From: David Laight <david.laight.linux@gmail.com>
To: Wen Yang <wen.yang@linux.dev>
Cc: Joel Granados <joel.granados@kernel.org>,
Kees Cook <kees@kernel.org>,
linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [RFC PATCH v2 0/6] sysctl: replace overloaded extra pointers with typed bounds and core stride
Date: Tue, 17 Nov 2026 18:10:04 +0000 [thread overview]
Message-ID: <20261117181004.6d806dae@pumpkin> (raw)
In-Reply-To: <a66d0c11-f13b-42ec-8ddb-2cbddf50682b@linux.dev>
On Wed, 18 Nov 2026 01:01:46 +0800
Wen Yang <wen.yang@linux.dev> wrote:
> On 11/17/26 05:40, David Laight wrote:
> > On Tue, 17 Nov 2026 00:07:32 +0800
> > wen.yang@linux.dev wrote:
> >
> >> From: Wen Yang <wen.yang@linux.dev>
> >>
> >> struct ctl_table has long relied on extra1 and extra2 pointers to serve
> >> multiple unrelated duties, including range limits and caller-specific
> >> context. This series addresses both:
> >>
> >> 1. Replace bound pointers with plain unsigned long min and max fields
> >> directly in struct ctl_table.
> >>
> >> 2. Replace per-entry context pointers with a core stride mechanism,
> >> allowing callers to register arrays of structures embedding struct
> >> ctl_table and recover enclosing context via standard container_of().
> >>
> >> This RFC focuses on the core stride infrastructure and two representative
> >> complex users (netfilter nf_log and IPv4 devinet), as suggested by Joel
> >> Granados. The remaining treewide callers are converted in a single
> >> squashed commit to ensure the tree builds and tests pass cleanly.
> >
> > I think I would have tried to avoid the indentation changes - just to
> > make the actual changes easier to see.
> >
>
> Good point. Looking closer at Patch 4, keeping the original
> 'static struct devinet_sysctl_table { ... } devinet_sysctl' layout not
> only avoids re-indenting the 33 DEVINET_SYSCTL_* entries (cutting the
> diff from +104/-86 down to +41/-23), it also fixes a bug where
> kmemdup(&devinet_sysctl, ...) was copying a bare array into a
> struct devinet_sysctl_table that starts with sysctl_header.
>
> With the original layout preserved, the table definition change is just:
>
> static struct devinet_sysctl_table {
> struct ctl_table_header *sysctl_header;
> - struct ctl_table devinet_vars[IPV4_DEVCONF_MAX];
> + struct devinet_sysctl_entry devinet_vars[IPV4_DEVCONF_MAX];
> } devinet_sysctl = {
> .devinet_vars = {
> DEVINET_SYSCTL_COMPLEX_ENTRY(FORWARDING,
> ...
>
> > ...
> >> 4. Baseline rebase & conflict resolution:
> >> - Rebased onto next-20261111 (aaa2f954c21b).
> >> - Added proc_dointvec_jiffies_minmax and proc_dointvec_ms_jiffies_minmax
> >> to sysctl_check_table().
> >
> > I didn't spot those in here, but are they even needed?
> > Without the bounds don't the minmax variants behave like the original functions?
> > So it was only necessary to add support for bounds to the existing functions.
> > (The only think that half makes sense is if a caller uses extra1/2 for
> > something else.)
> >
>
> Sorry for the confusion: I omitted the squashed treewide patch (6/6)
> from the v2 posting to keep the mailing list traffic small, and that
> bullet in the cover letter referred to a change in that patch.
> I have moved the sysctl_check_table() addition into Patch 1 for v3.
>
> proc_dointvec_ms_jiffies_minmax() was already in the tree, and
> proc_dointvec_jiffies_minmax() was recently added in linux-next by
> commit 75f5722df2b4 ("net: restrict message_cost and message_burst to
> non-negative values"). You are right about the historical reason: the
> separate *_minmax handlers only existed because callers of
> proc_dointvec() and friends could stash unrelated context in
> extra1/extra2. Once extra1/extra2 are gone and bounds live in dedicated
> min/max fields, merging the *_minmax variants into the base handlers
> will be a clean follow-up.
You can probably also delete the 'vec' part of the names while doing
the global change.
A lot of the cases actually error any attempt to supply a vector.
I found an old commit saying there were only a handful of cases,
and most of those were strings.
So the (IIRC) maxlen field can't be deleted (yet), or repurposed to avoid
the need for separate u8/u16/u32/u64 functions.
>
> > Related, I think you can reduce the size of the final patch and the likely hood
> > of some silly errors by making the 'dointvec' functions treat min == 0, max == 0
> > (ie no initialisers) as 'no limit' rather than 'only zero allowed'.
