xref: /linux/drivers/dma/idxd/cdev.c (revision a516c618a627e30b5613fadd264d4b4498254aeb)
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 
32 /*
33  * ictx is an array based off of accelerator types. enum idxd_type
34  * is used as index
35  */
36 static struct idxd_cdev_context ictx[IDXD_TYPE_MAX] = {
37 	{ .name = "dsa" },
38 	{ .name = "iax" }
39 };
40 
41 struct idxd_user_context {
42 	struct idxd_wq *wq;
43 	struct task_struct *task;
44 	unsigned int pasid;
45 	struct mm_struct *mm;
46 	unsigned int flags;
47 	struct iommu_sva *sva;
48 	struct idxd_dev idxd_dev;
49 	u64 counters[COUNTER_MAX];
50 	int id;
51 	pid_t pid;
52 };
53 
54 static void idxd_cdev_evl_drain_pasid(struct idxd_wq *wq, u32 pasid);
55 static void idxd_xa_pasid_remove(struct idxd_user_context *ctx);
56 
dev_to_uctx(struct device * dev)57 static inline struct idxd_user_context *dev_to_uctx(struct device *dev)
58 {
59 	struct idxd_dev *idxd_dev = confdev_to_idxd_dev(dev);
60 
61 	return container_of(idxd_dev, struct idxd_user_context, idxd_dev);
62 }
63 
cr_faults_show(struct device * dev,struct device_attribute * attr,char * buf)64 static ssize_t cr_faults_show(struct device *dev, struct device_attribute *attr, char *buf)
65 {
66 	struct idxd_user_context *ctx = dev_to_uctx(dev);
67 
68 	return sysfs_emit(buf, "%llu\n", ctx->counters[COUNTER_FAULTS]);
69 }
70 static DEVICE_ATTR_RO(cr_faults);
71 
cr_fault_failures_show(struct device * dev,struct device_attribute * attr,char * buf)72 static ssize_t cr_fault_failures_show(struct device *dev,
73 				      struct device_attribute *attr, char *buf)
74 {
75 	struct idxd_user_context *ctx = dev_to_uctx(dev);
76 
77 	return sysfs_emit(buf, "%llu\n", ctx->counters[COUNTER_FAULT_FAILS]);
78 }
79 static DEVICE_ATTR_RO(cr_fault_failures);
80 
pid_show(struct device * dev,struct device_attribute * attr,char * buf)81 static ssize_t pid_show(struct device *dev, struct device_attribute *attr, char *buf)
82 {
83 	struct idxd_user_context *ctx = dev_to_uctx(dev);
84 
85 	return sysfs_emit(buf, "%u\n", ctx->pid);
86 }
87 static DEVICE_ATTR_RO(pid);
88 
89 static struct attribute *cdev_file_attributes[] = {
90 	&dev_attr_cr_faults.attr,
91 	&dev_attr_cr_fault_failures.attr,
92 	&dev_attr_pid.attr,
93 	NULL
94 };
95 
cdev_file_attr_visible(struct kobject * kobj,struct attribute * a,int n)96 static umode_t cdev_file_attr_visible(struct kobject *kobj, struct attribute *a, int n)
97 {
98 	struct device *dev = container_of(kobj, typeof(*dev), kobj);
99 	struct idxd_user_context *ctx = dev_to_uctx(dev);
100 	struct idxd_wq *wq = ctx->wq;
101 
102 	if (!wq_pasid_enabled(wq))
103 		return 0;
104 
105 	return a->mode;
106 }
107 
108 static const struct attribute_group cdev_file_attribute_group = {
109 	.attrs = cdev_file_attributes,
110 	.is_visible = cdev_file_attr_visible,
111 };
112 
113 static const struct attribute_group *cdev_file_attribute_groups[] = {
114 	&cdev_file_attribute_group,
115 	NULL
116 };
117 
idxd_file_dev_release(struct device * dev)118 static void idxd_file_dev_release(struct device *dev)
119 {
120 	struct idxd_user_context *ctx = dev_to_uctx(dev);
121 	struct idxd_wq *wq = ctx->wq;
122 	struct idxd_device *idxd = wq->idxd;
123 	int rc;
124 
125 	ida_free(&file_ida, ctx->id);
126 
127 	/* Wait for in-flight operations to complete. */
128 	if (wq_shared(wq)) {
129 		idxd_device_drain_pasid(idxd, ctx->pasid);
130 	} else {
131 		if (device_user_pasid_enabled(idxd)) {
132 			/* The wq disable in the disable pasid function will drain the wq */
133 			rc = idxd_wq_disable_pasid(wq);
134 			if (rc < 0)
135 				dev_err(dev, "wq disable pasid failed.\n");
136 		} else {
137 			idxd_wq_drain(wq);
138 		}
139 	}
140 
141 	if (ctx->sva) {
142 		idxd_cdev_evl_drain_pasid(wq, ctx->pasid);
143 		iommu_sva_unbind_device(ctx->sva);
144 		idxd_xa_pasid_remove(ctx);
145 	}
146 	kfree(ctx);
147 	mutex_lock(&wq->wq_lock);
148 	idxd_wq_put(wq);
149 	mutex_unlock(&wq->wq_lock);
150 }
151 
152 static const struct device_type idxd_cdev_file_type = {
153 	.name = "idxd_file",
154 	.release = idxd_file_dev_release,
155 	.groups = cdev_file_attribute_groups,
156 };
157 
idxd_cdev_dev_release(struct device * dev)158 static void idxd_cdev_dev_release(struct device *dev)
159 {
160 	struct idxd_cdev *idxd_cdev = dev_to_cdev(dev);
161 
162 	kfree(idxd_cdev);
163 }
164 
165 static const struct device_type idxd_cdev_device_type = {
166 	.name = "idxd_cdev",
167 	.release = idxd_cdev_dev_release,
168 };
169 
inode_idxd_cdev(struct inode * inode)170 static inline struct idxd_cdev *inode_idxd_cdev(struct inode *inode)
171 {
172 	struct cdev *cdev = inode->i_cdev;
173 
174 	return container_of(cdev, struct idxd_cdev, cdev);
175 }
176 
inode_wq(struct inode * inode)177 static inline struct idxd_wq *inode_wq(struct inode *inode)
178 {
179 	struct idxd_cdev *idxd_cdev = inode_idxd_cdev(inode);
180 
181 	return idxd_cdev->wq;
182 }
183 
idxd_xa_pasid_remove(struct idxd_user_context * ctx)184 static void idxd_xa_pasid_remove(struct idxd_user_context *ctx)
185 {
186 	struct idxd_wq *wq = ctx->wq;
187 	void *ptr;
188 
189 	mutex_lock(&wq->uc_lock);
190 	ptr = xa_cmpxchg(&wq->upasid_xa, ctx->pasid, ctx, NULL, GFP_KERNEL);
191 	if (ptr != (void *)ctx)
192 		dev_warn(&wq->idxd->pdev->dev, "xarray cmpxchg failed for pasid %u\n",
193 			 ctx->pasid);
194 	mutex_unlock(&wq->uc_lock);
195 }
196 
idxd_user_counter_increment(struct idxd_wq * wq,u32 pasid,int index)197 void idxd_user_counter_increment(struct idxd_wq *wq, u32 pasid, int index)
198 {
199 	struct idxd_user_context *ctx;
200 
201 	if (index >= COUNTER_MAX)
202 		return;
203 
204 	mutex_lock(&wq->uc_lock);
205 	ctx = xa_load(&wq->upasid_xa, pasid);
206 	if (!ctx) {
207 		mutex_unlock(&wq->uc_lock);
208 		return;
209 	}
210 	ctx->counters[index]++;
211 	mutex_unlock(&wq->uc_lock);
212 }
213 
idxd_cdev_open(struct inode * inode,struct file * filp)214 static int idxd_cdev_open(struct inode *inode, struct file *filp)
215 {
216 	struct idxd_user_context *ctx;
217 	struct idxd_device *idxd;
218 	struct idxd_wq *wq;
219 	struct device *dev, *fdev;
220 	int rc = 0;
221 	struct iommu_sva *sva = NULL;
222 	unsigned int pasid;
223 	struct idxd_cdev *idxd_cdev;
224 
225 	wq = inode_wq(inode);
226 	idxd = wq->idxd;
227 	dev = &idxd->pdev->dev;
228 
229 	dev_dbg(dev, "%s called: %d\n", __func__, idxd_wq_refcount(wq));
230 
231 	ctx = kzalloc_obj(*ctx);
232 	if (!ctx)
233 		return -ENOMEM;
234 
235 	mutex_lock(&wq->wq_lock);
236 
237 	if (idxd_wq_refcount(wq) > 0 && wq_dedicated(wq)) {
238 		rc = -EBUSY;
239 		goto failed;
240 	}
241 
242 	ctx->wq = wq;
243 	filp->private_data = ctx;
244 	ctx->pid = current->pid;
245 
246 	if (device_user_pasid_enabled(idxd)) {
247 		sva = iommu_sva_bind_device(dev, current->mm);
248 		if (IS_ERR(sva)) {
249 			rc = PTR_ERR(sva);
250 			dev_err(dev, "pasid allocation failed: %d\n", rc);
251 			goto failed;
252 		}
253 
254 		pasid = iommu_sva_get_pasid(sva);
255 		if (pasid == IOMMU_PASID_INVALID) {
256 			rc = -EINVAL;
257 			goto failed_get_pasid;
258 		}
259 
260 		ctx->sva = sva;
261 		ctx->pasid = pasid;
262 		ctx->mm = current->mm;
263 
264 		mutex_lock(&wq->uc_lock);
265 		rc = xa_insert(&wq->upasid_xa, pasid, ctx, GFP_KERNEL);
266 		mutex_unlock(&wq->uc_lock);
267 		if (rc < 0)
268 			dev_warn(dev, "PASID entry already exist in xarray.\n");
269 
270 		if (wq_dedicated(wq)) {
271 			rc = idxd_wq_set_pasid(wq, pasid);
272 			if (rc < 0) {
273 				dev_err(dev, "wq set pasid failed: %d\n", rc);
274 				goto failed_set_pasid;
275 			}
276 		}
277 	}
278 
279 	idxd_cdev = wq->idxd_cdev;
280 	ctx->id = ida_alloc(&file_ida, GFP_KERNEL);
281 	if (ctx->id < 0) {
282 		dev_warn(dev, "ida alloc failure\n");
283 		goto failed_ida;
284 	}
285 	ctx->idxd_dev.type  = IDXD_DEV_CDEV_FILE;
286 	fdev = user_ctx_dev(ctx);
287 	device_initialize(fdev);
288 	fdev->parent = cdev_dev(idxd_cdev);
289 	fdev->bus = &dsa_bus_type;
290 	fdev->type = &idxd_cdev_file_type;
291 
292 	rc = dev_set_name(fdev, "file%d", ctx->id);
293 	if (rc < 0) {
294 		dev_warn(dev, "set name failure\n");
295 		goto failed_dev_name;
296 	}
297 
298 	rc = device_add(fdev);
299 	if (rc < 0) {
300 		dev_warn(dev, "file device add failure\n");
301 		goto failed_dev_add;
302 	}
303 
304 	idxd_wq_get(wq);
305 	mutex_unlock(&wq->wq_lock);
306 	return 0;
307 
308 failed_dev_add:
309 failed_dev_name:
310 	put_device(fdev);
311 failed_ida:
312 failed_set_pasid:
313 	if (device_user_pasid_enabled(idxd))
314 		idxd_xa_pasid_remove(ctx);
315 failed_get_pasid:
316 	if (device_user_pasid_enabled(idxd) && !IS_ERR_OR_NULL(sva))
317 		iommu_sva_unbind_device(sva);
318 failed:
319 	mutex_unlock(&wq->wq_lock);
320 	kfree(ctx);
321 	return rc;
322 }
323 
idxd_cdev_evl_drain_pasid(struct idxd_wq * wq,u32 pasid)324 static void idxd_cdev_evl_drain_pasid(struct idxd_wq *wq, u32 pasid)
325 {
326 	struct idxd_device *idxd = wq->idxd;
327 	struct idxd_evl *evl = idxd->evl;
328 	union evl_status_reg status;
329 	u16 h, t, size;
330 	int ent_size = evl_ent_size(idxd);
331 	struct __evl_entry *entry_head;
332 
333 	if (!evl)
334 		return;
335 
336 	mutex_lock(&evl->lock);
337 	status.bits = ioread64(idxd->reg_base + IDXD_EVLSTATUS_OFFSET);
338 	t = status.tail;
339 	h = status.head;
340 	size = evl->size;
341 
342 	while (h != t) {
343 		entry_head = (struct __evl_entry *)(evl->log + (h * ent_size));
344 		if (entry_head->pasid == pasid && entry_head->wq_idx == wq->id)
345 			set_bit(h, evl->bmap);
346 		h = (h + 1) % size;
347 	}
348 	if (wq->wq)
349 		drain_workqueue(wq->wq);
350 
351 	mutex_unlock(&evl->lock);
352 }
353 
idxd_cdev_release(struct inode * node,struct file * filep)354 static int idxd_cdev_release(struct inode *node, struct file *filep)
355 {
356 	struct idxd_user_context *ctx = filep->private_data;
357 	struct idxd_wq *wq = ctx->wq;
358 	struct idxd_device *idxd = wq->idxd;
359 	struct device *dev = &idxd->pdev->dev;
360 
361 	dev_dbg(dev, "%s called\n", __func__);
362 	filep->private_data = NULL;
363 
364 	device_unregister(user_ctx_dev(ctx));
365 
366 	return 0;
367 }
368 
check_vma(struct idxd_wq * wq,struct vm_area_struct * vma,const char * func)369 static int check_vma(struct idxd_wq *wq, struct vm_area_struct *vma,
370 		     const char *func)
371 {
372 	struct device *dev = &wq->idxd->pdev->dev;
373 
374 	if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
375 		dev_info_ratelimited(dev,
376 				     "%s: %s: mapping too large: %lu\n",
377 				     current->comm, func,
378 				     vma->vm_end - vma->vm_start);
379 		return -EINVAL;
380 	}
381 
382 	return 0;
383 }
384 
idxd_cdev_mmap(struct file * filp,struct vm_area_struct * vma)385 static int idxd_cdev_mmap(struct file *filp, struct vm_area_struct *vma)
386 {
387 	struct idxd_user_context *ctx = filp->private_data;
388 	struct idxd_wq *wq = ctx->wq;
389 	struct idxd_device *idxd = wq->idxd;
390 	struct pci_dev *pdev = idxd->pdev;
391 	phys_addr_t base = pci_resource_start(pdev, IDXD_WQ_BAR);
392 	unsigned long pfn;
393 	int rc;
394 
395 	dev_dbg(&pdev->dev, "%s called\n", __func__);
396 
397 	/*
398 	 * Due to an erratum in some of the devices supported by the driver,
399 	 * direct user submission to the device can be unsafe.
400 	 * (See the INTEL-SA-01084 security advisory)
401 	 *
402 	 * For the devices that exhibit this behavior, require that the user
403 	 * has CAP_SYS_RAWIO capabilities.
404 	 */
405 	if (!idxd->user_submission_safe && !capable(CAP_SYS_RAWIO))
406 		return -EPERM;
407 
408 	if (current->mm != ctx->mm)
409 		return -EPERM;
410 
411 	rc = check_vma(wq, vma, __func__);
412 	if (rc < 0)
413 		return rc;
414 
415 	vm_flags_set(vma, VM_DONTCOPY);
416 	pfn = (base + idxd_get_wq_portal_full_offset(wq->id,
417 				IDXD_PORTAL_LIMITED)) >> PAGE_SHIFT;
418 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
419 	vma->vm_private_data = ctx;
420 
421 	return io_remap_pfn_range(vma, vma->vm_start, pfn, PAGE_SIZE,
422 			vma->vm_page_prot);
423 }
424 
idxd_submit_user_descriptor(struct idxd_user_context * ctx,struct dsa_hw_desc __user * udesc)425 static int idxd_submit_user_descriptor(struct idxd_user_context *ctx,
426 				       struct dsa_hw_desc __user *udesc)
427 {
428 	struct idxd_wq *wq = ctx->wq;
429 	struct idxd_dev *idxd_dev = &wq->idxd->idxd_dev;
430 	const uint64_t comp_addr_align = is_dsa_dev(idxd_dev) ? 0x20 : 0x40;
431 	void __iomem *portal = idxd_wq_portal_addr(wq);
432 	struct dsa_hw_desc descriptor __aligned(64);
433 	int rc;
434 
435 	rc = copy_from_user(&descriptor, udesc, sizeof(descriptor));
436 	if (rc)
437 		return -EFAULT;
438 
439 	/*
440 	 * DSA devices are capable of indirect ("batch") command submission.
441 	 * On devices where direct user submissions are not safe, we cannot
442 	 * allow this since there is no good way for us to verify these
443 	 * indirect commands. Narrow the restriction of operations with the
444 	 * BATCH opcode to only DSA version 1 devices.
