2 * Intel I/OAT DMA Linux driver
3 * Copyright(c) 2004 - 2007 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 * You should have received a copy of the GNU General Public License along with
15 * this program; if not, write to the Free Software Foundation, Inc.,
16 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
18 * The full GNU General Public License is included in this distribution in
19 * the file called "COPYING".
24 * This driver supports an Intel I/OAT DMA engine, which does asynchronous
28 #include <linux/init.h>
29 #include <linux/module.h>
30 #include <linux/pci.h>
31 #include <linux/interrupt.h>
32 #include <linux/dmaengine.h>
33 #include <linux/delay.h>
34 #include <linux/dma-mapping.h>
36 #include "ioatdma_registers.h"
37 #include "ioatdma_hw.h"
39 #define to_ioat_chan(chan) container_of(chan, struct ioat_dma_chan, common)
40 #define to_ioatdma_device(dev) container_of(dev, struct ioatdma_device, common)
41 #define to_ioat_desc(lh) container_of(lh, struct ioat_desc_sw, node)
42 #define tx_to_ioat_desc(tx) container_of(tx, struct ioat_desc_sw, async_tx)
44 static int ioat_pending_level
= 4;
45 module_param(ioat_pending_level
, int, 0644);
46 MODULE_PARM_DESC(ioat_pending_level
,
47 "high-water mark for pushing ioat descriptors (default: 4)");
49 /* internal functions */
50 static void ioat_dma_start_null_desc(struct ioat_dma_chan
*ioat_chan
);
51 static void ioat_dma_memcpy_cleanup(struct ioat_dma_chan
*ioat_chan
);
53 static struct ioat_desc_sw
*
54 ioat1_dma_get_next_descriptor(struct ioat_dma_chan
*ioat_chan
);
55 static struct ioat_desc_sw
*
56 ioat2_dma_get_next_descriptor(struct ioat_dma_chan
*ioat_chan
);
58 static inline struct ioat_dma_chan
*ioat_lookup_chan_by_index(
59 struct ioatdma_device
*device
,
62 return device
->idx
[index
];
66 * ioat_dma_do_interrupt - handler used for single vector interrupt mode
68 * @data: interrupt data
70 static irqreturn_t
ioat_dma_do_interrupt(int irq
, void *data
)
72 struct ioatdma_device
*instance
= data
;
73 struct ioat_dma_chan
*ioat_chan
;
74 unsigned long attnstatus
;
78 intrctrl
= readb(instance
->reg_base
+ IOAT_INTRCTRL_OFFSET
);
80 if (!(intrctrl
& IOAT_INTRCTRL_MASTER_INT_EN
))
83 if (!(intrctrl
& IOAT_INTRCTRL_INT_STATUS
)) {
84 writeb(intrctrl
, instance
->reg_base
+ IOAT_INTRCTRL_OFFSET
);
88 attnstatus
= readl(instance
->reg_base
+ IOAT_ATTNSTATUS_OFFSET
);
89 for_each_bit(bit
, &attnstatus
, BITS_PER_LONG
) {
90 ioat_chan
= ioat_lookup_chan_by_index(instance
, bit
);
91 tasklet_schedule(&ioat_chan
->cleanup_task
);
94 writeb(intrctrl
, instance
->reg_base
+ IOAT_INTRCTRL_OFFSET
);
99 * ioat_dma_do_interrupt_msix - handler used for vector-per-channel interrupt mode
101 * @data: interrupt data
103 static irqreturn_t
ioat_dma_do_interrupt_msix(int irq
, void *data
)
105 struct ioat_dma_chan
*ioat_chan
= data
;
107 tasklet_schedule(&ioat_chan
->cleanup_task
);
112 static void ioat_dma_cleanup_tasklet(unsigned long data
);
115 * ioat_dma_enumerate_channels - find and initialize the device's channels
116 * @device: the device to be enumerated
118 static int ioat_dma_enumerate_channels(struct ioatdma_device
*device
)
123 struct ioat_dma_chan
*ioat_chan
;
125 device
->common
.chancnt
= readb(device
->reg_base
+ IOAT_CHANCNT_OFFSET
);
126 xfercap_scale
= readb(device
->reg_base
+ IOAT_XFERCAP_OFFSET
);
127 xfercap
= (xfercap_scale
== 0 ? -1 : (1UL << xfercap_scale
));
129 for (i
= 0; i
< device
->common
.chancnt
; i
++) {
130 ioat_chan
= kzalloc(sizeof(*ioat_chan
), GFP_KERNEL
);
132 device
->common
.chancnt
= i
;
136 ioat_chan
->device
= device
;
137 ioat_chan
->reg_base
= device
->reg_base
+ (0x80 * (i
+ 1));
138 ioat_chan
->xfercap
= xfercap
;
139 ioat_chan
->desccount
= 0;
140 if (ioat_chan
->device
->version
!= IOAT_VER_1_2
) {
141 writel(IOAT_DCACTRL_CMPL_WRITE_ENABLE
142 | IOAT_DMA_DCA_ANY_CPU
,
143 ioat_chan
->reg_base
+ IOAT_DCACTRL_OFFSET
);
145 spin_lock_init(&ioat_chan
->cleanup_lock
);
146 spin_lock_init(&ioat_chan
->desc_lock
);
147 INIT_LIST_HEAD(&ioat_chan
->free_desc
);
148 INIT_LIST_HEAD(&ioat_chan
->used_desc
);
149 /* This should be made common somewhere in dmaengine.c */
150 ioat_chan
->common
.device
= &device
->common
;
151 list_add_tail(&ioat_chan
->common
.device_node
,
152 &device
->common
.channels
);
153 device
->idx
[i
] = ioat_chan
;
