| Message ID | 20231122034420.1158898-1-kuba@kernel.org (mailing list archive) |
|---|---|
| Headers | show |
| Series | net: page_pool: add netlink-based introspection | expand |
On Tue, Nov 21, 2023 at 07:44:07PM -0800, Jakub Kicinski wrote: > We recently started to deploy newer kernels / drivers at Meta, > making significant use of page pools for the first time. > We immediately run into page pool leaks both real and false positive > warnings. As Eric pointed out/predicted there's no guarantee that > applications will read / close their sockets so a page pool page > may be stuck in a socket (but not leaked) forever. This happens > a lot in our fleet. Most of these are obviously due to application > bugs but we should not be printing kernel warnings due to minor > application resource leaks. > > Conversely the page pool memory may get leaked at runtime, and > we have no way to detect / track that, unless someone reconfigures > the NIC and destroys the page pools which leaked the pages. > > The solution presented here is to expose the memory use of page > pools via netlink. This allows for continuous monitoring of memory > used by page pools, regardless if they were destroyed or not. > Sample in patch 15 can print the memory use and recycling > efficiency: > > $ ./page-pool > eth0[2] page pools: 10 (zombies: 0) > refs: 41984 bytes: 171966464 (refs: 0 bytes: 0) > recycling: 90.3% (alloc: 656:397681 recycle: 89652:270201) Hi Jakub, I am wondering if you considered to expose these metrics through meminfo/vmstat as well, is that a bad idea or is this/netlink more of a preference? thanks, Shakeel
On Sat, 25 Nov 2023 20:57:24 +0000 Shakeel Butt wrote: > > $ ./page-pool > > eth0[2] page pools: 10 (zombies: 0) > > refs: 41984 bytes: 171966464 (refs: 0 bytes: 0) > > recycling: 90.3% (alloc: 656:397681 recycle: 89652:270201) > > Hi Jakub, I am wondering if you considered to expose these metrics > through meminfo/vmstat as well, is that a bad idea or is this/netlink > more of a preference? If that's net-namespaced we can add the basics there. We'll still need the netlink interface, tho, it's currently per-interface and per-queue (simplifying a bit). But internally the recycling stats are also per-CPU, which could be of interest at some stage.
On Sun, Nov 26, 2023 at 02:43:00PM -0800, Jakub Kicinski wrote: > On Sat, 25 Nov 2023 20:57:24 +0000 Shakeel Butt wrote: > > > $ ./page-pool > > > eth0[2] page pools: 10 (zombies: 0) > > > refs: 41984 bytes: 171966464 (refs: 0 bytes: 0) > > > recycling: 90.3% (alloc: 656:397681 recycle: 89652:270201) > > > > Hi Jakub, I am wondering if you considered to expose these metrics > > through meminfo/vmstat as well, is that a bad idea or is this/netlink > > more of a preference? > > If that's net-namespaced we can add the basics there. We'll still need > the netlink interface, tho, it's currently per-interface and per-queue > (simplifying a bit). But internally the recycling stats are also > per-CPU, which could be of interest at some stage. Not really net-namespaced but rather system level stats in those interfaces. Anyways if having system level makes sense then these stats can be added later.
We recently started to deploy newer kernels / drivers at Meta, making significant use of page pools for the first time. We immediately run into page pool leaks both real and false positive warnings. As Eric pointed out/predicted there's no guarantee that applications will read / close their sockets so a page pool page may be stuck in a socket (but not leaked) forever. This happens a lot in our fleet. Most of these are obviously due to application bugs but we should not be printing kernel warnings due to minor application resource leaks. Conversely the page pool memory may get leaked at runtime, and we have no way to detect / track that, unless someone reconfigures the NIC and destroys the page pools which leaked the pages. The solution presented here is to expose the memory use of page pools via netlink. This allows for continuous monitoring of memory used by page pools, regardless if they were destroyed or not. Sample in patch 15 can print the memory use and recycling efficiency: $ ./page-pool eth0[2] page pools: 10 (zombies: 0) refs: 41984 bytes: 171966464 (refs: 0 bytes: 0) recycling: 90.3% (alloc: 656:397681 recycle: 89652:270201) v3: - ID is still here, can't decide if it matters - rename destroyed -> detach-time, good enough? - fix build for netsec v2: https://lore.kernel.org/r/20231121000048.789613-1-kuba@kernel.org - hopefully fix build with PAGE_POOL=n v1: https://lore.kernel.org/all/20231024160220.3973311-1-kuba@kernel.org/ - The main change compared to the RFC is that the API now exposes outstanding references and byte counts even for "live" page pools. The warning is no longer printed if page pool is accessible via netlink. RFC: https://lore.kernel.org/all/20230816234303.3786178-1-kuba@kernel.org/ Jakub Kicinski (13): net: page_pool: factor out uninit net: page_pool: id the page pools net: page_pool: record pools per netdev net: page_pool: stash the NAPI ID for easier access eth: link netdev to page_pools in drivers net: page_pool: add nlspec for basic access to page pools net: page_pool: implement GET in the netlink API net: page_pool: add netlink notifications for state changes net: page_pool: report amount of memory held by page pools net: page_pool: report when page pool was destroyed net: page_pool: expose page pool stats via netlink net: page_pool: mute the periodic warning for visible page pools tools: ynl: add sample for getting page-pool information Documentation/netlink/specs/netdev.yaml | 170 +++++++ Documentation/networking/page_pool.rst | 10 +- drivers/net/ethernet/broadcom/bnxt/bnxt.c | 1 + .../net/ethernet/mellanox/mlx5/core/en_main.c | 1 + drivers/net/ethernet/microsoft/mana/mana_en.c | 1 + drivers/net/ethernet/socionext/netsec.c | 2 + include/linux/list.h | 20 + include/linux/netdevice.h | 4 + include/linux/poison.h | 2 + include/net/page_pool/helpers.h | 8 +- include/net/page_pool/types.h | 10 + include/uapi/linux/netdev.h | 36 ++ net/core/Makefile | 2 +- net/core/netdev-genl-gen.c | 60 +++ net/core/netdev-genl-gen.h | 11 + net/core/page_pool.c | 69 ++- net/core/page_pool_priv.h | 12 + net/core/page_pool_user.c | 414 +++++++++++++++++ tools/include/uapi/linux/netdev.h | 36 ++ tools/net/ynl/generated/netdev-user.c | 419 ++++++++++++++++++ tools/net/ynl/generated/netdev-user.h | 171 +++++++ tools/net/ynl/lib/ynl.h | 2 +- tools/net/ynl/samples/.gitignore | 1 + tools/net/ynl/samples/Makefile | 2 +- tools/net/ynl/samples/page-pool.c | 147 ++++++ 25 files changed, 1578 insertions(+), 33 deletions(-) create mode 100644 net/core/page_pool_priv.h create mode 100644 net/core/page_pool_user.c create mode 100644 tools/net/ynl/samples/page-pool.c