2 * offload engine driver for the Marvell XOR engine
3 * Copyright (C) 2007, 2008, Marvell International Ltd.
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 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
15 #include <linux/init.h>
16 #include <linux/slab.h>
17 #include <linux/delay.h>
18 #include <linux/dma-mapping.h>
19 #include <linux/spinlock.h>
20 #include <linux/interrupt.h>
21 #include <linux/of_device.h>
22 #include <linux/platform_device.h>
23 #include <linux/memory.h>
24 #include <linux/clk.h>
26 #include <linux/of_irq.h>
27 #include <linux/irqdomain.h>
28 #include <linux/cpumask.h>
29 #include <linux/platform_data/dma-mv_xor.h>
31 #include "dmaengine.h"
39 static void mv_xor_issue_pending(struct dma_chan *chan);
41 #define to_mv_xor_chan(chan) \
42 container_of(chan, struct mv_xor_chan, dmachan)
44 #define to_mv_xor_slot(tx) \
45 container_of(tx, struct mv_xor_desc_slot, async_tx)
47 #define mv_chan_to_devp(chan) \
50 static void mv_desc_init(struct mv_xor_desc_slot *desc,
51 dma_addr_t addr, u32 byte_count,
52 enum dma_ctrl_flags flags)
54 struct mv_xor_desc *hw_desc = desc->hw_desc;
56 hw_desc->status = XOR_DESC_DMA_OWNED;
57 hw_desc->phy_next_desc = 0;
58 /* Enable end-of-descriptor interrupts only for DMA_PREP_INTERRUPT */
59 hw_desc->desc_command = (flags & DMA_PREP_INTERRUPT) ?
60 XOR_DESC_EOD_INT_EN : 0;
61 hw_desc->phy_dest_addr = addr;
62 hw_desc->byte_count = byte_count;
65 static void mv_desc_set_mode(struct mv_xor_desc_slot *desc)
67 struct mv_xor_desc *hw_desc = desc->hw_desc;
72 hw_desc->desc_command |= XOR_DESC_OPERATION_XOR;
75 hw_desc->desc_command |= XOR_DESC_OPERATION_MEMCPY;
83 static void mv_desc_set_next_desc(struct mv_xor_desc_slot *desc,
86 struct mv_xor_desc *hw_desc = desc->hw_desc;
87 BUG_ON(hw_desc->phy_next_desc);
88 hw_desc->phy_next_desc = next_desc_addr;
91 static void mv_desc_set_src_addr(struct mv_xor_desc_slot *desc,
92 int index, dma_addr_t addr)
94 struct mv_xor_desc *hw_desc = desc->hw_desc;
95 hw_desc->phy_src_addr[mv_phy_src_idx(index)] = addr;
96 if (desc->type == DMA_XOR)
97 hw_desc->desc_command |= (1 << index);
100 static u32 mv_chan_get_current_desc(struct mv_xor_chan *chan)
102 return readl_relaxed(XOR_CURR_DESC(chan));
105 static void mv_chan_set_next_descriptor(struct mv_xor_chan *chan,
108 writel_relaxed(next_desc_addr, XOR_NEXT_DESC(chan));
111 static void mv_chan_unmask_interrupts(struct mv_xor_chan *chan)
113 u32 val = readl_relaxed(XOR_INTR_MASK(chan));
114 val |= XOR_INTR_MASK_VALUE << (chan->idx * 16);
115 writel_relaxed(val, XOR_INTR_MASK(chan));
118 static u32 mv_chan_get_intr_cause(struct mv_xor_chan *chan)
120 u32 intr_cause = readl_relaxed(XOR_INTR_CAUSE(chan));
121 intr_cause = (intr_cause >> (chan->idx * 16)) & 0xFFFF;
125 static void mv_chan_clear_eoc_cause(struct mv_xor_chan *chan)
129 val = XOR_INT_END_OF_DESC | XOR_INT_END_OF_CHAIN | XOR_INT_STOPPED;
130 val = ~(val << (chan->idx * 16));
131 dev_dbg(mv_chan_to_devp(chan), "%s, val 0x%08x\n", __func__, val);
132 writel_relaxed(val, XOR_INTR_CAUSE(chan));
135 static void mv_chan_clear_err_status(struct mv_xor_chan *chan)
137 u32 val = 0xFFFF0000 >> (chan->idx * 16);
138 writel_relaxed(val, XOR_INTR_CAUSE(chan));
141 static void mv_chan_set_mode(struct mv_xor_chan *chan,
142 enum dma_transaction_type type)
145 u32 config = readl_relaxed(XOR_CONFIG(chan));
149 op_mode = XOR_OPERATION_MODE_XOR;
152 op_mode = XOR_OPERATION_MODE_MEMCPY;
155 dev_err(mv_chan_to_devp(chan),
156 "error: unsupported operation %d\n",
165 #if defined(__BIG_ENDIAN)
166 config |= XOR_DESCRIPTOR_SWAP;
168 config &= ~XOR_DESCRIPTOR_SWAP;
171 writel_relaxed(config, XOR_CONFIG(chan));
172 chan->current_type = type;
175 static void mv_chan_set_mode_to_desc(struct mv_xor_chan *chan)
178 u32 config = readl_relaxed(XOR_CONFIG(chan));
180 op_mode = XOR_OPERATION_MODE_IN_DESC;
185 #if defined(__BIG_ENDIAN)
186 config |= XOR_DESCRIPTOR_SWAP;
188 config &= ~XOR_DESCRIPTOR_SWAP;
191 writel_relaxed(config, XOR_CONFIG(chan));
194 static void mv_chan_activate(struct mv_xor_chan *chan)
196 dev_dbg(mv_chan_to_devp(chan), " activate chan.\n");
198 /* writel ensures all descriptors are flushed before activation */
