- 06 Mar, 2024 21 commits
-
-
Niklas Söderlund authored
The EtherAVB device requires the SKB data to be aligned to 128 bytes. The alignment is done by allocating an skb 128 bytes larger than the maximum frame size supported by the device and adjusting the headroom to fit the requirement. This code has been refactored a few times and small issues have been added along the way. The issues are not harmful but prevent merging parts of the Rx code which have been split in two implementations with the addition of RZ/G2L support, a device that supports larger frame sizes. This change removes the need for duplicated and somewhat inaccurate hardware alignment constrains stored in the hardware information struct by creating a helper to handle the allocation of an skb and alignment of an skb data. For the R-Car class of devices the maximum frame size is 4K and each descriptor is limited to 2K of data. The current implementation does not support split descriptors, this limits the frame size to 2K. The current hardware information however records the descriptor size just under 2K due to bad understanding of the device when larger MTUs where added. For the RZ/G2L device the maximum frame size is 8K and each descriptor is limited to 4K of data. The current hardware information records this correctly, but it gets the alignment constrains wrong as just aligns it by 128, it does not extend it by 128 bytes to allow the full frame to be stored. This works because the RZ/G2L device supports split descriptors and allocates each skb to 8K and aligns each 4K descriptor in this space. Signed-off-by:
Niklas Söderlund <niklas.soderlund+renesas@ragnatech.se> Reviewed-by:
Paul Barker <paul.barker.ct@bp.renesas.com> Reviewed-by:
Sergey Shtylyov <s.shtylyov@omp.ru> Signed-off-by:
David S. Miller <davem@davemloft.net>
-
Niklas Söderlund authored
The struct member rx_max_buf_size was added before split descriptor support was added. It is unclear if the value describes the full skb frame buffer or the data descriptor buffer which can be combined into a single skb. Rename it to make it clear it referees to the maximum frame size and can cover multiple descriptors. Signed-off-by:
-
Niklas Söderlund authored
The Rx ring can either be made up of normal or extended descriptors, not a mix of the two at the same time. Make this explicit by grouping the two variables in a rx_ring union. The extension of the storage for more than one queue of normal descriptors from a single to NUM_RX_QUEUE queues have no practical effect. But aids in making the code readable as the code that uses it already piggyback on other members of struct ravb_private that are arrays of max length NUM_RX_QUEUE, e.g. rx_desc_dma. This will also make further refactoring easier. While at it, rename the normal descriptor Rx ring to make it clear it's not strictly related to the GbEthernet E-MAC IP found in RZ/G2L, normal descriptors could be used on R-Car SoCs too. Signed-off-by:
-
git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/next-queueDavid S. Miller authored
From: Tony Nguyen <anthony.l.nguyen@intel.com> To: davem@davemloft.net, kuba@kernel.org, pabeni@redhat.com, edumazet@google.com, netdev@vger.kernel.org Cc: Tony Nguyen <anthony.l.nguyen@intel.com>, alan.brady@intel.com Tony Nguyen says: ==================== idpf: refactor virtchnl messages Alan Brady says: The motivation for this series has two primary goals. We want to enable support of multiple simultaneous messages and make the channel more robust. The way it works right now, the driver can only send and receive a single message at a time and if something goes really wrong, it can lead to data corruption and strange bugs. To start the series, we introduce an idpf_virtchnl.h file. This reduces the burden on idpf.h which is overloaded with struct and function declarations. The conversion works by conceptualizing a send and receive as a "virtchnl transaction" (idpf_vc_xn) and introducing a "transaction manager" (idpf_vc_xn_manager). The vcxn_mngr will init a ring of transactions from which the driver will pop from a bitmap of free transactions to track in-flight messages. Instead of needing to handle a complicated send/recv for every a message, the driver now just needs to fill out a xn_params struct and hand it over to idpf_vc_xn_exec which will take care of all the messy bits. Once a message is sent and receives a reply, we leverage the completion API to signal the received buffer is