From: Sanjay R Mehta <[email protected]>
This patch series adds support for AMD PTDMA controller which
performs high bandwidth memory-to-memory and IO copy operation,
performs DMA transfer through queue based descriptor management.
AMD Processor has multiple ptdma device instances with each controller
having single queue. The driver also adds support for for multiple PTDMA
instances, each device will get an unique identifier and uniquely
named resources.
v7:
- Fixed module warnings reported ( by kernel test robot <[email protected]> ).
v6:
- Removed debug artifacts and made the suggested cosmetic changes.
- implemented and used to_pt_chan and to_pt_desc inline functions.
- Removed src and dst address check as framework does this.
- Removed devm_kcalloc() usage and used devm_kzalloc() api.
- Using framework debugfs directory to store dma info.
v5:
- modified code to submit next tranction in ISR itself and removed the tasklet.
- implemented .device_synchronize API.
- converted debugfs code by using DEFINE_SHOW_ATTRIBUTE()
- using dbg_dev_root for debugfs root directory.
- removed dma_status from pt_dma_chan
- removed module parameter cmd_queue_lenght.
- removed global device list for multiple devics.
- removed code related to dynamic adding/deleting to device list
- removed pt_add_device and pt_del_device functions
v4:
- modified DMA channel and descriptor management using virt-dma layer
instead of list based management.
- return only status of the cookie from pt_tx_status
- copyright year changed from 2019 to 2020
- removed dummy code for suspend & resume
- used bitmask and genmask
v3:
- Fixed the sparse warnings.
v2:
- Added controller description in cover letter
- Removed "default m" from Kconfig
- Replaced low_address() and high_address() functions with kernel
API's lower_32_bits & upper_32_bits().
- Removed the BH handler function pt_core_irq_bh() and instead
handling transaction in irq handler itself.
- Moved presetting of command queue registers into new function
"init_cmdq_regs()"
- Removed the kernel thread dependency to submit transaction.
- Increased the hardware command queue size to 32 and adding it
as a module parameter.
- Removed backlog command queue handling mechanism.
- Removed software command queue handling and instead submitting
transaction command directly to
hardware command queue.
- Added tasklet structure variable in "struct pt_device".
This is used to invoke pt_do_cmd_complete() upon receiving interrupt
for command completion.
- pt_core_perform_passthru() function parameters are modified and it is
now used to submit command directly to hardware from dmaengine framew
- Removed below structures, enums, macros and functions, as these value
constants. Making command submission simple,
- Removed "union pt_function" and several macros like PT_VERSION,
PT_BYTESWAP, PT_CMD_* etc..
- enum pt_passthru_bitwise, enum pt_passthru_byteswap, enum pt_memty
struct pt_dma_info, struct pt_data, struct pt_mem, struct pt_passt
struct pt_op,
Links of the review comments for v5:
1. https://lkml.org/lkml/2020/7/3/154
2. https://lkml.org/lkml/2020/8/25/431
3. https://lkml.org/lkml/2020/7/3/177
4. https://lkml.org/lkml/2020/7/3/186
Links of the review comments for v5:
1. https://lkml.org/lkml/2020/5/4/42
2. https://lkml.org/lkml/2020/5/4/45
3. https://lkml.org/lkml/2020/5/4/38
4. https://lkml.org/lkml/2020/5/26/70
Links of the review comments for v4:
1. https://lkml.org/lkml/2020/1/24/12
2. https://lkml.org/lkml/2020/1/24/17
Links of the review comments for v2:
1https://lkml.org/lkml/2019/12/27/630
2. https://lkml.org/lkml/2020/1/3/23
3. https://lkml.org/lkml/2020/1/3/314
4. https://lkml.org/lkml/2020/1/10/100
Links of the review comments for v1:
1. https://lkml.org/lkml/2019/9/24/490
2. https://lkml.org/lkml/2019/9/24/399
3. https://lkml.org/lkml/2019/9/24/862
4. https://lkml.org/lkml/2019/9/24/122
Sanjay R Mehta (3):
dmaengine: ptdma: Initial driver for the AMD PTDMA
dmaengine: ptdma: register PTDMA controller as a DMA resource
dmaengine: ptdma: Add debugfs entries for PTDMA information
MAINTAINERS | 6 +
drivers/dma/Kconfig | 2 +
drivers/dma/Makefile | 1 +
drivers/dma/ptdma/Kconfig | 13 +
drivers/dma/ptdma/Makefile | 10 +
drivers/dma/ptdma/ptdma-debugfs.c | 115 ++++++++
drivers/dma/ptdma/ptdma-dev.c | 333 ++++++++++++++++++++++
drivers/dma/ptdma/ptdma-dmaengine.c | 554 ++++++++++++++++++++++++++++++++++++
drivers/dma/ptdma/ptdma-pci.c | 252 ++++++++++++++++
drivers/dma/ptdma/ptdma.h | 352 +++++++++++++++++++++++
10 files changed, 1638 insertions(+)
create mode 100644 drivers/dma/ptdma/Kconfig
create mode 100644 drivers/dma/ptdma/Makefile
create mode 100644 drivers/dma/ptdma/ptdma-debugfs.c
create mode 100644 drivers/dma/ptdma/ptdma-dev.c
create mode 100644 drivers/dma/ptdma/ptdma-dmaengine.c
create mode 100644 drivers/dma/ptdma/ptdma-pci.c
create mode 100644 drivers/dma/ptdma/ptdma.h
--
2.7.4
From: Sanjay R Mehta <[email protected]>
Adding support for AMD PTDMA controller. It performs high-bandwidth
memory to memory and IO copy operation. Device commands are managed
via a circular queue of 'descriptors', each of which specifies source
and destination addresses for copying a single buffer of data.
Signed-off-by: Sanjay R Mehta <[email protected]>
---
MAINTAINERS | 6 +
drivers/dma/Kconfig | 2 +
drivers/dma/Makefile | 1 +
drivers/dma/ptdma/Kconfig | 11 ++
drivers/dma/ptdma/Makefile | 10 ++
drivers/dma/ptdma/ptdma-dev.c | 296 +++++++++++++++++++++++++++++++++++++++
drivers/dma/ptdma/ptdma-pci.c | 252 +++++++++++++++++++++++++++++++++
drivers/dma/ptdma/ptdma.h | 316 ++++++++++++++++++++++++++++++++++++++++++
8 files changed, 894 insertions(+)
create mode 100644 drivers/dma/ptdma/Kconfig
create mode 100644 drivers/dma/ptdma/Makefile
create mode 100644 drivers/dma/ptdma/ptdma-dev.c
create mode 100644 drivers/dma/ptdma/ptdma-pci.c
create mode 100644 drivers/dma/ptdma/ptdma.h
diff --git a/MAINTAINERS b/MAINTAINERS
index 33b27e6..f6ae0d1 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -943,6 +943,12 @@ S: Supported
F: arch/arm64/boot/dts/amd/amd-seattle-xgbe*.dtsi
F: drivers/net/ethernet/amd/xgbe/
+AMD PTDMA DRIVER
+M: Sanjay R Mehta <[email protected]>
+L: [email protected]
+S: Maintained
+F: drivers/dma/ptdma/
+
ANALOG DEVICES INC AD5686 DRIVER
M: Michael Hennerich <[email protected]>
L: [email protected]
diff --git a/drivers/dma/Kconfig b/drivers/dma/Kconfig
index 518a143..a21c983 100644
--- a/drivers/dma/Kconfig
+++ b/drivers/dma/Kconfig
@@ -748,6 +748,8 @@ source "drivers/dma/ti/Kconfig"
source "drivers/dma/fsl-dpaa2-qdma/Kconfig"
+source "drivers/dma/ptdma/Kconfig"
+
# clients
comment "DMA Clients"
depends on DMA_ENGINE
diff --git a/drivers/dma/Makefile b/drivers/dma/Makefile
index e60f813..2785756 100644
--- a/drivers/dma/Makefile
+++ b/drivers/dma/Makefile
@@ -83,6 +83,7 @@ obj-$(CONFIG_XGENE_DMA) += xgene-dma.o
obj-$(CONFIG_ZX_DMA) += zx_dma.o
obj-$(CONFIG_ST_FDMA) += st_fdma.o
obj-$(CONFIG_FSL_DPAA2_QDMA) += fsl-dpaa2-qdma/
+obj-$(CONFIG_AMD_PTDMA) += ptdma/
obj-y += mediatek/
obj-y += qcom/
diff --git a/drivers/dma/ptdma/Kconfig b/drivers/dma/ptdma/Kconfig
new file mode 100644
index 0000000..f93f9c2
--- /dev/null
+++ b/drivers/dma/ptdma/Kconfig
@@ -0,0 +1,11 @@
+# SPDX-License-Identifier: GPL-2.0-only
+config AMD_PTDMA
+ tristate "AMD PassThru DMA Engine"
+ depends on X86_64 && PCI
+ help
+ Enable support for the AMD PTDMA controller. This controller
+ provides DMA capabilities & performs high bandwidth memory to
+ memory and IO copy operation and performs DMA transfer through
+ queue based descriptor management. This DMA controller is intended
+ to use with AMD Non-Transparent Bridge devices and not for general
+ purpose slave DMA.
diff --git a/drivers/dma/ptdma/Makefile b/drivers/dma/ptdma/Makefile
new file mode 100644
index 0000000..320fa82
--- /dev/null
+++ b/drivers/dma/ptdma/Makefile
@@ -0,0 +1,10 @@
+# SPDX-License-Identifier: GPL-2.0-only
+#
+# AMD Passthru DMA driver
+#
+
+obj-$(CONFIG_AMD_PTDMA) += ptdma.o
+
+ptdma-objs := ptdma-dev.o
+
+ptdma-$(CONFIG_PCI) += ptdma-pci.o
diff --git a/drivers/dma/ptdma/ptdma-dev.c b/drivers/dma/ptdma/ptdma-dev.c
new file mode 100644
index 0000000..d1e9892
--- /dev/null
+++ b/drivers/dma/ptdma/ptdma-dev.c
@@ -0,0 +1,296 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * AMD Passthru DMA device driver
+ * -- Based on the CCP driver
+ *
+ * Copyright (C) 2016,2020 Advanced Micro Devices, Inc.