154 tasklet_init(&ioat_chan
->cleanup_task
,
155 ioat_dma_cleanup_tasklet
,
156 (unsigned long) ioat_chan
);
157 tasklet_disable(&ioat_chan
->cleanup_task
);
159 return device
->common
.chancnt
;
162 static void ioat_set_src(dma_addr_t addr
,
163 struct dma_async_tx_descriptor
*tx
,
166 tx_to_ioat_desc(tx
)->src
= addr
;
169 static void ioat_set_dest(dma_addr_t addr
,
170 struct dma_async_tx_descriptor
*tx
,
173 tx_to_ioat_desc(tx
)->dst
= addr
;
177 * ioat_dma_memcpy_issue_pending - push potentially unrecognized appended
179 * @chan: DMA channel handle
181 static inline void __ioat1_dma_memcpy_issue_pending(
182 struct ioat_dma_chan
*ioat_chan
)
184 ioat_chan
->pending
= 0;
185 writeb(IOAT_CHANCMD_APPEND
, ioat_chan
->reg_base
+ IOAT1_CHANCMD_OFFSET
);
188 static void ioat1_dma_memcpy_issue_pending(struct dma_chan
*chan
)
190 struct ioat_dma_chan
*ioat_chan
= to_ioat_chan(chan
);
192 if (ioat_chan
->pending
!= 0) {
193 spin_lock_bh(&ioat_chan
->desc_lock
);
194 __ioat1_dma_memcpy_issue_pending(ioat_chan
);
195 spin_unlock_bh(&ioat_chan
->desc_lock
);
199 static inline void __ioat2_dma_memcpy_issue_pending(
200 struct ioat_dma_chan
*ioat_chan
)
202 ioat_chan
->pending
= 0;
203 writew(ioat_chan
->dmacount
,
204 ioat_chan
->reg_base
+ IOAT_CHAN_DMACOUNT_OFFSET
);
207 static void ioat2_dma_memcpy_issue_pending(struct dma_chan
*chan
)
209 struct ioat_dma_chan
*ioat_chan
= to_ioat_chan(chan
);
211 if (ioat_chan
->pending
!= 0) {
212 spin_lock_bh(&ioat_chan
->desc_lock
);
213 __ioat2_dma_memcpy_issue_pending(ioat_chan
);
214 spin_unlock_bh(&ioat_chan
->desc_lock
);
218 static dma_cookie_t
ioat1_tx_submit(struct dma_async_tx_descriptor
*tx
)
220 struct ioat_dma_chan
*ioat_chan
= to_ioat_chan(tx
->chan
);
221 struct ioat_desc_sw
*first
= tx_to_ioat_desc(tx
);
222 struct ioat_desc_sw
*prev
, *new;
223 struct ioat_dma_descriptor
*hw
;
225 LIST_HEAD(new_chain
);
230 unsigned int desc_count
= 0;
232 /* src and dest and len are stored in the initial descriptor */
236 orig_ack
= first
->async_tx
.ack
;
239 spin_lock_bh(&ioat_chan
->desc_lock
);
240 prev
= to_ioat_desc(ioat_chan
->used_desc
.prev
);
243 copy
= min_t(size_t, len
, ioat_chan
->xfercap
);
245 new->async_tx
.ack
= 1;
254 /* chain together the physical address list for the HW */
256 prev
->hw
->next
= (u64
) new->async_tx
.phys
;
262 list_add_tail(&new->node
, &new_chain
);
265 } while (len
&& (new = ioat1_dma_get_next_descriptor(ioat_chan
)));
267 hw
->ctl
= IOAT_DMA_DESCRIPTOR_CTL_CP_STS
;
268 if (new->async_tx
.callback
) {
269 hw
->ctl
|= IOAT_DMA_DESCRIPTOR_CTL_INT_GN
;
271 /* move callback into to last desc */
272 new->async_tx
.callback
= first
->async_tx
.callback
;
273 new->async_tx
.callback_param
274 = first
->async_tx
.callback_param
;
275 first
->async_tx
.callback
= NULL
;
276 first
->async_tx
.callback_param
= NULL
;
280 new->tx_cnt
= desc_count
;
281 new->async_tx
.ack
= orig_ack
; /* client is in control of this ack */
283 /* store the original values for use in later cleanup */
285 new->src
= first
->src
;
286 new->dst
= first
->dst
;
287 new->len
= first
->len
;
290 /* cookie incr and addition to used_list must be atomic */
291 cookie
= ioat_chan
->common
.cookie
;
295 ioat_chan
->common
.cookie
= new->async_tx
.cookie
= cookie
;
297 /* write address into NextDescriptor field of last desc in chain */
298 to_ioat_desc(ioat_chan
->used_desc
.prev
)->hw
->next
=
299 first
->async_tx
.phys
;
300 __list_splice(&new_chain
, ioat_chan
->used_desc
.prev
);
302 ioat_chan
->dmacount
+= desc_count
;
303 ioat_chan
->pending
+= desc_count
;
304 if (ioat_chan
->pending
>= ioat_pending_level
)
305 __ioat1_dma_memcpy_issue_pending(ioat_chan
);
306 spin_unlock_bh(&ioat_chan
->desc_lock
);
311 static dma_cookie_t
ioat2_tx_submit(struct dma_async_tx_descriptor
*tx
)
313 struct ioat_dma_chan
*ioat_chan
= to_ioat_chan(tx
->chan
);
314 struct ioat_desc_sw
*first
= tx_to_ioat_desc(tx
);
315 struct ioat_desc_sw
*new;
316 struct ioat_dma_descriptor
*hw
;
322 unsigned int desc_count
= 0;
324 /* src and dest and len are stored in the initial descriptor */
328 orig_ack
= first
->async_tx
.ack
;
332 * ioat_chan->desc_lock is still in force in version 2 path
333 * it gets unlocked at end of this function
336 copy
= min_t(size_t, len
, ioat_chan
->xfercap
);
338 new->async_tx
.ack
= 1;
350 } while (len
