On Thu, Jul 29, 2021 at 6:56 AM fengchengwen [off-list ref] wrote:

Thanks, inline comment

On 2021/7/28 19:13, Bruce Richardson wrote:

quoted

On Tue, Jul 27, 2021 at 11:39:59AM +0800, Chengwen Feng wrote:

quoted

This patch introduce 'dmadevice' which is a generic type of DMA
device.

The APIs of dmadev library exposes some generic operations which can
enable configuration and I/O with the DMA devices.

Signed-off-by: Chengwen Feng <redacted>
Acked-by: Bruce Richardson <redacted>
Acked-by: Morten Brørup <redacted>
---

Thanks for this. Before it gets merged, I believe it needs to be split
further into multiple patches (say 4 or so) rather than adding the whole
lib in one go.

Normally, I believe the split would be something like:
* basic device structures and infrastructure e.g. alloc and release
  functions
* device config functions (and structures to go along with them)
  such as configure and queue_setup
* data plane functions

I will try for it
Maybe one patch for public file, one for pmd header file, one for
implementation, and last for doc.

+1.

quoted

Documentation would be included in each of the patches, rather than done as
a block at the end.

Besides that, I have one small additional requests for the API. Based off
feedback for ioat driver, we added in the following function to that API,
and we probably need something similar in dmadev:

rte_ioat_burst_capacity()

For our implementation this returns the number of elements that can be
enqueued to the ring, at least for the current burst/batch of packets. We
did the API this way because there can be additional limits beyond ring
size on each individual burst beyond just the raw ring capacity, e.g. even
if there are 4k ring elements free, there may be limits on the max burst
size the hardware can do, or limits on the number of outstanding
batches etc.

Therefore can I request the addition of rte_dmadev_burst_capacity() [or
something similarly named] to the basic dmadev API set. For most hardware,
I think this will likely be the remaining free ring size, but I don't
believe the API should commit to that. The use case it was added for was to
enable an application which needs to do a multi-copy operation to check
that all copies can fit or not before enqueuing the first one. This is
important for hardware that doesn't have scatter-gather list support.

Yes. Could you add the following to enable scatter-gather support in
rte_dmadev_info::sge_max
/**<Maximum allowed number of scatter-gather entries in a single sg call. */
        uint16_t sge_max;

Remaining capacity can be inferred by ring_idx which return from enqueue and
dequeue APIs.
So I don't think this API needs to be added.

For scatter-gather list, there maybe a hardware limit for max src or dst entry
size, I prefer add 'max_sges' filed in struct rte_dmadev_info to indicate it.

quoted

/Bruce

PS: One typo in code flagged below too.

<snip>

quoted

+ */
+enum rte_dma_status_code {
+    RTE_DMA_STATUS_SUCCESSFUL,
+    /**< The operation completed successfully. */
+    RTE_DMA_STATUS_USRER_ABORT,

              Typo here ^^^

OK, USRER->USER will fix later

quoted

.