From: fengchengwen <fengchengwen@huawei.com>
To: Vamsi Krishna <vattunuru@marvell.com>, <dev@dpdk.org>
Cc: <thomas@monjalon.net>, <bruce.richardson@intel.com>,
<anatoly.burakov@intel.com>, <kevin.laatz@intel.com>,
<jerinj@marvell.com>
Subject: Re: [PATCH v2 1/1] lib/dma: add control-plane APIs for inter-domain DMA transfers
Date: Wed, 15 Oct 2025 14:59:07 +0800 [thread overview]
Message-ID: <c8041ddd-0ad9-4fa9-93f9-9cd26cb9ce4c@huawei.com> (raw)
In-Reply-To: <20251013181049.2396041-1-vattunuru@marvell.com>
On 10/14/2025 2:10 AM, Vamsi Krishna wrote:
> From: Vamsi Attunuru <vattunuru@marvell.com>
>
> Secure and controlled inter-domain DMA transfers require dedicated
> control-plane APIs to create and manage access groups.
>
> DMA devices used for inter-process data transfer can be categorized as
> follows:
>
> Class A: Both endpoints require a DMA device for data transfer (e.g.,
> Marvell DMA devices).
> Class B: Only one endpoint requires a DMA device; the other does not.
> Class C: Other device types not currently classified.
>
> Providing a unified API for all these categories is complex, as Linux
> and other operating systems do not offer native control-plane APIs for
> this purpose. Therefore, DPDK can implement its own control-plane
> mechanisms to support Class A, B, and C devices.
>
> This commit introduces the necessary APIs for Class A DMA devices.
> Devices can create or join access groups using token-based
> authentication, ensuring that only authorized devices within the same
> group can perform DMA transfers across processes or OS domains.
>
> API Usage Flow:
>
> Process 1 (Group Creator):
> Invokes rte_dma_access_pair_group_create() to establish a new access
> pair group, then shares the group_id and token with Process 2 via IPC.
>
> Process 2 (Group Joiner):
> Receives the group_id and token from Process 1 and calls
> rte_dma_access_pair_group_join() to join the group.
>
> Both Processes:
> Use rte_dma_access_pair_group_handle_get() to obtain handler information
> for domains in the group.
> Perform DMA transfers as required.
>
> Process 2 (when finished):
> Calls rte_dma_access_pair_group_leave() to exit the group.
>
> Process 1:
> Monitors for group updates and confirms group membership as needed.
>
> Depends-on: patch-36335 (lib/dma: introduce inter-process and inter-OS DMA)
>
> Signed-off-by: Vamsi Attunuru <vattunuru@marvell.com>
> ---
> lib/dmadev/rte_dmadev.c | 137 ++++++++++++++++++++++++++++++++++
> lib/dmadev/rte_dmadev.h | 100 +++++++++++++++++++++++++
> lib/dmadev/rte_dmadev_pmd.h | 28 +++++++
> lib/dmadev/rte_dmadev_trace.h | 42 +++++++++++
> 4 files changed, 307 insertions(+)
>
> diff --git a/lib/dmadev/rte_dmadev.c b/lib/dmadev/rte_dmadev.c
> index d6792449bf..521b372b86 100644
> --- a/lib/dmadev/rte_dmadev.c
> +++ b/lib/dmadev/rte_dmadev.c
> @@ -804,6 +804,143 @@ rte_dma_vchan_status(int16_t dev_id, uint16_t vchan, enum rte_dma_vchan_status *
> return dev->dev_ops->vchan_status(dev, vchan, status);
> }
>
> +int
> +rte_dma_access_pair_group_create(int16_t dev_id, rte_uuid_t domain_id, rte_uuid_t token,
> + int16_t *group_id)
> +{
> + struct rte_dma_info dev_info;
> + struct rte_dma_dev *dev;
> +
> + if (!rte_dma_is_valid(dev_id) || group_id == NULL)
> + return -EINVAL;
> + dev = &rte_dma_devices[dev_id];
> +
> + if (rte_dma_info_get(dev_id, &dev_info)) {
> + RTE_DMA_LOG(ERR, "Device %d get device info fail", dev_id);
> + return -EINVAL;
> + }
> +
> + if (!((dev_info.dev_capa & RTE_DMA_CAPA_INTER_PROCESS_DOMAIN) ||
> + (dev_info.dev_capa & RTE_DMA_CAPA_INTER_OS_DOMAIN))) {
> + RTE_DMA_LOG(ERR, "Device %d don't support inter-process or inter-os domain",
> + dev_id);
> + return -EINVAL;
> + }
please add one blank line here
> + if (*dev->dev_ops->access_pair_group_create == NULL)
> + return -ENOTSUP;
> + return (*dev->dev_ops->access_pair_group_create)(dev, domain_id, token, group_id);
> +}
> +
> +int
> +rte_dma_access_pair_group_destroy(int16_t dev_id, int16_t group_id)
> +{
> + struct rte_dma_info dev_info;
> + struct rte_dma_dev *dev;
> +
> + if (!rte_dma_is_valid(dev_id))
> + return -EINVAL;
> + dev = &rte_dma_devices[dev_id];
> +
> + if (rte_dma_info_get(dev_id, &dev_info)) {
> + RTE_DMA_LOG(ERR, "Device %d get device info fail", dev_id);
> + return -EINVAL;
> + }
> +
> + if (!((dev_info.dev_capa & RTE_DMA_CAPA_INTER_PROCESS_DOMAIN) ||
> + (dev_info.dev_capa & RTE_DMA_CAPA_INTER_OS_DOMAIN))) {
> + RTE_DMA_LOG(ERR, "Device %d don't support inter-process or inter-os domain",
> + dev_id);
> + return -EINVAL;
> + }
> +
> + if (*dev->dev_ops->access_pair_group_destroy == NULL)
> + return -ENOTSUP;
> + return (*dev->dev_ops->access_pair_group_destroy)(dev, group_id);
> +}
> +
> +int
> +rte_dma_access_pair_group_join(int16_t dev_id, int16_t group_id, rte_uuid_t token,
> + rte_dma_access_pair_leave_cb_t leave_cb)
> +{
> + struct rte_dma_info dev_info;
> + struct rte_dma_dev *dev;
> +
> + if (!rte_dma_is_valid(dev_id))
> + return -EINVAL;
> + dev = &rte_dma_devices[dev_id];
> +
> + if (rte_dma_info_get(dev_id, &dev_info)) {
> + RTE_DMA_LOG(ERR, "Device %d get device info fail", dev_id);
> + return -EINVAL;
> + }
> +
> + if (!((dev_info.dev_capa & RTE_DMA_CAPA_INTER_PROCESS_DOMAIN) ||
> + (dev_info.dev_capa & RTE_DMA_CAPA_INTER_OS_DOMAIN))) {
> + RTE_DMA_LOG(ERR, "Device %d don't support inter-process or inter-os domain",
> + dev_id);
> + return -EINVAL;
> + }
> +
> + if (*dev->dev_ops->access_pair_group_join == NULL)
> + return -ENOTSUP;
> + return (*dev->dev_ops->access_pair_group_join)(dev, group_id, token, leave_cb);
> +}
> +
> +int
> +rte_dma_access_pair_group_leave(int16_t dev_id, int16_t group_id)
> +{
> + struct rte_dma_info dev_info;
> + struct rte_dma_dev *dev;
> +
> + if (!rte_dma_is_valid(dev_id))
> + return -EINVAL;
> + dev = &rte_dma_devices[dev_id];
> +
> + if (rte_dma_info_get(dev_id, &dev_info)) {
> + RTE_DMA_LOG(ERR, "Device %d get device info fail", dev_id);
> + return -EINVAL;
> + }
> +
> + if (!((dev_info.dev_capa & RTE_DMA_CAPA_INTER_PROCESS_DOMAIN) ||
> + (dev_info.dev_capa & RTE_DMA_CAPA_INTER_OS_DOMAIN))) {
> + RTE_DMA_LOG(ERR, "Device %d don't support inter-process or inter-os domain",
> + dev_id);
> + return -EINVAL;
> + }
> +
> + if (*dev->dev_ops->access_pair_group_leave == NULL)
> + return -ENOTSUP;
> + return (*dev->dev_ops->access_pair_group_leave)(dev, group_id);
> +}
> +
> +int
> +rte_dma_access_pair_group_handle_get(int16_t dev_id, int16_t group_id, rte_uuid_t domain_id,
> + uint16_t *handle)
> +{
> + struct rte_dma_info dev_info;
> + struct rte_dma_dev *dev;
> +
> + if (!rte_dma_is_valid(dev_id) || handle == NULL)
> + return -EINVAL;
> + dev = &rte_dma_devices[dev_id];
> +
> + if (rte_dma_info_get(dev_id, &dev_info)) {
> + RTE_DMA_LOG(ERR, "Device %d get device info fail", dev_id);
> + return -EINVAL;
> + }
> +
> + if (!((dev_info.dev_capa & RTE_DMA_CAPA_INTER_PROCESS_DOMAIN) ||
> + (dev_info.dev_capa & RTE_DMA_CAPA_INTER_OS_DOMAIN))) {
> + RTE_DMA_LOG(ERR, "Device %d don't support inter-process or inter-os domain",
> + dev_id);
> + return -EINVAL;
> + }
> +
> + if (*dev->dev_ops->access_pair_group_handle_get == NULL)
> + return -ENOTSUP;
> + return (*dev->dev_ops->access_pair_group_handle_get)(dev, group_id, domain_id, handle);
> +}
> +
> static const char *
> dma_capability_name(uint64_t capability)
> {
> diff --git a/lib/dmadev/rte_dmadev.h b/lib/dmadev/rte_dmadev.h
> index 4f735bb0c9..a6e73337eb 100644
> --- a/lib/dmadev/rte_dmadev.h
> +++ b/lib/dmadev/rte_dmadev.h
> @@ -148,6 +148,7 @@
>
> #include <rte_bitops.h>
> #include <rte_common.h>
> +#include <rte_uuid.h>
>
> #ifdef __cplusplus
> extern "C" {
> @@ -779,6 +780,105 @@ rte_dma_vchan_status(int16_t dev_id, uint16_t vchan, enum rte_dma_vchan_status *
> */
> int rte_dma_dump(int16_t dev_id, FILE *f);
>
> +/**
> + * Callback function invoked when a DMA device leaves an access group or when the group is
> + * destroyed due to some exception or failure.
> + *
> + * @param dma_id
> + * Identifier of the DMA device that is leaving the group.
> + * @param group_id
> + * ID of the access group being exited.
> + * @param user_data
> + * User-defined data passed to the callback for context.
What's the user_data define? vendor specific?
> + */
> +typedef void (*rte_dma_access_pair_leave_cb_t)(int16_t dma_id, int16_t group_id, void *user_data);
> +
> +/**
> + * Create an access pair group to enable secure DMA transfers between devices across different
> + * processes or operating system domains.
> + *
> + * @param dev_id
> + * Identifier of the DMA device initiating the group.
> + * @param domain_id
> + * Unique identifier representing the process or OS domain.
> + * @param token
> + * Authentication token used to establish the access group.
> + * @param[out] group_id
> + * Pointer to store the ID of the newly created access group.
> + *
> + * @return
> + * 0 on success,
> + * negative error code on failure.
> + */
> +int rte_dma_access_pair_group_create(int16_t dev_id, rte_uuid_t domain_id, rte_uuid_t token,
> + int16_t *group_id);
add blank line here
> +/**
> + * Destroy an access pair group if all participating devices have exited. This operation is only
> + * permitted by the device that originally created the group; attempts by other devices will result
> + * in failure.
> + *
> + * @param dev_id
> + * Identifier of the device requesting group destruction.
> + * @param group_id
> + * ID of the access group to be destroyed.
> + * @return
> + * 0 on success,
> + * negative value on failure indicating the error code.
> + */
> +int rte_dma_access_pair_group_destroy(int16_t dev_id, int16_t group_id);
add blank line here
> +/**
> + * Join an existing access group to enable secure DMA transfers between devices across different
> + * processes or OS domains.
> + *
> + * @param dev_id
> + * Identifier of the DMA device attempting to join the group.
> + * @param group_id
> + * ID of the access group to join.
> + * @param token
> + * Authentication token used to validate group membership.
> + * @param leave_cb
> + * Callback function to be invoked when the device leaves the group or when the group
> + * is destroyed due to some exception or failure.
> + *
> + * @return
> + * 0 on success,
> + * negative value on failure indicating the error code.
> + */
> +int rte_dma_access_pair_group_join(int16_t dev_id, int16_t group_id, rte_uuid_t token,
> + rte_dma_access_pair_leave_cb_t leave_cb);
add blank line here
> +/**
> + * Leave an access group, removing the device's entry from the group table and disabling
> + * inter-domain DMA transfers to and from this device. This operation is not permitted for the
> + * device that originally created the group.
> + *
> + * @param dev_id
> + * Identifier of the device requesting to leave the group.