199 writel(BIT(0), XOR_ACTIVATION(chan));
202 static char mv_chan_is_busy(struct mv_xor_chan *chan)
204 u32 state = readl_relaxed(XOR_ACTIVATION(chan));
206 state = (state >> 4) & 0x3;
208 return (state == 1) ? 1 : 0;
212 * mv_chan_start_new_chain - program the engine to operate on new
213 * chain headed by sw_desc
214 * Caller must hold &mv_chan->lock while calling this function
216 static void mv_chan_start_new_chain(struct mv_xor_chan *mv_chan,
217 struct mv_xor_desc_slot *sw_desc)
219 dev_dbg(mv_chan_to_devp(mv_chan), "%s %d: sw_desc %p\n",
220 __func__, __LINE__, sw_desc);
222 /* set the hardware chain */
223 mv_chan_set_next_descriptor(mv_chan, sw_desc->async_tx.phys);
226 mv_xor_issue_pending(&mv_chan->dmachan);
230 mv_desc_run_tx_complete_actions(struct mv_xor_desc_slot *desc,
231 struct mv_xor_chan *mv_chan,
234 BUG_ON(desc->async_tx.cookie < 0);
236 if (desc->async_tx.cookie > 0) {
237 cookie = desc->async_tx.cookie;
239 /* call the callback (must not sleep or submit new
240 * operations to this channel)
242 if (desc->async_tx.callback)
243 desc->async_tx.callback(
244 desc->async_tx.callback_param);
246 dma_descriptor_unmap(&desc->async_tx);
249 /* run dependent operations */
250 dma_run_dependencies(&desc->async_tx);
256 mv_chan_clean_completed_slots(struct mv_xor_chan *mv_chan)
258 struct mv_xor_desc_slot *iter, *_iter;
260 dev_dbg(mv_chan_to_devp(mv_chan), "%s %d\n", __func__, __LINE__);
261 list_for_each_entry_safe(iter, _iter, &mv_chan->completed_slots,
264 if (async_tx_test_ack(&iter->async_tx))
265 list_move_tail(&iter->node, &mv_chan->free_slots);
271 mv_desc_clean_slot(struct mv_xor_desc_slot *desc,
272 struct mv_xor_chan *mv_chan)
274 dev_dbg(mv_chan_to_devp(mv_chan), "%s %d: desc %p flags %d\n",
275 __func__, __LINE__, desc, desc->async_tx.flags);
277 /* the client is allowed to attach dependent operations
280 if (!async_tx_test_ack(&desc->async_tx))
281 /* move this slot to the completed_slots */
282 list_move_tail(&desc->node, &mv_chan->completed_slots);
284 list_move_tail(&desc->node, &mv_chan->free_slots);
289 /* This function must be called with the mv_xor_chan spinlock held */
290 static void mv_chan_slot_cleanup(struct mv_xor_chan *mv_chan)
292 struct mv_xor_desc_slot *iter, *_iter;
293 dma_cookie_t cookie = 0;
294 int busy = mv_chan_is_busy(mv_chan);
295 u32 current_desc = mv_chan_get_current_desc(mv_chan);
296 int current_cleaned = 0;
297 struct mv_xor_desc *hw_desc;
299 dev_dbg(mv_chan_to_devp(mv_chan), "%s %d\n", __func__, __LINE__);
300 dev_dbg(mv_chan_to_devp(mv_chan), "current_desc %x\n", current_desc);
301 mv_chan_clean_completed_slots(mv_chan);
303 /* free completed slots from the chain starting with
304 * the oldest descriptor
307 list_for_each_entry_safe(iter, _iter, &mv_chan->chain,
310 /* clean finished descriptors */
311 hw_desc = iter->hw_desc;
312 if (hw_desc->status & XOR_DESC_SUCCESS) {
313 cookie = mv_desc_run_tx_complete_actions(iter, mv_chan,
316 /* done processing desc, clean slot */
317 mv_desc_clean_slot(iter, mv_chan);
319 /* break if we did cleaned the current */
320 if (iter->async_tx.phys == current_desc) {
325 if (iter->async_tx.phys == current_desc) {
332 if ((busy == 0) && !list_empty(&mv_chan->chain)) {
333 if (current_cleaned) {
335 * current descriptor cleaned and removed, run
338 iter = list_entry(mv_chan->chain.next,
339 struct mv_xor_desc_slot,
341 mv_chan_start_new_chain(mv_chan, iter);
343 if (!list_is_last(&iter->node, &mv_chan->chain)) {
345 * descriptors are still waiting after
346 * current, trigger them
348 iter = list_entry(iter->node.next,
349 struct mv_xor_desc_slot,
351 mv_chan_start_new_chain(mv_chan, iter);
354 * some descriptors are still waiting
357 tasklet_schedule(&mv_chan->irq_tasklet);
363 mv_chan->dmachan.completed_cookie = cookie;
366 static void mv_xor_tasklet(unsigned long data)
368 struct mv_xor_chan *chan = (struct mv_xor_chan *) data;
370 spin_lock_bh(&chan->lock);
371 mv_chan_slot_cleanup(chan);
372 spin_unlock_bh(&chan->lock);
375 static struct mv_xor_desc_slot *
376 mv_chan_alloc_slot(struct mv_xor_chan *mv_chan)
378 struct mv_xor_desc_slot *iter;
380 spin_lock_bh(&mv_chan->lock);
