xref: /openbmc/linux/drivers/iommu/arm/arm-smmu/arm-smmu-qcom.c (revision 19b438592238b3b40c3f945bb5f9c4ca971c0c45)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (c) 2019, The Linux Foundation. All rights reserved.
4  */
5 
6 #include <linux/adreno-smmu-priv.h>
7 #include <linux/of_device.h>
8 #include <linux/qcom_scm.h>
9 
10 #include "arm-smmu.h"
11 
12 struct qcom_smmu {
13 	struct arm_smmu_device smmu;
14 	bool bypass_quirk;
15 	u8 bypass_cbndx;
16 	u32 stall_enabled;
17 };
18 
19 static struct qcom_smmu *to_qcom_smmu(struct arm_smmu_device *smmu)
20 {
21 	return container_of(smmu, struct qcom_smmu, smmu);
22 }
23 
24 static void qcom_adreno_smmu_write_sctlr(struct arm_smmu_device *smmu, int idx,
25 		u32 reg)
26 {
27 	struct qcom_smmu *qsmmu = to_qcom_smmu(smmu);
28 
29 	/*
30 	 * On the GPU device we want to process subsequent transactions after a
31 	 * fault to keep the GPU from hanging
32 	 */
33 	reg |= ARM_SMMU_SCTLR_HUPCF;
34 
35 	if (qsmmu->stall_enabled & BIT(idx))
36 		reg |= ARM_SMMU_SCTLR_CFCFG;
37 
38 	arm_smmu_cb_write(smmu, idx, ARM_SMMU_CB_SCTLR, reg);
39 }
40 
41 static void qcom_adreno_smmu_get_fault_info(const void *cookie,
42 		struct adreno_smmu_fault_info *info)
43 {
44 	struct arm_smmu_domain *smmu_domain = (void *)cookie;
45 	struct arm_smmu_cfg *cfg = &smmu_domain->cfg;
46 	struct arm_smmu_device *smmu = smmu_domain->smmu;
47 
48 	info->fsr = arm_smmu_cb_read(smmu, cfg->cbndx, ARM_SMMU_CB_FSR);
49 	info->fsynr0 = arm_smmu_cb_read(smmu, cfg->cbndx, ARM_SMMU_CB_FSYNR0);
50 	info->fsynr1 = arm_smmu_cb_read(smmu, cfg->cbndx, ARM_SMMU_CB_FSYNR1);
51 	info->far = arm_smmu_cb_readq(smmu, cfg->cbndx, ARM_SMMU_CB_FAR);
52 	info->cbfrsynra = arm_smmu_gr1_read(smmu, ARM_SMMU_GR1_CBFRSYNRA(cfg->cbndx));
53 	info->ttbr0 = arm_smmu_cb_read(smmu, cfg->cbndx, ARM_SMMU_CB_TTBR0);
54 	info->contextidr = arm_smmu_cb_read(smmu, cfg->cbndx, ARM_SMMU_CB_CONTEXTIDR);
55 }
56 
57 static void qcom_adreno_smmu_set_stall(const void *cookie, bool enabled)
58 {
59 	struct arm_smmu_domain *smmu_domain = (void *)cookie;
60 	struct arm_smmu_cfg *cfg = &smmu_domain->cfg;
61 	struct qcom_smmu *qsmmu = to_qcom_smmu(smmu_domain->smmu);
62 
63 	if (enabled)
64 		qsmmu->stall_enabled |= BIT(cfg->cbndx);
65 	else
66 		qsmmu->stall_enabled &= ~BIT(cfg->cbndx);
67 }
68 
69 static void qcom_adreno_smmu_resume_translation(const void *cookie, bool terminate)
70 {
71 	struct arm_smmu_domain *smmu_domain = (void *)cookie;
72 	struct arm_smmu_cfg *cfg = &smmu_domain->cfg;
73 	struct arm_smmu_device *smmu = smmu_domain->smmu;
74 	u32 reg = 0;
75 
76 	if (terminate)
77 		reg |= ARM_SMMU_RESUME_TERMINATE;
78 
79 	arm_smmu_cb_write(smmu, cfg->cbndx, ARM_SMMU_CB_RESUME, reg);
80 }
81 
82 #define QCOM_ADRENO_SMMU_GPU_SID 0
83 
84 static bool qcom_adreno_smmu_is_gpu_device(struct device *dev)
85 {
86 	struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
87 	int i;
88 
89 	/*
90 	 * The GPU will always use SID 0 so that is a handy way to uniquely
91 	 * identify it and configure it for per-instance pagetables
92 	 */
93 	for (i = 0; i < fwspec->num_ids; i++) {
94 		u16 sid = FIELD_GET(ARM_SMMU_SMR_ID, fwspec->ids[i]);
95 
96 		if (sid == QCOM_ADRENO_SMMU_GPU_SID)
97 			return true;
98 	}
99 
100 	return false;
101 }
102 
103 static const struct io_pgtable_cfg *qcom_adreno_smmu_get_ttbr1_cfg(
104 		const void *cookie)
105 {
106 	struct arm_smmu_domain *smmu_domain = (void *)cookie;
107 	struct io_pgtable *pgtable =
108 		io_pgtable_ops_to_pgtable(smmu_domain->pgtbl_ops);
109 	return &pgtable->cfg;
110 }
111 
112 /*
113  * Local implementation to configure TTBR0 with the specified pagetable config.
114  * The GPU driver will call this to enable TTBR0 when per-instance pagetables
115  * are active
116  */
117 
118 static int qcom_adreno_smmu_set_ttbr0_cfg(const void *cookie,
119 		const struct io_pgtable_cfg *pgtbl_cfg)
120 {
121 	struct arm_smmu_domain *smmu_domain = (void *)cookie;
122 	struct io_pgtable *pgtable = io_pgtable_ops_to_pgtable(smmu_domain->pgtbl_ops);
123 	struct arm_smmu_cfg *cfg = &smmu_domain->cfg;
124 	struct arm_smmu_cb *cb = &smmu_domain->smmu->cbs[cfg->cbndx];
125 
126 	/* The domain must have split pagetables already enabled */
127 	if (cb->tcr[0] & ARM_SMMU_TCR_EPD1)
128 		return -EINVAL;
129 
130 	/* If the pagetable config is NULL, disable TTBR0 */
131 	if (!pgtbl_cfg) {
132 		/* Do nothing if it is already disabled */
133 		if ((cb->tcr[0] & ARM_SMMU_TCR_EPD0))
134 			return -EINVAL;
135 
136 		/* Set TCR to the original configuration */
137 		cb->tcr[0] = arm_smmu_lpae_tcr(&pgtable->cfg);
138 		cb->ttbr[0] = FIELD_PREP(ARM_SMMU_TTBRn_ASID, cb->cfg->asid);
139 	} else {
140 		u32 tcr = cb->tcr[0];
141 
142 		/* Don't call this again if TTBR0 is already enabled */
143 		if (!(cb->tcr[0] & ARM_SMMU_TCR_EPD0))
144 			return -EINVAL;
145 
146 		tcr |= arm_smmu_lpae_tcr(pgtbl_cfg);
147 		tcr &= ~(ARM_SMMU_TCR_EPD0 | ARM_SMMU_TCR_EPD1);
148 
149 		cb->tcr[0] = tcr;
150 		cb->ttbr[0] = pgtbl_cfg->arm_lpae_s1_cfg.ttbr;
151 		cb->ttbr[0] |= FIELD_PREP(ARM_SMMU_TTBRn_ASID, cb->cfg->asid);
152 	}
153 
154 	arm_smmu_write_context_bank(smmu_domain->smmu, cb->cfg->cbndx);
155 
156 	return 0;
157 }
158 
159 static int qcom_adreno_smmu_alloc_context_bank(struct arm_smmu_domain *smmu_domain,
160 					       struct arm_smmu_device *smmu,
161 					       struct device *dev, int start)
162 {
163 	int count;
164 
165 	/*
166 	 * Assign context bank 0 to the GPU device so the GPU hardware can
167 	 * switch pagetables
168 	 */
169 	if (qcom_adreno_smmu_is_gpu_device(dev)) {
170 		start = 0;
171 		count = 1;
172 	} else {
173 		start = 1;
174 		count = smmu->num_context_banks;
175 	}
176 
177 	return __arm_smmu_alloc_bitmap(smmu->context_map, start, count);
178 }
179 
180 static int qcom_adreno_smmu_init_context(struct arm_smmu_domain *smmu_domain,
181 		struct io_pgtable_cfg *pgtbl_cfg, struct device *dev)
182 {
183 	struct adreno_smmu_priv *priv;
184 
185 	/* Only enable split pagetables for the GPU device (SID 0) */
186 	if (!qcom_adreno_smmu_is_gpu_device(dev))
187 		return 0;
188 
189 	/*
190 	 * All targets that use the qcom,adreno-smmu compatible string *should*
191 	 * be AARCH64 stage 1 but double check because the arm-smmu code assumes
192 	 * that is the case when the TTBR1 quirk is enabled
193 	 */
194 	if ((smmu_domain->stage == ARM_SMMU_DOMAIN_S1) &&
195 	    (smmu_domain->cfg.fmt == ARM_SMMU_CTX_FMT_AARCH64))
196 		pgtbl_cfg->quirks |= IO_PGTABLE_QUIRK_ARM_TTBR1;
197 
198 	/*
199 	 * Initialize private interface with GPU:
200 	 */
201 
202 	priv = dev_get_drvdata(dev);
203 	priv->cookie = smmu_domain;
204 	priv->get_ttbr1_cfg = qcom_adreno_smmu_get_ttbr1_cfg;
205 	priv->set_ttbr0_cfg = qcom_adreno_smmu_set_ttbr0_cfg;
206 	priv->get_fault_info = qcom_adreno_smmu_get_fault_info;
207 	priv->set_stall = qcom_adreno_smmu_set_stall;
208 	priv->resume_translation = qcom_adreno_smmu_resume_translation;
209 
210 	return 0;
211 }
212 
213 static const struct of_device_id qcom_smmu_client_of_match[] __maybe_unused = {
214 	{ .compatible = "qcom,adreno" },
215 	{ .compatible = "qcom,mdp4" },
216 	{ .compatible = "qcom,mdss" },
217 	{ .compatible = "qcom,sc7180-mdss" },
218 	{ .compatible = "qcom,sc7180-mss-pil" },
219 	{ .compatible = "qcom,sc8180x-mdss" },
220 	{ .compatible = "qcom,sdm845-mdss" },
221 	{ .compatible = "qcom,sdm845-mss-pil" },
222 	{ }
223 };
224 
225 static int qcom_smmu_cfg_probe(struct arm_smmu_device *smmu)
226 {
227 	unsigned int last_s2cr = ARM_SMMU_GR0_S2CR(smmu->num_mapping_groups - 1);
228 	struct qcom_smmu *qsmmu = to_qcom_smmu(smmu);
229 	u32 reg;
230 	u32 smr;
231 	int i;
232 
233 	/*
234 	 * With some firmware versions writes to S2CR of type FAULT are
235 	 * ignored, and writing BYPASS will end up written as FAULT in the
236 	 * register. Perform a write to S2CR to detect if this is the case and
237 	 * if so reserve a context bank to emulate bypass streams.
