kernel memory, whichever is lower. The overall mbuf related memory
limit must be set so that mbufs (and clusters of various sizes)
can't exhaust physical RAM or KVM.
The limit is set to half of the physical RAM or KVM (whichever is
lower) as the baseline. In any normal scenario we want to leave
at least half of the physmem/kvm for other kernel functions and
userspace to prevent it from swapping too easily. Via a tunable
kern.maxmbufmem the limit can be upped to at most 3/4 of physmem/kvm.
At the same time divorce maxfiles from maxusers and set maxfiles to
physpages / 8 with a floor based on maxusers. This way busy servers
can make use of the significantly increased mbuf limits with a much
larger number of open sockets.
Tidy up ordering in init_param2() and check up on some users of
those values calculated here.
Out of the overall mbuf memory limit 2K clusters and 4K (page size)
clusters to get 1/4 each because these are the most heavily used mbuf
sizes. 2K clusters are used for MTU 1500 ethernet inbound packets.
4K clusters are used whenever possible for sends on sockets and thus
outbound packets. The larger cluster sizes of 9K and 16K are limited
to 1/6 of the overall mbuf memory limit. When jumbo MTU's are used
these large clusters will end up only on the inbound path. They are
not used on outbound, there it's still 4K. Yes, that will stay that
way because otherwise we run into lots of complications in the
stack. And it really isn't a problem, so don't make a scene.
Normal mbufs (256B) weren't limited at all previously. This was
problematic as there are certain places in the kernel that on
allocation failure of clusters try to piece together their packet
from smaller mbufs.
The mbuf limit is the number of all other mbuf sizes together plus
some more to allow for standalone mbufs (ACK for example) and to
send off a copy of a cluster. Unfortunately there isn't a way to
set an overall limit for all mbuf memory together as UMA doesn't
support such a limiting.
NB: Every cluster also has an mbuf associated with it.
Two examples on the revised mbuf sizing limits:
1GB KVM:
512MB limit for mbufs
419,430 mbufs
65,536 2K mbuf clusters
32,768 4K mbuf clusters
9,709 9K mbuf clusters
5,461 16K mbuf clusters
16GB RAM:
8GB limit for mbufs
33,554,432 mbufs
1,048,576 2K mbuf clusters
524,288 4K mbuf clusters
155,344 9K mbuf clusters
87,381 16K mbuf clusters
These defaults should be sufficient for even the most demanding
network loads.
MFC after: 1 month
accept queues a new socket/connection may be added to the queue
due to a race on the ACCEPT_LOCK.
The submitted patch is slightly changed in comments, teardown
and locking order and extended with KASSERT's.
Submitted by: Vijay Singh <vijju.singh-at-gmail-dot-com>
Found by: His team.
MFC after: 1 week
overwriting the return mbuf pointer with newly received data after
a loop. Instead append the new mbuf chain to the existing one.
Fix up sb_lastrecord when dequeuing mbuf's so that sbappend_stream()
doesn't get confused.
For the remainder copy case in the mbuf delivery part deduct the
copied length len instead of the whole mbuf length. Additionally
don't depend on 'n' being being available which isn't true in the
case of MSG_PEEK.
Fix the MSG_WAITALL case by comparing against sb_hiwat. Before
it was looping for every receive as sb_lowat normally is zero.
Add comment about issue with (MSG_WAITALL | MSG_PEEK) which isn't
properly handled.
Submitted by: trociny (except for the change in last paragraph)
more appropriate named kernel options for the very distinct
send and receive path.
"options SOCKET_SEND_COW" enables VM page copy-on-write based
sending of data on an outbound socket.
NB: The COW based send mechanism is not safe and may result
in kernel crashes.
"options SOCKET_RECV_PFLIP" enables VM kernel/userspace page
flipping for special disposable pages attached as external
storage to mbufs.
Only the naming of the kernel options is changed and their
corresponding #ifdef sections are adjusted. No functionality
is added or removed.
Discussed with: alc (mechanism and limitations of send side COW)
output and replace it with a new visible sysctl kern.ipc.acceptqueue
of the same functionality. It specifies the maximum length of the
accept queue on a listen socket.
