]> git.kernelconcepts.de Git - karo-tx-linux.git/blobdiff - drivers/dma/ioat/dma.c
Merge branch 'dmaengine' into async-tx-next
[karo-tx-linux.git] / drivers / dma / ioat / dma.c
index edf4f5e5de735d8fbf97ee76a2a24299a108abee..c524d36d3c2e199db83e5b2a9a8a7b7bf1939a21 100644 (file)
@@ -99,23 +99,26 @@ static void ioat1_cleanup_tasklet(unsigned long data);
 /* common channel initialization */
 void ioat_init_channel(struct ioatdma_device *device,
                       struct ioat_chan_common *chan, int idx,
-                      work_func_t work_fn, void (*tasklet)(unsigned long),
-                      unsigned long tasklet_data)
+                      void (*timer_fn)(unsigned long),
+                      void (*tasklet)(unsigned long),
+                      unsigned long ioat)
 {
        struct dma_device *dma = &device->common;
 
        chan->device = device;
        chan->reg_base = device->reg_base + (0x80 * (idx + 1));
-       INIT_DELAYED_WORK(&chan->work, work_fn);
        spin_lock_init(&chan->cleanup_lock);
        chan->common.device = dma;
        list_add_tail(&chan->common.device_node, &dma->channels);
        device->idx[idx] = chan;
-       tasklet_init(&chan->cleanup_task, tasklet, tasklet_data);
+       init_timer(&chan->timer);
+       chan->timer.function = timer_fn;
+       chan->timer.data = ioat;
+       tasklet_init(&chan->cleanup_task, tasklet, ioat);
        tasklet_disable(&chan->cleanup_task);
 }
 
-static void ioat1_reset_part2(struct work_struct *work);
+static void ioat1_timer_event(unsigned long data);
 
 /**
  * ioat1_dma_enumerate_channels - find and initialize the device's channels
@@ -132,7 +135,14 @@ static int ioat1_enumerate_channels(struct ioatdma_device *device)
 
        INIT_LIST_HEAD(&dma->channels);
        dma->chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);
+       dma->chancnt &= 0x1f; /* bits [4:0] valid */
+       if (dma->chancnt > ARRAY_SIZE(device->idx)) {
+               dev_warn(dev, "(%d) exceeds max supported channels (%zu)\n",
+                        dma->chancnt, ARRAY_SIZE(device->idx));
+               dma->chancnt = ARRAY_SIZE(device->idx);
+       }
        xfercap_scale = readb(device->reg_base + IOAT_XFERCAP_OFFSET);
+       xfercap_scale &= 0x1f; /* bits [4:0] valid */
        xfercap = (xfercap_scale == 0 ? -1 : (1UL << xfercap_scale));
        dev_dbg(dev, "%s: xfercap = %d\n", __func__, xfercap);
 
@@ -146,7 +156,7 @@ static int ioat1_enumerate_channels(struct ioatdma_device *device)
                        break;
 
                ioat_init_channel(device, &ioat->base, i,
-                                 ioat1_reset_part2,
+                                 ioat1_timer_event,
                                  ioat1_cleanup_tasklet,
                                  (unsigned long) ioat);
                ioat->xfercap = xfercap;
@@ -185,62 +195,6 @@ static void ioat1_dma_memcpy_issue_pending(struct dma_chan *chan)
        }
 }
 
-/**
- * ioat1_reset_part2 - reinit the channel after a reset
- */
-static void ioat1_reset_part2(struct work_struct *work)
-{
-       struct ioat_chan_common *chan;
-       struct ioat_dma_chan *ioat;
-       struct ioat_desc_sw *desc;
-       int dmacount;
-       bool start_null = false;
-
-       chan = container_of(work, struct ioat_chan_common, work.work);
-       ioat = container_of(chan, struct ioat_dma_chan, base);
-       spin_lock_bh(&chan->cleanup_lock);
-       spin_lock_bh(&ioat->desc_lock);
-
-       chan->completion_virt->low = 0;
-       chan->completion_virt->high = 0;
-       ioat->pending = 0;
-
-       /* count the descriptors waiting */
-       dmacount = 0;
-       if (ioat->used_desc.prev) {
-               desc = to_ioat_desc(ioat->used_desc.prev);
-               do {
-                       dmacount++;
-                       desc = to_ioat_desc(desc->node.next);
-               } while (&desc->node != ioat->used_desc.next);
-       }
-
-       if (dmacount) {
-               /*
-                * write the new starting descriptor address
-                * this puts channel engine into ARMED state
-                */
-               desc = to_ioat_desc(ioat->used_desc.prev);
-               writel(((u64) desc->txd.phys) & 0x00000000FFFFFFFF,
-                      chan->reg_base + IOAT1_CHAINADDR_OFFSET_LOW);
-               writel(((u64) desc->txd.phys) >> 32,
-                      chan->reg_base + IOAT1_CHAINADDR_OFFSET_HIGH);
-
-               writeb(IOAT_CHANCMD_START, chan->reg_base
-                       + IOAT_CHANCMD_OFFSET(chan->device->version));
-       } else
-               start_null = true;
-       spin_unlock_bh(&ioat->desc_lock);
-       spin_unlock_bh(&chan->cleanup_lock);
-
-       dev_err(to_dev(chan),
-               "chan%d reset - %d descs waiting, %d total desc\n",
-               chan_num(chan), dmacount, ioat->desccount);
-
-       if (start_null)
-               ioat1_dma_start_null_desc(ioat);
-}
-
 /**
  * ioat1_reset_channel - restart a channel
  * @ioat: IOAT DMA channel handle
@@ -251,13 +205,9 @@ static void ioat1_reset_channel(struct ioat_dma_chan *ioat)
        void __iomem *reg_base = chan->reg_base;
        u32 chansts, chanerr;
 