238 	 */
239 	reg = FIELD_PREP(ARM_SMMU_S2CR_TYPE, S2CR_TYPE_BYPASS) |
240 	      FIELD_PREP(ARM_SMMU_S2CR_CBNDX, 0xff) |
241 	      FIELD_PREP(ARM_SMMU_S2CR_PRIVCFG, S2CR_PRIVCFG_DEFAULT);
242 	arm_smmu_gr0_write(smmu, last_s2cr, reg);
243 	reg = arm_smmu_gr0_read(smmu, last_s2cr);
244 	if (FIELD_GET(ARM_SMMU_S2CR_TYPE, reg) != S2CR_TYPE_BYPASS) {
245 		qsmmu->bypass_quirk = true;
246 		qsmmu->bypass_cbndx = smmu->num_context_banks - 1;
247 
248 		set_bit(qsmmu->bypass_cbndx, smmu->context_map);
249 
250 		arm_smmu_cb_write(smmu, qsmmu->bypass_cbndx, ARM_SMMU_CB_SCTLR, 0);
251 
252 		reg = FIELD_PREP(ARM_SMMU_CBAR_TYPE, CBAR_TYPE_S1_TRANS_S2_BYPASS);
253 		arm_smmu_gr1_write(smmu, ARM_SMMU_GR1_CBAR(qsmmu->bypass_cbndx), reg);
254 	}
255 
256 	for (i = 0; i < smmu->num_mapping_groups; i++) {
257 		smr = arm_smmu_gr0_read(smmu, ARM_SMMU_GR0_SMR(i));
258 
259 		if (FIELD_GET(ARM_SMMU_SMR_VALID, smr)) {
260 			/* Ignore valid bit for SMR mask extraction. */
261 			smr &= ~ARM_SMMU_SMR_VALID;
262 			smmu->smrs[i].id = FIELD_GET(ARM_SMMU_SMR_ID, smr);
263 			smmu->smrs[i].mask = FIELD_GET(ARM_SMMU_SMR_MASK, smr);
264 			smmu->smrs[i].valid = true;
265 
266 			smmu->s2crs[i].type = S2CR_TYPE_BYPASS;
267 			smmu->s2crs[i].privcfg = S2CR_PRIVCFG_DEFAULT;
268 			smmu->s2crs[i].cbndx = 0xff;
269 		}
270 	}
271 
272 	return 0;
273 }
274 
275 static void qcom_smmu_write_s2cr(struct arm_smmu_device *smmu, int idx)
276 {
277 	struct arm_smmu_s2cr *s2cr = smmu->s2crs + idx;
278 	struct qcom_smmu *qsmmu = to_qcom_smmu(smmu);
279 	u32 cbndx = s2cr->cbndx;
280 	u32 type = s2cr->type;
281 	u32 reg;
282 
283 	if (qsmmu->bypass_quirk) {
284 		if (type == S2CR_TYPE_BYPASS) {
285 			/*
286 			 * Firmware with quirky S2CR handling will substitute
287 			 * BYPASS writes with FAULT, so point the stream to the
288 			 * reserved context bank and ask for translation on the
289 			 * stream
290 			 */
291 			type = S2CR_TYPE_TRANS;
292 			cbndx = qsmmu->bypass_cbndx;
293 		} else if (type == S2CR_TYPE_FAULT) {
294 			/*
295 			 * Firmware with quirky S2CR handling will ignore FAULT
296 			 * writes, so trick it to write FAULT by asking for a
297 			 * BYPASS.
298 			 */
299 			type = S2CR_TYPE_BYPASS;
300 			cbndx = 0xff;
301 		}
302 	}
303 
304 	reg = FIELD_PREP(ARM_SMMU_S2CR_TYPE, type) |
305 	      FIELD_PREP(ARM_SMMU_S2CR_CBNDX, cbndx) |
306 	      FIELD_PREP(ARM_SMMU_S2CR_PRIVCFG, s2cr->privcfg);
307 	arm_smmu_gr0_write(smmu, ARM_SMMU_GR0_S2CR(idx), reg);
308 }
309 
310 static int qcom_smmu_def_domain_type(struct device *dev)
311 {
312 	const struct of_device_id *match =
313 		of_match_device(qcom_smmu_client_of_match, dev);
314 
315 	return match ? IOMMU_DOMAIN_IDENTITY : 0;
316 }
317 
318 static int qcom_sdm845_smmu500_reset(struct arm_smmu_device *smmu)
319 {
320 	int ret;
321 
322 	/*
323 	 * To address performance degradation in non-real time clients,
324 	 * such as USB and UFS, turn off wait-for-safe on sdm845 based boards,
325 	 * such as MTP and db845, whose firmwares implement secure monitor
326 	 * call handlers to turn on/off the wait-for-safe logic.