+ *
+ * Author: Sanjay R Mehta <[email protected]>
+ * Author: Gary R Hook <[email protected]>
+ */
+
+#include <linux/dma-mapping.h>
+#include <linux/debugfs.h>
+#include <linux/interrupt.h>
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/pci.h>
+
+#include "ptdma.h"
+
+/* Human-readable error strings */
+static char *pt_error_codes[] = {
+ "",
+ "ERR 01: ILLEGAL_ENGINE",
+ "ERR 03: ILLEGAL_FUNCTION_TYPE",
+ "ERR 04: ILLEGAL_FUNCTION_MODE",
+ "ERR 06: ILLEGAL_FUNCTION_SIZE",
+ "ERR 08: ILLEGAL_FUNCTION_RSVD",
+ "ERR 09: ILLEGAL_BUFFER_LENGTH",
+ "ERR 10: VLSB_FAULT",
+ "ERR 11: ILLEGAL_MEM_ADDR",
+ "ERR 12: ILLEGAL_MEM_SEL",
+ "ERR 13: ILLEGAL_CONTEXT_ID",
+ "ERR 15: 0xF Reserved",
+ "ERR 18: CMD_TIMEOUT",
+ "ERR 19: IDMA0_AXI_SLVERR",
+ "ERR 20: IDMA0_AXI_DECERR",
+ "ERR 21: 0x15 Reserved",
+ "ERR 22: IDMA1_AXI_SLAVE_FAULT",
+ "ERR 23: IDMA1_AIXI_DECERR",
+ "ERR 24: 0x18 Reserved",
+ "ERR 27: 0x1B Reserved",
+ "ERR 38: ODMA0_AXI_SLVERR",
+ "ERR 39: ODMA0_AXI_DECERR",
+ "ERR 40: 0x28 Reserved",
+ "ERR 41: ODMA1_AXI_SLVERR",
+ "ERR 42: ODMA1_AXI_DECERR",
+ "ERR 43: LSB_PARITY_ERR",
+};
+
+static void pt_log_error(struct pt_device *d, int e)
+{
+ dev_err(d->dev, "PTDMA error: %s (0x%x)\n", pt_error_codes[e], e);
+}
+
+void pt_start_queue(struct pt_cmd_queue *cmd_q)
+{
+ /* Turn on the run bit */
+ iowrite32(cmd_q->qcontrol | CMD_Q_RUN, cmd_q->reg_control);
+}
+
+void pt_stop_queue(struct pt_cmd_queue *cmd_q)
+{
+ /* Turn off the run bit */
+ iowrite32(cmd_q->qcontrol & ~CMD_Q_RUN, cmd_q->reg_control);
+}
+
+static int pt_core_execute_cmd(struct ptdma_desc *desc,
+ struct pt_cmd_queue *cmd_q)
+{
+ __le32 *mp;
+ u32 *dp;
+ u32 tail;
+ int i;
+
+ if (desc->dw0.soc) {
+ desc->dw0.ioc = 1;
+ desc->dw0.soc = 0;
+ }
+ mutex_lock(&cmd_q->q_mutex);
+
+ mp = (__le32 *)&cmd_q->qbase[cmd_q->qidx];
+ dp = (u32 *)desc;
+ for (i = 0; i < 8; i++)
+ mp[i] = cpu_to_le32(dp[i]); /* handle endianness */
+
+ cmd_q->qidx = (cmd_q->qidx + 1) % CMD_Q_LEN;
+
+ /* The data used by this command must be flushed to memory */
+ wmb();
+
+ /* Write the new tail address back to the queue register */
+ tail = lower_32_bits(cmd_q->qdma_tail + cmd_q->qidx * Q_DESC_SIZE);
+ iowrite32(tail, cmd_q->reg_tail_lo);
+
+ /* Turn the queue back on using our cached control register */
+ pt_start_queue(cmd_q);
+ mutex_unlock(&cmd_q->q_mutex);
+
+ return 0;
+}
+
+int pt_core_perform_passthru(struct pt_cmd_queue *cmd_q,
+ struct pt_passthru_engine *pt_engine)
+{
+ struct ptdma_desc desc;
+
+ cmd_q->cmd_error = 0;
+ memset(&desc, 0, sizeof(desc));
+ desc.dw0.val = CMD_DESC_DW0_VAL;
+ desc.length = pt_engine->src_len;
+ desc.src_lo = lower_32_bits(pt_engine->src_dma);
+ desc.dw3.src_hi = upper_32_bits(pt_engine->src_dma);
+ desc.dst_lo = lower_32_bits(pt_engine->dst_dma);
+ desc.dw5.dst_hi = upper_32_bits(pt_engine->dst_dma);
+
+ return pt_core_execute_cmd(&desc, cmd_q);
+}
+
+static inline void pt_core_disable_queue_interrupts(struct pt_device *pt)
+{
+ iowrite32(0, pt->cmd_q.reg_int_enable);
+}
+
+static inline void pt_core_enable_queue_interrupts(struct pt_device *pt)
+{
+ iowrite32(SUPPORTED_INTERRUPTS, pt->cmd_q.reg_int_enable);
+}
+
+static irqreturn_t pt_core_irq_handler(int irq, void *data)
+{
+ struct pt_device *pt = (struct pt_device *)data;
+ struct pt_cmd_queue *cmd_q = &pt->cmd_q;
+ u32 status;
+
+ pt_core_disable_queue_interrupts(pt);
+
+ status = ioread32(cmd_q->reg_interrupt_status);
+ if (status) {
+ cmd_q->int_status = status;
+ cmd_q->q_status = ioread32(cmd_q->reg_status);
+ cmd_q->q_int_status = ioread32(cmd_q->reg_int_status);
+
+ /* On error, only save the first error value */
+ if ((status & INT_ERROR) && !cmd_q->cmd_error)
+ cmd_q->cmd_error = CMD_Q_ERROR(cmd_q->q_status);
+
+ /* Acknowledge the interrupt */
+ iowrite32(status, cmd_q->reg_interrupt_status);
+ }
+
+ pt_core_enable_queue_interrupts(pt);
+
+ return IRQ_HANDLED;
+}
+
+static void pt_init_cmdq_regs(struct pt_cmd_queue *cmd_q)
+{
+ void __iomem *io_regs = cmd_q->reg_control;
+
+ cmd_q->reg_tail_lo = io_regs + CMD_Q_TAIL_LO_BASE;
+ cmd_q->reg_head_lo = io_regs + CMD_Q_HEAD_LO_BASE;
+ cmd_q->reg_status = io_regs + CMD_Q_STATUS_BASE;
+ cmd_q->reg_int_enable = io_regs + CMD_Q_INT_ENABLE_BASE;
+ cmd_q->reg_int_status = io_regs + CMD_Q_INT_STATUS_BASE;
+ cmd_q->reg_dma_status = io_regs + CMD_Q_DMA_STATUS_BASE;
+ cmd_q->reg_dma_read_status = io_regs + CMD_Q_DMA_READ_STATUS_BASE;
+ cmd_q->reg_dma_write_status = io_regs + CMD_Q_DMA_WRITE_STATUS_BASE;
+ cmd_q->reg_interrupt_status = io_regs + CMD_Q_INTERRUPT_STATUS_BASE;
+}
+
+int pt_core_init(struct pt_device *pt)
+{
+ char dma_pool_name[MAX_DMAPOOL_NAME_LEN];
+ struct pt_cmd_queue *cmd_q = &pt->cmd_q;
+ u32 dma_addr_lo, dma_addr_hi;
+ struct device *dev = pt->dev;
+ struct dma_pool *dma_pool;
+ int ret;
+
+ /* Allocate a dma pool for the queue */
+ snprintf(dma_pool_name, sizeof(dma_pool_name), "%s_q", pt->name);
+
+ dma_pool = dma_pool_create(dma_pool_name, dev,
+ PT_DMAPOOL_MAX_SIZE,
+ PT_DMAPOOL_ALIGN, 0);
+ if (!dma_pool) {
+ dev_err(dev, "unable to allocate dma pool\n");
+ ret = -ENOMEM;
+ return ret;
+ }
+
+ /* ptdma core initialisation */
+ iowrite32(CMD_CONFIG_VHB_EN, pt->io_regs + CMD_CONFIG_OFFSET);
+ iowrite32(CMD_QUEUE_PRIO, pt->io_regs + CMD_QUEUE_PRIO_OFFSET);
+ iowrite32(CMD_TIMEOUT_DISABLE, pt->io_regs + CMD_TIMEOUT_OFFSET);
+ iowrite32(CMD_CLK_GATE_CONFIG, pt->io_regs + CMD_CLK_GATE_CTL_OFFSET);
+ iowrite32(CMD_CONFIG_REQID, pt->io_regs + CMD_REQID_CONFIG_OFFSET);
+
+ cmd_q->pt = pt;
+ cmd_q->dma_pool = dma_pool;
+ mutex_init(&cmd_q->q_mutex);
+
+ /* Page alignment satisfies our needs for N <= 128 */
+ cmd_q->qsize = Q_SIZE(Q_DESC_SIZE);
+ cmd_q->qbase = dma_alloc_coherent(dev, cmd_q->qsize,
+ &cmd_q->qbase_dma,
+ GFP_KERNEL);
+ if (!cmd_q->qbase) {
+ dev_err(dev, "unable to allocate command queue\n");
+ ret = -ENOMEM;
+ goto e_dma_alloc;
+ }
+
+ cmd_q->qidx = 0;
+
+ /* Preset some register values */
+ cmd_q->reg_control = pt->io_regs + CMD_Q_STATUS_INCR;
+ pt_init_cmdq_regs(cmd_q);
+
+ /* Turn off the queues and disable interrupts until ready */
+ pt_core_disable_queue_interrupts(pt);
+
+ cmd_q->qcontrol = 0; /* Start with nothing */
+ iowrite32(cmd_q->qcontrol, cmd_q->reg_control);
+
+ ioread32(cmd_q->reg_int_status);
+ ioread32(cmd_q->reg_status);
+
+ /* Clear the interrupt status */
+ iowrite32(SUPPORTED_INTERRUPTS, cmd_q->reg_interrupt_status);
+
+ /* Request an irq */
+ ret = request_irq(pt->pt_irq, pt_core_irq_handler, 0, pt->name, pt);
+ if (ret) {
+ dev_err(dev, "unable to allocate an IRQ\n");
+ goto e_pool;
+ }
+
+ /* Update the device registers with queue information. */
+
+ cmd_q->qcontrol &= ~(CMD_Q_SIZE << CMD_Q_SHIFT);
+ cmd_q->qcontrol |= QUEUE_SIZE_VAL << CMD_Q_SHIFT;
+
+ cmd_q->qdma_tail = cmd_q->qbase_dma;
+ dma_addr_lo = lower_32_bits(cmd_q->qdma_tail);
+ iowrite32((u32)dma_addr_lo, cmd_q->reg_tail_lo);
+ iowrite32((u32)dma_addr_lo, cmd_q->reg_head_lo);
+
+ dma_addr_hi = upper_32_bits(cmd_q->qdma_tail);
+ cmd_q->qcontrol |= (dma_addr_hi << 16);
+ iowrite32(cmd_q->qcontrol, cmd_q->reg_control);
+
+ pt_core_enable_queue_interrupts(pt);
+
+ return 0;
+
+e_dma_alloc:
+ dma_free_coherent(dev, cmd_q->qsize, cmd_q->qbase, cmd_q->qbase_dma);
+
+e_pool:
+ dma_pool_destroy(pt->cmd_q.dma_pool);
+
+ return ret;
+}
+
+void pt_core_destroy(struct pt_device *pt)
+{
+ struct device *dev = pt->dev;
+ struct pt_cmd_queue *cmd_q = &pt->cmd_q;
+ struct pt_cmd *cmd;
+
+ /* Disable and clear interrupts */
+ pt_core_disable_queue_interrupts(pt);
+
+ /* Turn off the run bit */
+ pt_stop_queue(cmd_q);
+
+ /* Clear the interrupt status */
+ iowrite32(SUPPORTED_INTERRUPTS, cmd_q->reg_interrupt_status);
+ ioread32(cmd_q->reg_int_status);
+ ioread32(cmd_q->reg_status);
+
+ free_irq(pt->pt_irq, pt);
+
+ dma_free_coherent(dev, cmd_q->qsize, cmd_q->qbase,
+ cmd_q->qbase_dma);
+
+ /* Flush the cmd queue */
+ while (!list_empty(&pt->cmd)) {
+ /* Invoke the callback directly with an error code */
+ cmd = list_first_entry(&pt->cmd, struct pt_cmd, entry);
+ list_del(&cmd->entry);
+ cmd->pt_cmd_callback(cmd->data, -ENODEV);
+ }
+}
diff --git a/drivers/dma/ptdma/ptdma-pci.c b/drivers/dma/ptdma/ptdma-pci.c
new file mode 100644
index 0000000..5cbb6f3
--- /dev/null
+++ b/drivers/dma/ptdma/ptdma-pci.c
@@ -0,0 +1,252 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * AMD Passthru DMA device driver
+ * -- Based on the CCP driver
+ *
+ * Copyright (C) 2016,2020 Advanced Micro Devices, Inc.
