* [PATCH net-next v3 1/3] net: dsa: microchip: ksz9477: apply a shorter ageing time at once
2026-11-17 13:02 [PATCH net-next v3 0/3] net: dsa: make a shorter bridge ageing time take effect Oleksij Rempel
@ 2026-11-17 13:02 ` Oleksij Rempel
2026-11-19 1:03 ` netdev-bot+sashiko
2026-11-17 13:02 ` [PATCH net-next v3 2/3] net: dsa: ignore the old ageing time of the port being changed Oleksij Rempel
2026-11-17 13:02 ` [PATCH net-next v3 3/3] net: dsa: forget the port ageing time when the port leaves the bridge Oleksij Rempel
2 siblings, 1 reply; 9+ messages in thread
From: Oleksij Rempel @ 2026-11-17 13:02 UTC (permalink / raw)
To: Woojung Huh, UNGLinuxDriver, Andrew Lunn, Vladimir Oltean,
David S. Miller, Eric Dumazet, Jakub Kicinski, Paolo Abeni,
Simon Horman, Arun Ramadoss, Jiri Pirko, Linus Walleij,
Florian Fainelli, Vivien Didelot
Cc: netdev, linux-kernel, kernel, Jens Emil Schulz Østergaard,
Oleksij Rempel
Pulse Fast Aging when the programmed ageing time gets shorter or the Age
Count gets smaller.
The ageing time is Age Count times Age Period. An entry stores the Age
Count when it is updated, as the datasheet says. The datasheet does not
say when a new Age Period applies.
Without the pulse, stale entries outlived an STP topology change. In my
STP tests the switch ran at 300 s as 4 x 75 s, its reset value. With
forward_delay 4 s, the bridge asks for an ageing time of 8 s. Read back
right after the write, the register showed 8 s. In three tests, an
entry learned before the change still stayed for more than a minute. I
think the switch uses the new Age Period only after the running one
ends.
The pulse is also needed when the ageing time grows but the Age Count
shrinks. From 8 s (4 x 2 s) to 31 s (1 x 31 s), an entry stayed about
63 s in one test.
Fast Aging is a control for the whole switch. On a KSZ9893 the pulse
removed the dynamic entries of both user ports, also entries still in
use, and the new period started right away. Unicast to those addresses
is flooded until the switch learns them again.
I found this in STP tests on a KSZ9893. I tested the fix there with
kernel STP and with STP off. With mstpd I checked only the programmed
ageing time, not the FDB.
Fixes: 2c119d9982b1 ("net: dsa: microchip: add the support for set_ageing_time")
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Oleksij Rempel <o.rempel@pengutronix.de>
---
Changes in v3:
- Say that Fast Aging removes the entries of all ports (Sashiko)
Changes in v2:
- Fix the tense in the comment (Andrew)
- Add new_period, and compute old_secs next to new_secs
- Rephrase the comments and the commit message
- Add a Fixes tag
---
drivers/net/dsa/microchip/ksz9477.c | 41 ++++++++++++++++++++++++++++++++-----
1 file changed, 36 insertions(+), 5 deletions(-)
diff --git a/drivers/net/dsa/microchip/ksz9477.c b/drivers/net/dsa/microchip/ksz9477.c
index 8d4647e23cda153bc68255a44c44e0d1d92469f9..972703fe7454f2892b98dc507a1cba0a5c91dabd 100644
--- a/drivers/net/dsa/microchip/ksz9477.c
+++ b/drivers/net/dsa/microchip/ksz9477.c
@@ -1289,13 +1289,16 @@ static void ksz9477_phylink_get_caps(struct dsa_switch *ds, int port,
ksz_phylink_get_caps(ds, port, config);
}
+/* The datasheet gives no pulse length. A 2 ms pulse worked on a KSZ9893. */
+#define KSZ9477_FAST_AGE_PULSE_US (2 * USEC_PER_MSEC)
+
static int ksz9477_set_ageing_time(struct dsa_switch *ds, unsigned int msecs)
{
+ u8 data, mult, value, old_period, new_period;
struct ksz_device *dev = ds->priv;
+ u32 old_secs, new_secs, max_val;
u32 secs = msecs / 1000;
- u8 data, mult, value;
- u32 max_val;
- int ret;
+ int ret, err;
#define MAX_TIMER_VAL ((1 << 8) - 1)
@@ -1314,6 +1317,10 @@ static int ksz9477_set_ageing_time(struct dsa_switch *ds, unsigned int msecs)
if (ret < 0)
return ret;
+ ret = ksz_read8(dev, REG_SW_LUE_CTRL_3, &old_period);
+ if (ret < 0)
+ return ret;
+
/* Check whether there is need to update the multiplier. */
mult = FIELD_GET(SW_AGE_CNT_M, value);
max_val = MAX_TIMER_VAL;
@@ -1336,8 +1343,32 @@ static int ksz9477_set_ageing_time(struct dsa_switch *ds, unsigned int msecs)
return ret;
}
- value = DIV_ROUND_UP(secs, data);
- return ksz_write8(dev, REG_SW_LUE_CTRL_3, value);
+ new_period = DIV_ROUND_UP(secs, data);
+ ret = ksz_write8(dev, REG_SW_LUE_CTRL_3, new_period);
+ if (ret < 0)
+ return ret;
+
+ /* data is the new Age Count and mult is the old one. An entry
+ * stores the Age Count when it is updated. The datasheet does not
+ * say when a new Age Period applies. On a KSZ9893, entries outlived
+ * the new ageing time after it got shorter, and after the Age Count
+ * got smaller. Pulse Fast Aging in both cases. Fast Aging is a
+ * control for the whole switch. On a KSZ9893 the pulse removes the
+ * dynamic entries of all ports and starts the new period.