&& (new = ioat2_dma_get_next_descriptor(ioat_chan
)));
352 hw
->ctl
= IOAT_DMA_DESCRIPTOR_CTL_CP_STS
;
353 if (new->async_tx
.callback
) {
354 hw
->ctl
|= IOAT_DMA_DESCRIPTOR_CTL_INT_GN
;
356 /* move callback into to last desc */
357 new->async_tx
.callback
= first
->async_tx
.callback
;
358 new->async_tx
.callback_param
359 = first
->async_tx
.callback_param
;
360 first
->async_tx
.callback
= NULL
;
361 first
->async_tx
.callback_param
= NULL
;
365 new->tx_cnt
= desc_count
;
366 new->async_tx
.ack
= orig_ack
; /* client is in control of this ack */
368 /* store the original values for use in later cleanup */
370 new->src
= first
->src
;
371 new->dst
= first
->dst
;
372 new->len
= first
->len
;
375 /* cookie incr and addition to used_list must be atomic */
376 cookie
= ioat_chan
->common
.cookie
;
380 ioat_chan
->common
.cookie
= new->async_tx
.cookie
= cookie
;
382 ioat_chan
->dmacount
+= desc_count
;
383 ioat_chan
->pending
+= desc_count
;
384 if (ioat_chan
->pending
>= ioat_pending_level
)
385 __ioat2_dma_memcpy_issue_pending(ioat_chan
);
386 spin_unlock_bh(&ioat_chan
->desc_lock
);
392 * ioat_dma_alloc_descriptor - allocate and return a sw and hw descriptor pair
393 * @ioat_chan: the channel supplying the memory pool for the descriptors
394 * @flags: allocation flags
396 static struct ioat_desc_sw
*ioat_dma_alloc_descriptor(
397 struct ioat_dma_chan
*ioat_chan
,
400 struct ioat_dma_descriptor
*desc
;
401 struct ioat_desc_sw
*desc_sw
;
402 struct ioatdma_device
*ioatdma_device
;
405 ioatdma_device
= to_ioatdma_device(ioat_chan
->common
.device
);
406 desc
= pci_pool_alloc(ioatdma_device
->dma_pool
, flags
, &phys
);
410 desc_sw
= kzalloc(sizeof(*desc_sw
), flags
);
411 if (unlikely(!desc_sw
)) {
412 pci_pool_free(ioatdma_device
->dma_pool
, desc
, phys
);
416 memset(desc
, 0, sizeof(*desc
));
417 dma_async_tx_descriptor_init(&desc_sw
->async_tx
, &ioat_chan
->common
);
418 desc_sw
->async_tx
.tx_set_src
= ioat_set_src
;
419 desc_sw
->async_tx
.tx_set_dest
= ioat_set_dest
;
420 switch (ioat_chan
->device
->version
) {
422 desc_sw
->async_tx
.tx_submit
= ioat1_tx_submit
;
425 desc_sw
->async_tx
.tx_submit
= ioat2_tx_submit
;
428 INIT_LIST_HEAD(&desc_sw
->async_tx
.tx_list
);
431 desc_sw
->async_tx
.phys
= phys
;
436 static int ioat_initial_desc_count
= 256;
437 module_param(ioat_initial_desc_count
, int, 0644);
438 MODULE_PARM_DESC(ioat_initial_desc_count
,
439 "initial descriptors per channel (default: 256)");
442 * ioat2_dma_massage_chan_desc - link the descriptors into a circle
443 * @ioat_chan: the channel to be massaged
445 static void ioat2_dma_massage_chan_desc(struct ioat_dma_chan
*ioat_chan
)
447 struct ioat_desc_sw
*desc
, *_desc
;
449 /* setup used_desc */
450 ioat_chan
->used_desc
.next
= ioat_chan
->free_desc
.next
;
451 ioat_chan
->used_desc
.prev
= NULL
;
453 /* pull free_desc out of the circle so that every node is a hw
454 * descriptor, but leave it pointing to the list
456 ioat_chan
->free_desc
.prev
->next
= ioat_chan
->free_desc
.next
;
457 ioat_chan
->free_desc
.next
->prev
= ioat_chan
->free_desc
.prev
;
459 /* circle link the hw descriptors */
460 desc
= to_ioat_desc(ioat_chan
->free_desc
.next
);
461 desc
->hw
->next
= to_ioat_desc(desc
->node
.next
)->async_tx
.phys
;
462 list_for_each_entry_safe(desc
, _desc
, ioat_chan
->free_desc
.next
, node
) {
463 desc
->hw
->next
= to_ioat_desc(desc
->node
.next
)->async_tx
.phys
;
468 * ioat_dma_alloc_chan_resources - returns the number of allocated descriptors
469 * @chan: the channel to be filled out
471 static int ioat_dma_alloc_chan_resources(struct dma_chan
*chan
)
473 struct ioat_dma_chan
*ioat_chan
= to_ioat_chan(chan
);
474 struct ioat_desc_sw
*desc
;
480 /* have we already been set up? */
481 if (!list_empty(&ioat_chan
->free_desc
))
482 return ioat_chan
->desccount
;
484 /* Setup register to interrupt and write completion status on error */
485 chanctrl
= IOAT_CHANCTRL_ERR_INT_EN
|
486 IOAT_CHANCTRL_ANY_ERR_ABORT_EN
|
487 IOAT_CHANCTRL_ERR_COMPLETION_EN
;
488 writew(chanctrl
, ioat_chan
->reg_base
+ IOAT_CHANCTRL_OFFSET
);
490 chanerr
= readl(ioat_chan
->reg_base
+ IOAT_CHANERR_OFFSET
);
492 dev_err(&ioat_chan
->device
->pdev
->dev
,
493 "CHANERR = %x, clearing\n", chanerr
);
494 writel(chanerr
, ioat_chan
->reg_base
+ IOAT_CHANERR_OFFSET
);
497 /* Allocate descriptors */