445 	 */
446 	if (is_dsa_dev(idxd_dev) && descriptor.opcode == DSA_OPCODE_BATCH &&
447 	    wq->idxd->hw.version == DEVICE_VERSION_1 &&
448 	    !wq->idxd->user_submission_safe)
449 		return -EINVAL;
450 	/*
451 	 * As per the programming specification, the completion address must be
452 	 * aligned to 32 or 64 bytes. If this is violated the hardware
453 	 * engine can get very confused (security issue).
454 	 */
455 	if (!IS_ALIGNED(descriptor.completion_addr, comp_addr_align))
456 		return -EINVAL;
457 
458 	if (wq_dedicated(wq))
459 		iosubmit_cmds512(portal, &descriptor, 1);
460 	else {
461 		descriptor.priv = 0;
462 		descriptor.pasid = ctx->pasid;
463 		rc = idxd_enqcmds(wq, portal, &descriptor);
464 		if (rc < 0)
465 			return rc;
466 	}
467 
468 	return 0;
469 }
470 
idxd_cdev_write(struct file * filp,const char __user * buf,size_t len,loff_t * unused)471 static ssize_t idxd_cdev_write(struct file *filp, const char __user *buf, size_t len,
472 			       loff_t *unused)
473 {
474 	struct dsa_hw_desc __user *udesc = (struct dsa_hw_desc __user *)buf;
475 	struct idxd_user_context *ctx = filp->private_data;
476 	ssize_t written = 0;
477 	int i;
478 
479 	if (current->mm != ctx->mm)
480 		return -EPERM;
481 
482 	for (i = 0; i < len/sizeof(struct dsa_hw_desc); i++) {
483 		int rc = idxd_submit_user_descriptor(ctx, udesc + i);
484 
485 		if (rc)
486 			return written ? written : rc;
487 
488 		written += sizeof(struct dsa_hw_desc);
489 	}
490 
491 	return written;
492 }
493 
idxd_cdev_poll(struct file * filp,struct poll_table_struct * wait)494 static __poll_t idxd_cdev_poll(struct file *filp,
495 			       struct poll_table_struct *wait)
496 {
497 	struct idxd_user_context *ctx = filp->private_data;
498 	struct idxd_wq *wq = ctx->wq;
499 	struct idxd_device *idxd = wq->idxd;
500 	__poll_t out = 0;
501 
502 	if (current->mm != ctx->mm)
503 		return POLLNVAL;
504 
505 	poll_wait(filp, &wq->err_queue, wait);
506 	spin_lock(&idxd->dev_lock);
507 	if (idxd->sw_err.valid)
508 		out = EPOLLIN | EPOLLRDNORM;
509 	spin_unlock(&idxd->dev_lock);
510 
511 	return out;
512 }
513 
514 static const struct file_operations idxd_cdev_fops = {
515 	.owner = THIS_MODULE,
516 	.open = idxd_cdev_open,
517 	.release = idxd_cdev_release,
518 	.mmap = idxd_cdev_mmap,
519 	.write = idxd_cdev_write,
520 	.poll = idxd_cdev_poll,
521 };
522 
idxd_cdev_get_major(struct idxd_device * idxd)523 int idxd_cdev_get_major(struct idxd_device *idxd)
524 {
525 	return MAJOR(ictx[idxd->data->type].devt);
526 }
527 
idxd_wq_add_cdev(struct idxd_wq * wq)528 int idxd_wq_add_cdev(struct idxd_wq *wq)
529 {
530 	struct idxd_device *idxd = wq->idxd;
531 	struct idxd_cdev *idxd_cdev;
532 	struct cdev *cdev;
533 	struct device *dev;
534 	struct idxd_cdev_context *cdev_ctx;
535 	int rc, minor;
536 
537 	idxd_cdev = kzalloc_obj(*idxd_cdev);
538 	if (!idxd_cdev)
539 		return -ENOMEM;
540 
541 	idxd_cdev->idxd_dev.type = IDXD_DEV_CDEV;
542 	idxd_cdev->wq = wq;
543 	cdev = &idxd_cdev->cdev;
544 	dev = cdev_dev(idxd_cdev);
545 	cdev_ctx = &ictx[wq->idxd->data->type];
546 	minor = ida_alloc_max(&cdev_ctx->minor_ida, MINORMASK, GFP_KERNEL);
547 	if (minor < 0) {
548 		kfree(idxd_cdev);
549 		return minor;
550 	}
551 	idxd_cdev->minor = minor;
552 
