins[job->num_unpins].dev = dev;
178
+			job->unpins[job->num_unpins].dir = dir;
179
+			phys_addr = sg_dma_address(sgt->sgl);
180
+		}
130181 
131182 		job->addr_phys[job->num_unpins] = phys_addr;
132183 		job->unpins[job->num_unpins].bo = reloc->target.bo;
....@@ -134,26 +185,48 @@
134185 		job->num_unpins++;
135186 	}
136187 
188
+	/*
189
+	 * We will copy gathers BO content later, so there is no need to
190
+	 * hold and pin them.
191
+	 */
192
+	if (IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL))
193
+		return 0;
194
+
137195 	for (i = 0; i < job->num_gathers; i++) {
138
-		struct host1x_job_gather *g = &job->gathers[i];
139196 		size_t gather_size = 0;
140197 		struct scatterlist *sg;
141198 		struct sg_table *sgt;
142199 		dma_addr_t phys_addr;
143200 		unsigned long shift;
144201 		struct iova *alloc;
202
+		dma_addr_t *phys;
145203 		unsigned int j;
146204 
205
+		g = &job->gathers[i];
147206 		g->bo = host1x_bo_get(g->bo);
148207 		if (!g->bo) {
149208 			err = -EINVAL;
150209 			goto unpin;
151210 		}
152211 
153
-		phys_addr = host1x_bo_pin(g->bo, &sgt);
212
+		/**
213
+		 * If the host1x is not attached to an IOMMU, there is no need
214
+		 * to map the buffer object for the host1x, since the physical
215
+		 * address can simply be used.
216
+		 */
217
+		if (!iommu_get_domain_for_dev(host->dev))
218
+			phys = &phys_addr;
219
+		else
220
+			phys = NULL;
154221 
155
-		if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL) && host->domain) {
156
-			for_each_sg(sgt->sgl, sg, sgt->nents, j)
222
+		sgt = host1x_bo_pin(host->dev, g->bo, phys);
223
+		if (IS_ERR(sgt)) {
224
+			err = PTR_ERR(sgt);
225
+			goto put;
226
+		}
227
+
228
+		if (host->domain) {
229
+			for_each_sgtable_sg(sgt, sg, j)
157230 				gather_size += sg->length;
158231 			gather_size = iova_align(&host->iova, gather_size);
159232 
....@@ -162,26 +235,32 @@
162235 					   host->iova_end >> shift, true);
163236 			if (!alloc) {
164237 				err = -ENOMEM;
165
-				goto unpin;
238
+				goto put;
166239 			}
167240 
168
-			err = iommu_map_sg(host->domain,
241
+			err = iommu_map_sgtable(host->domain,
169242 					iova_dma_addr(&host->iova, alloc),
170
-					sgt->sgl, sgt->nents, IOMMU_READ);
243
+					sgt, IOMMU_READ);
171244 			if (err == 0) {
172245 				__free_iova(&host->iova, alloc);
173246 				err = -EINVAL;
174
-				goto unpin;
247
+				goto put;
175248 			}
176249 
177
-			job->addr_phys[job->num_unpins] =
178
-				iova_dma_addr(&host->iova, alloc);
179250 			job->unpins[job->num_unpins].size = gather_size;
180
-		} else {
181
-			job->addr_phys[job->num_unpins] = phys_addr;
251
+			phys_addr = iova_dma_addr(&host->iova, alloc);
252
+		} else if (sgt) {
253
+			err = dma_map_sgtable(host->dev, sgt, DMA_TO_DEVICE, 0);
254
+			if (err)
255
+				goto put;
256
+
257
+			job->unpins[job->num_unpins].dir = DMA_TO_DEVICE;
258
+			job->unpins[job->num_unpins].dev = host->dev;
259
+			phys_addr = sg_dma_address(sgt->sgl);
182260 		}
183261 
184
-		job->gather_addr_phys[i] = job->addr_phys[job->num_unpins];
262
+		job->addr_phys[job->num_unpins] = phys_addr;
263
+		job->gather_addr_phys[i] = phys_addr;
185264 
186265 		job->unpins[job->num_unpins].bo = g->bo;
187266 		job->unpins[job->num_unpins].sgt = sgt;
....@@ -190,6 +269,8 @@
190269 
191270 	return 0;
192271 
272
+put:
273
+	host1x_bo_put(g->bo);
193274 unpin:
194275 	host1x_job_unpin(job);
195276 	return err;
....@@ -197,8 +278,7 @@
197278 
198279 static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g)
199280 {
200
-	u32 last_page = ~0;
201
-	void *cmdbuf_page_addr = NULL;
281
+	void *cmdbuf_addr = NULL;
202282 	struct host1x_bo *cmdbuf = g->bo;
203283 	unsigned int i;
204284 
....@@ -220,28 +300,22 @@
220300 			goto patch_reloc;
221301 		}
222302 
223
-		if (last_page != reloc->cmdbuf.offset >> PAGE_SHIFT) {
224
-			if (cmdbuf_page_addr)
225
-				host1x_bo_kunmap(cmdbuf, last_page,
226
-						 cmdbuf_page_addr);
303
+		if (!cmdbuf_addr) {
304
+			cmdbuf_addr = host1x_bo_mmap(cmdbuf);
227305 
228
-			cmdbuf_page_addr = host1x_bo_kmap(cmdbuf,
229
-					reloc->cmdbuf.offset >> PAGE_SHIFT);
230
-			last_page = reloc->cmdbuf.offset >> PAGE_SHIFT;
231
-
232
-			if (unlikely(!cmdbuf_page_addr)) {
306
+			if (unlikely(!cmdbuf_addr)) {
233307 				pr_err("Could not map cmdbuf for relocation\n");
234308 				return -ENOMEM;
235309 			}
236310 		}
237311 
238
-		target = cmdbuf_page_addr + (reloc->cmdbuf.offset & ~PAGE_MASK);
312
+		target = cmdbuf_addr + reloc->cmdbuf.offset;
239313 patch_reloc:
240314 		*target = reloc_addr;
241315 	}
242316 
243
-	if (cmdbuf_page_addr)
244
-		host1x_bo_kunmap(cmdbuf, last_page, cmdbuf_page_addr);
317
+	if (cmdbuf_addr)
318
+		host1x_bo_munmap(cmdbuf, cmdbuf_addr);
245319 
246320 	return 0;
247321 }
....@@ -569,6 +643,8 @@
569643 
570644 	for (i = 0; i < job->num_unpins; i++) {
571645 		struct host1x_job_unpin_data *unpin = &job->unpins[i];
646
+		struct device *dev = unpin->dev ?: host->dev;
647
+		struct sg_table *sgt = unpin->sgt;
572648 
573649 		if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL) &&
574650 		    unpin->size && host->domain) {
....@@ -578,7 +654,10 @@
578654 				iova_pfn(&host->iova, job->addr_phys[i]));
579655 		}
580656 
581
-		host1x_bo_unpin(unpin->bo, unpin->sgt);
657
+		if (unpin->dev && sgt)
658
+			dma_unmap_sgtable(unpin->dev, sgt, unpin->dir, 0);
659
+
660
+		host1x_bo_unpin(dev, unpin->bo, sgt);
582661 		host1x_bo_put(unpin->bo);
583662 	}
584663