2 * Intel I/OAT DMA Linux driver
3 * Copyright(c) 2004 - 2015 Intel Corporation.
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 * The full GNU General Public License is included in this distribution in
15 * the file called "COPYING".
20 * This driver supports an Intel I/OAT DMA engine, which does asynchronous
24 #include <linux/init.h>
25 #include <linux/module.h>
26 #include <linux/slab.h>
27 #include <linux/pci.h>
28 #include <linux/interrupt.h>
29 #include <linux/dmaengine.h>
30 #include <linux/delay.h>
31 #include <linux/dma-mapping.h>
32 #include <linux/workqueue.h>
33 #include <linux/prefetch.h>
35 #include "registers.h"
38 #include "../dmaengine.h"
40 static void ioat_eh(struct ioatdma_chan
*ioat_chan
);
43 * ioat_dma_do_interrupt - handler used for single vector interrupt mode
45 * @data: interrupt data
47 irqreturn_t
ioat_dma_do_interrupt(int irq
, void *data
)
49 struct ioatdma_device
*instance
= data
;
50 struct ioatdma_chan
*ioat_chan
;
51 unsigned long attnstatus
;
55 intrctrl
= readb(instance
->reg_base
+ IOAT_INTRCTRL_OFFSET
);
57 if (!(intrctrl
& IOAT_INTRCTRL_MASTER_INT_EN
))
60 if (!(intrctrl
& IOAT_INTRCTRL_INT_STATUS
)) {
61 writeb(intrctrl
, instance
->reg_base
+ IOAT_INTRCTRL_OFFSET
);
65 attnstatus
= readl(instance
->reg_base
+ IOAT_ATTNSTATUS_OFFSET
);
66 for_each_set_bit(bit
, &attnstatus
, BITS_PER_LONG
) {
67 ioat_chan
= ioat_chan_by_index(instance
, bit
);
68 if (test_bit(IOAT_RUN
, &ioat_chan
->state
))
69 tasklet_schedule(&ioat_chan
->cleanup_task
);
72 writeb(intrctrl
, instance
->reg_base
+ IOAT_INTRCTRL_OFFSET
);
77 * ioat_dma_do_interrupt_msix - handler used for vector-per-channel interrupt mode
79 * @data: interrupt data
81 irqreturn_t
ioat_dma_do_interrupt_msix(int irq
, void *data
)
83 struct ioatdma_chan
*ioat_chan
= data
;
85 if (test_bit(IOAT_RUN
, &ioat_chan
->state
))
86 tasklet_schedule(&ioat_chan
->cleanup_task
);
91 void ioat_stop(struct ioatdma_chan
*ioat_chan
)
93 struct ioatdma_device
*ioat_dma
= ioat_chan
->ioat_dma
;
94 struct pci_dev
*pdev
= ioat_dma
->pdev
;
95 int chan_id
= chan_num(ioat_chan
);
96 struct msix_entry
*msix
;
98 /* 1/ stop irq from firing tasklets
99 * 2/ stop the tasklet from re-arming irqs
101 clear_bit(IOAT_RUN
, &ioat_chan
->state
);
103 /* flush inflight interrupts */
104 switch (ioat_dma
->irq_mode
) {
106 msix
= &ioat_dma
->msix_entries
[chan_id
];
107 synchronize_irq(msix
->vector
);
111 synchronize_irq(pdev
->irq
);
117 /* flush inflight timers */
118 del_timer_sync(&ioat_chan
->timer
);
120 /* flush inflight tasklet runs */
121 tasklet_kill(&ioat_chan
->cleanup_task
);
123 /* final cleanup now that everything is quiesced and can't re-arm */
124 ioat_cleanup_event((unsigned long)&ioat_chan
->dma_chan
);
127 static void __ioat_issue_pending(struct ioatdma_chan
*ioat_chan
)
129 ioat_chan
->dmacount
+= ioat_ring_pending(ioat_chan
);
130 ioat_chan
->issued
= ioat_chan
->head
;
131 writew(ioat_chan
->dmacount
,
132 ioat_chan
->reg_base
+ IOAT_CHAN_DMACOUNT_OFFSET
);
133 dev_dbg(to_dev(ioat_chan
),
134 "%s: head: %#x tail: %#x issued: %#x count: %#x\n",
135 __func__
, ioat_chan
->head
, ioat_chan
->tail
,
136 ioat_chan
->issued
, ioat_chan
->dmacount
);