The old kern.ipc.somaxconn remains available for reading and writing
for compatibility reasons so that existing programs, scripts and
configurations continue to work. There no plans to ever remove the
orginal and now hidden kern.ipc.somaxconn.
MSG_WAITALL is set, and it is possible to do the entire receive
operation at once if we block (resid <= hiwat). Actually it might make
the recv(2) with MSG_WAITALL flag get stuck when there is enough space
in the receiver buffer to satisfy the request but not enough to open
the window closed previously due to the buffer being full.
The issue can be reproduced using the following scenario:
On the sender side do 2 send(2) requests:
1) data of size much smaller than SOBUF_SIZE (e.g. SOBUF_SIZE / 10);
2) data of size equal to SOBUF_SIZE.
On the receiver side do 2 recv(2) requests with MSG_WAITALL flag set:
1) recv() data of SOBUF_SIZE / 10 size;
2) recv() data of SOBUF_SIZE size;
We totally fill the receiver buffer with one SOBUF_SIZE/10 size request
and partial SOBUF_SIZE request. When the first request is processed we
get SOBUF_SIZE/10 free space. It is just enough to receive the rest of
bytes for the second request, and soreceive_generic() blocks in the
part that is a subject of this change waiting for the rest. But the
window was closed when the buffer was filled and to avoid silly window
syndrome it opens only when available space is larger than sb_hiwat/4
or maxseg. So it is stuck and pending data is only sent via TCP window
probes.
Discussed with: kib (long ago)
MFC after: 2 weeks
check it for MT_CONTROL type too, otherwise the assertion
"m->m_type == MT_DATA" below may be triggered by the following scenario:
- the sender sends some data (MT_DATA) and then a file descriptor
(MT_CONTROL);
- the receiver calls recv(2) with a MSG_WAITALL asking for data larger
than the receive buffer (uio_resid > hiwat).
MFC after: 2 week
socket protocol number. This is useful since the socket type can
be implemented by different protocols in the same protocol family,
e.g. SOCK_STREAM may be provided by both TCP and SCTP.
Submitted by: Jukka A. Ukkonen <jau iki fi>
PR: kern/162352
Discussed with: bz
Reviewed by: glebius
MFC after: 2 weeks
Add the sysctl debug.iosize_max_clamp, enabled by default. Setting the
sysctl to zero allows to perform the SSIZE_MAX-sized i/o requests from
the usermode.
Discussed with: bde, das (previous versions)
MFC after: 1 month
If a selinfo object is recorded (via selrecord()) and then it is
quickly destroyed, with the waiters missing the opportunity to awake,
at the next iteration they will find the selinfo object destroyed,
causing a PF#.
That happens because the selinfo interface has no way to drain the
waiters before to destroy the registered selinfo object. Also this
race is quite rare to get in practice, because it would require a
selrecord(), a poll request by another thread and a quick destruction
of the selrecord()'ed selinfo object.
Fix this by adding the seldrain() routine which should be called
before to destroy the selinfo objects (in order to avoid such case),
and fix the present cases where it might have already been called.
Sometimes, the context is safe enough to prevent this type of race,
like it happens in device drivers which installs selinfo objects on
poll callbacks. There, the destruction of the selinfo object happens
at driver detach time, when all the filedescriptors should be already
closed, thus there cannot be a race.
For this case, mfi(4) device driver can be set as an example, as it
implements a full correct logic for preventing this from happening.
Sponsored by: Sandvine Incorporated
Reported by: rstone
Tested by: pluknet
Reviewed by: jhb, kib
Approved by: re (bz)
MFC after: 3 weeks
o Move the non-blocking socket test below the SBS_CANTRCVMORE so that EOF
is correctly returned on a remote connection close.
o In the non-blocking socket test compare SS_NBIO against the so->so_state
field instead of the incorrect sb->sb_state field.
o Simplify the ENOTCONN test by removing cases that can't occur.
Submitted by: trociny (with some further tweaks by committer)
Tested by: trociny
larger than the receive buffer, we have to receive in sections.
