[PATCH] dma: Add Keystone Packet DMA Engine driver

[Santosh Shilimkar]

On Tuesday 11 March 2014 06:23 PM, Vinod Koul wrote:
> On Fri, Feb 28, 2014 at 05:56:40PM -0500, Santosh Shilimkar wrote:
>> From: Sandeep Nair <sandeep_n at ti.com>
>>
>> The Packet DMA driver sets up the dma channels and flows for the
>> QMSS(Queue Manager SubSystem) who triggers the actual data movements
>> across clients using destination queues. Every client modules like
>> NETCP(Network Coprocessor), SRIO(Serial Rapid IO) and CRYPTO
>> Engines has its own instance of packet dma hardware. QMSS has also
>> an internal packet DMA module which is used as an infrastructure
>> DMA with zero copy.
>>
>> Patch adds DMAEngine driver for Keystone Packet DMA hardware.
>> The specifics on the device tree bindings for Packet DMA can be
>> found in:
>> 	Documentation/devicetree/bindings/dma/keystone-pktdma.txt
>>
>> The driver implements the configuration functions using standard DMAEngine
>> apis. The data movement is managed by QMSS device driver.
> Pls use subsystem appropriate name so here would have been dmaengine: ...
> 
> So i am still missing stuff like prepare calls, irq, descriptor management to
> call this a dmaengine driver.
>
> I guess you need to explain a bit more why the data movement is handled by some
> other driver and not by this one
>
To expand above statement, Packet DMA hardware blocks on Keystone SOCs
are DMAEngines. QMSS is centralised subsystem manages multiple functionalities
including triggering dma transfers, descriptor management and completion
irqs. There are separate instance of packet DMA hardware block per client
device. We program the DMA hardware to allocate channels and flows. So
the packet DMA resouces like dma channels and dma flows are configured
and managed through the DMAEngine driver. Thats why we implement only
device_alloc_chan_resources, device_free_chan_resources and device_control
DMAEngine APIs.
 
