Unmapped mbufs allow sendfile to carry multiple pages of data in a
single mbuf, without mapping those pages. It is a requirement for
Netflix's in-kernel TLS, and provides a 5-10% CPU savings on heavy web
serving workloads when used by sendfile, due to effectively
compressing socket buffers by an order of magnitude, and hence
reducing cache misses.
For this new external mbuf buffer type (EXT_PGS), the ext_buf pointer
now points to a struct mbuf_ext_pgs structure instead of a data
buffer. This structure contains an array of physical addresses (this
reduces cache misses compared to an earlier version that stored an
array of vm_page_t pointers). It also stores additional fields needed
for in-kernel TLS such as the TLS header and trailer data that are
currently unused. To more easily detect these mbufs, the M_NOMAP flag
is set in m_flags in addition to M_EXT.
Various functions like m_copydata() have been updated to safely access
packet contents (using uiomove_fromphys()), to make things like BPF
safe.
NIC drivers advertise support for unmapped mbufs on transmit via a new
IFCAP_NOMAP capability. This capability can be toggled via the new
'nomap' and '-nomap' ifconfig(8) commands. For NIC drivers that only
transmit packet contents via DMA and use bus_dma, adding the
capability to if_capabilities and if_capenable should be all that is
required.
If a NIC does not support unmapped mbufs, they are converted to a
chain of mapped mbufs (using sf_bufs to provide the mapping) in
ip_output or ip6_output. If an unmapped mbuf requires software
checksums, it is also converted to a chain of mapped mbufs before
computing the checksum.
Submitted by: gallatin (earlier version)
Reviewed by: gallatin, hselasky, rrs
Discussed with: ae, kp (firewalls)
Relnotes: yes
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D20616
- Perform ifp mismatch checks (to determine if a send tag is allocated
for a different ifp than the one the packet is being output on), in
ip_output() and ip6_output(). This avoids sending packets with send
tags to ifnet drivers that don't support send tags.
Since we are now checking for ifp mismatches before invoking
if_output, we can now try to allocate a new tag before invoking
if_output sending the original packet on the new tag if allocation
succeeds.
To avoid code duplication for the fragment and unfragmented cases,
add ip_output_send() and ip6_output_send() as wrappers around
if_output and nd6_output_ifp, respectively. All of the logic for
setting send tags and dealing with send tag-related errors is done
in these wrapper functions.
For pseudo interfaces that wrap other network interfaces (vlan and
lagg), wrapper send tags are now allocated so that ip*_output see
the wrapper ifp as the ifp in the send tag. The if_transmit
routines rewrite the send tags after performing an ifp mismatch
check. If an ifp mismatch is detected, the transmit routines fail
with EAGAIN.
- To provide clearer life cycle management of send tags, especially
in the presence of vlan and lagg wrapper tags, add a reference count
to send tags managed via m_snd_tag_ref() and m_snd_tag_rele().
Provide a helper function (m_snd_tag_init()) for use by drivers
supporting send tags. m_snd_tag_init() takes care of the if_ref
on the ifp meaning that code alloating send tags via if_snd_tag_alloc
no longer has to manage that manually. Similarly, m_snd_tag_rele
drops the refcount on the ifp after invoking if_snd_tag_free when
the last reference to a send tag is dropped.
This also closes use after free races if there are pending packets in
driver tx rings after the socket is closed (e.g. from tcpdrop).
In order for m_free to work reliably, add a new CSUM_SND_TAG flag in
csum_flags to indicate 'snd_tag' is set (rather than 'rcvif').
Drivers now also check this flag instead of checking snd_tag against
NULL. This avoids false positive matches when a forwarded packet
has a non-NULL rcvif that was treated as a send tag.
- cxgbe was relying on snd_tag_free being called when the inp was
detached so that it could kick the firmware to flush any pending
work on the flow. This is because the driver doesn't require ACK
messages from the firmware for every request, but instead does a
kind of manual interrupt coalescing by only setting a flag to
request a completion on a subset of requests. If all of the
in-flight requests don't have the flag when the tag is detached from
the inp, the flow might never return the credits. The current
snd_tag_free command issues a flush command to force the credits to
return. However, the credit return is what also frees the mbufs,
and since those mbufs now hold references on the tag, this meant
that snd_tag_free would never be called.
To fix, explicitly drop the mbuf's reference on the snd tag when the
mbuf is queued in the firmware work queue. This means that once the
inp's reference on the tag goes away and all in-flight mbufs have
been queued to the firmware, tag's refcount will drop to zero and
snd_tag_free will kick in and send the flush request. Note that we
need to avoid doing this in the middle of ethofld_tx(), so the
driver grabs a temporary reference on the tag around that loop to
defer the free to the end of the function in case it sends the last
mbuf to the queue after the inp has dropped its reference on the
tag.
- mlx5 preallocates send tags and was using the ifp pointer even when
the send tag wasn't in use. Explicitly use the ifp from other data
structures instead.
- Sprinkle some assertions in various places to assert that received
packets don't have a send tag, and that other places that overwrite
rcvif (e.g. 802.11 transmit) don't clobber a send tag pointer.
Reviewed by: gallatin, hselasky, rgrimes, ae
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D20117
Drivers can now pass up numa domain information via the
mbuf numa domain field. This information is then used
by TCP syncache_socket() to associate that information
with the inpcb. The domain information is then fed back
into transmitted mbufs in ip{6}_output(). This mechanism
is nearly identical to what is done to track RSS hash values
in the inp_flowid.
Follow on changes will use this information for lacp egress
port selection, binding TCP pacers to the appropriate NUMA
domain, etc.
Reviewed by: markj, kib, slavash, bz, scottl, jtl, tuexen
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D20028
When large SPDs are used, we face two problems:
- too many CPU cycles are spent during the linear searches in the SPD
for each packet
- too much contention on multi socket systems, since we use a single
shared lock.
Main changes:
- added the sysctl tree 'net.key.spdcache' to control the SPD cache
(disabled by default).
- cache the sp indexes that are used to perform SP lookups.
- use a range of dedicated mutexes to protect the cache lines.
Submitted by: Emeric Poupon <emeric.poupon@stormshield.eu>
Reviewed by: ae
Sponsored by: Stormshield
Differential Revision: https://reviews.freebsd.org/D15050
Mainly focus on files that use BSD 3-Clause license.
The Software Package Data Exchange (SPDX) group provides a specification
to make it easier for automated tools to detect and summarize well known
opensource licenses. We are gradually adopting the specification, noting
that the tags are considered only advisory and do not, in any way,
superceed or replace the license texts.
Special thanks to Wind River for providing access to "The Duke of
Highlander" tool: an older (2014) run over FreeBSD tree was useful as a
starting point.
- Fix it by replacing m_cat() with m_prev->m_next = m_new
(m_cat() will try to append data - as a result, there will be no
fragmentation).
- Move some constants out of the loop.
Was previously tested with D4077.
Differential Revision: https://reviews.freebsd.org/D4090
in mb_free_ext() always use fields from the original refcount holding
mbuf (see. r296242) mbuf. Cuts another cache miss from mb_free_ext().
However, treat EXT_EXTREF mbufs differently, since they are different -
they don't have a refcount holding mbuf.
Provide longer comments in m_ext declaration to explain this change
and change from r296242.
In collaboration with: gallatin
Differential Revision: https://reviews.freebsd.org/D12615
m_unshare passes on the source mbuf's flags as-is to m_getcl and this
results in a leak if the flags include M_NOFREE. The fix is to clear
the bits not listed in M_COPYALL before calling m_getcl. M_RDONLY
should probably be filtered out too but that's outside the scope of this
fix.
Add assertions in the zone_mbuf and zone_pack ctors to catch similar
bugs.
Update netmap_get_mbuf to not pass M_NOFREE to m_getcl. It's not clear
what the original code was trying to do but it's likely incorrect.
Updated code is no different functionally but it avoids the newly added
assertions.
Reviewed by: gnn@
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D5698
The mbuf provider is made up of a set of Statically Defined Tracepoints
which help us look into mbufs as they are allocated and freed. This can be
used to inspect the buffers or for a simplified mbuf leak detector.
New tracepoints are:
mbuf:::m-init
mbuf:::m-gethdr
mbuf:::m-get
mbuf:::m-getcl
mbuf:::m-clget
mbuf:::m-cljget
mbuf:::m-cljset
mbuf:::m-free
mbuf:::m-freem
There is also a translator for mbufs which gives some visibility into the structure,
see mbuf.d for more details.
Reviewed by: bz, markj
MFC after: 2 weeks
Sponsored by: Rubicon Communications (Netgate)
Differential Revision: https://reviews.freebsd.org/D5682
The m_ext.ext_cnt pointer becomes a union. It can now hold the refcount
value itself. To tell that m_ext.ext_flags flag EXT_FLAG_EMBREF is used.
The first mbuf to attach a cluster stores the refcount. The further mbufs
to reference the cluster point at refcount in the first mbuf. The first
mbuf is freed only when the last reference is freed.
The benefit over refcounts stored in separate slabs is that now refcounts
of different, unrelated mbufs do not share a cache line.
For EXT_EXTREF mbufs the zone_ext_refcnt is no longer needed, and m_extadd()
becomes void, making widely used M_EXTADD macro safe.
For EXT_SFBUF mbufs the sf_ext_ref() is removed, which was an optimization
exactly against the cache aliasing problem with regular refcounting.
Discussed with: rrs, rwatson, gnn, hiren, sbruno, np
Reviewed by: rrs
Differential Revision: https://reviews.freebsd.org/D5396
Sponsored by: Netflix
mbuf(9) manipulation functions into uipc_mbuf.c. This looks like
the initial intent, but had diffused in the last decade.
o Gather all declarations in mbuf.h in one place and sort them.
o Uninline m_clget() and m_cljget().
There are no functional changes in this patch.
The patch comes from a larger version, where all mbuf(9) allocation was
uninlined, which allowed to make mbuf(9) UMA zones private to kern_mbuf.c.
The performance impact of the total uninlining is still unclear, so we
are holding on now with larger version.
Together with: melifaro, olivier
up to now.
The new sendfile is the code that Netflix uses to send their multiple tens
of gigabits of data per second. The new implementation features asynchronous
I/O, when I/O operations are launched, but not awaited to be complete. An
explanation of why such behavior is beneficial compared to old one is
going to be too long for a commit message, so we will skip it here.
Additional features of new syscall are extra flags, which provide an
application more control over data sent. The SF_NOCACHE flag tells
kernel that data shouldn't be cached after it was sent. The SF_READAHEAD()
macro allows to specify readahead size in pages.
The new syscalls is a drop in replacement. No modifications are required
to applications. One can take nginx binary for stable/10 and run it
successfully on head. Although SF_NODISKIO lost its original sense, as now
sendfile doesn't block, and now means something completely different (tm),
using the new sendfile the old way is absolutely safe.
Celebrates: Netflix global launch!
Sponsored by: Nginx, Inc.
Sponsored by: Netflix
Relnotes: yes
packets and/or state transitions from each TCP socket. That would help with
narrowing down certain problems we see in the field that are hard to reproduce
without understanding the history of how we got into a certain state. This
change provides just that.
It saves copies of the last N packets in a list in the tcpcb. When the tcpcb is
destroyed, the list is freed. I thought this was likely to be more
performance-friendly than saving copies of the tcpcb. Plus, with the packets,
you should be able to reverse-engineer what happened to the tcpcb.
To enable the feature, you will need to compile a kernel with the TCPPCAP
option. Even then, the feature defaults to being deactivated. You can activate
it by setting a positive value for the number of captured packets. You can do
that on either a global basis or on a per-socket basis (via a setsockopt call).
There is no way to get the packets out of the kernel other than using kmem or
getting a coredump. I thought that would help some of the legal/privacy concerns
regarding such a feature. However, it should be possible to add a future effort
to export them in PCAP format.
I tested this at low scale, and found that there were no mbuf leaks and the peak
mbuf usage appeared to be unchanged with and without the feature.
The main performance concern I can envision is the number of mbufs that would be
used on systems with a large number of sockets. If you save five packets per
direction per socket and have 3,000 sockets, that will consume at least 30,000
mbufs just to keep these packets. I tried to reduce the concerns associated with
this by limiting the number of clusters (not mbufs) that could be used for this
feature. Again, in my testing, that appears to work correctly.
Differential Revision: D3100
Submitted by: Jonathan Looney <jlooney at juniper dot net>
Reviewed by: gnn, hiren
mbuf. Unlike in the pre-r248371 code, assert that M_PKTHDR is set
only on a first mbuf.
Reported & tested by: Andriy Voskoboinyk <s3erios gmail.com>
Sponsored by: Nginx, Inc.
m_dup_pkthdr(), that uses M_COPYFLAGS mask to copy m_flags field.
If original mbuf chain has M_RDONLY flag, its copy also will have it.
Reset this flag explicitly.
MFC after: 2 weeks
instead of preprocessor macros.
This will make debugger output of 'print *m' exactly match the names
we use in code, making life of a kernel hacker way more pleasant. And
this also allows to rename struct_m_ext back to m_ext.
STAILQs and SLISTs using the same structure field as good old m_next
and m_nextpkt linkage occupy.