-       if (!ioat->used_desc.prev)
-               return;
-
-       dev_dbg(to_dev(chan), "%s\n", __func__);
+       dev_warn(to_dev(chan), "reset\n");
        chanerr = readl(reg_base + IOAT_CHANERR_OFFSET);
-       chansts = (chan->completion_virt->low
-                                       & IOAT_CHANSTS_DMA_TRANSFER_STATUS);
+       chansts = *chan->completion & IOAT_CHANSTS_STATUS;
        if (chanerr) {
                dev_err(to_dev(chan),
                        "chan%d, CHANSTS = 0x%08x CHANERR = 0x%04x, clearing\n",
@@ -273,100 +223,11 @@ static void ioat1_reset_channel(struct ioat_dma_chan *ioat)
         * while we're waiting.
         */
 
-       spin_lock_bh(&ioat->desc_lock);
        ioat->pending = INT_MIN;
        writeb(IOAT_CHANCMD_RESET,
               reg_base + IOAT_CHANCMD_OFFSET(chan->device->version));
-       spin_unlock_bh(&ioat->desc_lock);
-
-       /* schedule the 2nd half instead of sleeping a long time */
-       schedule_delayed_work(&chan->work, RESET_DELAY);
-}
-
-/**
- * ioat1_chan_watchdog - watch for stuck channels
- */
-static void ioat1_chan_watchdog(struct work_struct *work)
-{
-       struct ioatdma_device *device =
-               container_of(work, struct ioatdma_device, work.work);
-       struct ioat_dma_chan *ioat;
-       struct ioat_chan_common *chan;
-       int i;
-
-       union {
-               u64 full;
-               struct {
-                       u32 low;
-                       u32 high;
-               };
-       } completion_hw;
-       unsigned long compl_desc_addr_hw;
-
-       for (i = 0; i < device->common.chancnt; i++) {
-               chan = ioat_chan_by_index(device, i);
-               ioat = container_of(chan, struct ioat_dma_chan, base);
-
-               if (/* have we started processing anything yet */
-                   chan->last_completion
-                   /* have we completed any since last watchdog cycle? */
-                   && (chan->last_completion == chan->watchdog_completion)
-                   /* has TCP stuck on one cookie since last watchdog? */
-                   && (chan->watchdog_tcp_cookie == chan->watchdog_last_tcp_cookie)
-                   && (chan->watchdog_tcp_cookie != chan->completed_cookie)
-                   /* is there something in the chain to be processed? */
-                   /* CB1 chain always has at least the last one processed */
-                   && (ioat->used_desc.prev != ioat->used_desc.next)
-                   && ioat->pending == 0) {
-
-                       /*
-                        * check CHANSTS register for completed
-                        * descriptor address.
-                        * if it is different than completion writeback,
-                        * it is not zero
-                        * and it has changed since the last watchdog
-                        *     we can assume that channel
-                        *     is still working correctly
-                        *     and the problem is in completion writeback.
-                        *     update completion writeback
-                        *     with actual CHANSTS value
-                        * else
-                        *     try resetting the channel
-                        */
-
-                       completion_hw.low = readl(chan->reg_base +
-                               IOAT_CHANSTS_OFFSET_LOW(chan->device->version));
-                       completion_hw.high = readl(chan->reg_base +
-                               IOAT_CHANSTS_OFFSET_HIGH(chan->device->version));
-#if (BITS_PER_LONG == 64)
-                       compl_desc_addr_hw =
-                               completion_hw.full
-                               & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
-#else
-                       compl_desc_addr_hw =
-                               completion_hw.low & IOAT_LOW_COMPLETION_MASK;
-#endif
-
-                       if ((compl_desc_addr_hw != 0)
-                          && (compl_desc_addr_hw != chan->watchdog_completion)
-                          && (compl_desc_addr_hw != chan->last_compl_desc_addr_hw)) {
-                               chan->last_compl_desc_addr_hw = compl_desc_addr_hw;
-                               chan->completion_virt->low = completion_hw.low;
-                               chan->completion_virt->high = completion_hw.high;
-                       } else {
-                               ioat1_reset_channel(ioat);
-                               chan->watchdog_completion = 0;
-                               chan->last_compl_desc_addr_hw = 0;
-                       }
-               } else {
-                       chan->last_compl_desc_addr_hw = 0;
-                       chan->watchdog_completion = chan->last_completion;
-               }
-
-               chan->watchdog_last_tcp_cookie = chan->watchdog_tcp_cookie;
-       }
-
-       schedule_delayed_work(&device->work, WATCHDOG_DELAY);
+       set_bit(IOAT_RESET_PENDING, &chan->state);
+       mod_timer(&chan->timer, jiffies + RESET_DELAY);
 }
 