382 if (!list_empty(&mv_chan->free_slots)) {
383 iter = list_first_entry(&mv_chan->free_slots,
384 struct mv_xor_desc_slot,
387 list_move_tail(&iter->node, &mv_chan->allocated_slots);
389 spin_unlock_bh(&mv_chan->lock);
391 /* pre-ack descriptor */
392 async_tx_ack(&iter->async_tx);
393 iter->async_tx.cookie = -EBUSY;
399 spin_unlock_bh(&mv_chan->lock);
401 /* try to free some slots if the allocation fails */
402 tasklet_schedule(&mv_chan->irq_tasklet);
407 /************************ DMA engine API functions ****************************/
409 mv_xor_tx_submit(struct dma_async_tx_descriptor *tx)
411 struct mv_xor_desc_slot *sw_desc = to_mv_xor_slot(tx);
412 struct mv_xor_chan *mv_chan = to_mv_xor_chan(tx->chan);
413 struct mv_xor_desc_slot *old_chain_tail;
415 int new_hw_chain = 1;
417 dev_dbg(mv_chan_to_devp(mv_chan),
418 "%s sw_desc %p: async_tx %p\n",
419 __func__, sw_desc, &sw_desc->async_tx);
421 spin_lock_bh(&mv_chan->lock);
422 cookie = dma_cookie_assign(tx);
424 if (list_empty(&mv_chan->chain))
425 list_move_tail(&sw_desc->node, &mv_chan->chain);
429 old_chain_tail = list_entry(mv_chan->chain.prev,
430 struct mv_xor_desc_slot,
432 list_move_tail(&sw_desc->node, &mv_chan->chain);
434 dev_dbg(mv_chan_to_devp(mv_chan), "Append to last desc %pa\n",
435 &old_chain_tail->async_tx.phys);
437 /* fix up the hardware chain */
438 mv_desc_set_next_desc(old_chain_tail, sw_desc->async_tx.phys);
440 /* if the channel is not busy */
441 if (!mv_chan_is_busy(mv_chan)) {
442 u32 current_desc = mv_chan_get_current_desc(mv_chan);
444 * and the curren desc is the end of the chain before
445 * the append, then we need to start the channel
447 if (current_desc == old_chain_tail->async_tx.phys)
453 mv_chan_start_new_chain(mv_chan, sw_desc);
455 spin_unlock_bh(&mv_chan->lock);
460 /* returns the number of allocated descriptors */
461 static int mv_xor_alloc_chan_resources(struct dma_chan *chan)
466 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
467 struct mv_xor_desc_slot *slot = NULL;
468 int num_descs_in_pool = MV_XOR_POOL_SIZE/MV_XOR_SLOT_SIZE;
470 /* Allocate descriptor slots */
471 idx = mv_chan->slots_allocated;
472 while (idx < num_descs_in_pool) {
473 slot = kzalloc(sizeof(*slot), GFP_KERNEL);
475 dev_info(mv_chan_to_devp(mv_chan),
476 "channel only initialized %d descriptor slots",
480 virt_desc = mv_chan->dma_desc_pool_virt;
481 slot->hw_desc = virt_desc + idx * MV_XOR_SLOT_SIZE;
483 dma_async_tx_descriptor_init(&slot->async_tx, chan);
484 slot->async_tx.tx_submit = mv_xor_tx_submit;
485 INIT_LIST_HEAD(&slot->node);
486 dma_desc = mv_chan->dma_desc_pool;
487 slot->async_tx.phys = dma_desc + idx * MV_XOR_SLOT_SIZE;
490 spin_lock_bh(&mv_chan->lock);
491 mv_chan->slots_allocated = idx;
492 list_add_tail(&slot->node, &mv_chan->free_slots);
493 spin_unlock_bh(&mv_chan->lock);
496 dev_dbg(mv_chan_to_devp(mv_chan),
497 "allocated %d descriptor slots\n",
498 mv_chan->slots_allocated);
500 return mv_chan->slots_allocated ? : -ENOMEM;
503 static struct dma_async_tx_descriptor *
504 mv_xor_prep_dma_xor(struct dma_chan *chan, dma_addr_t dest, dma_addr_t *src,
505 unsigned int src_cnt, size_t len, unsigned long flags)
507 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
508 struct mv_xor_desc_slot *sw_desc;
510 if (unlikely(len < MV_XOR_MIN_BYTE_COUNT))
513 BUG_ON(len > MV_XOR_MAX_BYTE_COUNT);
515 dev_dbg(mv_chan_to_devp(mv_chan),
516 "%s src_cnt: %d len: %u dest %pad flags: %ld\n",
517 __func__, src_cnt, len, &dest, flags);
519 sw_desc = mv_chan_alloc_slot(mv_chan);
521 sw_desc->type = DMA_XOR;
522 sw_desc->async_tx.flags = flags;
523 mv_desc_init(sw_desc, dest, len, flags);
524 if (mv_chan->op_in_desc == XOR_MODE_IN_DESC)
525 mv_desc_set_mode(sw_desc);
527 mv_desc_set_src_addr(sw_desc, src_cnt, src[src_cnt]);
530 dev_dbg(mv_chan_to_devp(mv_chan),
531 "%s sw_desc %p async_tx %p \n",
532 __func__, sw_desc, &sw_desc->async_tx);
533 return sw_desc ? &sw_desc->async_tx : NULL;
536 static struct dma_async_tx_descriptor *
537 mv_xor_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
538 size_t len, unsigned long flags)
541 * A MEMCPY operation is identical to an XOR operation with only
542 * a single source address.