> + * @param group_id
> + * ID of the access group to leave.
> + * @return
> + * 0 on success,
> + * negative value on failure indicating the error code.
> + */
> +int rte_dma_access_pair_group_leave(int16_t dev_id, int16_t group_id);
add blank line here
> +/**
> + * Retrieve the handler associated with a specific domain ID, which is used by the application to
> + * query source or destinationin handler to initiate inter-process or inter-OS DMA transfers.
> + *
> + * @param dev_id
> + * Identifier of the DMA device requesting the handler.
> + * @param group_id
> + * ID of the access group to query.
> + * @param group_tbl
> + * Unique identifier of the target process or OS domain.
> + * @param[out] handle
> + * Pointer to store the retrieved handler value.
> + * @return
> + * 0 on success,
> + * negative value on failure indicating the error code.
> + */
> +int rte_dma_access_pair_group_handle_get(int16_t dev_id, int16_t group_id, rte_uuid_t domain_id,
> + uint16_t *handle);
handle should be handler which keep consistency with depend commit.
if there are three OS join the same domain, then how this API get specific handler?
What I think is:
1. every domain has its own unique domain_id
2. Process 1 create one group, and share group_id and token with Process 2/3 via IPC.
3. Process 2 join the group (pass its domain_id)
4. Process 3 also join the group (pass its domain_id)
5. When Process 1 want to transfer between diffent domain:
it first invoke rte_dma_access_pair_group_handle_get() to get target handler
do vchan setup. there have three type: local to remote, remote to local, remote to remote. but how to describe the type I have see in depend commit.
do transfer.
6. If one process leave the group, then other process will be invoke by a callback.
this has a problem, the Process 1 hasn't register a callback for such event.
> +
> /**
> * DMA transfer result status code defines.
> *
> diff --git a/lib/dmadev/rte_dmadev_pmd.h b/lib/dmadev/rte_dmadev_pmd.h
> index 58729088ff..8be4485593 100644
> --- a/lib/dmadev/rte_dmadev_pmd.h
> +++ b/lib/dmadev/rte_dmadev_pmd.h
> @@ -64,6 +64,28 @@ typedef int (*rte_dma_vchan_status_t)(const struct rte_dma_dev *dev, uint16_t vc
> /** @internal Used to dump internal information. */
> typedef int (*rte_dma_dump_t)(const struct rte_dma_dev *dev, FILE *f);
>
> +/** @internal Used to create an access pair group for inter-process or inter-OS DMA transfers. */
> +typedef int (*rte_dma_access_pair_group_create_t)(const struct rte_dma_dev *dev,
> + rte_uuid_t domain_id, rte_uuid_t token,
> + int16_t *group_id);
> +
> +/** @internal Used to destroy an access pair group when all other devices have exited. */
> +typedef int (*rte_dma_access_pair_group_destroy_t)(const struct rte_dma_dev *dev,
> + int16_t group_id);
> +
> +/** @internal Used to join an access pair group for inter-process or inter-OS DMA transfers. */
> +typedef int (*rte_dma_access_pair_group_join_t)(const struct rte_dma_dev *dev, int16_t group_id,
> + rte_uuid_t token,
> + rte_dma_access_pair_leave_cb_t leave_cb);
> +
> +/** @internal Used to leave an access pair group, removing the device from the group. */
> +typedef int (*rte_dma_access_pair_group_leave_t)(const struct rte_dma_dev *dev, int16_t group_id);
> +
> +/** @internal Used to retrieve handler information of the domain_id present in the group. */
> +typedef int (*rte_dma_access_pair_group_handle_get_t)(const struct rte_dma_dev *dev,
> + int16_t group_id, rte_uuid_t domain_id,
> + uint16_t *handle);
> +
> /**
> * DMA device operations function pointer table.
