Skip to content

softirq: Per vector masking v2

Frederic Weisbecker lkml 1 version In review

Get series with b4
b4 am [email protected]

Versions newest first

  1. v1 Latest32 patches 18 repliesIn review

Patches in v1 No applies recorded yet

#Subject and filesSizeRepliesApplied
1/32locking/lockdep: Use expanded masks on find_usage_*() functionskernel/locking/lockdep.c+15 -151 reply
2/32locking/lockdep: Introduce struct lock_usagekernel/locking/lockdep.c kernel/locking/lockdep_internals.h+79 -482 replies
3/32locking/lockdep: Convert usage_mask to u64include/linux/lockdep.h kernel/locking/lockdep.c+5 -52 replies
4/32locking/lockdep: Test all incompatible scenario at once in check_irq_usage()kernel/locking/lockdep.c+137 -670 replies
5/32locking/lockdep: Prepare valid_state() to handle plain maskskernel/locking/lockdep.c+39 -54 replies
6/32locking/lockdep: Prepare check_usage_*() to handle plain maskskernel/locking/lockdep.c+24 -60 replies
7/32locking/lockdep: Prepare state_verbose() to handle all softirqskernel/locking/lockdep.c+4 -80 replies
8/32locking/lockdep: Make mark_lock() fastpath to work with multiple usage at oncekernel/locking/lockdep.c+2 -20 replies
9/32locking/lockdep: Save stack trace for each softirq vector involvedinclude/linux/lockdep.h kernel/locking/lockdep.c+19 -22 replies
10/32locking/lockdep: Make mark_lock() verbosity aware of vectorkernel/locking/lockdep.c+1 -10 replies
11/32softirq: Macrofy softirq vectorsinclude/linux/interrupt.h include/linux/softirq_vector.h+14 -132 replies
12/32locking/lockdep: Define per vector softirq lock usage stateskernel/locking/lockdep_internals.h kernel/locking/lockdep_states.h+42 -50 replies
13/32softirq: Pass softirq vector number to lockdep on vector executioninclude/linux/irqflags.h kernel/softirq.c lib/locking-selftest.c+12 -120 replies
14/32x86: Revert "x86/irq: Demote irq_cpustat_t::__softirq_pending to u16"arch/x86/include/asm/hardirq.h+1 -10 replies
15/32arch/softirq: Rename softirq_pending fields to softirq_dataarch/arm/include/asm/hardirq.h arch/arm64/include/asm/hardirq.h arch/h8300/kernel/asm-offsets.c +14 more+24 -240 replies
16/32softirq: Normalize softirq_pending naming schemearch/s390/include/asm/hardirq.h include/linux/interrupt.h kernel/softirq.c+16 -160 replies
17/32softirq: Convert softirq_pending_*() to set/clear mask schemearch/s390/include/asm/hardirq.h include/linux/interrupt.h kernel/softirq.c+7 -60 replies
18/32softirq: Introduce disabled softirq vectors bitsarch/s390/include/asm/hardirq.h include/linux/interrupt.h+56 -60 replies
19/32softirq: Rename _local_bh_enable() to local_bh_enable_no_softirq()arch/s390/lib/delay.c drivers/s390/char/sclp.c drivers/s390/cio/cio.c +2 more+10 -100 replies
20/32softirq: Move vectors bits to bottom_half.hinclude/linux/bottom_half.h include/linux/interrupt.h+24 -210 replies
21/32x86: Init softirq enabled fieldarch/x86/kernel/irq.c include/linux/bottom_half.h+6 -10 replies
22/32softirq: Check enabled vectors before processinginclude/linux/interrupt.h kernel/softirq.c+16 -100 replies
23/32softirq: Remove stale commentkernel/softirq.c+0 -42 replies
24/32softirq: Uninline !CONFIG_TRACE_IRQFLAGS __local_bh_disable_ip()include/linux/bottom_half.h kernel/softirq.c+6 -142 replies
25/32softirq: Prepare for mixing all/per-vector maskingkernel/softirq.c+22 -110 replies
26/32softirq: Support per vector maskinginclude/linux/bottom_half.h kernel/softirq.c+80 -120 replies
27/32locking/lockdep: Remove redundant softirqs on checkkernel/locking/lockdep.c kernel/locking/lockdep_internals.h kernel/locking/lockdep_proc.c+1 -70 replies
28/32locking/lockdep: Update check_flags() according to new layoutkernel/locking/lockdep.c+1 -60 replies
29/32locking/lockdep: Branch the new vec-finegrained softirq masking to lockdepkernel/locking/lockdep.c+11 -10 replies
30/32softirq: Allow to soft interrupt vector-specific masked contextskernel/softirq.c+5 -50 replies
31/32locking: Introduce spin_[un]lock_bh_mask()include/linux/spinlock.h include/linux/spinlock_api_smp.h include/linux/spinlock_api_up.h +1 more+72 -00 replies
32/32net: Make softirq vector masking finegrained on release_sock()net/core/sock.c+4 -20 replies