544 return mv_xor_prep_dma_xor(chan, dest, &src, 1, len, flags);
547 static struct dma_async_tx_descriptor *
548 mv_xor_prep_dma_interrupt(struct dma_chan *chan, unsigned long flags)
550 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
551 dma_addr_t src, dest;
554 src = mv_chan->dummy_src_addr;
555 dest = mv_chan->dummy_dst_addr;
556 len = MV_XOR_MIN_BYTE_COUNT;
559 * We implement the DMA_INTERRUPT operation as a minimum sized
560 * XOR operation with a single dummy source address.
562 return mv_xor_prep_dma_xor(chan, dest, &src, 1, len, flags);
565 static void mv_xor_free_chan_resources(struct dma_chan *chan)
567 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
568 struct mv_xor_desc_slot *iter, *_iter;
569 int in_use_descs = 0;
571 spin_lock_bh(&mv_chan->lock);
573 mv_chan_slot_cleanup(mv_chan);
575 list_for_each_entry_safe(iter, _iter, &mv_chan->chain,
578 list_move_tail(&iter->node, &mv_chan->free_slots);
580 list_for_each_entry_safe(iter, _iter, &mv_chan->completed_slots,
583 list_move_tail(&iter->node, &mv_chan->free_slots);
585 list_for_each_entry_safe(iter, _iter, &mv_chan->allocated_slots,
588 list_move_tail(&iter->node, &mv_chan->free_slots);
590 list_for_each_entry_safe_reverse(
591 iter, _iter, &mv_chan->free_slots, node) {
592 list_del(&iter->node);
594 mv_chan->slots_allocated--;
597 dev_dbg(mv_chan_to_devp(mv_chan), "%s slots_allocated %d\n",
598 __func__, mv_chan->slots_allocated);
599 spin_unlock_bh(&mv_chan->lock);
602 dev_err(mv_chan_to_devp(mv_chan),
603 "freeing %d in use descriptors!\n", in_use_descs);
607 * mv_xor_status - poll the status of an XOR transaction
608 * @chan: XOR channel handle
609 * @cookie: XOR transaction identifier
610 * @txstate: XOR transactions state holder (or NULL)
612 static enum dma_status mv_xor_status(struct dma_chan *chan,
614 struct dma_tx_state *txstate)
616 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
619 ret = dma_cookie_status(chan, cookie, txstate);
620 if (ret == DMA_COMPLETE)
623 spin_lock_bh(&mv_chan->lock);
624 mv_chan_slot_cleanup(mv_chan);
625 spin_unlock_bh(&mv_chan->lock);
627 return dma_cookie_status(chan, cookie, txstate);
630 static void mv_chan_dump_regs(struct mv_xor_chan *chan)
634 val = readl_relaxed(XOR_CONFIG(chan));
635 dev_err(mv_chan_to_devp(chan), "config 0x%08x\n", val);
637 val = readl_relaxed(XOR_ACTIVATION(chan));
638 dev_err(mv_chan_to_devp(chan), "activation 0x%08x\n", val);
640 val = readl_relaxed(XOR_INTR_CAUSE(chan));
641 dev_err(mv_chan_to_devp(chan), "intr cause 0x%08x\n", val);
643 val = readl_relaxed(XOR_INTR_MASK(chan));
644 dev_err(mv_chan_to_devp(chan), "intr mask 0x%08x\n", val);
646 val = readl_relaxed(XOR_ERROR_CAUSE(chan));
647 dev_err(mv_chan_to_devp(chan), "error cause 0x%08x\n", val);
649 val = readl_relaxed(XOR_ERROR_ADDR(chan));
650 dev_err(mv_chan_to_devp(chan), "error addr 0x%08x\n", val);
653 static void mv_chan_err_interrupt_handler(struct mv_xor_chan *chan,
656 if (intr_cause & XOR_INT_ERR_DECODE) {
657 dev_dbg(mv_chan_to_devp(chan), "ignoring address decode error\n");
661 dev_err(mv_chan_to_devp(chan), "error on chan %d. intr cause 0x%08x\n",
662 chan->idx, intr_cause);
664 mv_chan_dump_regs(chan);
668 static irqreturn_t mv_xor_interrupt_handler(int irq, void *data)
670 struct mv_xor_chan *chan = data;
671 u32 intr_cause = mv_chan_get_intr_cause(chan);
673 dev_dbg(mv_chan_to_devp(chan), "intr cause %x\n", intr_cause);
675 if (intr_cause & XOR_INTR_ERRORS)
676 mv_chan_err_interrupt_handler(chan, intr_cause);
678 tasklet_schedule(&chan->irq_tasklet);
680 mv_chan_clear_eoc_cause(chan);
685 static void mv_xor_issue_pending(struct dma_chan *chan)
687 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
689 if (mv_chan->pending >= MV_XOR_THRESHOLD) {
690 mv_chan->pending = 0;
691 mv_chan_activate(mv_chan);
696 * Perform a transaction to verify the HW works.