+ */
+ old_secs = mult * old_period;
+ new_secs = data * new_period;
+ if (new_secs >= old_secs && data >= mult)
+ return 0;
+
+ ret = ksz_rmw8(dev, REG_SW_LUE_CTRL_1, SW_FAST_AGING, SW_FAST_AGING);
+ if (!ret)
+ fsleep(KSZ9477_FAST_AGE_PULSE_US);
+
+ /* Clear the bit even after an error. */
+ err = ksz_rmw8(dev, REG_SW_LUE_CTRL_1, SW_FAST_AGING, 0);
+
+ return ret ?: err;
}
void ksz9477_port_queue_split(struct ksz_device *dev, int port)
--
2.47.3
^ permalink raw reply [flat|nested] 9+ messages in thread* Re: [PATCH net-next v3 1/3] net: dsa: microchip: ksz9477: apply a shorter ageing time at once
2026-11-17 13:02 ` [PATCH net-next v3 1/3] net: dsa: microchip: ksz9477: apply a shorter ageing time at once Oleksij Rempel
@ 2026-11-19 1:03 ` netdev-bot+sashiko
2026-11-19 7:19 ` Oleksij Rempel
0 siblings, 1 reply; 9+ messages in thread
From: netdev-bot+sashiko @ 2026-11-19 1:03 UTC (permalink / raw)
To: o.rempel
Cc: woojung.huh, UNGLinuxDriver, andrew, olteanv, davem, edumazet,
kuba, pabeni, horms, arun.ramadoss, jiri, linusw, f.fainelli,
vivien.didelot, netdev, linux-kernel, kernel,
jensemil.schulzostergaard
Thank you for your contribution! Sashiko AI review found 2 potential
issue(s) to consider.
Critical: 0 · High: 0 · Medium: 1 · Low: 1
If you believe none of the issues are legitimate, please reply to each
issue of Medium severity or higher and explain why the AI is wrong.
If any of them are legitimate and you plan to respin, please let patchwork
know by including "pw-bot: cr" as a separate line at the end of your reply
(one such reply per series is enough).
- [Medium] The fix depends on a fixed Fast Aging pulse length
(KSZ9477_FAST_AGE_PULSE_US = 2 ms, ksz9477.c:1293) that was found by…
- [Low] ksz9477_set_ageing_time()
(drivers/net/dsa/microchip/ksz9477.c:1359-1371) now pulses the switch-…
--
net: dsa: microchip: ksz9477: apply a shorter ageing time at once
This patch makes ksz9477_set_ageing_time() pulse the switch-wide Fast
Aging bit in REG_SW_LUE_CTRL_1 when the programmed ageing time gets
shorter or the Age Count gets smaller. Without the pulse, stale FDB
entries outlived an STP topology change on a KSZ9893. The bit is held
for a fixed 2 ms and then cleared.
> diff --git a/drivers/net/dsa/microchip/ksz9477.c b/drivers/net/dsa/microchip/ksz9477.c
> index 8d4647e23cda1..972703fe7454f 100644
> --- a/drivers/net/dsa/microchip/ksz9477.c
> +++ b/drivers/net/dsa/microchip/ksz9477.c
> @@ -1289,13 +1289,16 @@ static void ksz9477_phylink_get_caps(struct dsa_switch *ds, int port,
> ksz_phylink_get_caps(ds, port, config);
> }
>
> +/* The datasheet gives no pulse length. A 2 ms pulse worked on a KSZ9893. */
> +#define KSZ9477_FAST_AGE_PULSE_US (2 * USEC_PER_MSEC)
[Severity: Medium]
Is a fixed 2 ms pulse enough on every chip in this family and for every
stored Age Count?
The value was found by trial on one KSZ9893, where the entries stored Age
Count 4. Each ALU entry holds a 3-bit stored age count
(ALU_V_PRIO_AGE_CNT_M, decoded in ksz9477_convert_alu()). That count is
copied from SW_AGE_CNT_M when the entry is learned or refreshed, so it can
be as high as 7.
The pulse matters most when the old count (mult) was larger, and that is
exactly when entries hold the highest counts. One example: 1785 s
(7 x 255), then an STP topology change to 8 s. That gives data = 1 and
triggers the pulse.
Suppose Fast Aging here speeds up ageing rather than purging the table at
once. Could entries with stored counts 5 to 7 then survive a 2 ms pulse?
The same question applies to the other chips that use ksz9477_switch_ops
(KSZ9477, KSZ9897, KSZ9896, KSZ9567, KSZ9563, KSZ8563, LAN9646 and so on).
ksz9477_set_ageing_time() sets SW_FAST_AGING, sleeps, clears the bit and
returns 0. It never checks that the dynamic entries are gone. If they
survive, the stale FDB after a topology change would still happen, and
nothing would report it.
The KSZ8995XA path in ksz8.c describes an "800 us fast-age cycle", which
hints at sweep behaviour. It doesn't show how the KSZ9477 family behaves.
[ ... ]
> + old_secs = mult * old_period;
> + new_secs = data * new_period;
> + if (new_secs >= old_secs && data >= mult)
> + return 0;
> +
> + ret = ksz_rmw8(dev, REG_SW_LUE_CTRL_1, SW_FAST_AGING, SW_FAST_AGING);
> + if (!ret)
> + fsleep(KSZ9477_FAST_AGE_PULSE_US);
> +
> + /* Clear the bit even after an error. */
> + err = ksz_rmw8(dev, REG_SW_LUE_CTRL_1, SW_FAST_AGING, 0);
> +
> + return ret ?: err;
> }
[Severity: Low]
Can an ageing time change in one bridge now flush the dynamic FDB entries
of every other bridge on the same switch?
The value programmed here comes from dsa_switch_fastest_ageing_time(),
which takes the minimum over all ports of the switch. Before this patch,
one bridge could only shorten the ageing timer for the others, and entries
still in use kept being refreshed. With the pulse, entries still in use in
unrelated bridges are removed too. Their unicast traffic floods until the
addresses are learned again.
Nothing limits the flush to one bridge or limits how often it runs. Here
are two ways to trigger it again and again.
With kernel STP, __br_set_topology_change() lowers the ageing time when
received BPDUs or TCNs set TC, and restores it when TC clears:
net/bridge/br_stp.c:__br_set_topology_change() {
...
if (val) {
t = 2 * br->forward_delay;
...
} else {
t = br->bridge_ageing_time;
...
err = __set_ageing_time(br->dev, t);
...
}
For example, 300 s (4 x 75) -> 30 s pulses. Going back to 300 s
(2 x 150) does not. The next TC (30 s = 2 x 15) pulses again, so every
TC on/off cycle causes one switch-wide flush.
DSA user ports are not netns-immutable. A CAP_NET_ADMIN holder in another
netns that owns a bridged switch port could toggle IFLA_BR_AGEING_TIME.
Each toggle would flush the FDB of bridges owned by the host or by other
tenants. The same happens when a port joins a bridge with a shorter
ageing time, since the attribute is replayed on join.
The commit message says Fast Aging is a control for the whole switch.
Was the effect on other bridges sharing the switch intended?
--
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261117-ksz-stp-ageing-v3-0-135de6b9f8ad%40pengutronix.de
^ permalink raw reply [flat|nested] 9+ messages in thread* Re: [PATCH net-next v3 1/3] net: dsa: microchip: ksz9477: apply a shorter ageing time at once
2026-11-19 1:03 ` netdev-bot+sashiko
@ 2026-11-19 7:19 ` Oleksij Rempel
0 siblings, 0 replies; 9+ messages in thread
From: Oleksij Rempel @ 2026-11-19 7:19 UTC (permalink / raw)
To: netdev-bot+sashiko
Cc: woojung.huh, UNGLinuxDriver, andrew, olteanv, davem, edumazet,
kuba, pabeni, horms, arun.ramadoss, jiri, linusw, f.fainelli,
vivien.didelot, netdev, linux-kernel, kernel,
jensemil.schulzostergaard
On Thu, Nov 19, 2026 at 01:03:20AM +0000, netdev-bot+sashiko@kernel.org wrote:
> Is a fixed 2 ms pulse enough on every chip in this family and for every
> stored Age Count?
