On 17-12-19, 16:33, Dave Jiang wrote:
> In order to provide a lockless submission path, the request context needs
> to be pre-allocated rather than pulling from a memory pool.
> Use the common request allocation call request_from_pages_alloc() to
> accomplish this. The sbitmap code will be used to get the next
> free request context. This is a simplified version of what blk-mq does
> (not sbitmap_queue). The config option DMA_ENGINE_REQUEST is added so that
> only drivers that supports dma request would enable the code.
Can you give more context on this requirement of lockless submission
path? I see this and next patch are adding another set of dma APIs, so
we need a good justification, documentation and why this cant be added
to existing code :)
>
> Signed-off-by: Dave Jiang <dave.jiang@xxxxxxxxx>
> ---
> drivers/dma/Kconfig | 6 +++
> drivers/dma/Makefile | 1
> drivers/dma/dma-request.c | 96 +++++++++++++++++++++++++++++++++++++++++++++
> include/linux/dmaengine.h | 57 +++++++++++++++++++++++++++
> 4 files changed, 160 insertions(+)
> create mode 100644 drivers/dma/dma-request.c
>
> diff --git a/drivers/dma/Kconfig b/drivers/dma/Kconfig
> index 6fa1eba9d477..52a3c2086dcb 100644
> --- a/drivers/dma/Kconfig
> +++ b/drivers/dma/Kconfig
> @@ -56,6 +56,12 @@ config DMA_OF
> depends on OF
> select DMA_ENGINE
>
> +config DMA_ENGINE_REQUEST
> + def_bool n
> + depends on DMA_ENGINE
> + select SBITMAP
> + select CONTEXT_ALLOC
> +
> #devices
> config ALTERA_MSGDMA
> tristate "Altera / Intel mSGDMA Engine"
> diff --git a/drivers/dma/Makefile b/drivers/dma/Makefile
> index 42d7e2fc64fa..f80720075399 100644
> --- a/drivers/dma/Makefile
> +++ b/drivers/dma/Makefile
> @@ -8,6 +8,7 @@ obj-$(CONFIG_DMA_ENGINE) += dmaengine.o
> obj-$(CONFIG_DMA_VIRTUAL_CHANNELS) += virt-dma.o
> obj-$(CONFIG_DMA_ACPI) += acpi-dma.o
> obj-$(CONFIG_DMA_OF) += of-dma.o
> +obj-$(CONFIG_DMA_ENGINE_REQUEST) += dma-request.o
>
> #dmatest
> obj-$(CONFIG_DMATEST) += dmatest.o
> diff --git a/drivers/dma/dma-request.c b/drivers/dma/dma-request.c
> new file mode 100644
> index 000000000000..43462fadf777
> --- /dev/null
> +++ b/drivers/dma/dma-request.c
> @@ -0,0 +1,96 @@
> +// SPDX-License-Identifier: GPL-2.0-or-later
> +/* Copyright(c) 2019 Intel Corporation. All rights reserved. */
> +#include <linux/init.h>
> +#include <linux/module.h>
> +#include <linux/mm.h>
> +#include <linux/device.h>
> +#include <linux/dmaengine.h>
> +#include <linux/mempool.h>
> +
> +struct dma_request *dma_chan_alloc_request(struct dma_chan *chan)
> +{
> + int nr;
> + struct dma_request *req;
> +
> + nr = sbitmap_get(&chan->sbmap, 0, false);
> + if (nr < 0)
> + return NULL;
> +
> + req = chan->rqs[nr];
> + req->rq_private = NULL;
> + req->callback = NULL;
> + memset(&req->result, 0, sizeof(struct dmaengine_result));
> + return req;
> +}
> +EXPORT_SYMBOL_GPL(dma_chan_alloc_request);
> +
> +void dma_chan_free_request(struct dma_chan *chan, struct dma_request *rq)
> +{
> + sbitmap_clear_bit(&chan->sbmap, rq->id);
> +}
> +EXPORT_SYMBOL_GPL(dma_chan_free_request);
> +
> +void dma_chan_free_request_resources(struct dma_chan *chan)
> +{
> + context_free_from_pages(&chan->page_list);
> + kfree(chan->rqs);
> +}
> +EXPORT_SYMBOL_GPL(dma_chan_free_request_resources);
> +
> +static void dma_chan_assign_request(void *ctx, void *ptr, int idx)
> +{
> + struct dma_chan *chan = (struct dma_chan *)ctx;
> + struct dma_request *rq = ptr;
> +
> + chan->rqs[idx] = rq;
> +}
> +
> +int dma_chan_alloc_request_resources(struct dma_chan *chan)
> +{
> + int i, node, rc, id = 0;
> + size_t rq_size;
> +
> + /* Requests are already allocated */
> + if (chan->rqs)
> + return 0;
> +
> + node = dev_to_node(chan->device->dev);
> + rc = sbitmap_init_node(&chan->sbmap, chan->depth, -1,
