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/io-64-nonatomic-lo-hi.h>
8 #include <linux/dmaengine.h>
9 #include <linux/delay.h>
10 #include <uapi/linux/idxd.h>
11 #include "../dmaengine.h"
12 #include "idxd.h"
13 #include "registers.h"
14
15 enum irq_work_type {
16 IRQ_WORK_NORMAL = 0,
17 IRQ_WORK_PROCESS_FAULT,
18 };
19
20 struct idxd_resubmit {
21 struct work_struct work;
22 struct idxd_desc *desc;
23 };
24
25 struct idxd_int_handle_revoke {
26 struct work_struct work;
27 struct idxd_device *idxd;
28 };
29
idxd_device_reinit(struct work_struct * work)30 static void idxd_device_reinit(struct work_struct *work)
31 {
32 struct idxd_device *idxd = container_of(work, struct idxd_device, work);
33 struct device *dev = &idxd->pdev->dev;
34 int rc, i;
35
36 idxd_device_reset(idxd);
37 rc = idxd_device_config(idxd);
38 if (rc < 0)
39 goto out;
40
41 rc = idxd_device_enable(idxd);
42 if (rc < 0)
43 goto out;
44
45 for (i = 0; i < idxd->max_wqs; i++) {
46 if (test_bit(i, idxd->wq_enable_map)) {
47 struct idxd_wq *wq = idxd->wqs[i];
48
49 rc = idxd_wq_enable(wq);
50 if (rc < 0) {
51 clear_bit(i, idxd->wq_enable_map);
52 dev_warn(dev, "Unable to re-enable wq %s\n",
53 dev_name(wq_confdev(wq)));
54 }
55 }
56 }
57
58 return;
59
60 out:
61 idxd_device_clear_state(idxd);
62 }
63
64 /*
65 * The function sends a drain descriptor for the interrupt handle. The drain ensures
66 * all descriptors with this interrupt handle is flushed and the interrupt
67 * will allow the cleanup of the outstanding descriptors.
68 */
idxd_int_handle_revoke_drain(struct idxd_irq_entry * ie)69 static void idxd_int_handle_revoke_drain(struct idxd_irq_entry *ie)
70 {
71 struct idxd_wq *wq = ie_to_wq(ie);
72 struct idxd_device *idxd = wq->idxd;
73 struct device *dev = &idxd->pdev->dev;
74 struct dsa_hw_desc desc = {};
75 void __iomem *portal;
76 int rc;
77
78 /* Issue a simple drain operation with interrupt but no completion record */
79 desc.flags = IDXD_OP_FLAG_RCI;
80 desc.opcode = DSA_OPCODE_DRAIN;
81 desc.priv = 1;
82
83 if (ie->pasid != INVALID_IOASID)
84 desc.pasid = ie->pasid;
85 desc.int_handle = ie->int_handle;
86 portal = idxd_wq_portal_addr(wq);
87
88 /*
89 * The wmb() makes sure that the descriptor is all there before we
90 * issue.