 static dma_cookie_t ioat1_tx_submit(struct dma_async_tx_descriptor *tx)
@@ -374,6 +235,7 @@ static dma_cookie_t ioat1_tx_submit(struct dma_async_tx_descriptor *tx)
        struct dma_chan *c = tx->chan;
        struct ioat_dma_chan *ioat = to_ioat_chan(c);
        struct ioat_desc_sw *desc = tx_to_ioat_desc(tx);
+       struct ioat_chan_common *chan = &ioat->base;
        struct ioat_desc_sw *first;
        struct ioat_desc_sw *chain_tail;
        dma_cookie_t cookie;
@@ -389,16 +251,20 @@ static dma_cookie_t ioat1_tx_submit(struct dma_async_tx_descriptor *tx)
        dev_dbg(to_dev(&ioat->base), "%s: cookie: %d\n", __func__, cookie);
 
        /* write address into NextDescriptor field of last desc in chain */
-       first = to_ioat_desc(tx->tx_list.next);
+       first = to_ioat_desc(desc->tx_list.next);
        chain_tail = to_ioat_desc(ioat->used_desc.prev);
        /* make descriptor updates globally visible before chaining */
        wmb();
        chain_tail->hw->next = first->txd.phys;
-       list_splice_tail_init(&tx->tx_list, &ioat->used_desc);
+       list_splice_tail_init(&desc->tx_list, &ioat->used_desc);
        dump_desc_dbg(ioat, chain_tail);
        dump_desc_dbg(ioat, first);
 
-       ioat->pending += desc->tx_cnt;
+       if (!test_and_set_bit(IOAT_COMPLETION_PENDING, &chan->state))
+               mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
+
+       ioat->active += desc->hw->tx_cnt;
+       ioat->pending += desc->hw->tx_cnt;
        if (ioat->pending >= ioat_pending_level)
                __ioat1_dma_memcpy_issue_pending(ioat);
        spin_unlock_bh(&ioat->desc_lock);
@@ -432,6 +298,7 @@ ioat_dma_alloc_descriptor(struct ioat_dma_chan *ioat, gfp_t flags)
 
        memset(desc, 0, sizeof(*desc));
 
+       INIT_LIST_HEAD(&desc_sw->tx_list);
        dma_async_tx_descriptor_init(&desc_sw->txd, &ioat->base.common);
        desc_sw->txd.tx_submit = ioat1_tx_submit;
        desc_sw->hw = desc;
@@ -454,7 +321,6 @@ static int ioat1_dma_alloc_chan_resources(struct dma_chan *c)
        struct ioat_dma_chan *ioat = to_ioat_chan(c);
        struct ioat_chan_common *chan = &ioat->base;
        struct ioat_desc_sw *desc;
-       u16 chanctrl;
        u32 chanerr;
        int i;
        LIST_HEAD(tmp_list);
@@ -464,10 +330,7 @@ static int ioat1_dma_alloc_chan_resources(struct dma_chan *c)
                return ioat->desccount;
 
        /* Setup register to interrupt and write completion status on error */
-       chanctrl = IOAT_CHANCTRL_ERR_INT_EN |
-               IOAT_CHANCTRL_ANY_ERR_ABORT_EN |
-               IOAT_CHANCTRL_ERR_COMPLETION_EN;
-       writew(chanctrl, chan->reg_base + IOAT_CHANCTRL_OFFSET);
+       writew(IOAT_CHANCTRL_RUN, chan->reg_base + IOAT_CHANCTRL_OFFSET);
 
        chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
        if (chanerr) {
@@ -492,14 +355,12 @@ static int ioat1_dma_alloc_chan_resources(struct dma_chan *c)
 