139 void ioat_issue_pending(struct dma_chan
*c
)
141 struct ioatdma_chan
*ioat_chan
= to_ioat_chan(c
);
143 if (ioat_ring_pending(ioat_chan
)) {
144 spin_lock_bh(&ioat_chan
->prep_lock
);
145 __ioat_issue_pending(ioat_chan
);
146 spin_unlock_bh(&ioat_chan
->prep_lock
);
151 * ioat_update_pending - log pending descriptors
152 * @ioat: ioat+ channel
154 * Check if the number of unsubmitted descriptors has exceeded the
155 * watermark. Called with prep_lock held
157 static void ioat_update_pending(struct ioatdma_chan
*ioat_chan
)
159 if (ioat_ring_pending(ioat_chan
) > ioat_pending_level
)
160 __ioat_issue_pending(ioat_chan
);
163 static void __ioat_start_null_desc(struct ioatdma_chan
*ioat_chan
)
165 struct ioat_ring_ent
*desc
;
166 struct ioat_dma_descriptor
*hw
;
168 if (ioat_ring_space(ioat_chan
) < 1) {
169 dev_err(to_dev(ioat_chan
),
170 "Unable to start null desc - ring full\n");
174 dev_dbg(to_dev(ioat_chan
),
175 "%s: head: %#x tail: %#x issued: %#x\n",
176 __func__
, ioat_chan
->head
, ioat_chan
->tail
, ioat_chan
->issued
);
177 desc
= ioat_get_ring_ent(ioat_chan
, ioat_chan
->head
);
182 hw
->ctl_f
.int_en
= 1;
183 hw
->ctl_f
.compl_write
= 1;
184 /* set size to non-zero value (channel returns error when size is 0) */
185 hw
->size
= NULL_DESC_BUFFER_SIZE
;
188 async_tx_ack(&desc
->txd
);
189 ioat_set_chainaddr(ioat_chan
, desc
->txd
.phys
);
190 dump_desc_dbg(ioat_chan
, desc
);
191 /* make sure descriptors are written before we submit */
193 ioat_chan
->head
+= 1;
194 __ioat_issue_pending(ioat_chan
);
197 void ioat_start_null_desc(struct ioatdma_chan
*ioat_chan
)
199 spin_lock_bh(&ioat_chan
->prep_lock
);
200 __ioat_start_null_desc(ioat_chan
);
201 spin_unlock_bh(&ioat_chan
->prep_lock
);
204 static void __ioat_restart_chan(struct ioatdma_chan
*ioat_chan
)
206 /* set the tail to be re-issued */
207 ioat_chan
->issued
= ioat_chan
->tail
;
208 ioat_chan
->dmacount
= 0;
209 mod_timer(&ioat_chan
->timer
, jiffies
+ COMPLETION_TIMEOUT
);
211 dev_dbg(to_dev(ioat_chan
),
212 "%s: head: %#x tail: %#x issued: %#x count: %#x\n",
213 __func__
, ioat_chan
->head
, ioat_chan
->tail
,
214 ioat_chan
->issued
, ioat_chan
->dmacount
);
216 if (ioat_ring_pending(ioat_chan
)) {
217 struct ioat_ring_ent
*desc
;
219 desc
= ioat_get_ring_ent(ioat_chan
, ioat_chan
->tail
);
220 ioat_set_chainaddr(ioat_chan
, desc
->txd
.phys
);
221 __ioat_issue_pending(ioat_chan
);
223 __ioat_start_null_desc(ioat_chan
);
226 static int ioat_quiesce(struct ioatdma_chan
*ioat_chan
, unsigned long tmo
)
228 unsigned long end
= jiffies
+ tmo
;
232 status
= ioat_chansts(ioat_chan
);
233 if (is_ioat_active(status
) || is_ioat_idle(status
))
234 ioat_suspend(ioat_chan
);
235 while (is_ioat_active(status
) || is_ioat_idle(status
)) {
236 if (tmo
&& time_after(jiffies
, end
)) {
240 status
= ioat_chansts(ioat_chan
);
247 static int ioat_reset_sync(struct ioatdma_chan
*ioat_chan
, unsigned long tmo
)
249 unsigned long end
= jiffies
+ tmo
;
252 ioat_reset(ioat_chan
);
253 while (ioat_reset_pending(ioat_chan
)) {
254 if (end
&& time_after(jiffies
, end
)) {
264 static dma_cookie_t
ioat_tx_submit_unlock(struct dma_async_tx_descriptor
*tx
)
265 __releases(&ioat_chan
->prep_lock
