• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2019 Intel Corporation. All rights rsvd. */
3 #include <linux/init.h>
4 #include <linux/kernel.h>
5 #include <linux/module.h>
6 #include <linux/pci.h>
7 #include <linux/device.h>
8 #include <linux/sched/task.h>
9 #include <linux/io-64-nonatomic-lo-hi.h>
10 #include <linux/cdev.h>
11 #include <linux/fs.h>
12 #include <linux/poll.h>
13 #include <linux/iommu.h>
14 #include <linux/highmem.h>
15 #include <uapi/linux/idxd.h>
16 #include <linux/xarray.h>
17 #include "registers.h"
18 #include "idxd.h"
19 
20 struct idxd_cdev_context {
21 	const char *name;
22 	dev_t devt;
23 	struct ida minor_ida;
24 };
25 
26 /*
27  * Since user file names are global in DSA devices, define their ida's as
28  * global to avoid conflict file names.
29  */
30 static DEFINE_IDA(file_ida);
31 static DEFINE_MUTEX(ida_lock);
32 
33 /*
34  * ictx is an array based off of accelerator types. enum idxd_type
35  * is used as index
36  */
37 static struct idxd_cdev_context ictx[IDXD_TYPE_MAX] = {
38 	{ .name = "dsa" },
39 	{ .name = "iax" }
40 };
41 
42 struct idxd_user_context {
43 	struct idxd_wq *wq;
44 	struct task_struct *task;
45 	unsigned int pasid;
46 	struct mm_struct *mm;
47 	unsigned int flags;
48 	struct iommu_sva *sva;
49 	struct idxd_dev idxd_dev;
50 	u64 counters[COUNTER_MAX];
51 	int id;
52 	pid_t pid;
53 };
54 
55 static void idxd_cdev_evl_drain_pasid(struct idxd_wq *wq, u32 pasid);
56 static void idxd_xa_pasid_remove(struct idxd_user_context *ctx);
57 
dev_to_uctx(struct device * dev)58 static inline struct idxd_user_context *dev_to_uctx(struct device *dev)
59 {
60 	struct idxd_dev *idxd_dev = confdev_to_idxd_dev(dev);
61 
62 	return container_of(idxd_dev, struct idxd_user_context, idxd_dev);
63 }
64 
cr_faults_show(struct device * dev,struct device_attribute * attr,char * buf)65 static ssize_t cr_faults_show(struct device *dev, struct device_attribute *attr, char *buf)
66 {
67 	struct idxd_user_context *ctx = dev_to_uctx(dev);
68 
69 	return sysfs_emit(buf, "%llu\n", ctx->counters[COUNTER_FAULTS]);
70 }
71 static DEVICE_ATTR_RO(cr_faults);
72 
cr_fault_failures_show(struct device * dev,struct device_attribute * attr,char * buf)73 static ssize_t cr_fault_failures_show(struct device *dev,
74 				      struct device_attribute *attr, char *buf)
75 {
76 	struct idxd_user_context *ctx = dev_to_uctx(dev);
77 
78 	return sysfs_emit(buf, "%llu\n", ctx->counters[COUNTER_FAULT_FAILS]);
79 }
80 static DEVICE_ATTR_RO(cr_fault_failures);
81 
pid_show(struct device * dev,struct device_attribute * attr,char * buf)82 static ssize_t pid_show(struct device *dev, struct device_attribute *attr, char *buf)
83 {
84 	struct idxd_user_context *ctx = dev_to_uctx(dev);
85 
86 	return sysfs_emit(buf, "%u\n", ctx->pid);
87 }
88 static DEVICE_ATTR_RO(pid);
89 
90 static struct attribute *cdev_file_attributes[] = {
91 	&dev_attr_cr_faults.attr,
92 	&dev_attr_cr_fault_failures.attr,
93 	&dev_attr_pid.attr,
94 	NULL
95 };
96 