        /* allocate a completion writeback area */
        /* doing 2 32bit writes to mmio since 1 64b write doesn't work */
-       chan->completion_virt = pci_pool_alloc(chan->device->completion_pool,
-                                              GFP_KERNEL,
-                                              &chan->completion_addr);
-       memset(chan->completion_virt, 0,
-              sizeof(*chan->completion_virt));
-       writel(((u64) chan->completion_addr) & 0x00000000FFFFFFFF,
+       chan->completion = pci_pool_alloc(chan->device->completion_pool,
+                                         GFP_KERNEL, &chan->completion_dma);
+       memset(chan->completion, 0, sizeof(*chan->completion));
+       writel(((u64) chan->completion_dma) & 0x00000000FFFFFFFF,
               chan->reg_base + IOAT_CHANCMP_OFFSET_LOW);
-       writel(((u64) chan->completion_addr) >> 32,
+       writel(((u64) chan->completion_dma) >> 32,
               chan->reg_base + IOAT_CHANCMP_OFFSET_HIGH);
 
        tasklet_enable(&chan->cleanup_task);
@@ -528,6 +389,7 @@ static void ioat1_dma_free_chan_resources(struct dma_chan *c)
                return;
 
        tasklet_disable(&chan->cleanup_task);
+       del_timer_sync(&chan->timer);
        ioat1_cleanup(ioat);
 
        /* Delay 100ms after reset to allow internal DMA logic to quiesce
@@ -558,18 +420,16 @@ static void ioat1_dma_free_chan_resources(struct dma_chan *c)
        spin_unlock_bh(&ioat->desc_lock);
 
        pci_pool_free(ioatdma_device->completion_pool,
-                     chan->completion_virt,
-                     chan->completion_addr);
+                     chan->completion,
+                     chan->completion_dma);
 
        /* one is ok since we left it on there on purpose */
        if (in_use_descs > 1)
                dev_err(to_dev(chan), "Freeing %d in use descriptors!\n",
                        in_use_descs - 1);
 
-       chan->last_completion = chan->completion_addr = 0;
-       chan->watchdog_completion = 0;
-       chan->last_compl_desc_addr_hw = 0;
-       chan->watchdog_tcp_cookie = chan->watchdog_last_tcp_cookie = 0;
+       chan->last_completion = 0;
+       chan->completion_dma = 0;
        ioat->pending = 0;
        ioat->desccount = 0;
 }
@@ -662,11 +522,11 @@ ioat1_dma_prep_memcpy(struct dma_chan *c, dma_addr_t dma_dest,
        spin_unlock_bh(&ioat->desc_lock);
 
        desc->txd.flags = flags;
-       desc->tx_cnt = tx_cnt;
        desc->len = total_len;
-       list_splice(&chain, &desc->txd.tx_list);
+       list_splice(&chain, &desc->tx_list);
        hw->ctl_f.int_en = !!(flags & DMA_PREP_INTERRUPT);
        hw->ctl_f.compl_write = 1;
+       hw->tx_cnt = tx_cnt;
        dump_desc_dbg(ioat, desc);
 
        return &desc->txd;
@@ -675,22 +535,11 @@ ioat1_dma_prep_memcpy(struct dma_chan *c, dma_addr_t dma_dest,
 static void ioat1_cleanup_tasklet(unsigned long data)
 {
        struct ioat_dma_chan *chan = (void *)data;
-       ioat1_cleanup(chan);
-       writew(IOAT_CHANCTRL_INT_DISABLE,
-              chan->base.reg_base + IOAT_CHANCTRL_OFFSET);
-}
 
-static void ioat_unmap(struct pci_dev *pdev, dma_addr_t addr, size_t len,
-                      int direction, enum dma_ctrl_flags flags, bool dst)
-{
-       if ((dst && (flags & DMA_COMPL_DEST_UNMAP_SINGLE)) ||
-           (!dst && (flags & DMA_COMPL_SRC_UNMAP_SINGLE)))
-               pci_unmap_single(pdev, addr, len, direction);
-       else
-               pci_unmap_page(pdev, addr, len, direction);
+       ioat1_cleanup(chan);
+       writew(IOAT_CHANCTRL_RUN, chan->base.reg_base + IOAT_CHANCTRL_OFFSET);
 }
 