+ *
+ * Author: Sanjay R Mehta <[email protected]>
+ * Author: Tom Lendacky <[email protected]>
+ * Author: Gary R Hook <[email protected]>
+ */
+
+#include <linux/device.h>
+#include <linux/dma-mapping.h>
+#include <linux/delay.h>
+#include <linux/interrupt.h>
+#include <linux/kernel.h>
+#include <linux/kthread.h>
+#include <linux/module.h>
+#include <linux/pci_ids.h>
+#include <linux/pci.h>
+#include <linux/spinlock.h>
+#include <linux/sched.h>
+
+#include "ptdma.h"
+
+/* Ever-increasing value to produce unique unit numbers */
+static atomic_t pt_ordinal;
+
+struct pt_msix {
+ int msix_count;
+ struct msix_entry msix_entry;
+};
+
+/*
+ * pt_alloc_struct - allocate and initialize the pt_device struct
+ *
+ * @dev: device struct of the PTDMA
+ */
+static struct pt_device *pt_alloc_struct(struct device *dev)
+{
+ struct pt_device *pt;
+
+ pt = devm_kzalloc(dev, sizeof(*pt), GFP_KERNEL);
+
+ if (!pt)
+ return NULL;
+ pt->dev = dev;
+ pt->ord = atomic_inc_return(&pt_ordinal);
+
+ INIT_LIST_HEAD(&pt->cmd);
+
+ snprintf(pt->name, MAX_PT_NAME_LEN, "pt-%u", pt->ord);
+
+ return pt;
+}
+
+static int pt_get_msix_irqs(struct pt_device *pt)
+{
+ struct pt_msix *pt_msix = pt->pt_msix;
+ struct device *dev = pt->dev;
+ struct pci_dev *pdev = to_pci_dev(dev);
+ int ret;
+
+ pt_msix->msix_entry.entry = 0;
+
+ ret = pci_enable_msix_range(pdev, &pt_msix->msix_entry, 1, 1);
+ if (ret < 0)
+ return ret;
+
+ pt_msix->msix_count = ret;
+
+ pt->pt_irq = pt_msix->msix_entry.vector;
+
+ return 0;
+}
+
+static int pt_get_msi_irq(struct pt_device *pt)
+{
+ struct device *dev = pt->dev;
+ struct pci_dev *pdev = to_pci_dev(dev);
+ int ret;
+
+ ret = pci_enable_msi(pdev);
+ if (ret)
+ return ret;
+
+ pt->pt_irq = pdev->irq;
+
+ return 0;
+}
+
+static int pt_get_irqs(struct pt_device *pt)
+{
+ struct device *dev = pt->dev;
+ int ret;
+
+ ret = pt_get_msix_irqs(pt);
+ if (!ret)
+ return 0;
+
+ /* Couldn't get MSI-X vectors, try MSI */
+ dev_dbg(dev, "could not enable MSI-X (%d), trying MSI\n", ret);
+ ret = pt_get_msi_irq(pt);
+ if (!ret)
+ return 0;
+
+ /* Couldn't get MSI interrupt */
+ dev_dbg(dev, "could not enable MSI (%d)\n", ret);
+
+ return ret;
+}
+
+static void pt_free_irqs(struct pt_device *pt)
+{
+ struct pt_msix *pt_msix = pt->pt_msix;
+ struct device *dev = pt->dev;
+ struct pci_dev *pdev = to_pci_dev(dev);
+
+ if (pt_msix->msix_count)
+ pci_disable_msix(pdev);
+ else if (pt->pt_irq)
+ pci_disable_msi(pdev);
+
+ pt->pt_irq = 0;
+}
+
+static int pt_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
+{
+ struct pt_device *pt;
+ struct pt_msix *pt_msix;
+ struct device *dev = &pdev->dev;
+ void __iomem * const *iomap_table;
+ int bar_mask;
+ int ret = -ENOMEM;
+
+ pt = pt_alloc_struct(dev);
+ if (!pt)
+ goto e_err;
+
+ pt_msix = devm_kzalloc(dev, sizeof(*pt_msix), GFP_KERNEL);
+ if (!pt_msix)
+ goto e_err;
+
+ pt->pt_msix = pt_msix;
+ pt->dev_vdata = (struct pt_dev_vdata *)id->driver_data;
+ if (!pt->dev_vdata) {
+ ret = -ENODEV;
+ dev_err(dev, "missing driver data\n");
+ goto e_err;
+ }
+
+ ret = pcim_enable_device(pdev);
+ if (ret) {
+ dev_err(dev, "pcim_enable_device failed (%d)\n", ret);
+ goto e_err;
+ }
+
+ bar_mask = pci_select_bars(pdev, IORESOURCE_MEM);
+ ret = pcim_iomap_regions(pdev, bar_mask, "ptdma");
+ if (ret) {
+ dev_err(dev, "pcim_iomap_regions failed (%d)\n", ret);
+ goto e_err;
+ }
+
+ iomap_table = pcim_iomap_table(pdev);
+ if (!iomap_table) {
+ dev_err(dev, "pcim_iomap_table failed\n");
+ ret = -ENOMEM;
+ goto e_err;
+ }
+
+ pt->io_regs = iomap_table[pt->dev_vdata->bar];
+ if (!pt->io_regs) {
+ dev_err(dev, "ioremap failed\n");
+ ret = -ENOMEM;
+ goto e_err;
+ }
+
+ ret = pt_get_irqs(pt);
+ if (ret)
+ goto e_err;
+
+ pci_set_master(pdev);
+
+ ret = dma_set_mask_and_coherent(dev, DMA_BIT_MASK(48));
+ if (ret) {
+ ret = dma_set_mask_and_coherent(dev, DMA_BIT_MASK(32));
+ if (ret) {
+ dev_err(dev, "dma_set_mask_and_coherent failed (%d)\n",
+ ret);
+ goto e_err;
+ }
+ }
+
+ dev_set_drvdata(dev, pt);
+
+ if (pt->dev_vdata)
+ ret = pt_core_init(pt);
+
+ if (ret)
+ goto e_err;
+
+ dev_dbg(dev, "PTDMA enabled\n");
+
+ return 0;
+
+e_err:
+ dev_err(dev, "initialization failed\n");
+
+ return ret;
+}
+
+static void pt_pci_remove(struct pci_dev *pdev)
+{
+ struct device *dev = &pdev->dev;
+ struct pt_device *pt = dev_get_drvdata(dev);
+
+ if (!pt)
+ return;
+
+ if (pt->dev_vdata)
+ pt_core_destroy(pt);
+
+ pt_free_irqs(pt);
+}
+
+static const struct pt_dev_vdata dev_vdata[] = {
+ {
+ .bar = 2,
+ },
+};
+
+static const struct pci_device_id pt_pci_table[] = {
+ { PCI_VDEVICE(AMD, 0x1498), (kernel_ulong_t)&dev_vdata[0] },
+ /* Last entry must be zero */
+ { 0, }
+};
+MODULE_DEVICE_TABLE(pci, pt_pci_table);
+
+static struct pci_driver pt_pci_driver = {
+ .name = "ptdma",
+ .id_table = pt_pci_table,
+ .probe = pt_pci_probe,
+ .remove = pt_pci_remove,
+};
+
+module_pci_driver(pt_pci_driver);
+
+MODULE_AUTHOR("Sanjay R Mehta <[email protected]>");
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("AMD PassThru DMA driver");
diff --git a/drivers/dma/ptdma/ptdma.h b/drivers/dma/ptdma/ptdma.h
new file mode 100644
index 0000000..9e5819f
--- /dev/null
+++ b/drivers/dma/ptdma/ptdma.h
@@ -0,0 +1,316 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * AMD Passthru DMA device driver
+ * -- Based on the CCP driver
+ *
+ * Copyright (C) 2016,2020 Advanced Micro Devices, Inc.
+ *
+ * Author: Sanjay R Mehta <[email protected]>
+ * Author: Tom Lendacky <[email protected]>
+ * Author: Gary R Hook <[email protected]>
+ */
+
+#ifndef __PT_DEV_H__
+#define __PT_DEV_H__
+
+#include <linux/device.h>
+#include <linux/pci.h>
+#include <linux/spinlock.h>
+#include <linux/mutex.h>
+#include <linux/list.h>
+#include <linux/wait.h>
+#include <linux/dmapool.h>
+
+#define MAX_PT_NAME_LEN 16
+#define MAX_DMAPOOL_NAME_LEN 32
+
+#define MAX_HW_QUEUES 1
+#define MAX_CMD_QLEN 100
+
+#define PT_ENGINE_PASSTHRU 5
+#define PT_OFFSET 0x0
+
+#define PT_VSIZE 16
+#define PT_VMASK ((unsigned int)((1 << PT_VSIZE) - 1))
+
+/* Register Mappings */
+#define IRQ_MASK_REG 0x040
+#define IRQ_STATUS_REG 0x200
+
+#define CMD_Q_ERROR(__qs) ((__qs) & 0x0000003f)
+
+#define CMD_QUEUE_PRIO_OFFSET 0x00
+#define CMD_REQID_CONFIG_OFFSET 0x04
+#define CMD_TIMEOUT_OFFSET 0x08
+#define CMD_PT_VERSION 0x10
+
+#define CMD_Q_CONTROL_BASE 0x0000
+#define CMD_Q_TAIL_LO_BASE 0x0004
+#define CMD_Q_HEAD_LO_BASE 0x0008
+#define CMD_Q_INT_ENABLE_BASE 0x000C
+#define CMD_Q_INTERRUPT_STATUS_BASE 0x0010
+
+#define CMD_Q_STATUS_BASE 0x0100
+#define CMD_Q_INT_STATUS_BASE 0x0104
+#define CMD_Q_DMA_STATUS_BASE 0x0108
+#define CMD_Q_DMA_READ_STATUS_BASE 0x010C
+#define CMD_Q_DMA_WRITE_STATUS_BASE 0x0110
+#define CMD_Q_ABORT_BASE 0x0114
+#define CMD_Q_AX_CACHE_BASE 0x0118
+
+#define CMD_CONFIG_OFFSET 0x1120
+#define CMD_CLK_GATE_CTL_OFFSET 0x6004
+
+#define CMD_DESC_DW0_VAL 0x500012
+
+/* Address offset for virtual queue registers */
+#define CMD_Q_STATUS_INCR 0x1000
+
+/* Bit masks */
+#define CMD_CONFIG_REQID 0
+#define CMD_TIMEOUT_DISABLE 0
+#define CMD_CLK_DYN_GATING_DIS 0
+#define CMD_CLK_SW_GATE_MODE 0
+#define CMD_CLK_GATE_CTL 0
+#define CMD_QUEUE_PRIO GENMASK(2, 1)
+#define CMD_CONFIG_VHB_EN BIT(0)
+#define CMD_CLK_DYN_GATING_EN BIT(0)
+#define CMD_CLK_HW_GATE_MODE BIT(0)
+#define CMD_CLK_GATE_ON_DELAY BIT(12)
+#define CMD_CLK_GATE_OFF_DELAY BIT(12)
+
+#define CMD_CLK_GATE_CONFIG (CMD_CLK_GATE_CTL | \
+ CMD_CLK_HW_GATE_MODE | \
+ CMD_CLK_GATE_ON_DELAY | \
+ CMD_CLK_DYN_GATING_EN | \
+ CMD_CLK_GATE_OFF_DELAY)
+
+#define CMD_Q_LEN 32
+#define CMD_Q_RUN BIT(0)
+#define CMD_Q_HALT BIT(1)
+#define CMD_Q_MEM_LOCATION BIT(2)
+#define CMD_Q_SIZE GENMASK(4, 0)
+#define CMD_Q_SHIFT GENMASK(1, 0)
+#define QUEUE_SIZE_VAL ((ffs(CMD_Q_LEN) - 2) & \
+ CMD_Q_SIZE)
+#define Q_PTR_MASK (2 << (QUEUE_SIZE_VAL + 5) - 1)
+#define Q_DESC_SIZE sizeof(struct ptdma_desc)
+#define Q_SIZE(n) (CMD_Q_LEN * (n))
+
+#define INT_COMPLETION BIT(0)
+#define INT_ERROR BIT(1)
+#define INT_QUEUE_STOPPED BIT(2)
+#define INT_EMPTY_QUEUE BIT(3)
+#define SUPPORTED_INTERRUPTS (INT_COMPLETION | INT_ERROR)
+
+/****** Local Storage Block ******/
+#define LSB_START 0
+#define LSB_END 127
+#define LSB_COUNT (LSB_END - LSB_START + 1)
+
+#define PT_DMAPOOL_MAX_SIZE 64
+#define PT_DMAPOOL_ALIGN BIT(5)
+
+#define PT_PASSTHRU_BLOCKSIZE 512
+
+struct pt_device;
+
+struct pt_tasklet_data {
+ struct completion completion;
+ struct pt_cmd *cmd;
+};
+
+/*
+ * struct pt_passthru_engine - pass-through operation
+ * without performing DMA mapping
+ * @mask: mask to be applied to data
+ * @mask_len: length in bytes of mask
+ * @src_dma: data to be used for this operation
+ * @dst_dma: data produced by this operation
+ * @src_len: length in bytes of data used for this operation
+ *
+ * Variables required to be set when calling pt_enqueue_cmd():
+ * - bit_mod, byte_swap, src, dst, src_len
+ * - mask, mask_len if bit_mod is not PT_PASSTHRU_BITWISE_NOOP
+ */
+struct pt_passthru_engine {
+ dma_addr_t mask;
+ u32 mask_len; /* In bytes */
+
+ dma_addr_t src_dma, dst_dma;
+ u64 src_len; /* In bytes */
+};
+
+/*
+ * struct pt_cmd - PTDMA operation request
+ * @entry: list element
+ * @work: work element used for callbacks
+ * @pt: PT device to be run on
+ * @ret: operation return code
+ * @flags: cmd processing flags
+ * @engine: PTDMA operation to perform (passthru)
+ * @engine_error: PT engine return code
+ * @passthru: engine specific structures, refer to specific engine struct below
+ * @callback: operation completion callback function
+ * @data: parameter value to be supplied to the callback function
+ *
+ * Variables required to be set when calling pt_enqueue_cmd():
+ * - engine, callback
+ * - See the operation structures below for what is required for each
+ * operation.