> few more comments inline
> 
>>
>> Cc: Vinod Koul <vinod.koul at intel.com>
>> Cc: Russell King <linux at arm.linux.org.uk>
>> Cc: Grant Likely <grant.likely at linaro.org>
>> Cc: Rob Herring <robh+dt at kernel.org>
>> Cc: Mark Rutland <mark.rutland at arm.com>
>> Signed-off-by: Sandeep Nair <sandeep_n at ti.com>
>> Signed-off-by: Santosh Shilimkar <santosh.shilimkar at ti.com>
>> ---
>>  .../devicetree/bindings/dma/keystone-pktdma.txt    |   72 ++
>>  drivers/dma/Kconfig                                |    8 +
>>  drivers/dma/Makefile                               |    1 +
>>  drivers/dma/keystone-pktdma.c                      |  795 ++++++++++++++++++++
>>  include/dt-bindings/dma/keystone.h                 |   33 +
>>  include/linux/keystone-pktdma.h                    |  168 +++++
>>  6 files changed, 1077 insertions(+)
>>  create mode 100644 Documentation/devicetree/bindings/dma/keystone-pktdma.txt
>>  create mode 100644 drivers/dma/keystone-pktdma.c
>>  create mode 100644 include/dt-bindings/dma/keystone.h
>>  create mode 100644 include/linux/keystone-pktdma.h
>>
>> diff --git a/Documentation/devicetree/bindings/dma/keystone-pktdma.txt b/Documentation/devicetree/bindings/dma/keystone-pktdma.txt
>> new file mode 100644
>> index 0000000..ea061d5
>> --- /dev/null
>> +++ b/Documentation/devicetree/bindings/dma/keystone-pktdma.txt
>> @@ -0,0 +1,72 @@
>> +Keystone Packet DMA Controller
>> +
>> +This document explains the device tree bindings for the packet dma
>> +on keystone devices. The the Network coprocessor, Cypto engines
>> +and the SRIO on Keystone devices all have their own packet dma modules.
>> +Each individual packet dma has a certain number of RX channels,
>> +RX flows and TX channels. Each instance of the packet DMA is being
>> +initialized through device specific bindings.
>> +
>> +* DMA controller
>> +
>> +Required properties:
>> + - compatible: Should be "ti,keystone-pktdma"
>> + - reg: Should contain register location and length of the following pktdma
>> +	register regions. The value for "reg-names" property of the respective
>> +	region is specified in parenthesis.
>> +	- Global control register region (global).
>> +	- Tx DMA channel configuration register region (txchan).
>> +	- Rx DMA channel configuration register region (rxchan).
>> +	- Tx DMA channel Scheduler configuration register region (txsched).
>> +	- Rx DMA flow configuration register region (rxflow).
>> + - reg-names: Names for the above regions. The name to be used is specified in
>> +	      the above description of "reg" property.
>> + - qm-base-address: Base address of the logical queue managers for pktdma.
>> + - #dma-cells: Has to be 1. Keystone-pktdma doesn't support anything else.
>> +
>> +Optional properties:
>> + - enable-all: Enable all DMA channels.
>> + - loop-back: To loopback Tx streaming I/F to Rx streaming I/F. Used for
>> +	      infrastructure transfers.
>> + - rx-retry-timeout: Number of pktdma cycles to wait before retry on buffer
>> +		     starvation.
>> +
>> +Example:
>> +	netcp-dma: pktdma at 2004000 {
>> +		compatible = "ti,keystone-pktdma";
>> +		reg =	<0x2004000 0x100>,
>> +			<0x2004400 0x120>,
>> +			<0x2004800 0x300>,
>> +			<0x2004c00 0x120>,
>> +			<0x2005000 0x400>;
>> +		reg-names = "global", "txchan", "rxchan", "txsched",
>> +			     "rxflow";
>> +		qm-base-address = <0x23a80000 0x23a90000
>> +				   0x23aa0000 0x23ab0000>;
>> +		#dma-cells = <1>;
>> +		/* enable-all; */
>> +		rx-retry-timeout = <3500>;
>> +		/* loop-back; */
>> +	};
>> +
>> +
>> +* DMA client
>> +
>> +Required properties:
>> +- dmas: One DMA request specifier consisting of a phandle to the DMA controller
>> +	followed by the integer specifying the channel identifier. The channel
>> +	identifier is encoded as follows:
>> +	- bits 7-0: Tx DMA channel number or the Rx flow number.
>> +	- bits 31-24: Channel type. 0xff for Tx DMA channel & 0xfe for Rx flow.
>> +- dma-names: List of string identifiers for the DMA requests.
>> +
>> +Example:
>> +
>> +	netcp: netcp at 2090000 {
>> +		...
>> +		dmas =	<&netcpdma KEYSTONE_DMA_RX_FLOW(22)>,
>> +			<&netcpdma KEYSTONE_DMA_RX_FLOW(23)>,
>> +			<&netcpdma KEYSTONE_DMA_TX_CHAN(8)>;
>> +			dma-names = "netrx0", "netrx1", "nettx";
>> +		...
>> +	};
> Can you pls separate the binding to separate patch and also this needs ack from DT
> folks
>
Will do that in next version. DT folks have been CC'ed on the patch. Will get
their ack on DT bindings.

 
[..]