Reviews trailers per patch

None of the 32 patches has a Reviewed-by or Acked-by of their own; the series has one for all of them.

PatchDavid S. Miller
SeriesR-b
1/32
2/32
3/32
4/32
5/32
6/32
7/32
8/32
9/32
10/32
11/32
12/32
13/32
14/32
15/32
16/32
17/32
18/32
19/32
20/32
21/32
22/32
23/32
24/32
25/32
26/32
27/32
28/32
29/32
30/32
31/32
32/32

Trailers given in replies are filled; those carried in the patches from earlier versions are outlined with dashes.

Read the v1 thread

Cover letter [PATCH 00/32] softirq: Per vector masking v2

For those who missed the infinitely invasive and carpal tunnel unfriendly v1: https://lwn.net/Articles/768157/

Softirqs masking is an all or nothing operation. It's currently not possible to disable a single vector. Yet some workloads are interested in deterministic latencies for vectors execution. Reducing their interdependencies is a first step toward that.

Unlike the previous take, the current APIs aren't changed, as advised by reviewers. But new APIs have been introduced. An individual vector can be disabled and that behaviour can self-nest and nest with existing APIs:

    bh = local_bh_disable_mask(BIT(TASKLET_SOFTIRQ));
	    bh2 = spin_lock_bh_mask(lock, BIT(NET_RX_SOFTIRQ));
	        local_bh_disable();
            [...]
            local_bh_enable();
        spin_unlock_bh_mask(lock, bh2);
    local_bh_enable_mask(bh));

Also mandatory, the new version provides lockdep validation in a per vector finegrained way.

The next step could be to allow soft-interrupting softirq vectors with other vectors. We'll need to be careful about stack usage and interdependencies though. But that could solve issues with long lasting vectors running at the expense of others.

A few details need improvement:

* We need to set all vectors of local_softirq_enabled() on boot for all archs. Only x86 does it for now. Shouldn't be too hard to achieve though.

* Handle multiple usage on lockdep verbose debugging (see patch PATCH 10/32) Also easy to fix.

* Restore redundant softirqs on tracking on softirq. See "locking/lockdep: Remove redundant softirqs on check". Also shouldn't be too hard to fix.

git://git.kernel.org/pub/scm/linux/kernel/git/frederic/linux-dynticks.git
softirq/soft-interruptible