+ */
+struct pt_cmd {
+ struct list_head entry;
+ struct work_struct work;
+ struct pt_device *pt;
+ int ret;
+ u32 engine;
+ u32 engine_error;
+ struct pt_passthru_engine passthru;
+ /* Completion callback support */
+ void (*pt_cmd_callback)(void *data, int err);
+ void *data;
+};
+
+struct pt_cmd_queue {
+ struct pt_device *pt;
+
+ /* Queue dma pool */
+ struct dma_pool *dma_pool;
+
+ /* Queue base address (not neccessarily aligned)*/
+ struct ptdma_desc *qbase;
+
+ /* Aligned queue start address (per requirement) */
+ struct mutex q_mutex ____cacheline_aligned;
+ unsigned int qidx;
+
+ unsigned int qsize;
+ dma_addr_t qbase_dma;
+ dma_addr_t qdma_tail;
+
+ unsigned int active;
+ unsigned int suspended;
+
+ /* Register addresses for queue */
+ void __iomem *reg_control;
+ void __iomem *reg_tail_lo;
+ void __iomem *reg_head_lo;
+ void __iomem *reg_int_enable;
+ void __iomem *reg_interrupt_status;
+ void __iomem *reg_status;
+ void __iomem *reg_int_status;
+ void __iomem *reg_dma_status;
+ void __iomem *reg_dma_read_status;
+ void __iomem *reg_dma_write_status;
+ u32 qcontrol; /* Cached control register */
+
+ /* Status values from job */
+ u32 int_status;
+ u32 q_status;
+ u32 q_int_status;
+ u32 cmd_error;
+} ____cacheline_aligned;
+
+struct pt_device {
+ struct list_head entry;
+
+ unsigned int ord;
+ char name[MAX_PT_NAME_LEN];
+
+ struct device *dev;
+
+ /* Bus specific device information */
+ struct pt_msix *pt_msix;
+
+ struct pt_dev_vdata *dev_vdata;
+
+ unsigned int pt_irq;
+
+ /* I/O area used for device communication */
+ void __iomem *io_regs;
+
+ spinlock_t cmd_lock ____cacheline_aligned;
+ unsigned int cmd_count;
+ struct list_head cmd;
+
+ /*
+ * The command queue. This represent the queue available on the
+ * PTDMA that are available for processing cmds
+ */
+ struct pt_cmd_queue cmd_q;
+
+ wait_queue_head_t lsb_queue;
+
+ struct pt_tasklet_data tdata;
+};
+
+/*
+ * descriptor for PTDMA commands
+ * 8 32-bit words:
+ * word 0: function; engine; control bits
+ * word 1: length of source data
+ * word 2: low 32 bits of source pointer
+ * word 3: upper 16 bits of source pointer; source memory type
+ * word 4: low 32 bits of destination pointer
+ * word 5: upper 16 bits of destination pointer; destination memory type
+ * word 6: reserved 32 bits
+ * word 7: reserved 32 bits
+ */
+
+union dword0 {
+ struct {
+ unsigned int soc:1;
+ unsigned int ioc:1;
+ unsigned int rsvd1:1;
+ unsigned int init:1;
+ unsigned int eom:1;
+ unsigned int function:15;
+ unsigned int engine:4;
+ unsigned int prot:1;
+ unsigned int rsvd2:7;
+ };
+ u32 val;
+};
+
+struct dword3 {
+ unsigned int src_hi:16;
+ unsigned int src_mem:2;
+ unsigned int lsb_cxt_id:8;
+ unsigned int rsvd1:5;
+ unsigned int fixed:1;
+};
+
+struct dword5 {
+ unsigned int dst_hi:16;
+ unsigned int dst_mem:2;
+ unsigned int rsvd1:13;
+ unsigned int fixed:1;
+};
+
+struct ptdma_desc {
+ union dword0 dw0;
+ u32 length;
+ u32 src_lo;
+ struct dword3 dw3;
+ u32 dst_lo;
+ struct dword5 dw5;
+ __le32 rsvd1;
+ __le32 rsvd2;
+};
+
+/* Structure to hold PT device data */
+struct pt_dev_vdata {
+ const unsigned int bar;
+};
+
+int pt_core_init(struct pt_device *pt);
+void pt_core_destroy(struct pt_device *pt);
+
+int pt_core_perform_passthru(struct pt_cmd_queue *cmd_q,
+ struct pt_passthru_engine *pt_engine);
+
+void pt_start_queue(struct pt_cmd_queue *cmd_q);
+void pt_stop_queue(struct pt_cmd_queue *cmd_q);
+
+#endif
--
2.7.4
From: Sanjay R Mehta <[email protected]>
Register ptdma queue to Linux dmaengine framework as general-purpose
DMA channels.
Signed-off-by: Sanjay R Mehta <[email protected]>
---
drivers/dma/ptdma/Kconfig | 2 +
drivers/dma/ptdma/Makefile | 2 +-
drivers/dma/ptdma/ptdma-dev.c | 32 +++
drivers/dma/ptdma/ptdma-dmaengine.c | 554 ++++++++++++++++++++++++++++++++++++
drivers/dma/ptdma/ptdma.h | 31 ++
5 files changed, 620 insertions(+), 1 deletion(-)
create mode 100644 drivers/dma/ptdma/ptdma-dmaengine.c
diff --git a/drivers/dma/ptdma/Kconfig b/drivers/dma/ptdma/Kconfig
index f93f9c2..694ed27 100644
--- a/drivers/dma/ptdma/Kconfig
+++ b/drivers/dma/ptdma/Kconfig
@@ -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 & performs high bandwidth memory to
diff --git a/drivers/dma/ptdma/Makefile b/drivers/dma/ptdma/Makefile
index 320fa82..a528cb0 100644
--- a/drivers/dma/ptdma/Makefile
+++ b/drivers/dma/ptdma/Makefile
@@ -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
diff --git a/drivers/dma/ptdma/ptdma-dev.c b/drivers/dma/ptdma/ptdma-dev.c
index d1e9892..4bdad35 100644
--- a/drivers/dma/ptdma/ptdma-dev.c
+++ b/drivers/dma/ptdma/ptdma-dev.c
@@ -127,6 +127,26 @@ static inline void pt_core_enable_queue_interrupts(struct pt_device *pt)
iowrite32(SUPPORTED_INTERRUPTS, pt->cmd_q.reg_int_enable);
}
+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;
+
+ tail = lower_32_bits(cmd_q->qdma_tail + cmd_q->qidx * Q_DESC_SIZE);
+ if (cmd_q->cmd_error) {
+ /*
+ * Log the error and flush the queue by
+ * moving the head pointer
+ */
+ pt_log_error(cmd_q->pt, cmd_q->cmd_error);
+ iowrite32(tail, cmd_q->reg_head_lo);
+ }
+
+ cmd->pt_cmd_callback(cmd->data, cmd->ret);
+}
+
static irqreturn_t pt_core_irq_handler(int irq, void *data)
{
struct pt_device *pt = (struct pt_device *)data;
@@ -150,6 +170,7 @@ static irqreturn_t pt_core_irq_handler(int irq, void *data)
}
pt_core_enable_queue_interrupts(pt);
+ pt_do_cmd_complete((ulong)&pt->tdata);
return IRQ_HANDLED;
}
@@ -253,8 +274,16 @@ 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);
@@ -270,6 +299,9 @@ 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
index 0000000..9b9b77c
--- /dev/null
+++ b/drivers/dma/ptdma/ptdma-dmaengine.c
@@ -0,0 +1,554 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * AMD Passthrough DMA device driver
+ * -- Based on the CCP driver
+ *
+ * Copyright (C) 2016,2020 Advanced Micro Devices, Inc.
+ *
+ * Author: Sanjay R Mehta <[email protected]>
+ * Author: Gary R Hook <[email protected]>
+ */
+
+#include "ptdma.h"
+#include "../dmaengine.h"
+#include "../virt-dma.h"
+
+#define PT_DMA_WIDTH(_mask) \
+({ \
+ u64 mask = (_mask) + 1; \
+ (mask == 0) ? 64 : fls64(mask); \
+})
+
+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_cmd_resources(struct pt_device *pt,
+ struct list_head *list)
+{
+ struct pt_dma_cmd *cmd, *ctmp;
+
+ list_for_each_entry_safe(cmd, ctmp, list, entry) {
+ list_del(&cmd->entry);
+ kmem_cache_free(pt->dma_cmd_cache, cmd);
+ }
+}
+
+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;
+
+ pt_free_cmd_resources(pt, &desc->cmdlist);
+ kmem_cache_free(pt->dma_desc_cache, desc);
+}
+
+static int pt_issue_next_cmd(struct pt_dma_desc *desc)
+{
+ struct pt_passthru_engine *pt_engine;
+ struct pt_dma_cmd *cmd;
+ struct pt_device *pt;
+ struct pt_cmd *pt_cmd;
+ struct pt_cmd_queue *cmd_q;
+
+ cmd = list_first_entry(&desc->cmdlist, struct pt_dma_cmd, entry);
+ desc->issued_to_hw = 1;
+
+ pt_cmd = &cmd->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 void pt_free_active_cmd(struct pt_dma_desc *desc)
+{
+ struct pt_dma_cmd *cmd = NULL;
+
+ if (desc->issued_to_hw)
+ cmd = list_first_entry_or_null(&desc->cmdlist, struct pt_dma_cmd,
+ entry);
+ if (!cmd)
+ return;
+
+ list_del(&cmd->entry);
+ kmem_cache_free(desc->pt->dma_cmd_cache, cmd);
+}
+
+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_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);
+
+ if (list_empty(&chan->vc.desc_submitted))
+ return NULL;
+
+ vd = list_empty(&chan->vc.desc_issued) ?