-
 void ioat_dma_unmap(struct ioat_chan_common *chan, enum dma_ctrl_flags flags,
                    size_t len, struct ioat_dma_descriptor *hw)
 {
@@ -709,28 +558,18 @@ void ioat_dma_unmap(struct ioat_chan_common *chan, enum dma_ctrl_flags flags,
 unsigned long ioat_get_current_completion(struct ioat_chan_common *chan)
 {
        unsigned long phys_complete;
+       u64 completion;
 
-       /* The completion writeback can happen at any time,
-          so reads by the driver need to be atomic operations
-          The descriptor physical addresses are limited to 32-bits
-          when the CPU can only do a 32-bit mov */
-
-#if (BITS_PER_LONG == 64)
-       phys_complete =
-               chan->completion_virt->full
-               & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
-#else
-       phys_complete = chan->completion_virt->low & IOAT_LOW_COMPLETION_MASK;
-#endif
+       completion = *chan->completion;
+       phys_complete = ioat_chansts_to_addr(completion);
 
        dev_dbg(to_dev(chan), "%s: phys_complete: %#llx\n", __func__,
                (unsigned long long) phys_complete);
 
-       if ((chan->completion_virt->full
-               & IOAT_CHANSTS_DMA_TRANSFER_STATUS) ==
-                               IOAT_CHANSTS_DMA_TRANSFER_STATUS_HALTED) {
+       if (is_ioat_halted(completion)) {
+               u32 chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
                dev_err(to_dev(chan), "Channel halted, chanerr = %x\n",
-                       readl(chan->reg_base + IOAT_CHANERR_OFFSET));
+                       chanerr);
 
                /* TODO do something to salvage the situation */
        }
@@ -738,48 +577,31 @@ unsigned long ioat_get_current_completion(struct ioat_chan_common *chan)
        return phys_complete;
 }
 
-/**
- * ioat1_cleanup - cleanup up finished descriptors
- * @chan: ioat channel to be cleaned up
- */
-static void ioat1_cleanup(struct ioat_dma_chan *ioat)
+bool ioat_cleanup_preamble(struct ioat_chan_common *chan,
+                          unsigned long *phys_complete)
 {
-       struct ioat_chan_common *chan = &ioat->base;
-       unsigned long phys_complete;
-       struct ioat_desc_sw *desc, *_desc;
-       dma_cookie_t cookie = 0;
-       struct dma_async_tx_descriptor *tx;
-
-       prefetch(chan->completion_virt);
+       *phys_complete = ioat_get_current_completion(chan);
+       if (*phys_complete == chan->last_completion)
+               return false;
+       clear_bit(IOAT_COMPLETION_ACK, &chan->state);
+       mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
 
-       if (!spin_trylock_bh(&chan->cleanup_lock))
-               return;
-
-       phys_complete = ioat_get_current_completion(chan);
-       if (phys_complete == chan->last_completion) {
-               spin_unlock_bh(&chan->cleanup_lock);
-               /*
-                * perhaps we're stuck so hard that the watchdog can't go off?
-                * try to catch it after 2 seconds
-                */
-               if (time_after(jiffies,
-                              chan->last_completion_time + HZ*WATCHDOG_DELAY)) {
-                       ioat1_chan_watchdog(&(chan->device->work.work));
-                       chan->last_completion_time = jiffies;
-               }
-               return;
-       }
-       chan->last_completion_time = jiffies;
+       return true;
+}
 
-       cookie = 0;
-       if (!spin_trylock_bh(&ioat->desc_lock)) {
-               spin_unlock_bh(&chan->cleanup_lock);
-               return;
-       }
+static void __cleanup(struct ioat_dma_chan *ioat, unsigned long phys_complete)
+{
+       struct ioat_chan_common *chan = &ioat->base;
+       struct list_head *_desc, *n;
+       struct dma_async_tx_descriptor *tx;
 
