[PATCH v2 0/2] scsi: ufs: batch device-init reads with UFS 5.0 aggregated read

[PATCH v2 0/2] scsi: ufs: batch device-init reads with UFS 5.0 aggregated read

[Hyeoncheol Jeong]

UFS 5.0 adds the AGGREGATED READ query (opcode 0x9): one QUERY RESPONSE
UPIU returns many descriptors, attributes and flags as a chain of typed
groups.

Patch 1 adds the opcode, the request mask, the group header and the reply
plumbing. Patch 2 uses it during device init: a single aggregated read
replaces the per-item queries for the attributes, flags and ID strings,
falling back to individual queries on older devices or a rejected opcode.

No functional change for pre-5.0 devices.

Motivation: on a Qualcomm SM8850 platform with a Samsung UFS 5.0 device,
timing raw UPIU queries over BSG (100 iterations) shows that replacing the
device-init read burst with a single aggregated read reduces the query
time by 75-78% (min 75.3%, avg 76.4%, max 78.5%).

(That setup batched 8 init reads; on current mainline this series batches
the 6 device-init reads that remain individual queries.)

Changes since v1:
- Rename UFS_AGG_TYPE_UNIT_DESC to UFS_AGG_TYPE_UNIT_RPMB_DESC (Peter Wang)
- Add UFS_AGG_GROUP_RESERVED = 0x00 (Peter Wang)
- Rename hdr_off/next_off to group_off/next_group_off (Peter Wang)
- Rebased onto v7.3-rc5

Hyeoncheol Jeong (2):
  scsi: ufs: Add the aggregated read query opcode and its reply format
  scsi: ufs: Serve the device init reads from one aggregated read

 drivers/ufs/core/ufshcd.c | 364 +++++++++++++++++++++++++++++++++++---
 include/ufs/ufs.h         |  47 +++++
 include/ufs/ufshcd.h      |  11 ++
 3 files changed, 402 insertions(+), 20 deletions(-)

--
2.25.1

Re: [PATCH v2 0/2] scsi: ufs: batch device-init reads with UFS 5.0 aggregated read

[Bean Huo]

On Wed, 2026-09-30 at 14:02 +0900, Hyeoncheol Jeong wrote:
> No functional change for pre-5.0 devices.
> 
> Motivation: on a Qualcomm SM8850 platform with a Samsung UFS 5.0 device,
> timing raw UPIU queries over BSG (100 iterations) shows that replacing the
> device-init read burst with a single aggregated read reduces the query
> time by 75-78% (min 75.3%, avg 76.4%, max 78.5%).


Hi Jeong,

I don't know if you have plan to enable this feature in UEFI or any bootloaders
to eliminate the boot overhead becaused of query iterations.

Kind regards, 
Bean

Re: [PATCH v2 0/2] scsi: ufs: batch device-init reads with UFS 5.0 aggregated read

[Hyeoncheol Jeong]

Hi Bean,

On Fri, 2 Oct 2026 11:24:15 +0200, Bean Huo wrote:
> On Wed, 2026-09-30 at 14:02 +0900, Hyeoncheol Jeong wrote:
> > Motivation: on a Qualcomm SM8850 platform with a Samsung UFS 5.0 device,
> > timing raw UPIU queries over BSG (100 iterations) shows that replacing the
> > device-init read burst with a single aggregated read reduces the query
> > time by 75-78% (min 75.3%, avg 76.4%, max 78.5%).
>
> I don't know if you have plan to enable this feature in UEFI or any bootloaders
> to eliminate the boot overhead becaused of query iterations.

For now, my focus is on the kernel side. I do plan to bring the same
approach to bootloaders such as EDK II as a follow-up.

Thanks,
Hyeoncheol