On Mon, Sep 01, 2025 at 11:03:36AM +0530, Ekansh Gupta wrote:
> For any remote call to DSP, after sending an invocation message,
> fastRPC driver waits for glink response and during this time the
> CPU can go into low power modes. This adds latency to overall fastrpc
> call as CPU wakeup and scheduling latencies are included. Adding a
s/Adding/Add/, see Documentation/process/submitting-patches.rst
> polling mode support with which fastRPC driver will poll continuously
> on a memory after sending a message to remote subsystem which will
> eliminate CPU wakeup and scheduling latencies and reduce fastRPC
> overhead.
Describe your design decisions: when it is enabled, why, etc.
>
> Signed-off-by: Ekansh Gupta <ekansh.gu...@oss.qualcomm.com>
> ---
> drivers/misc/fastrpc.c | 121 ++++++++++++++++++++++++++++++++++++++---
> 1 file changed, 114 insertions(+), 7 deletions(-)
>
> diff --git a/drivers/misc/fastrpc.c b/drivers/misc/fastrpc.c
> index 57e118de6e4a..939a3e3d29e2 100644
> --- a/drivers/misc/fastrpc.c
> +++ b/drivers/misc/fastrpc.c
> @@ -22,6 +22,8 @@
> #include <linux/firmware/qcom/qcom_scm.h>
> #include <uapi/misc/fastrpc.h>
> #include <linux/of_reserved_mem.h>
> +#include <linux/compiler.h>
> +#include <linux/iopoll.h>
>
> #define ADSP_DOMAIN_ID (0)
> #define MDSP_DOMAIN_ID (1)
> @@ -37,6 +39,7 @@
> #define FASTRPC_CTX_MAX (256)
> #define FASTRPC_INIT_HANDLE 1
> #define FASTRPC_DSP_UTILITIES_HANDLE 2
> +#define FASTRPC_MAX_STATIC_HANDLE (20)
What is this?
> #define FASTRPC_CTXID_MASK (0xFF00)
> #define INIT_FILELEN_MAX (2 * 1024 * 1024)
> #define INIT_FILE_NAMELEN_MAX (128)
> @@ -105,6 +108,20 @@
>
> #define miscdev_to_fdevice(d) container_of(d, struct fastrpc_device, miscdev)
>
> +/* Poll response number from remote processor for call completion */
> +#define FASTRPC_POLL_RESPONSE (0xdecaf)
> +
> +/* Polling mode timeout limit */
> +#define FASTRPC_POLL_MAX_TIMEOUT_US (10000)
> +
> +/* Response types supported for RPC calls */
> +enum fastrpc_response_flags {
> + /* normal job completion glink response */
> + NORMAL_RESPONSE = 0,
> + /* process updates poll memory instead of glink response */
> + POLL_MODE = 1,
> +};
bool is_polled;
OR
unsigned long is_polled : 1;
> +
> struct fastrpc_phy_page {
> u64 addr; /* physical address */
> u64 size; /* size of contiguous region */
> @@ -235,8 +252,14 @@ struct fastrpc_invoke_ctx {
> u32 sc;
> u64 *fdlist;
> u32 *crc;
> + /* Poll memory that DSP updates */
> + u32 *poll;
> u64 ctxid;
> u64 msg_sz;
> + /* work done status flag */
> + bool is_work_done;
> + /* response flags from remote processor */
> + enum fastrpc_response_flags rsp_flags;
> struct kref refcount;
> struct list_head node; /* list of ctxs */
> struct completion work;
> @@ -891,7 +914,8 @@ static int fastrpc_get_meta_size(struct
> fastrpc_invoke_ctx *ctx)
> sizeof(struct fastrpc_invoke_buf) +
> sizeof(struct fastrpc_phy_page)) * ctx->nscalars +
> sizeof(u64) * FASTRPC_MAX_FDLIST +
> - sizeof(u32) * FASTRPC_MAX_CRCLIST;
> + sizeof(u32) * FASTRPC_MAX_CRCLIST +
> + sizeof(u32);
>
> return size;
> }
> @@ -987,6 +1011,8 @@ static int fastrpc_get_args(u32 kernel, struct
> fastrpc_invoke_ctx *ctx)
> list = fastrpc_invoke_buf_start(rpra, ctx->nscalars);
> pages = fastrpc_phy_page_start(list, ctx->nscalars);
> ctx->fdlist = (u64 *)(pages + ctx->nscalars);
> + ctx->crc = (u32 *)(ctx->fdlist + FASTRPC_MAX_FDLIST);
Why?