)
267 struct dma_chan
*c
= tx
->chan
;
268 struct ioatdma_chan
*ioat_chan
= to_ioat_chan(c
);
271 cookie
= dma_cookie_assign(tx
);
272 dev_dbg(to_dev(ioat_chan
), "%s: cookie: %d\n", __func__
, cookie
);
274 if (!test_and_set_bit(IOAT_CHAN_ACTIVE
, &ioat_chan
->state
))
275 mod_timer(&ioat_chan
->timer
, jiffies
+ COMPLETION_TIMEOUT
);
277 /* make descriptor updates visible before advancing ioat->head,
278 * this is purposefully not smp_wmb() since we are also
279 * publishing the descriptor updates to a dma device
283 ioat_chan
->head
+= ioat_chan
->produce
;
285 ioat_update_pending(ioat_chan
);
286 spin_unlock_bh(&ioat_chan
->prep_lock
);
291 static struct ioat_ring_ent
*
292 ioat_alloc_ring_ent(struct dma_chan
*chan
, gfp_t flags
)
294 struct ioat_dma_descriptor
*hw
;
295 struct ioat_ring_ent
*desc
;
296 struct ioatdma_device
*ioat_dma
;
299 ioat_dma
= to_ioatdma_device(chan
->device
);
300 hw
= pci_pool_alloc(ioat_dma
->dma_pool
, flags
, &phys
);
303 memset(hw
, 0, sizeof(*hw
));
305 desc
= kmem_cache_zalloc(ioat_cache
, flags
);
307 pci_pool_free(ioat_dma
->dma_pool
, hw
, phys
);
311 dma_async_tx_descriptor_init(&desc
->txd
, chan
);
312 desc
->txd
.tx_submit
= ioat_tx_submit_unlock
;
314 desc
->txd
.phys
= phys
;
318 void ioat_free_ring_ent(struct ioat_ring_ent
*desc
, struct dma_chan
*chan
)
320 struct ioatdma_device
*ioat_dma
;
322 ioat_dma
= to_ioatdma_device(chan
->device
);
323 pci_pool_free(ioat_dma
->dma_pool
, desc
->hw
, desc
->txd
.phys
);
324 kmem_cache_free(ioat_cache
, desc
);
327 struct ioat_ring_ent
**
328 ioat_alloc_ring(struct dma_chan
*c
, int order
, gfp_t flags
)
330 struct ioat_ring_ent
**ring
;
331 int descs
= 1 << order
;
334 if (order
> ioat_get_max_alloc_order())
337 /* allocate the array to hold the software ring */
338 ring
= kcalloc(descs
, sizeof(*ring
), flags
);
341 for (i
= 0; i
< descs
; i
++) {
342 ring
[i
] = ioat_alloc_ring_ent(c
, flags
);
345 ioat_free_ring_ent(ring
[i
], c
);
349 set_desc_id(ring
[i
], i
);
353 for (i
= 0; i
< descs
-1; i
++) {
354 struct ioat_ring_ent
*next
= ring
[i
+1];
355 struct ioat_dma_descriptor
*hw
= ring
[i
]->hw
;
357 hw
->next
= next
->txd
.phys
;
359 ring
[i
]->hw
->next
= ring
[0]->txd
.phys
;
364 static bool reshape_ring(struct ioatdma_chan
*ioat_chan
, int order
)
366 /* reshape differs from normal ring allocation in that we want
367 * to allocate a new software ring while only
368 * extending/truncating the hardware ring
370 struct dma_chan
*c
= &ioat_chan
->dma_chan
;
371 const u32 curr_size
= ioat_ring_size(ioat_chan
);
372 const u16 active
= ioat_ring_active(ioat_chan
);
373 const u32 new_size
= 1 << order
;
374 struct ioat_ring_ent
**ring
;
377 if (order
> ioat_get_max_alloc_order())
380 /* double check that we have at least 1 free descriptor */
381 if (active
== curr_size
)
384 /* when shrinking, verify that we can hold the current active
385 * set in the new ring
387 if (active
>= new_size
)
390 /* allocate the array to hold the software ring */
391 ring
= kcalloc(new_size
, sizeof(*ring
), GFP_NOWAIT
);
395 /* allocate/trim descriptors as needed */
396 if (new_size
> curr_size
) {
397 /* copy current descriptors to the new ring */
398 for (i