91 */
92 wmb();
93 if (wq_dedicated(wq)) {
94 iosubmit_cmds512(portal, &desc, 1);
95 } else {
96 rc = idxd_enqcmds(wq, portal, &desc);
97 /* This should not fail unless hardware failed. */
98 if (rc < 0)
99 dev_warn(dev, "Failed to submit drain desc on wq %d\n", wq->id);
100 }
101 }
102
idxd_abort_invalid_int_handle_descs(struct idxd_irq_entry * ie)103 static void idxd_abort_invalid_int_handle_descs(struct idxd_irq_entry *ie)
104 {
105 LIST_HEAD(flist);
106 struct idxd_desc *d, *t;
107 struct llist_node *head;
108
109 spin_lock(&ie->list_lock);
110 head = llist_del_all(&ie->pending_llist);
111 if (head) {
112 llist_for_each_entry_safe(d, t, head, llnode)
113 list_add_tail(&d->list, &ie->work_list);
114 }
115
116 list_for_each_entry_safe(d, t, &ie->work_list, list) {
117 if (d->completion->status == DSA_COMP_INT_HANDLE_INVAL)
118 list_move_tail(&d->list, &flist);
119 }
120 spin_unlock(&ie->list_lock);
121
122 list_for_each_entry_safe(d, t, &flist, list) {
123 list_del(&d->list);
124 idxd_dma_complete_txd(d, IDXD_COMPLETE_ABORT, true);
125 }
126 }
127
idxd_int_handle_revoke(struct work_struct * work)128 static void idxd_int_handle_revoke(struct work_struct *work)
129 {
130 struct idxd_int_handle_revoke *revoke =
131 container_of(work, struct idxd_int_handle_revoke, work);
132 struct idxd_device *idxd = revoke->idxd;
133 struct pci_dev *pdev = idxd->pdev;
134 struct device *dev = &pdev->dev;
135 int i, new_handle, rc;
136
137 if (!idxd->request_int_handles) {
138 kfree(revoke);
139 dev_warn(dev, "Unexpected int handle refresh interrupt.\n");
140 return;
141 }
142
143 /*
144 * The loop attempts to acquire new interrupt handle for all interrupt
145 * vectors that supports a handle. If a new interrupt handle is acquired and the
146 * wq is kernel type, the driver will kill the percpu_ref to pause all
147 * ongoing descriptor submissions. The interrupt handle is then changed.
148 * After change, the percpu_ref is revived and all the pending submissions
149 * are woken to try again. A drain is sent to for the interrupt handle
150 * at the end to make sure all invalid int handle descriptors are processed.
151 */
152 for (i = 1; i < idxd->irq_cnt; i++) {
153 struct idxd_irq_entry *ie = idxd_get_ie(idxd, i);
154 struct idxd_wq *wq = ie_to_wq(ie);
155
156 if (ie->int_handle == INVALID_INT_HANDLE)
157 continue;
158
159 rc = idxd_device_request_int_handle(idxd, i, &new_handle, IDXD_IRQ_MSIX);
160 if (rc < 0) {
161 dev_warn(dev, "get int handle %d failed: %d\n", i, rc);
162 /*
163 * Failed to acquire new interrupt handle. Kill the WQ
164 * and release all the pending submitters. The submitters will
165 * get error return code and handle appropriately.
166 */
167 ie->int_handle = INVALID_INT_HANDLE;
168 idxd_wq_quiesce(wq);
169 idxd_abort_invalid_int_handle_descs(ie);
170 continue;
171 }
172
173 /* No change in interrupt handle, nothing needs to be done */
174 if (ie->int_handle == new_handle)
175 continue;
176
177 if (wq->state != IDXD_WQ_ENABLED || wq->type != IDXD_WQT_KERNEL) {
178 /*
179 * All the MSIX interrupts are allocated at once during probe.
180 * Therefore we need to update all interrupts even if the WQ
181 * isn't supporting interrupt operations.
182 */
183 ie->int_handle = new_handle;
184 continue;
185 }
186
187 mutex_lock(&wq->wq_lock);
188 reinit_completion(&wq->wq_resurrect);
189
190 /* Kill percpu_ref to pause additional descriptor submissions */
191 percpu_ref_kill(&wq->wq_active);
192
193 /* Wait for all submitters quiesce before we change interrupt handle */
194 wait_for_completion(&wq->wq_dead);
195
196 ie->int_handle = new_handle;
197
198 /* Revive percpu ref and wake up all the waiting submitters */
199 percpu_ref_reinit(&wq->wq_active);
200 complete_all(&wq->wq_resurrect);
201 mutex_unlock(&wq->wq_lock);
202
203 /*
204 * The delay here is to wait for all possible MOVDIR64B that
205 * are issued before percpu_ref_kill() has happened to have
206 * reached the PCIe domain before the drain is issued. The driver
207 * needs to ensure that the drain descriptor issued does not pass
208 * all the other issued descriptors that contain the invalid
209 * interrupt handle in order to ensure that the drain descriptor
210 * interrupt will allow the cleanup of all the descriptors with
211 * invalid interrupt handle.