cdev_file_attr_visible(struct kobject * kobj,struct attribute * a,int n)97 static umode_t cdev_file_attr_visible(struct kobject *kobj, struct attribute *a, int n)
98 {
99 	struct device *dev = container_of(kobj, typeof(*dev), kobj);
100 	struct idxd_user_context *ctx = dev_to_uctx(dev);
101 	struct idxd_wq *wq = ctx->wq;
102 
103 	if (!wq_pasid_enabled(wq))
104 		return 0;
105 
106 	return a->mode;
107 }
108 
109 static const struct attribute_group cdev_file_attribute_group = {
110 	.attrs = cdev_file_attributes,
111 	.is_visible = cdev_file_attr_visible,
112 };
113 
114 static const struct attribute_group *cdev_file_attribute_groups[] = {
115 	&cdev_file_attribute_group,
116 	NULL
117 };
118 
idxd_file_dev_release(struct device * dev)119 static void idxd_file_dev_release(struct device *dev)
120 {
121 	struct idxd_user_context *ctx = dev_to_uctx(dev);
122 	struct idxd_wq *wq = ctx->wq;
123 	struct idxd_device *idxd = wq->idxd;
124 	int rc;
125 
126 	mutex_lock(&ida_lock);
127 	ida_free(&file_ida, ctx->id);
128 	mutex_unlock(&ida_lock);
129 
130 	/* Wait for in-flight operations to complete. */
131 	if (wq_shared(wq)) {
132 		idxd_device_drain_pasid(idxd, ctx->pasid);
133 	} else {
134 		if (device_user_pasid_enabled(idxd)) {
135 			/* The wq disable in the disable pasid function will drain the wq */
136 			rc = idxd_wq_disable_pasid(wq);
137 			if (rc < 0)
138 				dev_err(dev, "wq disable pasid failed.\n");
139 		} else {
140 			idxd_wq_drain(wq);
141 		}
142 	}
143 
144 	if (ctx->sva) {
145 		idxd_cdev_evl_drain_pasid(wq, ctx->pasid);
146 		iommu_sva_unbind_device(ctx->sva);
147 		idxd_xa_pasid_remove(ctx);
148 	}
149 	kfree(ctx);
150 	mutex_lock(&wq->wq_lock);
151 	idxd_wq_put(wq);
152 	mutex_unlock(&wq->wq_lock);
153 }
154 
155 static struct device_type idxd_cdev_file_type = {
156 	.name = "idxd_file",
157 	.release = idxd_file_dev_release,
158 	.groups = cdev_file_attribute_groups,
159 };
160 
idxd_cdev_dev_release(struct device * dev)161 static void idxd_cdev_dev_release(struct device *dev)
162 {
163 	struct idxd_cdev *idxd_cdev = dev_to_cdev(dev);
164 	struct idxd_cdev_context *cdev_ctx;
165 	struct idxd_wq *wq = idxd_cdev->wq;
166 
167 	cdev_ctx = &ictx[wq->idxd->data->type];
168 	ida_simple_remove(&cdev_ctx->minor_ida, idxd_cdev->minor);
169 	kfree(idxd_cdev);
170 }
171 
172 static struct device_type idxd_cdev_device_type = {
173 	.name = "idxd_cdev",
174 	.release = idxd_cdev_dev_release,
175 };
176 
inode_idxd_cdev(struct inode * inode)177 static inline struct idxd_cdev *inode_idxd_cdev(struct inode *inode)
178 {
179 	struct cdev *cdev = inode->i_cdev;
180 
181 	return container_of(cdev, struct idxd_cdev, cdev);
182 }
183 
inode_wq(struct inode * inode)184 static inline struct idxd_wq *inode_wq(struct inode *inode)
185 {
186 	struct idxd_cdev *idxd_cdev = inode_idxd_cdev(inode);
187 
188 	return idxd_cdev->wq;
189 }
190 
idxd_xa_pasid_remove(struct idxd_user_context * ctx)191 static void idxd_xa_pasid_remove(struct idxd_user_context *ctx)
192 {
193 	struct idxd_wq *wq = ctx->wq;
194 	void *ptr;
195 
196 	mutex_lock(&wq->uc_lock);
197 	ptr = xa_cmpxchg(&wq->upasid_xa, ctx->pasid, ctx, NULL, GFP_KERNEL);