ready to be used (assuming success, or an error code otherwise). At a low-level, this implements the "sw cookie" field of the virtchnl message descriptor to enable this. We have 16 bits we can put whatever we want and the recipient is required to apply the same cookie to the reply for that message. We use the first 8 bits as an index into the array of transactions to enable fast lookups and we use the second 8 bits as a salt to make sure each cookie is unique for that message. As transactions are received in arbitrary order, it's possible to reuse a transaction index and the salt guards against index conflicts to make certain the lookup is correct. As a primitive example, say index 1 is used with salt 1. The message times out without receiving a reply so index 1 is renewed to be ready for a new transaction, we report the timeout, and send the message again. Since index 1 is free to be used again now, index 1 is again sent but now salt is 2. This time we do get a reply, however it could be that the reply is _actually_ for the previous send index 1 with salt 1. Without the salt we would have no way of knowing for sure if it's the correct reply, but with we will know for certain. Through this conversion we also get several other benefits. We can now more appropriately handle asynchronously sent messages by providing space for a callback to be defined. This notably allows us to handle MAC filter failures better; previously we could potentially have stale, failed filters in our list, which shouldn't really have a major impact but is obviously not correct. I also managed to remove fairly significant more lines than I added which is a win in my book. Additionally, this converts some variables to use auto-variables where appropriate. This makes the alloc paths much cleaner and less prone to memory leaks. We also fix a few virtchnl related bugs while we're here. ==================== Signed-off-by: -
David S. Miller authored
Jakub Kicinski says: ==================== netlink: handle EMSGSIZE errors in the core Ido discovered some time back that we usually force NLMSG_DONE to be delivered in a separate recv() syscall, even if it would fit into the same skb as data messages. He made nexthop try to fit DONE with data in commit 8743aeff ("nexthop: Fix infinite nexthop bucket dump when using maximum nexthop ID"), and nobody has complained so far. We have since also tried to follow the same pattern in new genetlink families, but explaining to people, or even remembering the correct handling ourselves is tedious. Let the netlink socket layer consume -EMSGSIZE errors. Practically speaking most families use this error code as "dump needs more space", anyway. v2: - init err to 0 in last patch v1: https://lore.kernel.org/all/20240301012845.2951053-1-kuba@kernel.org/ ==================== Signed-off-by:
-
Jakub Kicinski authored
Make sure ctrl_fill_info() returns sensible error codes and propagate them out to netlink core. Let netlink core decide when to return skb->len and when to treat the exit as an error. Netlink core does better job at it, if we always return skb->len the core doesn't know when we're done dumping and NLMSG_DONE ends up in a separate read(). Reviewed-by:
Eric Dumazet <edumazet@google.com> Signed-off-by:
Jakub Kicinski <kuba@kernel.org> Reviewed-by:
Ido Schimmel <idosch@nvidia.com> Signed-off-by:
-
Jakub Kicinski authored
Previous change added -EMSGSIZE handling to af_netlink, we don't have to hide these errors any longer. Theoretically the error handling changes from: if (err == -EMSGSIZE) to if (err == -EMSGSIZE && skb->len) everywhere, but in practice it doesn't matter. All messages fit into NLMSG_GOODSIZE, so overflow of an empty skb cannot happen. Reviewed-by:
-
Jakub Kicinski authored
Eric points out that our current suggested way of handling EMSGSIZE errors ((err == -EMSGSIZE) ? skb->len : err) will break if we didn't fit even a single object into the buffer provided by the user. This should not happen for well behaved applications, but we can fix that, and free netlink families from dealing with that completely by moving error handling into the core. Let's assume from now on that all EMSGSIZE errors in dumps are because we run out of skb space. Families can now propagate the error nla_put_*() etc generated and not worry about any return value magic. If some family really wants to send EMSGSIZE to user space, assuming it generates the same error on the next dump iteration the skb->len should be 0, and user space should still see the EMSGSIZE. This should simplify families and prevent mistakes in return values which lead to DONE being forced into a separate recv() call as discovered by Ido some time ago. Reviewed-by:
-
Jakub Kicinski authored
selftest harness uses various exit codes to signal test results. Avoid calling exit() directly, otherwise tests may get broken by harness refactoring (like the commit under Fixes). SKIP() will instruct the harness that the test shouldn't run, it used to not be the case, but that has been fixed. So just return, no need to exit. Note that for hmm-tests this actually changes the result from pass to skip. Which seems fair, the test is skipped, after all. Reported-by:
Mark Brown <broonie@kernel.org> Link: https://lore.kernel.org/all/05f7bf89-04a5-4b65-bf59-c19456aeb1f0@sirena.org.uk Fixes: a7247079 ("selftests: kselftest_harness: use KSFT_* exit codes") Reviewed-by:
Kees Cook <keescook@chromium.org> Reviewed-by:
Przemek Kitszel <przemyslaw.kitszel@intel.com> Link: https://lore.kernel.org/r/20240304233621.646054-1-kuba@kernel.orgSigned-off-by:
-
Jakub Kicinski authored
Oleksij Rempel says: ==================== net: ethernet: Rework EEE with Andrew's permission I'll continue mainlining this patches: ============================================================== Most MAC drivers get EEE wrong. The API to the PHY is not very obvious, which is probably why. Rework the API, pushing most of the EEE handling into phylib core, leaving the MAC drivers to just enable/disable support for EEE in there change_link call back. MAC drivers are now expect to indicate to phylib if they support EEE. This will allow future patches to configure the PHY to advertise no EEE link modes when EEE is not supported. The information could also be used to enable SmartEEE if the PHY supports it. With these changes, the uAPI configuration eee_enable becomes a global on/off. tx-lpi must also be enabled before EEE is enabled. This fits the discussion here: https://lore.kernel.org/netdev/af880ce8-a7b8-138e-1ab9-8c89e662eecf@gmail.com/T/ This patchset puts in place all the infrastructure, and converts one MAC driver to the new API. Following patchsets will convert other MAC drivers, extend support into phylink, and when all MAC drivers are converted to the new scheme, clean up some unneeded code. ==================== Link: https://lore.kernel.org/r/20240302195306.3207716-1-o.rempel@pengutronix.deSigned-off-by:
-
Andrew Lunn authored
The enabling/disabling of EEE in the MAC should happen as a result of auto negotiation. So move the enable/disable into fec_enet_adjust_link() which gets called by phylib when there is a change in link status. fec_enet_set_eee() now just stores away the LPI timer value. Everything else is passed to phylib, so it can correctly setup the PHY. fec_enet_get_eee() relies on phylib doing most of the work, the MAC driver just adds the LPI timer value. Call phy_support_eee() if the quirk is present to indicate the MAC actually supports EEE. Signed-off-by:
Andrew Lunn <andrew@lunn.ch> Tested-by: Oleksij Rempel <o.rempel@pengutronix.de> (On iMX8MP debix) Signed-off-by:
Oleksij Rempel <o.rempel@pengutronix.de> Reviewed-by:
Wei Fang <wei.fang@nxp.com> Link: https://lore.kernel.org/r/20240302195306.3207716-8-o.rempel@pengutronix.deSigned-off-by:
-
Andrew Lunn authored
FEC is about to get its EEE code re-written. To allow this, move fec_enet_eee_mode_set() before fec_enet_adjust_link() which will need to call it. Signed-off-by:
-
Andrew Lunn authored
In order for EEE to operate, both the MAC and the PHY need to support it, similar to how pause works. With some exception - a number of PHYs have SmartEEE or AutoGrEEEn support in order to provide some EEE-like power savings with non-EEE capable MACs. Copy the pause concept and add the call phy_support_eee() which the MAC makes after connecting the PHY to indicate it supports EEE. phylib will then advertise EEE when auto-neg is performed. Signed-off-by:
-
Andrew Lunn authored