= 0; i
< curr_size
; i
++) {
399 u16 curr_idx
= (ioat_chan
->tail
+i
) & (curr_size
-1);
400 u16 new_idx
= (ioat_chan
->tail
+i
) & (new_size
-1);
402 ring
[new_idx
] = ioat_chan
->ring
[curr_idx
];
403 set_desc_id(ring
[new_idx
], new_idx
);
406 /* add new descriptors to the ring */
407 for (i
= curr_size
; i
< new_size
; i
++) {
408 u16 new_idx
= (ioat_chan
->tail
+i
) & (new_size
-1);
410 ring
[new_idx
] = ioat_alloc_ring_ent(c
, GFP_NOWAIT
);
411 if (!ring
[new_idx
]) {
413 u16 new_idx
= (ioat_chan
->tail
+i
) &
416 ioat_free_ring_ent(ring
[new_idx
], c
);
421 set_desc_id(ring
[new_idx
], new_idx
);
424 /* hw link new descriptors */
425 for (i
= curr_size
-1; i
< new_size
; i
++) {
426 u16 new_idx
= (ioat_chan
->tail
+i
) & (new_size
-1);
427 struct ioat_ring_ent
*next
=
428 ring
[(new_idx
+1) & (new_size
-1)];
429 struct ioat_dma_descriptor
*hw
= ring
[new_idx
]->hw
;
431 hw
->next
= next
->txd
.phys
;
434 struct ioat_dma_descriptor
*hw
;
435 struct ioat_ring_ent
*next
;
437 /* copy current descriptors to the new ring, dropping the
438 * removed descriptors
440 for (i
= 0; i
< new_size
; i
++) {
441 u16 curr_idx
= (ioat_chan
->tail
+i
) & (curr_size
-1);
442 u16 new_idx
= (ioat_chan
->tail
+i
) & (new_size
-1);
444 ring
[new_idx
] = ioat_chan
->ring
[curr_idx
];
445 set_desc_id(ring
[new_idx
], new_idx
);
448 /* free deleted descriptors */
449 for (i
= new_size
; i
< curr_size
; i
++) {
450 struct ioat_ring_ent
*ent
;
452 ent
= ioat_get_ring_ent(ioat_chan
, ioat_chan
->tail
+i
);
453 ioat_free_ring_ent(ent
, c
);
456 /* fix up hardware ring */
457 hw
= ring
[(ioat_chan
->tail
+new_size
-1) & (new_size
-1)]->hw
;
458 next
= ring
[(ioat_chan
->tail
+new_size
) & (new_size
-1)];
459 hw
->next
= next
->txd
.phys
;
462 dev_dbg(to_dev(ioat_chan
), "%s: allocated %d descriptors\n",
465 kfree(ioat_chan
->ring
);
466 ioat_chan
->ring
= ring
;
467 ioat_chan
->alloc_order
= order
;
473 * ioat_check_space_lock - verify space and grab ring producer lock
474 * @ioat: ioat,3 channel (ring) to operate on
475 * @num_descs: allocation length
477 int ioat_check_space_lock(struct ioatdma_chan
*ioat_chan
, int num_descs
)
478 __acquires(&ioat_chan
->prep_lock
)
483 spin_lock_bh(&ioat_chan
->prep_lock
);
484 /* never allow the last descriptor to be consumed, we need at
485 * least one free at all times to allow for on-the-fly ring
488 if (likely(ioat_ring_space(ioat_chan
) > num_descs
)) {
489 dev_dbg(to_dev(ioat_chan
), "%s: num_descs: %d (%x:%x:%x)\n",
490 __func__
, num_descs
, ioat_chan
->head
,
491 ioat_chan
->tail
, ioat_chan
->issued
);
492 ioat_chan
->produce
= num_descs
;
493 return 0; /* with ioat->prep_lock held */
495 retry
= test_and_set_bit(IOAT_RESHAPE_PENDING
, &ioat_chan
->state
);
496 spin_unlock_bh(&ioat_chan
->prep_lock
);
498 /* is another cpu already trying to expand the ring? */
502 spin_lock_bh(&ioat_chan
->cleanup_lock
);
503 spin_lock_bh(&ioat_chan
->prep_lock
);
504 retry
= reshape_ring(ioat_chan
, ioat_chan
->alloc_order
+ 1);
505 clear_bit(IOAT_RESHAPE_PENDING
, &ioat_chan
->state
);
506 spin_unlock_bh(&ioat_chan
->prep_lock
);
507 spin_unlock_bh(&ioat_chan
->cleanup_lock
);
509 /* if we were able to expand the ring retry the allocation */
