dmaengine: qcom_hidma: implement lower level hardware interface
authorSinan Kaya <okaya@codeaurora.org>
Sun, 1 May 2016 04:25:26 +0000 (00:25 -0400)
committerVinod Koul <vinod.koul@intel.com>
Sat, 14 May 2016 06:24:45 +0000 (11:54 +0530)
This patch implements the hardware hooks for the HIDMA channel driver.

The main functions of interest are:
- hidma_ll_init
- hidma_ll_request
- hidma_ll_queue_request
- hidma_ll_hw_start

OS layer calls the hidma_ll_init function during probe to set up the
hardware. At this moment, the number of supported descriptors are also
given. On each request, a descriptor is allocated from the free pool and
filled in with the transfer parameters. Multiple requests can be queued
into the hardware via the OS interface. When client is ready for requests
to be executed, start method is called.

Completions are delivered via callbacks via tasklet.

Signed-off-by: Sinan Kaya <okaya@codeaurora.org>
Signed-off-by: Vinod Koul <vinod.koul@intel.com>
drivers/dma/qcom/Makefile
drivers/dma/qcom/hidma.c
drivers/dma/qcom/hidma.h
drivers/dma/qcom/hidma_ll.c [new file with mode: 0644]

index bfea699..6bf9267 100644 (file)
@@ -1,3 +1,5 @@
 obj-$(CONFIG_QCOM_BAM_DMA) += bam_dma.o
 obj-$(CONFIG_QCOM_HIDMA_MGMT) += hdma_mgmt.o
 hdma_mgmt-objs  := hidma_mgmt.o hidma_mgmt_sys.o
+obj-$(CONFIG_QCOM_HIDMA) +=  hdma.o
+hdma-objs        := hidma_ll.o hidma.o
index cccc78e..af5e542 100644 (file)
@@ -404,7 +404,7 @@ static int hidma_terminate_channel(struct dma_chan *chan)
        spin_unlock_irqrestore(&mchan->lock, irqflags);
 
        /* this suspends the existing transfer */
-       rc = hidma_ll_pause(dmadev->lldev);
+       rc = hidma_ll_disable(dmadev->lldev);
        if (rc) {
                dev_err(dmadev->ddev.dev, "channel did not pause\n");
                goto out;
@@ -427,7 +427,7 @@ static int hidma_terminate_channel(struct dma_chan *chan)
                list_move(&mdesc->node, &mchan->free);
        }
 
-       rc = hidma_ll_resume(dmadev->lldev);
+       rc = hidma_ll_enable(dmadev->lldev);
 out:
        pm_runtime_mark_last_busy(dmadev->ddev.dev);
        pm_runtime_put_autosuspend(dmadev->ddev.dev);
@@ -488,7 +488,7 @@ static int hidma_pause(struct dma_chan *chan)
        dmadev = to_hidma_dev(mchan->chan.device);
        if (!mchan->paused) {
                pm_runtime_get_sync(dmadev->ddev.dev);
-               if (hidma_ll_pause(dmadev->lldev))
+               if (hidma_ll_disable(dmadev->lldev))
                        dev_warn(dmadev->ddev.dev, "channel did not stop\n");
                mchan->paused = true;
                pm_runtime_mark_last_busy(dmadev->ddev.dev);
@@ -507,7 +507,7 @@ static int hidma_resume(struct dma_chan *chan)
        dmadev = to_hidma_dev(mchan->chan.device);
        if (mchan->paused) {
                pm_runtime_get_sync(dmadev->ddev.dev);
-               rc = hidma_ll_resume(dmadev->lldev);
+               rc = hidma_ll_enable(dmadev->lldev);
                if (!rc)
                        mchan->paused = false;
                else
index 231e306..b1362ef 100644 (file)
@@ -1,7 +1,7 @@
 /*
  * Qualcomm Technologies HIDMA data structures
  *
- * Copyright (c) 2014, The Linux Foundation. All rights reserved.
+ * Copyright (c) 2014-2016, The Linux Foundation. All rights reserved.