New code is encouraged to use queue(3) macros, instead of implementing
the wheel. However, better not to have a mixture of old style and
queue(3) in one file or subsystem.
Reviewed by: rwatson, rrs, rpaulo
Differential Revision: D1499
more generally make it easier to extend 'struct mbuf in the future', make
a number of changes to the data structure:
- As we anticipate embedding mbufs headers within variable-size regions of
memory in the future, change the definitions of byte arrays embedded in
mbufs to be of size [0] rather than [MLEN] and [MHLEN]. In fact, the
cxgbe driver already uses 'struct mbuf' on the front of other storage
sizes, but we would like the global mbuf allocator do be able to do this
as well.
- Fold 'struct m_hdr' into 'struct mbuf' itself, eliminating a set of
macros that aliased 'mh_foo' field names to 'm_foo' names such as
'm_next'. These present a particular problem as we would like to add
new mbuf-header fields -- e.g., 'm_size' -- that, if similarly named via
macros, would introduce collisions with many other variable names in the
kernel.
- Rename 'struct m_ext' to 'struct struct_m_ext' so that we can add
compile-time assertions without bumping into the still-extant 'm_ext'
macro.
- Remove the MSIZE compile-time assertion for 'struct mbuf', but add new
assertions for alignment of embedded data arrays (64-bit alignment even
on 32-bit platforms), and for the sizes the mbuf header, packet header,
and m_ext structure.
- Document that these assertions exist in comments in mbuf.h.
This change is not intended to cause (non-trivial) behavioural
differences, but is a precursor to further mbuf-allocator work.
Differential Revision: https://reviews.freebsd.org/D1483
Reviewed by: bz, gnn, np, glebius ("go ahead, I trust you")
Sponsored by: EMC / Isilon Storage Division
in FreeBSD 7 that has not been used since. It contains a number
of unresolved bugs including an inverted bcopy() and incorrect
handling of read-only mbufs using internal storage. Removing this
unused code is substantially essier than fixing it in order to
update it to the coming mbuf world order -- but it can always be
restored from revision history if it turns out to prove useful for
future work.
Pointed out by: jmallett
Sponsored by: EMC / Isilon Storage Division
with calls to the centralised macros, reducing direct use of MLEN and
MHLEN.
Differential Revision: https://reviews.freebsd.org/D1444
Reviewed by: bz
Sponsored by: EMC / Isilon Storage Division
the knowledge of mbuf layout, and in particular constants such as M_EXT,
MLEN, MHLEN, and so on, in mbuf consumers by unifying various alignment
utility functions (M_ALIGN(), MH_ALIGN(), MEXT_ALIGN() in a single
M_ALIGN() macro, implemented by a now-inlined m_align() function:
- Move m_align() from uipc_mbuf.c to mbuf.h; mark as __inline.
- Reimplement M_ALIGN(), MH_ALIGN(), and MEXT_ALIGN() using m_align().
- Update consumers around the tree to simply use M_ALIGN().
This change eliminates a number of cases where mbuf consumers must be aware
of whether or not mbufs returned by the allocator use external storage, but
also assumptions about the size of the returned mbuf. This will make it
easier to introduce changes in how we use external storage, as well as
features such as variable-size mbufs.
Differential Revision: https://reviews.freebsd.org/D1436
Reviewed by: glebius, trasz, gnn, bz
Sponsored by: EMC / Isilon Storage Division
sending not ready data:
o Add new flag to pru_send() flags - PRUS_NOTREADY.
o Add new protocol method pru_ready().
Sponsored by: Nginx, Inc.
Sponsored by: Netflix
o Do not use UMA refcount zone. The problem with this zone is that
several refcounting words (16 on amd64) share the same cache line,
and issueing atomic(9) updates on them creates cache line contention.
Also, allocating and freeing them is extra CPU cycles.
Instead, refcount the page directly via vm_page_wire() and the sfbuf
via sf_buf_alloc(sf_buf_page(sf)) [1].
o Call refcounting/freeing function for EXT_SFBUF via direct function
call, instead of function pointer. This removes barrier for CPU
branch predictor.
o Do not cleanup the mbuf to be freed in mb_free_ext(), merely to
satisfy assertion in mb_dtor_mbuf(). Remove the assertion from
mb_dtor_mbuf(). Use bcopy() instead of manual assignments to
copy m_ext in mb_dupcl().
[1] This has some problems for now. Using sf_buf_alloc() merely to
increase refcount is expensive, and is broken on sparc64. To be
fixed.