513 dev_dbg_ratelimited(to_dev(ioat_chan
),
514 "%s: ring full! num_descs: %d (%x:%x:%x)\n",
515 __func__
, num_descs
, ioat_chan
->head
,
516 ioat_chan
->tail
, ioat_chan
->issued
);
518 /* progress reclaim in the allocation failure case we may be
519 * called under bh_disabled so we need to trigger the timer
522 if (time_is_before_jiffies(ioat_chan
->timer
.expires
)
523 && timer_pending(&ioat_chan
->timer
)) {
524 mod_timer(&ioat_chan
->timer
, jiffies
+ COMPLETION_TIMEOUT
);
525 ioat_timer_event((unsigned long)ioat_chan
);
531 static bool desc_has_ext(struct ioat_ring_ent
*desc
)
533 struct ioat_dma_descriptor
*hw
= desc
->hw
;
535 if (hw
->ctl_f
.op
== IOAT_OP_XOR
||
536 hw
->ctl_f
.op
== IOAT_OP_XOR_VAL
) {
537 struct ioat_xor_descriptor
*xor = desc
->xor;
539 if (src_cnt_to_sw(xor->ctl_f
.src_cnt
) > 5)
541 } else if (hw
->ctl_f
.op
== IOAT_OP_PQ
||
542 hw
->ctl_f
.op
== IOAT_OP_PQ_VAL
) {
543 struct ioat_pq_descriptor
*pq
= desc
->pq
;
545 if (src_cnt_to_sw(pq
->ctl_f
.src_cnt
) > 3)
553 ioat_free_sed(struct ioatdma_device
*ioat_dma
, struct ioat_sed_ent
*sed
)
558 dma_pool_free(ioat_dma
->sed_hw_pool
[sed
->hw_pool
], sed
->hw
, sed
->dma
);
559 kmem_cache_free(ioat_sed_cache
, sed
);
562 static u64
ioat_get_current_completion(struct ioatdma_chan
*ioat_chan
)
567 completion
= *ioat_chan
->completion
;
568 phys_complete
= ioat_chansts_to_addr(completion
);
570 dev_dbg(to_dev(ioat_chan
), "%s: phys_complete: %#llx\n", __func__
,
571 (unsigned long long) phys_complete
);
573 return phys_complete
;
576 static bool ioat_cleanup_preamble(struct ioatdma_chan
*ioat_chan
,
579 *phys_complete
= ioat_get_current_completion(ioat_chan
);
580 if (*phys_complete
== ioat_chan
->last_completion
)
583 clear_bit(IOAT_COMPLETION_ACK
, &ioat_chan
->state
);
584 mod_timer(&ioat_chan
->timer
, jiffies
+ COMPLETION_TIMEOUT
);
590 desc_get_errstat(struct ioatdma_chan
*ioat_chan
, struct ioat_ring_ent
*desc
)
592 struct ioat_dma_descriptor
*hw
= desc
->hw
;
594 switch (hw
->ctl_f
.op
) {
596 case IOAT_OP_PQ_VAL_16S
:
598 struct ioat_pq_descriptor
*pq
= desc
->pq
;
600 /* check if there's error written */
601 if (!pq
->dwbes_f
.wbes
)
604 /* need to set a chanerr var for checking to clear later */
606 if (pq
->dwbes_f
.p_val_err
)
607 *desc
->result
|= SUM_CHECK_P_RESULT
;
609 if (pq
->dwbes_f
.q_val_err
)
610 *desc
->result
|= SUM_CHECK_Q_RESULT
;
620 * __cleanup - reclaim used descriptors
621 * @ioat: channel (ring) to clean
623 static void __cleanup(struct ioatdma_chan
*ioat_chan
, dma_addr_t phys_complete
)
625 struct ioatdma_device
*ioat_dma
= ioat_chan
->ioat_dma
;
626 struct ioat_ring_ent
*desc
;
627 bool seen_current
= false;
628 int idx
= ioat_chan
->tail
, i
;
631 dev_dbg(to_dev(ioat_chan
), "%s: head: %#x tail: %#x issued: %#x\n",
632 __func__
, ioat_chan
->head
, ioat_chan
->tail
, ioat_chan
->issued
);
635 * At restart of the channel, the completion address and the
636 * channel status will be 0 due to starting a new chain. Since
637 * it's new chain and the first descriptor "fails", there is
638 * nothing to clean up. We do not want to reap the entire submitted
639 * chain due to this 0 address value and then BUG.
