Message ID | 20210707122201.14618-1-alexander.mikhalitsyn@virtuozzo.com (mailing list archive) |
---|---|
State | Accepted |
Commit | 459ce6e3d792afe3fe8f7f4b78baa68bb72b8439 |
Delegated to: | David Ahern |
Headers | show |
Series | [PATCHv5,iproute2] ip route: ignore ENOENT during save if RT_TABLE_MAIN is being dumped | expand |
Context | Check | Description |
---|---|---|
netdev/tree_selection | success | Not a local patch |
On Wed, 7 Jul 2021 15:22:01 +0300 Alexander Mikhalitsyn <alexander.mikhalitsyn@virtuozzo.com> wrote: > We started to use in-kernel filtering feature which allows to get only > needed tables (see iproute_dump_filter()). From the kernel side it's > implemented in net/ipv4/fib_frontend.c (inet_dump_fib), net/ipv6/ip6_fib.c > (inet6_dump_fib). The problem here is that behaviour of "ip route save" > was changed after > c7e6371bc ("ip route: Add protocol, table id and device to dump request"). > If filters are used, then kernel returns ENOENT error if requested table > is absent, but in newly created net namespace even RT_TABLE_MAIN table > doesn't exist. It is really allocated, for instance, after issuing > "ip l set lo up". > > Reproducer is fairly simple: > $ unshare -n ip route save > dump > Error: ipv4: FIB table does not exist. > Dump terminated > > Expected result here is to get empty dump file (as it was before this > change). > > v2: reworked, so, now it takes into account NLMSGERR_ATTR_MSG > (see nl_dump_ext_ack_done() function). We want to suppress error messages > in stderr about absent FIB table from kernel too. > > v3: reworked to make code clearer. Introduced rtnl_suppressed_errors(), > rtnl_suppress_error() helpers. User may suppress up to 3 errors (may be > easily extended by changing SUPPRESS_ERRORS_INIT macro). > > v4: reworked, rtnl_dump_filter_errhndlr() was introduced. Thanks > to Stephen Hemminger for comments and suggestions > > v5: space fixes, commit message reformat, empty initializers > > Fixes: c7e6371bc ("ip route: Add protocol, table id and device to dump request") > Cc: David Ahern <dsahern@gmail.com> > Cc: Stephen Hemminger <stephen@networkplumber.org> > Cc: Andrei Vagin <avagin@gmail.com> > Cc: Alexander Mikhalitsyn <alexander@mihalicyn.com> > Signed-off-by: Alexander Mikhalitsyn <alexander.mikhalitsyn@virtuozzo.com> Applied this version
On Wed, 7 Jul 2021 07:35:05 -0700 Stephen Hemminger <stephen@networkplumber.org> wrote: > On Wed, 7 Jul 2021 15:22:01 +0300 > Alexander Mikhalitsyn <alexander.mikhalitsyn@virtuozzo.com> wrote: > > > We started to use in-kernel filtering feature which allows to get only > > needed tables (see iproute_dump_filter()). From the kernel side it's > > implemented in net/ipv4/fib_frontend.c (inet_dump_fib), net/ipv6/ip6_fib.c > > (inet6_dump_fib). The problem here is that behaviour of "ip route save" > > was changed after > > c7e6371bc ("ip route: Add protocol, table id and device to dump request"). > > If filters are used, then kernel returns ENOENT error if requested table > > is absent, but in newly created net namespace even RT_TABLE_MAIN table > > doesn't exist. It is really allocated, for instance, after issuing > > "ip l set lo up". > > > > Reproducer is fairly simple: > > $ unshare -n ip route save > dump > > Error: ipv4: FIB table does not exist. > > Dump terminated > > > > Expected result here is to get empty dump file (as it was before this > > change). > > > > v2: reworked, so, now it takes into account NLMSGERR_ATTR_MSG > > (see nl_dump_ext_ack_done() function). We want to suppress error messages > > in stderr about absent FIB table from kernel too. > > > > v3: reworked to make code clearer. Introduced rtnl_suppressed_errors(), > > rtnl_suppress_error() helpers. User may suppress up to 3 errors (may be > > easily extended by changing SUPPRESS_ERRORS_INIT macro). > > > > v4: reworked, rtnl_dump_filter_errhndlr() was introduced. Thanks > > to Stephen Hemminger for comments and suggestions > > > > v5: space fixes, commit message reformat, empty initializers > > > > Fixes: c7e6371bc ("ip route: Add protocol, table id and device to dump request") > > Cc: David Ahern <dsahern@gmail.com> > > Cc: Stephen Hemminger <stephen@networkplumber.org> > > Cc: Andrei Vagin <avagin@gmail.com> > > Cc: Alexander Mikhalitsyn <alexander@mihalicyn.com> > > Signed-off-by: Alexander Mikhalitsyn <alexander.mikhalitsyn@virtuozzo.com> > > Applied this version Stephen, Thank you for your review and suggestions! Alex
Hello: This patch was applied to iproute2/iproute2.git (refs/heads/main): On Wed, 7 Jul 2021 15:22:01 +0300 you wrote: > We started to use in-kernel filtering feature which allows to get only > needed tables (see iproute_dump_filter()). From the kernel side it's > implemented in net/ipv4/fib_frontend.c (inet_dump_fib), net/ipv6/ip6_fib.c > (inet6_dump_fib). The problem here is that behaviour of "ip route save" > was changed after > c7e6371bc ("ip route: Add protocol, table id and device to dump request"). > If filters are used, then kernel returns ENOENT error if requested table > is absent, but in newly created net namespace even RT_TABLE_MAIN table > doesn't exist. It is really allocated, for instance, after issuing > "ip l set lo up". > > [...] Here is the summary with links: - [PATCHv5,iproute2] ip route: ignore ENOENT during save if RT_TABLE_MAIN is being dumped https://git.kernel.org/pub/scm/network/iproute2/iproute2.git/commit/?id=459ce6e3d792 You are awesome, thank you! -- Deet-doot-dot, I am a bot. https://korg.docs.kernel.org/patchwork/pwbot.html
On 2021-07-07 5:50 PM, patchwork-bot+netdevbpf@kernel.org wrote: > Hello: > > This patch was applied to iproute2/iproute2.git (refs/heads/main): > > On Wed, 7 Jul 2021 15:22:01 +0300 you wrote: >> We started to use in-kernel filtering feature which allows to get only >> needed tables (see iproute_dump_filter()). From the kernel side it's >> implemented in net/ipv4/fib_frontend.c (inet_dump_fib), net/ipv6/ip6_fib.c >> (inet6_dump_fib). The problem here is that behaviour of "ip route save" >> was changed after >> c7e6371bc ("ip route: Add protocol, table id and device to dump request"). >> If filters are used, then kernel returns ENOENT error if requested table >> is absent, but in newly created net namespace even RT_TABLE_MAIN table >> doesn't exist. It is really allocated, for instance, after issuing >> "ip l set lo up". >> >> [...] > > Here is the summary with links: > - [PATCHv5,iproute2] ip route: ignore ENOENT during save if RT_TABLE_MAIN is being dumped > https://git.kernel.org/pub/scm/network/iproute2/iproute2.git/commit/?id=459ce6e3d792 > > You are