212 */
213 if (wq_dedicated(wq))
214 udelay(100);
215 idxd_int_handle_revoke_drain(ie);
216 }
217 kfree(revoke);
218 }
219
process_misc_interrupts(struct idxd_device * idxd,u32 cause)220 static int process_misc_interrupts(struct idxd_device *idxd, u32 cause)
221 {
222 struct device *dev = &idxd->pdev->dev;
223 union gensts_reg gensts;
224 u32 val = 0;
225 int i;
226 bool err = false;
227
228 if (cause & IDXD_INTC_HALT_STATE)
229 goto halt;
230
231 if (cause & IDXD_INTC_ERR) {
232 spin_lock(&idxd->dev_lock);
233 for (i = 0; i < 4; i++)
234 idxd->sw_err.bits[i] = ioread64(idxd->reg_base +
235 IDXD_SWERR_OFFSET + i * sizeof(u64));
236
237 iowrite64(idxd->sw_err.bits[0] & IDXD_SWERR_ACK,
238 idxd->reg_base + IDXD_SWERR_OFFSET);
239
240 if (idxd->sw_err.valid && idxd->sw_err.wq_idx_valid) {
241 int id = idxd->sw_err.wq_idx;
242 struct idxd_wq *wq = idxd->wqs[id];
243
244 if (wq->type == IDXD_WQT_USER)
245 wake_up_interruptible(&wq->err_queue);
246 } else {
247 int i;
248
249 for (i = 0; i < idxd->max_wqs; i++) {
250 struct idxd_wq *wq = idxd->wqs[i];
251
252 if (wq->type == IDXD_WQT_USER)
253 wake_up_interruptible(&wq->err_queue);
254 }
255 }
256
257 spin_unlock(&idxd->dev_lock);
258 val |= IDXD_INTC_ERR;
259
260 for (i = 0; i < 4; i++)
261 dev_warn(dev, "err[%d]: %#16.16llx\n",
262 i, idxd->sw_err.bits[i]);
263 err = true;
264 }
265
266 if (cause & IDXD_INTC_INT_HANDLE_REVOKED) {
267 struct idxd_int_handle_revoke *revoke;
268
269 val |= IDXD_INTC_INT_HANDLE_REVOKED;
270
271 revoke = kzalloc(sizeof(*revoke), GFP_ATOMIC);
272 if (revoke) {
273 revoke->idxd = idxd;
274 INIT_WORK(&revoke->work, idxd_int_handle_revoke);
275 queue_work(idxd->wq, &revoke->work);
276
277 } else {
278 dev_err(dev, "Failed to allocate work for int handle revoke\n");
279 idxd_wqs_quiesce(idxd);
280 }
281 }
282
283 if (cause & IDXD_INTC_CMD) {
284 val |= IDXD_INTC_CMD;
285 complete(idxd->cmd_done);
286 }
287
288 if (cause & IDXD_INTC_OCCUPY) {
289 /* Driver does not utilize occupancy interrupt */
290 val |= IDXD_INTC_OCCUPY;
291 }
292
293 if (cause & IDXD_INTC_PERFMON_OVFL) {
294 val |= IDXD_INTC_PERFMON_OVFL;
295 perfmon_counter_overflow(idxd);
296 }
297
298 val ^= cause;
299 if (val)
300 dev_warn_once(dev, "Unexpected interrupt cause bits set: %#x\n",
301 val);
302
303 if (!err)
304 return 0;
305
306 halt:
307 gensts.bits = ioread32(idxd->reg_base + IDXD_GENSTATS_OFFSET);
308 if (gensts.state == IDXD_DEVICE_STATE_HALT) {
309 idxd->state = IDXD_DEV_HALTED;
310 if (gensts.reset_type == IDXD_DEVICE_RESET_SOFTWARE) {
311 /*
312 * If we need a software reset, we will throw the work
313 * on a system workqueue in order to allow interrupts
314 * for the device command completions.