When notifying the protocol that some data has been drained the
lock is released for a moment. Returning we block waiting for the
rest of data. There is a race, when data could arrive while the
lock was released and then the connection stalls in sbwait.
Fix this by checking for data before blocking and skip blocking
if there are some.
PR: kern/154504
Reported by: Andrey Simonenko <simon@comsys.ntu-kpi.kiev.ua>
Tested by: Andrey Simonenko <simon@comsys.ntu-kpi.kiev.ua>
Reviewed by: rwatson
Approved by: kib (co-mentor)
MFC after: 2 weeks
VNET socket push back:
try to minimize the number of places where we have to switch vnets
and narrow down the time we stay switched. Add assertions to the
socket code to catch possibly unset vnets as seen in r204147.
While this reduces the number of vnet recursion in some places like
NFS, POSIX local sockets and some netgraph, .. recursions are
impossible to fix.
The current expectations are documented at the beginning of
uipc_socket.c along with the other information there.
Sponsored by: The FreeBSD Foundation
Sponsored by: CK Software GmbH
Reviewed by: jhb
Tested by: zec
Tested by: Mikolaj Golub (to.my.trociny gmail.com)
MFC after: 2 weeks
Make VNET_ASSERT() available with either VNET_DEBUG or INVARIANTS.
Change the syntax to match KASSERT() to allow more flexible panic
messages rather than having a printf with hardcoded arguments
before panic.
Adjust the few assertions we have to the new format (and enhance
the output).
Sponsored by: The FreeBSD Foundation
Sponsored by: CK Software GmbH
Reviewed by: jhb
MFC after: 2 weeks
you tag a socket with an uint32_t value. The cookie can then be
used by the kernel for various purposes, e.g. setting the skipto
rule or pipe number in ipfw (this is the reason SO_USER_COOKIE has
been implemented; however there is nothing ipfw-specific in its
implementation).
The ipfw-related code that uses the optopn will be committed separately.
This change adds a field to 'struct socket', but the struct is not
part of any driver or userland-visible ABI so the change should be
harmless.
See the discussion at
http://lists.freebsd.org/pipermail/freebsd-ipfw/2009-October/004001.html
Idea and code from Paul Joe, small modifications and manpage
changes by myself.
Submitted by: Paul Joe
MFC after: 1 week
was eliminated: all references to sockets are explicitly managed by sorele()
and the protocols. As such, garbage collect sotryfree(), and update
sofree() comments to make the new world order more clear.
MFC after: 3 days
Reported by: Anuranjan Shukla <anshukla at juniper dot net>
SOCK_DGRAM socket. MSG_TRUNC was only returned when some mbufs could
not be copied to the application. If some data was left in the last
mbuf, it was correctly discarded, but MSG_TRUNC was not set.
Reviewed by: bz
MFC after: 3 weeks
set but be cleared before the call to sodisconnect(). In this case,
ENOTCONN is returned: suppress this error rather than returning it to
userspace so that close() doesn't report an error improperly.
PR: kern/144061
Reported by: Matt Reimer <mreimer at vpop.net>,
Nikolay Denev <ndenev at gmail.com>,
Mikolaj Golub <to.my.trociny at gmail.com>
MFC after: 3 days
for upcoming 64-bit PowerPC and MIPS support. This renames the COMPAT_IA32
option to COMPAT_FREEBSD32, removes some IA32-specific code from MI parts
of the kernel and enhances the freebsd32 compatibility code to support
big-endian platforms.
Reviewed by: kib, jhb
TCP_SORECEIVE_STREAM for the time being.
Requested by: brooks
Once compiled in make it easily switchable for testers by using a tuneable
net.inet.tcp.soreceive_stream
and a corresponding read-only sysctl to report the current state.
Suggested by: rwatson
MFC after: 2 days
This reverts part of r196460, so that sockets only return POLLHUP if both
directions are closed/error. Fifos get POLLHUP by closing the unused
direction immediately after creating the sockets.
The tools/regression/poll/*poll.c tests now pass except for two other things:
- if POLLHUP is returned, POLLIN is always returned as well instead of only
when there is data left in the buffer to be read
- fifo old/new reader distinction does not work the way POSIX specs it
Reviewed by: kib, bde
returning POLLHUP instead of POLLIN for several cases. Now, the
tools/regression/poll results for FreeBSD are closer to that of the
Solaris and Linux.
Also, improve the POSIX conformance by explicitely clearing POLLOUT
when POLLHUP is reported in pollscan(), making the fix global.
Submitted by: bde
Reviewed by: rwatson
MFC after: 1 week
vnet.h, we now use jails (rather than vimages) as the abstraction
for virtualization management, and what remained was specific to
virtual network stacks. Minor cleanups are done in the process,
and comments updated to reflect these changes.
Reviewed by: bz
Approved by: re (vimage blanket)
when all writers, observed by reader, exited. Use writer generation
counter for fifo, and store the snapshot of the fifo generation in the
f_seqcount field of struct file, that is otherwise unused for fifos.
Set FreeBSD-undocumented POLLINIGNEOF flag only when file f_seqcount is
equal to fifo' fi_wgen, and revert r89376.
Fix POLLINIGNEOF for sockets and pipes, and return POLLHUP for them.
Note that the patch does not fix not returning POLLHUP for fifos.
PR: kern/94772
Submitted by: bde (original version)
Reviewed by: rwatson, jilles
Approved by: re (kensmith)
MFC after: 6 weeks (might be)
stream (TCP) sockets.
It is functionally identical to generic soreceive() but has a
number stream specific optimizations:
o does only one sockbuf unlock/lock per receive independent of
the length of data to be moved into the uio compared to
soreceive() which unlocks/locks per *mbuf*.
o uses m_mbuftouio() instead of its own copy(out) variant.
o much more compact code flow as a large number of special
cases is removed.
o much improved reability.
It offers significantly reduced CPU usage and lock contention
when receiving fast TCP streams. Additional gains are obtained
when the receiving application is using SO_RCVLOWAT to batch up
some data before a read (and wakeup) is done.
This function was written by "reverse engineering" and is not
just a stripped down variant of soreceive().
It is not yet enabled by default on TCP sockets. Instead it is
commented out in the protocol initialization in tcp_usrreq.c
until more widespread testing has been done.
Testers, especially with 10GigE gear, are welcome.
MFP4: r164817 //depot/user/andre/soreceive_stream/
vnode interlock to protect the knote fields [1]. The locking assumes
that shared vnode lock is held, thus we get exclusive access to knote
either by exclusive vnode lock protection, or by shared vnode lock +
vnode interlock.
Do not use kl_locked() method to assert either lock ownership or the
fact that curthread does not own the lock. For shared locks, ownership
is not recorded, e.g. VOP_ISLOCKED can return LK_SHARED for the shared
lock not owned by curthread, causing false positives in kqueue subsystem
assertions about knlist lock.
Remove kl_locked method from knlist lock vector, and add two separate
assertion methods kl_assert_locked and kl_assert_unlocked, that are
supposed to use proper asserts. Change knlist_init accordingly.
Add convenience function knlist_init_mtx to reduce number of arguments
for typical knlist initialization.
Submitted by: jhb [1]
Noted by: jhb [2]
Reviewed by: jhb
Tested by: rnoland
and used in a large number of files, but also because an increasing number
of incorrect uses of MAC calls were sneaking in due to copy-and-paste of
MAC-aware code without the associated opt_mac.h include.
Discussed with: pjd
count of the number of registered policies.
Rather than unconditionally locking sockets before passing them into MAC,
lock them in the MAC entry points only if mac_policy_count is non-zero.
This avoids locking overhead for a number of socket system calls when no
policies are registered, eliminating measurable overhead for the MAC
Framework for the socket subsystem when there are no active policies.
Possibly socket locks should be acquired by policies if they are required
for socket labels, which would further avoid locking overhead when there
are policies but they don't require labeling of sockets, or possibly
don't even implement socket controls.
Obtained from: TrustedBSD Project