553 	device_initialize(dev);
554 	dev->parent = wq_confdev(wq);
555 	dev->bus = &dsa_bus_type;
556 	dev->type = &idxd_cdev_device_type;
557 	dev->devt = MKDEV(MAJOR(cdev_ctx->devt), minor);
558 
559 	rc = dev_set_name(dev, "%s/wq%u.%u", idxd->data->name_prefix, idxd->id, wq->id);
560 	if (rc < 0)
561 		goto err;
562 
563 	wq->idxd_cdev = idxd_cdev;
564 	cdev_init(cdev, &idxd_cdev_fops);
565 	rc = cdev_device_add(cdev, dev);
566 	if (rc) {
567 		dev_dbg(&wq->idxd->pdev->dev, "cdev_add failed: %d\n", rc);
568 		goto err;
569 	}
570 
571 	return 0;
572 
573  err:
574 	put_device(dev);
575 	wq->idxd_cdev = NULL;
576 	return rc;
577 }
578 
idxd_wq_del_cdev(struct idxd_wq * wq)579 void idxd_wq_del_cdev(struct idxd_wq *wq)
580 {
581 	struct idxd_cdev_context *cdev_ctx;
582 	struct idxd_cdev *idxd_cdev;
583 
584 	idxd_cdev = wq->idxd_cdev;
585 	wq->idxd_cdev = NULL;
586 	cdev_device_del(&idxd_cdev->cdev, cdev_dev(idxd_cdev));
587 
588 	cdev_ctx = &ictx[wq->idxd->data->type];
589 	ida_free(&cdev_ctx->minor_ida, idxd_cdev->minor);
590 	put_device(cdev_dev(idxd_cdev));
591 }
592 
idxd_user_drv_probe(struct idxd_dev * idxd_dev)593 static int idxd_user_drv_probe(struct idxd_dev *idxd_dev)
594 {
595 	struct device *dev = &idxd_dev->conf_dev;
596 	struct idxd_wq *wq = idxd_dev_to_wq(idxd_dev);
597 	struct idxd_device *idxd = wq->idxd;
598 	int rc;
599 
600 	if (idxd->state != IDXD_DEV_ENABLED)
601 		return -ENXIO;
602 
603 	mutex_lock(&wq->wq_lock);
604 
605 	if (!idxd_wq_driver_name_match(wq, dev)) {
606 		idxd->cmd_status = IDXD_SCMD_WQ_NO_DRV_NAME;
607 		rc = -ENODEV;
608 		goto wq_err;
609 	}
610 
611 	/*
612 	 * User type WQ is enabled only when SVA is enabled for two reasons:
613 	 *   - If no IOMMU or IOMMU Passthrough without SVA, userspace
614 	 *     can directly access physical address through the WQ.
615 	 *   - The IDXD cdev driver does not provide any ways to pin
616 	 *     user pages and translate the address from user VA to IOVA or
617 	 *     PA without IOMMU SVA. Therefore the application has no way
618 	 *     to instruct the device to perform DMA function. This makes
619 	 *     the cdev not usable for normal application usage.
620 	 */
621 	if (!device_user_pasid_enabled(idxd)) {
622 		idxd->cmd_status = IDXD_SCMD_WQ_USER_NO_IOMMU;
623 		dev_dbg(&idxd->pdev->dev,
624 			"User type WQ cannot be enabled without SVA.\n");
625 
626 		rc = -EOPNOTSUPP;
627 		goto wq_err;
628 	}
629 
630 	wq->wq = create_workqueue(dev_name(wq_confdev(wq)));
631 	if (!wq->wq) {
632 		rc = -ENOMEM;
633 		goto wq_err;
634 	}
635 
636 	wq->type = IDXD_WQT_USER;
637 	rc = idxd_drv_enable_wq(wq);
638 	if (rc < 0)
639 		goto err;
640 
641 	rc = idxd_wq_add_cdev(wq);
642 	if (rc < 0) {
643 		idxd->cmd_status = IDXD_SCMD_CDEV_ERR;
644 		goto err_cdev;
645 	}
646 
647 	idxd->cmd_status = 0;
648 	mutex_unlock(&wq->wq_lock);
649 	return 0;
650 
651 err_cdev:
652 	idxd_drv_disable_wq(wq);
653 err:
654 	destroy_workqueue(wq->wq);
655 	wq->type = IDXD_WQT_NONE;
656 wq_err:
657 	mutex_unlock(&wq->wq_lock);
658 	return rc;
659 }
660 
idxd_user_drv_remove(struct idxd_dev * idxd_dev)661 static void idxd_user_drv_remove(struct idxd_dev *idxd_dev)
662 {
663 	struct idxd_wq *wq = idxd_dev_to_wq(idxd_dev);
664 
665 	mutex_lock(&wq->wq_lock);
666 	idxd_wq_del_cdev(wq);
667 	idxd_drv_disable_wq(wq);
668 	wq->type = IDXD_WQT_NONE;
669 	destroy_workqueue(wq->wq);
670 	wq->wq = NULL;
671 	mutex_unlock(&wq->wq_lock);
672 }
673 
674 static enum idxd_dev_type dev_types[] = {
675 	IDXD_DEV_WQ,
676 	IDXD_DEV_NONE,
677 };
678 
679 struct idxd_device_driver idxd_user_drv = {
680 	.probe = idxd_user_drv_probe,
681 	.remove = idxd_user_drv_remove,
682 	.name = "user",
683 	.type = dev_types,
684 };
685 EXPORT_SYMBOL_GPL(idxd_user_drv);
686 
idxd_cdev_register(void)687 int idxd_cdev_register(void)
688 {
689 	int rc, i;
690 
691 	for (i = 0; i < IDXD_TYPE_MAX; i++) {
692 		ida_init(&ictx[i].minor_ida);
693 		rc = alloc_chrdev_region(&ictx[i].devt, 0, MINORMASK,
694 					 ictx[i].name);
695 		if (rc)
696 			goto err_free_chrdev_region;
697 	}
698 
699 	return 0;
700 
701 err_free_chrdev_region:
702 	for (i--; i >= 0; i--)
703 		unregister_chrdev_region(ictx[i].devt, MINORMASK);
704 
705 	return rc;
706 }
707 
idxd_cdev_remove(void)708 void idxd_cdev_remove(void)
709 {
710 	int i;
711 
712 	for (i = 0; i < IDXD_TYPE_MAX; i++) {
713 		unregister_chrdev_region(ictx[i].devt, MINORMASK);
714 		ida_destroy(&ictx[i].minor_ida);
715 	}
716 }
717 
718 /**
719  * idxd_copy_cr - copy completion record to user address space found by wq and
720  *		  PASID
721  * @wq:		work queue
722  * @pasid:	PASID
723  * @addr:	user fault address to write
724  * @cr:		completion record
725  * @len:	number of bytes to copy
726  *
727  * This is called by a work that handles completion record fault.
728  *
729  * Return: number of bytes copied.
730  */
idxd_copy_cr(struct idxd_wq * wq,ioasid_t pasid,unsigned long addr,void * cr,int len)731 int idxd_copy_cr(struct idxd_wq *wq, ioasid_t pasid, unsigned long addr,
732 		 void *cr, int len)
733 {
734 	struct device *dev = &wq->idxd->pdev->dev;
735 	int left = len, status_size = 1;
736 	struct idxd_user_context *ctx;
737 	struct mm_struct *mm;
738 
739 	mutex_lock(&wq->uc_lock);
740 
741 	ctx = xa_load(&wq->upasid_xa, pasid);
742 	if (!ctx) {
743 		dev_warn(dev, "No user context\n");
744 		goto out;
745 	}
746 
747 	mm = ctx->mm;
748 	/*
749 	 * The completion record fault handling work is running in kernel
750 	 * thread context. It temporarily switches to the mm to copy cr
751 	 * to addr in the mm.
752 	 */
753 	kthread_use_mm(mm);
754 	left = copy_to_user((void __user *)addr + status_size, cr + status_size,
755 			    len - status_size);
756 	/*
757 	 * Copy status only after the rest of completion record is copied
758 	 * successfully so that the user gets the complete completion record
759 	 * when a non-zero status is polled.
760 	 */
761 	if (!left) {
762 		u8 status;
763 
764 		/*
765 		 * Ensure that the completion record's status field is written
766 		 * after the rest of the completion record has been written.
767 		 * This ensures that the user receives the correct completion
768 		 * record information once polling for a non-zero status.
769 		 */
770 		wmb();
771 		status = *(u8 *)cr;
772 		if (put_user(status, (u8 __user *)addr))
773 			left += status_size;
774 	} else {
775 		left += status_size;
776 	}
777 	kthread_unuse_mm(mm);
778 
779 out:
780 	mutex_unlock(&wq->uc_lock);
781 
782 	return len - left;
783 }
784