kernel/sysctl.c | 237 ++++++++++++++++++++----------- tools/testing/selftests/sysctl/sysctl.sh | 60 ++++++++ 2 files changed, 217 insertions(+), 80 deletions(-)
What?
=====
Stage table->data when writing INT & ULONG ctl_tables. Commit staged
data to kernel variable only when all the conversions have succeeded.
This is applicable only to variables that represent a vector; paths
pertaining to scalers are left unchanged. Notice that partial updates
can still happen if less than the size of the vector are passed and
correctly formatted.
This is the behavior we are protecting against:
# echo "4 4 1 7" > /proc/sys/kernel/printk
# echo "1 x" > /proc/sys/kernel/printk
-bash: echo: write error: Invalid argument
# cat /proc/sys/kernel/printk
1 4 1 7 <- the write failed, the first element changed anyway
Why?
====
Allowing a partial change to a vector after returning from an erroneous
proc_handler is just plain wrong. This should be handled within sysctl
to avoid users having to do it for themselves [1].
Behavioral Changes
==================
1. A failed write will not update the vector
2. Vector writes can now fail with -ENOMEM
Testing
=======
This went through regular sysctl kunit and self test. Also is posted to
0-day.
I always find it difficult to know who wants to receive this. Please let
me know if you want to be removed from the Cc/To.
Best
[1] https://lore.kernel.org/all/tencent_A860C873956A52E26AD8D309A308A241BA08@qq.com/
Signed-off-by: Joel Granados <joel.granados@kernel.org>
---
Changes in v2:
- Avoid doc compilation warnings related to apply_conv_on_vec function
- Initialize rc in testing to avoid a false positive when executing 0013
individually
- Link to v1: https://lore.kernel.org/r/20260813-lklm-partial_ctlvec-v1-0-df9e51c13704@kernel.org
---
Joel Granados (4):
sysctl: Split data conversion and file position handling
sysctl: Reject uint arrays before calling the general proc_vec
sysctl: Disallow partial updates for erroneous sysctl vectors
sysctl: Add 0013 to test partially updated vectors
kernel/sysctl.c | 237 ++++++++++++++++++++-----------
tools/testing/selftests/sysctl/sysctl.sh | 60 ++++++++
2 files changed, 217 insertions(+), 80 deletions(-)
---
base-commit: 4c12287001da60f3c022bb25932aa7a5590fc0b1
change-id: 20260813-lklm-partial_ctlvec-bd8867b70d5c
Best regards,
--
Joel Granados <joel.granados@kernel.org>
On Fri, Aug 14, 2026 at 12:41:11PM +0200, Joel Granados wrote: > What? > ===== > > Stage table->data when writing INT & ULONG ctl_tables. Commit staged > data to kernel variable only when all the conversions have succeeded. > This is applicable only to variables that represent a vector; paths > pertaining to scalers are left unchanged. Notice that partial updates > can still happen if less than the size of the vector are passed and > correctly formatted. ... snip ... > sysctl: Split data conversion and file position handling > sysctl: Reject uint arrays before calling the general proc_vec > sysctl: Disallow partial updates for erroneous sysctl vectors > sysctl: Add 0013 to test partially updated vectors > > kernel/sysctl.c | 237 ++++++++++++++++++++----------- > tools/testing/selftests/sysctl/sysctl.sh | 60 ++++++++ > 2 files changed, 217 insertions(+), 80 deletions(-) > --- > base-commit: 4c12287001da60f3c022bb25932aa7a5590fc0b1 > change-id: 20260813-lklm-partial_ctlvec-bd8867b70d5c > > Best regards, > -- > Joel Granados <joel.granados@kernel.org> > > I feel the superficial changes mentioned in the review (up until now) do not merit me sending a V3. The modified tree is here [1]. Best [1] https://git.kernel.org/pub/scm/linux/kernel/git/joel.granados/linux.git/log/?h=jag/partial_ctlvec
Copy it, only write back once every value converted. That's the right way to do this. Went through the whole series against the tree, notes in the patch replies. Some nits on the series: - "miss-formatted" in the subject. mis-formatted, or just malformed. - akpm's scalars nit still applies. - Sashiko saw something on this one (akpm linked it). Otherwise fine by me, the split is clean. Reviewed-by: Bradley Morgan <include@grrlz.net> also Tested-by: Bradley Morgan <include@grrlz.net> for the test Thanks!
