xref: /linux/drivers/iommu/intel/prq.c (revision 7f4f3b14e8079ecde096bd734af10e30d40c27b7)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2015 Intel Corporation
4  *
5  * Originally split from drivers/iommu/intel/svm.c
6  */
7 
8 #include <linux/pci.h>
9 #include <linux/pci-ats.h>
10 
11 #include "iommu.h"
12 #include "pasid.h"
13 #include "../iommu-pages.h"
14 #include "trace.h"
15 
16 /* Page request queue descriptor */
17 struct page_req_dsc {
18 	union {
19 		struct {
20 			u64 type:8;
21 			u64 pasid_present:1;
22 			u64 rsvd:7;
23 			u64 rid:16;
24 			u64 pasid:20;
25 			u64 exe_req:1;
26 			u64 pm_req:1;
27 			u64 rsvd2:10;
28 		};
29 		u64 qw_0;
30 	};
31 	union {
32 		struct {
33 			u64 rd_req:1;
34 			u64 wr_req:1;
35 			u64 lpig:1;
36 			u64 prg_index:9;
37 			u64 addr:52;
38 		};
39 		u64 qw_1;
40 	};
41 	u64 qw_2;
42 	u64 qw_3;
43 };
44 
45 /**
46  * intel_iommu_drain_pasid_prq - Drain page requests and responses for a pasid
47  * @dev: target device
48  * @pasid: pasid for draining
49  *
50  * Drain all pending page requests and responses related to @pasid in both
51  * software and hardware. This is supposed to be called after the device
52  * driver has stopped DMA, the pasid entry has been cleared, and both IOTLB
53  * and DevTLB have been invalidated.
54  *
55  * It waits until all pending page requests for @pasid in the page fault
56  * queue are completed by the prq handling thread. Then follow the steps
57  * described in VT-d spec CH7.10 to drain all page requests and page
58  * responses pending in the hardware.
59  */
60 void intel_iommu_drain_pasid_prq(struct device *dev, u32 pasid)
61 {
62 	struct device_domain_info *info;
63 	struct dmar_domain *domain;
64 	struct intel_iommu *iommu;
65 	struct qi_desc desc[3];
66 	int head, tail;
67 	u16 sid, did;
68 
69 	info = dev_iommu_priv_get(dev);
70 	if (!info->pri_enabled)
71 		return;
72 
73 	iommu = info->iommu;
74 	domain = info->domain;
75 	sid = PCI_DEVID(info->bus, info->devfn);
76 	did = domain ? domain_id_iommu(domain, iommu) : FLPT_DEFAULT_DID;
77 
78 	/*
79 	 * Check and wait until all pending page requests in the queue are
80 	 * handled by the prq handling thread.
81 	 */
82 prq_retry:
83 	reinit_completion(&iommu->prq_complete);
84 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
85 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
86 	while (head != tail) {
87 		struct page_req_dsc *req;
88 
89 		req = &iommu->prq[head / sizeof(*req)];
90 		if (!req->pasid_present || req->pasid != pasid) {
91 			head = (head + sizeof(*req)) & PRQ_RING_MASK;
92 			continue;
93 		}
94 
95 		wait_for_completion(&iommu->prq_complete);
96 		goto prq_retry;
97 	}
98 
99 	iopf_queue_flush_dev(dev);
100 
101 	/*
102 	 * Perform steps described in VT-d spec CH7.10 to drain page
103 	 * requests and responses in hardware.
104 	 */
105 	memset(desc, 0, sizeof(desc));
106 	desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
107 			QI_IWD_FENCE |
108 			QI_IWD_TYPE;
109 	if (pasid == IOMMU_NO_PASID) {
110 		qi_desc_iotlb(iommu, did, 0, 0, DMA_TLB_DSI_FLUSH, &desc[1]);
111 		qi_desc_dev_iotlb(sid, info->pfsid, info->ats_qdep, 0,
112 				  MAX_AGAW_PFN_WIDTH, &desc[2]);
113 	} else {
114 		qi_desc_piotlb(did, pasid, 0, -1, 0, &desc[1]);
115 		qi_desc_dev_iotlb_pasid(sid, info->pfsid, pasid, info->ats_qdep,
116 					0, MAX_AGAW_PFN_WIDTH, &desc[2]);
117 	}
118 qi_retry:
119 	reinit_completion(&iommu->prq_complete);
120 	qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
121 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
122 		wait_for_completion(&iommu->prq_complete);
123 		goto qi_retry;
124 	}
125 }
126 
127 static bool is_canonical_address(u64 addr)
128 {
129 	int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
130 	long saddr = (long)addr;
131 
132 	return (((saddr << shift) >> shift) == saddr);
133 }
134 
135 static void handle_bad_prq_event(struct intel_iommu *iommu,
136 				 struct page_req_dsc *req, int result)
137 {
138 	struct qi_desc desc = { };
139 
140 	pr_err("%s: Invalid page request: %08llx %08llx\n",
141 	       iommu->name, ((unsigned long long *)req)[0],
142 	       ((unsigned long long *)req)[1]);
143 
144 	if (!req->lpig)
145 		return;
146 
147 	desc.qw0 = QI_PGRP_PASID(req->pasid) |
148 			QI_PGRP_DID(req->rid) |
149 			QI_PGRP_PASID_P(req->pasid_present) |
150 			QI_PGRP_RESP_CODE(result) |
151 			QI_PGRP_RESP_TYPE;
152 	desc.qw1 = QI_PGRP_IDX(req->prg_index) |
153 			QI_PGRP_LPIG(req->lpig);
154 
155 	qi_submit_sync(iommu, &desc, 1, 0);
156 }
157 
158 static int prq_to_iommu_prot(struct page_req_dsc *req)
159 {
160 	int prot = 0;
161 
162 	if (req->rd_req)
163 		prot |= IOMMU_FAULT_PERM_READ;
164 	if (req->wr_req)
165 		prot |= IOMMU_FAULT_PERM_WRITE;
166 	if (req->exe_req)
167 		prot |= IOMMU_FAULT_PERM_EXEC;
168 	if (req->pm_req)
169 		prot |= IOMMU_FAULT_PERM_PRIV;
170 
171 	return prot;
172 }
173 
174 static void intel_prq_report(struct intel_iommu *iommu, struct device *dev,
175 			     struct page_req_dsc *desc)
176 {
177 	struct iopf_fault event = { };
178 
179 	/* Fill in event data for device specific processing */
180 	event.fault.type = IOMMU_FAULT_PAGE_REQ;
181 	event.fault.prm.addr = (u64)desc->addr << VTD_PAGE_SHIFT;
182 	event.fault.prm.pasid = desc->pasid;
183 	event.fault.prm.grpid = desc->prg_index;
184 	event.fault.prm.perm = prq_to_iommu_prot(desc);
185 
186 	if (desc->lpig)
187 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
188 	if (desc->pasid_present) {
189 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
190 		event.fault.prm.flags |= IOMMU_FAULT_PAGE_RESPONSE_NEEDS_PASID;
191 	}
192 
193 	iommu_report_device_fault(dev, &event);
194 }
195 
196 static irqreturn_t prq_event_thread(int irq, void *d)
197 {
198 	struct intel_iommu *iommu = d;
199 	struct page_req_dsc *req;
200 	int head, tail, handled;
201 	struct device *dev;
202 	u64 address;
203 
204 	/*
205 	 * Clear PPR bit before reading head/tail registers, to ensure that
206 	 * we get a new interrupt if needed.
207 	 */
208 	writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
209 
210 	tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
211 	head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
212 	handled = (head != tail);
213 	while (head != tail) {
214 		req = &iommu->prq[head / sizeof(*req)];
215 		address = (u64)req->addr << VTD_PAGE_SHIFT;
216 
217 		if (unlikely(!is_canonical_address(address))) {
218 			pr_err("IOMMU: %s: Address is not canonical\n",
219 			       iommu->name);
220 bad_req:
221 			handle_bad_prq_event(iommu, req, QI_RESP_INVALID);
222 			goto prq_advance;
223 		}
224 
225 		if (unlikely(req->pm_req && (req->rd_req | req->wr_req))) {
226 			pr_err("IOMMU: %s: Page request in Privilege Mode\n",
227 			       iommu->name);
228 			goto bad_req;
229 		}
230 
231 		if (unlikely(req->exe_req && req->rd_req)) {
232 			pr_err("IOMMU: %s: Execution request not supported\n",
233 			       iommu->name);
234 			goto bad_req;
235 		}
236 
237 		/* Drop Stop Marker message. No need for a response. */
238 		if (unlikely(req->lpig && !req->rd_req && !req->wr_req))
239 			goto prq_advance;
240 
241 		/*
242 		 * If prq is to be handled outside iommu driver via receiver of
243 		 * the fault notifiers, we skip the page response here.
244 		 */
245 		mutex_lock(&iommu->iopf_lock);
246 		dev = device_rbtree_find(iommu, req->rid);
247 		if (!dev) {
248 			mutex_unlock(&iommu->iopf_lock);
249 			goto bad_req;
250 		}
251 
252 		intel_prq_report(iommu, dev, req);
253 		trace_prq_report(iommu, dev, req->qw_0, req->qw_1,
254 				 req->qw_2, req->qw_3,
255 				 iommu->prq_seq_number++);
256 		mutex_unlock(&iommu->iopf_lock);
257 prq_advance:
258 		head = (head + sizeof(*req)) & PRQ_RING_MASK;
259 	}
260 
261 	dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
262 
263 	/*
264 	 * Clear the page request overflow bit and wake up all threads that
265 	 * are waiting for the completion of this handling.