699 static int mv_chan_memcpy_self_test(struct mv_xor_chan *mv_chan)
703 dma_addr_t src_dma, dest_dma;
704 struct dma_chan *dma_chan;
706 struct dma_async_tx_descriptor *tx;
707 struct dmaengine_unmap_data *unmap;
710 src = kmalloc(sizeof(u8) * PAGE_SIZE, GFP_KERNEL);
714 dest = kzalloc(sizeof(u8) * PAGE_SIZE, GFP_KERNEL);
720 /* Fill in src buffer */
721 for (i = 0; i < PAGE_SIZE; i++)
722 ((u8 *) src)[i] = (u8)i;
724 dma_chan = &mv_chan->dmachan;
725 if (mv_xor_alloc_chan_resources(dma_chan) < 1) {
730 unmap = dmaengine_get_unmap_data(dma_chan->device->dev, 2, GFP_KERNEL);
736 src_dma = dma_map_page(dma_chan->device->dev, virt_to_page(src), 0,
737 PAGE_SIZE, DMA_TO_DEVICE);
738 unmap->addr[0] = src_dma;
740 ret = dma_mapping_error(dma_chan->device->dev, src_dma);
747 dest_dma = dma_map_page(dma_chan->device->dev, virt_to_page(dest), 0,
748 PAGE_SIZE, DMA_FROM_DEVICE);
749 unmap->addr[1] = dest_dma;
751 ret = dma_mapping_error(dma_chan->device->dev, dest_dma);
757 unmap->len = PAGE_SIZE;
759 tx = mv_xor_prep_dma_memcpy(dma_chan, dest_dma, src_dma,
762 dev_err(dma_chan->device->dev,
763 "Self-test cannot prepare operation, disabling\n");
768 cookie = mv_xor_tx_submit(tx);
769 if (dma_submit_error(cookie)) {
770 dev_err(dma_chan->device->dev,
771 "Self-test submit error, disabling\n");
776 mv_xor_issue_pending(dma_chan);
780 if (mv_xor_status(dma_chan, cookie, NULL) !=
782 dev_err(dma_chan->device->dev,
783 "Self-test copy timed out, disabling\n");
788 dma_sync_single_for_cpu(dma_chan->device->dev, dest_dma,
789 PAGE_SIZE, DMA_FROM_DEVICE);
790 if (memcmp(src, dest, PAGE_SIZE)) {
791 dev_err(dma_chan->device->dev,
792 "Self-test copy failed compare, disabling\n");
798 dmaengine_unmap_put(unmap);
799 mv_xor_free_chan_resources(dma_chan);
806 #define MV_XOR_NUM_SRC_TEST 4 /* must be <= 15 */
808 mv_chan_xor_self_test(struct mv_xor_chan *mv_chan)
812 struct page *xor_srcs[MV_XOR_NUM_SRC_TEST];
813 dma_addr_t dma_srcs[MV_XOR_NUM_SRC_TEST];
815 struct dma_async_tx_descriptor *tx;
816 struct dmaengine_unmap_data *unmap;
817 struct dma_chan *dma_chan;
822 int src_count = MV_XOR_NUM_SRC_TEST;
824 for (src_idx = 0; src_idx < src_count; src_idx++) {
825 xor_srcs[src_idx] = alloc_page(GFP_KERNEL);
826 if (!xor_srcs[src_idx]) {
828 __free_page(xor_srcs[src_idx]);
833 dest = alloc_page(GFP_KERNEL);
836 __free_page(xor_srcs[src_idx]);
840 /* Fill in src buffers */
841 for (src_idx = 0; src_idx < src_count; src_idx++) {
842 u8 *ptr = page_address(xor_srcs[src_idx]);
843 for (i = 0; i < PAGE_SIZE; i++)
844 ptr[i] = (1 << src_idx);
847 for (src_idx = 0; src_idx < src_count; src_idx++)
848 cmp_byte ^= (u8) (1 << src_idx);
850 cmp_word = (cmp_byte << 24) | (cmp_byte << 16) |
851 (cmp_byte << 8) | cmp_byte;
853 memset(page_address(dest), 0, PAGE_SIZE);
855 dma_chan = &mv_chan->dmachan;