+ list_first_entry(&chan->vc.desc_submitted,
+ struct virt_dma_desc, node) : NULL;
+
+ vchan_issue_pending(&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) {
+ /* Remove the DMA command from the list and free it */
+ pt_free_active_cmd(desc);
+ if (!desc->issued_to_hw) {
+ /* No errors, keep going */
+ if (desc->status != DMA_ERROR)
+ return desc;
+ /* Error, free remaining commands and move on */
+ pt_free_cmd_resources(desc->pt,
+ &desc->cmdlist);
+ }
+
+ 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_issue_next_cmd(desc);
+ if (!ret)
+ break;
+
+ desc->status = DMA_ERROR;
+ }
+}
+
+static struct pt_dma_cmd *pt_alloc_dma_cmd(struct pt_dma_chan *chan)
+{
+ struct pt_dma_cmd *cmd;
+
+ cmd = kmem_cache_zalloc(chan->pt->dma_cmd_cache, GFP_NOWAIT);
+
+ return cmd;
+}
+
+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;
+ INIT_LIST_HEAD(&desc->cmdlist);
+ desc->status = DMA_IN_PROGRESS;
+
+ return desc;
+}
+
+static struct pt_dma_desc *pt_create_desc(struct dma_chan *dma_chan,
+ struct scatterlist *dst_sg,
+ unsigned int dst_nents,
+ struct scatterlist *src_sg,
+ unsigned int src_nents,
+ unsigned long flags)
+{
+ struct pt_dma_chan *chan = to_pt_chan(dma_chan);
+ struct pt_device *pt = chan->pt;
+ struct pt_dma_desc *desc;
+ struct pt_dma_cmd *cmd;
+ struct pt_cmd *pt_cmd;
+ struct pt_passthru_engine *pt_engine;
+ unsigned int src_offset, src_len;
+ unsigned int dst_offset, dst_len;
+ unsigned int len;
+ size_t total_len;
+
+ if (!dst_sg || !src_sg)
+ return NULL;
+
+ if (!dst_nents || !src_nents)
+ return NULL;
+
+ desc = pt_alloc_dma_desc(chan, flags);
+ if (!desc)
+ return NULL;
+
+ total_len = 0;
+
+ src_len = sg_dma_len(src_sg);
+ src_offset = 0;
+
+ dst_len = sg_dma_len(dst_sg);
+ dst_offset = 0;
+
+ while (true) {
+ if (!src_len) {
+ src_nents--;
+ if (!src_nents)
+ break;
+
+ src_sg = sg_next(src_sg);
+ if (!src_sg)
+ break;
+
+ src_len = sg_dma_len(src_sg);
+ src_offset = 0;
+ continue;
+ }
+
+ if (!dst_len) {
+ dst_nents--;
+ if (!dst_nents)
+ break;
+
+ dst_sg = sg_next(dst_sg);
+ if (!dst_sg)
+ break;
+
+ dst_len = sg_dma_len(dst_sg);
+ dst_offset = 0;
+ continue;
+ }
+
+ len = min(dst_len, src_len);
+
+ cmd = pt_alloc_dma_cmd(chan);
+ if (!cmd)
+ goto err;
+
+ pt_cmd = &cmd->pt_cmd;
+ pt_cmd->pt = chan->pt;
+ pt_engine = &pt_cmd->passthru;
+ pt_cmd->engine = PT_ENGINE_PASSTHRU;
+ pt_engine->src_dma = sg_dma_address(src_sg) + src_offset;
+ pt_engine->dst_dma = sg_dma_address(dst_sg) + dst_offset;
+ pt_engine->src_len = len;
+ pt_cmd->pt_cmd_callback = pt_cmd_callback;
+ pt_cmd->data = desc;
+
+ list_add_tail(&cmd->entry, &desc->cmdlist);
+
+ total_len += len;
+
+ src_len -= len;
+ src_offset += len;
+
+ dst_len -= len;
+ dst_offset += len;
+ }
+
+ desc->len = total_len;
+
+ if (list_empty(&desc->cmdlist))
+ goto err;
+
+ return desc;
+
+err:
+ pt_free_cmd_resources(pt, &desc->cmdlist);
+ kmem_cache_free(pt->dma_desc_cache, desc);
+
+ return NULL;
+}
+
+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 scatterlist dst_sg, src_sg;
+ struct pt_dma_desc *desc;
+
+ sg_init_table(&dst_sg, 1);
+ sg_dma_address(&dst_sg) = dst;
+ sg_dma_len(&dst_sg) = len;
+
+ sg_init_table(&src_sg, 1);
+ sg_dma_address(&src_sg) = src;
+ sg_dma_len(&src_sg) = len;
+
+ desc = pt_create_desc(dma_chan, &dst_sg, 1, &src_sg, 1, 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);
+
+ 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 enum dma_status pt_tx_status(struct dma_chan *dma_chan,
+ dma_cookie_t cookie,
+ struct dma_tx_state *state)
+{
+ struct pt_dma_chan *chan = to_pt_chan(dma_chan);
+ struct pt_dma_desc *desc;
+ enum dma_status ret;
+ unsigned long flags;
+ struct virt_dma_desc *vd;
+
+ ret = dma_cookie_status(dma_chan, cookie, state);
+ if (ret == DMA_COMPLETE)
+ return ret;
+
+ spin_lock_irqsave(&chan->vc.lock, flags);
+ vd = vchan_find_desc(&chan->vc, cookie);
+ desc = vd ? to_pt_desc(vd) : NULL;
+ ret = desc->status;
+ spin_unlock_irqrestore(&chan->vc.lock, flags);
+
+ return ret;
+}
+
+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);
+
+ 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 *dma_cmd_cache_name;
+ char *dma_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;
+
+ dma_cmd_cache_name = devm_kasprintf(pt->dev, GFP_KERNEL,
+ "%s-dmaengine-cmd-cache",
+ pt->name);
+ if (!dma_cmd_cache_name)
+ return -ENOMEM;
+
+ pt->dma_cmd_cache = kmem_cache_create(dma_cmd_cache_name,
+ sizeof(struct pt_dma_cmd),
+ sizeof(void *),
+ SLAB_HWCACHE_ALIGN, NULL);
+ if (!pt->dma_cmd_cache)
+ return -ENOMEM;
+
+ dma_desc_cache_name = devm_kasprintf(pt->dev, GFP_KERNEL,
+ "%s-dmaengine-desc-cache",
+ pt->name);
+ if (!dma_desc_cache_name) {
+ ret = -ENOMEM;
+ goto err_cache;
+ }
+
+ pt->dma_desc_cache = kmem_cache_create(dma_desc_cache_name,
+ sizeof(struct pt_dma_desc),
+ sizeof(void *),
+ SLAB_HWCACHE_ALIGN, NULL);
+ if (!pt->dma_desc_cache) {
+ ret = -ENOMEM;
+ goto err_cache;
+ }
+
+ dma_dev->dev = pt->dev;
+ dma_dev->src_addr_widths = PT_DMA_WIDTH(dma_get_mask(pt->dev));
+ dma_dev->dst_addr_widths = PT_DMA_WIDTH(dma_get_mask(pt->dev));
+ 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);
+ dma_cap_set(DMA_PRIVATE, dma_dev->cap_mask);
+
+ INIT_LIST_HEAD(&dma_dev->channels);
+
+ chan = pt->pt_dma_chan;
+ chan->pt = pt;
+
+ 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 = pt_tx_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);
+
+ 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);
+}
diff --git a/drivers/dma/ptdma/ptdma.h b/drivers/dma/ptdma/ptdma.h
index 9e5819f..67c24bd 100644
--- a/drivers/dma/ptdma/ptdma.h
+++ b/drivers/dma/ptdma/ptdma.h
@@ -20,6 +20,9 @@
#include <linux/list.h>
#include <linux/wait.h>
#include <linux/dmapool.h>
+#include <linux/dmaengine.h>
+
+#include "../virt-dma.h"
#define MAX_PT_NAME_LEN 16
#define MAX_DMAPOOL_NAME_LEN 32
@@ -172,6 +175,25 @@ struct pt_cmd {
void *data;
};
+struct pt_dma_cmd {
+ struct list_head entry;
+ struct pt_cmd pt_cmd;
+};
+
+struct pt_dma_desc {
+ struct virt_dma_desc vd;
+ struct pt_device *pt;
+ struct list_head cmdlist;
+ enum dma_status status;
+ size_t len;
+ bool issued_to_hw;
+};
+
+struct pt_dma_chan {
+ struct virt_dma_chan vc;
+ struct pt_device *pt;
+};
+
struct pt_cmd_queue {
struct pt_device *pt;
@@ -240,6 +262,12 @@ 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;
@@ -304,6 +332,9 @@ 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);
--
2.7.4
From: Sanjay R Mehta <[email protected]>
Expose data about the configuration and operation of the
PTDMA through debugfs entries: device name, capabilities,
configuration, statistics.
Signed-off-by: Sanjay R Mehta <[email protected]>
---
drivers/dma/ptdma/Makefile | 2 +-
drivers/dma/ptdma/ptdma-debugfs.c | 115 ++++++++++++++++++++++++++++++++++++++
drivers/dma/ptdma/ptdma-dev.c | 5 ++
drivers/dma/ptdma/ptdma.h | 5 ++
4 files changed, 126 insertions(+), 1 deletion(-)
create mode 100644 drivers/dma/ptdma/ptdma-debugfs.c
diff --git a/drivers/dma/ptdma/Makefile b/drivers/dma/ptdma/Makefile
index a528cb0..ce54102 100644
--- a/drivers/dma/ptdma/Makefile
+++ b/drivers/dma/ptdma/Makefile
@@ -5,6 +5,6 @@
obj-$(CONFIG_AMD_PTDMA) += ptdma.o
-ptdma-objs := ptdma-dev.o ptdma-dmaengine.o
+ptdma-objs := ptdma-dev.o ptdma-dmaengine.o ptdma-debugfs.o
ptdma-$(CONFIG_PCI) += ptdma-pci.o
diff --git a/drivers/dma/ptdma/ptdma-debugfs.c b/drivers/dma/ptdma/ptdma-debugfs.c
new file mode 100644
index 0000000..03d47cd
--- /dev/null
+++ b/drivers/dma/ptdma/ptdma-debugfs.c
@@ -0,0 +1,115 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * AMD Passthrough DMA device driver
+ * -- Based on the CCP driver
+ *
+ * Copyright (C) 2016,2020 Advanced Micro Devices, Inc.
+ *
+ * Author: Sanjay R Mehta <[email protected]>
+ * Author: Gary R Hook <[email protected]>
+ */
+
+#include <linux/debugfs.h>
+#include <linux/seq_file.h>
+
+#include "ptdma.h"
+
+/* DebugFS helpers */
+#define MAX_NAME_LEN 20
+#define RI_VERSION_NUM 0x0000003F
+
+#define RI_NUM_VQM 0x00078000
+#define RI_NVQM_SHIFT 15
+
+static DEFINE_MUTEX(pt_debugfs_lock);
+
+static int pt_debugfs_info_show(struct seq_file *s, void *p)
+{
+ struct pt_device *pt = s->private;
+ unsigned int regval;
+
+ if (!pt)
+ return 0;
+
+ seq_printf(s, "Device name: %s\n", pt->name);
+ seq_printf(s, " # Queues: %d\n", 1);
+ seq_printf(s, " # Cmds: %d\n", pt->cmd_count);
+
+ regval = ioread32(pt->io_regs + CMD_PT_VERSION);
+
+ seq_printf(s, " Version: %d\n", regval & RI_VERSION_NUM);
+ seq_puts(s, " Engines:");
+ seq_puts(s, "\n");
+ seq_printf(s, " Queues: %d\n", (regval & RI_NUM_VQM) >> RI_NVQM_SHIFT);
+
+ return 0;
+}
+
+/*
+ * Return a formatted buffer containing the current
+ * statistics of queue for PTDMA
+ */
+static int pt_debugfs_stats_show(struct seq_file *s, void *p)
+{
+ struct pt_device *pt = s->private;
+
+ seq_printf(s, "Total Interrupts Handled: %ld\n", pt->total_interrupts);
+
+ return 0;
+}
+
+static int pt_debugfs_queue_show(struct seq_file *s, void *p)
+{
+ struct pt_cmd_queue *cmd_q = s->private;
+ unsigned int regval;
+
+ if (!cmd_q)
+ return 0;
+
+ seq_printf(s, " Pass-Thru: %ld\n", cmd_q->total_pt_ops);
+
+ regval = ioread32(cmd_q->reg_int_enable);
+
+ seq_puts(s, " Enabled Interrupts:");
+ if (regval & INT_EMPTY_QUEUE)
+ seq_puts(s, " EMPTY");
+ if (regval & INT_QUEUE_STOPPED)
+ seq_puts(s, " STOPPED");
+ if (regval & INT_ERROR)
+ seq_puts(s, " ERROR");
+ if (regval & INT_COMPLETION)
+ seq_puts(s, " COMPLETION");
+ seq_puts(s, "\n");
+
+ return 0;
+}
+
+DEFINE_SHOW_ATTRIBUTE(pt_debugfs_info);
+DEFINE_SHOW_ATTRIBUTE(pt_debugfs_queue);
+DEFINE_SHOW_ATTRIBUTE(pt_debugfs_stats);
+
+void ptdma_debugfs_setup(struct pt_device *pt)
+{
+ struct pt_cmd_queue *cmd_q;
+ char name[MAX_NAME_LEN + 1];
+ struct dentry *debugfs_q_instance;
+
+ if (!debugfs_initialized())
+ return;
+
+ debugfs_create_file("info", 0400, pt->dma_dev.dbg_dev_root, pt,
+ &pt_debugfs_info_fops);
+
+ debugfs_create_file("stats", 0600, pt->dma_dev.dbg_dev_root, pt,
+ &pt_debugfs_stats_fops);
+
+ cmd_q = &pt->cmd_q;
+
+ snprintf(name, MAX_NAME_LEN - 1, "q");
+
+ debugfs_q_instance =
+ debugfs_create_dir(name, pt->dma_dev.dbg_dev_root);
+
+ debugfs_create_file("stats", 0600, debugfs_q_instance, cmd_q,
+ &pt_debugfs_queue_fops);
+}
diff --git a/drivers/dma/ptdma/ptdma-dev.c b/drivers/dma/ptdma/ptdma-dev.c
index 4bdad35..0f44944 100644
--- a/drivers/dma/ptdma/ptdma-dev.c
+++ b/drivers/dma/ptdma/ptdma-dev.c
@@ -106,6 +106,7 @@ int pt_core_perform_passthru(struct pt_cmd_queue *cmd_q,
struct ptdma_desc desc;
cmd_q->cmd_error = 0;
+ cmd_q->total_pt_ops++;
memset(&desc, 0, sizeof(desc));
desc.dw0.val = CMD_DESC_DW0_VAL;
desc.length = pt_engine->src_len;
@@ -154,6 +155,7 @@ static irqreturn_t pt_core_irq_handler(int irq, void *data)