644 active
= ioat_ring_active(ioat_chan
);
645 for (i
= 0; i
< active
&& !seen_current
; i
++) {
646 struct dma_async_tx_descriptor
*tx
;
648 smp_read_barrier_depends();
649 prefetch(ioat_get_ring_ent(ioat_chan
, idx
+ i
+ 1));
650 desc
= ioat_get_ring_ent(ioat_chan
, idx
+ i
);
651 dump_desc_dbg(ioat_chan
, desc
);
653 /* set err stat if we are using dwbes */
654 if (ioat_dma
->cap
& IOAT_CAP_DWBES
)
655 desc_get_errstat(ioat_chan
, desc
);
659 dma_cookie_complete(tx
);
660 dma_descriptor_unmap(tx
);
662 tx
->callback(tx
->callback_param
);
667 if (tx
->phys
== phys_complete
)
670 /* skip extended descriptors */
671 if (desc_has_ext(desc
)) {
672 BUG_ON(i
+ 1 >= active
);
676 /* cleanup super extended descriptors */
678 ioat_free_sed(ioat_dma
, desc
->sed
);
683 /* finish all descriptor reads before incrementing tail */
685 ioat_chan
->tail
= idx
+ i
;
686 /* no active descs have written a completion? */
687 BUG_ON(active
&& !seen_current
);
688 ioat_chan
->last_completion
= phys_complete
;
690 if (active
- i
== 0) {
691 dev_dbg(to_dev(ioat_chan
), "%s: cancel completion timeout\n",
693 mod_timer(&ioat_chan
->timer
, jiffies
+ IDLE_TIMEOUT
);
696 /* 5 microsecond delay per pending descriptor */
697 writew(min((5 * (active
- i
)), IOAT_INTRDELAY_MASK
),
698 ioat_chan
->ioat_dma
->reg_base
+ IOAT_INTRDELAY_OFFSET
);
701 static void ioat_cleanup(struct ioatdma_chan
*ioat_chan
)
705 spin_lock_bh(&ioat_chan
->cleanup_lock
);
707 if (ioat_cleanup_preamble(ioat_chan
, &phys_complete
))
708 __cleanup(ioat_chan
, phys_complete
);
710 if (is_ioat_halted(*ioat_chan
->completion
)) {
711 u32 chanerr
= readl(ioat_chan
->reg_base
+ IOAT_CHANERR_OFFSET
);
713 if (chanerr
& IOAT_CHANERR_HANDLE_MASK
) {
714 mod_timer(&ioat_chan
->timer
, jiffies
+ IDLE_TIMEOUT
);
719 spin_unlock_bh(&ioat_chan
->cleanup_lock
);
722 void ioat_cleanup_event(unsigned long data
)
724 struct ioatdma_chan
*ioat_chan
= to_ioat_chan((void *)data
);
726 ioat_cleanup(ioat_chan
);
727 if (!test_bit(IOAT_RUN
, &ioat_chan
->state
))
729 writew(IOAT_CHANCTRL_RUN
, ioat_chan
->reg_base
+ IOAT_CHANCTRL_OFFSET
);
732 static void ioat_restart_channel(struct ioatdma_chan
*ioat_chan
)
736 ioat_quiesce(ioat_chan
, 0);
737 if (ioat_cleanup_preamble(ioat_chan
, &phys_complete
))
738 __cleanup(ioat_chan
, phys_complete
);
740 __ioat_restart_chan(ioat_chan
);
743 static void ioat_eh(struct ioatdma_chan
*ioat_chan
)
745 struct pci_dev
*pdev
= to_pdev(ioat_chan
);
746 struct ioat_dma_descriptor
*hw
;
747 struct dma_async_tx_descriptor
*tx
;
749 struct ioat_ring_ent
*desc
;
754 /* cleanup so tail points to descriptor that caused the error */
755 if (ioat_cleanup_preamble(ioat_chan
, &phys_complete
))
756 __cleanup(ioat_chan
, phys_complete
);
758 chanerr
= readl(ioat_chan
->reg_base
+ IOAT_CHANERR_OFFSET
);
759 pci_read_config_dword(pdev
, IOAT_PCI_CHANERR_INT_OFFSET
, &chanerr_int
);
761 dev_dbg(to_dev(ioat_chan