856 if (mv_xor_alloc_chan_resources(dma_chan) < 1) {
861 unmap = dmaengine_get_unmap_data(dma_chan->device->dev, src_count + 1,
869 for (i = 0; i < src_count; i++) {
870 unmap->addr[i] = dma_map_page(dma_chan->device->dev, xor_srcs[i],
871 0, PAGE_SIZE, DMA_TO_DEVICE);
872 dma_srcs[i] = unmap->addr[i];
873 ret = dma_mapping_error(dma_chan->device->dev, unmap->addr[i]);
881 unmap->addr[src_count] = dma_map_page(dma_chan->device->dev, dest, 0, PAGE_SIZE,
883 dest_dma = unmap->addr[src_count];
884 ret = dma_mapping_error(dma_chan->device->dev, unmap->addr[src_count]);
890 unmap->len = PAGE_SIZE;
892 tx = mv_xor_prep_dma_xor(dma_chan, dest_dma, dma_srcs,
893 src_count, PAGE_SIZE, 0);
895 dev_err(dma_chan->device->dev,
896 "Self-test cannot prepare operation, disabling\n");
901 cookie = mv_xor_tx_submit(tx);
902 if (dma_submit_error(cookie)) {
903 dev_err(dma_chan->device->dev,
904 "Self-test submit error, disabling\n");
909 mv_xor_issue_pending(dma_chan);
913 if (mv_xor_status(dma_chan, cookie, NULL) !=
915 dev_err(dma_chan->device->dev,
916 "Self-test xor timed out, disabling\n");
921 dma_sync_single_for_cpu(dma_chan->device->dev, dest_dma,
922 PAGE_SIZE, DMA_FROM_DEVICE);
923 for (i = 0; i < (PAGE_SIZE / sizeof(u32)); i++) {
924 u32 *ptr = page_address(dest);
925 if (ptr[i] != cmp_word) {
926 dev_err(dma_chan->device->dev,
927 "Self-test xor failed compare, disabling. index %d, data %x, expected %x\n",
928 i, ptr[i], cmp_word);
935 dmaengine_unmap_put(unmap);
936 mv_xor_free_chan_resources(dma_chan);
940 __free_page(xor_srcs[src_idx]);
945 static int mv_xor_channel_remove(struct mv_xor_chan *mv_chan)
947 struct dma_chan *chan, *_chan;
948 struct device *dev = mv_chan->dmadev.dev;
950 dma_async_device_unregister(&mv_chan->dmadev);
952 dma_free_coherent(dev, MV_XOR_POOL_SIZE,
953 mv_chan->dma_desc_pool_virt, mv_chan->dma_desc_pool);
954 dma_unmap_single(dev, mv_chan->dummy_src_addr,
955 MV_XOR_MIN_BYTE_COUNT, DMA_FROM_DEVICE);
956 dma_unmap_single(dev, mv_chan->dummy_dst_addr,
957 MV_XOR_MIN_BYTE_COUNT, DMA_TO_DEVICE);
959 list_for_each_entry_safe(chan, _chan, &mv_chan->dmadev.channels,
961 list_del(&chan->device_node);
964 free_irq(mv_chan->irq, mv_chan);
969 static struct mv_xor_chan *
970 mv_xor_channel_add(struct mv_xor_device *xordev,
971 struct platform_device *pdev,
972 int idx, dma_cap_mask_t cap_mask, int irq, int op_in_desc)
975 struct mv_xor_chan *mv_chan;
976 struct dma_device *dma_dev;
978 mv_chan = devm_kzalloc(&pdev->dev, sizeof(*mv_chan), GFP_KERNEL);
980 return ERR_PTR(-ENOMEM);
984 mv_chan->op_in_desc = op_in_desc;
986 dma_dev = &mv_chan->dmadev;
989 * These source and destination dummy buffers are used to implement
990 * a DMA_INTERRUPT operation as a minimum-sized XOR operation.
991 * Hence, we only need to map the buffers at initialization-time.