198 	if (ptr != (void *)ctx)
199 		dev_warn(&wq->idxd->pdev->dev, "xarray cmpxchg failed for pasid %u\n",
200 			 ctx->pasid);
201 	mutex_unlock(&wq->uc_lock);
202 }
203 
idxd_user_counter_increment(struct idxd_wq * wq,u32 pasid,int index)204 void idxd_user_counter_increment(struct idxd_wq *wq, u32 pasid, int index)
205 {
206 	struct idxd_user_context *ctx;
207 
208 	if (index >= COUNTER_MAX)
209 		return;
210 
211 	mutex_lock(&wq->uc_lock);
212 	ctx = xa_load(&wq->upasid_xa, pasid);
213 	if (!ctx) {
214 		mutex_unlock(&wq->uc_lock);
215 		return;
216 	}
217 	ctx->counters[index]++;
218 	mutex_unlock(&wq->uc_lock);
219 }
220 
idxd_cdev_open(struct inode * inode,struct file * filp)221 static int idxd_cdev_open(struct inode *inode, struct file *filp)
222 {
223 	struct idxd_user_context *ctx;
224 	struct idxd_device *idxd;
225 	struct idxd_wq *wq;
226 	struct device *dev, *fdev;
227 	int rc = 0;
228 	struct iommu_sva *sva = NULL;
229 	unsigned int pasid;
230 	struct idxd_cdev *idxd_cdev;
231 
232 	wq = inode_wq(inode);
233 	idxd = wq->idxd;
234 	dev = &idxd->pdev->dev;
235 
236 	dev_dbg(dev, "%s called: %d\n", __func__, idxd_wq_refcount(wq));
237 
238 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
239 	if (!ctx)
240 		return -ENOMEM;
241 
242 	mutex_lock(&wq->wq_lock);
243 
244 	if (idxd_wq_refcount(wq) > 0 && wq_dedicated(wq)) {
245 		rc = -EBUSY;
246 		goto failed;
247 	}
248 
249 	ctx->wq = wq;
250 	filp->private_data = ctx;
251 	ctx->pid = current->pid;
252 
253 	if (device_user_pasid_enabled(idxd)) {
254 		sva = iommu_sva_bind_device(dev, current->mm);
255 		if (IS_ERR(sva)) {
256 			rc = PTR_ERR(sva);
257 			dev_err(dev, "pasid allocation failed: %d\n", rc);
258 			goto failed;
259 		}
260 
261 		pasid = iommu_sva_get_pasid(sva);
262 		if (pasid == IOMMU_PASID_INVALID) {
263 			rc = -EINVAL;
264 			goto failed_get_pasid;
265 		}
266 
267 		ctx->sva = sva;
268 		ctx->pasid = pasid;
269 		ctx->mm = current->mm;
270 
271 		mutex_lock(&wq->uc_lock);
272 		rc = xa_insert(&wq->upasid_xa, pasid, ctx, GFP_KERNEL);
273 		mutex_unlock(&wq->uc_lock);
274 		if (rc < 0)
275 			dev_warn(dev, "PASID entry already exist in xarray.\n");
276 
277 		if (wq_dedicated(wq)) {
278 			rc = idxd_wq_set_pasid(wq, pasid);
279 			if (rc < 0) {
280 				dev_err(dev, "wq set pasid failed: %d\n", rc);
281 				goto failed_set_pasid;
282 			}
283 		}
284 	}
285 
286 	idxd_cdev = wq->idxd_cdev;
287 	mutex_lock(&ida_lock);
288 	ctx->id = ida_alloc(&file_ida, GFP_KERNEL);
289 	mutex_unlock(&ida_lock);
290 	if (ctx->id < 0) {
291 		dev_warn(dev, "ida alloc failure\n");
292 		goto failed_ida;
293 	}
294 	ctx->idxd_dev.type  = IDXD_DEV_CDEV_FILE;
295 	fdev = user_ctx_dev(ctx);
296 	device_initialize(fdev);
297 	fdev->parent = cdev_dev(idxd_cdev);
298 	fdev->bus = &dsa_bus_type;
299 	fdev->type = &idxd_cdev_file_type;
300 
301 	rc = dev_set_name(fdev, "file%d", ctx->id);
302 	if (rc < 0) {
303 		dev_warn(dev, "set name failure\n");
304 		goto failed_dev_name;
305 	}
306 
307 	rc = device_add(fdev);