        dev_dbg(to_dev(chan), "%s: phys_complete: %lx\n",
                 __func__, phys_complete);
-       list_for_each_entry_safe(desc, _desc, &ioat->used_desc, node) {
+       list_for_each_safe(_desc, n, &ioat->used_desc) {
+               struct ioat_desc_sw *desc;
+
+               prefetch(n);
+               desc = list_entry(_desc, typeof(*desc), node);
                tx = &desc->txd;
                /*
                 * Incoming DMA requests may use multiple descriptors,
@@ -788,8 +610,10 @@ static void ioat1_cleanup(struct ioat_dma_chan *ioat)
                 */
                dump_desc_dbg(ioat, desc);
                if (tx->cookie) {
-                       cookie = tx->cookie;
+                       chan->completed_cookie = tx->cookie;
+                       tx->cookie = 0;
                        ioat_dma_unmap(chan, tx->flags, desc->len, desc->hw);
+                       ioat->active -= desc->hw->tx_cnt;
                        if (tx->callback) {
                                tx->callback(tx->callback_param);
                                tx->callback = NULL;
@@ -803,27 +627,110 @@ static void ioat1_cleanup(struct ioat_dma_chan *ioat)
                         */
                        if (async_tx_test_ack(tx))
                                list_move_tail(&desc->node, &ioat->free_desc);
-                       else
-                               tx->cookie = 0;
                } else {
                        /*
                         * last used desc. Do not remove, so we can
-                        * append from it, but don't look at it next
-                        * time, either
+                        * append from it.
                         */
-                       tx->cookie = 0;
+
+                       /* if nothing else is pending, cancel the
+                        * completion timeout
+                        */
+                       if (n == &ioat->used_desc) {
+                               dev_dbg(to_dev(chan),
+                                       "%s cancel completion timeout\n",
+                                       __func__);
+                               clear_bit(IOAT_COMPLETION_PENDING, &chan->state);
+                       }
 
                        /* TODO check status bits? */
                        break;
                }
        }
 
+       chan->last_completion = phys_complete;
+}
+
+/**
+ * ioat1_cleanup - cleanup up finished descriptors
+ * @chan: ioat channel to be cleaned up
+ *
+ * To prevent lock contention we defer cleanup when the locks are
+ * contended with a terminal timeout that forces cleanup and catches
+ * completion notification errors.
+ */
+static void ioat1_cleanup(struct ioat_dma_chan *ioat)
+{
+       struct ioat_chan_common *chan = &ioat->base;
+       unsigned long phys_complete;
+
+       prefetch(chan->completion);
+
+       if (!spin_trylock_bh(&chan->cleanup_lock))
+               return;
+
+       if (!ioat_cleanup_preamble(chan, &phys_complete)) {
+               spin_unlock_bh(&chan->cleanup_lock);
+               return;
+       }
+
+       if (!spin_trylock_bh(&ioat->desc_lock)) {
+               spin_unlock_bh(&chan->cleanup_lock);
+               return;
+       }
+
+       __cleanup(ioat, phys_complete);
+
        spin_unlock_bh(&ioat->desc_lock);
+       spin_unlock_bh(&chan->cleanup_lock);
+}
 
-       chan->last_completion = phys_complete;
-       if (cookie != 0)
-               chan->completed_cookie = cookie;
+static void ioat1_timer_event(unsigned long data)
+{
+       struct ioat_dma_chan *ioat = (void *) data;
+       struct ioat_chan_common *chan = &ioat->base;
+
+       dev_dbg(to_dev(chan), "%s: state: %lx\n", __func__, chan->state);
+
+       spin_lock_bh(&chan->cleanup_lock);
+       if (test_and_clear_bit(IOAT_RESET_PENDING, &chan->state)) {
+               struct ioat_desc_sw *desc;
+
+               spin_lock_bh(&ioat->desc_lock);
+
+               /* restart active descriptors */
+               desc = to_ioat_desc(ioat->used_desc.prev);
+               ioat_set_chainaddr(ioat, desc->txd.phys);
+               ioat_start(chan);
 
+               ioat->pending = 0;
+               set_bit(IOAT_COMPLETION_PENDING, &chan->state);
+               mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
+               spin_unlock_bh(&ioat->desc_lock);
+       } else if (test_bit(IOAT_COMPLETION_PENDING, &chan->state)) {
+               unsigned long phys_complete;
+
+               spin_lock_bh(&ioat->desc_lock);
+               /* if we haven't made progress and we have already
+                * acknowledged a pending completion once, then be more
+                * forceful with a restart
+                */
+               if (ioat_cleanup_preamble(chan, &phys_complete))
+                       __cleanup(ioat, phys_complete);
+               else if (test_bit(IOAT_COMPLETION_ACK, &chan->state))
+                       ioat1_reset_channel(ioat);
+               else {
+                       u64 status = ioat_chansts(chan);
+
+                       /* manually update the last completion address */
+                       if (ioat_chansts_to_addr(status) != 0)
+                               *chan->completion = status;
+
+                       set_bit(IOAT_COMPLETION_ACK, &chan->state);
+                       mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
+               }
+               spin_unlock_bh(&ioat->desc_lock);
+       }
        spin_unlock_bh(&chan->cleanup_lock);
 }
 
@@ -872,13 +779,8 @@ static void ioat1_dma_start_null_desc(struct ioat_dma_chan *ioat)
        list_add_tail(&desc->node, &ioat->used_desc);
        dump_desc_dbg(ioat, desc);
 
