dmaengine: ptdma: register PTDMA controller as a DMA resource
authorSanjay R Mehta <sanju.mehta@amd.com>
Tue, 17 Aug 2021 13:55:58 +0000 (08:55 -0500)
committerVinod Koul <vkoul@kernel.org>
Sun, 29 Aug 2021 13:44:20 +0000 (19:14 +0530)
Register ptdma queue to Linux dmaengine framework as general-purpose
DMA channels.

Signed-off-by: Sanjay R Mehta <sanju.mehta@amd.com>
Link: https://lore.kernel.org/r/1629208559-51964-3-git-send-email-Sanju.Mehta@amd.com
Signed-off-by: Vinod Koul <vkoul@kernel.org>
drivers/dma/ptdma/Kconfig
drivers/dma/ptdma/Makefile
drivers/dma/ptdma/ptdma-dev.c
drivers/dma/ptdma/ptdma-dmaengine.c [new file with mode: 0644]
drivers/dma/ptdma/ptdma.h

index e6b8ca1..b430edd 100644 (file)
@@ -2,6 +2,8 @@
 config AMD_PTDMA
        tristate  "AMD PassThru DMA Engine"
        depends on X86_64 && PCI
+       select DMA_ENGINE
+       select DMA_VIRTUAL_CHANNELS
        help
          Enable support for the AMD PTDMA controller. This controller
          provides DMA capabilities to perform high bandwidth memory to
index 320fa82..a528cb0 100644 (file)
@@ -5,6 +5,6 @@
 
 obj-$(CONFIG_AMD_PTDMA) += ptdma.o
 
-ptdma-objs := ptdma-dev.o
+ptdma-objs := ptdma-dev.o ptdma-dmaengine.o
 
 ptdma-$(CONFIG_PCI) += ptdma-pci.o
index 953cdd0..46e7eff 100644 (file)
@@ -123,6 +123,26 @@ static inline void pt_core_enable_queue_interrupts(struct pt_device *pt)
        iowrite32(SUPPORTED_INTERRUPTS, pt->cmd_q.reg_control + 0x000C);
 }
 
+static void pt_do_cmd_complete(unsigned long data)
+{
+       struct pt_tasklet_data *tdata = (struct pt_tasklet_data *)data;
+       struct pt_cmd *cmd = tdata->cmd;
+       struct pt_cmd_queue *cmd_q = &cmd->pt->cmd_q;
+       u32 tail;
+
+       if (cmd_q->cmd_error) {
+              /*
+               * Log the error and flush the queue by
+               * moving the head pointer
+               */
+               tail = lower_32_bits(cmd_q->qdma_tail + cmd_q->qidx * Q_DESC_SIZE);
+               pt_log_error(cmd_q->pt, cmd_q->cmd_error);
+               iowrite32(tail, cmd_q->reg_control + 0x0008);
+       }
+
+       cmd->pt_cmd_callback(cmd->data, cmd->ret);
+}
+
 static irqreturn_t pt_core_irq_handler(int irq, void *data)
 {
        struct pt_device *pt = data;
@@ -143,6 +163,7 @@ static irqreturn_t pt_core_irq_handler(int irq, void *data)
                /* Acknowledge the interrupt */
                iowrite32(status, cmd_q->reg_control + 0x0010);
                pt_core_enable_queue_interrupts(pt);
+               pt_do_cmd_complete((ulong)&pt->tdata);
        }
        return IRQ_HANDLED;
 }
@@ -224,8 +245,16 @@ int pt_core_init(struct pt_device *pt)
 
        pt_core_enable_queue_interrupts(pt);
 
+       /* Register the DMA engine support */
+       ret = pt_dmaengine_register(pt);
+       if (ret)
+               goto e_dmaengine;
+
        return 0;
 
+e_dmaengine:
+       free_irq(pt->pt_irq, pt);
+
 e_dma_alloc:
        dma_free_coherent(dev, cmd_q->qsize, cmd_q->qbase, cmd_q->qbase_dma);
 
@@ -242,6 +271,9 @@ void pt_core_destroy(struct pt_device *pt)
        struct pt_cmd_queue *cmd_q = &pt->cmd_q;
        struct pt_cmd *cmd;
 
+       /* Unregister the DMA engine */
+       pt_dmaengine_unregister(pt);
+
        /* Disable and clear interrupts */
        pt_core_disable_queue_interrupts(pt);
 
diff --git a/drivers/dma/ptdma/ptdma-dmaengine.c b/drivers/dma/ptdma/ptdma-dmaengine.c
new file mode 100644 (file)
index 0000000..c9e52f6
--- /dev/null
@@ -0,0 +1,389 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * AMD Passthrough DMA device driver
+ * -- Based on the CCP driver
+ *
+ * Copyright (C) 2016,2021 Advanced Micro Devices, Inc.
+ *
+ * Author: Sanjay R Mehta <sanju.mehta@amd.com>
+ * Author: Gary R Hook <gary.hook@amd.com>
+ */
+
+#include "ptdma.h"
+#include "../dmaengine.h"
+#include "../virt-dma.h"
+
+static inline struct pt_dma_chan *to_pt_chan(struct dma_chan *dma_chan)
+{
+       return container_of(dma_chan, struct pt_dma_chan, vc.chan);
+}
+
+static inline struct pt_dma_desc *to_pt_desc(struct virt_dma_desc *vd)
+{
+       return container_of(vd, struct pt_dma_desc, vd);
+}
+
+static void pt_free_chan_resources(struct dma_chan *dma_chan)
+{
+       struct pt_dma_chan *chan = to_pt_chan(dma_chan);
+
+       vchan_free_chan_resources(&chan->vc);
+}
+
+static void pt_synchronize(struct dma_chan *dma_chan)
+{
+       struct pt_dma_chan *chan = to_pt_chan(dma_chan);
+
+       vchan_synchronize(&chan->vc);
+}
+
+static void pt_do_cleanup(struct virt_dma_desc *vd)
+{
+       struct pt_dma_desc *desc = to_pt_desc(vd);
+       struct pt_device *pt = desc->pt;
+
+       kmem_cache_free(pt->dma_desc_cache, desc);
+}
+
+static int pt_dma_start_desc(struct pt_dma_desc *desc)
+{
+       struct pt_passthru_engine *pt_engine;
+       struct pt_device *pt;
+       struct pt_cmd *pt_cmd;
+       struct pt_cmd_queue *cmd_q;
+
+       desc->issued_to_hw = 1;
+
+       pt_cmd = &desc->pt_cmd;
+       pt = pt_cmd->pt;
+       cmd_q = &pt->cmd_q;
+       pt_engine = &pt_cmd->passthru;
+
+       pt->tdata.cmd = pt_cmd;
+
+       /* Execute the command */
+       pt_cmd->ret = pt_core_perform_passthru(cmd_q, pt_engine);
+
+       return 0;
+}
+
+static struct pt_dma_desc *pt_next_dma_desc(struct pt_dma_chan *chan)
+{
+       /* Get the next DMA descriptor on the active list */
+       struct virt_dma_desc *vd = vchan_next_desc(&chan->vc);
+
+       return vd ? to_pt_desc(vd) : NULL;
+}
+
+static struct pt_dma_desc *pt_handle_active_desc(struct pt_dma_chan *chan,
+                                                struct pt_dma_desc *desc)
+{
+       struct dma_async_tx_descriptor *tx_desc;
+       struct virt_dma_desc *vd;
+       unsigned long flags;
+
+       /* Loop over descriptors until one is found with commands */
+       do {
+               if (desc) {
+                       if (!desc->issued_to_hw) {
+                               /* No errors, keep going */
+                               if (desc->status != DMA_ERROR)
+                                       return desc;
+                       }
+
+                       tx_desc = &desc->vd.tx;
+                       vd = &desc->vd;
+               } else {
+                       tx_desc = NULL;
+               }
+
+               spin_lock_irqsave(&chan->vc.lock, flags);
+
+               if (desc) {
+                       if (desc->status != DMA_ERROR)
+                               desc->status = DMA_COMPLETE;
+
+                       dma_cookie_complete(tx_desc);
+                       dma_descriptor_unmap(tx_desc);
+                       list_del(&desc->vd.node);
+               }
+
+               desc = pt_next_dma_desc(chan);
+
+               spin_unlock_irqrestore(&chan->vc.lock, flags);
+
+               if (tx_desc) {
+                       dmaengine_desc_get_callback_invoke(tx_desc, NULL);
+                       dma_run_dependencies(tx_desc);
+                       vchan_vdesc_fini(vd);
+               }
+       } while (desc);
+