HEAD: 0d35ad1a4b13b62e135672dfe86d362b49f41abf

Thanks,
Frederic
---

Frederic Weisbecker (32):
locking/lockdep: Use expanded masks on find_usage_*() functions locking/lockdep: Introduce struct lock_usage
locking/lockdep: Convert usage_mask to u64
locking/lockdep: Test all incompatible scenario at once in check_irq_usage() locking/lockdep: Prepare valid_state() to handle plain masks locking/lockdep: Prepare check_usage_*() to handle plain masks locking/lockdep: Prepare state_verbose() to handle all softirqs locking/lockdep: Make mark_lock() fastpath to work with multiple usage at once locking/lockdep: Save stack trace for each softirq vector involved locking/lockdep: Make mark_lock() verbosity aware of vector
softirq: Macrofy softirq vectors
locking/lockdep: Define per vector softirq lock usage states
softirq: Pass softirq vector number to lockdep on vector execution
x86: Revert "x86/irq: Demote irq_cpustat_t::__softirq_pending to u16" arch/softirq: Rename softirq_pending fields to softirq_data
softirq: Normalize softirq_pending naming scheme
softirq: Convert softirq_pending_*() to set/clear mask scheme
softirq: Introduce disabled softirq vectors bits
softirq: Rename _local_bh_enable() to local_bh_enable_no_softirq()
softirq: Move vectors bits to bottom_half.h
x86: Init softirq enabled field
softirq: Check enabled vectors before processing
softirq: Remove stale comment
softirq: Uninline !CONFIG_TRACE_IRQFLAGS __local_bh_disable_ip()
softirq: Prepare for mixing all/per-vector masking
softirq: Support per vector masking
locking/lockdep: Remove redundant softirqs on check locking/lockdep: Update check_flags() according to new layout locking/lockdep: Branch the new vec-finegrained softirq masking to lockdep
softirq: Allow to soft interrupt vector-specific masked contexts
locking: Introduce spin_[un]lock_bh_mask()
net: Make softirq vector masking finegrained on release_sock()

 arch/arm/include/asm/hardirq.h      |   2 +-
 arch/arm64/include/asm/hardirq.h    |   2 +-
 arch/h8300/kernel/asm-offsets.c     |   2 +-
 arch/ia64/include/asm/hardirq.h     |   2 +-
 arch/ia64/include/asm/processor.h   |   2 +-
 arch/m68k/include/asm/hardirq.h     |   2 +-
 arch/m68k/kernel/asm-offsets.c      |   2 +-
 arch/parisc/include/asm/hardirq.h   |   2 +-
 arch/powerpc/include/asm/hardirq.h  |   2 +-
 arch/s390/include/asm/hardirq.h     |  11 +-
 arch/s390/lib/delay.c               |   2 +-
 arch/sh/include/asm/hardirq.h       |   2 +-
 arch/sparc/include/asm/cpudata_64.h |   2 +-
 arch/sparc/include/asm/hardirq_64.h |   4 +-
 arch/um/include/asm/hardirq.h       |   2 +-
 arch/x86/include/asm/hardirq.h      |   2 +-
 arch/x86/kernel/irq.c               |   5 +-
 drivers/s390/char/sclp.c            |   2 +-
 drivers/s390/cio/cio.c              |   2 +-
 include/asm-generic/hardirq.h       |   2 +-
 include/linux/bottom_half.h         |  41 +++-
 include/linux/interrupt.h           |  87 ++++---
 include/linux/irqflags.h            |  12 +-
 include/linux/lockdep.h             |   5 +-
 include/linux/softirq_vector.h      |  10 +
 include/linux/spinlock.h            |  14 ++
 include/linux/spinlock_api_smp.h    |  26 ++
 include/linux/spinlock_api_up.h     |  13 +
 kernel/locking/lockdep.c            | 465 ++++++++++++++++++++++++------------
 kernel/locking/lockdep_internals.h  |  50 +++-
 kernel/locking/lockdep_proc.c       |   2 +-
 kernel/locking/lockdep_states.h     |   4 +-
 kernel/locking/spinlock.c           |  19 ++
 kernel/softirq.c                    | 159 ++++++++----
 lib/locking-selftest.c              |   4 +-
 net/core/sock.c                     |   6 +-
 36 files changed, 690 insertions(+), 281 deletions(-)