Sponsored by: Netflix
Sponsored by: Nginx, Inc.
problems with the way MLEN, MHLEN, and struct mbuf are set up.
CTASSERT's are provided to detect such issues at compile time in the
future.
The #define MLEN and MHLEN calculation do not take actual compiler-
induced alignment and padding inside the complete struct mbuf into
account. Accordingly appropriate attention is required when changing
members of struct mbuf.
Ideally one would calculate MLEN as (MSIZE - sizeof(((struct mbuf *)0)->m_hdr)
but that doesn't work as the compiler refuses to operate on an as of
yet incomplete structure.
In particular ARM 32bit has more strict alignment requirements which
caused 4 bytes of padding between m_hdr and pkthdr in struct mbuf
because of the 64bit members in pkthdr. This wasn't picked up by MLEN
and MHLEN causing an overflow of the mbuf provided data storage by
overestimating its size.
I386 didn't show this problem because it handles unaligned access just
fine, albeit at a small performance penalty.
On 64bit architectures the struct mbuf layout is 64bit aligned in all
places.
Reported by: Thomas Skibo <ThomasSkibo-at-sbcglobal-dot-net>
Tested by: tuexen, ian, Thomas Skibo (extended patch)
Sponsored by: The FreeBSD Foundation
external mbuf buffer management capabilities in the future.
For now only EXT_FREE_OK is defined with current legacy behavior.
Sponsored by: The FreeBSD Foundation
features. The changes in particular are:
o Remove rarely used "header" pointer and replace it with a 64bit protocol/
layer specific union PH_loc for local use. Protocols can flexibly overlay
their own 8 to 64 bit fields to store information while the packet is
worked on.
o Mechanically convert IP reassembly, IGMP/MLD and ATM to use pkthdr.PH_loc
instead of pkthdr.header.
o Extend csum_flags to 64bits to allow for additional future offload
information to be carried (e.g. iSCSI, IPsec offload, and others).
o Move the RSS hash type enumerator from abusing m_flags to its own 8bit
rsstype field. Adjust accessor macros.
o Add cosqos field to store Class of Service / Quality of Service information
with the packet. It is not yet supported in any drivers but allows us to
get on par with Cisco/Juniper in routing applications (plus MPLS QoS) with
a modernized ALTQ.
o Add four 8 bit fields l[2-5]hlen to store the relative header offsets
from the start of the packet. This is important for various offload
capabilities and to relieve the drivers from having to parse the packet
and protocol headers to find out location of checksums and other
information. Header parsing in drivers is a lot of copy-paste and
unhandled corner cases which we want to avoid.
o Add another flexible 64bit union to map various additional persistent
packet information, like ether_vtag, tso_segsz and csum fields.
Depending on the csum_flags settings some fields may have different usage
making it very flexible and adaptable to future capabilities.
o Restructure the CSUM flags to better signify their outbound (down the
stack) and inbound (up the stack) use. The CSUM flags used to be a bit
chaotic and rather poorly documented leading to incorrect use in many
places. Bring clarity into their use through better naming.
Compatibility mappings are provided to preserve the API. The drivers
can be corrected one by one and MFC'd without issue.
o The size of pkthdr stays the same at 48/56bytes (32/64bit architectures).
Sponsored by: The FreeBSD Foundation
free function access to the mbuf the external memory was attached
to.
Mechanically adjust all users to include the mbuf parameter.
This fixes a long standing annoyance for external free functions.
Before one had to sacrifice one of the argument pointers for this.
Sponsored by: The FreeBSD Foundation
to 8 bits. ext_type is an enumerator and the number of types we
have is a mere dozen.
A couple of ext_types are renumbered to fit within 8 bits.
EXT_VENDOR[1-4] and EXT_EXP[1-4] types for vendor-internal and
experimental local mapping.
The ext_flags field is currently unused but has a couple of flags
already defined for future use. Again vendor and experimental
flags are provided for local mapping.
EXT_FLAG_BITS is provided for the printf(9) %b identifier.
Initialize and copy ext_flags in the relevant mbuf functions.
Improve alignment and packing of struct m_ext on 32 and 64 archs
by carefully sorting the fields.
for a very long time, if ever.
Should such a functionality ever be needed again the appropriate and
much better way to do it is through a custom EXT_SOMETHING external mbuf
type together with a dedicated *ext_free function.
Discussed with: trociny, glebius
Now that r253351 moved sendfile() stats to a separate struct, the
last field used in mbstat is m_mcfail, which is updated, but never
read or obtained from userland.