+       return NULL;
+}
+
+static void pt_cmd_callback(void *data, int err)
+{
+       struct pt_dma_desc *desc = data;
+       struct dma_chan *dma_chan;
+       struct pt_dma_chan *chan;
+       int ret;
+
+       if (err == -EINPROGRESS)
+               return;
+
+       dma_chan = desc->vd.tx.chan;
+       chan = to_pt_chan(dma_chan);
+
+       if (err)
+               desc->status = DMA_ERROR;
+
+       while (true) {
+               /* Check for DMA descriptor completion */
+               desc = pt_handle_active_desc(chan, desc);
+
+               /* Don't submit cmd if no descriptor or DMA is paused */
+               if (!desc)
+                       break;
+
+               ret = pt_dma_start_desc(desc);
+               if (!ret)
+                       break;
+
+               desc->status = DMA_ERROR;
+       }
+}
+
+static struct pt_dma_desc *pt_alloc_dma_desc(struct pt_dma_chan *chan,
+                                            unsigned long flags)
+{
+       struct pt_dma_desc *desc;
+
+       desc = kmem_cache_zalloc(chan->pt->dma_desc_cache, GFP_NOWAIT);
+       if (!desc)
+               return NULL;
+
+       vchan_tx_prep(&chan->vc, &desc->vd, flags);
+
+       desc->pt = chan->pt;
+       desc->issued_to_hw = 0;
+       desc->status = DMA_IN_PROGRESS;
+
+       return desc;
+}
+
+static struct pt_dma_desc *pt_create_desc(struct dma_chan *dma_chan,
+                                         dma_addr_t dst,
+                                         dma_addr_t src,
+                                         unsigned int len,
+                                         unsigned long flags)
+{
+       struct pt_dma_chan *chan = to_pt_chan(dma_chan);
+       struct pt_passthru_engine *pt_engine;
+       struct pt_dma_desc *desc;
+       struct pt_cmd *pt_cmd;
+
+       desc = pt_alloc_dma_desc(chan, flags);
+       if (!desc)
+               return NULL;
+
+       pt_cmd = &desc->pt_cmd;
+       pt_cmd->pt = chan->pt;
+       pt_engine = &pt_cmd->passthru;
+       pt_cmd->engine = PT_ENGINE_PASSTHRU;
+       pt_engine->src_dma = src;
+       pt_engine->dst_dma = dst;
+       pt_engine->src_len = len;
+       pt_cmd->pt_cmd_callback = pt_cmd_callback;
+       pt_cmd->data = desc;
+
+       desc->len = len;
+
+       return desc;
+}
+
+static struct dma_async_tx_descriptor *
+pt_prep_dma_memcpy(struct dma_chan *dma_chan, dma_addr_t dst,
+                  dma_addr_t src, size_t len, unsigned long flags)
+{
+       struct pt_dma_desc *desc;
+
+       desc = pt_create_desc(dma_chan, dst, src, len, flags);
+       if (!desc)
+               return NULL;
+
+       return &desc->vd.tx;
+}
+
+static struct dma_async_tx_descriptor *
+pt_prep_dma_interrupt(struct dma_chan *dma_chan, unsigned long flags)
+{
+       struct pt_dma_chan *chan = to_pt_chan(dma_chan);
+       struct pt_dma_desc *desc;
+
+       desc = pt_alloc_dma_desc(chan, flags);
+       if (!desc)
+               return NULL;
+
+       return &desc->vd.tx;
+}
+
+static void pt_issue_pending(struct dma_chan *dma_chan)
+{
+       struct pt_dma_chan *chan = to_pt_chan(dma_chan);
+       struct pt_dma_desc *desc;
+       unsigned long flags;
+
+       spin_lock_irqsave(&chan->vc.lock, flags);
+
+       vchan_issue_pending(&chan->vc);
+
+       desc = pt_next_dma_desc(chan);
+
+       spin_unlock_irqrestore(&chan->vc.lock, flags);
+
+       /* If there was nothing active, start processing */
+       if (desc)
+               pt_cmd_callback(desc, 0);
+}
+
+static int pt_pause(struct dma_chan *dma_chan)
+{
+       struct pt_dma_chan *chan = to_pt_chan(dma_chan);
+       unsigned long flags;
+
+       spin_lock_irqsave(&chan->vc.lock, flags);
+       pt_stop_queue(&chan->pt->cmd_q);
+       spin_unlock_irqrestore(&chan->vc.lock, flags);
+
+       return 0;
+}
+
+static int pt_resume(struct dma_chan *dma_chan)
+{
+       struct pt_dma_chan *chan = to_pt_chan(dma_chan);
+       struct pt_dma_desc *desc = NULL;
+       unsigned long flags;
+
+       spin_lock_irqsave(&chan->vc.lock, flags);