308 	if (rc < 0) {
309 		dev_warn(dev, "file device add failure\n");
310 		goto failed_dev_add;
311 	}
312 
313 	idxd_wq_get(wq);
314 	mutex_unlock(&wq->wq_lock);
315 	return 0;
316 
317 failed_dev_add:
318 failed_dev_name:
319 	put_device(fdev);
320 failed_ida:
321 failed_set_pasid:
322 	if (device_user_pasid_enabled(idxd))
323 		idxd_xa_pasid_remove(ctx);
324 failed_get_pasid:
325 	if (device_user_pasid_enabled(idxd) && !IS_ERR_OR_NULL(sva))
326 		iommu_sva_unbind_device(sva);
327 failed:
328 	mutex_unlock(&wq->wq_lock);
329 	kfree(ctx);
330 	return rc;
331 }
332 
idxd_cdev_evl_drain_pasid(struct idxd_wq * wq,u32 pasid)333 static void idxd_cdev_evl_drain_pasid(struct idxd_wq *wq, u32 pasid)
334 {
335 	struct idxd_device *idxd = wq->idxd;
336 	struct idxd_evl *evl = idxd->evl;
337 	union evl_status_reg status;
338 	u16 h, t, size;
339 	int ent_size = evl_ent_size(idxd);
340 	struct __evl_entry *entry_head;
341 
342 	if (!evl)
343 		return;
344 
345 	mutex_lock(&evl->lock);
346 	status.bits = ioread64(idxd->reg_base + IDXD_EVLSTATUS_OFFSET);
347 	t = status.tail;
348 	h = status.head;
349 	size = evl->size;
350 
351 	while (h != t) {
352 		entry_head = (struct __evl_entry *)(evl->log + (h * ent_size));
353 		if (entry_head->pasid == pasid && entry_head->wq_idx == wq->id)
354 			set_bit(h, evl->bmap);
355 		h = (h + 1) % size;
356 	}
357 	if (wq->wq)
358 		drain_workqueue(wq->wq);
359 
360 	mutex_unlock(&evl->lock);
361 }
362 
idxd_cdev_release(struct inode * node,struct file * filep)363 static int idxd_cdev_release(struct inode *node, struct file *filep)
364 {
365 	struct idxd_user_context *ctx = filep->private_data;
366 	struct idxd_wq *wq = ctx->wq;
367 	struct idxd_device *idxd = wq->idxd;
368 	struct device *dev = &idxd->pdev->dev;
369 
370 	dev_dbg(dev, "%s called\n", __func__);
371 	filep->private_data = NULL;
372 
373 	device_unregister(user_ctx_dev(ctx));
374 
375 	return 0;
376 }
377 
check_vma(struct idxd_wq * wq,struct vm_area_struct * vma,const char * func)378 static int check_vma(struct idxd_wq *wq, struct vm_area_struct *vma,
379 		     const char *func)
380 {
381 	struct device *dev = &wq->idxd->pdev->dev;
382 
383 	if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
384 		dev_info_ratelimited(dev,
385 				     "%s: %s: mapping too large: %lu\n",
386 				     current->comm, func,
387 				     vma->vm_end - vma->vm_start);
388 		return -EINVAL;
389 	}
390 
391 	return 0;
392 }
393 
idxd_cdev_mmap(struct file * filp,struct vm_area_struct * vma)394 static int idxd_cdev_mmap(struct file *filp, struct vm_area_struct *vma)
395 {
396 	struct idxd_user_context *ctx = filp->private_data;
397 	struct idxd_wq *wq = ctx->wq;
398 	struct idxd_device *idxd = wq->idxd;
399 	struct pci_dev *pdev = idxd->pdev;
400 	phys_addr_t base = pci_resource_start(pdev, IDXD_WQ_BAR);
401 	unsigned long pfn;
402 	int rc;
403 
404 	dev_dbg(&pdev->dev, "%s called\n", __func__);
405 
406 	/*
407 	 * Due to an erratum in some of the devices supported by the driver,
408 	 * direct user submission to the device can be unsafe.