u32 status;
pt_core_disable_queue_interrupts(pt);
+ pt->total_interrupts++;
status = ioread32(cmd_q->reg_interrupt_status);
if (status) {
@@ -279,6 +281,9 @@ int pt_core_init(struct pt_device *pt)
if (ret)
goto e_dmaengine;
+ /* Set up debugfs entries */
+ ptdma_debugfs_setup(pt);
+
return 0;
e_dmaengine:
diff --git a/drivers/dma/ptdma/ptdma.h b/drivers/dma/ptdma/ptdma.h
index 67c24bd..67027d8 100644
--- a/drivers/dma/ptdma/ptdma.h
+++ b/drivers/dma/ptdma/ptdma.h
@@ -232,6 +232,8 @@ struct pt_cmd_queue {
u32 q_status;
u32 q_int_status;
u32 cmd_error;
+ /* queue Statistics */
+ unsigned long total_pt_ops;
} ____cacheline_aligned;
struct pt_device {
@@ -270,6 +272,8 @@ struct pt_device {
wait_queue_head_t lsb_queue;
+ /* Device Statistics */
+ unsigned long total_interrupts;
struct pt_tasklet_data tdata;
};
@@ -335,6 +339,7 @@ struct pt_dev_vdata {
int pt_dmaengine_register(struct pt_device *pt);
void pt_dmaengine_unregister(struct pt_device *pt);
+void ptdma_debugfs_setup(struct pt_device *pt);
int pt_core_init(struct pt_device *pt);
void pt_core_destroy(struct pt_device *pt);
--
2.7.4
> Sanjay R Mehta (3):
> dmaengine: ptdma: Initial driver for the AMD PTDMA
> dmaengine: ptdma: register PTDMA controller as a DMA resource
> dmaengine: ptdma: Add debugfs entries for PTDMA information
>
> MAINTAINERS | 6 +
> drivers/dma/Kconfig | 2 +
> drivers/dma/Makefile | 1 +
> drivers/dma/ptdma/Kconfig | 13 +
> drivers/dma/ptdma/Makefile | 10 +
> drivers/dma/ptdma/ptdma-debugfs.c | 115 ++++++++
> drivers/dma/ptdma/ptdma-dev.c | 333 ++++++++++++++++++++++
> drivers/dma/ptdma/ptdma-dmaengine.c | 554 ++++++++++++++++++++++++++++++++++++
> drivers/dma/ptdma/ptdma-pci.c | 252 ++++++++++++++++
> drivers/dma/ptdma/ptdma.h | 352 +++++++++++++++++++++++
> 10 files changed, 1638 insertions(+)
> create mode 100644 drivers/dma/ptdma/Kconfig
> create mode 100644 drivers/dma/ptdma/Makefile
> create mode 100644 drivers/dma/ptdma/ptdma-debugfs.c
> create mode 100644 drivers/dma/ptdma/ptdma-dev.c
> create mode 100644 drivers/dma/ptdma/ptdma-dmaengine.c
> create mode 100644 drivers/dma/ptdma/ptdma-pci.c
> create mode 100644 drivers/dma/ptdma/ptdma.h
Hi Vinod,
Have incorporated all the changes suggested by you in this patch series.
Any further comments?
Thanks,
Sanjay
>
On 10/16/2020 1:09 PM, Sanjay R Mehta wrote:
> From: Sanjay R Mehta <[email protected]>
>
> This patch series adds support for AMD PTDMA controller which
> performs high bandwidth memory-to-memory and IO copy operation,
> performs DMA transfer through queue based descriptor management.
>
> AMD Processor has multiple ptdma device instances with each controller
> having single queue. The driver also adds support for for multiple PTDMA
> instances, each device will get an unique identifier and uniquely
> named resources.
>
> v7:
> - Fixed module warnings reported ( by kernel test robot <[email protected]> ).
>
> v6:
> - Removed debug artifacts and made the suggested cosmetic changes.
> - implemented and used to_pt_chan and to_pt_desc inline functions.
> - Removed src and dst address check as framework does this.
> - Removed devm_kcalloc() usage and used devm_kzalloc() api.
> - Using framework debugfs directory to store dma info.
>
> v5:
> - modified code to submit next tranction in ISR itself and removed the tasklet.
> - implemented .device_synchronize API.
> - converted debugfs code by using DEFINE_SHOW_ATTRIBUTE()
> - using dbg_dev_root for debugfs root directory.
> - removed dma_status from pt_dma_chan
> - removed module parameter cmd_queue_lenght.
> - removed global device list for multiple devics.
> - removed code related to dynamic adding/deleting to device list
> - removed pt_add_device and pt_del_device functions
>
> v4:
> - modified DMA channel and descriptor management using virt-dma layer
> instead of list based management.
> - return only status of the cookie from pt_tx_status
> - copyright year changed from 2019 to 2020
> - removed dummy code for suspend & resume
> - used bitmask and genmask
>
> v3:
> - Fixed the sparse warnings.
>
> v2:
> - Added controller description in cover letter
> - Removed "default m" from Kconfig
> - Replaced low_address() and high_address() functions with kernel
> API's lower_32_bits & upper_32_bits().
> - Removed the BH handler function pt_core_irq_bh() and instead
> handling transaction in irq handler itself.
> - Moved presetting of command queue registers into new function
> "init_cmdq_regs()"
> - Removed the kernel thread dependency to submit transaction.
> - Increased the hardware command queue size to 32 and adding it
> as a module parameter.
> - Removed backlog command queue handling mechanism.
> - Removed software command queue handling and instead submitting
> transaction command directly to
> hardware command queue.
> - Added tasklet structure variable in "struct pt_device".
> This is used to invoke pt_do_cmd_complete() upon receiving interrupt
> for command completion.
> - pt_core_perform_passthru() function parameters are modified and it is
> now used to submit command directly to hardware from dmaengine framew
> - Removed below structures, enums, macros and functions, as these value
> constants. Making command submission simple,
> - Removed "union pt_function" and several macros like PT_VERSION,
> PT_BYTESWAP, PT_CMD_* etc..
> - enum pt_passthru_bitwise, enum pt_passthru_byteswap, enum pt_memty
> struct pt_dma_info, struct pt_data, struct pt_mem, struct pt_passt
> struct pt_op,
>
> Links of the review comments for v5:
> 1. https://lkml.org/lkml/2020/7/3/154
> 2. https://lkml.org/lkml/2020/8/25/431
> 3. https://lkml.org/lkml/2020/7/3/177
> 4. https://lkml.org/lkml/2020/7/3/186
>
> Links of the review comments for v5:
> 1. https://lkml.org/lkml/2020/5/4/42
> 2. https://lkml.org/lkml/2020/5/4/45
> 3. https://lkml.org/lkml/2020/5/4/38
> 4. https://lkml.org/lkml/2020/5/26/70
>
> Links of the review comments for v4:
> 1. https://lkml.org/lkml/2020/1/24/12
> 2. https://lkml.org/lkml/2020/1/24/17
>
> Links of the review comments for v2:
> 1https://lkml.org/lkml/2019/12/27/630
> 2. https://lkml.org/lkml/2020/1/3/23
> 3. https://lkml.org/lkml/2020/1/3/314
> 4. https://lkml.org/lkml/2020/1/10/100
>
> Links of the review comments for v1:
> 1. https://lkml.org/lkml/2019/9/24/490
> 2. https://lkml.org/lkml/2019/9/24/399
> 3. https://lkml.org/lkml/2019/9/24/862
> 4. https://lkml.org/lkml/2019/9/24/122
>
> Sanjay R Mehta (3):
> dmaengine: ptdma: Initial driver for the AMD PTDMA
> dmaengine: ptdma: register PTDMA controller as a DMA resource
> dmaengine: ptdma: Add debugfs entries for PTDMA information
>
> MAINTAINERS | 6 +
> drivers/dma/Kconfig | 2 +
> drivers/dma/Makefile | 1 +
> drivers/dma/ptdma/Kconfig | 13 +
> drivers/dma/ptdma/Makefile | 10 +
> drivers/dma/ptdma/ptdma-debugfs.c | 115 ++++++++
> drivers/dma/ptdma/ptdma-dev.c | 333 ++++++++++++++++++++++
> drivers/dma/ptdma/ptdma-dmaengine.c | 554 ++++++++++++++++++++++++++++++++++++
> drivers/dma/ptdma/ptdma-pci.c | 252 ++++++++++++++++
> drivers/dma/ptdma/ptdma.h | 352 +++++++++++++++++++++++
> 10 files changed, 1638 insertions(+)
> create mode 100644 drivers/dma/ptdma/Kconfig
> create mode 100644 drivers/dma/ptdma/Makefile
> create mode 100644 drivers/dma/ptdma/ptdma-debugfs.c
> create mode 100644 drivers/dma/ptdma/ptdma-dev.c
> create mode 100644 drivers/dma/ptdma/ptdma-dmaengine.c
> create mode 100644 drivers/dma/ptdma/ptdma-pci.c
> create mode 100644 drivers/dma/ptdma/ptdma.h
Hey Vinod,
Hope you are doing good.
I had submitted this patch series 3 weeks ago with the suggested changes.
Did you got chance to review?
Thanks,
Sanjay
>
On 16-10-20, 02:39, Sanjay R Mehta wrote:
> From: Sanjay R Mehta <[email protected]>
>
> Adding support for AMD PTDMA controller. It performs high-bandwidth
Add support or Adds support.. would be apt!
> memory to memory and IO copy operation. Device commands are managed
> via a circular queue of 'descriptors', each of which specifies source
> and destination addresses for copying a single buffer of data.
>
> Signed-off-by: Sanjay R Mehta <[email protected]>
> ---
> MAINTAINERS | 6 +
> drivers/dma/Kconfig | 2 +
> drivers/dma/Makefile | 1 +
> drivers/dma/ptdma/Kconfig | 11 ++
> drivers/dma/ptdma/Makefile | 10 ++
> drivers/dma/ptdma/ptdma-dev.c | 296 +++++++++++++++++++++++++++++++++++++++
> drivers/dma/ptdma/ptdma-pci.c | 252 +++++++++++++++++++++++++++++++++
> drivers/dma/ptdma/ptdma.h | 316 ++++++++++++++++++++++++++++++++++++++++++
> 8 files changed, 894 insertions(+)
> create mode 100644 drivers/dma/ptdma/Kconfig
> create mode 100644 drivers/dma/ptdma/Makefile
> create mode 100644 drivers/dma/ptdma/ptdma-dev.c
> create mode 100644 drivers/dma/ptdma/ptdma-pci.c
> create mode 100644 drivers/dma/ptdma/ptdma.h
>
> diff --git a/MAINTAINERS b/MAINTAINERS
> index 33b27e6..f6ae0d1 100644
> --- a/MAINTAINERS
> +++ b/MAINTAINERS
> @@ -943,6 +943,12 @@ S: Supported
> F: arch/arm64/boot/dts/amd/amd-seattle-xgbe*.dtsi
> F: drivers/net/ethernet/amd/xgbe/
>
> +AMD PTDMA DRIVER
> +M: Sanjay R Mehta <[email protected]>
> +L: [email protected]
> +S: Maintained
> +F: drivers/dma/ptdma/
> +
> ANALOG DEVICES INC AD5686 DRIVER
> M: Michael Hennerich <[email protected]>
> L: [email protected]
> diff --git a/drivers/dma/Kconfig b/drivers/dma/Kconfig
> index 518a143..a21c983 100644
> --- a/drivers/dma/Kconfig
> +++ b/drivers/dma/Kconfig
> @@ -748,6 +748,8 @@ source "drivers/dma/ti/Kconfig"
>
> source "drivers/dma/fsl-dpaa2-qdma/Kconfig"
>
> +source "drivers/dma/ptdma/Kconfig"
> +
> # clients
> comment "DMA Clients"
> depends on DMA_ENGINE
> diff --git a/drivers/dma/Makefile b/drivers/dma/Makefile
> index e60f813..2785756 100644
> --- a/drivers/dma/Makefile
> +++ b/drivers/dma/Makefile
> @@ -83,6 +83,7 @@ obj-$(CONFIG_XGENE_DMA) += xgene-dma.o
> obj-$(CONFIG_ZX_DMA) += zx_dma.o
> obj-$(CONFIG_ST_FDMA) += st_fdma.o
> obj-$(CONFIG_FSL_DPAA2_QDMA) += fsl-dpaa2-qdma/
> +obj-$(CONFIG_AMD_PTDMA) += ptdma/
Please keep this sorted :(
> obj-y += mediatek/
> obj-y += qcom/
> diff --git a/drivers/dma/ptdma/Kconfig b/drivers/dma/ptdma/Kconfig
> new file mode 100644
> index 0000000..f93f9c2
> --- /dev/null
> +++ b/drivers/dma/ptdma/Kconfig
> @@ -0,0 +1,11 @@
> +# SPDX-License-Identifier: GPL-2.0-only
> +config AMD_PTDMA
> + tristate "AMD PassThru DMA Engine"
> + depends on X86_64 && PCI
> + help
> + Enable support for the AMD PTDMA controller. This controller
> + provides DMA capabilities & performs high bandwidth memory to
> + memory and IO copy operation and performs DMA transfer through
> + queue based descriptor management. This DMA controller is intended
> + to use with AMD Non-Transparent Bridge devices and not for general
> + purpose slave DMA.