awesome, thank you! > -- > Deet-doot-dot, I am a bot. > https://korg.docs.kernel.org/patchwork/pwbot.html > > Hi, I didn't get a chance to check but after this commit the utility ss is crashing from libnetlink.c (gdb) bt #0 0x0000000000000000 in ?? () #1 0x0000000000418980 in rtnl_dump_done (a=0x7fffffffdc00, h=0x449c80) at libnetlink.c:734 #2 rtnl_dump_filter_l (rth=rth@entry=0x7fffffffdcf0, arg=arg@entry=0x7fffffffdc00) at libnetlink.c:893 #3 0x00000000004197a2 in rtnl_dump_filter_nc (rth=rth@entry=0x7fffffffdcf0, filter=filter@entry=0x40cdf0 <show_one_inet_sock>, arg1=arg1@entry=0x7fffffffdca0, nc_flags=nc_flags@entry=0) at libnetlink.c:957 #4 0x0000000000406dc7 in inet_show_netlink (dump_fp=dump_fp@entry=0x0, protocol=protocol@entry=255, f=0x42e900 <current_filter>) at ss.c:3638 #5 0x0000000000404db3 in raw_show (f=0x42e900 <current_filter>) at ss.c:3939 if I revert this commit I can run the ss utility ok. Thanks, Roi
On 2021-07-11 1:59 PM, Roi Dayan wrote: > > > On 2021-07-07 5:50 PM, patchwork-bot+netdevbpf@kernel.org wrote: >> Hello: >> >> This patch was applied to iproute2/iproute2.git (refs/heads/main): >> >> On Wed, 7 Jul 2021 15:22:01 +0300 you wrote: >>> We started to use in-kernel filtering feature which allows to get only >>> needed tables (see iproute_dump_filter()). From the kernel side it's >>> implemented in net/ipv4/fib_frontend.c (inet_dump_fib), >>> net/ipv6/ip6_fib.c >>> (inet6_dump_fib). The problem here is that behaviour of "ip route save" >>> was changed after >>> c7e6371bc ("ip route: Add protocol, table id and device to dump >>> request"). >>> If filters are used, then kernel returns ENOENT error if requested table >>> is absent, but in newly created net namespace even RT_TABLE_MAIN table >>> doesn't exist. It is really allocated, for instance, after issuing >>> "ip l set lo up". >>> >>> [...] >> >> Here is the summary with links: >> - [PATCHv5,iproute2] ip route: ignore ENOENT during save if >> RT_TABLE_MAIN is being dumped >> >> https://git.kernel.org/pub/scm/network/iproute2/iproute2.git/commit/?id=459ce6e3d792 >> >> >> You are awesome, thank you! >> -- >> Deet-doot-dot, I am a bot. >> https://korg.docs.kernel.org/patchwork/pwbot.html >> >> > > > Hi, > > I didn't get a chance to check but after this commit the utility ss > is crashing from libnetlink.c > > (gdb) bt > #0 0x0000000000000000 in ?? () > #1 0x0000000000418980 in rtnl_dump_done (a=0x7fffffffdc00, h=0x449c80) > at libnetlink.c:734 > #2 rtnl_dump_filter_l (rth=rth@entry=0x7fffffffdcf0, > arg=arg@entry=0x7fffffffdc00) at libnetlink.c:893 > #3 0x00000000004197a2 in rtnl_dump_filter_nc > (rth=rth@entry=0x7fffffffdcf0, filter=filter@entry=0x40cdf0 > <show_one_inet_sock>, > arg1=arg1@entry=0x7fffffffdca0, nc_flags=nc_flags@entry=0) at > libnetlink.c:957 > #4 0x0000000000406dc7 in inet_show_netlink (dump_fp=dump_fp@entry=0x0, > protocol=protocol@entry=255, f=0x42e900 <current_filter>) at ss.c:3638 > #5 0x0000000000404db3 in raw_show (f=0x42e900 <current_filter>) at > ss.c:3939 > > > if I revert this commit I can run the ss utility ok. > > Thanks, > Roi I found for me ss crashed using a->errhndlr but it could be null in rtnl_dump_done() I see a second usage of a->errhndlr in rtnl_dump_error()