993 mv_chan->dummy_src_addr = dma_map_single(dma_dev->dev,
994 mv_chan->dummy_src, MV_XOR_MIN_BYTE_COUNT, DMA_FROM_DEVICE);
995 mv_chan->dummy_dst_addr = dma_map_single(dma_dev->dev,
996 mv_chan->dummy_dst, MV_XOR_MIN_BYTE_COUNT, DMA_TO_DEVICE);
998 /* allocate coherent memory for hardware descriptors
999 * note: writecombine gives slightly better performance, but
1000 * requires that we explicitly flush the writes
1002 mv_chan->dma_desc_pool_virt =
1003 dma_alloc_writecombine(&pdev->dev, MV_XOR_POOL_SIZE,
1004 &mv_chan->dma_desc_pool, GFP_KERNEL);
1005 if (!mv_chan->dma_desc_pool_virt)
1006 return ERR_PTR(-ENOMEM);
1008 /* discover transaction capabilites from the platform data */
1009 dma_dev->cap_mask = cap_mask;
1011 INIT_LIST_HEAD(&dma_dev->channels);
1013 /* set base routines */
1014 dma_dev->device_alloc_chan_resources = mv_xor_alloc_chan_resources;
1015 dma_dev->device_free_chan_resources = mv_xor_free_chan_resources;
1016 dma_dev->device_tx_status = mv_xor_status;
1017 dma_dev->device_issue_pending = mv_xor_issue_pending;
1018 dma_dev->dev = &pdev->dev;
1020 /* set prep routines based on capability */
1021 if (dma_has_cap(DMA_INTERRUPT, dma_dev->cap_mask))
1022 dma_dev->device_prep_dma_interrupt = mv_xor_prep_dma_interrupt;
1023 if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask))
1024 dma_dev->device_prep_dma_memcpy = mv_xor_prep_dma_memcpy;
1025 if (dma_has_cap(DMA_XOR, dma_dev->cap_mask)) {
1026 dma_dev->max_xor = 8;
1027 dma_dev->device_prep_dma_xor = mv_xor_prep_dma_xor;
1030 mv_chan->mmr_base = xordev->xor_base;
1031 mv_chan->mmr_high_base = xordev->xor_high_base;
1032 tasklet_init(&mv_chan->irq_tasklet, mv_xor_tasklet, (unsigned long)
1035 /* clear errors before enabling interrupts */
1036 mv_chan_clear_err_status(mv_chan);
1038 ret = request_irq(mv_chan->irq, mv_xor_interrupt_handler,
1039 0, dev_name(&pdev->dev), mv_chan);
1043 mv_chan_unmask_interrupts(mv_chan);
1045 if (mv_chan->op_in_desc == XOR_MODE_IN_DESC)
1046 mv_chan_set_mode_to_desc(mv_chan);
1048 mv_chan_set_mode(mv_chan, DMA_XOR);
1050 spin_lock_init(&mv_chan->lock);
1051 INIT_LIST_HEAD(&mv_chan->chain);
1052 INIT_LIST_HEAD(&mv_chan->completed_slots);
1053 INIT_LIST_HEAD(&mv_chan->free_slots);
1054 INIT_LIST_HEAD(&mv_chan->allocated_slots);
1055 mv_chan->dmachan.device = dma_dev;
1056 dma_cookie_init(&mv_chan->dmachan);
1058 list_add_tail(&mv_chan->dmachan.device_node, &dma_dev->channels);
1060 if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask)) {
1061 ret = mv_chan_memcpy_self_test(mv_chan);
1062 dev_dbg(&pdev->dev, "memcpy self test returned %d\n", ret);
1067 if (dma_has_cap(DMA_XOR, dma_dev->cap_mask)) {
1068 ret = mv_chan_xor_self_test(mv_chan);
1069 dev_dbg(&pdev->dev, "xor self test returned %d\n", ret);
1074 dev_info(&pdev->dev, "Marvell XOR (%s): ( %s%s%s)\n",
1075 mv_chan->op_in_desc ? "Descriptor Mode" : "Registers Mode",
1076 dma_has_cap(DMA_XOR, dma_dev->cap_mask) ? "xor " : "",
1077 dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask) ? "cpy " : "",
1078 dma_has_cap(DMA_INTERRUPT, dma_dev->cap_mask) ? "intr " : "");
1080 dma_async_device_register(dma_dev);
1084 free_irq(mv_chan->irq, mv_chan);
1086 dma_free_coherent(&pdev->dev, MV_XOR_POOL_SIZE,
1087 mv_chan->dma_desc_pool_virt, mv_chan->dma_desc_pool);
1088 return ERR_PTR(ret);
1092 mv_xor_conf_mbus_windows(struct mv_xor_device *xordev,
1093 const struct mbus_dram_target_info *dram)
1095 void __iomem *base = xordev->xor_high_base;
1099 for (i = 0; i < 8; i++) {
1100 writel(0, base + WINDOW_BASE(i));
1101 writel(0, base + WINDOW_SIZE(i));
1103 writel(0, base + WINDOW_REMAP_HIGH(i));
1106 for (i = 0; i < dram->num_cs; i++) {
1107 const struct mbus_dram_window *cs = dram->cs + i;
1109 writel((cs->base & 0xffff0000) |
1110 (cs->mbus_attr << 8) |
1111 dram->mbus_dram_target_id, base + WINDOW_BASE(i));
1112 writel((cs->size - 1) & 0xffff0000, base + WINDOW_SIZE(i));
1114 win_enable |= (1 << i);
1115 win_enable |= 3 << (16 + (2 * i));
1118 writel(win_enable, base + WINDOW_BAR_ENABLE(0));
1119 writel(win_enable, base + WINDOW_BAR_ENABLE(1));
1120 writel(0, base + WINDOW_OVERRIDE_CTRL(0));
1121 writel(0, base + WINDOW_OVERRIDE_CTRL(1));
1124 static const struct of_device_id mv_xor_dt_ids[] = {
1125 { .compatible = "marvell,orion-xor", .data = (void *)XOR_MODE_IN_REG },
1126 { .compatible = "marvell,armada-380-xor", .data = (void *)XOR_MODE_IN_DESC },
1130 static unsigned int mv_xor_engine_count;
1132 static int mv_xor_probe(struct platform_device *pdev)
1134 const struct mbus_dram_target_info *dram;
1135 struct mv_xor_device *xordev;
1136 struct mv_xor_platform_data *pdata = dev_get_platdata(&pdev->dev);
1137 struct resource *res;
1138 unsigned int max_engines, max_channels;
1142 dev_notice(&pdev->dev, "Marvell shared XOR driver\n");
1144 xordev = devm_kzalloc(&pdev->dev, sizeof(*xordev), GFP_KERNEL);
1148 res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1152 xordev->xor_base = devm_ioremap(&pdev->dev, res->start,
1153 resource_size(res));
1154 if (!xordev->xor_base)
1157 res = platform_get_resource(pdev, IORESOURCE_MEM, 1);
1161 xordev->xor_high_base = devm_ioremap(&pdev->dev, res->start,
1162 resource_size(res));
1163 if (!xordev->xor_high_base)
1166 platform_set_drvdata(pdev, xordev);
1169 * (Re-)program MBUS remapping windows if we are asked to.