498 for (i
= 0; i
< ioat_initial_desc_count
; i
++) {
499 desc
= ioat_dma_alloc_descriptor(ioat_chan
, GFP_KERNEL
);
501 dev_err(&ioat_chan
->device
->pdev
->dev
,
502 "Only %d initial descriptors\n", i
);
505 list_add_tail(&desc
->node
, &tmp_list
);
507 spin_lock_bh(&ioat_chan
->desc_lock
);
508 ioat_chan
->desccount
= i
;
509 list_splice(&tmp_list
, &ioat_chan
->free_desc
);
510 if (ioat_chan
->device
->version
!= IOAT_VER_1_2
)
511 ioat2_dma_massage_chan_desc(ioat_chan
);
512 spin_unlock_bh(&ioat_chan
->desc_lock
);
514 /* allocate a completion writeback area */
515 /* doing 2 32bit writes to mmio since 1 64b write doesn't work */
516 ioat_chan
->completion_virt
=
517 pci_pool_alloc(ioat_chan
->device
->completion_pool
,
519 &ioat_chan
->completion_addr
);
520 memset(ioat_chan
->completion_virt
, 0,
521 sizeof(*ioat_chan
->completion_virt
));
522 writel(((u64
) ioat_chan
->completion_addr
) & 0x00000000FFFFFFFF,
523 ioat_chan
->reg_base
+ IOAT_CHANCMP_OFFSET_LOW
);
524 writel(((u64
) ioat_chan
->completion_addr
) >> 32,
525 ioat_chan
->reg_base
+ IOAT_CHANCMP_OFFSET_HIGH
);
527 tasklet_enable(&ioat_chan
->cleanup_task
);
528 ioat_dma_start_null_desc(ioat_chan
); /* give chain to dma device */
529 return ioat_chan
->desccount
;
533 * ioat_dma_free_chan_resources - release all the descriptors
534 * @chan: the channel to be cleaned
536 static void ioat_dma_free_chan_resources(struct dma_chan
*chan
)
538 struct ioat_dma_chan
*ioat_chan
= to_ioat_chan(chan
);
539 struct ioatdma_device
*ioatdma_device
= to_ioatdma_device(chan
->device
);
540 struct ioat_desc_sw
*desc
, *_desc
;
541 int in_use_descs
= 0;
543 tasklet_disable(&ioat_chan
->cleanup_task
);
544 ioat_dma_memcpy_cleanup(ioat_chan
);
546 /* Delay 100ms after reset to allow internal DMA logic to quiesce
547 * before removing DMA descriptor resources.
549 writeb(IOAT_CHANCMD_RESET
,
551 + IOAT_CHANCMD_OFFSET(ioat_chan
->device
->version
));
554 spin_lock_bh(&ioat_chan
->desc_lock
);
555 switch (ioat_chan
->device
->version
) {
557 list_for_each_entry_safe(desc
, _desc
,
558 &ioat_chan
->used_desc
, node
) {
560 list_del(&desc
->node
);
561 pci_pool_free(ioatdma_device
->dma_pool
, desc
->hw
,
562 desc
->async_tx
.phys
);
565 list_for_each_entry_safe(desc
, _desc
,
566 &ioat_chan
->free_desc
, node
) {
567 list_del(&desc
->node
);
568 pci_pool_free(ioatdma_device
->dma_pool
, desc
->hw
,
569 desc
->async_tx
.phys
);
574 list_for_each_entry_safe(desc
, _desc
,
575 ioat_chan
->free_desc
.next
, node
) {
576 list_del(&desc
->node
);
577 pci_pool_free(ioatdma_device
->dma_pool
, desc
->hw
,
578 desc
->async_tx
.phys
);
581 desc
= to_ioat_desc(ioat_chan
->free_desc
.next
);
582 pci_pool_free(ioatdma_device
->dma_pool
, desc
->hw
,
583 desc
->async_tx
.phys
);
585 INIT_LIST_HEAD(&ioat_chan
->free_desc
);
586 INIT_LIST_HEAD(&ioat_chan
->used_desc
);
589 spin_unlock_bh(&ioat_chan
->desc_lock
);
591 pci_pool_free(ioatdma_device
->completion_pool
,
592 ioat_chan
->completion_virt
,
593 ioat_chan
->completion_addr
);
595 /* one is ok since we left it on there on purpose */
596 if (in_use_descs
> 1)
597 dev_err(&ioat_chan
->device
->pdev
->dev
,
598 "Freeing %d in use descriptors!\n",
601 ioat_chan
->last_completion
= ioat_chan
->completion_addr
= 0;
602 ioat_chan
->pending
= 0;
603 ioat_chan
->dmacount
= 0;
607 * ioat_dma_get_next_descriptor - return the next available descriptor
608 * @ioat_chan: IOAT DMA channel handle
610 * Gets the next descriptor from the chain, and must be called with the
611 * channel's desc_lock held. Allocates more descriptors if the channel
614 static struct ioat_desc_sw
*
615 ioat1_dma_get_next_descriptor(struct ioat_dma_chan
*ioat_chan
)
617 struct ioat_desc_sw
*new;
619 if (!list_empty(&ioat_chan
->free_desc
)) {
620 new = to_ioat_desc(ioat_chan
->free_desc
.next
);
621 list_del(&new->node
);
623 /* try to get another desc */
624 new = ioat_dma_alloc_descriptor(ioat_chan
, GFP_ATOMIC
);
626 dev_err(&ioat_chan
->device
->pdev
->dev
,
636 static struct ioat_desc_sw
*
637 ioat2_dma_get_next_descriptor(struct ioat_dma_chan
*ioat_chan
)
639 struct ioat_desc_sw
*new;
642 * used.prev points to where to start processing
643 * used.next points to next free descriptor
644 * if used.prev == NULL, there are none waiting to be processed
645 * if used.next == used.prev.prev, there is only one free descriptor,