), "%s: error = %x:%x\n",
762 __func__
, chanerr
, chanerr_int
);
764 desc
= ioat_get_ring_ent(ioat_chan
, ioat_chan
->tail
);
766 dump_desc_dbg(ioat_chan
, desc
);
768 switch (hw
->ctl_f
.op
) {
769 case IOAT_OP_XOR_VAL
:
770 if (chanerr
& IOAT_CHANERR_XOR_P_OR_CRC_ERR
) {
771 *desc
->result
|= SUM_CHECK_P_RESULT
;
772 err_handled
|= IOAT_CHANERR_XOR_P_OR_CRC_ERR
;
776 case IOAT_OP_PQ_VAL_16S
:
777 if (chanerr
& IOAT_CHANERR_XOR_P_OR_CRC_ERR
) {
778 *desc
->result
|= SUM_CHECK_P_RESULT
;
779 err_handled
|= IOAT_CHANERR_XOR_P_OR_CRC_ERR
;
781 if (chanerr
& IOAT_CHANERR_XOR_Q_ERR
) {
782 *desc
->result
|= SUM_CHECK_Q_RESULT
;
783 err_handled
|= IOAT_CHANERR_XOR_Q_ERR
;
788 /* fault on unhandled error or spurious halt */
789 if (chanerr
^ err_handled
|| chanerr
== 0) {
790 dev_err(to_dev(ioat_chan
), "%s: fatal error (%x:%x)\n",
791 __func__
, chanerr
, err_handled
);
793 } else { /* cleanup the faulty descriptor */
796 dma_cookie_complete(tx
);
797 dma_descriptor_unmap(tx
);
799 tx
->callback(tx
->callback_param
);
805 writel(chanerr
, ioat_chan
->reg_base
+ IOAT_CHANERR_OFFSET
);
806 pci_write_config_dword(pdev
, IOAT_PCI_CHANERR_INT_OFFSET
, chanerr_int
);
808 /* mark faulting descriptor as complete */
809 *ioat_chan
->completion
= desc
->txd
.phys
;
811 spin_lock_bh(&ioat_chan
->prep_lock
);
812 ioat_restart_channel(ioat_chan
);
813 spin_unlock_bh(&ioat_chan
->prep_lock
);
816 static void check_active(struct ioatdma_chan
*ioat_chan
)
818 if (ioat_ring_active(ioat_chan
)) {
819 mod_timer(&ioat_chan
->timer
, jiffies
+ COMPLETION_TIMEOUT
);
823 if (test_and_clear_bit(IOAT_CHAN_ACTIVE
, &ioat_chan
->state
))
824 mod_timer(&ioat_chan
->timer
, jiffies
+ IDLE_TIMEOUT
);
825 else if (ioat_chan
->alloc_order
> ioat_get_alloc_order()) {
826 /* if the ring is idle, empty, and oversized try to step
829 reshape_ring(ioat_chan
, ioat_chan
->alloc_order
- 1);
831 /* keep shrinking until we get back to our minimum
834 if (ioat_chan
->alloc_order
> ioat_get_alloc_order())
835 mod_timer(&ioat_chan
->timer
, jiffies
+ IDLE_TIMEOUT
);
840 void ioat_timer_event(unsigned long data
)
842 struct ioatdma_chan
*ioat_chan
= to_ioat_chan((void *)data
);
843 dma_addr_t phys_complete
;
846 status
= ioat_chansts(ioat_chan
);
848 /* when halted due to errors check for channel
849 * programming errors before advancing the completion state
851 if (is_ioat_halted(status
)) {
854 chanerr
= readl(ioat_chan
->reg_base
+ IOAT_CHANERR_OFFSET
);
855 dev_err(to_dev(ioat_chan
), "%s: Channel halted (%x)\n",
857 if (test_bit(IOAT_RUN
, &ioat_chan
->state
))
858 BUG_ON(is_ioat_bug(chanerr
));
859 else /* we never got off the ground */
863 /* if we haven't made progress and we have already
864 * acknowledged a pending completion once, then be more
865 * forceful with a restart
867 spin_lock_bh(&ioat_chan
->cleanup_lock
);
868 if (ioat_cleanup_preamble(ioat_chan
, &phys_complete