Thanks, Roi. I will send a patch soon. On Sun, Jul 11, 2021 at 2:09 PM Roi Dayan <roid@nvidia.com> wrote: > > > > On 2021-07-11 1:59 PM, Roi Dayan wrote: > > > > > > On 2021-07-07 5:50 PM, patchwork-bot+netdevbpf@kernel.org wrote: > >> Hello: > >> > >> This patch was applied to iproute2/iproute2.git (refs/heads/main): > >> > >> On Wed, 7 Jul 2021 15:22:01 +0300 you wrote: > >>> We started to use in-kernel filtering feature which allows to get only > >>> needed tables (see iproute_dump_filter()). From the kernel side it's > >>> implemented in net/ipv4/fib_frontend.c (inet_dump_fib), > >>> net/ipv6/ip6_fib.c > >>> (inet6_dump_fib). The problem here is that behaviour of "ip route save" > >>> was changed after > >>> c7e6371bc ("ip route: Add protocol, table id and device to dump > >>> request"). > >>> If filters are used, then kernel returns ENOENT error if requested table > >>> is absent, but in newly created net namespace even RT_TABLE_MAIN table > >>> doesn't exist. It is really allocated, for instance, after issuing > >>> "ip l set lo up". > >>> > >>> [...] > >> > >> Here is the summary with links: > >> - [PATCHv5,iproute2] ip route: ignore ENOENT during save if > >> RT_TABLE_MAIN is being dumped > >> > >> https://git.kernel.org/pub/scm/network/iproute2/iproute2.git/commit/?id=459ce6e3d792 > >> > >> > >> You are awesome, thank you! > >> -- > >> Deet-doot-dot, I am a bot. > >> https://korg.docs.kernel.org/patchwork/pwbot.html > >> > >> > > > > > > Hi, > > > > I didn't get a chance to check but after this commit the utility ss > > is crashing from libnetlink.c > > > > (gdb) bt > > #0 0x0000000000000000 in ?? () > > #1 0x0000000000418980 in rtnl_dump_done (a=0x7fffffffdc00, h=0x449c80) > > at libnetlink.c:734 > > #2 rtnl_dump_filter_l (rth=rth@entry=0x7fffffffdcf0, > > arg=arg@entry=0x7fffffffdc00) at libnetlink.c:893 > > #3 0x00000000004197a2 in rtnl_dump_filter_nc > > (rth=rth@entry=0x7fffffffdcf0, filter=filter@entry=0x40cdf0 > > <show_one_inet_sock>, > > arg1=arg1@entry=0x7fffffffdca0, nc_flags=nc_flags@entry=0) at > > libnetlink.c:957 > > #4 0x0000000000406dc7 in inet_show_netlink (dump_fp=dump_fp@entry=0x0, > > protocol=protocol@entry=255, f=0x42e900 <current_filter>) at ss.c:3638 > > #5 0x0000000000404db3 in raw_show (f=0x42e900 <current_filter>) at > > ss.c:3939 > > > > > > if I revert this commit I can run the ss utility ok. > > > > Thanks, > > Roi > > I found for me ss crashed using a->errhndlr but it could be null > in rtnl_dump_done() > > I see a second usage of a->errhndlr in rtnl_dump_error() > >
diff --git a/include/libnetlink.h b/include/libnetlink.h index b9073a6a..4401551f 100644 --- a/include/libnetlink.h +++ b/include/libnetlink.h @@ -104,6 +104,27 @@ struct rtnl_ctrl_data { typedef int (*rtnl_filter_t)(struct nlmsghdr *n, void *); +/** + * rtnl error handler called from + * rtnl_dump_done() + * rtnl_dump_error() + * + * Return value is a bitmask of the following values: + * RTNL_LET_NLERR + * error handled as usual + * RTNL_SUPPRESS_NLMSG_DONE_NLERR + * error in nlmsg_type == NLMSG_DONE will be suppressed + * RTNL_SUPPRESS_NLMSG_ERROR_NLERR + * error in nlmsg_type == NLMSG_ERROR will be suppressed + * and nlmsg will be skipped + * RTNL_SUPPRESS_NLERR - suppress error in both previous cases + */ +#define RTNL_LET_NLERR 0x01 +#define RTNL_SUPPRESS_NLMSG_DONE_NLERR 0x02 +#define RTNL_SUPPRESS_NLMSG_ERROR_NLERR 