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License version 2 and
 #include <linux/interrupt.h>
 #include <linux/dmaengine.h>
 
-#define TRE_SIZE                       32 /* each TRE is 32 bytes  */
-#define TRE_CFG_IDX                    0
-#define TRE_LEN_IDX                    1
-#define TRE_SRC_LOW_IDX                2
-#define TRE_SRC_HI_IDX                 3
-#define TRE_DEST_LOW_IDX               4
-#define TRE_DEST_HI_IDX                5
-
-struct hidma_tx_status {
-       u8 err_info;                    /* error record in this transfer    */
-       u8 err_code;                    /* completion code                  */
-};
+#define HIDMA_TRE_SIZE                 32 /* each TRE is 32 bytes  */
+#define HIDMA_TRE_CFG_IDX              0
+#define HIDMA_TRE_LEN_IDX              1
+#define HIDMA_TRE_SRC_LOW_IDX          2
+#define HIDMA_TRE_SRC_HI_IDX           3
+#define HIDMA_TRE_DEST_LOW_IDX         4
+#define HIDMA_TRE_DEST_HI_IDX          5
 
 struct hidma_tre {
        atomic_t allocated;             /* if this channel is allocated     */
        bool queued;                    /* flag whether this is pending     */
        u16 status;                     /* status                           */
-       u32 chidx;                      /* index of the tre                 */
+       u32 idx;                        /* index of the tre                 */
        u32 dma_sig;                    /* signature of the tre             */
        const char *dev_name;           /* name of the device               */
        void (*callback)(void *data);   /* requester callback               */
        void *data;                     /* Data associated with this channel*/
        struct hidma_lldev *lldev;      /* lldma device pointer             */
-       u32 tre_local[TRE_SIZE / sizeof(u32) + 1]; /* TRE local copy        */
+       u32 tre_local[HIDMA_TRE_SIZE / sizeof(u32) + 1]; /* TRE local copy  */
        u32 tre_index;                  /* the offset where this was written*/
        u32 int_flags;                  /* interrupt flags                  */
+       u8 err_info;                    /* error record in this transfer    */
+       u8 err_code;                    /* completion code                  */
 };
 
 struct hidma_lldev {
@@ -61,22 +58,21 @@ struct hidma_lldev {
        void __iomem *evca;             /* Event Channel address          */
        struct hidma_tre
                **pending_tre_list;     /* Pointers to pending TREs       */
-       struct hidma_tx_status
-               *tx_status_list;        /* Pointers to pending TREs status*/
        s32 pending_tre_count;          /* Number of TREs pending         */
 
        void *tre_ring;                 /* TRE ring                       */
-       dma_addr_t tre_ring_handle;     /* TRE ring to be shared with HW  */
+       dma_addr_t tre_dma;             /* TRE ring to be shared with HW  */
        u32 tre_ring_size;              /* Byte size of the ring          */
        u32 tre_processed_off;          /* last processed TRE             */
 
        void *evre_ring;                /* EVRE ring                       */
-       dma_addr_t evre_ring_handle;    /* EVRE ring to be shared with HW  */
+       dma_addr_t evre_dma;            /* EVRE ring to be shared with HW  */
        u32 evre_ring_size;             /* Byte size of the ring           */
        u32 evre_processed_off;         /* last processed EVRE             */
 
        u32 tre_write_offset;           /* TRE write location              */
        struct tasklet_struct task;     /* task delivering notifications   */
+       struct tasklet_struct rst_task; /* task to reset HW                */
        DECLARE_KFIFO_PTR(handoff_fifo,
                struct hidma_tre *);    /* pending TREs FIFO               */
 };
@@ -145,8 +141,8 @@ enum dma_status hidma_ll_status(struct hidma_lldev *llhndl, u32 tre_ch);
 bool hidma_ll_isenabled(struct hidma_lldev *llhndl);
 void hidma_ll_queue_request(struct hidma_lldev *llhndl, u32 tre_ch);
 void hidma_ll_start(struct hidma_lldev *llhndl);
-int hidma_ll_pause(struct hidma_lldev *llhndl);
-int hidma_ll_resume(struct hidma_lldev *llhndl);
+int hidma_ll_disable(struct hidma_lldev *lldev);
+int hidma_ll_enable(struct hidma_lldev *llhndl);
 void hidma_ll_set_transfer_params(struct hidma_lldev *llhndl, u32 tre_ch,
        dma_addr_t src, dma_addr_t dest, u32 len, u32 flags);
 int hidma_ll_setup(struct hidma_lldev *lldev);
diff --git a/drivers/dma/qcom/hidma_ll.c b/drivers/dma/qcom/hidma_ll.c
new file mode 100644 (file)
index 0000000..f392900
--- /dev/null
@@ -0,0 +1,872 @@
+/*
+ * Qualcomm Technologies HIDMA DMA engine low level code
+ *
+ * Copyright (c) 2015-2016, The Linux Foundation. All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 and
+ * only version 2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ */
+
+#include <linux/dmaengine.h>
+#include <linux/slab.h>
+#include <linux/interrupt.h>
+#include <linux/mm.h>
+#include <linux/highmem.h>
+#include <linux/dma-mapping.h>
+#include <linux/delay.h>
+#include <linux/atomic.h>
+#include <linux/iopoll.h>
+#include <linux/kfifo.h>
+#include <linux/bitops.h>
+
+#include "hidma.h"
+
+#define HIDMA_EVRE_SIZE                        16      /* each EVRE is 16 bytes */
