Reclaim reserved buffer for unused ports

[stephenxs]

201911 PRs

PR title	state	context
Upgrade Mellanox-SAI to 1.19.3 to support reclaiming reserved buffer on admin down ports
[Buffer Manager] Reclaim unused buffer for admin-down ports
[Buffer Orch] Handle DEL_COMMAND for BUFFER_PG and BUFFER_QUEUE table
Regression test case for traditional buffer manager in 201911
[201911] [Submodule] Update sonic-swss
Verify buffer priority groups for all platforms
[201911] Verify buffer priority groups for all platforms

Master PRs (currently draft)

PR title	state	context
[Reclaim buffer] Reclaim unused buffers by applying zero buffer profiles
[Reclaim buffer] Common infrastructure update for reclaiming buffer
[Reclaim buffer] Reclaim unused buffer for dynamic buffer model
[Reclaim buffer] Reclaim unused buffer for traditional buffer model
[Reclaim buffer] Common code update for swss
[Reclaim buffer] [Mellanox] Db migrator support reclaiming reserved buffer for unused ports
[Reclaiming buffer] Test cases for reclaiming buffer on both traditional and dynamic model

202106 PRs

PR title	state	context
[Reclaim buffer][202106] Reclaim unused buffers by applying zero buffer profiles
[Reclaim buffer][202106] Reclaim unused buffer for dynamic buffer model
[Reclaim buffer][202106] Reclaim unused buffer in traditional buffer model
[Reclaim buffer][202106] Common code update
[Reclaim buffer][202106] Db migrator support reclaiming reserved buffer for unused ports

202012 PRs

PR title	state	context
[Reclaim buffer][202012] Reclaim unused buffers by applying zero buffer profiles
[Reclaim buffer][202012] Reclaim unused buffer for both traditional and dynamic buffer model
[Reclaim buffer][202012] Reclaim unused buffer in traditional buffer model
[Reclaim buffer][202012] Common code update
[Reclaim buffer][202012] Db migrator support reclaiming reserved buffer for unused ports

Reclaim reserved buffer for unused ports.
Originally, the reserved buffer is reclaimed by removing buffer objects of the unused ports. However, this introduces inconsistency. Consider the following scenarios:

1. System starts without any buffer configuration. The buffer configuration in the ASIC is the SDK default value whose reserved size is not 0. This is to guarantee the traffic can be forwarded without buffer configuration in SONiC.
2. After the system started and the buffer was configured, some buffer objects are removed from ports. There is still no buffer configuration in SONiC. However, in the ASIC the reserved size of some of the buffer objects is 0.

Now we have two scenarios with the same SONiC buffer configuration, resulting in different buffer configurations in ASIC.

To resolve this issue:

• zero buffer profiles are introduced to indicate 0 reserved size of a buffer object.
• Removing a buffer object indicates setting the buffer object to SDK default value.

The solution:

• For lossy PG,
  • SONiC to create
    • a buffer pool with static threshold mode and zero size
    • a zero buffer profile with static threshold == 0, size == 0, referencing the pool created in above step
  • SONiC to configure the zero buffer profile on the unused port
  • SAI to set the reserved buffer size to 0 if a zero buffer profile is configured
• For lossless PG, to remove the PG from SAI
• For port buffer pool and queues,
  • SONiC to create a zero buffer profile with size == 0 and dynamic mode
  • SAI to set the reserved buffer size to 0 (no change on SAI)

The zero profiles and pools will be handled in different ways between the dynamic and traditional buffer models.

• Traditional buffer model:
• The zero profiles and pools will be configured into CONFIG_DB if there is at least one inactive port.
• The zero profiles will be configured to corresponding buffer items of inactive ports, like priority groups, queues, ingress/egress profile lists.
• Dynamic buffer model:
  • The normal profiles will always be configured to corresponding buffer items of all ports, including active ones and inactive ones.
  • The buffer manager will insert the zero profiles and pools into APPL_DB when the first port becomes admin down
  • The buffer manager will apply the zero profiles to corresponding buffer items of inactive ports into APPL_DB

[zhangyanzhao]

@neethajohn would you please help to approve this HLD PR? Thanks.

[stephenxs]

Suggest merging order:

1. Swss common PR ([Reclaiming buffer] Common code update sonic-swss#1996) and then submodule update PR which will be opened once the swss PR has been merged
2. Build image common PR (#9133)
3. Swss dynamic mode PR ([Reclaim buffer] Reclaim unused buffer for dynamic buffer model sonic-swss#1910)
4. Db migrator PR ([Reclaim buffer] [Mellanox] Db migrator support reclaiming reserved buffer for unused ports sonic-utilities#1822)
5. Submodule update PR which will be opened after swss/utilities PRs have been merged.
6. Build image Mellanox specific PR ([Reclaim buffer] Reclaim unused buffers by applying zero buffer profiles sonic-buildimage#8768)

The swss static mode PR (sonic-net/sonic-swss#2011) doesn't depend on any PR and can be merged at any time.

Db migrator PR (sonic-net/sonic-utilities#1822) and build image Mellanox specific PR (sonic-net/sonic-buildimage#8768) can also be merged before swss dynamic mode PR (sonic-net/sonic-swss#1910).

[liat-grozovik]

@neethajohn could you please help to review?
Code changes are already in place and are in review. Better to have the HLD approved before all code is in

[liat-grozovik]

@neethajohn kindly reminder, this PR needs to get it as most of the code is already merged.