-       writel(((u64) desc->txd.phys) & 0x00000000FFFFFFFF,
-              chan->reg_base + IOAT1_CHAINADDR_OFFSET_LOW);
-       writel(((u64) desc->txd.phys) >> 32,
-              chan->reg_base + IOAT1_CHAINADDR_OFFSET_HIGH);
-
-       writeb(IOAT_CHANCMD_START, chan->reg_base
-               + IOAT_CHANCMD_OFFSET(chan->device->version));
+       ioat_set_chainaddr(ioat, desc->txd.phys);
+       ioat_start(chan);
        spin_unlock_bh(&ioat->desc_lock);
 }
 
@@ -887,7 +789,7 @@ static void ioat1_dma_start_null_desc(struct ioat_dma_chan *ioat)
  */
 #define IOAT_TEST_SIZE 2000
 
-static void ioat_dma_test_callback(void *dma_async_param)
+static void __devinit ioat_dma_test_callback(void *dma_async_param)
 {
        struct completion *cmp = dma_async_param;
 
@@ -898,7 +800,7 @@ static void ioat_dma_test_callback(void *dma_async_param)
  * ioat_dma_self_test - Perform a IOAT transaction to verify the HW works.
  * @device: device to be tested
  */
-static int ioat_dma_self_test(struct ioatdma_device *device)
+int __devinit ioat_dma_self_test(struct ioatdma_device *device)
 {
        int i;
        u8 *src;
@@ -1099,7 +1001,7 @@ static void ioat_disable_interrupts(struct ioatdma_device *device)
        writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
 }
 
-int ioat_probe(struct ioatdma_device *device)
+int __devinit ioat_probe(struct ioatdma_device *device)
 {
        int err = -ENODEV;
        struct dma_device *dma = &device->common;
@@ -1129,13 +1031,8 @@ int ioat_probe(struct ioatdma_device *device)
        dma_cap_set(DMA_MEMCPY, dma->cap_mask);
        dma->dev = &pdev->dev;
 
-       dev_err(dev, "Intel(R) I/OAT DMA Engine found,"
-               " %d channels, device version 0x%02x, driver version %s\n",
-               dma->chancnt, device->version, IOAT_DMA_VERSION);
-
        if (!dma->chancnt) {
-               dev_err(dev, "Intel(R) I/OAT DMA Engine problem found: "
-                       "zero channels detected\n");
+               dev_err(dev, "zero channels detected\n");
                goto err_setup_interrupts;
        }
 
@@ -1143,7 +1040,7 @@ int ioat_probe(struct ioatdma_device *device)
        if (err)
                goto err_setup_interrupts;
 
-       err = ioat_dma_self_test(device);
+       err = device->self_test(device);
        if (err)
                goto err_self_test;
 
@@ -1159,7 +1056,7 @@ err_dma_pool:
        return err;
 }
 
-int ioat_register(struct ioatdma_device *device)
+int __devinit ioat_register(struct ioatdma_device *device)
 {
        int err = dma_async_device_register(&device->common);
 
@@ -1186,7 +1083,114 @@ static void ioat1_intr_quirk(struct ioatdma_device *device)
        pci_write_config_dword(pdev, IOAT_PCI_DMACTRL_OFFSET, dmactrl);
 }
 