315 */
316 INIT_WORK(&idxd->work, idxd_device_reinit);
317 queue_work(idxd->wq, &idxd->work);
318 } else {
319 idxd->state = IDXD_DEV_HALTED;
320 idxd_wqs_quiesce(idxd);
321 idxd_wqs_unmap_portal(idxd);
322 idxd_device_clear_state(idxd);
323 dev_err(&idxd->pdev->dev,
324 "idxd halted, need %s.\n",
325 gensts.reset_type == IDXD_DEVICE_RESET_FLR ?
326 "FLR" : "system reset");
327 return -ENXIO;
328 }
329 }
330
331 return 0;
332 }
333
idxd_misc_thread(int vec,void * data)334 irqreturn_t idxd_misc_thread(int vec, void *data)
335 {
336 struct idxd_irq_entry *irq_entry = data;
337 struct idxd_device *idxd = ie_to_idxd(irq_entry);
338 int rc;
339 u32 cause;
340
341 cause = ioread32(idxd->reg_base + IDXD_INTCAUSE_OFFSET);
342 if (cause)
343 iowrite32(cause, idxd->reg_base + IDXD_INTCAUSE_OFFSET);
344
345 while (cause) {
346 rc = process_misc_interrupts(idxd, cause);
347 if (rc < 0)
348 break;
349 cause = ioread32(idxd->reg_base + IDXD_INTCAUSE_OFFSET);
350 if (cause)
351 iowrite32(cause, idxd->reg_base + IDXD_INTCAUSE_OFFSET);
352 }
353
354 return IRQ_HANDLED;
355 }
356
idxd_int_handle_resubmit_work(struct work_struct * work)357 static void idxd_int_handle_resubmit_work(struct work_struct *work)
358 {
359 struct idxd_resubmit *irw = container_of(work, struct idxd_resubmit, work);
360 struct idxd_desc *desc = irw->desc;
361 struct idxd_wq *wq = desc->wq;
362 int rc;
363
364 desc->completion->status = 0;
365 rc = idxd_submit_desc(wq, desc);
366 if (rc < 0) {
367 dev_dbg(&wq->idxd->pdev->dev, "Failed to resubmit desc %d to wq %d.\n",
368 desc->id, wq->id);
369 /*
370 * If the error is not -EAGAIN, it means the submission failed due to wq
371 * has been killed instead of ENQCMDS failure. Here the driver needs to
372 * notify the submitter of the failure by reporting abort status.
373 *
374 * -EAGAIN comes from ENQCMDS failure. idxd_submit_desc() will handle the
375 * abort.
376 */
377 if (rc != -EAGAIN) {
378 desc->completion->status = IDXD_COMP_DESC_ABORT;
379 idxd_dma_complete_txd(desc, IDXD_COMPLETE_ABORT, false);
380 }
381 idxd_free_desc(wq, desc);
382 }
383 kfree(irw);
384 }
385
idxd_queue_int_handle_resubmit(struct idxd_desc * desc)386 bool idxd_queue_int_handle_resubmit(struct idxd_desc *desc)
387 {
388 struct idxd_wq *wq = desc->wq;
389 struct idxd_device *idxd = wq->idxd;
390 struct idxd_resubmit *irw;
391
392 irw = kzalloc(sizeof(*irw), GFP_KERNEL);
393 if (!irw)
394 return false;
395
396 irw->desc = desc;
397 INIT_WORK(&irw->work, idxd_int_handle_resubmit_work);
398 queue_work(idxd->wq, &irw->work);
399 return true;
400 }
401
irq_process_pending_llist(struct idxd_irq_entry * irq_entry)402 static void irq_process_pending_llist(struct idxd_irq_entry *irq_entry)
403 {
404 struct idxd_desc *desc, *t;
405 struct llist_node *head;
406
407 head = llist_del_all(&irq_entry->pending_llist);
408 if (!head)
409 return;
410
411 llist_for_each_entry_safe(desc, t, head, llnode) {
412 u8 status = desc->completion->status & DSA_COMP_STATUS_MASK;
413
414 if (status) {
415 /*
416 * Check against the original status as ABORT is software defined
417 * and 0xff, which DSA_COMP_STATUS_MASK can mask out.