+
+#define HIDMA_TRCA_CTRLSTS_REG                 0x000
+#define HIDMA_TRCA_RING_LOW_REG                0x008
+#define HIDMA_TRCA_RING_HIGH_REG               0x00C
+#define HIDMA_TRCA_RING_LEN_REG                0x010
+#define HIDMA_TRCA_DOORBELL_REG                0x400
+
+#define HIDMA_EVCA_CTRLSTS_REG                 0x000
+#define HIDMA_EVCA_INTCTRL_REG                 0x004
+#define HIDMA_EVCA_RING_LOW_REG                0x008
+#define HIDMA_EVCA_RING_HIGH_REG               0x00C
+#define HIDMA_EVCA_RING_LEN_REG                0x010
+#define HIDMA_EVCA_WRITE_PTR_REG               0x020
+#define HIDMA_EVCA_DOORBELL_REG                0x400
+
+#define HIDMA_EVCA_IRQ_STAT_REG                0x100
+#define HIDMA_EVCA_IRQ_CLR_REG                 0x108
+#define HIDMA_EVCA_IRQ_EN_REG                  0x110
+
+#define HIDMA_EVRE_CFG_IDX                     0
+
+#define HIDMA_EVRE_ERRINFO_BIT_POS             24
+#define HIDMA_EVRE_CODE_BIT_POS                28
+
+#define HIDMA_EVRE_ERRINFO_MASK                GENMASK(3, 0)
+#define HIDMA_EVRE_CODE_MASK                   GENMASK(3, 0)
+
+#define HIDMA_CH_CONTROL_MASK                  GENMASK(7, 0)
+#define HIDMA_CH_STATE_MASK                    GENMASK(7, 0)
+#define HIDMA_CH_STATE_BIT_POS                 0x8
+
+#define HIDMA_IRQ_EV_CH_EOB_IRQ_BIT_POS        0
+#define HIDMA_IRQ_EV_CH_WR_RESP_BIT_POS        1
+#define HIDMA_IRQ_TR_CH_TRE_RD_RSP_ER_BIT_POS  9
+#define HIDMA_IRQ_TR_CH_DATA_RD_ER_BIT_POS     10
+#define HIDMA_IRQ_TR_CH_DATA_WR_ER_BIT_POS     11
+#define HIDMA_IRQ_TR_CH_INVALID_TRE_BIT_POS    14
+
+#define ENABLE_IRQS (BIT(HIDMA_IRQ_EV_CH_EOB_IRQ_BIT_POS)      | \
+                    BIT(HIDMA_IRQ_EV_CH_WR_RESP_BIT_POS)       | \
+                    BIT(HIDMA_IRQ_TR_CH_TRE_RD_RSP_ER_BIT_POS) | \
+                    BIT(HIDMA_IRQ_TR_CH_DATA_RD_ER_BIT_POS)    | \
+                    BIT(HIDMA_IRQ_TR_CH_DATA_WR_ER_BIT_POS)    | \
+                    BIT(HIDMA_IRQ_TR_CH_INVALID_TRE_BIT_POS))
+
+#define HIDMA_INCREMENT_ITERATOR(iter, size, ring_size)        \
+do {                                                           \
+       iter += size;                                           \
+       if (iter >= ring_size)                                  \
+               iter -= ring_size;                              \
+} while (0)
+
+#define HIDMA_CH_STATE(val)    \
+       ((val >> HIDMA_CH_STATE_BIT_POS) & HIDMA_CH_STATE_MASK)
+
+#define HIDMA_ERR_INT_MASK                             \
+       (BIT(HIDMA_IRQ_TR_CH_INVALID_TRE_BIT_POS)   |   \
+        BIT(HIDMA_IRQ_TR_CH_TRE_RD_RSP_ER_BIT_POS) |   \
+        BIT(HIDMA_IRQ_EV_CH_WR_RESP_BIT_POS)       |   \
+        BIT(HIDMA_IRQ_TR_CH_DATA_RD_ER_BIT_POS)    |   \
+        BIT(HIDMA_IRQ_TR_CH_DATA_WR_ER_BIT_POS))
+
+enum ch_command {
+       HIDMA_CH_DISABLE = 0,
+       HIDMA_CH_ENABLE = 1,
+       HIDMA_CH_SUSPEND = 2,
+       HIDMA_CH_RESET = 9,
+};
+
+enum ch_state {
+       HIDMA_CH_DISABLED = 0,
+       HIDMA_CH_ENABLED = 1,
+       HIDMA_CH_RUNNING = 2,
+       HIDMA_CH_SUSPENDED = 3,
+       HIDMA_CH_STOPPED = 4,
+};
+
+enum tre_type {
+       HIDMA_TRE_MEMCPY = 3,
+};
+
+enum err_code {
+       HIDMA_EVRE_STATUS_COMPLETE = 1,
+       HIDMA_EVRE_STATUS_ERROR = 4,
+};
+
+static int hidma_is_chan_enabled(int state)
+{
+       switch (state) {
+       case HIDMA_CH_ENABLED:
+       case HIDMA_CH_RUNNING:
+               return true;
+       default:
+               return false;
+       }
+}
+
+void hidma_ll_free(struct hidma_lldev *lldev, u32 tre_ch)
+{
+       struct hidma_tre *tre;
+
+       if (tre_ch >= lldev->nr_tres) {
+               dev_err(lldev->dev, "invalid TRE number in free:%d", tre_ch);
+               return;
+       }
+
+       tre = &lldev->trepool[tre_ch];
+       if (atomic_read(&tre->allocated) != true) {
+               dev_err(lldev->dev, "trying to free an unused TRE:%d", tre_ch);
+               return;
+       }
+
+       atomic_set(&tre->allocated, 0);
+}
+
+int hidma_ll_request(struct hidma_lldev *lldev, u32 sig, const char *dev_name,
+                    void (*callback)(void *data), void *data, u32 *tre_ch)
+{
+       unsigned int i;
+       struct hidma_tre *tre;
+       u32 *tre_local;
+
+       if (!tre_ch || !lldev)
+               return -EINVAL;
+
+       /* need to have at least one empty spot in the queue */
+       for (i = 0; i < lldev->nr_tres - 1; i++) {
+               if (atomic_add_unless(&lldev->trepool[i].allocated, 1, 1))
+                       break;
+       }
+
+       if (i == (lldev->nr_tres - 1))
+               return -ENOMEM;
+
+       tre = &lldev->trepool[i];
+       tre->dma_sig = sig;
+       tre->dev_name = dev_name;
+       tre->callback = callback;
+       tre->data = data;
+       tre->idx = i;
+       tre->status = 0;
+       tre->queued = 0;
+       tre->err_code = 0;
+       tre->err_info = 0;
+       tre->lldev = lldev;
+       tre_local = &tre->tre_local[0];
+       tre_local[HIDMA_TRE_CFG_IDX] = HIDMA_TRE_MEMCPY;
+       tre_local[HIDMA_TRE_CFG_IDX] |= (lldev->chidx & 0xFF) << 8;
+       tre_local[HIDMA_TRE_CFG_IDX] |= BIT(16);        /* set IEOB */
+       *tre_ch = i;
+       if (callback)
+               callback(data);
+       return 0;
+}
+
+/*
+ * Multiple TREs may be queued and waiting in the pending queue.