The MAC driver changes its EEE hardware configuration in its adjust_link callback. This is called when auto-neg completes. Disabling EEE via eee_enabled false will trigger an autoneg, and as a result the adjust_link callback will be called with phydev->enable_tx_lpi set to false. Similarly, eee_enabled set to true and with a change of advertised link modes will result in a new autoneg, and a call the adjust_link call. If set_eee is called with only a change to tx_lpi_enabled which does not trigger an auto-neg, it is necessary to call the adjust_link callback so that the MAC is reconfigured to take this change into account. When setting phydev->enable_tx_lpi, take both eee_enabled and tx_lpi_enabled into account, so the MAC drivers just needs to act on phydev->enable_tx_lpi and not the whole EEE configuration. The same check should be done for tx_lpi_timer too. Signed-off-by:
Florian Fainelli <florian.fainelli@broadcom.com> Signed-off-by:
-
Andrew Lunn authored
Have phylib keep track of the EEE configuration. This simplifies the MAC drivers, in that they don't need to store it. Future patches to phylib will also make use of this information to further simplify the MAC drivers. Reviewed-by:
Russell King (Oracle) <rmk+kernel@armlinux.org.uk> Signed-off-by:
-
Andrew Lunn authored
MAC drivers which support EEE need to know the results of the EEE auto-neg in order to program the hardware to perform EEE or not. The oddly named phy_init_eee() can be used to determine this, it returns 0 if EEE should be used, or a negative error code, e.g. -EOPPROTONOTSUPPORT if the PHY does not support EEE or negotiate resulted in it not being used. However, many MAC drivers get this wrong. Add phydev->enable_tx_lpi which indicates the result of the autoneg for EEE, including if EEE is administratively disabled with ethtool. The MAC driver can then access this in the same way as link speed and duplex in the adjust link callback. If enable_tx_lpi is true, the MAC should send low power indications and does not need to consider anything else with respect to EEE. Reviewed-by:
-
Russell King authored
Add helpers that phylib and phylink can use to manage EEE configuration and determine whether the MAC should be permitted to use LPI based on that configuration. Signed-off-by:
-
Heiner Kallweit authored
This function is used with the set_eee() ethtool operation. Certain fields of struct ethtool_keee() are relevant only for the get_eee() operation. In addition, in case of the ioctl interface, we have no guarantee that userspace sends sane values in struct ethtool_eee. Therefore explicitly ignore all fields not needed for set_eee(). This protects from drivers trying to use unchecked and unreliable data, relying on specific userspace behavior. Note: Such unsafe driver behavior has been found and fixed in the tg3 driver. Signed-off-by:
Heiner Kallweit <hkallweit1@gmail.com> Reviewed-by:
-
Kees Cook authored
While testing for places where zero-sized destinations were still showing up in the kernel, sock_copy() and inet_reqsk_clone() were found, which are using very specific memcpy() offsets for both avoiding a portion of struct sock, and copying beyond the end of it (since struct sock is really just a common header before the protocol-specific allocation). Instead of trying to unravel this historical lack of container_of(), just switch to unsafe_memcpy(), since that's effectively what was happening already (memcpy() wasn't checking 0-sized destinations while the code base was being converted away from fake flexible arrays). Avoid the following false positive warning with future changes to CONFIG_FORTIFY_SOURCE: memcpy: detected field-spanning write (size 3068) of destination "&nsk->__sk_common.skc_dontcopy_end" at net/core/sock.c:2057 (size 0) Signed-off-by:
Simon Horman <horms@kernel.org> Link: https://lore.kernel.org/r/20240304212928.make.772-kees@kernel.orgSigned-off-by:
-
Breno Leitao authored
Commit 3e2f544d ("net: get stats64 if device if driver is configured") moved the callback to dev_get_tstats64() to net core, so, unless the driver is doing some custom stats collection, it does not need to set .ndo_get_stats64. Since this driver is now relying in NETDEV_PCPU_STAT_TSTATS, then, it doesn't need to set the dev_get_tstats64() generic .ndo_get_stats64 function pointer. Signed-off-by:
Breno Leitao <leitao@debian.org> Reviewed-by:
Willem de Bruijn <willemb@google.com> Link: https://lore.kernel.org/r/20240304183810.1474883-2-leitao@debian.orgSigned-off-by:
-
Breno Leitao authored
With commit 34d21de9 ("net: Move {l,t,d}stats allocation to core and convert veth & vrf"), stats allocation could be done on net core instead of in this driver. With this new approach, the driver doesn't have to bother with error handling (allocation failure checking, making sure free happens in the right spot, etc). This is core responsibility now. Remove the allocation in the tun/tap driver and leverage the network core allocation instead. Signed-off-by:
-
- 05 Mar, 2024 19 commits
-
-
Uwe Kleine-König authored
The .remove() callback for a platform driver returns an int which makes many driver authors wrongly assume it's possible to do error handling by returning an error code. However the value returned is ignored (apart from emitting a warning) and this typically results in resource leaks. To improve here there is a quest to make the remove callback return void. In the first step of this quest all drivers are converted to .remove_new(), which already returns void. Eventually after all drivers are converted, .remove_new() will be renamed to .remove(). Trivially convert this driver from always returning zero in the remove callback to the void returning variant. Signed-off-by:
Uwe Kleine-König <u.kleine-koenig@pengutronix.de> Reviewed-by:
-
Jakub Kicinski authored
Ricardo B. Marliere says: ==================== net: constify struct class usage This is a simple and straight forward cleanup series that aims to make the class structures in net constant. This has been possible since 2023 [1]. [1]: https://lore.kernel.org/all/2023040248-customary-release-4aec@gregkh/ ==================== Link: https://lore.kernel.org/r/20240302-class_cleanup-net-next-v1-0-8fa378595b93@marliere.netSigned-off-by:
-
Ricardo B. Marliere authored
Since commit 43a7206b ("driver core: class: make class_register() take a const *"), the driver core allows for struct class to be in read-only memory, so move the nfc_class structure to be declared at build time placing it into read-only memory, instead of having to be dynamically allocated at boot time. Suggested-by:
Greg Kroah-Hartman <gregkh@linuxfoundation.org> Signed-off-by:
Ricardo B. Marliere <ricardo@marliere.net> Reviewed-by:
-
Ricardo B. Marliere authored
Since commit 43a7206b ("driver core: class: make class_register() take a const *"), the driver core allows for struct class to be in read-only memory, so move the wwan_class structure to be declared at build time placing it into read-only memory, instead of having to be dynamically allocated at boot time. Suggested-by:
Sergey Ryazanov <ryazanov.s.a@gmail.com> Link: https://lore.kernel.org/r/20240302-class_cleanup-net-next-v1-5-8fa378595b93@marliere.netSigned-off-by:
-
Ricardo B. Marliere authored
Since commit 43a7206b ("driver core: class: make class_register() take a const *"), the driver core allows for struct class to be in read-only memory, so move the wwan_hwsim_class structure to be declared at build time placing it into read-only memory, instead of having to be dynamically allocated at boot time. Suggested-by:
-
Ricardo B. Marliere authored
Since commit 43a7206b ("driver core: class: make class_register() take a const *"), the driver core allows for struct class to be in read-only memory, so move the ppp_class structure to be declared at build time placing it into read-only memory, instead of having to be dynamically allocated at boot time. Suggested-by:
-
Ricardo B. Marliere authored
Since commit 43a7206b ("driver core: class: make class_register() take a const *"), the driver core allows for struct class to be in read-only memory, so move the framer_class structure to be declared at build time placing it into read-only memory, instead of having to be dynamically allocated at boot time. Suggested-by:
Herve Codina <herve.codina@bootlin.com> Link: https://lore.kernel.org/r/20240302-class_cleanup-net-next-v1-2-8fa378595b93@marliere.netSigned-off-by:
-
Ricardo B. Marliere authored
Since commit 43a7206b ("driver core: class: make class_register() take a const *"), the driver core allows for struct class to be in read-only memory, so move the hnae_class structure to be declared at build time placing it into read-only memory, instead of having to be dynamically allocated at boot time. Suggested-by:
-
Jakub Kicinski authored
Horatiu Vultur says: ==================== net: phy: micrel: lan8814 erratas Add two erratas for lan8814. First one fix the led which might stay on even that there is no link. The second one improves increases length of the cable that can be used when used in 1000Base-T. ==================== Link: https://lore.kernel.org/r/20240304091548.1386022-1-horatiu.vultur@microchip.comSigned-off-by:
-
Horatiu Vultur authored