418 */
419 if (unlikely(desc->completion->status == IDXD_COMP_DESC_ABORT)) {
420 idxd_dma_complete_txd(desc, IDXD_COMPLETE_ABORT, true);
421 continue;
422 }
423
424 idxd_dma_complete_txd(desc, IDXD_COMPLETE_NORMAL, true);
425 } else {
426 spin_lock(&irq_entry->list_lock);
427 list_add_tail(&desc->list,
428 &irq_entry->work_list);
429 spin_unlock(&irq_entry->list_lock);
430 }
431 }
432 }
433
irq_process_work_list(struct idxd_irq_entry * irq_entry)434 static void irq_process_work_list(struct idxd_irq_entry *irq_entry)
435 {
436 LIST_HEAD(flist);
437 struct idxd_desc *desc, *n;
438
439 /*
440 * This lock protects list corruption from access of list outside of the irq handler
441 * thread.
442 */
443 spin_lock(&irq_entry->list_lock);
444 if (list_empty(&irq_entry->work_list)) {
445 spin_unlock(&irq_entry->list_lock);
446 return;
447 }
448
449 list_for_each_entry_safe(desc, n, &irq_entry->work_list, list) {
450 if (desc->completion->status) {
451 list_move_tail(&desc->list, &flist);
452 }
453 }
454
455 spin_unlock(&irq_entry->list_lock);
456
457 list_for_each_entry(desc, &flist, list) {
458 /*
459 * Check against the original status as ABORT is software defined
460 * and 0xff, which DSA_COMP_STATUS_MASK can mask out.
461 */
462 if (unlikely(desc->completion->status == IDXD_COMP_DESC_ABORT)) {
463 idxd_dma_complete_txd(desc, IDXD_COMPLETE_ABORT, true);
464 continue;
465 }
466
467 idxd_dma_complete_txd(desc, IDXD_COMPLETE_NORMAL, true);
468 }
469 }
470
idxd_wq_thread(int irq,void * data)471 irqreturn_t idxd_wq_thread(int irq, void *data)
472 {
473 struct idxd_irq_entry *irq_entry = data;
474
475 /*
476 * There are two lists we are processing. The pending_llist is where
477 * submmiter adds all the submitted descriptor after sending it to
478 * the workqueue. It's a lockless singly linked list. The work_list
479 * is the common linux double linked list. We are in a scenario of
480 * multiple producers and a single consumer. The producers are all
481 * the kernel submitters of descriptors, and the consumer is the
482 * kernel irq handler thread for the msix vector when using threaded
483 * irq. To work with the restrictions of llist to remain lockless,
484 * we are doing the following steps:
485 * 1. Iterate through the work_list and process any completed
486 * descriptor. Delete the completed entries during iteration.
487 * 2. llist_del_all() from the pending list.
488 * 3. Iterate through the llist that was deleted from the pending list
489 * and process the completed entries.
490 * 4. If the entry is still waiting on hardware, list_add_tail() to
491 * the work_list.
492 */
493 irq_process_work_list(irq_entry);
494 irq_process_pending_llist(irq_entry);
495
496 return IRQ_HANDLED;
497 }
498