+       pt_start_queue(&chan->pt->cmd_q);
+       desc = pt_next_dma_desc(chan);
+       spin_unlock_irqrestore(&chan->vc.lock, flags);
+
+       /* If there was something active, re-start */
+       if (desc)
+               pt_cmd_callback(desc, 0);
+
+       return 0;
+}
+
+static int pt_terminate_all(struct dma_chan *dma_chan)
+{
+       struct pt_dma_chan *chan = to_pt_chan(dma_chan);
+       unsigned long flags;
+       LIST_HEAD(head);
+
+       spin_lock_irqsave(&chan->vc.lock, flags);
+       vchan_get_all_descriptors(&chan->vc, &head);
+       spin_unlock_irqrestore(&chan->vc.lock, flags);
+
+       vchan_dma_desc_free_list(&chan->vc, &head);
+       vchan_free_chan_resources(&chan->vc);
+
+       return 0;
+}
+
+int pt_dmaengine_register(struct pt_device *pt)
+{
+       struct pt_dma_chan *chan;
+       struct dma_device *dma_dev = &pt->dma_dev;
+       char *cmd_cache_name;
+       char *desc_cache_name;
+       int ret;
+
+       pt->pt_dma_chan = devm_kzalloc(pt->dev, sizeof(*pt->pt_dma_chan),
+                                      GFP_KERNEL);
+       if (!pt->pt_dma_chan)
+               return -ENOMEM;
+
+       cmd_cache_name = devm_kasprintf(pt->dev, GFP_KERNEL,
+                                       "%s-dmaengine-cmd-cache",
+                                       dev_name(pt->dev));
+       if (!cmd_cache_name)
+               return -ENOMEM;
+
+       desc_cache_name = devm_kasprintf(pt->dev, GFP_KERNEL,
+                                        "%s-dmaengine-desc-cache",
+                                        dev_name(pt->dev));
+       if (!desc_cache_name) {
+               ret = -ENOMEM;
+               goto err_cache;
+       }
+
+       pt->dma_desc_cache = kmem_cache_create(desc_cache_name,
+                                              sizeof(struct pt_dma_desc), 0,
+                                              SLAB_HWCACHE_ALIGN, NULL);
+       if (!pt->dma_desc_cache) {
+               ret = -ENOMEM;
+               goto err_cache;
+       }
+
+       dma_dev->dev = pt->dev;
+       dma_dev->src_addr_widths = DMA_SLAVE_BUSWIDTH_64_BYTES;
+       dma_dev->dst_addr_widths = DMA_SLAVE_BUSWIDTH_64_BYTES;
+       dma_dev->directions = DMA_MEM_TO_MEM;
+       dma_dev->residue_granularity = DMA_RESIDUE_GRANULARITY_DESCRIPTOR;
+       dma_cap_set(DMA_MEMCPY, dma_dev->cap_mask);
+       dma_cap_set(DMA_INTERRUPT, dma_dev->cap_mask);
+
+       /*
+        * PTDMA is intended to be used with the AMD NTB devices, hence
+        * marking it as DMA_PRIVATE.
+        */
+       dma_cap_set(DMA_PRIVATE, dma_dev->cap_mask);
+
+       INIT_LIST_HEAD(&dma_dev->channels);
+
+       chan = pt->pt_dma_chan;
+       chan->pt = pt;
+
+       /* Set base and prep routines */
+       dma_dev->device_free_chan_resources = pt_free_chan_resources;
+       dma_dev->device_prep_dma_memcpy = pt_prep_dma_memcpy;
+       dma_dev->device_prep_dma_interrupt = pt_prep_dma_interrupt;
+       dma_dev->device_issue_pending = pt_issue_pending;
+       dma_dev->device_tx_status = dma_cookie_status;
+       dma_dev->device_pause = pt_pause;
+       dma_dev->device_resume = pt_resume;
+       dma_dev->device_terminate_all = pt_terminate_all;
+       dma_dev->device_synchronize = pt_synchronize;
+
+       chan->vc.desc_free = pt_do_cleanup;
+       vchan_init(&chan->vc, dma_dev);
+
+       dma_set_mask_and_coherent(pt->dev, DMA_BIT_MASK(64));
+
+       ret = dma_async_device_register(dma_dev);
+       if (ret)
+               goto err_reg;
+
+       return 0;
+
+err_reg:
+       kmem_cache_destroy(pt->dma_desc_cache);
+
+err_cache:
+       kmem_cache_destroy(pt->dma_cmd_cache);
+
+       return ret;
+}
+
+void pt_dmaengine_unregister(struct pt_device *pt)
+{
+       struct dma_device *dma_dev = &pt->dma_dev;
+
+       dma_async_device_unregister(dma_dev);
+
+       kmem_cache_destroy(pt->dma_desc_cache);
+       kmem_cache_destroy(pt->dma_cmd_cache);
+}
index f195b15..1329c55 100644 (file)
@@ -14,6 +14,7 @@
 #define __PT_DEV_H__
 