> + GFP_KERNEL, node);
> + if (rc < 0)
> + return rc;
> +
> + chan->rqs = kcalloc_node(chan->depth, sizeof(struct dma_request *),
> + GFP_KERNEL, node);
> + if (!chan->rqs) {
> + rc = -ENOMEM;
> + goto fail;
> + }
> +
> + INIT_LIST_HEAD(&chan->page_list);
> +
> + rq_size = round_up(sizeof(struct dma_request) +
> + chan->max_sgs * sizeof(struct scatterlist),
> + cache_line_size());
> +
> + rc = context_alloc_from_pages((void *)chan, chan->depth, rq_size,
> + &chan->page_list, 4, node,
> + dma_chan_assign_request);
> + if (rc < 0)
> + goto fail;
> +
> + for (i = 0; i < rc; i++) {
> + struct dma_request *rq = chan->rqs[i];
> +
> + rq->id = id++;
> + rq->chan = chan;
> + }
> +
> + return 0;
> +
> + fail:
> + sbitmap_free(&chan->sbmap);
> + dma_chan_free_request_resources(chan);
> + return rc;
> +}
> +EXPORT_SYMBOL_GPL(dma_chan_alloc_request_resources);
> diff --git a/include/linux/dmaengine.h b/include/linux/dmaengine.h
> index 0202d44a17a5..7bc8c3f8283f 100644
> --- a/include/linux/dmaengine.h
> +++ b/include/linux/dmaengine.h
> @@ -12,6 +12,8 @@
> #include <linux/scatterlist.h>
> #include <linux/bitmap.h>
> #include <linux/types.h>
> +#include <linux/sbitmap.h>
> +#include <linux/bvec.h>
> #include <asm/page.h>
>
> /**
> @@ -176,6 +178,8 @@ struct dma_interleaved_template {
> * @DMA_PREP_CMD: tell the driver that the data passed to DMA API is command
> * data and the descriptor should be in different format from normal
> * data descriptors.
> + * @DMA_SUBMIT_NONBLOCK: tell the driver do not wait for resources if submit
> + * is not possible.
> */
> enum dma_ctrl_flags {
> DMA_PREP_INTERRUPT = (1 << 0),
> @@ -186,6 +190,7 @@ enum dma_ctrl_flags {
> DMA_PREP_FENCE = (1 << 5),
> DMA_CTRL_REUSE = (1 << 6),
> DMA_PREP_CMD = (1 << 7),
> + DMA_SUBMIT_NONBLOCK = (1 << 8),
> };
>
> /**
> @@ -268,6 +273,13 @@ struct dma_chan {
> struct dma_router *router;
> void *route_data;
>
> + /* DMA request */
> + int max_sgs;
> + int depth;
> + struct sbitmap sbmap;
> + struct dma_request **rqs;
> + struct list_head page_list;
> +
> void *private;
> };
>
> @@ -511,6 +523,25 @@ struct dma_async_tx_descriptor {
> #endif
> };
>
> +struct dma_request {
> + int id;
> + struct dma_chan *chan;
> + enum dma_transaction_type cmd;
> + enum dma_ctrl_flags flags;
> + struct bio_vec bvec;
> + dma_addr_t pg_dma;
> + int sg_nents;
> + void *rq_private;
> +
> + /* Set by driver */
> + dma_async_tx_callback_result callback;
> + struct dmaengine_result result;
> + void *callback_param;
> +
> + /* Leave as last member for flexible array of scatterlist */
> + struct scatterlist sg[];
> +};
> +
> #ifdef CONFIG_DMA_ENGINE
> static inline void dma_set_unmap(struct dma_async_tx_descriptor *tx,
> struct dmaengine_unmap_data *unmap)
> @@ -1359,6 +1390,32 @@ static inline int dma_get_slave_caps(struct dma_chan *chan,
> }
> #endif
>
> +#ifdef CONFIG_DMA_ENGINE_REQUEST
> +struct dma_request *dma_chan_alloc_request(struct dma_chan *chan);
> +void dma_chan_free_request(struct dma_chan *chan, struct dma_request *rq);
> +void dma_chan_free_request_resources(struct dma_chan *chan);
> +int dma_chan_alloc_request_resources(struct dma_chan *chan);
> +#else
> +static inline struct dma_request *dma_chan_alloc_request(struct dma_chan *chan)
> +{
> + return NULL;
> +}
> +
> +static inline void dma_chan_free_request(struct dma_chan *chan,
> + struct dma_request *rq)
> +{
> +}
> +
> +static inline void dma_chan_free_request_resources(struct dma_chan *chan)
> +{
> +}
> +
> +static inline int dma_chan_alloc_request_resources(struct dma_chan *chan)
> +{
> + return -EOPNOTSUPP;
> +}
> +#endif
> +
> #define dma_request_slave_channel_reason(dev, name) dma_request_chan(dev, name)
>
> static inline int dmaengine_desc_set_reuse(struct dma_async_tx_descriptor *tx)
--
~Vinod