646 * and we need to use it to as a noop descriptor before
647 * linking in a new set of descriptors, since the device
648 * has probably already read the pointer to it
650 if (ioat_chan
->used_desc
.prev
&&
651 ioat_chan
->used_desc
.next
== ioat_chan
->used_desc
.prev
->prev
) {
653 struct ioat_desc_sw
*desc
;
654 struct ioat_desc_sw
*noop_desc
;
657 /* set up the noop descriptor */
658 noop_desc
= to_ioat_desc(ioat_chan
->used_desc
.next
);
659 noop_desc
->hw
->size
= 0;
660 noop_desc
->hw
->ctl
= IOAT_DMA_DESCRIPTOR_NUL
;
661 noop_desc
->hw
->src_addr
= 0;
662 noop_desc
->hw
->dst_addr
= 0;
664 ioat_chan
->used_desc
.next
= ioat_chan
->used_desc
.next
->next
;
665 ioat_chan
->pending
++;
666 ioat_chan
->dmacount
++;
668 /* try to get a few more descriptors */
669 for (i
= 16; i
; i
--) {
670 desc
= ioat_dma_alloc_descriptor(ioat_chan
, GFP_ATOMIC
);
672 dev_err(&ioat_chan
->device
->pdev
->dev
,
676 list_add_tail(&desc
->node
, ioat_chan
->used_desc
.next
);
679 = to_ioat_desc(desc
->node
.next
)->async_tx
.phys
;
680 to_ioat_desc(desc
->node
.prev
)->hw
->next
681 = desc
->async_tx
.phys
;
682 ioat_chan
->desccount
++;
685 ioat_chan
->used_desc
.next
= noop_desc
->node
.next
;
687 new = to_ioat_desc(ioat_chan
->used_desc
.next
);
689 ioat_chan
->used_desc
.next
= new->node
.next
;
691 if (ioat_chan
->used_desc
.prev
== NULL
)
692 ioat_chan
->used_desc
.prev
= &new->node
;
698 static struct ioat_desc_sw
*ioat_dma_get_next_descriptor(
699 struct ioat_dma_chan
*ioat_chan
)
704 switch (ioat_chan
->device
->version
) {
706 return ioat1_dma_get_next_descriptor(ioat_chan
);
709 return ioat2_dma_get_next_descriptor(ioat_chan
);
715 static struct dma_async_tx_descriptor
*ioat1_dma_prep_memcpy(
716 struct dma_chan
*chan
,
720 struct ioat_dma_chan
*ioat_chan
= to_ioat_chan(chan
);
721 struct ioat_desc_sw
*new;
723 spin_lock_bh(&ioat_chan
->desc_lock
);
724 new = ioat_dma_get_next_descriptor(ioat_chan
);
725 spin_unlock_bh(&ioat_chan
->desc_lock
);
729 return &new->async_tx
;
734 static struct dma_async_tx_descriptor
*ioat2_dma_prep_memcpy(
735 struct dma_chan
*chan
,
739 struct ioat_dma_chan
*ioat_chan
= to_ioat_chan(chan
);
740 struct ioat_desc_sw
*new;
742 spin_lock_bh(&ioat_chan
->desc_lock
);
743 new = ioat2_dma_get_next_descriptor(ioat_chan
);
746 * leave ioat_chan->desc_lock set in ioat 2 path
747 * it will get unlocked at end of tx_submit
752 return &new->async_tx
;
757 static void ioat_dma_cleanup_tasklet(unsigned long data
)
759 struct ioat_dma_chan
*chan
= (void *)data
;
760 ioat_dma_memcpy_cleanup(chan
);
761 writew(IOAT_CHANCTRL_INT_DISABLE
,
762 chan
->reg_base
+ IOAT_CHANCTRL_OFFSET
);
766 * ioat_dma_memcpy_cleanup - cleanup up finished descriptors
767 * @chan: ioat channel to be cleaned up
769 static void ioat_dma_memcpy_cleanup(struct ioat_dma_chan
*ioat_chan
)
771 unsigned long phys_complete
;
772 struct ioat_desc_sw
*desc
, *_desc
;
773 dma_cookie_t cookie
= 0;
774 unsigned long desc_phys
;
775 struct ioat_desc_sw
*latest_desc
;
777 prefetch(ioat_chan
->completion_virt
);
779 if (!spin_trylock_bh(&ioat_chan
->cleanup_lock
))
782 /* The completion writeback can happen at any time,
783 so reads by the driver need to be atomic operations
784 The descriptor physical addresses are limited to 32-bits
785 when the CPU can only do a 32-bit mov */
787 #if (BITS_PER_LONG == 64)
789 ioat_chan
->completion_virt
->full
790 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR
;
793 ioat_chan
->completion_virt
->low
& IOAT_LOW_COMPLETION_MASK
;
796 if ((ioat_chan
->completion_virt
->full
797 & IOAT_CHANSTS_DMA_TRANSFER_STATUS
) ==
798 IOAT_CHANSTS_DMA_TRANSFER_STATUS_HALTED
) {
799 dev_err(&ioat_chan
->device
->pdev
->dev
,
800 "Channel halted, chanerr = %x\n",
801 readl(ioat_chan
->reg_base
+ IOAT_CHANERR_OFFSET
));
803 /* TODO do something to salvage the situation */
806 if (phys_complete
== ioat_chan
->last_completion
) {
807 spin_unlock_bh(&ioat_chan
->cleanup_lock
);
812 spin_lock_bh(&ioat_chan
->desc_lock
);
813 switch (ioat_chan
->device
->version
) {
815 list_for_each_entry_safe(desc
, _desc
,
816 &ioat_chan
->used_desc
, node
) {
819 * Incoming DMA requests may use multiple descriptors,
820 * due to exceeding xfercap, perhaps. If so, only the
821 * last one will have a cookie, and require unmapping.