327 	 */
328 	ret = qcom_scm_qsmmu500_wait_safe_toggle(0);
329 	if (ret)
330 		dev_warn(smmu->dev, "Failed to turn off SAFE logic\n");
331 
332 	return ret;
333 }
334 
335 static int qcom_smmu500_reset(struct arm_smmu_device *smmu)
336 {
337 	const struct device_node *np = smmu->dev->of_node;
338 
339 	arm_mmu500_reset(smmu);
340 
341 	if (of_device_is_compatible(np, "qcom,sdm845-smmu-500"))
342 		return qcom_sdm845_smmu500_reset(smmu);
343 
344 	return 0;
345 }
346 
347 static const struct arm_smmu_impl qcom_smmu_impl = {
348 	.cfg_probe = qcom_smmu_cfg_probe,
349 	.def_domain_type = qcom_smmu_def_domain_type,
350 	.reset = qcom_smmu500_reset,
351 	.write_s2cr = qcom_smmu_write_s2cr,
352 };
353 
354 static const struct arm_smmu_impl qcom_adreno_smmu_impl = {
355 	.init_context = qcom_adreno_smmu_init_context,
356 	.def_domain_type = qcom_smmu_def_domain_type,
357 	.reset = qcom_smmu500_reset,
358 	.alloc_context_bank = qcom_adreno_smmu_alloc_context_bank,
359 	.write_sctlr = qcom_adreno_smmu_write_sctlr,
360 };
361 
362 static struct arm_smmu_device *qcom_smmu_create(struct arm_smmu_device *smmu,
363 		const struct arm_smmu_impl *impl)
364 {
365 	struct qcom_smmu *qsmmu;
366 
367 	/* Check to make sure qcom_scm has finished probing */
368 	if (!qcom_scm_is_available())
369 		return ERR_PTR(-EPROBE_DEFER);
370 
371 	qsmmu = devm_krealloc(smmu->dev, smmu, sizeof(*qsmmu), GFP_KERNEL);
372 	if (!qsmmu)
373 		return ERR_PTR(-ENOMEM);
374 
375 	qsmmu->smmu.impl = impl;
376 
377 	return &qsmmu->smmu;
378 }
379 
380 static const struct of_device_id __maybe_unused qcom_smmu_impl_of_match[] = {
381 	{ .compatible = "qcom,msm8998-smmu-v2" },
382 	{ .compatible = "qcom,sc7180-smmu-500" },
383 	{ .compatible = "qcom,sc8180x-smmu-500" },
384 	{ .compatible = "qcom,sdm630-smmu-v2" },
385 	{ .compatible = "qcom,sdm845-smmu-500" },
386 	{ .compatible = "qcom,sm8150-smmu-500" },
387 	{ .compatible = "qcom,sm8250-smmu-500" },
388 	{ .compatible = "qcom,sm8350-smmu-500" },
389 	{ }
390 };
391 
392 struct arm_smmu_device *qcom_smmu_impl_init(struct arm_smmu_device *smmu)
393 {
394 	const struct device_node *np = smmu->dev->of_node;
395 
396 	if (of_match_node(qcom_smmu_impl_of_match, np))
397 		return qcom_smmu_create(smmu, &qcom_smmu_impl);
398 
399 	if (of_device_is_compatible(np, "qcom,adreno-smmu"))
400 		return qcom_smmu_create(smmu, &qcom_adreno_smmu_impl);
401 
402 	return smmu;
403 }
404