s/slave/peripheral
> +static int pt_core_execute_cmd(struct ptdma_desc *desc,
> + struct pt_cmd_queue *cmd_q)
> +{
> + __le32 *mp;
> + u32 *dp;
> + u32 tail;
> + int i;
> +
> + if (desc->dw0.soc) {
> + desc->dw0.ioc = 1;
> + desc->dw0.soc = 0;
> + }
> + mutex_lock(&cmd_q->q_mutex);
> +
> + mp = (__le32 *)&cmd_q->qbase[cmd_q->qidx];
why not make the qidx __le32 instead of int ?
> + dp = (u32 *)desc;
> + for (i = 0; i < 8; i++)
why 8..?
> + mp[i] = cpu_to_le32(dp[i]); /* handle endianness */
Where is this used..? Why not drop this..
> +static irqreturn_t pt_core_irq_handler(int irq, void *data)
> +{
> + struct pt_device *pt = (struct pt_device *)data;
why do you need cast away from void *
> +int pt_core_init(struct pt_device *pt)
> +{
> + char dma_pool_name[MAX_DMAPOOL_NAME_LEN];
> + struct pt_cmd_queue *cmd_q = &pt->cmd_q;
> + u32 dma_addr_lo, dma_addr_hi;
> + struct device *dev = pt->dev;
> + struct dma_pool *dma_pool;
> + int ret;
> +
> + /* Allocate a dma pool for the queue */
> + snprintf(dma_pool_name, sizeof(dma_pool_name), "%s_q", pt->name);
> +
> + dma_pool = dma_pool_create(dma_pool_name, dev,
> + PT_DMAPOOL_MAX_SIZE,
> + PT_DMAPOOL_ALIGN, 0);
> + if (!dma_pool) {
> + dev_err(dev, "unable to allocate dma pool\n");
> + ret = -ENOMEM;
> + return ret;
return -ENOMEM?
> + }
> +
> + /* ptdma core initialisation */
> + iowrite32(CMD_CONFIG_VHB_EN, pt->io_regs + CMD_CONFIG_OFFSET);
> + iowrite32(CMD_QUEUE_PRIO, pt->io_regs + CMD_QUEUE_PRIO_OFFSET);
> + iowrite32(CMD_TIMEOUT_DISABLE, pt->io_regs + CMD_TIMEOUT_OFFSET);
> + iowrite32(CMD_CLK_GATE_CONFIG, pt->io_regs + CMD_CLK_GATE_CTL_OFFSET);
> + iowrite32(CMD_CONFIG_REQID, pt->io_regs + CMD_REQID_CONFIG_OFFSET);
> +
> + cmd_q->pt = pt;
> + cmd_q->dma_pool = dma_pool;
> + mutex_init(&cmd_q->q_mutex);
> +
> + /* Page alignment satisfies our needs for N <= 128 */
> + cmd_q->qsize = Q_SIZE(Q_DESC_SIZE);
> + cmd_q->qbase = dma_alloc_coherent(dev, cmd_q->qsize,
> + &cmd_q->qbase_dma,
> + GFP_KERNEL);
> + if (!cmd_q->qbase) {
> + dev_err(dev, "unable to allocate command queue\n");
> + ret = -ENOMEM;
> + goto e_dma_alloc;
> + }
> +
> + cmd_q->qidx = 0;
> +
> + /* Preset some register values */
> + cmd_q->reg_control = pt->io_regs + CMD_Q_STATUS_INCR;
> + pt_init_cmdq_regs(cmd_q);
> +
> + /* Turn off the queues and disable interrupts until ready */
> + pt_core_disable_queue_interrupts(pt);
> +
> + cmd_q->qcontrol = 0; /* Start with nothing */
> + iowrite32(cmd_q->qcontrol, cmd_q->reg_control);
> +
> + ioread32(cmd_q->reg_int_status);
> + ioread32(cmd_q->reg_status);
> +
> + /* Clear the interrupt status */
> + iowrite32(SUPPORTED_INTERRUPTS, cmd_q->reg_interrupt_status);
> +
> + /* Request an irq */
> + ret = request_irq(pt->pt_irq, pt_core_irq_handler, 0, pt->name, pt);
> + if (ret) {
> + dev_err(dev, "unable to allocate an IRQ\n");
Do log the error while you are it!
> + goto e_pool;
> + }
> +
> + /* Update the device registers with queue information. */
> +
> + cmd_q->qcontrol &= ~(CMD_Q_SIZE << CMD_Q_SHIFT);
> + cmd_q->qcontrol |= QUEUE_SIZE_VAL << CMD_Q_SHIFT;
Please use helpers in bitfield.h to do the shifts for you, no need to
define the shift values
> +#include <linux/device.h>
> +#include <linux/dma-mapping.h>
> +#include <linux/delay.h>
> +#include <linux/interrupt.h>
> +#include <linux/kernel.h>
> +#include <linux/kthread.h>
> +#include <linux/module.h>
> +#include <linux/pci_ids.h>
> +#include <linux/pci.h>
> +#include <linux/spinlock.h>
> +#include <linux/sched.h>
why do you need sched.h here?
> +
> +#include "ptdma.h"
> +
> +/* Ever-increasing value to produce unique unit numbers */
> +static atomic_t pt_ordinal;
What is the need of that?
> +static int pt_get_irqs(struct pt_device *pt)
> +{
> + struct device *dev = pt->dev;
> + int ret;
> +
> + ret = pt_get_msix_irqs(pt);
> + if (!ret)
> + return 0;
> +
> + /* Couldn't get MSI-X vectors, try MSI */
> + dev_dbg(dev, "could not enable MSI-X (%d), trying MSI\n", ret);
> + ret = pt_get_msi_irq(pt);
> + if (!ret)
> + return 0;
> +
> + /* Couldn't get MSI interrupt */
> + dev_dbg(dev, "could not enable MSI (%d)\n", ret);
Should this not be an error?
> +/*
> + * descriptor for PTDMA commands
> + * 8 32-bit words:
> + * word 0: function; engine; control bits
> + * word 1: length of source data
> + * word 2: low 32 bits of source pointer
> + * word 3: upper 16 bits of source pointer; source memory type
> + * word 4: low 32 bits of destination pointer
> + * word 5: upper 16 bits of destination pointer; destination memory type
> + * word 6: reserved 32 bits
> + * word 7: reserved 32 bits
> + */
> +
> +union dword0 {
> + struct {
> + unsigned int soc:1;
> + unsigned int ioc:1;
> + unsigned int rsvd1:1;
> + unsigned int init:1;
> + unsigned int eom:1;
> + unsigned int function:15;
> + unsigned int engine:4;
> + unsigned int prot:1;
> + unsigned int rsvd2:7;
> + };
Do you really want to use union and bitfields here, this seem like HW
description, best to use FIELD_PREP or FIELD_GET for these
Been there, done that would advise to avoid this approach.
--
~Vinod
On 16-10-20, 02:39, Sanjay R Mehta wrote:
> diff --git a/drivers/dma/ptdma/ptdma-dmaengine.c b/drivers/dma/ptdma/ptdma-dmaengine.c
> new file mode 100644
> index 0000000..9b9b77c
> --- /dev/null
> +++ b/drivers/dma/ptdma/ptdma-dmaengine.c
> @@ -0,0 +1,554 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/*
> + * AMD Passthrough DMA device driver
> + * -- Based on the CCP driver
> + *
> + * Copyright (C) 2016,2020 Advanced Micro Devices, Inc.
> + *
> + * Author: Sanjay R Mehta <[email protected]>
> + * Author: Gary R Hook <[email protected]>
> + */
> +
> +#include "ptdma.h"
> +#include "../dmaengine.h"
> +#include "../virt-dma.h"
> +
> +#define PT_DMA_WIDTH(_mask) \
> +({ \
> + u64 mask = (_mask) + 1; \
> + (mask == 0) ? 64 : fls64(mask); \
how would mask be 0 here..?
> +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) {
> + /* Remove the DMA command from the list and free it */
> + pt_free_active_cmd(desc);
> + if (!desc->issued_to_hw) {
> + /* No errors, keep going */
> + if (desc->status != DMA_ERROR)
> + return desc;
> + /* Error, free remaining commands and move on */
> + pt_free_cmd_resources(desc->pt,
> + &desc->cmdlist);
this should be single line
> + }
> +
> + 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);
So IIUC, this has two functions, one to find active desc and second to
complete and invoke callback. These two are different tasks and I am not
sure why these are handled here. Can you split these up into different
routines, that may help making this read better
> +static struct pt_dma_desc *pt_create_desc(struct dma_chan *dma_chan,
> + struct scatterlist *dst_sg,
> + unsigned int dst_nents,
> + struct scatterlist *src_sg,
> + unsigned int src_nents,
> + unsigned long flags)
> +{
> + struct pt_dma_chan *chan = to_pt_chan(dma_chan);
> + struct pt_device *pt = chan->pt;
> + struct pt_dma_desc *desc;
> + struct pt_dma_cmd *cmd;
> + struct pt_cmd *pt_cmd;
> + struct pt_passthru_engine *pt_engine;
> + unsigned int src_offset, src_len;
> + unsigned int dst_offset, dst_len;
> + unsigned int len;
> + size_t total_len;
> +
> + if (!dst_sg || !src_sg)
> + return NULL;
> +
> + if (!dst_nents || !src_nents)
> + return NULL;
> +
> + desc = pt_alloc_dma_desc(chan, flags);
> + if (!desc)
> + return NULL;
> +
> + total_len = 0;
> +
> + src_len = sg_dma_len(src_sg);
> + src_offset = 0;
> +
> + dst_len = sg_dma_len(dst_sg);
> + dst_offset = 0;
> +
> + while (true) {
> + if (!src_len) {
> + src_nents--;
> + if (!src_nents)
> + break;
> +
> + src_sg = sg_next(src_sg);
> + if (!src_sg)
> + break;
> +
> + src_len = sg_dma_len(src_sg);
> + src_offset = 0;
> + continue;
> + }
> +
> + if (!dst_len) {
> + dst_nents--;
> + if (!dst_nents)
> + break;
> +
> + dst_sg = sg_next(dst_sg);
> + if (!dst_sg)
> + break;
> +
> + dst_len = sg_dma_len(dst_sg);
> + dst_offset = 0;
> + continue;
> + }
> +
> + len = min(dst_len, src_len);
Ah why not use for_each_sg()
> +
> + cmd = pt_alloc_dma_cmd(chan);
> + if (!cmd)
> + goto err;
> +
> + pt_cmd = &cmd->pt_cmd;
> + pt_cmd->pt = chan->pt;
> + pt_engine = &pt_cmd->passthru;
> + pt_cmd->engine = PT_ENGINE_PASSTHRU;
> + pt_engine->src_dma = sg_dma_address(src_sg) + src_offset;
> + pt_engine->dst_dma = sg_dma_address(dst_sg) + dst_offset;
> + pt_engine->src_len = len;
> + pt_cmd->pt_cmd_callback = pt_cmd_callback;
> + pt_cmd->data = desc;
> +
> + list_add_tail(&cmd->entry, &desc->cmdlist);
> +
> + total_len += len;
> +
> + src_len -= len;
> + src_offset += len;
> +
> + dst_len -= len;
> + dst_offset += len;
> + }
since you have both src and dst sgl that needs to be checked as well,
but where are the two lists coming from..? something does not sound
right here
> +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 scatterlist dst_sg, src_sg;
> + struct pt_dma_desc *desc;
> +
> + sg_init_table(&dst_sg, 1);
> + sg_dma_address(&dst_sg) = dst;
> + sg_dma_len(&dst_sg) = len;
> +
> + sg_init_table(&src_sg, 1);
> + sg_dma_address(&src_sg) = src;
> + sg_dma_len(&src_sg) = len;
Why do you need this overhead, why not pass on the addresses here?
