On 9/27/23 09:08, Maxime Coquelin wrote:
On 9/21/23 19:18, Chautru, Nicolas wrote:
Hi David,
-----Original Message-----
From: David Marchand <david.march...@redhat.com>
Sent: Thursday, September 21, 2023 12:13 AM
To: Chautru, Nicolas <nicolas.chau...@intel.com>
Cc: dev@dpdk.org; maxime.coque...@redhat.com;
hemant.agra...@nxp.com; Vargas, Hernan <hernan.var...@intel.com>;
Thomas Monjalon <tho...@monjalon.net>
Subject: Re: [PATCH v1 3/7] baseband/acc: remove the 4G SO capability
for
VRB1
On Tue, Sep 19, 2023 at 10:32 PM Chautru, Nicolas
<nicolas.chau...@intel.com> wrote:
Hi David,
-----Original Message-----
From: David Marchand <david.march...@redhat.com>
Sent: Tuesday, September 19, 2023 8:20 AM
To: Chautru, Nicolas <nicolas.chau...@intel.com>
Cc: dev@dpdk.org; maxime.coque...@redhat.com;
hemant.agra...@nxp.com; Vargas, Hernan <hernan.var...@intel.com>
Subject: Re: [PATCH v1 3/7] baseband/acc: remove the 4G SO
capability for
VRB1
On Tue, Sep 19, 2023 at 3:25 AM Nicolas Chautru
<nicolas.chau...@intel.com> wrote:
This removes the specific capability and support of LTE Decoder
Soft Output option on the VRB1 PMD.
Please explain why such support is removed for this hw.
The decision is made to defeature this optional capability as under
certain
race conditions enabling this may potentially cause reliability
issues which
would not be acceptable.
Note that this is an optional additional output information (soft
output
information) independent of the actual decoding operation.
More details below next to your other comments.
This must be explained in the commitlog.
OK will add now.
Signed-off-by: Nicolas Chautru <nicolas.chau...@intel.com>
---
drivers/baseband/acc/rte_vrb_pmd.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/drivers/baseband/acc/rte_vrb_pmd.c
b/drivers/baseband/acc/rte_vrb_pmd.c
index 3c8f3409ed..e0f50460bd 100644
--- a/drivers/baseband/acc/rte_vrb_pmd.c
+++ b/drivers/baseband/acc/rte_vrb_pmd.c
@@ -1019,14 +1019,11 @@ vrb_dev_info_get(struct rte_bbdev *dev,
struct
rte_bbdev_driver_info *dev_info)
RTE_BBDEV_TURBO_CRC_TYPE_24B |
RTE_BBDEV_TURBO_DEC_CRC_24B_DROP |
RTE_BBDEV_TURBO_EQUALIZER |
-
RTE_BBDEV_TURBO_SOFT_OUT_SATURATE |
RTE_BBDEV_TURBO_HALF_ITERATION_EVEN |
RTE_BBDEV_TURBO_CONTINUE_CRC_MATCH |
- RTE_BBDEV_TURBO_SOFT_OUTPUT |
RTE_BBDEV_TURBO_EARLY_TERMINATION |
RTE_BBDEV_TURBO_DEC_INTERRUPTS |
RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN |
-
RTE_BBDEV_TURBO_NEG_LLR_1_BIT_SOFT_OUT |
RTE_BBDEV_TURBO_MAP_DEC |
RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP |
RTE_BBDEV_TURBO_DEC_SCATTER_GATHER,
@@ -1975,6 +1972,9 @@ enqueue_dec_one_op_cb(struct acc_queue
*q,
struct rte_bbdev_dec_op *op,
struct rte_mbuf *input, *h_output_head, *h_output,
*s_output_head, *s_output;
+ /* Disable explictly SO for VRB 1. */
+ op->turbo_dec.op_flags &= ~RTE_BBDEV_TURBO_SOFT_OUTPUT;
Can you explain why it is needed to filter this out?
I did not find a clear description in the bbdev API.
It would help if there were explicits references in doxygen of which
capability is necessary for using flags/API.
I was expecting that asking for RTE_BBDEV_TURBO_SOFT_OUTPUT to a
driver is only allowed if rte_bbdev_op_cap contains it.
With this assumption, it would be invalid for an application to
request RTE_BBDEV_TURBO_SOFT_OUTPUT through
rte_bbdev_enqueue_dec_ops.
You may arguably expect this from a well behaved user application
but still
there is nothing that enforces it explicitly, ie. notably under
negative scenario
conditions which we still need to manage gracefully.
If your application is buggy (not reading / complying with the device
capabilities), fix it.
Supporting negative scenario is within the scope of the PMD, whatever
the application throws at us in cannot cause any HW issue.
Fixing application issues is outside of DPDK control obviously.
I don't think it is not the role of the PMD to workaround application
bugs.
Of course I meant:
I don't think it is the role of the PMD to workaround application bugs.
The PMD driver reports capabilities for a given device variant. The
application ignores that and forces the capability, the PMD driver
should fail. It is better for the application the PMD driver let it know
it is misbehaving than trying to hide it.
Here we want to make sure that in case the optional operational flag is
included, we fall back to default mode when using the VRB1 variant.
Keep in mind that the unified driver can support multiple HW variant
(see
rest of the serie) and may support this option for other variants
using same
code.
Whatever the HW variant, the API should be respected: exposing
capabilities
is done on a per device basis.
It should be ideally, but in practice in case this is not done for
whatever reason (negative scenario, bug in user application)
then we want the PMD to still avoid misbehaving.
In term of documentation, I believe that capability/flag (ie. note
that the
enum maps to a capability when retrieved from info_get, and to an
operation
flag when provided to the bbdev api) is already captured explicitly
for many
generations. Basically this an optional output of the LTE decoding
processing,
to provide APP LLR which can be potentially be useful for the user
application
(separate optional mbuf). It may or may not be supported by a bb
device, and
it may or may not be requested to be provided through the API.
Typically this
is not enabled.
Being optional does not mean that a driver can ignore it.
Otherwise, there is no point in exposing a capability.
I am not sure I follow your concern. Capability are critical for
application to enumerate what the underlying device can do.
Here we are only stating that this is valuable to harden the PMD so
that it can operate even if an unexpected API is provided, notably to
guarantee the unified code is not used in an unintended manner.
Note that no PMD to my knowledge enforces checking explicitly the
op_flag matches with the capability (like a bitmask check),
and I don’t really think we have to, these other flags are just meant
to have effect since not supported.
Thanks.
--
David Marchand
