The legacy printer kthread uses console_lock and
__console_flush_and_unlock to flush records to the console which
holds the console_lock being held for the entire flush. This
results in large waiting times for console_lock waiters
especially where there is a large volume of records or where the
console is slow (e.g. serial). During boot, this contention causes
delays in the filp_open call in console_on_rootfs.
   
Let's instead release and reacquire console_lock in between
printing individual records.

Signed-off-by: Andrew Murray <amurray@thegoodpenguin.co.uk>
---
Changes in v3:
- Rebased on v6.18-rc2
- Reordering of variables and minor tweak to comments in console_flush_one_record
  in first patch
- Improve semantics of console_flush_one_record in the second patch through different
  use of return value and by replacing any_usable with try_again.
- Update third patch to use newer version of console_flush_one_record
- Link to v2: https://lore.kernel.org/r/20250927-printk_legacy_thread_console_lock-v2-0-cff9f063071a@thegoodpenguin.co.uk

Changes in v2:
- Move any_usable=false to console_flush_all in the 'introduce
  console_flush_one_record' patch to match original implementation.
- Add Petr's console_flush_one_record() code cleanup patch
- Open code flushing implementation in legacy_kthread_func instead
  of introducing new console_flush functions.
- Link to v1: https://lore.kernel.org/r/20250915-printk_legacy_thread_console_lock-v1-0-f34d42a9bcb3@thegoodpenguin.co.uk

---
Andrew Murray (2):
      printk: Introduce console_flush_one_record
      printk: Use console_flush_one_record for legacy printer kthread

Petr Mladek (1):
      printk: console_flush_one_record() code cleanup

 kernel/printk/printk.c | 186 +++++++++++++++++++++++++++++++------------------
 1 file changed, 119 insertions(+), 67 deletions(-)
---
base-commit: 211ddde0823f1442e4ad052a2f30f050145ccada
change-id: 20250914-printk_legacy_thread_console_lock-1c27f59bf990

Best regards,
-- 
Andrew Murray <amurray@thegoodpenguin.co.uk>

On Mon 2025-10-20 16:38:04, Andrew Murray wrote:
> The legacy printer kthread uses console_lock and
> __console_flush_and_unlock to flush records to the console which
> holds the console_lock being held for the entire flush. This
> results in large waiting times for console_lock waiters
> especially where there is a large volume of records or where the
> console is slow (e.g. serial). During boot, this contention causes
> delays in the filp_open call in console_on_rootfs.
>    
> Let's instead release and reacquire console_lock in between
> printing individual records.
> 
> Signed-off-by: Andrew Murray <amurray@thegoodpenguin.co.uk>
> ---
> Changes in v3:
> - Rebased on v6.18-rc2
> - Reordering of variables and minor tweak to comments in console_flush_one_record
>   in first patch
> - Improve semantics of console_flush_one_record in the second patch through different
>   use of return value and by replacing any_usable with try_again.
> - Update third patch to use newer version of console_flush_one_record
> - Link to v2: https://lore.kernel.org/r/20250927-printk_legacy_thread_console_lock-v2-0-cff9f063071a@thegoodpenguin.co.uk
> 
> Changes in v2:
> - Move any_usable=false to console_flush_all in the 'introduce
>   console_flush_one_record' patch to match original implementation.
> - Add Petr's console_flush_one_record() code cleanup patch
> - Open code flushing implementation in legacy_kthread_func instead
>   of introducing new console_flush functions.
> - Link to v1: https://lore.kernel.org/r/20250915-printk_legacy_thread_console_lock-v1-0-f34d42a9bcb3@thegoodpenguin.co.uk
> 
> ---
> Andrew Murray (2):
>       printk: Introduce console_flush_one_record
>       printk: Use console_flush_one_record for legacy printer kthread
> 
> Petr Mladek (1):
>       printk: console_flush_one_record() code cleanup
> 
>  kernel/printk/printk.c | 186 +++++++++++++++++++++++++++++++------------------
>  1 file changed, 119 insertions(+), 67 deletions(-)
> ---
> base-commit: 211ddde0823f1442e4ad052a2f30f050145ccada
> change-id: 20250914-printk_legacy_thread_console_lock-1c27f59bf990

JFYI, the patchset has been committed into printk/linux.git,
branch rework/preempt-legacy-kthread. It is intended for 6.19.

Best Regards,
Petr