))
869 __cleanup(ioat_chan
, phys_complete
);
870 else if (test_bit(IOAT_COMPLETION_ACK
, &ioat_chan
->state
)) {
871 spin_lock_bh(&ioat_chan
->prep_lock
);
872 ioat_restart_channel(ioat_chan
);
873 spin_unlock_bh(&ioat_chan
->prep_lock
);
874 spin_unlock_bh(&ioat_chan
->cleanup_lock
);
877 set_bit(IOAT_COMPLETION_ACK
, &ioat_chan
->state
);
878 mod_timer(&ioat_chan
->timer
, jiffies
+ COMPLETION_TIMEOUT
);
882 if (ioat_ring_active(ioat_chan
))
883 mod_timer(&ioat_chan
->timer
, jiffies
+ COMPLETION_TIMEOUT
);
885 spin_lock_bh(&ioat_chan
->prep_lock
);
886 check_active(ioat_chan
);
887 spin_unlock_bh(&ioat_chan
->prep_lock
);
889 spin_unlock_bh(&ioat_chan
->cleanup_lock
);
893 ioat_tx_status(struct dma_chan
*c
, dma_cookie_t cookie
,
894 struct dma_tx_state
*txstate
)
896 struct ioatdma_chan
*ioat_chan
= to_ioat_chan(c
);
899 ret
= dma_cookie_status(c
, cookie
, txstate
);
900 if (ret
== DMA_COMPLETE
)
903 ioat_cleanup(ioat_chan
);
905 return dma_cookie_status(c
, cookie
, txstate
);
908 static int ioat_irq_reinit(struct ioatdma_device
*ioat_dma
)
910 struct pci_dev
*pdev
= ioat_dma
->pdev
;
911 int irq
= pdev
->irq
, i
;
913 if (!is_bwd_ioat(pdev
))
916 switch (ioat_dma
->irq_mode
) {
918 for (i
= 0; i
< ioat_dma
->dma_dev
.chancnt
; i
++) {
919 struct msix_entry
*msix
= &ioat_dma
->msix_entries
[i
];
920 struct ioatdma_chan
*ioat_chan
;
922 ioat_chan
= ioat_chan_by_index(ioat_dma
, i
);
923 devm_free_irq(&pdev
->dev
, msix
->vector
, ioat_chan
);
926 pci_disable_msix(pdev
);
929 pci_disable_msi(pdev
);
932 devm_free_irq(&pdev
->dev
, irq
, ioat_dma
);
937 ioat_dma
->irq_mode
= IOAT_NOIRQ
;
939 return ioat_dma_setup_interrupts(ioat_dma
);
942 int ioat_reset_hw(struct ioatdma_chan
*ioat_chan
)
944 /* throw away whatever the channel was doing and get it
945 * initialized, with ioat3 specific workarounds
947 struct ioatdma_device
*ioat_dma
= ioat_chan
->ioat_dma
;
948 struct pci_dev
*pdev
= ioat_dma
->pdev
;
953 ioat_quiesce(ioat_chan
, msecs_to_jiffies(100));
955 chanerr
= readl(ioat_chan
->reg_base
+ IOAT_CHANERR_OFFSET
);
956 writel(chanerr
, ioat_chan
->reg_base
+ IOAT_CHANERR_OFFSET
);
958 if (ioat_dma
->version
< IOAT_VER_3_3
) {
959 /* clear any pending errors */
960 err
= pci_read_config_dword(pdev
,
961 IOAT_PCI_CHANERR_INT_OFFSET
, &chanerr
);
964 "channel error register unreachable\n");
967 pci_write_config_dword(pdev
,
968 IOAT_PCI_CHANERR_INT_OFFSET
, chanerr
);
970 /* Clear DMAUNCERRSTS Cfg-Reg Parity Error status bit
971 * (workaround for spurious config parity error after restart)
973 pci_read_config_word(pdev
, IOAT_PCI_DEVICE_ID_OFFSET
, &dev_id
);
974 if (dev_id
== PCI_DEVICE_ID_INTEL_IOAT_TBG0
) {
975 pci_write_config_dword(pdev
,
976 IOAT_PCI_DMAUNCERRSTS_OFFSET
,
981 err
= ioat_reset_sync(ioat_chan
, msecs_to_jiffies(200));
983 err
= ioat_irq_reinit(ioat_dma
);
986 dev_err(&pdev
->dev
, "Failed to reset: %d\n", err
);