+ */
+static void hidma_ll_tre_complete(unsigned long arg)
+{
+       struct hidma_lldev *lldev = (struct hidma_lldev *)arg;
+       struct hidma_tre *tre;
+
+       while (kfifo_out(&lldev->handoff_fifo, &tre, 1)) {
+               /* call the user if it has been read by the hardware */
+               if (tre->callback)
+                       tre->callback(tre->data);
+       }
+}
+
+static int hidma_post_completed(struct hidma_lldev *lldev, int tre_iterator,
+                               u8 err_info, u8 err_code)
+{
+       struct hidma_tre *tre;
+       unsigned long flags;
+
+       spin_lock_irqsave(&lldev->lock, flags);
+       tre = lldev->pending_tre_list[tre_iterator / HIDMA_TRE_SIZE];
+       if (!tre) {
+               spin_unlock_irqrestore(&lldev->lock, flags);
+               dev_warn(lldev->dev, "tre_index [%d] and tre out of sync\n",
+                        tre_iterator / HIDMA_TRE_SIZE);
+               return -EINVAL;
+       }
+       lldev->pending_tre_list[tre->tre_index] = NULL;
+
+       /*
+        * Keep track of pending TREs that SW is expecting to receive
+        * from HW. We got one now. Decrement our counter.
+        */
+       lldev->pending_tre_count--;
+       if (lldev->pending_tre_count < 0) {
+               dev_warn(lldev->dev, "tre count mismatch on completion");
+               lldev->pending_tre_count = 0;
+       }
+
+       spin_unlock_irqrestore(&lldev->lock, flags);
+
+       tre->err_info = err_info;
+       tre->err_code = err_code;
+       tre->queued = 0;
+
+       kfifo_put(&lldev->handoff_fifo, tre);
+       tasklet_schedule(&lldev->task);
+
+       return 0;
+}
+
+/*
+ * Called to handle the interrupt for the channel.
+ * Return a positive number if TRE or EVRE were consumed on this run.
+ * Return a positive number if there are pending TREs or EVREs.
+ * Return 0 if there is nothing to consume or no pending TREs/EVREs found.
+ */
+static int hidma_handle_tre_completion(struct hidma_lldev *lldev)
+{
+       u32 evre_ring_size = lldev->evre_ring_size;
+       u32 tre_ring_size = lldev->tre_ring_size;
+       u32 err_info, err_code, evre_write_off;
+       u32 tre_iterator, evre_iterator;
+       u32 num_completed = 0;
+
+       evre_write_off = readl_relaxed(lldev->evca + HIDMA_EVCA_WRITE_PTR_REG);
+       tre_iterator = lldev->tre_processed_off;
+       evre_iterator = lldev->evre_processed_off;
+
+       if ((evre_write_off > evre_ring_size) ||
+           (evre_write_off % HIDMA_EVRE_SIZE)) {
+               dev_err(lldev->dev, "HW reports invalid EVRE write offset\n");
+               return 0;
+       }
+
+       /*
+        * By the time control reaches here the number of EVREs and TREs
+        * may not match. Only consume the ones that hardware told us.
+        */
+       while ((evre_iterator != evre_write_off)) {
+               u32 *current_evre = lldev->evre_ring + evre_iterator;
+               u32 cfg;
+
+               cfg = current_evre[HIDMA_EVRE_CFG_IDX];
+               err_info = cfg >> HIDMA_EVRE_ERRINFO_BIT_POS;
+               err_info &= HIDMA_EVRE_ERRINFO_MASK;
+               err_code =
+                   (cfg >> HIDMA_EVRE_CODE_BIT_POS) & HIDMA_EVRE_CODE_MASK;
+
+               if (hidma_post_completed(lldev, tre_iterator, err_info,
+                                        err_code))
+                       break;
+
+               HIDMA_INCREMENT_ITERATOR(tre_iterator, HIDMA_TRE_SIZE,
+                                        tre_ring_size);
+               HIDMA_INCREMENT_ITERATOR(evre_iterator, HIDMA_EVRE_SIZE,
+                                        evre_ring_size);
+
+               /*
+                * Read the new event descriptor written by the HW.
+                * As we are processing the delivered events, other events
+                * get queued to the SW for processing.
+                */
+               evre_write_off =
+                   readl_relaxed(lldev->evca + HIDMA_EVCA_WRITE_PTR_REG);
+               num_completed++;
+       }
+
+       if (num_completed) {
+               u32 evre_read_off = (lldev->evre_processed_off +
+                                    HIDMA_EVRE_SIZE * num_completed);
+               u32 tre_read_off = (lldev->tre_processed_off +
+                                   HIDMA_TRE_SIZE * num_completed);
+
+               evre_read_off = evre_read_off % evre_ring_size;
+               tre_read_off = tre_read_off % tre_ring_size;
+
+               writel(evre_read_off, lldev->evca + HIDMA_EVCA_DOORBELL_REG);
+
+               /* record the last processed tre offset */
+               lldev->tre_processed_off = tre_read_off;
+               lldev->evre_processed_off = evre_read_off;
+       }
+
+       return num_completed;
+}
+
+void hidma_cleanup_pending_tre(struct hidma_lldev *lldev, u8 err_info,
+                              u8 err_code)
+{
+       u32 tre_iterator;
+       u32 tre_ring_size = lldev->tre_ring_size;
+       int num_completed = 0;
+       u32 tre_read_off;
+
+       tre_iterator = lldev->tre_processed_off;
+       while (lldev->pending_tre_count) {
+               if (hidma_post_completed(lldev, tre_iterator, err_info,
+                                        err_code))
+                       break;
+               HIDMA_INCREMENT_ITERATOR(tre_iterator, HIDMA_TRE_SIZE,
+                                        tre_ring_size);
+               num_completed++;
+       }
+       tre_read_off = (lldev->tre_processed_off +
+                       HIDMA_TRE_SIZE * num_completed);
+
+       tre_read_off = tre_read_off % tre_ring_size;
+
+       /* record the last processed tre offset */
+       lldev->tre_processed_off = tre_read_off;
+}
+
+static int hidma_ll_reset(struct hidma_lldev *lldev)
+{
+       u32 val;
+       int ret;
+
+       val = readl(lldev->trca + HIDMA_TRCA_CTRLSTS_REG);
+       val &= ~(HIDMA_CH_CONTROL_MASK << 16);
+       val |= HIDMA_CH_RESET << 16;
+       writel(val, lldev->trca + HIDMA_TRCA_CTRLSTS_REG);
+
+       /*
+        * Delay 10ms after reset to allow DMA logic to quiesce.
+        * Do a polled read up to 1ms and 10ms maximum.
+        */
+       ret = readl_poll_timeout(lldev->trca + HIDMA_TRCA_CTRLSTS_REG, val,
+                                HIDMA_CH_STATE(val) == HIDMA_CH_DISABLED,
+                                1000, 10000);
+       if (ret) {
+               dev_err(lldev->dev, "transfer channel did not reset\n");
+               return ret;
+       }
+
+       val = readl(lldev->evca + HIDMA_EVCA_CTRLSTS_REG);
+       val &= ~(HIDMA_CH_CONTROL_MASK << 16);
+       val |= HIDMA_CH_RESET << 16;
+       writel(val, lldev->evca + HIDMA_EVCA_CTRLSTS_REG);
+
+       /*
+        * Delay 10ms after reset to allow DMA logic to quiesce.
+        * Do a polled read up to 1ms and 10ms maximum.
+        */
+       ret = readl_poll_timeout(lldev->evca + HIDMA_EVCA_CTRLSTS_REG, val,
+                                HIDMA_CH_STATE(val) == HIDMA_CH_DISABLED,
+                                1000, 10000);
+       if (ret)
+               return ret;
+
+       lldev->trch_state = HIDMA_CH_DISABLED;
+       lldev->evch_state = HIDMA_CH_DISABLED;
+       return 0;
+}
+
+/*
+ * Abort all transactions and perform a reset.
+ */
+static void hidma_ll_abort(unsigned long arg)
+{
+       struct hidma_lldev *lldev = (struct hidma_lldev *)arg;
+       u8 err_code = HIDMA_EVRE_STATUS_ERROR;
+       u8 err_info = 0xFF;
+       int rc;
+
+       hidma_cleanup_pending_tre(lldev, err_info, err_code);
+
+       /* reset the channel for recovery */
+       rc = hidma_ll_setup(lldev);
+       if (rc) {
+               dev_err(lldev->dev, "channel reinitialize failed after error\n");
+               return;
+       }
+       writel(ENABLE_IRQS, lldev->evca + HIDMA_EVCA_IRQ_EN_REG);
+}
+
+/*
+ * The interrupt handler for HIDMA will try to consume as many pending
+ * EVRE from the event queue as possible. Each EVRE has an associated
+ * TRE that holds the user interface parameters. EVRE reports the
+ * result of the transaction. Hardware guarantees ordering between EVREs
+ * and TREs. We use last processed offset to figure out which TRE is
+ * associated with which EVRE. If two TREs are consumed by HW, the EVREs
+ * are in order in the event ring.
+ *
+ * This handler will do a one pass for consuming EVREs. Other EVREs may
+ * be delivered while we are working. It will try to consume incoming
+ * EVREs one more time and return.
+ *
+ * For unprocessed EVREs, hardware will trigger another interrupt until
+ * all the interrupt bits are cleared.
+ *
+ * Hardware guarantees that by the time interrupt is observed, all data
+ * transactions in flight are delivered to their respective places and
+ * are visible to the CPU.
+ *
+ * On demand paging for IOMMU is only supported for PCIe via PRI
+ * (Page Request Interface) not for HIDMA. All other hardware instances
+ * including HIDMA work on pinned DMA addresses.
+ *
+ * HIDMA is not aware of IOMMU presence since it follows the DMA API. All
+ * IOMMU latency will be built into the data movement time. By the time
+ * interrupt happens, IOMMU lookups + data movement has already taken place.
+ *
+ * While the first read in a typical PCI endpoint ISR flushes all outstanding
+ * requests traditionally to the destination, this concept does not apply
+ * here for this HW.
+ */
+irqreturn_t hidma_ll_inthandler(int chirq, void *arg)
+{
+       struct hidma_lldev *lldev = arg;
+       u32 status;
+       u32 enable;
+       u32 cause;
+
+       /*
+        * Fine tuned for this HW...
+        *
+        * This ISR has been designed for this particular hardware. Relaxed
+        * read and write accessors are used for performance reasons due to
+        * interrupt delivery guarantees. Do not copy this code blindly and
+        * expect that to work.
+        */
+       status = readl_relaxed(lldev->evca + HIDMA_EVCA_IRQ_STAT_REG);
+       enable = readl_relaxed(lldev->evca + HIDMA_EVCA_IRQ_EN_REG);
+       cause = status & enable;
+
+       while (cause) {
+               if (cause & HIDMA_ERR_INT_MASK) {
+                       dev_err(lldev->dev, "error 0x%x, resetting...\n",
+                                       cause);
+
+                       /* Clear out pending interrupts */
+                       writel(cause, lldev->evca + HIDMA_EVCA_IRQ_CLR_REG);
+
+                       tasklet_schedule(&lldev->rst_task);
+                       goto out;
+               }
+
+               /*
+                * Try to consume as many EVREs as possible.
+                */
+               hidma_handle_tre_completion(lldev);
+
+               /* We consumed TREs or there are pending TREs or EVREs. */
+               writel_relaxed(cause, lldev->evca + HIDMA_EVCA_IRQ_CLR_REG);
+
+               /*
+                * Another interrupt might have arrived while we are
+                * processing this one. Read the new cause.
+                */
+               status = readl_relaxed(lldev->evca + HIDMA_EVCA_IRQ_STAT_REG);
+               enable = readl_relaxed(lldev->evca + HIDMA_EVCA_IRQ_EN_REG);
+               cause = status & enable;
+       }
+
+out:
+       return IRQ_HANDLED;
+}
+
+int hidma_ll_enable(struct hidma_lldev *lldev)
+{
+       u32 val;
+       int ret;
+
+       val = readl(lldev->evca + HIDMA_EVCA_CTRLSTS_REG);
+       val &= ~(HIDMA_CH_CONTROL_MASK << 16);
+       val |= HIDMA_CH_ENABLE << 16;
+       writel(val, lldev->evca + HIDMA_EVCA_CTRLSTS_REG);
+
+       ret = readl_poll_timeout(lldev->evca + HIDMA_EVCA_CTRLSTS_REG, val,
+                                hidma_is_chan_enabled(HIDMA_CH_STATE(val)),
+                                1000, 10000);
+       if (ret) {
+               dev_err(lldev->dev, "event channel did not get enabled\n");
+               return ret;
+       }
+
+       val = readl(lldev->trca + HIDMA_TRCA_CTRLSTS_REG);
+       val &= ~(HIDMA_CH_CONTROL_MASK << 16);
+       val |= HIDMA_CH_ENABLE << 16;
+       writel(val, lldev->trca + HIDMA_TRCA_CTRLSTS_REG);
+
+       ret = readl_poll_timeout(lldev->trca + HIDMA_TRCA_CTRLSTS_REG, val,
+                                hidma_is_chan_enabled(HIDMA_CH_STATE(val)),
+                                1000, 10000);
+       if (ret) {
+               dev_err(lldev->dev, "transfer channel did not get enabled\n");
+               return ret;
+       }
+
+       lldev->trch_state = HIDMA_CH_ENABLED;
+       lldev->evch_state = HIDMA_CH_ENABLED;
+
+       return 0;
+}
+
+void hidma_ll_start(struct hidma_lldev *lldev)
+{
+       unsigned long irqflags;
+
+       spin_lock_irqsave(&lldev->lock, irqflags);
+       writel(lldev->tre_write_offset, lldev->trca + HIDMA_TRCA_DOORBELL_REG);
+       spin_unlock_irqrestore(&lldev->lock, irqflags);
+}
+
+bool hidma_ll_isenabled(struct hidma_lldev *lldev)
+{
+       u32 val;
+
+       val = readl(lldev->trca + HIDMA_TRCA_CTRLSTS_REG);
+       lldev->trch_state = HIDMA_CH_STATE(val);
+       val = readl(lldev->evca + HIDMA_EVCA_CTRLSTS_REG);
+       lldev->evch_state = HIDMA_CH_STATE(val);
+
+       /* both channels have to be enabled before calling this function */
+       if (hidma_is_chan_enabled(lldev->trch_state) &&
+           hidma_is_chan_enabled(lldev->evch_state))
+               return true;
+
+       return false;
+}
+
+void hidma_ll_queue_request(struct hidma_lldev *lldev, u32 tre_ch)
+{
+       struct hidma_tre *tre;
+       unsigned long flags;
+
+       tre = &lldev->trepool[tre_ch];
+
+       /* copy the TRE into its location in the TRE ring */
+       spin_lock_irqsave(&lldev->lock, flags);
+       tre->tre_index = lldev->tre_write_offset / HIDMA_TRE_SIZE;
+       lldev->pending_tre_list[tre->tre_index] = tre;
+       memcpy(lldev->tre_ring + lldev->tre_write_offset,
+                       &tre->tre_local[0], HIDMA_TRE_SIZE);
+       tre->err_code = 0;
+       tre->err_info = 0;
+       tre->queued = 1;
+       lldev->pending_tre_count++;
+       lldev->tre_write_offset = (lldev->tre_write_offset + HIDMA_TRE_SIZE)
+                                       % lldev->tre_ring_size;
+       spin_unlock_irqrestore(&lldev->lock, flags);
+}
+
+/*
+ * Note that even though we stop this channel if there is a pending transaction
+ * in flight it will complete and follow the callback. This request will
+ * prevent further requests to be made.
+ */
+int hidma_ll_disable(struct hidma_lldev *lldev)
+{
+       u32 val;
+       int ret;
+
+       val = readl(lldev->evca + HIDMA_EVCA_CTRLSTS_REG);
+       lldev->evch_state = HIDMA_CH_STATE(val);
+       val = readl(lldev->trca + HIDMA_TRCA_CTRLSTS_REG);
+       lldev->trch_state = HIDMA_CH_STATE(val);
+
+       /* already suspended by this OS */
+       if ((lldev->trch_state == HIDMA_CH_SUSPENDED) ||
+           (lldev->evch_state == HIDMA_CH_SUSPENDED))
+               return 0;
+
+       /* already stopped by the manager */
+       if ((lldev->trch_state == HIDMA_CH_STOPPED) ||
+           (lldev->evch_state == HIDMA_CH_STOPPED))
+               return 0;
+
+       val = readl(lldev->trca + HIDMA_TRCA_CTRLSTS_REG);
+       val &= ~(HIDMA_CH_CONTROL_MASK << 16);
+       val |= HIDMA_CH_SUSPEND << 16;
+       writel(val, lldev->trca + HIDMA_TRCA_CTRLSTS_REG);
+
+       /*
+        * Start the wait right after the suspend is confirmed.
+        * Do a polled read up to 1ms and 10ms maximum.
+        */
+       ret = readl_poll_timeout(lldev->trca + HIDMA_TRCA_CTRLSTS_REG, val,
+                                HIDMA_CH_STATE(val) == HIDMA_CH_SUSPENDED,
+                                1000, 10000);
+       if (ret)
+               return ret;
+
+       val = readl(lldev->evca + HIDMA_EVCA_CTRLSTS_REG);
+       val &= ~(HIDMA_CH_CONTROL_MASK << 16);
+       val |= HIDMA_CH_SUSPEND << 16;
+       writel(val, lldev->evca + HIDMA_EVCA_CTRLSTS_REG);
+
+       /*
+        * Start the wait right after the suspend is confirmed
+        * Delay up to 10ms after reset to allow DMA logic to quiesce.
+        */
+       ret = readl_poll_timeout(lldev->evca + HIDMA_EVCA_CTRLSTS_REG, val,
+                                HIDMA_CH_STATE(val) == HIDMA_CH_SUSPENDED,
+                                1000, 10000);
+       if (ret)
+               return ret;
+
+       lldev->trch_state = HIDMA_CH_SUSPENDED;
+       lldev->evch_state = HIDMA_CH_SUSPENDED;
+       return 0;
+}
+
+void hidma_ll_set_transfer_params(struct hidma_lldev *lldev, u32 tre_ch,
+                                 dma_addr_t src, dma_addr_t dest, u32 len,
+                                 u32 flags)
+{
+       struct hidma_tre *tre;
+       u32 *tre_local;
+
+       if (tre_ch >= lldev->nr_tres) {
+               dev_err(lldev->dev, "invalid TRE number in transfer params:%d",
+                       tre_ch);
+               return;
+       }
+
+       tre = &lldev->trepool[tre_ch];
+       if (atomic_read(&tre->allocated) != true) {
+               dev_err(lldev->dev, "trying to set params on an unused TRE:%d",
+                       tre_ch);
+               return;
+       }
+
+       tre_local = &tre->tre_local[0];
+       tre_local[HIDMA_TRE_LEN_IDX] = len;
+       tre_local[HIDMA_TRE_SRC_LOW_IDX] = lower_32_bits(src);
+       tre_local[HIDMA_TRE_SRC_HI_IDX] = upper_32_bits(src);
+       tre_local[HIDMA_TRE_DEST_LOW_IDX] = lower_32_bits(dest);
+       tre_local[HIDMA_TRE_DEST_HI_IDX] = upper_32_bits(dest);
+       tre->int_flags = flags;
+}
+
+/*
+ * Called during initialization and after an error condition
+ * to restore hardware state.
+ */
+int hidma_ll_setup(struct hidma_lldev *lldev)
+{
+       int rc;
+       u64 addr;
+       u32 val;
+       u32 nr_tres = lldev->nr_tres;
+
+       lldev->pending_tre_count = 0;
+       lldev->tre_processed_off = 0;
+       lldev->evre_processed_off = 0;
+       lldev->tre_write_offset = 0;
+
+       /* disable interrupts */
+       writel(0, lldev->evca + HIDMA_EVCA_IRQ_EN_REG);
+
+       /* clear all pending interrupts */
+       val = readl(lldev->evca + HIDMA_EVCA_IRQ_STAT_REG);
+       writel(val, lldev->evca + HIDMA_EVCA_IRQ_CLR_REG);
+
+       rc = hidma_ll_reset(lldev);
+       if (rc)
+               return rc;
+
+       /*
+        * Clear all pending interrupts again.
+        * Otherwise, we observe reset complete interrupts.
+        */
+       val = readl(lldev->evca + HIDMA_EVCA_IRQ_STAT_REG);
+       writel(val, lldev->evca + HIDMA_EVCA_IRQ_CLR_REG);
+
+       /* disable interrupts again after reset */
+       writel(0, lldev->evca + HIDMA_EVCA_IRQ_EN_REG);
+
+       addr = lldev->tre_dma;
+       writel(lower_32_bits(addr), lldev->trca + HIDMA_TRCA_RING_LOW_REG);
+       writel(upper_32_bits(addr), lldev->trca + HIDMA_TRCA_RING_HIGH_REG);
+       writel(lldev->tre_ring_size, lldev->trca + HIDMA_TRCA_RING_LEN_REG);
+
+       addr = lldev->evre_dma;
+       writel(lower_32_bits(addr), lldev->evca + HIDMA_EVCA_RING_LOW_REG);
+       writel(upper_32_bits(addr), lldev->evca + HIDMA_EVCA_RING_HIGH_REG);
+       writel(HIDMA_EVRE_SIZE * nr_tres,
+                       lldev->evca + HIDMA_EVCA_RING_LEN_REG);
+
+       /* support IRQ only for now */
+       val = readl(lldev->evca + HIDMA_EVCA_INTCTRL_REG);
+       val &= ~0xF;
+       val |= 0x1;
+       writel(val, lldev->evca + HIDMA_EVCA_INTCTRL_REG);
+
+       /* clear all pending interrupts and enable them */
+       writel(ENABLE_IRQS, lldev->evca + HIDMA_EVCA_IRQ_CLR_REG);
+       writel(ENABLE_IRQS, lldev->evca + HIDMA_EVCA_IRQ_EN_REG);
+
+       return hidma_ll_enable(lldev);
+}
+
+struct hidma_lldev *hidma_ll_init(struct device *dev, u32 nr_tres,
+                                 void __iomem *trca, void __iomem *evca,
+                                 u8 chidx)
+{
+       u32 required_bytes;
+       struct hidma_lldev *lldev;
+       int rc;
+       size_t sz;
+
+       if (!trca || !evca || !dev || !nr_tres)
+               return NULL;
+