>> diff --git a/drivers/dma/keystone-pktdma.c b/drivers/dma/keystone-pktdma.c
>> new file mode 100644
>> index 0000000..b3f77e5
>> --- /dev/null
>> +++ b/drivers/dma/keystone-pktdma.c
>> @@ -0,0 +1,795 @@
>> +/*
>> + * Copyright (C) 2014 Texas Instruments Incorporated
>> + * Authors:	Sandeep Nair <sandeep_n at ti.com>
>> + *		Cyril Chemparathy <cyril at ti.com>
>> + *		Santosh Shilimkar <santosh.shilimkar at ti.com>
>> + *
>> + * This program is free software; you can redistribute it and/or
>> + * modify it under the terms of the GNU General Public License as
>> + * published by the Free Software Foundation version 2.
>> + *
>> + * This program is distributed "as is" WITHOUT ANY WARRANTY of any
>> + * kind, whether express or implied; without even the implied warranty
>> + * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
>> + * GNU General Public License for more details.
>> + */
>> +
>> +#include <linux/io.h>
>> +#include <linux/sched.h>
>> +#include <linux/module.h>
>> +#include <linux/dma-direction.h>
>> +#include <linux/dmaengine.h>
>> +#include <linux/interrupt.h>
>> +#include <linux/of_dma.h>
>> +#include <linux/of_address.h>
>> +#include <linux/platform_device.h>
>> +#include <linux/keystone-pktdma.h>
>> +#include <linux/pm_runtime.h>
>> +#include <dt-bindings/dma/keystone.h>
>> +
>> +#define BITS(x)			(BIT(x) - 1)
> this might get confusing, perhaps a better name could be given?
> 
Would "BIT_MASK" be ok with you ?

>> +#define REG_MASK		0xffffffff
>> +
>> +#define DMA_LOOPBACK		BIT(31)
>> +#define DMA_ENABLE		BIT(31)
>> +#define DMA_TEARDOWN		BIT(30)
>> +
>> +#define DMA_TX_FILT_PSWORDS	BIT(29)
>> +#define DMA_TX_FILT_EINFO	BIT(30)
>> +#define DMA_TX_PRIO_SHIFT	0
>> +#define DMA_RX_PRIO_SHIFT	16
>> +#define DMA_PRIO_MASK		BITS(3)
>> +#define DMA_PRIO_DEFAULT	0
>> +#define DMA_RX_TIMEOUT_DEFAULT	17500 /* cycles */
>> +#define DMA_RX_TIMEOUT_MASK	BITS(16)
>> +#define DMA_RX_TIMEOUT_SHIFT	0
>> +
>> +#define CHAN_HAS_EPIB		BIT(30)
>> +#define CHAN_HAS_PSINFO		BIT(29)
>> +#define CHAN_ERR_RETRY		BIT(28)
>> +#define CHAN_PSINFO_AT_SOP	BIT(25)
>> +#define CHAN_SOP_OFF_SHIFT	16
>> +#define CHAN_SOP_OFF_MASK	BITS(9)
>> +#define DESC_TYPE_SHIFT		26
>> +#define DESC_TYPE_MASK		BITS(2)
>> +
>> +/*
>> + * QMGR & QNUM together make up 14 bits with QMGR as the 2 MSb's in the logical
>> + * navigator cloud mapping scheme.
>> + * using the 14bit physical queue numbers directly maps into this scheme.
>> + */
>> +#define CHAN_QNUM_MASK		BITS(14)
>> +#define DMA_MAX_QMS		4
>> +#define DMA_TIMEOUT		1000	/* msecs */
>> +
>> +struct reg_global {
>> +	u32	revision;
>> +	u32	perf_control;
>> +	u32	emulation_control;
>> +	u32	priority_control;
>> +	u32	qm_base_address[4];
>> +};
>> +
>> +struct reg_chan {
>> +	u32	control;
>> +	u32	mode;
>> +	u32	__rsvd[6];
>> +};
>> +
>> +struct reg_tx_sched {
>> +	u32	prio;
>> +};
>> +
>> +struct reg_rx_flow {
>> +	u32	control;
>> +	u32	tags;
>> +	u32	tag_sel;
>> +	u32	fdq_sel[2];
>> +	u32	thresh[3];
>> +};
>> +
>> +#define BUILD_CHECK_REGS()						\
>> +	do {								\
>> +		BUILD_BUG_ON(sizeof(struct reg_global)   != 32);	\
>> +		BUILD_BUG_ON(sizeof(struct reg_chan)     != 32);	\
>> +		BUILD_BUG_ON(sizeof(struct reg_rx_flow)  != 32);	\
>> +		BUILD_BUG_ON(sizeof(struct reg_tx_sched) !=  4);	\
>> +	} while (0)
> why is this required, do you want to use __packed__ to ensure right size?
>
This is just to ensure the register sanity. We should use __packed__ as
well. We can take the BUILD_CHECK_REGS() out if you don't prefer it.
 