409 	 * (See the INTEL-SA-01084 security advisory)
410 	 *
411 	 * For the devices that exhibit this behavior, require that the user
412 	 * has CAP_SYS_RAWIO capabilities.
413 	 */
414 	if (!idxd->user_submission_safe && !capable(CAP_SYS_RAWIO))
415 		return -EPERM;
416 
417 	if (current->mm != ctx->mm)
418 		return -EPERM;
419 
420 	rc = check_vma(wq, vma, __func__);
421 	if (rc < 0)
422 		return rc;
423 
424 	vm_flags_set(vma, VM_DONTCOPY);
425 	pfn = (base + idxd_get_wq_portal_full_offset(wq->id,
426 				IDXD_PORTAL_LIMITED)) >> PAGE_SHIFT;
427 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
428 	vma->vm_private_data = ctx;
429 
430 	return io_remap_pfn_range(vma, vma->vm_start, pfn, PAGE_SIZE,
431 			vma->vm_page_prot);
432 }
433 
idxd_submit_user_descriptor(struct idxd_user_context * ctx,struct dsa_hw_desc __user * udesc)434 static int idxd_submit_user_descriptor(struct idxd_user_context *ctx,
435 				       struct dsa_hw_desc __user *udesc)
436 {
437 	struct idxd_wq *wq = ctx->wq;
438 	struct idxd_dev *idxd_dev = &wq->idxd->idxd_dev;
439 	const uint64_t comp_addr_align = is_dsa_dev(idxd_dev) ? 0x20 : 0x40;
440 	void __iomem *portal = idxd_wq_portal_addr(wq);
441 	struct dsa_hw_desc descriptor __aligned(64);
442 	int rc;
443 
444 	rc = copy_from_user(&descriptor, udesc, sizeof(descriptor));
445 	if (rc)
446 		return -EFAULT;
447 
448 	/*
449 	 * DSA devices are capable of indirect ("batch") command submission.
450 	 * On devices where direct user submissions are not safe, we cannot
451 	 * allow this since there is no good way for us to verify these
452 	 * indirect commands.
453 	 */
454 	if (is_dsa_dev(idxd_dev) && descriptor.opcode == DSA_OPCODE_BATCH &&
455 		!wq->idxd->user_submission_safe)
456 		return -EINVAL;
457 	/*
458 	 * As per the programming specification, the completion address must be
459 	 * aligned to 32 or 64 bytes. If this is violated the hardware
460 	 * engine can get very confused (security issue).
461 	 */
462 	if (!IS_ALIGNED(descriptor.completion_addr, comp_addr_align))
463 		return -EINVAL;
464 
465 	if (wq_dedicated(wq))
466 		iosubmit_cmds512(portal, &descriptor, 1);
467 	else {
468 		descriptor.priv = 0;
469 		descriptor.pasid = ctx->pasid;
470 		rc = idxd_enqcmds(wq, portal, &descriptor);
471 		if (rc < 0)
472 			return rc;
473 	}
474 
475 	return 0;
476 }
477 
idxd_cdev_write(struct file * filp,const char __user * buf,size_t len,loff_t * unused)478 static ssize_t idxd_cdev_write(struct file *filp, const char __user *buf, size_t len,
479 			       loff_t *unused)
480 {
481 	struct dsa_hw_desc __user *udesc = (struct dsa_hw_desc __user *)buf;
482 	struct idxd_user_context *ctx = filp->private_data;
483 	ssize_t written = 0;
484 	int i;
485 
486 	if (current->mm != ctx->mm)
487 		return -EPERM;
488 
489 	for (i = 0; i < len/sizeof(struct dsa_hw_desc); i++) {
490 		int rc = idxd_submit_user_descriptor(ctx, udesc + i);
491 
492 		if (rc)
493 			return written ? written : rc;
494 
495 		written += sizeof(struct dsa_hw_desc);
496 	}
497 
498 	return written;
499 }
500 
idxd_cdev_poll(struct file * filp,struct poll_table_struct * wait)501 static __poll_t idxd_cdev_poll(struct file *filp,
502 			       struct poll_table_struct *wait)
503 {
504 	struct idxd_user_context *ctx = filp->private_data;
505 	struct idxd_wq *wq = ctx->wq;
506 	struct idxd_device *idxd = wq->idxd;
507 	__poll_t out = 0;
508 
509 	if (current->mm != ctx->mm)
510 		return POLLNVAL;
511 
512 	poll_wait(filp, &wq->err_queue, wait);
513 	spin_lock(&idxd->dev_lock);
514 	if (idxd->sw_err.valid)
515 		out = EPOLLIN | EPOLLRDNORM;
516 	spin_unlock(&idxd->dev_lock);
517 
518 	return out;
519 }
520 
521 static const struct file_operations idxd_cdev_fops = {
522 	.owner = THIS_MODULE,
523 	.open = idxd_cdev_open,
524 	.release = idxd_cdev_release,
525 	.mmap = idxd_cdev_mmap,
526 	.write = idxd_cdev_write,
527 	.poll = idxd_cdev_poll,
528 };
529 
idxd_cdev_get_major(struct idxd_device * idxd)530 int idxd_cdev_get_major(struct idxd_device *idxd)
531 {
532 	return MAJOR(ictx[idxd->data->type].devt);
533 }
534 
idxd_wq_add_cdev(struct idxd_wq * wq)535 int idxd_wq_add_cdev(struct idxd_wq *wq)
536 {
537 	struct idxd_device *idxd = wq->idxd;
538 	struct idxd_cdev *idxd_cdev;
539 	struct cdev *cdev;
540 	struct device *dev;
541 	struct idxd_cdev_context *cdev_ctx;
542 	int rc, minor;
543 
544 	idxd_cdev = kzalloc(sizeof(*idxd_cdev), GFP_KERNEL);
545 	if (!idxd_cdev)
546 		return -ENOMEM;
547 
548 	idxd_cdev->idxd_dev.type = IDXD_DEV_CDEV;
549 	idxd_cdev->wq = wq;
550 	cdev = &idxd_cdev->cdev;
551 	dev = cdev_dev(idxd_cdev);
552 	cdev_ctx = &ictx[wq->idxd->data->type];
553 	minor = ida_simple_get(&cdev_ctx->minor_ida, 0, MINORMASK, GFP_KERNEL);
554 	if (minor < 0) {
555 		kfree(idxd_cdev);
556 		return minor;
557 	}
558 	idxd_cdev->minor = minor;
559 
560 	device_initialize(dev);
561 	dev->parent = wq_confdev(wq);
562 	dev->bus = &dsa_bus_type;
563 	dev->type = &idxd_cdev_device_type;
564 	dev->devt = MKDEV(MAJOR(cdev_ctx->devt), minor);
565 