0x04 +#define RTNL_SUPPRESS_NLERR 0x06 +typedef int (*rtnl_err_hndlr_t)(struct nlmsghdr *n, void *); + typedef int (*rtnl_listen_filter_t)(struct rtnl_ctrl_data *, struct nlmsghdr *n, void *); @@ -113,6 +134,8 @@ typedef int (*nl_ext_ack_fn_t)(const char *errmsg, uint32_t off, struct rtnl_dump_filter_arg { rtnl_filter_t filter; void *arg1; + rtnl_err_hndlr_t errhndlr; + void *arg2; __u16 nc_flags; }; @@ -121,6 +144,15 @@ int rtnl_dump_filter_nc(struct rtnl_handle *rth, void *arg, __u16 nc_flags); #define rtnl_dump_filter(rth, filter, arg) \ rtnl_dump_filter_nc(rth, filter, arg, 0) +int rtnl_dump_filter_errhndlr_nc(struct rtnl_handle *rth, + rtnl_filter_t filter, + void *arg1, + rtnl_err_hndlr_t errhndlr, + void *arg2, + __u16 nc_flags); +#define rtnl_dump_filter_errhndlr(rth, filter, farg, errhndlr, earg) \ + rtnl_dump_filter_errhndlr_nc(rth, filter, farg, errhndlr, earg, 0) + int rtnl_talk(struct rtnl_handle *rtnl, struct nlmsghdr *n, struct nlmsghdr **answer) __attribute__((warn_unused_result)); diff --git a/ip/iproute.c b/ip/iproute.c index 5853f026..e45f0bea 100644 --- a/ip/iproute.c +++ b/ip/iproute.c @@ -1727,6 +1727,18 @@ static int iproute_flush(int family, rtnl_filter_t filter_fn) } } +static int save_route_errhndlr(struct nlmsghdr *n, void *arg) +{ + int err = -*(int *)NLMSG_DATA(n); + + if (n->nlmsg_type == NLMSG_DONE && + filter.tb == RT_TABLE_MAIN && + err == ENOENT) + return RTNL_SUPPRESS_NLMSG_DONE_NLERR; + + return RTNL_LET_NLERR; +} + static int iproute_list_flush_or_save(int argc, char **argv, int action) { int dump_family = preferred_family; @@ -1939,7 +1951,8 @@ static int iproute_list_flush_or_save(int argc, char **argv, int action) new_json_obj(json); - if (rtnl_dump_filter(&rth, filter_fn, stdout) < 0) { + if (rtnl_dump_filter_errhndlr(&rth, filter_fn, stdout, + save_route_errhndlr, NULL) < 0) { fprintf(stderr, "Dump terminated\n"); return -2; } diff --git a/lib/libnetlink.c b/lib/libnetlink.c index c958aa57..e9b8c3bd 100644 --- a/lib/libnetlink.c +++ b/lib/libnetlink.c @@ -673,7 +673,8 @@ int rtnl_dump_request_n(struct rtnl_handle *rth, struct nlmsghdr *n) return sendmsg(rth->fd, &msg, 0); } -static int rtnl_dump_done(struct nlmsghdr *h) +static int rtnl_dump_done(struct nlmsghdr *h, + const struct rtnl_dump_filter_arg *a) { int len = *(int *)NLMSG_DATA(h); @@ -683,11 +684,15 @@ static int rtnl_dump_done(struct nlmsghdr *h) } if (len < 0) { + errno = -len; + + if (a->errhndlr(h, a->arg2) & RTNL_SUPPRESS_NLMSG_DONE_NLERR) + return 0; + /* check for any messages returned from kernel */ if (nl_dump_ext_ack_done(h, len)) return len; - errno = -len; switch (errno) { case ENOENT: case EOPNOTSUPP: @@ -708,8 +713,9 @@ static int rtnl_dump_done(struct nlmsghdr *h) return 0; } -static void rtnl_dump_error(const struct rtnl_handle *rth, - struct nlmsghdr *h) +static int rtnl_dump_error(const struct rtnl_handle *rth, + struct nlmsghdr *h, + const struct rtnl_dump_filter_arg *a) { if (h->nlmsg_len < NLMSG_LENGTH(sizeof(struct nlmsgerr))) { @@ -721,11 +727,16 @@ static void rtnl_dump_error(const struct rtnl_handle *rth, if (rth->proto == NETLINK_SOCK_DIAG && (errno == ENOENT || errno == EOPNOTSUPP)) - return; + return -1; + + if (a->errhndlr(h, a->arg2) & RTNL_SUPPRESS_NLMSG_ERROR_NLERR) + return 0; if (!