>> +
>> +enum keystone_chan_state {
>> +	/* stable states */
>> +	CHAN_STATE_OPENED,
>> +	CHAN_STATE_CLOSED,
>> +};
>> +
>> +struct keystone_dma_device {
>> +	struct dma_device		engine;
>> +	bool				loopback, enable_all;
>> +	unsigned			tx_priority, rx_priority, rx_timeout;
>> +	unsigned			logical_queue_managers;
>> +	unsigned			qm_base_address[DMA_MAX_QMS];
>> +	struct reg_global __iomem	*reg_global;
>> +	struct reg_chan __iomem		*reg_tx_chan;
>> +	struct reg_rx_flow __iomem	*reg_rx_flow;
>> +	struct reg_chan __iomem		*reg_rx_chan;
>> +	struct reg_tx_sched __iomem	*reg_tx_sched;
>> +	unsigned			max_rx_chan, max_tx_chan;
>> +	unsigned			max_rx_flow;
>> +	atomic_t			in_use;
>> +};
>> +#define to_dma(dma)	(&(dma)->engine)
>> +#define dma_dev(dma)	((dma)->engine.dev)
>> +
>> +struct keystone_dma_chan {
>> +	struct dma_chan			achan;
>> +	enum dma_transfer_direction	direction;
>> +	atomic_t			state;
>> +	struct keystone_dma_device	*dma;
>> +
>> +	/* registers */
>> +	struct reg_chan __iomem		*reg_chan;
>> +	struct reg_tx_sched __iomem	*reg_tx_sched;
>> +	struct reg_rx_flow __iomem	*reg_rx_flow;
>> +
>> +	/* configuration stuff */
>> +	unsigned			channel, flow;
>> +};
>> +#define from_achan(ch)	container_of(ch, struct keystone_dma_chan, achan)
>> +#define to_achan(ch)	(&(ch)->achan)
>> +#define chan_dev(ch)	(&to_achan(ch)->dev->device)
>> +#define chan_num(ch)	((ch->direction == DMA_MEM_TO_DEV) ? \
>> +			ch->channel : ch->flow)
>> +#define chan_vdbg(ch, format, arg...)				\
>> +			dev_vdbg(chan_dev(ch), format, ##arg);
>> +
>> +/**
>> + * dev_to_dma_chan - convert a device pointer to the its sysfs container object
>> + * @dev - device node
>> + */
>> +static inline struct dma_chan *dev_to_dma_chan(struct device *dev)
>> +{
>> +	struct dma_chan_dev *chan_dev;
>> +
>> +	chan_dev = container_of(dev, typeof(*chan_dev), device);
>> +	return chan_dev->chan;
>> +}
>> +
>> +static inline enum keystone_chan_state
>> +chan_get_state(struct keystone_dma_chan *chan)
>> +{
>> +	return atomic_read(&chan->state);
>> +}
>> +
>> +static int chan_start(struct keystone_dma_chan *chan,
>> +			struct dma_keystone_cfg *cfg)
>> +{
>> +	u32 v = 0;
>> +
>> +	if ((chan->direction == DMA_MEM_TO_DEV) && chan->reg_chan) {
> why second check?
In case we to enable all channels with 'enable_all', then we can
skip per channel configuration. Second check is for that case.