566 	rc = dev_set_name(dev, "%s/wq%u.%u", idxd->data->name_prefix, idxd->id, wq->id);
567 	if (rc < 0)
568 		goto err;
569 
570 	wq->idxd_cdev = idxd_cdev;
571 	cdev_init(cdev, &idxd_cdev_fops);
572 	rc = cdev_device_add(cdev, dev);
573 	if (rc) {
574 		dev_dbg(&wq->idxd->pdev->dev, "cdev_add failed: %d\n", rc);
575 		goto err;
576 	}
577 
578 	return 0;
579 
580  err:
581 	put_device(dev);
582 	wq->idxd_cdev = NULL;
583 	return rc;
584 }
585 
idxd_wq_del_cdev(struct idxd_wq * wq)586 void idxd_wq_del_cdev(struct idxd_wq *wq)
587 {
588 	struct idxd_cdev *idxd_cdev;
589 
590 	idxd_cdev = wq->idxd_cdev;
591 	wq->idxd_cdev = NULL;
592 	cdev_device_del(&idxd_cdev->cdev, cdev_dev(idxd_cdev));
593 	put_device(cdev_dev(idxd_cdev));
594 }
595 
idxd_user_drv_probe(struct idxd_dev * idxd_dev)596 static int idxd_user_drv_probe(struct idxd_dev *idxd_dev)
597 {
598 	struct device *dev = &idxd_dev->conf_dev;
599 	struct idxd_wq *wq = idxd_dev_to_wq(idxd_dev);
600 	struct idxd_device *idxd = wq->idxd;
601 	int rc;
602 
603 	if (idxd->state != IDXD_DEV_ENABLED)
604 		return -ENXIO;
605 
606 	/*
607 	 * User type WQ is enabled only when SVA is enabled for two reasons:
608 	 *   - If no IOMMU or IOMMU Passthrough without SVA, userspace
609 	 *     can directly access physical address through the WQ.
610 	 *   - The IDXD cdev driver does not provide any ways to pin
611 	 *     user pages and translate the address from user VA to IOVA or
612 	 *     PA without IOMMU SVA. Therefore the application has no way
613 	 *     to instruct the device to perform DMA function. This makes
614 	 *     the cdev not usable for normal application usage.
615 	 */
616 	if (!device_user_pasid_enabled(idxd)) {
617 		idxd->cmd_status = IDXD_SCMD_WQ_USER_NO_IOMMU;
618 		dev_dbg(&idxd->pdev->dev,
619 			"User type WQ cannot be enabled without SVA.\n");
620 
621 		return -EOPNOTSUPP;
622 	}
623 
624 	mutex_lock(&wq->wq_lock);
625 
626 	if (!idxd_wq_driver_name_match(wq, dev)) {
627 		idxd->cmd_status = IDXD_SCMD_WQ_NO_DRV_NAME;
628 		rc = -ENODEV;
629 		goto wq_err;
630 	}
631 
632 	wq->wq = create_workqueue(dev_name(wq_confdev(wq)));
633 	if (!wq->wq) {
634 		rc = -ENOMEM;
635 		goto wq_err;
636 	}
637 
638 	wq->type = IDXD_WQT_USER;
639 	rc = drv_enable_wq(wq);
640 	if (rc < 0)
641 		goto err;
642 
643 	rc = idxd_wq_add_cdev(wq);
644 	if (rc < 0) {
645 		idxd->cmd_status = IDXD_SCMD_CDEV_ERR;
646 		goto err_cdev;
647 	}
648 
649 	idxd->cmd_status = 0;
650 	mutex_unlock(&wq->wq_lock);
651 	return 0;
652 
653 err_cdev:
654 	drv_disable_wq(wq);
655 err:
656 	destroy_workqueue(wq->wq);
657 	wq->type = IDXD_WQT_NONE;
658 wq_err:
659 	mutex_unlock(&wq->wq_lock);
660 	return rc;
661 }
662 
idxd_user_drv_remove(struct idxd_dev * idxd_dev)663 static void idxd_user_drv_remove(struct idxd_dev *idxd_dev)
664 {
665 	struct idxd_wq *wq = idxd_dev_to_wq(idxd_dev);
666 
667 	mutex_lock(&wq->wq_lock);
668 	idxd_wq_del_cdev(wq);
669 	drv_disable_wq(wq);
670 	wq->type = IDXD_WQT_NONE;
671 	destroy_workqueue(wq->wq);
672 	wq->wq = NULL;
673 	mutex_unlock(&wq->wq_lock);
674 }
675 