-int ioat1_dma_probe(struct ioatdma_device *device, int dca)
+static ssize_t ring_size_show(struct dma_chan *c, char *page)
+{
+       struct ioat_dma_chan *ioat = to_ioat_chan(c);
+
+       return sprintf(page, "%d\n", ioat->desccount);
+}
+static struct ioat_sysfs_entry ring_size_attr = __ATTR_RO(ring_size);
+
+static ssize_t ring_active_show(struct dma_chan *c, char *page)
+{
+       struct ioat_dma_chan *ioat = to_ioat_chan(c);
+
+       return sprintf(page, "%d\n", ioat->active);
+}
+static struct ioat_sysfs_entry ring_active_attr = __ATTR_RO(ring_active);
+
+static ssize_t cap_show(struct dma_chan *c, char *page)
+{
+       struct dma_device *dma = c->device;
+
+       return sprintf(page, "copy%s%s%s%s%s%s\n",
+                      dma_has_cap(DMA_PQ, dma->cap_mask) ? " pq" : "",
+                      dma_has_cap(DMA_PQ_VAL, dma->cap_mask) ? " pq_val" : "",
+                      dma_has_cap(DMA_XOR, dma->cap_mask) ? " xor" : "",
+                      dma_has_cap(DMA_XOR_VAL, dma->cap_mask) ? " xor_val" : "",
+                      dma_has_cap(DMA_MEMSET, dma->cap_mask)  ? " fill" : "",
+                      dma_has_cap(DMA_INTERRUPT, dma->cap_mask) ? " intr" : "");
+
+}
+struct ioat_sysfs_entry ioat_cap_attr = __ATTR_RO(cap);
+
+static ssize_t version_show(struct dma_chan *c, char *page)
+{
+       struct dma_device *dma = c->device;
+       struct ioatdma_device *device = to_ioatdma_device(dma);
+
+       return sprintf(page, "%d.%d\n",
+                      device->version >> 4, device->version & 0xf);
+}
+struct ioat_sysfs_entry ioat_version_attr = __ATTR_RO(version);
+
+static struct attribute *ioat1_attrs[] = {
+       &ring_size_attr.attr,
+       &ring_active_attr.attr,
+       &ioat_cap_attr.attr,
+       &ioat_version_attr.attr,
+       NULL,
+};
+
+static ssize_t
+ioat_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
+{
+       struct ioat_sysfs_entry *entry;
+       struct ioat_chan_common *chan;
+
+       entry = container_of(attr, struct ioat_sysfs_entry, attr);
+       chan = container_of(kobj, struct ioat_chan_common, kobj);
+
+       if (!entry->show)
+               return -EIO;
+       return entry->show(&chan->common, page);
+}
+
+struct sysfs_ops ioat_sysfs_ops = {
+       .show   = ioat_attr_show,
+};
+
+static struct kobj_type ioat1_ktype = {
+       .sysfs_ops = &ioat_sysfs_ops,
+       .default_attrs = ioat1_attrs,
+};
+
+void ioat_kobject_add(struct ioatdma_device *device, struct kobj_type *type)
+{
+       struct dma_device *dma = &device->common;
+       struct dma_chan *c;
+
+       list_for_each_entry(c, &dma->channels, device_node) {
+               struct ioat_chan_common *chan = to_chan_common(c);
+               struct kobject *parent = &c->dev->device.kobj;
+               int err;
+
+               err = kobject_init_and_add(&chan->kobj, type, parent, "quickdata");
+               if (err) {
+                       dev_warn(to_dev(chan),
+                                "sysfs init error (%d), continuing...\n", err);
+                       kobject_put(&chan->kobj);
+                       set_bit(IOAT_KOBJ_INIT_FAIL, &chan->state);
+               }
+       }
+}
+
+void ioat_kobject_del(struct ioatdma_device *device)
+{
+       struct dma_device *dma = &device->common;
+       struct dma_chan *c;
+
+       list_for_each_entry(c, &dma->channels, device_node) {
+               struct ioat_chan_common *chan = to_chan_common(c);
+
+               if (!test_bit(IOAT_KOBJ_INIT_FAIL, &chan->state)) {
+                       kobject_del(&chan->kobj);
+                       kobject_put(&chan->kobj);
+               }
+       }
+}
+
+int __devinit ioat1_dma_probe(struct ioatdma_device *device, int dca)
 {
        struct pci_dev *pdev = device->pdev;
        struct dma_device *dma;
@@ -1194,6 +1198,7 @@ int ioat1_dma_probe(struct ioatdma_device *device, int dca)
 
        device->intr_quirk = ioat1_intr_quirk;
        device->enumerate_channels = ioat1_enumerate_channels;
+       device->self_test = ioat_dma_self_test;
        dma = &device->common;
        dma->device_prep_dma_memcpy = ioat1_dma_prep_memcpy;
        dma->device_issue_pending = ioat1_dma_memcpy_issue_pending;
@@ -1208,24 +1213,22 @@ int ioat1_dma_probe(struct ioatdma_device *device, int dca)
        err = ioat_register(device);
        if (err)
                return err;
+       ioat_kobject_add(device, &ioat1_ktype);
+
        if (dca)
                device->dca = ioat_dca_init(pdev, device->reg_base);
 
-       INIT_DELAYED_WORK(&device->work, ioat1_chan_watchdog);
-       schedule_delayed_work(&device->work, WATCHDOG_DELAY);
-
        return err;
 }
 
-void ioat_dma_remove(struct ioatdma_device *device)
+void __devexit ioat_dma_remove(struct ioatdma_device *device)
 {
        struct dma_device *dma = &device->common;
 
-       if (device->version != IOAT_VER_3_0)
-               cancel_delayed_work(&device->work);
-
        ioat_disable_interrupts(device);
 
+       ioat_kobject_del(device);
+
        dma_async_device_unregister(dma);
 
        pci_pool_destroy(device->dma_pool);