823 if (desc
->async_tx
.cookie
) {
824 cookie
= desc
->async_tx
.cookie
;
827 * yes we are unmapping both _page and _single
828 * alloc'd regions with unmap_page. Is this
831 pci_unmap_page(ioat_chan
->device
->pdev
,
832 pci_unmap_addr(desc
, dst
),
833 pci_unmap_len(desc
, len
),
835 pci_unmap_page(ioat_chan
->device
->pdev
,
836 pci_unmap_addr(desc
, src
),
837 pci_unmap_len(desc
, len
),
840 if (desc
->async_tx
.callback
) {
841 desc
->async_tx
.callback(desc
->async_tx
.callback_param
);
842 desc
->async_tx
.callback
= NULL
;
846 if (desc
->async_tx
.phys
!= phys_complete
) {
848 * a completed entry, but not the last, so clean
849 * up if the client is done with the descriptor
851 if (desc
->async_tx
.ack
) {
852 list_del(&desc
->node
);
853 list_add_tail(&desc
->node
,
854 &ioat_chan
->free_desc
);
856 desc
->async_tx
.cookie
= 0;
859 * last used desc. Do not remove, so we can
860 * append from it, but don't look at it next
863 desc
->async_tx
.cookie
= 0;
865 /* TODO check status bits? */
871 /* has some other thread has already cleaned up? */
872 if (ioat_chan
->used_desc
.prev
== NULL
)
875 /* work backwards to find latest finished desc */
876 desc
= to_ioat_desc(ioat_chan
->used_desc
.next
);
879 desc
= to_ioat_desc(desc
->node
.prev
);
880 desc_phys
= (unsigned long)desc
->async_tx
.phys
881 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR
;
882 if (desc_phys
== phys_complete
) {
886 } while (&desc
->node
!= ioat_chan
->used_desc
.prev
);
888 if (latest_desc
!= NULL
) {
890 /* work forwards to clear finished descriptors */
891 for (desc
= to_ioat_desc(ioat_chan
->used_desc
.prev
);
892 &desc
->node
!= latest_desc
->node
.next
&&
893 &desc
->node
!= ioat_chan
->used_desc
.next
;
894 desc
= to_ioat_desc(desc
->node
.next
)) {
895 if (desc
->async_tx
.cookie
) {
896 cookie
= desc
->async_tx
.cookie
;
897 desc
->async_tx
.cookie
= 0;
899 pci_unmap_page(ioat_chan
->device
->pdev
,
900 pci_unmap_addr(desc
, dst
),
901 pci_unmap_len(desc
, len
),
903 pci_unmap_page(ioat_chan
->device
->pdev
,
904 pci_unmap_addr(desc
, src
),
905 pci_unmap_len(desc
, len
),
908 if (desc
->async_tx
.callback
) {
909 desc
->async_tx
.callback(desc
->async_tx
.callback_param
);
910 desc
->async_tx
.callback
= NULL
;
915 /* move used.prev up beyond those that are finished */
916 if (&desc
->node
== ioat_chan
->used_desc
.next
)
917 ioat_chan
->used_desc
.prev
= NULL
;
919 ioat_chan
->used_desc
.prev
= &desc
->node
;
924 spin_unlock_bh(&ioat_chan
->desc_lock
);
926 ioat_chan
->last_completion
= phys_complete
;
928 ioat_chan
->completed_cookie
= cookie
;
930 spin_unlock_bh(&ioat_chan
->cleanup_lock
);
933 static void ioat_dma_dependency_added(struct dma_chan
*chan
)
935 struct ioat_dma_chan
*ioat_chan
= to_ioat_chan(chan
);
936 spin_lock_bh(&ioat_chan
->desc_lock
);
937 if (ioat_chan
->pending
== 0) {
938 spin_unlock_bh(&ioat_chan
->desc_lock
);
939 ioat_dma_memcpy_cleanup(ioat_chan
);
941 spin_unlock_bh(&ioat_chan
->desc_lock
);
945 * ioat_dma_is_complete - poll the status of a IOAT DMA transaction
946 * @chan: IOAT DMA channel handle
947 * @cookie: DMA transaction identifier
948 * @done: if not %NULL, updated with last completed transaction
949 * @used: if not %NULL, updated with last used transaction
951 static enum dma_status
ioat_dma_is_complete(struct dma_chan
*chan
,
956 struct ioat_dma_chan
*ioat_chan
= to_ioat_chan(chan
);
957 dma_cookie_t last_used
;
958 dma_cookie_t last_complete
;
961 last_used
= chan
->cookie
;
962 last_complete
= ioat_chan
->completed_cookie
;
965 *done
= last_complete
;
969 ret
= dma_async_is_complete(cookie
, last_complete
, last_used
);
970 if (ret
== DMA_SUCCESS
)
973 ioat_dma_memcpy_cleanup(ioat_chan
);
975 last_used
= chan
->cookie
;
976 last_complete
= ioat_chan
->completed_cookie
;
979 *done
= last_complete
;
983 return dma_async_is_complete(cookie
, last_complete
, last_used
);
986 static void ioat_dma_start_null_desc(struct ioat_dma_chan
*ioat_chan
)
988 struct ioat_desc_sw
*desc
;
990 spin_lock_bh(&ioat_chan
->desc_lock
);
992 desc
= ioat_dma_get_next_descriptor(ioat_chan
);
993 desc
->hw
->ctl
= IOAT_DMA_DESCRIPTOR_NUL
994 | IOAT_DMA_DESCRIPTOR_CTL_INT_GN
995 | IOAT_DMA_DESCRIPTOR_CTL_CP_STS
;
997 desc
->hw
->src_addr
= 0;
998 desc
->hw
->dst_addr
= 0;
999 desc
->async_tx
.ack
= 1;
1000 switch (ioat_chan
->device
->version
) {
1003 list_add_tail(&desc
->node
, &ioat_chan
->used_desc
);
1005 writel(((u64
) desc
->async_tx
.phys
) & 0x00000000FFFFFFFF,
1006 ioat_chan
->reg_base
+ IOAT1_CHAINADDR_OFFSET_LOW
);
1007 writel(((u64
) desc
->async_tx
.phys
) >> 32,
1008 ioat_chan
->reg_base
+ IOAT1_CHAINADDR_OFFSET_HIGH
);
1010 writeb(IOAT_CHANCMD_START
, ioat_chan
->reg_base
1011 + IOAT_CHANCMD_OFFSET(ioat_chan
->device
->version
));
1014 writel(((u64
) desc
->async_tx
.phys
) & 0x00000000FFFFFFFF,
1015 ioat_chan
->reg_base
+ IOAT2_CHAINADDR_OFFSET_LOW
);
1016 writel(((u64
) desc
->async_tx
.phys
) >> 32,
1017 ioat_chan
->reg_base
+ IOAT2_CHAINADDR_OFFSET_HIGH
);
1019 ioat_chan
->dmacount
++;
1020 __ioat2_dma_memcpy_issue_pending(ioat_chan
);
1023 spin_unlock_bh(&ioat_chan
->desc_lock
);
1027 * Perform a IOAT transaction to verify the HW works.