When the length of the cable is more than 100m and the lan8814 is configured to run in 1000Base-T Slave then the register of the device needs to be optimized. Workaround this by setting the measure time to a value of 0xb. This value can be set regardless of the configuration. This issue is described in 'LAN8814 Silicon Errata and Data Sheet Clarification' and according to that, this will not be corrected in a future silicon revision. Reviewed-by:
Wojciech Drewek <wojciech.drewek@intel.com> Signed-off-by:
Horatiu Vultur <horatiu.vultur@microchip.com> Reviewed-by:
Arun Ramadoss <arun.ramadoss@microchip.com> Link: https://lore.kernel.org/r/20240304091548.1386022-3-horatiu.vultur@microchip.comSigned-off-by:
-
Horatiu Vultur authored
Lan8814 phy led behavior is not correct. It was noticed that the led still remains ON when the cable is unplugged while there was traffic passing at that time. The fix consists in clearing bit 10 of register 0x38, in this way the led behaviour is correct and gets OFF when there is no link. Reviewed-by:
-
Jakub Kicinski authored
Ido Schimmel says: ==================== selftests: forwarding: Various improvements This patchset speeds up the multipath tests (patches #1-#2) and makes other tests more stable (patches #3-#6) so that they will not randomly fail in the netdev CI. On my system, after applying the first two patches, the run time of gre_multipath_nh_res.sh is reduced by over 90%. ==================== Link: https://lore.kernel.org/r/20240304095612.462900-1-idosch@nvidia.comSigned-off-by:
-
Ido Schimmel authored
These tests generate various IPv6 flows, encapsulate them in GRE packets and check that the encapsulated packets are distributed between the available nexthops according to the configured weights. Unlike the corresponding IPv4 tests, these tests sometimes fail in the netdev CI because of large discrepancies between the expected and measured ratios [1]. This can be explained by the fact that the IPv4 tests generate about 3,600 different flows whereas the IPv6 tests only generate about 784 different flows (potentially by mistake). Fix by aligning the IPv6 tests to the IPv4 ones and increase the number of generated flows. [1] [...] # TEST: ping [ OK ] # INFO: Running IPv6 over GRE over IPv4 multipath tests # TEST: ECMP [FAIL] # Too large discrepancy between expected and measured ratios # INFO: Expected ratio 1.00 Measured ratio 1.18 [...] Signed-off-by:
-
Ido Schimmel authored
These tests sometimes fail on the netdev CI because the expected number of packets is larger than expected [1]. Make the tests more robust by specifically matching on VXLAN encapsulated packets and allowing up to five stray packets instead of just two. [1] [...] # TEST: VXLAN: ECN encap: 0x00->0x00 [FAIL] # v1: Expected to capture 10 packets, got 13. # TEST: VXLAN: ECN encap: 0x01->0x01 [ OK ] # TEST: VXLAN: ECN encap: 0x02->0x02 [ OK ] # TEST: VXLAN: ECN encap: 0x03->0x02 [ OK ] [...] Signed-off-by:
-
Ido Schimmel authored
The ageing time used by the test is too short for debug kernels and results in entries being aged out prematurely [1]. Fix by increasing the ageing time. [1] # ./vxlan_bridge_1q.sh [...] INFO: learning vlan 10 TEST: VXLAN: flood before learning [ OK ] TEST: VXLAN: show learned FDB entry [ OK ] TEST: VXLAN: learned FDB entry [FAIL] swp4: Expected to capture 0 packets, got 10. RTNETLINK answers: No such file or directory TEST: VXLAN: deletion of learned FDB entry [ OK ] TEST: VXLAN: Ageing of learned FDB entry [FAIL] swp4: Expected to capture 0 packets, got 10. TEST: VXLAN: learning toggling on bridge port [ OK ] [...] Signed-off-by: -
Ido Schimmel authored
The test configures a policer with a rate of 80Mbps and expects to measure a rate close to it. This is a too high rate for debug kernels, causing the test to fail [1]. Fix by reducing the rate to 10Mbps. [1] # ./tc_police.sh TEST: police on rx [FAIL] Expected rate 76.2Mbps, got 29.6Mbps, which is -61% off. Required accuracy is +-10%. TEST: police on tx [FAIL] Expected rate 76.2Mbps, got 30.4Mbps, which is -60% off. Required accuracy is +-10%. Signed-off-by: -
Ido Schimmel authored
The various multipath tests use mausezahn to generate different flows and check how they are distributed between the available nexthops. The tool is currently invoked with an hard coded transmission delay of 1 ms. This is unnecessary when the tests are run with veth pairs and needlessly prolongs the tests. Parametrize this delay and default it to 0 us. It can be overridden using the forwarding.config file. On my system, this reduces the run time of router_multipath.sh by 93%. Signed-off-by:
-
Ido Schimmel authored
The multipath tests currently test both the L3 and L4 multipath hash policies for IPv6, but only the L4 policy for IPv4. The reason is mostly historic: When the initial multipath test was added (router_multipath.sh) the IPv6 L4 policy did not exist and was later added to the test. The other multipath tests copied this pattern although there is little value in testing both policies. Align the IPv4 and IPv6 tests and only test the L4 policy. On my system, this reduces the run time of router_multipath.sh by 89% because of the repeated ping6 invocations to randomize the flow label. Signed-off-by:
-
Eric Dumazet authored
Many syzbot reports show extreme rtnl pressure, and many of them hint that smc acquires rtnl in netns creation for no good reason [1] This patch returns early from smc_pnet_net_init() if there is no netdevice yet. I am not even sure why smc_pnet_create_pnetids_list() even exists, because smc_pnet_netdev_event() is also calling smc_pnet_add_base_pnetid() when handling NETDEV_UP event. [1] extract of typical syzbot reports 2 locks held by syz-executor.3/12252: #0: ffffffff8f369610 (pernet_ops_rwsem){++++}-{3:3}, at: copy_net_ns+0x4c7/0x7b0 net/core/net_namespace.c:491 #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_create_pnetids_list net/smc/smc_pnet.c:809 [inline] #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_net_init+0x10a/0x1e0 net/smc/smc_pnet.c:878 2 locks held by syz-executor.4/12253: #0: ffffffff8f369610 (pernet_ops_rwsem){++++}-{3:3}, at: copy_net_ns+0x4c7/0x7b0 net/core/net_namespace.c:491 #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_create_pnetids_list net/smc/smc_pnet.c:809 [inline] #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_net_init+0x10a/0x1e0 net/smc/smc_pnet.c:878 2 locks held by syz-executor.1/12257: #0: ffffffff8f369610 (pernet_ops_rwsem){++++}-{3:3}, at: copy_net_ns+0x4c7/0x7b0 net/core/net_namespace.c:491 #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_create_pnetids_list net/smc/smc_pnet.c:809 [inline] #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_net_init+0x10a/0x1e0 net/smc/smc_pnet.c:878 2 locks held by syz-executor.2/12261: #0: ffffffff8f369610 (pernet_ops_rwsem){++++}-{3:3}, at: copy_net_ns+0x4c7/0x7b0 net/core/net_namespace.c:491 #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_create_pnetids_list net/smc/smc_pnet.c:809 [inline] #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_net_init+0x10a/0x1e0 net/smc/smc_pnet.c:878 2 locks held by syz-executor.0/12265: #0: ffffffff8f369610 (pernet_ops_rwsem){++++}-{3:3}, at: copy_net_ns+0x4c7/0x7b0 net/core/net_namespace.c:491 #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_create_pnetids_list net/smc/smc_pnet.c:809 [inline] #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_net_init+0x10a/0x1e0 net/smc/smc_pnet.c:878 2 locks held by syz-executor.3/12268: #0: ffffffff8f369610 (pernet_ops_rwsem){++++}-{3:3}, at: copy_net_ns+0x4c7/0x7b0 net/core/net_namespace.c:491 #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_create_pnetids_list net/smc/smc_pnet.c:809 [inline] #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_net_init+0x10a/0x1e0 net/smc/smc_pnet.c:878 2 locks held by syz-executor.4/12271: #0: ffffffff8f369610 (pernet_ops_rwsem){++++}-{3:3}, at: copy_net_ns+0x4c7/0x7b0 net/core/net_namespace.c:491 #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_create_pnetids_list net/smc/smc_pnet.c:809 [inline] #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_net_init+0x10a/0x1e0 net/smc/smc_pnet.c:878 2 locks held by syz-executor.1/12274: #0: ffffffff8f369610 (pernet_ops_rwsem){++++}-{3:3}, at: copy_net_ns+0x4c7/0x7b0 net/core/net_namespace.c:491 #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_create_pnetids_list net/smc/smc_pnet.c:809 [inline] #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_net_init+0x10a/0x1e0 net/smc/smc_pnet.c:878 2 locks held by syz-executor.2/12280: #0: ffffffff8f369610 (pernet_ops_rwsem){++++}-{3:3}, at: copy_net_ns+0x4c7/0x7b0 net/core/net_namespace.c:491 #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_create_pnetids_list net/smc/smc_pnet.c:809 [inline] #1: ffffffff8f375b88 (rtnl_mutex){+.+.}-{3:3}, at: smc_pnet_net_init+0x10a/0x1e0 net/smc/smc_pnet.c:878 Signed-off-by:Wenjia Zhang <wenjia@linux.ibm.com> Link: https://lore.kernel.org/r/20240302100744.3868021-1-edumazet@google.comSigned-off-by:
Paolo Abeni <pabeni@redhat.com>
-