1171 dram = mv_mbus_dram_info();
1173 mv_xor_conf_mbus_windows(xordev, dram);
1175 /* Not all platforms can gate the clock, so it is not
1176 * an error if the clock does not exists.
1178 xordev->clk = clk_get(&pdev->dev, NULL);
1179 if (!IS_ERR(xordev->clk))
1180 clk_prepare_enable(xordev->clk);
1183 * We don't want to have more than one channel per CPU in
1184 * order for async_tx to perform well. So we limit the number
1185 * of engines and channels so that we take into account this
1186 * constraint. Note that we also want to use channels from
1187 * separate engines when possible.
1189 max_engines = num_present_cpus();
1190 max_channels = min_t(unsigned int,
1191 MV_XOR_MAX_CHANNELS,
1192 DIV_ROUND_UP(num_present_cpus(), 2));
1194 if (mv_xor_engine_count >= max_engines)
1197 if (pdev->dev.of_node) {
1198 struct device_node *np;
1200 const struct of_device_id *of_id =
1201 of_match_device(mv_xor_dt_ids,
1204 for_each_child_of_node(pdev->dev.of_node, np) {
1205 struct mv_xor_chan *chan;
1206 dma_cap_mask_t cap_mask;
1208 op_in_desc = (int)of_id->data;
1210 if (i >= max_channels)
1213 dma_cap_zero(cap_mask);
1214 dma_cap_set(DMA_MEMCPY, cap_mask);
1215 dma_cap_set(DMA_XOR, cap_mask);
1216 dma_cap_set(DMA_INTERRUPT, cap_mask);
1218 irq = irq_of_parse_and_map(np, 0);
1221 goto err_channel_add;
1224 chan = mv_xor_channel_add(xordev, pdev, i,
1225 cap_mask, irq, op_in_desc);
1227 ret = PTR_ERR(chan);
1228 irq_dispose_mapping(irq);
1229 goto err_channel_add;
1232 xordev->channels[i] = chan;
1235 } else if (pdata && pdata->channels) {
1236 for (i = 0; i < max_channels; i++) {
1237 struct mv_xor_channel_data *cd;
1238 struct mv_xor_chan *chan;
1241 cd = &pdata->channels[i];
1244 goto err_channel_add;
1247 irq = platform_get_irq(pdev, i);
1250 goto err_channel_add;
1253 chan = mv_xor_channel_add(xordev, pdev, i,
1257 ret = PTR_ERR(chan);
1258 goto err_channel_add;
1261 xordev->channels[i] = chan;
1268 for (i = 0; i < MV_XOR_MAX_CHANNELS; i++)
1269 if (xordev->channels[i]) {
1270 mv_xor_channel_remove(xordev->channels[i]);
1271 if (pdev->dev.of_node)
1272 irq_dispose_mapping(xordev->channels[i]->irq);
1275 if (!IS_ERR(xordev->clk)) {
1276 clk_disable_unprepare(xordev->clk);
1277 clk_put(xordev->clk);
1283 static struct platform_driver mv_xor_driver = {
1284 .probe = mv_xor_probe,
1286 .name = MV_XOR_NAME,
1287 .of_match_table = of_match_ptr(mv_xor_dt_ids),
1292 static int __init mv_xor_init(void)
1294 return platform_driver_register(&mv_xor_driver);
1296 device_initcall(mv_xor_init);
1299 MODULE_AUTHOR("Saeed Bishara <saeed@marvell.com>");
1300 MODULE_DESCRIPTION("DMA engine driver for Marvell's XOR engine");
1301 MODULE_LICENSE("GPL");