> >
>
> There are two reasons why treating (min == 0 && max == 0) as "no limit"
> in the *_minmax handlers turns out to be problematic:
>
> 1. It conflicts with dropping redundant '.min = 0' on signed handlers,
> and does not help non-zero lower bounds:
>
> - Most proc_dointvec_minmax() callers in the tree only set
> '.extra1 = SYSCTL_ZERO' (with .extra2 = NULL) to reject negative
> inputs, i.e. [0, INT_MAX]. Once we drop the redundant '.min = 0'
> initialiser, omitting '.max = INT_MAX' as well leaves the entry at
> (min == 0, max == 0). If proc_dointvec_minmax() treats (0, 0) as
> "no limit" ([INT_MIN, INT_MAX]), all of those non-negative entries
> silently accept negative values again.
That is the one I missed.
> - For entries with a positive lower bound and no upper bound (e.g.
> '.extra1 = SYSCTL_ONE, .extra2 = NULL' -> '.min = 1'), min is non-zero,
> so a (min == 0 && max == 0) check does not trigger anyway; omitting
> .max leaves max == 0 < min and rejects all writes.
That empty range could easily be handled.
>
> 2. Implicitly defaulting an omitted .max to the type limit is fail-open
> on handlers whose whole purpose is range checking:
> Under the old "extra2 == NULL means no upper bound" rule, several
> *_minmax entries in net-next had bugs fixed just in the past two weeks
> because .extra2 was left unset:
>
> - 850a7225bc61 ("tcp: limit tcp_tso_rtt_log to a valid shift count"):
> tcp_tso_rtt_log used proc_dou8vec_minmax with neither .extra1 nor
> .extra2 set (so both defaulted to NULL -> [0, 255]), allowing shift
> counts >= 32 and triggering UBSAN shift-out-of-bounds in
> tcp_tso_autosize().
>
> - 0bc7b5b2985f ("ipv4: restrict igmp_qrv to 255") and
> 6d6bdc1b943c ("ipv6: restrict mld_qrv to 255"):
> both entries used proc_dointvec_minmax with '.extra1 = SYSCTL_ONE'
> and left .extra2 unset (NULL -> INT_MAX); writing 256 silently
> wrapped the backing u8 field to 0 and triggered
> WARN_ON(idev->mc_qrv == 0) in mld_query_work().
>
> With an explicit [min, max] closed interval, .min defaults to 0
> (matching non-negative inputs), while forgetting .max on a *_minmax
> entry leaves .max = 0 (fail-closed, immediately rejecting non-zero
> writes during testing) rather than silently allowing up to INT_MAX.
>
> To eliminate the risk of hand-written mistakes when converting existing
> one-sided entries, I will updat the coccinelle script (Patch 5) in v3 so
> that whenever an entry uses a *_minmax handler with only a lower bound
> (.extra2 absent), the script automatically supplies the matching type
> maximum (INT_MAX/UINT_MAX/ULONG_MAX/U8_MAX) according to .proc_handler.
In some sense that hides the fact that possibly no one had thought
about a sane upper bound.
(The same is actually true if there is no lower bound, even zero may
not actually be a sane value in many cases.)
Possibly you need a different identifier for the implicit bounds rather
than any explicit bounds that someone has actually set.
So you might use [DEFAULT_ZERO .. DEFAULT_UINT_MAX] to differentiate from
an explicit [0 .. UINT_MAX] that you might hope someone that thought about.
I think I'd definitely do that if removing the 'minmax' functions and
just supplying a non-restrictive range.
Perhaps you should add a sample file updated by the script?
All of them is definitely OTT.
David
>
>
> Thanks,
> Wen
>
prev parent reply other threads:[~2026-11-17 18:10 UTC|newest]
Thread overview: 9+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-11-16 16:07 [RFC PATCH v2 0/6] sysctl: replace overloaded extra pointers with typed bounds and core stride wen.yang
2026-11-16 16:07 ` [RFC PATCH v2 1/5] sysctl: add numeric bounds and core stride infrastructure wen.yang
2026-11-16 16:07 ` [RFC PATCH v2 2/5] sysctl: cover numeric bounds and table containers in KUnit wen.yang
2026-11-16 16:07 ` [RFC PATCH v2 3/5] netfilter: embed logger sysctl context using stride wen.yang
2026-11-16 16:07 ` [RFC PATCH v2 4/5] net/ipv4: embed device " wen.yang
2026-11-16 16:07 ` [RFC PATCH v2 5/5] scripts/coccinelle: add script to detect and convert sysctl bounds wen.yang
2026-11-16 21:40 ` [RFC PATCH v2 0/6] sysctl: replace overloaded extra pointers with typed bounds and core stride David Laight
2026-11-17 17:01 ` Wen Yang
2026-11-17 18:10 ` David Laight [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20261117181004.6d806dae@pumpkin \
--to=david.laight.linux@gmail.com \
--cc=joel.granados@kernel.org \
--cc=kees@kernel.org \
--cc=linux-fsdevel@vger.kernel.org \
--cc=linux-kernel@vger.kernel.org \
--cc=wen.yang@linux.dev \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox