linux.git/drivers/net/dsa/ocelot, branch for-next Linux kernel source tree net: dsa: felix: fix stuck CPU-injected packets with short taprio windows 2024-12-12T04:24:56+00:00 Vladimir Oltean vladimir.oltean@nxp.com 2024-12-10T13:26:40+00:00 acfcdb78d5d4cdb78e975210c8825b9a112463f6 With this port schedule: tc qdisc replace dev $send_if parent root handle 100 taprio \ num_tc 8 queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \ map 0 1 2 3 4 5 6 7 \ base-time 0 cycle-time 10000 \ sched-entry S 01 1250 \ sched-entry S 02 1250 \ sched-entry S 04 1250 \ sched-entry S 08 1250 \ sched-entry S 10 1250 \ sched-entry S 20 1250 \ sched-entry S 40 1250 \ sched-entry S 80 1250 \ flags 2 ptp4l would fail to take TX timestamps of Pdelay_Resp messages like this: increasing tx_timestamp_timeout may correct this issue, but it is likely caused by a driver bug ptp4l[4134.168]: port 2: send peer delay response failed It turns out that the driver can't take their TX timestamps because it can't transmit them in the first place. And there's nothing special about the Pdelay_Resp packets - they're just regular 68 byte packets. But with this taprio configuration, the switch would refuse to send even the ETH_ZLEN minimum packet size. This should have definitely not been the case. When applying the taprio config, the driver prints: mscc_felix 0000:00:00.5: port 0 tc 0 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS mscc_felix 0000:00:00.5: port 0 tc 1 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS mscc_felix 0000:00:00.5: port 0 tc 2 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS mscc_felix 0000:00:00.5: port 0 tc 3 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS mscc_felix 0000:00:00.5: port 0 tc 4 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS mscc_felix 0000:00:00.5: port 0 tc 5 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS mscc_felix 0000:00:00.5: port 0 tc 6 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS mscc_felix 0000:00:00.5: port 0 tc 7 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS and thus, everything under 132 bytes - ETH_FCS_LEN should have been sent without problems. Yet it's not. For the forwarding path, the configuration is fine, yet packets injected from Linux get stuck with this schedule no matter what. The first hint that the static guard bands are the cause of the problem is that reverting Michael Walle's commit 297c4de6f780 ("net: dsa: felix: re-enable TAS guard band mode") made things work. It must be that the guard bands are calculated incorrectly. I remembered that there is a magic constant in the driver, set to 33 ns for no logical reason other than experimentation, which says "never let the static guard bands get so large as to leave less than this amount of remaining space in the time slot, because the queue system will refuse to schedule packets otherwise, and they will get stuck". I had a hunch that my previous experimentally-determined value was only good for packets coming from the forwarding path, and that the CPU injection path needed more. I came to the new value of 35 ns through binary search, after seeing that with 544 ns (the bit time required to send the Pdelay_Resp packet at gigabit) it works. Again, this is purely experimental, there's no logic and the manual doesn't say anything. The new driver prints for this schedule look like this: mscc_felix 0000:00:00.5: port 0 tc 0 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS mscc_felix 0000:00:00.5: port 0 tc 1 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS mscc_felix 0000:00:00.5: port 0 tc 2 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS mscc_felix 0000:00:00.5: port 0 tc 3 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS mscc_felix 0000:00:00.5: port 0 tc 4 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS mscc_felix 0000:00:00.5: port 0 tc 5 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS mscc_felix 0000:00:00.5: port 0 tc 6 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS mscc_felix 0000:00:00.5: port 0 tc 7 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS So yes, the maximum MTU is now even smaller by 1 byte than before. This is maybe counter-intuitive, but makes more sense with a diagram of one time slot. Before: Gate open Gate close | | v 1250 ns total time slot duration v <----------------------------------------------------> <----><----------------------------------------------> 33 ns 1217 ns static guard band useful Gate open Gate close | | v 1250 ns total time slot duration v <----------------------------------------------------> <-----><---------------------------------------------> 35 ns 1215 ns static guard band useful The static guard band implemented by this switch hardware directly determines the maximum allowable MTU for that traffic class. The larger it is, the earlier the switch will stop scheduling frames for transmission, because otherwise they might overrun the gate close time (and avoiding that is the entire purpose of Michael's patch). So, we now have guard bands smaller by 2 ns, thus, in this particular case, we lose a byte of the maximum MTU. Fixes: 11afdc6526de ("net: dsa: felix: tc-taprio intervals smaller than MTU should send at least one packet") Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> Reviewed-by: Michael Walle <mwalle@kernel.org> Link: https://patch.msgid.link/20241210132640.3426788-1-vladimir.oltean@nxp.com Signed-off-by: Jakub Kicinski <kuba@kernel.org> 
With this port schedule:

tc qdisc replace dev $send_if parent root handle 100 taprio \
	num_tc 8 queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \
	map 0 1 2 3 4 5 6 7 \
	base-time 0 cycle-time 10000 \
	sched-entry S 01 1250 \
	sched-entry S 02 1250 \
	sched-entry S 04 1250 \
	sched-entry S 08 1250 \
	sched-entry S 10 1250 \
	sched-entry S 20 1250 \
	sched-entry S 40 1250 \
	sched-entry S 80 1250 \
	flags 2

ptp4l would fail to take TX timestamps of Pdelay_Resp messages like this:

increasing tx_timestamp_timeout may correct this issue, but it is likely caused by a driver bug
ptp4l[4134.168]: port 2: send peer delay response failed

It turns out that the driver can't take their TX timestamps because it
can't transmit them in the first place. And there's nothing special
about the Pdelay_Resp packets - they're just regular 68 byte packets.
But with this taprio configuration, the switch would refuse to send even
the ETH_ZLEN minimum packet size.

This should have definitely not been the case. When applying the taprio
config, the driver prints:

mscc_felix 0000:00:00.5: port 0 tc 0 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS
mscc_felix 0000:00:00.5: port 0 tc 1 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS
mscc_felix 0000:00:00.5: port 0 tc 2 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS
mscc_felix 0000:00:00.5: port 0 tc 3 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS
mscc_felix 0000:00:00.5: port 0 tc 4 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS
mscc_felix 0000:00:00.5: port 0 tc 5 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS
mscc_felix 0000:00:00.5: port 0 tc 6 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS
mscc_felix 0000:00:00.5: port 0 tc 7 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 132 octets including FCS

and thus, everything under 132 bytes - ETH_FCS_LEN should have been sent
without problems. Yet it's not.

For the forwarding path, the configuration is fine, yet packets injected
from Linux get stuck with this schedule no matter what.

The first hint that the static guard bands are the cause of the problem
is that reverting Michael Walle's commit 297c4de6f780 ("net: dsa: felix:
re-enable TAS guard band mode") made things work. It must be that the
guard bands are calculated incorrectly.

I remembered that there is a magic constant in the driver, set to 33 ns
for no logical reason other than experimentation, which says "never let
the static guard bands get so large as to leave less than this amount of
remaining space in the time slot, because the queue system will refuse
to schedule packets otherwise, and they will get stuck". I had a hunch
that my previous experimentally-determined value was only good for
packets coming from the forwarding path, and that the CPU injection path
needed more.

I came to the new value of 35 ns through binary search, after seeing
that with 544 ns (the bit time required to send the Pdelay_Resp packet
at gigabit) it works. Again, this is purely experimental, there's no
logic and the manual doesn't say anything.

The new driver prints for this schedule look like this:

mscc_felix 0000:00:00.5: port 0 tc 0 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS
mscc_felix 0000:00:00.5: port 0 tc 1 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS
mscc_felix 0000:00:00.5: port 0 tc 2 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS
mscc_felix 0000:00:00.5: port 0 tc 3 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS
mscc_felix 0000:00:00.5: port 0 tc 4 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS
mscc_felix 0000:00:00.5: port 0 tc 5 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS
mscc_felix 0000:00:00.5: port 0 tc 6 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS
mscc_felix 0000:00:00.5: port 0 tc 7 min gate length 1250 ns not enough for max frame size 1526 at 1000 Mbps, dropping frames over 131 octets including FCS

So yes, the maximum MTU is now even smaller by 1 byte than before.
This is maybe counter-intuitive, but makes more sense with a diagram of
one time slot.

Before:

 Gate open                                   Gate close
 |                                                    |
 v           1250 ns total time slot duration         v
 <---------------------------------------------------->
 <----><---------------------------------------------->
  33 ns            1217 ns static guard band
  useful

 Gate open                                   Gate close
 |                                                    |
 v           1250 ns total time slot duration         v
 <---------------------------------------------------->
 <-----><--------------------------------------------->
  35 ns            1215 ns static guard band
  useful

The static guard band implemented by this switch hardware directly
determines the maximum allowable MTU for that traffic class. The larger
it is, the earlier the switch will stop scheduling frames for
transmission, because otherwise they might overrun the gate close time
(and avoiding that is the entire purpose of Michael's patch).
So, we now have guard bands smaller by 2 ns, thus, in this particular
case, we lose a byte of the maximum MTU.

Fixes: 11afdc6526de ("net: dsa: felix: tc-taprio intervals smaller than MTU should send at least one packet")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Reviewed-by: Michael Walle <mwalle@kernel.org>
Link: https://patch.msgid.link/20241210132640.3426788-1-vladimir.oltean@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
module: Convert symbol namespace to string literal 2024-12-02T19:34:44+00:00 Peter Zijlstra peterz@infradead.org 2024-12-02T14:59:47+00:00 cdd30ebb1b9f36159d66f088b61aee264e649d7a Clean up the existing export namespace code along the same lines of commit 33def8498fdd ("treewide: Convert macro and uses of __section(foo) to __section("foo")") and for the same reason, it is not desired for the namespace argument to be a macro expansion itself. Scripted using git grep -l -e MODULE_IMPORT_NS -e EXPORT_SYMBOL_NS | while read file; do awk -i inplace ' /^#define EXPORT_SYMBOL_NS/ { gsub(/__stringify\(ns\)/, "ns"); print; next; } /^#define MODULE_IMPORT_NS/ { gsub(/__stringify\(ns\)/, "ns"); print; next; } /MODULE_IMPORT_NS/ { $0 = gensub(/MODULE_IMPORT_NS\(([^)]*)\)/, "MODULE_IMPORT_NS(\"\\1\")", "g"); } /EXPORT_SYMBOL_NS/ { if ($0 ~ /(EXPORT_SYMBOL_NS[^(]*)\(([^,]+),/) { if ($0 !~ /(EXPORT_SYMBOL_NS[^(]*)\(([^,]+), ([^)]+)\)/ && $0 !~ /(EXPORT_SYMBOL_NS[^(]*)\(\)/ && $0 !~ /^my/) { getline line; gsub(/[[:space:]]*\\$/, ""); gsub(/[[:space:]]/, "", line); $0 = $0 " " line; } $0 = gensub(/(EXPORT_SYMBOL_NS[^(]*)\(([^,]+), ([^)]+)\)/, "\\1(\\2, \"\\3\")", "g"); } } { print }' $file; done Requested-by: Masahiro Yamada <masahiroy@kernel.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://mail.google.com/mail/u/2/#inbox/FMfcgzQXKWgMmjdFwwdsfgxzKpVHWPlc Acked-by: Greg KH <gregkh@linuxfoundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> 
Clean up the existing export namespace code along the same lines of
commit 33def8498fdd ("treewide: Convert macro and uses of __section(foo)
to __section("foo")") and for the same reason, it is not desired for the
namespace argument to be a macro expansion itself.

Scripted using

  git grep -l -e MODULE_IMPORT_NS -e EXPORT_SYMBOL_NS | while read file;
  do
    awk -i inplace '
      /^#define EXPORT_SYMBOL_NS/ {
        gsub(/__stringify\(ns\)/, "ns");
        print;
        next;
      }
      /^#define MODULE_IMPORT_NS/ {
        gsub(/__stringify\(ns\)/, "ns");
        print;
        next;
      }
      /MODULE_IMPORT_NS/ {
        $0 = gensub(/MODULE_IMPORT_NS\(([^)]*)\)/, "MODULE_IMPORT_NS(\"\\1\")", "g");
      }
      /EXPORT_SYMBOL_NS/ {
        if ($0 ~ /(EXPORT_SYMBOL_NS[^(]*)\(([^,]+),/) {
  	if ($0 !~ /(EXPORT_SYMBOL_NS[^(]*)\(([^,]+), ([^)]+)\)/ &&
  	    $0 !~ /(EXPORT_SYMBOL_NS[^(]*)\(\)/ &&
  	    $0 !~ /^my/) {
  	  getline line;
  	  gsub(/[[:space:]]*\\$/, "");
  	  gsub(/[[:space:]]/, "", line);
  	  $0 = $0 " " line;
  	}

  	$0 = gensub(/(EXPORT_SYMBOL_NS[^(]*)\(([^,]+), ([^)]+)\)/,
  		    "\\1(\\2, \"\\3\")", "g");
        }
      }
      { print }' $file;
  done

Requested-by: Masahiro Yamada <masahiroy@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://mail.google.com/mail/u/2/#inbox/FMfcgzQXKWgMmjdFwwdsfgxzKpVHWPlc
Acked-by: Greg KH <gregkh@linuxfoundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
net: dsa: Switch back to struct platform_driver::remove() 2024-10-04T23:39:57+00:00 Uwe Kleine-König u.kleine-koenig@baylibre.com 2024-10-03T10:01:04+00:00 4818016ded1c340e6bbce46f7ee87811dc41215b After commit 0edb555a65d1 ("platform: Make platform_driver::remove() return void") .remove() is (again) the right callback to implement for platform drivers. Convert all platform drivers below drivers/net/dsa to use .remove(), with the eventual goal to drop struct platform_driver::remove_new(). As .remove() and .remove_new() have the same prototypes, conversion is done by just changing the structure member name in the driver initializer. Signed-off-by: Uwe Kleine-König <u.kleine-koenig@baylibre.com> Reviewed-by: Andrew Lunn <andrew@lunn.ch> Link: https://patch.msgid.link/36da477cb9fa0bffec32d50c2cf3d18e94a0e7e3.1727949050.git.u.kleine-koenig@baylibre.com Signed-off-by: Jakub Kicinski <kuba@kernel.org> 
After commit 0edb555a65d1 ("platform: Make platform_driver::remove()
return void") .remove() is (again) the right callback to implement for
platform drivers.

Convert all platform drivers below drivers/net/dsa to use .remove(),
with the eventual goal to drop struct platform_driver::remove_new(). As
.remove() and .remove_new() have the same prototypes, conversion is done
by just changing the structure member name in the driver initializer.

Signed-off-by: Uwe Kleine-König <u.kleine-koenig@baylibre.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Link: https://patch.msgid.link/36da477cb9fa0bffec32d50c2cf3d18e94a0e7e3.1727949050.git.u.kleine-koenig@baylibre.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Merge git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net 2024-09-13T00:11:24+00:00 Jakub Kicinski kuba@kernel.org 2024-09-06T03:27:09+00:00 46ae4d0a489741565520195bddebc3414781e603 Cross-merge networking fixes after downstream PR. No conflicts (sort of) and no adjacent changes. This merge reverts commit b3c9e65eb227 ("net: hsr: remove seqnr_lock") from net, as it was superseded by commit 430d67bdcb04 ("net: hsr: Use the seqnr lock for frames received via interlink port.") in net-next. Signed-off-by: Jakub Kicinski <kuba@kernel.org> 
Cross-merge networking fixes after downstream PR.

No conflicts (sort of) and no adjacent changes.

This merge reverts commit b3c9e65eb227 ("net: hsr: remove seqnr_lock")
from net, as it was superseded by
commit 430d67bdcb04 ("net: hsr: Use the seqnr lock for frames received via interlink port.")
in net-next.

Signed-off-by: Jakub Kicinski <kuba@kernel.org>
net: dsa: felix: ignore pending status of TAS module when it's disabled 2024-09-11T22:54:51+00:00 Xiaoliang Yang xiaoliang.yang_1@nxp.com 2024-09-06T09:35:50+00:00 70654f4c212e83898feced125d91ebb3695950d8 The TAS module could not be configured when it's running in pending status. We need disable the module and configure it again. However, the pending status is not cleared after the module disabled. TC taprio set will always return busy even it's disabled. For example, a user uses tc-taprio to configure Qbv and a future basetime. The TAS module will run in a pending status. There is no way to reconfigure Qbv, it always returns busy. Actually the TAS module can be reconfigured when it's disabled. So it doesn't need to check the pending status if the TAS module is disabled. After the patch, user can delete the tc taprio configuration to disable Qbv and reconfigure it again. Fixes: de143c0e274b ("net: dsa: felix: Configure Time-Aware Scheduler via taprio offload") Signed-off-by: Xiaoliang Yang <xiaoliang.yang_1@nxp.com> Link: https://patch.msgid.link/20240906093550.29985-1-xiaoliang.yang_1@nxp.com Signed-off-by: Jakub Kicinski <kuba@kernel.org> 
The TAS module could not be configured when it's running in pending
status. We need disable the module and configure it again. However, the
pending status is not cleared after the module disabled. TC taprio set
will always return busy even it's disabled.

For example, a user uses tc-taprio to configure Qbv and a future
basetime. The TAS module will run in a pending status. There is no way
to reconfigure Qbv, it always returns busy.

Actually the TAS module can be reconfigured when it's disabled. So it
doesn't need to check the pending status if the TAS module is disabled.

After the patch, user can delete the tc taprio configuration to disable
Qbv and reconfigure it again.

Fixes: de143c0e274b ("net: dsa: felix: Configure Time-Aware Scheduler via taprio offload")
Signed-off-by: Xiaoliang Yang <xiaoliang.yang_1@nxp.com>
Link: https://patch.msgid.link/20240906093550.29985-1-xiaoliang.yang_1@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
net: dsa: felix: Annotate struct action_gate_entry with __counted_by 2024-09-05T11:01:01+00:00 Hongbo Li lihongbo22@huawei.com 2024-09-04T01:49:56+00:00 50ddaedeae7562b25e8eeca04f581d7c815be1db Add the __counted_by compiler attribute to the flexible array member entries to improve access bounds-checking via CONFIG_UBSAN_BOUNDS and CONFIG_FORTIFY_SOURCE. Signed-off-by: Hongbo Li <lihongbo22@huawei.com> Link: https://patch.msgid.link/20240904014956.2035117-1-lihongbo22@huawei.com Signed-off-by: Paolo Abeni <pabeni@redhat.com> 
Add the __counted_by compiler attribute to the flexible array member
entries to improve access bounds-checking via CONFIG_UBSAN_BOUNDS and
CONFIG_FORTIFY_SOURCE.

Signed-off-by: Hongbo Li <lihongbo22@huawei.com>
Link: https://patch.msgid.link/20240904014956.2035117-1-lihongbo22@huawei.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Merge git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net 2024-08-23T00:06:18+00:00 Jakub Kicinski kuba@kernel.org 2024-08-23T00:05:09+00:00 761d527d5d0036b98a2cecf4de1293d84d452aa1 Cross-merge networking fixes after downstream PR. No conflicts. Adjacent changes: drivers/net/ethernet/broadcom/bnxt/bnxt.h c948c0973df5 ("bnxt_en: Don't clear ntuple filters and rss contexts during ethtool ops") f2878cdeb754 ("bnxt_en: Add support to call FW to update a VNIC") Link: https://patch.msgid.link/20240822210125.1542769-1-kuba@kernel.org Signed-off-by: Jakub Kicinski <kuba@kernel.org> 
Cross-merge networking fixes after downstream PR.

No conflicts.

Adjacent changes:

drivers/net/ethernet/broadcom/bnxt/bnxt.h
  c948c0973df5 ("bnxt_en: Don't clear ntuple filters and rss contexts during ethtool ops")
  f2878cdeb754 ("bnxt_en: Add support to call FW to update a VNIC")

Link: https://patch.msgid.link/20240822210125.1542769-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
net: dsa: ocelot: Simplify with scoped for each OF child loop 2024-08-22T00:18:44+00:00 Jinjie Ruan ruanjinjie@huawei.com 2024-08-20T07:48:05+00:00 e58c3f3d5196952ffa29dd987a15405b7f559af9 Use scoped for_each_available_child_of_node_scoped() when iterating over device nodes to make code a bit simpler. Signed-off-by: Jinjie Ruan <ruanjinjie@huawei.com> Link: https://patch.msgid.link/20240820074805.680674-1-ruanjinjie@huawei.com Signed-off-by: Jakub Kicinski <kuba@kernel.org> 
Use scoped for_each_available_child_of_node_scoped() when iterating over
device nodes to make code a bit simpler.

Signed-off-by: Jinjie Ruan <ruanjinjie@huawei.com>
Link: https://patch.msgid.link/20240820074805.680674-1-ruanjinjie@huawei.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
net: dsa: felix: fix VLAN tag loss on CPU reception with ocelot-8021q 2024-08-16T08:59:32+00:00 Vladimir Oltean vladimir.oltean@nxp.com 2024-08-15T00:07:06+00:00 f1288fd7293b91442ad7420394c252a252ecaa30 There is a major design bug with ocelot-8021q, which is that it expects more of the hardware than the hardware can actually do. The short summary of the issue is that when a port is under a VLAN-aware bridge and we use this tagging protocol, VLAN upper interfaces of this port do not see RX traffic. We use VCAP ES0 (egress rewriter) rules towards the tag_8021q CPU port to encapsulate packets with an outer tag, later stripped by software, that depends on the source user port. We do this so that packets can be identified in ocelot_rcv(). To be precise, we create rules with push_outer_tag = OCELOT_ES0_TAG and push_inner_tag = 0. With this configuration, we expect the switch to keep the inner tag configuration as found in the packet (if it was untagged on user port ingress, keep it untagged, otherwise preserve the VLAN tag unmodified as the inner tag towards the tag_8021q CPU port). But this is not what happens. Instead, table "Tagging Combinations" from the user manual suggests that when the ES0 action is "PUSH_OUTER_TAG=1 and PUSH_INNER_TAG=0", there will be "no inner tag". Experimentation further clarifies what this means. It appears that this "inner tag" which is not pushed into the packet on its egress towards the CPU is none other than the classified VLAN. When the ingress user port is standalone or under a VLAN-unaware bridge, the classified VLAN is a discardable quantity: it is a fixed value - the result of ocelot_vlan_unaware_pvid()'s configuration, and actually independent of the VID from any 802.1Q header that may be in the frame. It is actually preferable to discard the "inner tag" in this case. The problem is when the ingress port is under a VLAN-aware bridge. Then, the classified VLAN is taken from the frame's 802.1Q header, with a fallback on the bridge port's PVID. It would be very good to not discard the "inner tag" here, because if we do, we break communication with any 8021q VLAN uppers that the port might have. These have a processing path outside the bridge. There seems to be nothing else we can do except to change the configuration for VCAP ES0 rules, to actually push the inner VLAN into the frame. There are 2 options for that, first is to push a fixed value specified in the rule, and second is to push a fixed value, plus (aka arithmetic +) the classified VLAN. We choose the second option, and we select that fixed value as 0. Thus, what is pushed in the inner tag is just the classified VLAN. From there, we need to perform software untagging, in the receive path, of stuff that was untagged on the wire. Fixes: 7c83a7c539ab ("net: dsa: add a second tagger for Ocelot switches based on tag_8021q") Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> Signed-off-by: David S. Miller <davem@davemloft.net> 
There is a major design bug with ocelot-8021q, which is that it expects
more of the hardware than the hardware can actually do. The short
summary of the issue is that when a port is under a VLAN-aware bridge
and we use this tagging protocol, VLAN upper interfaces of this port do
not see RX traffic.

We use VCAP ES0 (egress rewriter) rules towards the tag_8021q CPU port
to encapsulate packets with an outer tag, later stripped by software,
that depends on the source user port. We do this so that packets can be
identified in ocelot_rcv(). To be precise, we create rules with
push_outer_tag = OCELOT_ES0_TAG and push_inner_tag = 0.

With this configuration, we expect the switch to keep the inner tag
configuration as found in the packet (if it was untagged on user port
ingress, keep it untagged, otherwise preserve the VLAN tag unmodified
as the inner tag towards the tag_8021q CPU port). But this is not what
happens.

Instead, table "Tagging Combinations" from the user manual suggests
that when the ES0 action is "PUSH_OUTER_TAG=1 and PUSH_INNER_TAG=0",
there will be "no inner tag". Experimentation further clarifies what
this means.

It appears that this "inner tag" which is not pushed into the packet on
its egress towards the CPU is none other than the classified VLAN.

When the ingress user port is standalone or under a VLAN-unaware bridge,
the classified VLAN is a discardable quantity: it is a fixed value - the
result of ocelot_vlan_unaware_pvid()'s configuration, and actually
independent of the VID from any 802.1Q header that may be in the frame.
It is actually preferable to discard the "inner tag" in this case.

The problem is when the ingress port is under a VLAN-aware bridge.
Then, the classified VLAN is taken from the frame's 802.1Q header, with
a fallback on the bridge port's PVID. It would be very good to not
discard the "inner tag" here, because if we do, we break communication
with any 8021q VLAN uppers that the port might have. These have a
processing path outside the bridge.

There seems to be nothing else we can do except to change the
configuration for VCAP ES0 rules, to actually push the inner VLAN into
the frame. There are 2 options for that, first is to push a fixed value
specified in the rule, and second is to push a fixed value, plus
(aka arithmetic +) the classified VLAN. We choose the second option,
and we select that fixed value as 0. Thus, what is pushed in the inner
tag is just the classified VLAN.

From there, we need to perform software untagging, in the receive path,
of stuff that was untagged on the wire.

Fixes: 7c83a7c539ab ("net: dsa: add a second tagger for Ocelot switches based on tag_8021q")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net: mscc: ocelot: serialize access to the injection/extraction groups 2024-08-16T08:59:32+00:00 Vladimir Oltean vladimir.oltean@nxp.com 2024-08-15T00:07:04+00:00 c5e12ac3beb0dd3a718296b2d8af5528e9ab728e As explained by Horatiu Vultur in commit 603ead96582d ("net: sparx5: Add spinlock for frame transmission from CPU") which is for a similar hardware design, multiple CPUs can simultaneously perform injection or extraction. There are only 2 register groups for injection and 2 for extraction, and the driver only uses one of each. So we'd better serialize access using spin locks, otherwise frame corruption is possible. Note that unlike in sparx5, FDMA in ocelot does not have this issue because struct ocelot_fdma_tx_ring already contains an xmit_lock. I guess this is mostly a problem for NXP LS1028A, as that is dual core. I don't think VSC7514 is. So I'm blaming the commit where LS1028A (aka the felix DSA driver) started using register-based packet injection and extraction. Fixes: 0a6f17c6ae21 ("net: dsa: tag_ocelot_8021q: add support for PTP timestamping") Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> Signed-off-by: David S. Miller <davem@davemloft.net> 
As explained by Horatiu Vultur in commit 603ead96582d ("net: sparx5: Add
spinlock for frame transmission from CPU") which is for a similar
hardware design, multiple CPUs can simultaneously perform injection
or extraction. There are only 2 register groups for injection and 2
for extraction, and the driver only uses one of each. So we'd better
serialize access using spin locks, otherwise frame corruption is
possible.

Note that unlike in sparx5, FDMA in ocelot does not have this issue
because struct ocelot_fdma_tx_ring already contains an xmit_lock.

I guess this is mostly a problem for NXP LS1028A, as that is dual core.
I don't think VSC7514 is. So I'm blaming the commit where LS1028A (aka
the felix DSA driver) started using register-based packet injection and
extraction.

Fixes: 0a6f17c6ae21 ("net: dsa: tag_ocelot_8021q: add support for PTP timestamping")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>