Re: [PATCH 0/2] mark some slabs as visible not mergeable
From: Stephen Hemminger Date: Mon, 26 Feb 2018 13:46:13 -0800 > This is ancient original iproute2 code that dumpster dives into > slabinfo to get summary statistics on active objects. > > 1) open sockets (sock_inode_cache) The sockets inuse counter from /proc/net/sockstat is really sufficient for this. > 2) TCP ports bound (tcp_bind_buckets) [*] > 3) TCP time wait sockets (tw_sock_TCP) [*] Time wait is provided by /proc/net/sockstat as well. > 4) TCP syn sockets (request_sock_TCP) [*] It shouldn't be too hard to fill in the last two gaps, maintaining a counter for bind buckets and request socks, and exporting them in new /proc/net/sockstat field. That would be so much better than disabling a useful optimization in the SLAB allocator. Thank you.
Re: [PATCH 0/2] mark some slabs as visible not mergeable
On Mon, 26 Feb 2018 15:15:02 -0500 (EST) David Miller wrote: > From: Stephen Hemminger > Date: Sat, 24 Feb 2018 11:04:52 -0800 > > > This fixes an old bug in iproute2's ss command because it was > > reading slabinfo to get statistics. There isn't a better API > > to do this, and one can argue that /proc is a UAPI that must > > not change. > > Please elaborate what kind of statistics are needed. This is ancient original iproute2 code that dumpster dives into slabinfo to get summary statistics on active objects. 1) open sockets (sock_inode_cache) 2) TCP ports bound (tcp_bind_buckets) [*] 3) TCP time wait sockets (tw_sock_TCP) [*] 4) TCP syn sockets (request_sock_TCP) [*] From man page: -s, --summary Print summary statistics. This option does not parse socket lists obtaining summary from various sources. It is useful when amount of sockets is so huge that parsing /proc/net/tcp is painful. The items with * are currently broken. See 0 for timewait, synrecv, and ports. $ sudo ss -s Total: 1089 (kernel 1093) TCP: 33 (estab 4, closed 1, orphaned 0, synrecv 0, timewait 0/0), ports 0 Transport Total IPIPv6 * 1093 - - RAW 0 0 0 UDP 21138 TCP 32248 INET 533716 FRAG 0 0 0 > > > Therefore this patch set adds a flag to slab to give another > > reason to prevent merging, and then uses it in network code. > > > > The patches are against davem's linux-net tree and should also > > goto stable as well. > > Well, as has been pointed out this never worked with SLUB so > in some sense this was always broken. > > And the "UAPI" of slabinfo is to show the state of the various > slab caches. And that's it. > > If the implementation does merging or whatever, the UAPI is expressing > that and it's perfectly legitimate and not breaking UAPI in my > opinion. > > I think the better solution is to grab the information from somewhere > else, so let's move this conversation along with the answer to my > question about asking for more details about what is needed by > iproute2. > > Thank you. There is no where else that gives summary information. Both /proc/net/tcp and sock diag info require user space to read all the data, which is what I think Alexey was trying to avoid. Ideally there would be network namespace aware API to do this, but the code (iproute2 and kernel) are currently broken. Some values are missing (because they are merged) and some values have wrong cache name (because of acme's changes to make this generic).
Re: [PATCH 0/2] mark some slabs as visible not mergeable
From: Stephen Hemminger Date: Sat, 24 Feb 2018 11:04:52 -0800 > This fixes an old bug in iproute2's ss command because it was > reading slabinfo to get statistics. There isn't a better API > to do this, and one can argue that /proc is a UAPI that must > not change. Please elaborate what kind of statistics are needed. > Therefore this patch set adds a flag to slab to give another > reason to prevent merging, and then uses it in network code. > > The patches are against davem's linux-net tree and should also > goto stable as well. Well, as has been pointed out this never worked with SLUB so in some sense this was always broken. And the "UAPI" of slabinfo is to show the state of the various slab caches. And that's it. If the implementation does merging or whatever, the UAPI is expressing that and it's perfectly legitimate and not breaking UAPI in my opinion. I think the better solution is to grab the information from somewhere else, so let's move this conversation along with the answer to my question about asking for more details about what is needed by iproute2. Thank you.
Re: [PATCH 0/2] mark some slabs as visible not mergeable
On Sat, 24 Feb 2018 11:04:52 -0800 Stephen Hemminger wrote: > This fixes an old bug in iproute2's ss command because it was > reading slabinfo to get statistics. There isn't a better API > to do this, and one can argue that /proc is a UAPI that must > not change. > > Therefore this patch set adds a flag to slab to give another > reason to prevent merging, and then uses it in network code. > > The patches are against davem's linux-net tree and should also > goto stable as well. > > Stephen Hemminger (2): > slab: add flag to block merging of UAPI elements > net: mark slab's used by ss as UAPI > > include/linux/slab.h | 6 ++ > mm/slab_common.c | 2 +- > net/ipv4/tcp.c | 3 ++- > net/ipv4/tcp_ipv4.c | 2 +- > net/ipv6/tcp_ipv6.c | 2 +- > net/socket.c | 6 +++--- > 6 files changed, 14 insertions(+), 7 deletions(-) > The kbuild reports need more root cause investigation before applying.
Re: [PATCH 0/2] mark some slabs as visible not mergeable
On Sat, Feb 24, 2018 at 11:04:52AM -0800, Stephen Hemminger wrote: > This fixes an old bug in iproute2's ss command because it was > reading slabinfo to get statistics. There isn't a better API > to do this, and one can argue that /proc is a UAPI that must > not change. > > Therefore this patch set adds a flag to slab to give another > reason to prevent merging, and then uses it in network code. This is exactly the solution I would have suggested. Note that SLUB has always had slab merging, so this tool has been broken since 2.6.22 on any kernel with CONFIG_SLUB. Reviewed-by: Matthew Wilcox
[PATCH 0/2] mark some slabs as visible not mergeable
This fixes an old bug in iproute2's ss command because it was reading slabinfo to get statistics. There isn't a better API to do this, and one can argue that /proc is a UAPI that must not change. Therefore this patch set adds a flag to slab to give another reason to prevent merging, and then uses it in network code. The patches are against davem's linux-net tree and should also goto stable as well. Stephen Hemminger (2): slab: add flag to block merging of UAPI elements net: mark slab's used by ss as UAPI include/linux/slab.h | 6 ++ mm/slab_common.c | 2 +- net/ipv4/tcp.c | 3 ++- net/ipv4/tcp_ipv4.c | 2 +- net/ipv6/tcp_ipv6.c | 2 +- net/socket.c | 6 +++--- 6 files changed, 14 insertions(+), 7 deletions(-) -- 2.16.1