> + ctx->poll = (u32 *)(ctx->crc + FASTRPC_MAX_CRCLIST);
> args = (uintptr_t)ctx->buf->virt + metalen;
> rlen = pkt_size - metalen;
> ctx->rpra = rpra;
> @@ -1155,6 +1181,83 @@ static int fastrpc_invoke_send(struct
> fastrpc_session_ctx *sctx,
>
> }
>
> +static inline u32 fastrpc_poll_op(void *p)
> +{
> + struct fastrpc_invoke_ctx *ctx = p;
> +
> + dma_rmb();
> + return READ_ONCE(*ctx->poll);
Is this enough? Is the write by the DSP side going to invalidate the
cache for this memory location? Think about older platforms which
usually don't have dma-coherent property in the DSP / FastRPC nodes.
> +}
> +
> +static int poll_for_remote_response(struct fastrpc_invoke_ctx *ctx)
> +{
> + u32 val;
> + int ret;
> +
> + /*
> + * Poll until DSP writes FASTRPC_POLL_RESPONSE into *ctx->poll
> + * or until another path marks the work done.
> + */
> + ret = read_poll_timeout_atomic(fastrpc_poll_op, val,
> + (val == FASTRPC_POLL_RESPONSE) ||
> + ctx->is_work_done, 1,
> + FASTRPC_POLL_MAX_TIMEOUT_US, false, ctx);
> +
> + if (!ret && val == FASTRPC_POLL_RESPONSE) {
> + ctx->is_work_done = true;
> + ctx->retval = 0;
> + }
> +
> + if (ret == -ETIMEDOUT)
> + ret = -EIO;
> +
> + return ret;
> +}
> +
> +static inline int fastrpc_wait_for_response(struct fastrpc_invoke_ctx *ctx,
> + u32 kernel)
> +{
> + int err = 0;
> +
> + if (kernel) {
> + if (!wait_for_completion_timeout(&ctx->work, 10 * HZ))
> + err = -ETIMEDOUT;
> + } else {
> + err = wait_for_completion_interruptible(&ctx->work);
> + }
> +
> + return err;
> +}
> +
> +static int fastrpc_wait_for_completion(struct fastrpc_invoke_ctx *ctx,
> + u32 kernel)
> +{
> + int err;
> +
> + do {
> + switch (ctx->rsp_flags) {
> + case NORMAL_RESPONSE:
> + err = fastrpc_wait_for_response(ctx, kernel);
> + if (err || ctx->is_work_done)
> + return err;
> + break;
> + case POLL_MODE:
> + err = poll_for_remote_response(ctx);
> + /* If polling timed out, move to normal response mode */
> + if (err)
> + ctx->rsp_flags = NORMAL_RESPONSE;
> + break;
> + default:
What kind of response type can it be? Have you had checked for the flag
being set, you wouldn't have a false possibility of having another
response type.
> + err = -EBADR;
> + dev_dbg(ctx->fl->sctx->dev,
> + "unsupported response type:0x%x\n",
> ctx->rsp_flags);
> + break;
> + }
> + } while (!ctx->is_work_done);
> +
> + return err;
> +}
> +
> static int fastrpc_internal_invoke(struct fastrpc_user *fl, u32 kernel,
> u32 handle, u32 sc,
> struct fastrpc_invoke_args *args)
> @@ -1190,16 +1293,19 @@ static int fastrpc_internal_invoke(struct
> fastrpc_user *fl, u32 kernel,
> if (err)
> goto bail;
>
> - if (kernel) {
> - if (!wait_for_completion_timeout(&ctx->work, 10 * HZ))
> - err = -ETIMEDOUT;
> - } else {
> - err = wait_for_completion_interruptible(&ctx->work);
> - }
> + if (handle > FASTRPC_MAX_STATIC_HANDLE && fl->pd == USER_PD)
> + ctx->rsp_flags = POLL_MODE;
This definitely needs to be explained.
>
> + err = fastrpc_wait_for_completion(ctx, kernel);
> if (err)
> goto bail;
>
> + if (!ctx->is_work_done) {
> + err = -ETIMEDOUT;
> + dev_dbg(fl->sctx->dev, "Invalid workdone state for handle 0x%x,
> sc 0x%x\n",
> + handle, sc);
> + goto bail;
> + }
> /* make sure that all memory writes by DSP are seen by CPU */
> dma_rmb();
> /* populate all the output buffers with results */
> @@ -2462,6 +2568,7 @@ static int fastrpc_rpmsg_callback(struct rpmsg_device
> *rpdev, void *data,
>
> ctx->retval = rsp->retval;
> complete(&ctx->work);
> + ctx->is_work_done = true;
>
> /*
> * The DMA buffer associated with the context cannot be freed in
> --
> 2.34.1
>
--
With best wishes
Dmitry