266 	 */
267 	if (readl(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PRO) {
268 		pr_info_ratelimited("IOMMU: %s: PRQ overflow detected\n",
269 				    iommu->name);
270 		head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
271 		tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
272 		if (head == tail) {
273 			iopf_queue_discard_partial(iommu->iopf_queue);
274 			writel(DMA_PRS_PRO, iommu->reg + DMAR_PRS_REG);
275 			pr_info_ratelimited("IOMMU: %s: PRQ overflow cleared",
276 					    iommu->name);
277 		}
278 	}
279 
280 	if (!completion_done(&iommu->prq_complete))
281 		complete(&iommu->prq_complete);
282 
283 	return IRQ_RETVAL(handled);
284 }
285 
286 int intel_iommu_enable_prq(struct intel_iommu *iommu)
287 {
288 	struct iopf_queue *iopfq;
289 	int irq, ret;
290 
291 	iommu->prq = iommu_alloc_pages_node(iommu->node, GFP_KERNEL, PRQ_ORDER);
292 	if (!iommu->prq) {
293 		pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
294 			iommu->name);
295 		return -ENOMEM;
296 	}
297 
298 	irq = dmar_alloc_hwirq(IOMMU_IRQ_ID_OFFSET_PRQ + iommu->seq_id, iommu->node, iommu);
299 	if (irq <= 0) {
300 		pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
301 		       iommu->name);
302 		ret = -EINVAL;
303 		goto free_prq;
304 	}
305 	iommu->pr_irq = irq;
306 
307 	snprintf(iommu->iopfq_name, sizeof(iommu->iopfq_name),
308 		 "dmar%d-iopfq", iommu->seq_id);
309 	iopfq = iopf_queue_alloc(iommu->iopfq_name);
310 	if (!iopfq) {
311 		pr_err("IOMMU: %s: Failed to allocate iopf queue\n", iommu->name);
312 		ret = -ENOMEM;
313 		goto free_hwirq;
314 	}
315 	iommu->iopf_queue = iopfq;
316 
317 	snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
318 
319 	ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
320 				   iommu->prq_name, iommu);
321 	if (ret) {
322 		pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
323 		       iommu->name);
324 		goto free_iopfq;
325 	}
326 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
327 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
328 	dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
329 
330 	init_completion(&iommu->prq_complete);
331 
332 	return 0;
333 
334 free_iopfq:
335 	iopf_queue_free(iommu->iopf_queue);
336 	iommu->iopf_queue = NULL;
337 free_hwirq:
338 	dmar_free_hwirq(irq);
339 	iommu->pr_irq = 0;
340 free_prq:
341 	iommu_free_pages(iommu->prq, PRQ_ORDER);
342 	iommu->prq = NULL;
343 
344 	return ret;
345 }
346 
347 int intel_iommu_finish_prq(struct intel_iommu *iommu)
348 {
349 	dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
350 	dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
351 	dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
352 
353 	if (iommu->pr_irq) {
354 		free_irq(iommu->pr_irq, iommu);
355 		dmar_free_hwirq(iommu->pr_irq);
356 		iommu->pr_irq = 0;
357 	}
358 
359 	if (iommu->iopf_queue) {
360 		iopf_queue_free(iommu->iopf_queue);
361 		iommu->iopf_queue = NULL;
362 	}
363 
364 	iommu_free_pages(iommu->prq, PRQ_ORDER);
365 	iommu->prq = NULL;
366 
367 	return 0;
368 }
369 
370 void intel_iommu_page_response(struct device *dev, struct iopf_fault *evt,
371 			       struct iommu_page_response *msg)
372 {
373 	struct device_domain_info *info = dev_iommu_priv_get(dev);
374 	struct intel_iommu *iommu = info->iommu;
375 	u8 bus = info->bus, devfn = info->devfn;
376 	struct iommu_fault_page_request *prm;
377 	struct qi_desc desc;
378 	bool pasid_present;
379 	bool last_page;
380 	u16 sid;
381 
382 	prm = &evt->fault.prm;
383 	sid = PCI_DEVID(bus, devfn);
384 	pasid_present = prm->flags & IOMMU_FAULT_PAGE_REQUEST_PASID_VALID;
385 	last_page = prm->flags & IOMMU_FAULT_PAGE_REQUEST_LAST_PAGE;
386 
387 	desc.qw0 = QI_PGRP_PASID(prm->pasid) | QI_PGRP_DID(sid) |
388 			QI_PGRP_PASID_P(pasid_present) |
389 			QI_PGRP_RESP_CODE(msg->code) |
390 			QI_PGRP_RESP_TYPE;
391 	desc.qw1 = QI_PGRP_IDX(prm->grpid) | QI_PGRP_LPIG(last_page);
392 	desc.qw2 = 0;
393 	desc.qw3 = 0;
394 
395 	qi_submit_sync(iommu, &desc, 1, 0);
396 }
397