+       /* need at least four TREs */
+       if (nr_tres < 4)
+               return NULL;
+
+       /* need an extra space */
+       nr_tres += 1;
+
+       lldev = devm_kzalloc(dev, sizeof(struct hidma_lldev), GFP_KERNEL);
+       if (!lldev)
+               return NULL;
+
+       lldev->evca = evca;
+       lldev->trca = trca;
+       lldev->dev = dev;
+       sz = sizeof(struct hidma_tre);
+       lldev->trepool = devm_kcalloc(lldev->dev, nr_tres, sz, GFP_KERNEL);
+       if (!lldev->trepool)
+               return NULL;
+
+       required_bytes = sizeof(lldev->pending_tre_list[0]);
+       lldev->pending_tre_list = devm_kcalloc(dev, nr_tres, required_bytes,
+                                              GFP_KERNEL);
+       if (!lldev->pending_tre_list)
+               return NULL;
+
+       sz = (HIDMA_TRE_SIZE + 1) * nr_tres;
+       lldev->tre_ring = dmam_alloc_coherent(dev, sz, &lldev->tre_dma,
+                                             GFP_KERNEL);
+       if (!lldev->tre_ring)
+               return NULL;
+
+       memset(lldev->tre_ring, 0, (HIDMA_TRE_SIZE + 1) * nr_tres);
+       lldev->tre_ring_size = HIDMA_TRE_SIZE * nr_tres;
+       lldev->nr_tres = nr_tres;
+
+       /* the TRE ring has to be TRE_SIZE aligned */
+       if (!IS_ALIGNED(lldev->tre_dma, HIDMA_TRE_SIZE)) {
+               u8 tre_ring_shift;
+
+               tre_ring_shift = lldev->tre_dma % HIDMA_TRE_SIZE;
+               tre_ring_shift = HIDMA_TRE_SIZE - tre_ring_shift;
+               lldev->tre_dma += tre_ring_shift;
+               lldev->tre_ring += tre_ring_shift;
+       }
+
+       sz = (HIDMA_EVRE_SIZE + 1) * nr_tres;
+       lldev->evre_ring = dmam_alloc_coherent(dev, sz, &lldev->evre_dma,
+                                              GFP_KERNEL);
+       if (!lldev->evre_ring)
+               return NULL;
+
+       memset(lldev->evre_ring, 0, (HIDMA_EVRE_SIZE + 1) * nr_tres);
+       lldev->evre_ring_size = HIDMA_EVRE_SIZE * nr_tres;
+
+       /* the EVRE ring has to be EVRE_SIZE aligned */
+       if (!IS_ALIGNED(lldev->evre_dma, HIDMA_EVRE_SIZE)) {
+               u8 evre_ring_shift;
+
+               evre_ring_shift = lldev->evre_dma % HIDMA_EVRE_SIZE;
+               evre_ring_shift = HIDMA_EVRE_SIZE - evre_ring_shift;
+               lldev->evre_dma += evre_ring_shift;
+               lldev->evre_ring += evre_ring_shift;
+       }
+       lldev->nr_tres = nr_tres;
+       lldev->chidx = chidx;
+
+       sz = nr_tres * sizeof(struct hidma_tre *);
+       rc = kfifo_alloc(&lldev->handoff_fifo, sz, GFP_KERNEL);
+       if (rc)
+               return NULL;
+
+       rc = hidma_ll_setup(lldev);
+       if (rc)
+               return NULL;
+
+       spin_lock_init(&lldev->lock);
+       tasklet_init(&lldev->rst_task, hidma_ll_abort, (unsigned long)lldev);
+       tasklet_init(&lldev->task, hidma_ll_tre_complete, (unsigned long)lldev);
+       lldev->initialized = 1;
+       writel(ENABLE_IRQS, lldev->evca + HIDMA_EVCA_IRQ_EN_REG);
+       return lldev;
+}
+
+int hidma_ll_uninit(struct hidma_lldev *lldev)
+{
+       u32 required_bytes;
+       int rc = 0;
+       u32 val;
+
+       if (!lldev)
+               return -ENODEV;
+
+       if (!lldev->initialized)
+               return 0;
+
+       lldev->initialized = 0;
+
+       required_bytes = sizeof(struct hidma_tre) * lldev->nr_tres;
+       tasklet_kill(&lldev->task);
+       memset(lldev->trepool, 0, required_bytes);
+       lldev->trepool = NULL;
+       lldev->pending_tre_count = 0;
+       lldev->tre_write_offset = 0;
+
+       rc = hidma_ll_reset(lldev);
+
+       /*
+        * Clear all pending interrupts again.
+        * Otherwise, we observe reset complete interrupts.
+        */
+       val = readl(lldev->evca + HIDMA_EVCA_IRQ_STAT_REG);
+       writel(val, lldev->evca + HIDMA_EVCA_IRQ_CLR_REG);
+       writel(0, lldev->evca + HIDMA_EVCA_IRQ_EN_REG);
+       return rc;
+}
+
+enum dma_status hidma_ll_status(struct hidma_lldev *lldev, u32 tre_ch)
+{
+       enum dma_status ret = DMA_ERROR;
+       struct hidma_tre *tre;
+       unsigned long flags;
+       u8 err_code;
+
+       spin_lock_irqsave(&lldev->lock, flags);
+
+       tre = &lldev->trepool[tre_ch];
+       err_code = tre->err_code;
+
+       if (err_code & HIDMA_EVRE_STATUS_COMPLETE)
+               ret = DMA_COMPLETE;
+       else if (err_code & HIDMA_EVRE_STATUS_ERROR)
+               ret = DMA_ERROR;
+       else
+               ret = DMA_IN_PROGRESS;
+       spin_unlock_irqrestore(&lldev->lock, flags);
+
+       return ret;
+}