676 static enum idxd_dev_type dev_types[] = {
677 	IDXD_DEV_WQ,
678 	IDXD_DEV_NONE,
679 };
680 
681 struct idxd_device_driver idxd_user_drv = {
682 	.probe = idxd_user_drv_probe,
683 	.remove = idxd_user_drv_remove,
684 	.name = "user",
685 	.type = dev_types,
686 };
687 EXPORT_SYMBOL_GPL(idxd_user_drv);
688 
idxd_cdev_register(void)689 int idxd_cdev_register(void)
690 {
691 	int rc, i;
692 
693 	for (i = 0; i < IDXD_TYPE_MAX; i++) {
694 		ida_init(&ictx[i].minor_ida);
695 		rc = alloc_chrdev_region(&ictx[i].devt, 0, MINORMASK,
696 					 ictx[i].name);
697 		if (rc)
698 			goto err_free_chrdev_region;
699 	}
700 
701 	return 0;
702 
703 err_free_chrdev_region:
704 	for (i--; i >= 0; i--)
705 		unregister_chrdev_region(ictx[i].devt, MINORMASK);
706 
707 	return rc;
708 }
709 
idxd_cdev_remove(void)710 void idxd_cdev_remove(void)
711 {
712 	int i;
713 
714 	for (i = 0; i < IDXD_TYPE_MAX; i++) {
715 		unregister_chrdev_region(ictx[i].devt, MINORMASK);
716 		ida_destroy(&ictx[i].minor_ida);
717 	}
718 }
719 
720 /**
721  * idxd_copy_cr - copy completion record to user address space found by wq and
722  *		  PASID
723  * @wq:		work queue
724  * @pasid:	PASID
725  * @addr:	user fault address to write
726  * @cr:		completion record
727  * @len:	number of bytes to copy
728  *
729  * This is called by a work that handles completion record fault.
730  *
731  * Return: number of bytes copied.
732  */
idxd_copy_cr(struct idxd_wq * wq,ioasid_t pasid,unsigned long addr,void * cr,int len)733 int idxd_copy_cr(struct idxd_wq *wq, ioasid_t pasid, unsigned long addr,
734 		 void *cr, int len)
735 {
736 	struct device *dev = &wq->idxd->pdev->dev;
737 	int left = len, status_size = 1;
738 	struct idxd_user_context *ctx;
739 	struct mm_struct *mm;
740 
741 	mutex_lock(&wq->uc_lock);
742 
743 	ctx = xa_load(&wq->upasid_xa, pasid);
744 	if (!ctx) {
745 		dev_warn(dev, "No user context\n");
746 		goto out;
747 	}
748 
749 	mm = ctx->mm;
750 	/*
751 	 * The completion record fault handling work is running in kernel
752 	 * thread context. It temporarily switches to the mm to copy cr
753 	 * to addr in the mm.
754 	 */
755 	kthread_use_mm(mm);
756 	left = copy_to_user((void __user *)addr + status_size, cr + status_size,
757 			    len - status_size);
758 	/*
759 	 * Copy status only after the rest of completion record is copied
760 	 * successfully so that the user gets the complete completion record
761 	 * when a non-zero status is polled.
762 	 */
763 	if (!left) {
764 		u8 status;
765 
766 		/*
767 		 * Ensure that the completion record's status field is written
768 		 * after the rest of the completion record has been written.
769 		 * This ensures that the user receives the correct completion
770 		 * record information once polling for a non-zero status.
771 		 */
772 		wmb();
773 		status = *(u8 *)cr;
774 		if (put_user(status, (u8 __user *)addr))
775 			left += status_size;
776 	} else {
777 		left += status_size;
778 	}
779 	kthread_unuse_mm(mm);
780 
781 out:
782 	mutex_unlock(&wq->uc_lock);
783 
784 	return len - left;
785 }
786