(rth->flags & RTNL_HANDLE_F_SUPPRESS_NLERR)) perror("RTNETLINK answers"); } + + return -1; } static int __rtnl_recvmsg(int fd, struct msghdr *msg, int flags) @@ -834,7 +845,7 @@ static int rtnl_dump_filter_l(struct rtnl_handle *rth, dump_intr = 1; if (h->nlmsg_type == NLMSG_DONE) { - err = rtnl_dump_done(h); + err = rtnl_dump_done(h, a); if (err < 0) { free(buf); return -1; @@ -845,9 +856,13 @@ static int rtnl_dump_filter_l(struct rtnl_handle *rth, } if (h->nlmsg_type == NLMSG_ERROR) { - rtnl_dump_error(rth, h); - free(buf); - return -1; + err = rtnl_dump_error(rth, h, a); + if (err < 0) { + free(buf); + return -1; + } + + goto skip_it; } if (!rth->dump_fp) { @@ -887,8 +902,25 @@ int rtnl_dump_filter_nc(struct rtnl_handle *rth, void *arg1, __u16 nc_flags) { const struct rtnl_dump_filter_arg a[2] = { - { .filter = filter, .arg1 = arg1, .nc_flags = nc_flags, }, - { .filter = NULL, .arg1 = NULL, .nc_flags = 0, }, + { .filter = filter, .arg1 = arg1, + .errhndlr = NULL, .arg2 = NULL, .nc_flags = nc_flags, }, + { }, + }; + + return rtnl_dump_filter_l(rth, a); +} + +int rtnl_dump_filter_errhndlr_nc(struct rtnl_handle *rth, + rtnl_filter_t filter, + void *arg1, + rtnl_err_hndlr_t errhndlr, + void *arg2, + __u16 nc_flags) +{ + const struct rtnl_dump_filter_arg a[2] = { + { .filter = filter, .arg1 = arg1, + .errhndlr = errhndlr, .arg2 = arg2, .nc_flags = nc_flags, }, + { }, }; return rtnl_dump_filter_l(rth, a);
We started to use in-kernel filtering feature which allows to get only needed tables (see iproute_dump_filter()). From the kernel side it's implemented in net/ipv4/fib_frontend.c (inet_dump_fib), net/ipv6/ip6_fib.c (inet6_dump_fib). The problem here is that behaviour of "ip route save" was changed after c7e6371bc ("ip route: Add protocol, table id and device to dump request"). If filters are used, then kernel returns ENOENT error if requested table is absent, but in newly created net namespace even RT_TABLE_MAIN table doesn't exist. It is really allocated, for instance, after issuing "ip l set lo up". Reproducer is fairly simple: $ unshare -n ip route save > dump Error: ipv4: FIB table does not exist. Dump terminated Expected result here is to get empty dump file (as it was before this change). v2: reworked, so, now it takes into account NLMSGERR_ATTR_MSG (see nl_dump_ext_ack_done() function). We want to suppress error messages in stderr about absent FIB table from kernel too. v3: reworked to make code clearer. Introduced rtnl_suppressed_errors(), rtnl_suppress_error() helpers. User may suppress up to 3 errors (may be easily extended by changing SUPPRESS_ERRORS_INIT macro). v4: reworked, rtnl_dump_filter_errhndlr() was introduced. Thanks to Stephen Hemminger for comments and suggestions v5: space fixes, commit message reformat, empty initializers Fixes: c7e6371bc ("ip route: Add protocol, table id and device to dump request") Cc: David Ahern <dsahern@gmail.com> Cc: Stephen Hemminger <stephen@networkplumber.org> Cc: Andrei Vagin <avagin@gmail.com> Cc: Alexander Mikhalitsyn <alexander@mihalicyn.com> Signed-off-by: Alexander Mikhalitsyn <alexander.mikhalitsyn@virtuozzo.com> --- include/libnetlink.h | 32 ++++++++++++++++++++++++++ ip/iproute.c | 15 +++++++++++- lib/libnetlink.c | 54 +++++++++++++++++++++++++++++++++++--------- 3 files changed, 89 insertions(+), 12 deletions(-)