>> +		if (cfg->u.tx.filt_pswords)
>> +			v |= DMA_TX_FILT_PSWORDS;
>> +		if (cfg->u.tx.filt_einfo)
>> +			v |= DMA_TX_FILT_EINFO;
>> +		writel_relaxed(v, &chan->reg_chan->mode);
>> +		writel_relaxed(DMA_ENABLE, &chan->reg_chan->control);
>> +	}
> no else for DMA_DEV_TO_MEM?

'chan->reg_rx_flow' case below takes care of "DMA_DEV_TO_MEM" case.
Since it was explicit, the additional direction check isn't used.

>> +	if (chan->reg_tx_sched)
>> +		writel_relaxed(cfg->u.tx.priority, &chan->reg_tx_sched->prio);
>> +
>> +	if (chan->reg_rx_flow) {
>> +		v = 0;
>> +
>> +		if (cfg->u.rx.einfo_present)
>> +			v |= CHAN_HAS_EPIB;
>> +		if (cfg->u.rx.psinfo_present)
>> +			v |= CHAN_HAS_PSINFO;
>> +		if (cfg->u.rx.err_mode == DMA_RETRY)
>> +			v |= CHAN_ERR_RETRY;
>> +		v |= (cfg->u.rx.desc_type & DESC_TYPE_MASK) << DESC_TYPE_SHIFT;
>> +		if (cfg->u.rx.psinfo_at_sop)
>> +			v |= CHAN_PSINFO_AT_SOP;
>> +		v |= (cfg->u.rx.sop_offset & CHAN_SOP_OFF_MASK)
>> +			<< CHAN_SOP_OFF_SHIFT;
>> +		v |= cfg->u.rx.dst_q & CHAN_QNUM_MASK;
>> +
>> +		writel_relaxed(v, &chan->reg_rx_flow->control);
>> +		writel_relaxed(0, &chan->reg_rx_flow->tags);
>> +		writel_relaxed(0, &chan->reg_rx_flow->tag_sel);
>> +
>> +		v =  cfg->u.rx.fdq[0] << 16;
>> +		v |=  cfg->u.rx.fdq[1] & CHAN_QNUM_MASK;
>> +		writel_relaxed(v, &chan->reg_rx_flow->fdq_sel[0]);
>> +
>> +		v =  cfg->u.rx.fdq[2] << 16;
>> +		v |=  cfg->u.rx.fdq[3] & CHAN_QNUM_MASK;
>> +		writel_relaxed(v, &chan->reg_rx_flow->fdq_sel[1]);
>> +
>> +		writel_relaxed(0, &chan->reg_rx_flow->thresh[0]);
>> +		writel_relaxed(0, &chan->reg_rx_flow->thresh[1]);
>> +		writel_relaxed(0, &chan->reg_rx_flow->thresh[2]);
>> +	}
>> +
>> +	return 0;
>> +}
>> +

[...]

>> +static int chan_init(struct dma_chan *achan)
>> +{
>> +	struct keystone_dma_chan *chan = from_achan(achan);
>> +	struct keystone_dma_device *dma = chan->dma;
>> +
>> +	chan_vdbg(chan, "initializing %s channel\n",
>> +		  chan->direction == DMA_MEM_TO_DEV   ? "transmit" :
>> +		  chan->direction == DMA_DEV_TO_MEM ? "receive"  :
>> +		  "unknown");
>> +
>> +	if (chan->direction != DMA_MEM_TO_DEV &&
>> +	    chan->direction != DMA_DEV_TO_MEM) {
>> +		dev_err(chan_dev(chan), "bad direction\n");
>> +		return -EINVAL;
>> +	}
> is_slave_direction() pls
> 
ok.

[..]

>> +static int chan_keystone_config(struct keystone_dma_chan *chan,
>> +		struct dma_keystone_cfg *cfg)
>> +{
>> +	if (chan_get_state(chan) != CHAN_STATE_OPENED)
>> +		return -ENODEV;
>> +
>> +	if (cfg->sl_cfg.direction != chan->direction)
>> +		return -EINVAL;
> direction is deprecated...
ok will remove that. The above check is suprefluous any ways.