 #include <linux/device.h>
+#include <linux/dmaengine.h>
 #include <linux/pci.h>
 #include <linux/spinlock.h>
 #include <linux/mutex.h>
@@ -21,6 +22,8 @@
 #include <linux/wait.h>
 #include <linux/dmapool.h>
 
+#include "../virt-dma.h"
+
 #define MAX_PT_NAME_LEN                        16
 #define MAX_DMAPOOL_NAME_LEN           32
 
@@ -170,6 +173,20 @@ struct pt_cmd {
        void *data;
 };
 
+struct pt_dma_desc {
+       struct virt_dma_desc vd;
+       struct pt_device *pt;
+       enum dma_status status;
+       size_t len;
+       bool issued_to_hw;
+       struct pt_cmd pt_cmd;
+};
+
+struct pt_dma_chan {
+       struct virt_dma_chan vc;
+       struct pt_device *pt;
+};
+
 struct pt_cmd_queue {
        struct pt_device *pt;
 
@@ -229,6 +246,12 @@ struct pt_device {
         */
        struct pt_cmd_queue cmd_q;
 
+       /* Support for the DMA Engine capabilities */
+       struct dma_device dma_dev;
+       struct pt_dma_chan *pt_dma_chan;
+       struct kmem_cache *dma_cmd_cache;
+       struct kmem_cache *dma_desc_cache;
+
        wait_queue_head_t lsb_queue;
 
        struct pt_tasklet_data tdata;
@@ -281,6 +304,9 @@ struct pt_dev_vdata {
        const unsigned int bar;
 };
 
+int pt_dmaengine_register(struct pt_device *pt);
+void pt_dmaengine_unregister(struct pt_device *pt);
+
 int pt_core_init(struct pt_device *pt);
 void pt_core_destroy(struct pt_device *pt);