So you create sg list here and then unravel that for description
creation. Typically we would have a lower level API which would handle
dma_addr_t and that can be called from here and sgl users!
> +static enum dma_status pt_tx_status(struct dma_chan *dma_chan,
> + dma_cookie_t cookie,
> + struct dma_tx_state *state)
> +{
> + struct pt_dma_chan *chan = to_pt_chan(dma_chan);
> + struct pt_dma_desc *desc;
> + enum dma_status ret;
> + unsigned long flags;
> + struct virt_dma_desc *vd;
> +
> + ret = dma_cookie_status(dma_chan, cookie, state);
> + if (ret == DMA_COMPLETE)
> + return ret;
> +
> + spin_lock_irqsave(&chan->vc.lock, flags);
> + vd = vchan_find_desc(&chan->vc, cookie);
> + desc = vd ? to_pt_desc(vd) : NULL;
> + ret = desc->status;
> + spin_unlock_irqrestore(&chan->vc.lock, flags);
how does this block of code help?
> +int pt_dmaengine_register(struct pt_device *pt)
> +{
> + struct pt_dma_chan *chan;
> + struct dma_device *dma_dev = &pt->dma_dev;
> + char *dma_cmd_cache_name;
> + char *dma_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;
> +
> + dma_cmd_cache_name = devm_kasprintf(pt->dev, GFP_KERNEL,
maybe user shorter variable names
> + "%s-dmaengine-cmd-cache",
> + pt->name);
> + if (!dma_cmd_cache_name)
> + return -ENOMEM;
> +
> + pt->dma_cmd_cache = kmem_cache_create(dma_cmd_cache_name,
> + sizeof(struct pt_dma_cmd),
> + sizeof(void *),
> + SLAB_HWCACHE_ALIGN, NULL);
> + if (!pt->dma_cmd_cache)
> + return -ENOMEM;
> +
> + dma_desc_cache_name = devm_kasprintf(pt->dev, GFP_KERNEL,
> + "%s-dmaengine-desc-cache",
> + pt->name);
> + if (!dma_desc_cache_name) {
> + ret = -ENOMEM;
> + goto err_cache;
> + }
> +
> + pt->dma_desc_cache = kmem_cache_create(dma_desc_cache_name,
> + sizeof(struct pt_dma_desc),
> + sizeof(void *),
> + SLAB_HWCACHE_ALIGN, NULL);
> + if (!pt->dma_desc_cache) {
> + ret = -ENOMEM;
> + goto err_cache;
> + }
> +
> + dma_dev->dev = pt->dev;
> + dma_dev->src_addr_widths = PT_DMA_WIDTH(dma_get_mask(pt->dev));
> + dma_dev->dst_addr_widths = PT_DMA_WIDTH(dma_get_mask(pt->dev));
> + 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);
> + dma_cap_set(DMA_PRIVATE, dma_dev->cap_mask);
Why DMA_PRIVATE for a memcpy function?
--
~Vinod
On Tue, Nov 24, 2020 at 12:18 PM Vinod Koul <[email protected]> wrote:
> IIRC there were few issues that I would like to get fixed before we can
> apply.. I hope authors can reply to the comments received and we can
> discuss on this.
Sounds like a plan, thanks!
--
wbr, Vitaly
Hello Vitaly,
On 24-11-20, 09:23, Vitaly Mayatskih wrote:
> On Wed, Nov 18, 2020 at 7:20 AM Vinod Koul <[email protected]> wrote:
>
> > this should be single line
>
> Vinod, do you see any obvious functional defects still present in the
> driver, or can it be finally merged for us to start testing, while
> Sanjay is working on improvements and style fixes? I'm sure the driver
> has hidden bugs, as any other piece of code on the planet, and Sanjay
> will appreciate bug reports from the actual PTDMA users.
IIRC there were few issues that I would like to get fixed before we can
apply.. I hope authors can reply to the comments received and we can
discuss on this.
--
~Vinod
On 11/24/2020 10:48 PM, Vinod Koul wrote:
> [CAUTION: External Email]
>
> Hello Vitaly,
>
> On 24-11-20, 09:23, Vitaly Mayatskih wrote:
>> On Wed, Nov 18, 2020 at 7:20 AM Vinod Koul <[email protected]> wrote:
>>
>>> this should be single line
>>
>> Vinod, do you see any obvious functional defects still present in the
>> driver, or can it be finally merged for us to start testing, while
>> Sanjay is working on improvements and style fixes? I'm sure the driver
>> has hidden bugs, as any other piece of code on the planet, and Sanjay
>> will appreciate bug reports from the actual PTDMA users.
>
> IIRC there were few issues that I would like to get fixed before we can
> apply.. I hope authors can reply to the comments received and we can
> discuss on this.
>
Thanks Vinod. Am working on your suggestions.
> --
> ~Vinod
>
On 11/24/2020 7:53 PM, Vitaly Mayatskih wrote:
> [CAUTION: External Email]
>
> On Wed, Nov 18, 2020 at 7:20 AM Vinod Koul <[email protected]> wrote:
>
>> this should be single line
>
> Vinod, do you see any obvious functional defects still present in the
> driver, or can it be finally merged for us to start testing, while
> Sanjay is working on improvements and style fixes? I'm sure the driver
> has hidden bugs, as any other piece of code on the planet, and Sanjay
> will appreciate bug reports from the actual PTDMA users.
>
Thanks Vitaly. Appreciate your suggestion.
Currently, am working on Vinod's feedback and will send next version of patch set.
please let me know if you have any feedback, I'll try to incorporate them as well.
> Thanks!
> --
> wbr, Vitaly
>
On Wed, Nov 18, 2020 at 7:20 AM Vinod Koul <[email protected]> wrote:
> this should be single line
Vinod, do you see any obvious functional defects still present in the
driver, or can it be finally merged for us to start testing, while
Sanjay is working on improvements and style fixes? I'm sure the driver
has hidden bugs, as any other piece of code on the planet, and Sanjay
will appreciate bug reports from the actual PTDMA users.
Thanks!
--
wbr, Vitaly
>> +#include <linux/delay.h>
>> +#include <linux/interrupt.h>
>> +#include <linux/kernel.h>
>> +#include <linux/kthread.h>
>> +#include <linux/module.h>
>> +#include <linux/pci_ids.h>
>> +#include <linux/pci.h>
>> +#include <linux/spinlock.h>
>> +#include <linux/sched.h>
>
> why do you need sched.h here?
>
>> +
>> +#include "ptdma.h"
>> +
>> +/* Ever-increasing value to produce unique unit numbers */
>> +static atomic_t pt_ordinal;
>
> What is the need of that?
>
The "pt_ordinal" is incremented for each DMA instances and its number is used only to assign device name for each instances.
This same device name is passed as a string parameter in many places in code like while using request_irq(), dma_pool_create() and in debugfs.
Also, I have implemented all of the comments for this patch except this. if this is fine, will send the next version for review.
Thanks,
Sanjay Mehta
>> +static int pt_get_irqs(struct pt_device *pt)
>> +{
>> + struct device *dev = pt->dev;
>> + int ret;
>> +
>> + ret = pt_get_msix_irqs(pt);
>> + if (!ret)
>> + return 0;
>> +
>> + /* Couldn't get MSI-X vectors, try MSI */
>> + dev_dbg(dev, "could not enable MSI-X (%d), trying MSI\n", ret);
>> + ret = pt_get_msi_irq(pt);
>> + dma_dev->dst_addr_widths = PT_DMA_WIDTH(dma_get_mask(pt->dev));
>> + 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);
>> + dma_cap_set(DMA_PRIVATE, dma_dev->cap_mask);
>
> Why DMA_PRIVATE for a memcpy function?
This DMA engine is intended to use with AMD NTB IP and not for general purpose DMA, hence set the cap as DMA_PRIVATE.
Otherwise this DMA engine will be used for system DMA. Please correct me if my understanding is not right.
Also, I have implemented all of the comments for this patch except this. if this is fine, will send the next version for review.
Thanks,
Sanjay
> --
> ~Vinod
>
On 18-03-21, 16:16, Sanjay R Mehta wrote:
> >> +#include <linux/delay.h>
> >> +#include <linux/interrupt.h>
> >> +#include <linux/kernel.h>
> >> +#include <linux/kthread.h>
> >> +#include <linux/module.h>
> >> +#include <linux/pci_ids.h>
> >> +#include <linux/pci.h>
> >> +#include <linux/spinlock.h>
> >> +#include <linux/sched.h>
> >
> > why do you need sched.h here?
> >
> >> +
> >> +#include "ptdma.h"
> >> +
> >> +/* Ever-increasing value to produce unique unit numbers */
> >> +static atomic_t pt_ordinal;
> >
> > What is the need of that?
> >
>
[please wrap your emails within 80 chars]
> The "pt_ordinal" is incremented for each DMA instances and its number
> is used only to assign device name for each instances. This same
> device name is passed as a string parameter in many places in code
> like while using request_irq(), dma_pool_create() and in debugfs.
Why do you need that, why not use device name which is unique..?
> Also, I have implemented all of the comments for this patch except
> this. if this is fine, will send the next version for review.
Am not sure I remember all the comments I gave, it has been _quite_ a
while since the feedback was provided. In order to have effective review
it would be great to revert back on a reasonable timeline and discuss...
Thanks
--
~Vinod
On 3/22/2021 11:34 AM, Vinod Koul wrote:
> [CAUTION: External Email]
>
> On 18-03-21, 16:16, Sanjay R Mehta wrote:
>>>> +#include <linux/delay.h>
>>>> +#include <linux/interrupt.h>
>>>> +#include <linux/kernel.h>
>>>> +#include <linux/kthread.h>
>>>> +#include <linux/module.h>
>>>> +#include <linux/pci_ids.h>
>>>> +#include <linux/pci.h>
>>>> +#include <linux/spinlock.h>
>>>> +#include <linux/sched.h>
>>>
>>> why do you need sched.h here?
>>>
>>>> +
>>>> +#include "ptdma.h"
>>>> +
>>>> +/* Ever-increasing value to produce unique unit numbers */
>>>> +static atomic_t pt_ordinal;
>>>
>>> What is the need of that?
>>>
>>
>
> [please wrap your emails within 80 chars]
>
Sure Vinod.
>> The "pt_ordinal" is incremented for each DMA instances and its number
>> is used only to assign device name for each instances. This same
>> device name is passed as a string parameter in many places in code
>> like while using request_irq(), dma_pool_create() and in debugfs.
>
> Why do you need that, why not use device name which is unique..?
>
Can we take this as part of bug fixes series in future?
>> Also, I have implemented all of the comments for this patch except
>> this. if this is fine, will send the next version for review.
>
> Am not sure I remember all the comments I gave, it has been _quite_ a
> while since the feedback was provided. In order to have effective review
> it would be great to revert back on a reasonable timeline and discuss...
>
Apologies from my side. I understand that I have taken more time. But I assure it doesn't happen again.
I have already sent out v8, can you please have a look at and provide your valuable feedback.
> Thanks
> --
> ~Vinod
>