| Message ID | 20250307155725.219009-1-sdf@fomichev.me (mailing list archive) |
|---|---|
| Headers | show |
| Series | net: remove rtnl_lock from the callers of queue APIs | expand |
On Fri, 7 Mar 2025 07:57:21 -0800 Stanislav Fomichev wrote: > All drivers that use queue management APIs already depend on the netdev > lock. Ultimately, we want to have most of the paths that work with > specific netdev to be rtnl_lock-free (ethtool mostly in particular). > Queue API currently has a much smaller API surface, so start with > rtnl_lock from it: > > - add mutex to each dmabuf binding (to replace rtnl_lock) > - protect global net_devmem_dmabuf_bindings with a new (global) lock > - move netdev lock management to the callers of netdev_rx_queue_restart > and drop rtnl_lock One more note, looks like this silently conflicts with my: https://lore.kernel.org/all/20250307183006.2312761-1-kuba@kernel.org/ You need to add: #include <net/netdev_lock.h> to net/core/netdev_rx_queue.c, otherwise the series together break the build.
On 03/07, Jakub Kicinski wrote: > On Fri, 7 Mar 2025 07:57:21 -0800 Stanislav Fomichev wrote: > > All drivers that use queue management APIs already depend on the netdev > > lock. Ultimately, we want to have most of the paths that work with > > specific netdev to be rtnl_lock-free (ethtool mostly in particular). > > Queue API currently has a much smaller API surface, so start with > > rtnl_lock from it: > > > > - add mutex to each dmabuf binding (to replace rtnl_lock) > > - protect global net_devmem_dmabuf_bindings with a new (global) lock > > - move netdev lock management to the callers of netdev_rx_queue_restart > > and drop rtnl_lock > > One more note, looks like this silently conflicts with my: > https://lore.kernel.org/all/20250307183006.2312761-1-kuba@kernel.org/ > > You need to add: > > #include <net/netdev_lock.h> > > to net/core/netdev_rx_queue.c, otherwise the series together break > the build. Noted, thanks!
All drivers that use queue management APIs already depend on the netdev lock. Ultimately, we want to have most of the paths that work with specific netdev to be rtnl_lock-free (ethtool mostly in particular). Queue API currently has a much smaller API surface, so start with rtnl_lock from it: - add mutex to each dmabuf binding (to replace rtnl_lock) - protect global net_devmem_dmabuf_bindings with a new (global) lock - move netdev lock management to the callers of netdev_rx_queue_restart and drop rtnl_lock Cc: Mina Almasry <almasrymina@google.com> Stanislav Fomichev (4): net: create netdev_nl_sock to wrap bindings list net: protect net_devmem_dmabuf_bindings by new net_devmem_bindings_mutex net: add granular lock for the netdev netlink socket net: drop rtnl_lock for queue_mgmt operations Documentation/netlink/specs/netdev.yaml | 4 +-- drivers/net/ethernet/broadcom/bnxt/bnxt.c | 4 +-- drivers/net/netdevsim/netdev.c | 4 +-- include/net/netdev_netlink.h | 12 +++++++ net/core/devmem.c | 32 ++++++++--------- net/core/netdev-genl-gen.c | 4 +-- net/core/netdev-genl-gen.h | 6 ++-- net/core/netdev-genl.c | 42 ++++++++++------------- net/core/netdev_rx_queue.c | 15 +++----- 9 files changed, 63 insertions(+), 60 deletions(-) create mode 100644 include/net/netdev_netlink.h