1029 #define IOAT_TEST_SIZE 2000
1031 static void ioat_dma_test_callback(void *dma_async_param
)
1033 printk(KERN_ERR
"ioatdma: ioat_dma_test_callback(%p)\n",
1038 * ioat_dma_self_test - Perform a IOAT transaction to verify the HW works.
1039 * @device: device to be tested
1041 static int ioat_dma_self_test(struct ioatdma_device
*device
)
1046 struct dma_chan
*dma_chan
;
1047 struct dma_async_tx_descriptor
*tx
;
1049 dma_cookie_t cookie
;
1052 src
= kzalloc(sizeof(u8
) * IOAT_TEST_SIZE
, GFP_KERNEL
);
1055 dest
= kzalloc(sizeof(u8
) * IOAT_TEST_SIZE
, GFP_KERNEL
);
1061 /* Fill in src buffer */
1062 for (i
= 0; i
< IOAT_TEST_SIZE
; i
++)
1065 /* Start copy, using first DMA channel */
1066 dma_chan
= container_of(device
->common
.channels
.next
,
1069 if (device
->common
.device_alloc_chan_resources(dma_chan
) < 1) {
1070 dev_err(&device
->pdev
->dev
,
1071 "selftest cannot allocate chan resource\n");
1076 tx
= device
->common
.device_prep_dma_memcpy(dma_chan
, IOAT_TEST_SIZE
, 0);
1078 dev_err(&device
->pdev
->dev
,
1079 "Self-test prep failed, disabling\n");
1081 goto free_resources
;
1085 addr
= dma_map_single(dma_chan
->device
->dev
, src
, IOAT_TEST_SIZE
,
1087 tx
->tx_set_src(addr
, tx
, 0);
1088 addr
= dma_map_single(dma_chan
->device
->dev
, dest
, IOAT_TEST_SIZE
,
1090 tx
->tx_set_dest(addr
, tx
, 0);
1091 tx
->callback
= ioat_dma_test_callback
;
1092 tx
->callback_param
= (void *)0x8086;
1093 cookie
= tx
->tx_submit(tx
);
1095 dev_err(&device
->pdev
->dev
,
1096 "Self-test setup failed, disabling\n");
1098 goto free_resources
;
1100 device
->common
.device_issue_pending(dma_chan
);
1103 if (device
->common
.device_is_tx_complete(dma_chan
, cookie
, NULL
, NULL
)
1105 dev_err(&device
->pdev
->dev
,
1106 "Self-test copy timed out, disabling\n");
1108 goto free_resources
;
1110 if (memcmp(src
, dest
, IOAT_TEST_SIZE
)) {
1111 dev_err(&device
->pdev
->dev
,
1112 "Self-test copy failed compare, disabling\n");
1114 goto free_resources
;
1118 device
->common
.device_free_chan_resources(dma_chan
);
1125 static char ioat_interrupt_style
[32] = "msix";
1126 module_param_string(ioat_interrupt_style
, ioat_interrupt_style
,
1127 sizeof(ioat_interrupt_style
), 0644);
1128 MODULE_PARM_DESC(ioat_interrupt_style
,
1129 "set ioat interrupt style: msix (default), "
1130 "msix-single-vector, msi, intx)");
1133 * ioat_dma_setup_interrupts - setup interrupt handler
1134 * @device: ioat device
1136 static int ioat_dma_setup_interrupts(struct ioatdma_device
*device
)
1138 struct ioat_dma_chan
*ioat_chan
;
1139 int err
, i
, j
, msixcnt
;
1142 if (!strcmp(ioat_interrupt_style
, "msix"))
1144 if (!strcmp(ioat_interrupt_style
, "msix-single-vector"))
1145 goto msix_single_vector
;
1146 if (!strcmp(ioat_interrupt_style
, "msi"))
1148 if (!strcmp(ioat_interrupt_style
, "intx"))
1150 dev_err(&device
->pdev
->dev
, "invalid ioat_interrupt_style %s\n",
1151 ioat_interrupt_style
);
1155 /* The number of MSI-X vectors should equal the number of channels */
1156 msixcnt
= device
->common
.chancnt
;
1157 for (i
= 0; i
< msixcnt
; i
++)
1158 device
->msix_entries
[i
].entry
= i
;
1160 err
= pci_enable_msix(device
->pdev
, device
->msix_entries
, msixcnt
);
1164 goto msix_single_vector
;
1166 for (i
= 0; i
< msixcnt
; i
++) {
1167 ioat_chan
= ioat_lookup_chan_by_index(device
, i
);
1168 err
= request_irq(device
->msix_entries
[i
].vector
,
1169 ioat_dma_do_interrupt_msix
,
1170 0, "ioat-msix", ioat_chan
);
1172 for (j
= 0; j
< i
; j
++) {
1174 ioat_lookup_chan_by_index(device
, j
);
1175 free_irq(device
->msix_entries
[j
].vector
,
1178 goto msix_single_vector
;
1181 intrctrl
|= IOAT_INTRCTRL_MSIX_VECTOR_CONTROL
;
1182 device
->irq_mode
= msix_multi_vector
;
1186 device
->msix_entries
[0].entry
= 0;
1187 err
= pci_enable_msix(device
->pdev
, device
->msix_entries
, 1);
1191 err
= request_irq(device
->msix_entries
[0].vector
, ioat_dma_do_interrupt
,
1192 0, "ioat-msix", device
);
1194 pci_disable_msix(device
->pdev
);
1197 device
->irq_mode
= msix_single_vector
;
1201 err
= pci_enable_msi(device
->pdev
);
1205 err
= request_irq(device
->pdev
->irq
, ioat_dma_do_interrupt
,
1206 0, "ioat-msi", device
);
1208 pci_disable_msi(device
->pdev
);
1212 * CB 1.2 devices need a bit set in configuration space to enable MSI
1214 if (device
->version
== IOAT_VER_1_2
) {
1216 pci_read_config_dword(device
->pdev
,
1217 IOAT_PCI_DMACTRL_OFFSET
, &dmactrl
);
