| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
During any file locking interaction between an NFS client and server,
the client tells the server what hostname it will use as the mon_name
argument of the SM_NOTIFY request sent by the client when it reboots.
This is the "caller_name" argument of an NLMPROC_LOCK request.
The server, however, never tells the client what mon_name argument
it will use when sending an SM_NOTIFY request. In order to recognize
the server, clients usually guess what mon_name the server might
send, by using the server hostname provided by the user on the mount
command line.
Frequently, the user provides an unqualified server name on the mount
command. The server might then call the client back with a fully
qualified domain name, which might not match in some cases.
Solaris, and perhaps other implementations, attempt to mitigate this
problem by sending two SM_NOTIFY requests to each peer: one with an
unqualified mon_name argument, and one with a fully qualified mon_name.
Implement such a scheme for sm-notify.
Since my_name is almost always the fully-qualified hostname associated
with the local system, just wiping the left-most '.' in the my_name
argument and sending another SM_NOTIFY is nearly always sufficient.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The mon_name argument of an SM_NOTIFY request is a string that
identifies the rebooting host.
sm-notify should send the my_name provided by the local lockd at the
time the remote was monitored, rather than cocking up a mon_name
argument based on the present return value of gethostname(3). If the
local system's hostname happened to change after the last reboot, then
the string returned by gethostname(3) will not be recognized by the
remote. Thus the remote will never initiate lock recovery for the
original named host, possibly leaving stale locks.
The existing behavior of using the -v command line option as the
mon_name argument is preserved, but we now prevent sending an IP
presentation address, as some non-Linux implementations don't
recognize addresses as valid mon_names.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Jeff Layton <jlayton@redhat.com>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
To support IPv6, statd must support multi-homed remote peers. For our
purposes, "multi-homed peer" means that more than one unique IP
address maps to the one canonical host name for that peer.
An SM_MON request from the local lockd has a "mon_name" argument that
statd reverse maps to a canonical hostname (ie the A record for that
host). statd assumes the canonical hostname is unique enough that
it stores the callback data for this mon_name in a file named after
that canonical hostname.
Because lockd can't distinguish between two unique IP addresses that
may be from the same physical host, the kernel can hand statd a
mon_name that maps to the same canonical hostname as some previous
mon_name. So that the kernel can keep this instance of the mon_name
unique, it creates a fresh priv cookie for each new address.
Note that a mon_name can be a presentation address string, or the
caller_name string sent in each NLMPROC_LOCK request. There's
nothing that requires the caller_name to be a fully-qualified
hostname, thus it's uniqueness is not guaranteed. The current
design of statd assumes that canonical hostnames will be unique
enough.
When a mon_name for a fresh SM_MON request maps to the same canonical
hostname as an existing monitored peer, but the priv cookie is new,
statd will try to write the information for the fresh request into an
existing monitor record file, wiping out the contents of the file.
This is because the mon_name/cookie combination won't match any record
statd already has.
Currently, statd doesn't check if a record file already exists before
writing into it. statd's logic assumes that the svc routine has
already checked that no matching record exists in the in-core monitor
list. And, it doesn't use O_EXCL when opening the record file. Not
only is the old data in that file wiped out, but statd's in-core
monitor list will no longer match what's in the on-disk monitor list.
Note that IPv6 isn't needed to exercise multi-homed peer support.
Any IPv4 peer that has multiple addresses that map to its canonical
hostname will trigger this behavior. However, this scenario will
become quite common when all hosts on a network automatically get both
an IPv4 address and an IPv6 address.
I can think of a few ways to address this:
1. Replace the current on-disk format with a database that has a
uniqueness constraint on the monitor records
2. Create a new file naming scheme; eg. one that uses a truly
unique name such as a hash generated from the mon_name, my_name, and
priv cookie
3. Support multiple lines in each monitor record file
Since statd's on-disk format constitutes a formal API, options 1 and 2
are right out. This patch implements option 3. There are two parts:
adding a new line to an existing file; and deleting a line from a file
with more than one line. Interestingly, the existing code already
supports reading more than one line from these files, so we don't need
to add extra code here to do that.
One file may contain a line for every unique mon_name / priv cookie
where the mon_name reverse maps to the same canonical hostname. We
use the atomic write facility added by a previous patch to ensure the
on-disk monitor record list is updated atomically.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
|
| |
|
|
|
|
|
|
|
| |
Currently sm-notify does not use the mon_name and my_name strings
passed to smn_get_host(). Very soon we're going to need the mon_name
and my_name strings, so add code to store those strings in struct
nsm_host, and free them when each host is forgotten.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Introduce generic helpers for managing socket addresses. These are
general enough that they are useful for pretty much any component of
nfs-utils.
We also include the definition of nfs_sockaddr here, so it can be
shared. See:
https://bugzilla.redhat.com/show_bug.cgi?id=448743
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
|
| |
|
|
|
|
|
| |
When IPV6_SUPPORTED is enabled and the local system has IPv6 support,
request AF_INET6 and AF_INET addresses from the DNS resolver.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
|
| |
|
|
|
|
|
|
|
|
|
|
| |
This patch updates the "bind to a user-specified port" arm of
smn_create_socket() so it can deal with IPv6 bind addresses.
A single getaddrinfo(3) call can convert a user-specified bind address
or hostname to a socket address, optionally plant a provided port
number, or whip up an appropriate wildcard address for use as the main
socket's bind address.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
|
| |
|
|
|
|
|
| |
This patch updates the "bind to an arbitrary privileged port" arm of
smn_create_socket() so it can deal with IPv6 bind addresses.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
Socket creation is unfortunately complicated by the need to handle the
case where sm-notify is built with IPv6 support, but the local system
has disabled it entirely at run-time (ie, socket(3) returns
EAFNOSUPPORT when we try to create an AF_INET6 socket).
The run-time address family setting is made available in the global
variable nsm_family. This setting can control the family of the
socket's bind address and what kind of addresses we want returned by
smn_lookup(). Support for that is added in subsequent patches.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
|
| |
|
|
|
|
|
|
|
|
|
|
| |
The top half of the notify() function creates the main socket that
sm-notify uses to do its job. To make adding IPv6 support simpler,
refactor that piece into a separate function.
The logic is modified slightly so that exit(3) is invoked only in
main(). This is not required, but it makes the code slightly easier
to understand and maintain.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Replace the open code to construct SM_NOTIFY and PMAP_GETPORT RPC
requests with calls to our new library routines that support
IPv6 and RPCB_GETADDR as well.
This change allows sm-notify to send RPCB_GETADDR, but it won't do
that until the main sm-notify socket supports PF_INET6 and the DNS
resolution logic is updated to return IPv6 addresses.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Replace open-coded accesses to on-disk NSM data with calls to the new
libnsm.a API.
One major change is that sync(2) is no longer called when the NSM
state number is updated at boot time. Otherwise sm-notify should
behave much the same as it did before.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
|
| |
Use shared sockaddr port management functions instead of duplicating
this functionality in sm-notify. This is now easy because sm-notify
is linked with libnfs.a, where nfs_{get,set}_port() reside.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
Clean up: Get rid of a false positive compiler warning, seen with
-Wextra.
sm-notify.c: In function ¿record_pid¿:
sm-notify.c:690: warning: comparison between signed and unsigned integer
expressions
Document some ignored return codes while we're here.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
|
|
|
| |
To facilitate code sharing between statd and sm-notify (and with other
components of nfs-utils), replace sm-notify's nsm_log() with xlog().
Since opt_quiet is used in only a handful of insignificant cases, it
is removed.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Currently, if getaddrinfo(3) fails when trying to resolve a hostname,
sm-notify gives up immediately on that host. If sm-notify is started
before network service is available on a system, that means it quits
without notifying anyone. Or, if DNS service isn't available due to
a network partition or because the DNS server crashed, sm-notify will
simply remove all of its callback files and exit.
Really, sm-notify should try harder. We know that the hostnames
passed in to notify_host() have already been vetted by statd, which
won't monitor a hostname that it can't resolve. So it's likely that
any DNS failure we meet here is a temporary condition. If it isn't,
then sm-notify will stop trying to notify that host in 15 minutes
anyway.
[ The host's file is left in /var/lib/nfs/sm.bak in this case, but
sm.bak is not read again until the next time sm-notify runs. ]
sm-notify already has retry logic for handling RPC timeouts. We can
co-opt that to drive DNS resolution retries.
We also add AI_ADDRCONFIG because on systems whose network startup is
handled by NetworkManager, there appears to be a bug that causes
processes that started calling getaddinfo(3) before the network came
up to continue getting EAI_AGAIN even after the network is fully
operating.
As I understand it, legacy glibc (before AI_ADDRCONFIG was exposed in
headers) sets AI_ADDRCONFIG by default, although I haven't checked
this. In any event, pre-glibc-2.2 systems probably won't run
NetworkManager anyway, so this may not be much of a problem for them.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
sm-notify orphans an addrinfo struct in its address list rotation
logic if only a single result was returned from getaddrinfo(3).
For each host, the first time through notify_host(), we want to
send a PMAP_GETPORT request. ->ai is NULL, and retries is set to 100,
forcing a DNS lookup and an address rotation. If only a single
addrinfo struct is returned, the rotation logic causes a NULL to be
planted in ->ai, copied from the ai_next field of the returned result.
This means that the second time through notify_host() (to perform the
actual SM_NOTIFY call) we do a second DNS lookup, since ->ai is NULL.
The result of the first lookup has been orphaned, and extra network
traffic is generated.
This scenario is actually fairly common. Since we pass
.ai_protocol = IPPROTO_UDP,
to getaddrinfo(3), for most hosts, which have a single forward and
reverse pointer in the DNS database, we get back a single addrinfo
struct as a result.
To address this problem, only perform the address list rotation if
there is more than one element on the list returned by getaddrinfo(3).
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
| |
flag has been set. This cause warnings to be generated when
return values from reads/writes (and other calls) are not
checked. The patch address those warnings.
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
The recv_reply() function was referencing host->ai in a freeaddrinfo(3)
call after it had freed @host.
This is not likely to be harmful in a single-threaded user context,
but it's still bad form, and it will get called out if testing
sm-notify with poisoned free memory. The less noise, the better we
are able to see real problems.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
| |
Added curly brackets around the record_pid() check which
stop sm-notify from exiting when a pid file does not
exist.
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
| |
there are no hosts to notify. This also decreases
start up time by a few seconds.
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Clean up.
The sm-notify command is built from a single source file.
Some of its internal functions are appropriately defined as static.
However, some are declared static, but defined as global. Some are
declared and defined as global. None of them are used outside of
utils/statd/sm-notify.c.
Make all the internal functions in utils/statd/sm-notify.cstatic.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
|
| |
Clean up: replace "typedef struct sockaddr_storage nsm_address" with
standard socket address types. This makes sm-notify.c consistent with other
parts of nfs-utils, and with typical network application coding conventions.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Clean up a few issues with logging in sm-notify.c.
Sometimes in sm-notify, when a system call fails the problem is reported
to stderr but not logged, and then usually sm-notify exits. In cases like
this, there are probably more hosts to notify, but sm-notify dies silently.
Make sure these errors are logged, and that the log messages explain the
nature of the problem.
Also, if sm-notify exits prematurely, make sure this is always reported at
the LOG_ERR level, not at the LOG_WARNING level.
Remove a couple of unnecessary '\n' in the arguments of nsm_log() calls --
nsm_log() already appends an '\n' to the message.
Finally, use exit() consistently in main().
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
|
|
|
| |
Make sure the results of getaddrinfo(3) are properly freed in notify().
Note this is a one-time addrinfo allocation that would be automatically
freed when sm-notify exits anyway, so this is more of a nit than a real
bug fix.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
| |
Clean up: Include config.h as other source files do; instead of using
"config.h" use the HAVE_CONFIG_H macro and include <config.h>.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
If an NFS server has no network connectivity when it reboots,
it will block in sm-notify waiting for DNS lookup for a potentially
large number of hosts. This is not helpful and just annoys the
sysadmin.
So do the DNS lookup in the backgrounded phase of sm-notify,
before sending off the NOTIFY requests.
Acked-by: NeilBrown <neilb@suse.de>
Signed-off-by: Steve Dickson <steved@redhat.com>
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Ports < 1024 are a scarce resource and should not be used
carelessly. Technically they should be not used at all without
registration with IANA, but sometimes we need them despite that.
So: for the socket that RPC services listen on, don't use a <1024 port
by default. There is no need.
For sockets that we send messages on, that are long-lived, and that might
need to appear 'privileged', avoid using a number that is registered in
/etc/services if possible.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Both SM_STAT and SM_MON can return the state of an NSM, but it is
unclear which NSM they return the state of, so the value cannot be
used, and lockd doesn't use it.
Document this confusion, and give the current state to the kernel
via a sysctl if that sysctl is available (since about 2.6.19).
This should make is possible for the NFS server to detect a small
class of bad SM_NOTIFY packets and not flush locks in that case.
Signed-off-by: Neil Brown <neilb@suse.de>
|
| |
|
|
|
|
|
| |
When sending an SM_NOTIFY to multi-homed host, try all the addresses
in rotation. After 4 failures on one address, try the next.
Signed-off-by: Neil Brown <neilb@suse.de>
|
| |
|
|
|
|
| |
Make sure that sm-notify really runs only once per reboot.
Signed-off-by: Neil Brown <neilb@suse.de>
|
| | |
|
| | |
|
| |
|
|
|
| |
Add sm-notify to the compile/install scripts,
(and fix a compile warning).
|
| |
|
|
|
| |
If /var/lib/nfs/sm is owned by non-root, setuid to that uid
after opening sockets but before receiving answers.
|
| |
|
|
|
|
|
|
|
| |
As "mount.nfs" can start statd, and as statd can start sm-notify,
the risk of sm-notify being run multiple times increases.
As this is not normally appropriate, sm-notify now creates a
file in /var/run which will stop future instances from being
run (though ofcourse this behaviour can be controlled by a
new command line option).
|
| |
|
|
|
| |
This functionality is alreday present in getaddrinfo so it isn't
needed explicitly.
|
| |
|
|
| |
for compat with statd.
|
|
|
Not included in build yet.
|