On Sat, Aug 15, 2026 at 05:09:30PM +0100, Bradley Morgan wrote: > > Copy it, only write back once every value converted. That's the > right way to do this. Went through the whole series against the > tree, notes in the patch replies. Your feedback was useful, but it was riddled with a *lot* of hand waving. Please be succinct and to the point. The previous paragraph is not needed. > > Some nits on the series: > - "miss-formatted" in the subject. mis-formatted, or just malformed. Note that this comment is missing the why. It is good to know why you want the change > - akpm's scalars nit still applies. > - Sashiko saw something on this one (akpm linked it). In my opinion this is just noise. If you see that I have missed a comment, call me out on it. If you have additional information, answer to reviewers mail. > > Otherwise fine by me, the split is clean. > > > Reviewed-by: Bradley Morgan <include@grrlz.net> > also > > Tested-by: Bradley Morgan <include@grrlz.net> for the test How did you test? And what where your results? Best
On 18 August 2026 13:46:58 BST, Joel Granados <joel.granados@kernel.org> wrote: >On Sat, Aug 15, 2026 at 05:09:30PM +0100, Bradley Morgan wrote: >> >> Copy it, only write back once every value converted. That's the >> right way to do this. Went through the whole series against the >> tree, notes in the patch replies. >Your feedback was useful, but it was riddled with a *lot* of hand >waving. Please be succinct and to the point. The previous paragraph is >not needed. > >> >> Some nits on the series: >> - "miss-formatted" in the subject. mis-formatted, or just malformed. >Note that this comment is missing the why. It is good to know why you >want the change > >> - akpm's scalars nit still applies. >> - Sashiko saw something on this one (akpm linked it). >In my opinion this is just noise. If you see that I have missed a >comment, call me out on it. If you have additional information, answer >to reviewers mail. > >> >> Otherwise fine by me, the split is clean. >> >> >> Reviewed-by: Bradley Morgan <include@grrlz.net> >> also >> >> Tested-by: Bradley Morgan <include@grrlz.net> for the test >How did you test? And what where your results? > Built and booted the latest kernel. Its the summer holidays, I can do all this, it's magical >Best > Thanks!
On Fri, 14 Aug 2026 12:41:11 +0200 Joel Granados <joel.granados@kernel.org> wrote: > What? > ===== > > Stage table->data when writing INT & ULONG ctl_tables. Commit staged > data to kernel variable only when all the conversions have succeeded. > This is applicable only to variables that represent a vector; paths > pertaining to scalers are left unchanged. Notice that partial updates "scalars" > can still happen if less than the size of the vector are passed and > correctly formatted. > > This is the behavior we are protecting against: > > # echo "4 4 1 7" > /proc/sys/kernel/printk > # echo "1 x" > /proc/sys/kernel/printk > -bash: echo: write error: Invalid argument > # cat /proc/sys/kernel/printk > 1 4 1 7 <- the write failed, the first element changed anyway I like the change. > Why? > ==== > Allowing a partial change to a vector after returning from an erroneous > proc_handler is just plain wrong. This should be handled within sysctl > to avoid users having to do it for themselves [1]. > > Behavioral Changes > ================== > 1. A failed write will not update the vector > 2. Vector writes can now fail with -ENOMEM Yeah, this improves Linux. > Testing > ======= > This went through regular sysctl kunit and self test. Also is posted to > 0-day. > > I always find it difficult to know who wants to receive this. Please let > me know if you want to be removed from the Cc/To. Sashiko said one thing: https://sashiko.dev/#/patchset/20260814-lklm-partial_ctlvec-v2-0-9df50d26e477@kernel.org
On Fri, Aug 14, 2026 at 12:36:02PM -0700, Andrew Morton wrote: > On Fri, 14 Aug 2026 12:41:11 +0200 Joel Granados <joel.granados@kernel.org> wrote: > > > What? > > ===== > > > > Stage table->data when writing INT & ULONG ctl_tables. Commit staged > > data to kernel variable only when all the conversions have succeeded. > > This is applicable only to variables that represent a vector; paths > > pertaining to scalers are left unchanged. Notice that partial updates > > "scalars" Will change in the cover letter (If/when I send the next version). It wont actually show up in upstream as this is just in the cover letter. > > > can still happen if less than the size of the vector are passed and > > correctly formatted. > > > > This is the behavior we are protecting against: > > > > # echo "4 4 1 7" > /proc/sys/kernel/printk > > # echo "1 x" > /proc/sys/kernel/printk > > -bash: echo: write error: Invalid argument > > # cat /proc/sys/kernel/printk > > 1 4 1 7 <- the write failed, the first element changed anyway > > I like the change. > > > Why? > > ==== > > Allowing a partial change to a vector after returning from an erroneous > > proc_handler is just plain wrong. This should be handled within sysctl > > to avoid users having to do it for themselves [1]. > > > > Behavioral Changes > > ================== > > 1. A failed write will not update the vector > > 2. Vector writes can now fail with -ENOMEM > > Yeah, this improves Linux. > > > Testing > > ======= > > This went through regular sysctl kunit and self test. Also is posted to > > 0-day. > > > > I always find it difficult to know who wants to receive this. Please let > > me know if you want to be removed from the Cc/To. > > Sashiko said one thing: > https://sashiko.dev/#/patchset/20260814-lklm-partial_ctlvec-v2-0-9df50d26e477@kernel.org The tests that keep on giving :). I'm already working on a replacement for the sysctl selftests and prefer to leave these as they are and address them in the new version if they persist. Thx for the feedback Best Joel
© 2016 - 2026 Red Hat, Inc.