[..]

>> +static int pktdma_init_chan(struct keystone_dma_device *dma,
>> +				struct device_node *node,
>> +				enum dma_transfer_direction dir,
>> +				unsigned chan_num)
>> +{
>> +	struct device *dev = dma_dev(dma);
>> +	struct keystone_dma_chan *chan;
>> +	struct dma_chan *achan;
>> +	int ret = -EINVAL;
>> +
>> +	chan = devm_kzalloc(dev, sizeof(*chan), GFP_KERNEL);
>> +	if (!chan)
>> +		return -ENOMEM;
>> +
>> +	achan = to_achan(chan);
>> +	achan->device   = &dma->engine;
>> +	chan->dma	= dma;
>> +	chan->direction	= DMA_NONE;
>> +	atomic_set(&chan->state, CHAN_STATE_OPENED);
>> +
>> +	if (dir == DMA_MEM_TO_DEV) {
>> +		chan->direction = dir;
>> +		ret = pktdma_init_tx_chan(chan, node, chan_num);
>> +	} else if (dir == DMA_DEV_TO_MEM) {
>> +		chan->direction = dir;
>> +		ret = pktdma_init_rx_chan(chan, node, chan_num);
>> +	} else {
>> +		dev_err(dev, "channel(%d) direction unknown\n", chan_num);
>> +	}
>> +
>> +	if (ret < 0)
>> +		goto fail;
>> +
>> +	list_add_tail(&to_achan(chan)->device_node, &to_dma(dma)->channels);
>> +	return 0;
>> +
>> +fail:
>> +	devm_kfree(dev, chan);
> ??
Missed out. Will removed that.


>> +/* dummy function: feature not supported */
>> +static enum dma_status chan_xfer_status(struct dma_chan *achan,
>> +				      dma_cookie_t cookie,
>> +				      struct dma_tx_state *txstate)
>> +{
>> +	WARN(1, "xfer status not supported\n");
>> +	return DMA_ERROR;
>> +}
>> +
>> +/* dummy function: feature not supported */
>> +static void chan_issue_pending(struct dma_chan *chan)
>> +{
>> +	WARN(1, "issue pending not supported\n");
>> +}
> Supporting status is okay but not issue_pending. This breaks use of dmaengine
> API. What we expect is that user will do channel allocation, prepare a
> descriptor, then submit it. Submit doesnt start the transaction, this call does!
> So we need implementation here!
>
As explained in the begining, the data movement prep, trigger and
completion is managed by QMSS and hence we don't have any implementation
here.. Thats how the hardware is designed. 
 
>> +
>> +static ssize_t keystone_dma_show_chan_num(struct device *dev,
>> +			     struct device_attribute *attr, char *buf)
>> +{
>> +	struct dma_chan *achan = dev_to_dma_chan(dev);
>> +	struct keystone_dma_chan *chan = from_achan(achan);
>> +
>> +	return scnprintf(buf, PAGE_SIZE, "%u\n", chan->channel);
>> +}
> ???
> 
>> +
>> +static ssize_t keystone_dma_show_flow(struct device *dev,
>> +			     struct device_attribute *attr, char *buf)
>> +{
>> +	struct dma_chan *achan = dev_to_dma_chan(dev);
>> +	struct keystone_dma_chan *chan = from_achan(achan);
>> +
>> +	return scnprintf(buf, PAGE_SIZE, "%u\n", chan->flow);
>> +}
>> +
>> +static DEVICE_ATTR(chan_num, S_IRUSR, keystone_dma_show_chan_num, NULL);
>> +static DEVICE_ATTR(rx_flow, S_IRUSR, keystone_dma_show_flow, NULL);
> okay why do we need these?
> 
The sysfs was added just to see runtime which DMA resources have been
used. No major need as such.

Regards,
Santosh