I do not know. I tested only on a KSZ9893. 2 ms worked there, and I
did not try other lengths. The datasheet only says that the bit
enables fast aging. It gives no pulse length and does not say what the
bit removes. The
800 us figure in ksz8.c is for the KSZ8995XA. I have no such figure for
the KSZ9477 family.
I tested Age Counts 2 and 4. I write Age Count x Age Period, so
2 x 150 s is Age Count 2 and Age Period 150 s.
- From 2 x 150 s to 2 x 5 s, I read the FDB after the pulse. The
entries were gone within 2 s. On an earlier build with the same
code, an entry learned about 2 s after the change aged out after 5
to 10 s, so the new period applied at once.
- From 4 x 75 s to 4 x 2 s, in the kernel STP tests, a monitor read the
entries of lan1 every 1 to 3 s. The entries of stations that had
stopped sending were gone at the first read after the change. That
run used an earlier version of this patch with the same pulse code.
The tests do not show how the pulse works. It may remove all dynamic
entries at once. Or it may run a fast ageing pass, and entries with a
higher Age Count could survive it. I did not test Age Counts 5 to 7.
The driver does not check that the entries are gone after the pulse.
Starting from the reset value of 4 x 75 s, the driver picks Age Count
5 or more only for ageing times above 1020 s. I keep 2 ms for now. If
someone can test a higher Age Count on another chip of the family, I
can extend the pulse or check the table.
> Can an ageing time change in one bridge now flush the dynamic FDB entries
> of every other bridge on the same switch?
Yes. On the KSZ9893 the pulse removed the entries of both user ports.
Fast Aging, the Age Count and the Age Period are switch-wide, and
.set_ageing_time() gets no port or bridge argument. So the pulse also
hits other bridges on the switch. The commit message and the comment
say that the pulse removes the entries of all ports. They do not name
other bridges. IMO the cost is acceptable, see the next answer.
> For example, 300 s (4 x 75) -> 30 s pulses. Going back to 300 s
> (2 x 150) does not. The next TC (30 s = 2 x 15) pulses again, so every
> TC on/off cycle causes one switch-wide flush.
Yes. There is no rate limit. The pulse comes only when the programmed
ageing time or the Age Count gets smaller.
The cost is unicast flooding until the switch learns the addresses
again. The switch learns an address from the next frame that the
station sends through it.
The pulse also helps after a station moved. Without it, traffic to
that station goes to the old port and is lost until the station sends
a frame through the switch or the stale entry ages out. In my
worst-case test without this series, one-way UDP with static ARP
stopped for up to 56 s, 43 s of that after STP had converged. Such
fully one-way traffic is rare. Once its entry ages out, the switch
floods that traffic anyway.
On a topology change, kernel STP only shortens the ageing time. With the
pulse, the switch also removes the entries, as RSTP does. For RSTP and
MSTP, IEEE 802.1Q-2022, 13.19, NOTE 2 lets a bridge remove all Dynamic
Filtering Entries, not only those of the given ports: "This does not
result in incorrect operation, but will result in more flooding of
frames with unknown destination addresses." The note speaks of a single
bridge, not of several bridges on one switch. For other bridges on the
switch, I think the cost is the same: extra flooding.
> A CAP_NET_ADMIN holder in another netns that owns a bridged switch port
> could toggle IFLA_BR_AGEING_TIME.
Without this patch, that user can already lower the ageing time of the
whole switch. With this patch, that user can also cause a flush. A
flush happens when the programmed ageing time or the Age Count gets
smaller. It can also happen on a raise, for example from 8 s (2 x 4)
to 31 s (1 x 31). It also happens when a port joins a bridge with a
shorter ageing time than the switch has now. Other bridges then flood
unicast until they learn the addresses again.
Oleksij
^ permalink raw reply [flat|nested] 9+ messages in thread
* [PATCH net-next v3 2/3] net: dsa: ignore the old ageing time of the port being changed
2026-11-17 13:02 [PATCH net-next v3 0/3] net: dsa: make a shorter bridge ageing time take effect Oleksij Rempel
2026-11-17 13:02 ` [PATCH net-next v3 1/3] net: dsa: microchip: ksz9477: apply a shorter ageing time at once Oleksij Rempel
@ 2026-11-17 13:02 ` Oleksij Rempel
2026-11-17 21:37 ` Linus Walleij
2026-11-17 13:02 ` [PATCH net-next v3 3/3] net: dsa: forget the port ageing time when the port leaves the bridge Oleksij Rempel
2 siblings, 1 reply; 9+ messages in thread
From: Oleksij Rempel @ 2026-11-17 13:02 UTC (permalink / raw)
To: Woojung Huh, UNGLinuxDriver, Andrew Lunn, Vladimir Oltean,
David S. Miller, Eric Dumazet, Jakub Kicinski, Paolo Abeni,
Simon Horman, Arun Ramadoss, Jiri Pirko, Linus Walleij,
Florian Fainelli, Vivien Didelot
Cc: netdev, linux-kernel, kernel, Jens Emil Schulz Østergaard,
Oleksij Rempel
Pass the port to the ageing time notifier. When
dsa_switch_fastest_ageing_time() takes the minimum, skip the cached
dp->ageing_time of that port and use the requested value instead. DSA
updates the cache only after the notifier.
Without this, the minimum includes the port's old ageing time. During
an STP topology change, the bridge lowers the ageing time to
2 x forward_delay (8 s with forward_delay 4 s). When it restores the
configured value, DSA programs the cached 8 s again, and the switch
keeps ageing at 8 s.
An automated review of the lan9645x bridge series pointed this out [1].
I reproduced it on a KSZ9893 and tested the fix there.
Fixes: 77b61365ecef ("net: dsa: remove the transactional logic from ageing time notifiers")
Link: https://lore.kernel.org/netdev/179180428624.434549.14425623884223795761@kernel.org/ [1]
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Oleksij Rempel <o.rempel@pengutronix.de>
---
Changes in v3:
- No changes.
Changes in v2:
- Pass the port to the notifier and skip its old value, instead of
updating the cache first (Andrew)
- Rephrase the subject and the commit message, and move the sja1105
note to the cover letter
- Add a Fixes tag
---
net/dsa/port.c | 7 ++++---
net/dsa/switch.c | 17 +++++++++++++----
net/dsa/switch.h | 1 +
3 files changed, 18 insertions(+), 7 deletions(-)
diff --git a/net/dsa/port.c b/net/dsa/port.c
index 1f5536c0dffc2311db24015697e0c8f5c014ac74..15e7da20b0e2f470a0ce5b763fa73b9b66b4c6f2 100644
--- a/net/dsa/port.c
+++ b/net/dsa/port.c
@@ -876,11 +876,12 @@ int dsa_port_ageing_time(struct dsa_port *dp, clock_t ageing_clock)
{
unsigned long ageing_jiffies = clock_t_to_jiffies(ageing_clock);
unsigned int ageing_time = jiffies_to_msecs(ageing_jiffies);
- struct dsa_notifier_ageing_time_info info;
+ struct dsa_notifier_ageing_time_info info = {
+ .dp = dp,
+ .ageing_time = ageing_time,
+ };
int err;
- info.ageing_time = ageing_time;
-
err = dsa_port_notify(dp, DSA_NOTIFIER_AGEING_TIME, &info);
if (err)
return err;
diff --git a/net/dsa/switch.c b/net/dsa/switch.c
index a670a04edb7237a2606bc1959f6f2bb324c832ab..e9bef66fea3173c54fcde88da319cb8fd91195fc 100644
--- a/net/dsa/switch.c
+++ b/net/dsa/switch.c
@@ -20,14 +20,23 @@
#include "trace.h"
#include "user.h"
-static unsigned int dsa_switch_fastest_ageing_time(struct dsa_switch *ds,
- unsigned int ageing_time)
+static unsigned int
+dsa_switch_fastest_ageing_time(struct dsa_switch *ds,
+ const struct dsa_notifier_ageing_time_info *info)
{
+ unsigned int ageing_time = info->ageing_time;
struct dsa_port *dp;
- dsa_switch_for_each_port(dp, ds)
+ dsa_switch_for_each_port(dp, ds) {
+ /* info->dp is the port being changed. Its cached value is
+ * stale until the notifier returns. Skip it.
+ */
+ if (dp == info->dp)
+ continue;
+
if (dp->ageing_time && dp->ageing_time < ageing_time)
ageing_time = dp->ageing_time;
+ }
return ageing_time;
}
@@ -44,7 +53,7 @@ static int dsa_switch_ageing_time(struct dsa_switch *ds,
return -ERANGE;
/* Program the fastest ageing time in case of multiple bridges */
- ageing_time = dsa_switch_fastest_ageing_time(ds, ageing_time);
+ ageing_time = dsa_switch_fastest_ageing_time(ds, info);
if (ds->ops->set_ageing_time)
return ds->ops->set_ageing_time(ds, ageing_time);
diff --git a/net/dsa/switch.h b/net/dsa/switch.h
index be0a2749cd9791e19b8f03c6f6ceafb2d0432b14..5362ee9bfb68cfb52759c108d0626c55e19bb3a3 100644
--- a/net/dsa/switch.h
+++ b/net/dsa/switch.h
@@ -39,6 +39,7 @@ enum {
/* DSA_NOTIFIER_AGEING_TIME */
struct dsa_notifier_ageing_time_info {
+ const struct dsa_port *dp;
unsigned int ageing_time;
};
--
2.47.3
^ permalink raw reply [flat|nested] 9+ messages in thread* Re: [PATCH net-next v3 2/3] net: dsa: ignore the old ageing time of the port being changed
2026-11-17 13:02 ` [PATCH net-next v3 2/3] net: dsa: ignore the old ageing time of the port being changed Oleksij Rempel
@ 2026-11-17 21:37 ` Linus Walleij
0 siblings, 0 replies; 9+ messages in thread
From: Linus Walleij @ 2026-11-17 21:37 UTC (permalink / raw)
To: Oleksij Rempel
Cc: Woojung Huh, UNGLinuxDriver, Andrew Lunn, Vladimir Oltean,
David S. Miller, Eric Dumazet, Jakub Kicinski, Paolo Abeni,
Simon Horman, Arun Ramadoss, Jiri Pirko, Florian Fainelli,
Vivien Didelot, netdev, linux-kernel, kernel,
Jens Emil Schulz Østergaard
On Tue, Nov 17, 2026 at 2:02 PM Oleksij Rempel <o.rempel@pengutronix.de> wrote:
> Pass the port to the ageing time notifier. When
> dsa_switch_fastest_ageing_time() takes the minimum, skip the cached
> dp->ageing_time of that port and use the requested value instead. DSA
> updates the cache only after the notifier.
>
> Without this, the minimum includes the port's old ageing time. During
> an STP topology change, the bridge lowers the ageing time to
> 2 x forward_delay (8 s with forward_delay 4 s). When it restores the
> configured value, DSA programs the cached 8 s again, and the switch
> keeps ageing at 8 s.
>
> An automated review of the lan9645x bridge series pointed this out [1].
> I reproduced it on a KSZ9893 and tested the fix there.
>
> Fixes: 77b61365ecef ("net: dsa: remove the transactional logic from ageing time notifiers")
> Link: https://lore.kernel.org/netdev/179180428624.434549.14425623884223795761@kernel.org/ [1]
> Reviewed-by: Andrew Lunn <andrew@lunn.ch>
> Signed-off-by: Oleksij Rempel <o.rempel@pengutronix.de>
This looks right to me, good find from Sashiko!
Reviewed-by: Linus Walleij <linusw@kernel.org>
Yours,
Linus Walleij
^ permalink raw reply [flat|nested] 9+ messages in thread
* [PATCH net-next v3 3/3] net: dsa: forget the port ageing time when the port leaves the bridge
2026-11-17 13:02 [PATCH net-next v3 0/3] net: dsa: make a shorter bridge ageing time take effect Oleksij Rempel
2026-11-17 13:02 ` [PATCH net-next v3 1/3] net: dsa: microchip: ksz9477: apply a shorter ageing time at once Oleksij Rempel
2026-11-17 13:02 ` [PATCH net-next v3 2/3] net: dsa: ignore the old ageing time of the port being changed Oleksij Rempel
@ 2026-11-17 13:02 ` Oleksij Rempel
2026-11-19 1:03 ` netdev-bot+sashiko
2 siblings, 1 reply; 9+ messages in thread
From: Oleksij Rempel @ 2026-11-17 13:02 UTC (permalink / raw)
To: Woojung Huh, UNGLinuxDriver, Andrew Lunn, Vladimir Oltean,
David S. Miller, Eric Dumazet, Jakub Kicinski, Paolo Abeni,
Simon Horman, Arun Ramadoss, Jiri Pirko, Linus Walleij,
Florian Fainelli, Vivien Didelot
Cc: netdev, linux-kernel, kernel, Jens Emil Schulz Østergaard,
Oleksij Rempel
Clear the cached dp->ageing_time when the port leaves the bridge.
Otherwise dsa_switch_fastest_ageing_time() still uses the ageing time
of the port that left. If the port leaves with a shorter ageing time,
for example during an STP topology change, the switch stays at the
short value until that port joins a bridge again. With this change the
switch goes back to the long value when the bridge restores it at the
end of the topology change. The switch gets the restored value only if
another port of the same bridge is still in this DSA tree.
Found by code inspection. Tested on a KSZ9893 with lan1 and lan2 in a
bridge, ageing time 300 s:
- kernel STP, forward delay 4 s, lan2 removed during a topology change:
without this patch the switch stayed at 8 s (2 x 4 s). With it, the
switch went back to 300 s at the end of the topology change.
- no STP: I set the ageing time to 10 s, removed lan2, then set 300 s.
The switch stayed at 10 s without this patch and went to 300 s with
it.
Fixes: 34a79f63bbe4 ("net: dsa: support switchdev ageing time attr")
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Oleksij Rempel <o.rempel@pengutronix.de>
---
Changes in v3:
- No changes.
Changes in v2:
- Say how long the short ageing time stays, with an STP test (Andrew)
- Rephrase the commit message and the comment
- Move the case of the last bridge port leaving to the cover letter
- Add a Fixes tag
---
net/dsa/port.c | 6 ++++++
1 file changed, 6 insertions(+)
diff --git a/net/dsa/port.c b/net/dsa/port.c
index 15e7da20b0e2f470a0ce5b763fa73b9b66b4c6f2..54c9709840952210ecd1abdfde31f751d07ef2b6 100644
--- a/net/dsa/port.c
+++ b/net/dsa/port.c
@@ -415,6 +415,12 @@ static void dsa_port_switchdev_unsync_attrs(struct dsa_port *dp,
/* Ageing time may be global to the switch chip, so don't change it
* here because we have no good reason (or value) to change it to.
*/
+
+ /* Forget the ageing time cached for this port.
+ * dsa_switch_fastest_ageing_time() skips 0, so this port no longer
+ * counts towards the minimum.
+ */
+ dp->ageing_time = 0;
}
static int dsa_port_bridge_create(struct dsa_port *dp,
--
2.47.3
^ permalink raw reply [flat|nested] 9+ messages in thread* Re: [PATCH net-next v3 3/3] net: dsa: forget the port ageing time when the port leaves the bridge
2026-11-17 13:02 ` [PATCH net-next v3 3/3] net: dsa: forget the port ageing time when the port leaves the bridge Oleksij Rempel
@ 2026-11-19 1:03 ` netdev-bot+sashiko
2026-11-19 7:23 ` Oleksij Rempel
0 siblings, 1 reply; 9+ messages in thread
From: netdev-bot+sashiko @ 2026-11-19 1:03 UTC (permalink / raw)
To: o.rempel
Cc: woojung.huh, UNGLinuxDriver, andrew, olteanv, davem, edumazet,
kuba, pabeni, horms, arun.ramadoss, jiri, linusw, f.fainelli,
vivien.didelot, netdev, linux-kernel, kernel,
jensemil.schulzostergaard
Thank you for your contribution! Sashiko AI review found 1 potential
issue(s) to consider.
Critical: 0 · High: 0 · Medium: 1 · Low: 0
If you believe none of the issues are legitimate, please reply to each
issue of Medium severity or higher and explain why the AI is wrong.
If any of them are legitimate and you plan to respin, please let patchwork
know by including "pw-bot: cr" as a separate line at the end of your reply
(one such reply per series is enough).
- [Medium] The fix is incomplete because nothing reprograms the switch
when the port leaves.
--
net: dsa: forget the port ageing time when the port leaves the bridge
This patch clears the cached dp->ageing_time in
dsa_port_switchdev_unsync_attrs() when a port leaves a bridge. After that,
dsa_switch_fastest_ageing_time() no longer counts the departed port's
value when it computes the minimum. The aim is to keep the switch from
staying at a short ageing time (for example one set during an STP
topology change) after the port that requested it has left.
> diff --git a/net/dsa/port.c b/net/dsa/port.c
> index 15e7da20b0e2f..54c9709840952 100644
> --- a/net/dsa/port.c
> +++ b/net/dsa/port.c
> @@ -415,6 +415,12 @@ static void dsa_port_switchdev_unsync_attrs(struct dsa_port *dp,
> /* Ageing time may be global to the switch chip, so don't change it
> * here because we have no good reason (or value) to change it to.
> */
> +
> + /* Forget the ageing time cached for this port.
> + * dsa_switch_fastest_ageing_time() skips 0, so this port no longer
> + * counts towards the minimum.
> + */
> + dp->ageing_time = 0;
> }
[Severity: Medium]
Can the switch still stay at the departed port's short ageing time after
this change?
This only clears the cache. Nothing on the leave path reprograms the
hardware. dsa_port_bridge_leave() ends right after
dsa_port_switchdev_unsync_attrs(). The only code that writes the ageing
register is dsa_switch_ageing_time() in net/dsa/switch.c:
net/dsa/switch.c:dsa_switch_ageing_time() {
...
ageing_time = dsa_switch_fastest_ageing_time(ds, info);
if (ds->ops->set_ageing_time)
return ds->ops->set_ageing_time(ds, ageing_time);
...
}
That function runs only for DSA_NOTIFIER_AGEING_TIME. The only sender is
dsa_port_ageing_time(). It is called from the sync_attrs path at bridge
join, and from the SWITCHDEV_ATTR_ID_BRIDGE_AGEING_TIME handler, which
needs a port that still offloads the bridge.
Here is a sequence on a single chip, with br0={lan1} at 300 s and
br1={lan2} with STP enabled:
topology change on br1
__set_ageing_time(br1, 2 * fwd_delay)
dsa_port_ageing_time(lan2, 8 s)
programs min(8, lan1=300) = 8 s for the whole chip
ip link set lan2 nomaster
dsa_port_bridge_leave()
dsa_port_switchdev_unsync_attrs()
dp->ageing_time = 0; /* no reprogram */
end of the topology change on br1
br1 restores 300 s, but br1 has no DSA lower port left,
so no notifier is sent
lan1 in br0 would keep ageing at 8 s until br0's ageing time changes or
some port joins a bridge again. Would it make sense to recompute the
minimum of the remaining non-zero dp->ageing_time values on dp->ds and
program it here when the port leaves?
The existing comment just above the new code says there is "no good
reason (or value) to change it to". Once the cache is cleared, the
minimum of the remaining cached values is a usable value, so that
comment no longer seems accurate. The new comment also doesn't mention
that the hardware value stays unchanged until the next notification.
The commit message does cover part of this:
"The switch gets the restored value only if another port of the same
bridge is still in this DSA tree."
However, the problem the commit message describes, a switch stuck at the
departed port's short ageing time, can still happen through this leave
path when ports of a different bridge remain on the same chip.
--
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261117-ksz-stp-ageing-v3-0-135de6b9f8ad%40pengutronix.de
^ permalink raw reply [flat|nested] 9+ messages in thread* Re: [PATCH net-next v3 3/3] net: dsa: forget the port ageing time when the port leaves the bridge
2026-11-19 1:03 ` netdev-bot+sashiko
@ 2026-11-19 7:23 ` Oleksij Rempel
0 siblings, 0 replies; 9+ messages in thread
From: Oleksij Rempel @ 2026-11-19 7:23 UTC (permalink / raw)
To: netdev-bot+sashiko
Cc: woojung.huh, UNGLinuxDriver, andrew, olteanv, davem, edumazet,
kuba, pabeni, horms, arun.ramadoss, jiri, linusw, f.fainelli,
vivien.didelot, netdev, linux-kernel, kernel,
jensemil.schulzostergaard
On Thu, Nov 19, 2026 at 01:03:22AM +0000, netdev-bot+sashiko@kernel.org wrote:
> Can the switch still stay at the departed port's short ageing time after
> this change?
Yes. The v3 cover letter lists this case as not fixed: "The last port
of one bridge leaves the switch while another bridge stays on it."
In my reply to v2 I wrote that v3 would add a patch for this case. I
dropped it, and the v3 changelog should have said so. That patch
reprogrammed the switch when a port left. It did not cover the case
where the last bridged port of the switch leaves.
This patch only drops a cached value and writes nothing to the switch.
The dropped patch added a write to the switch, and with patch 1
sometimes a flush, to every port leave. That is new behaviour for
every driver with .set_ageing_time(). I want such a change reviewed on
its own, not in this series of fixes.
This patch does not make the case worse:
- Before this patch, DSA kept the ageing time of a port that left
until the port joined a bridge again. Every later update included
that stale value in the minimum. So the switch got the short ageing
time again at each update.
- With this patch, DSA forgets that value. The switch keeps the short
value until DSA computes the ageing time again. That happens when a
port of this DSA tree joins a bridge, or when a bridge with a port of
this tree sets its ageing time. If neither happens, the switch stays
at the short value.
> The existing comment just above the new code says there is "no good
> reason (or value) to change it to".
Yes, that comment is out of date. After this patch there is a value to
use when another port of the switch is still bridged: the minimum of
the remaining cached values. I still do not reprogram the switch on
leave in this series, for the reason above.
Oleksij
^ permalink raw reply [flat|nested] 9+ messages in thread