> *
> @@ -83,6 +105,12 @@ struct rte_dma_dev_ops {
>
> rte_dma_vchan_status_t vchan_status;
> rte_dma_dump_t dev_dump;
> +
> + rte_dma_access_pair_group_create_t access_pair_group_create;
> + rte_dma_access_pair_group_destroy_t access_pair_group_destroy;
> + rte_dma_access_pair_group_join_t access_pair_group_join;
> + rte_dma_access_pair_group_leave_t access_pair_group_leave;
> + rte_dma_access_pair_group_handle_get_t access_pair_group_handle_get;
> };
>
> /**
> diff --git a/lib/dmadev/rte_dmadev_trace.h b/lib/dmadev/rte_dmadev_trace.h
> index fa619c4090..c25cf8d60a 100644
> --- a/lib/dmadev/rte_dmadev_trace.h
> +++ b/lib/dmadev/rte_dmadev_trace.h
> @@ -101,6 +101,48 @@ RTE_TRACE_POINT(
> rte_trace_point_emit_int(ret);
> )
>
> +RTE_TRACE_POINT(
> + rte_dma_trace_access_pair_group_create,
> + RTE_TRACE_POINT_ARGS(int16_t dev_id, rte_uuid_t domain_id, rte_uuid_t token,
> + int16_t *group_id),
> + rte_trace_point_emit_i16(dev_id);
> + rte_trace_point_emit_u8_ptr(&domain_id[0]);
> + rte_trace_point_emit_u8_ptr(&token[0]);
> + rte_trace_point_emit_ptr(group_id);
> +)
> +
> +RTE_TRACE_POINT(
> + rte_dma_trace_access_pair_group_destroy,
> + RTE_TRACE_POINT_ARGS(int16_t dev_id, int16_t group_id),
> + rte_trace_point_emit_i16(dev_id);
> + rte_trace_point_emit_i16(group_id);
> +)
> +
> +RTE_TRACE_POINT(
> + rte_dma_trace_access_pair_group_join,
> + RTE_TRACE_POINT_ARGS(int16_t dev_id, int16_t group_id, rte_uuid_t token),
> + rte_trace_point_emit_i16(dev_id);
> + rte_trace_point_emit_i16(group_id);
> + rte_trace_point_emit_u8_ptr(&token[0]);
> +)
> +
> +RTE_TRACE_POINT(
> + rte_dma_trace_access_pair_group_leave,
> + RTE_TRACE_POINT_ARGS(int16_t dev_id, int16_t group_id),
> + rte_trace_point_emit_i16(dev_id);
> + rte_trace_point_emit_i16(group_id);
> +)
> +
> +RTE_TRACE_POINT(
> + rte_dma_trace_access_pair_group_handle_get,
> + RTE_TRACE_POINT_ARGS(int16_t dev_id, int16_t group_id, rte_uuid_t domain_id,
> + uint16_t *handle),
> + rte_trace_point_emit_i16(dev_id);
> + rte_trace_point_emit_i16(group_id);
> + rte_trace_point_emit_u8_ptr(&domain_id[0]);
> + rte_trace_point_emit_ptr(handle);
> +)
> +
> #ifdef __cplusplus
> }
> #endif
please add descript in DMAdev's doc.
prev parent reply other threads:[~2025-10-15 6:59 UTC|newest]
Thread overview: 17+ messages / expand[flat|nested] mbox.gz Atom feed top
2025-09-01 12:33 [RFC] lib/dma: introduce inter-process and inter-OS DMA Vamsi Krishna
2025-09-18 11:06 ` Vamsi Krishna Attunuru
2025-09-19 9:02 ` fengchengwen
2025-09-22 11:48 ` [EXTERNAL] " Vamsi Krishna Attunuru
2025-09-24 4:14 ` Vamsi Krishna Attunuru
2025-09-25 1:34 ` fengchengwen
2025-10-01 5:57 ` Jerin Jacob
2025-10-06 13:59 ` Vamsi Krishna Attunuru
2025-10-09 2:27 ` Vamsi Krishna Attunuru
2025-10-09 11:08 ` fengchengwen
2025-10-10 10:40 ` Jerin Jacob
2025-09-25 2:06 ` fengchengwen
2025-10-10 14:46 ` [PATCH v2 1/1] " Vamsi Krishna
2025-10-12 1:49 ` fengchengwen
2025-10-13 4:07 ` [EXTERNAL] " Vamsi Krishna Attunuru
2025-10-13 18:10 ` [PATCH v2 1/1] lib/dma: add control-plane APIs for inter-domain DMA transfers Vamsi Krishna
2025-10-15 6:59 ` fengchengwen [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=c8041ddd-0ad9-4fa9-93f9-9cd26cb9ce4c@huawei.com \
--to=fengchengwen@huawei.com \
--cc=anatoly.burakov@intel.com \
--cc=bruce.richardson@intel.com \
--cc=dev@dpdk.org \
--cc=jerinj@marvell.com \
--cc=kevin.laatz@intel.com \
--cc=thomas@monjalon.net \
--cc=vattunuru@marvell.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).