1218 dmactrl
|= IOAT_PCI_DMACTRL_MSI_EN
;
1219 pci_write_config_dword(device
->pdev
,
1220 IOAT_PCI_DMACTRL_OFFSET
, dmactrl
);
1222 device
->irq_mode
= msi
;
1226 err
= request_irq(device
->pdev
->irq
, ioat_dma_do_interrupt
,
1227 IRQF_SHARED
, "ioat-intx", device
);
1230 device
->irq_mode
= intx
;
1233 intrctrl
|= IOAT_INTRCTRL_MASTER_INT_EN
;
1234 writeb(intrctrl
, device
->reg_base
+ IOAT_INTRCTRL_OFFSET
);
1238 /* Disable all interrupt generation */
1239 writeb(0, device
->reg_base
+ IOAT_INTRCTRL_OFFSET
);
1240 dev_err(&device
->pdev
->dev
, "no usable interrupts\n");
1241 device
->irq_mode
= none
;
1246 * ioat_dma_remove_interrupts - remove whatever interrupts were set
1247 * @device: ioat device
1249 static void ioat_dma_remove_interrupts(struct ioatdma_device
*device
)
1251 struct ioat_dma_chan
*ioat_chan
;
1254 /* Disable all interrupt generation */
1255 writeb(0, device
->reg_base
+ IOAT_INTRCTRL_OFFSET
);
1257 switch (device
->irq_mode
) {
1258 case msix_multi_vector
:
1259 for (i
= 0; i
< device
->common
.chancnt
; i
++) {
1260 ioat_chan
= ioat_lookup_chan_by_index(device
, i
);
1261 free_irq(device
->msix_entries
[i
].vector
, ioat_chan
);
1263 pci_disable_msix(device
->pdev
);
1265 case msix_single_vector
:
1266 free_irq(device
->msix_entries
[0].vector
, device
);
1267 pci_disable_msix(device
->pdev
);
1270 free_irq(device
->pdev
->irq
, device
);
1271 pci_disable_msi(device
->pdev
);
1274 free_irq(device
->pdev
->irq
, device
);
1277 dev_warn(&device
->pdev
->dev
,
1278 "call to %s without interrupts setup\n", __func__
);
1280 device
->irq_mode
= none
;
1283 struct ioatdma_device
*ioat_dma_probe(struct pci_dev
*pdev
,
1284 void __iomem
*iobase
)
1287 struct ioatdma_device
*device
;
1289 device
= kzalloc(sizeof(*device
), GFP_KERNEL
);
1294 device
->pdev
= pdev
;
1295 device
->reg_base
= iobase
;
1296 device
->version
= readb(device
->reg_base
+ IOAT_VER_OFFSET
);
1298 /* DMA coherent memory pool for DMA descriptor allocations */
1299 device
->dma_pool
= pci_pool_create("dma_desc_pool", pdev
,
1300 sizeof(struct ioat_dma_descriptor
),
1302 if (!device
->dma_pool
) {
1307 device
->completion_pool
= pci_pool_create("completion_pool", pdev
,
1308 sizeof(u64
), SMP_CACHE_BYTES
,
1310 if (!device
->completion_pool
) {
1312 goto err_completion_pool
;
1315 INIT_LIST_HEAD(&device
->common
.channels
);
1316 ioat_dma_enumerate_channels(device
);
1318 device
->common
.device_alloc_chan_resources
=
1319 ioat_dma_alloc_chan_resources
;
1320 device
->common
.device_free_chan_resources
=
1321 ioat_dma_free_chan_resources
;
1322 device
->common
.dev
= &pdev
->dev
;
1324 dma_cap_set(DMA_MEMCPY
, device
->common
.cap_mask
);
1325 device
->common
.device_is_tx_complete
= ioat_dma_is_complete
;
1326 device
->common
.device_dependency_added
= ioat_dma_dependency_added
;
1327 switch (device
->version
) {
1329 device
->common
.device_prep_dma_memcpy
= ioat1_dma_prep_memcpy
;
1330 device
->common
.device_issue_pending
=
1331 ioat1_dma_memcpy_issue_pending
;
1334 device
->common
.device_prep_dma_memcpy
= ioat2_dma_prep_memcpy
;
1335 device
->common
.device_issue_pending
=
1336 ioat2_dma_memcpy_issue_pending
;
1340 dev_err(&device
->pdev
->dev
,
1341 "Intel(R) I/OAT DMA Engine found,"
1342 " %d channels, device version 0x%02x, driver version %s\n",
1343 device
->common
.chancnt
, device
->version
, IOAT_DMA_VERSION
);
1345 err
= ioat_dma_setup_interrupts(device
);
1347 goto err_setup_interrupts
;
1349 err
= ioat_dma_self_test(device
);
1353 dma_async_device_register(&device
->common
);
1358 ioat_dma_remove_interrupts(device
);
1359 err_setup_interrupts
:
1360 pci_pool_destroy(device
->completion_pool
);
1361 err_completion_pool
:
1362 pci_pool_destroy(device
->dma_pool
);
1367 "Intel(R) I/OAT DMA Engine initialization failed\n");
1371 void ioat_dma_remove(struct ioatdma_device
*device
)
1373 struct dma_chan
*chan
, *_chan
;
1374 struct ioat_dma_chan
*ioat_chan
;
1376 ioat_dma_remove_interrupts(device
);
1378 dma_async_device_unregister(&device
->common
);
1380 pci_pool_destroy(device
->dma_pool
);
1381 pci_pool_destroy(device
->completion_pool
);
1383 iounmap(device
->reg_base
);
1384 pci_release_regions(device
->pdev
);
1385 pci_disable_device(device
->pdev
);
1387 list_for_each_entry_safe(chan
, _chan
,
1388 &device
->common
.channels
, device_node
) {
1389 ioat_chan
= to_ioat_chan(chan
);
1390 list_del(&chan
->device_node
);