We've been seeing lots of cache line contention (but not lock contention!)
in our workloads between the various TX and RX threads going on.
The write lock is only grabbed when configuration changes are made - which
are infrequent.
With this patch, the contention and cycles spent waiting for updates
disappear.
Sponsored by: Netflix, Inc.
- Remove excessive parenthesis
- Use KNF continuation indentation
- Cut down on excessive continuation lines
- More consistent style in messages
- Use uprintf() instead of printf()
Submitted by: bde
calls ns8250_bus_ipend() almost immediately after ns8250_bus_attach().
As it appears, a line break condition is being signalled for almost
all received characters due to this. A delay of 150ms seems enough
to allow the H/W to settle and to avoid the problem.
More analysis is needed, but for now a regression has been addressed.
Reported by: kevlo@
Tested by: kevlo@
MADV_DONTNEED) and madvise(..., MADV_FREE). Specifically, introduce a new
pmap function, pmap_advise(), that operates on a range of virtual addresses
within the specified pmap, allowing for a more efficient implementation of
MADV_DONTNEED and MADV_FREE. Previously, the implementation of
MADV_DONTNEED and MADV_FREE relied on per-page pmap operations, such as
pmap_clear_reference(). Intuitively, the problem with this implementation
is that the pmap-level locks are acquired and released and the page table
traversed repeatedly, once for each resident page in the range
that was specified to madvise(2). A more subtle flaw with the previous
implementation is that pmap_clear_reference() would clear the reference bit
on all mappings to the specified page, not just the mapping in the range
specified to madvise(2).
Since our malloc(3) makes heavy use of madvise(2), this change can have a
measureable impact. For example, the system time for completing a parallel
"buildworld" on a 6-core amd64 machine was reduced by about 1.5% to 2.0%.
Note: This change only contains pmap_advise() implementations for a subset
of our supported architectures. I will commit implementations for the
remaining architectures after further testing. For now, a stub function is
sufficient because of the advisory nature of pmap_advise().
Discussed with: jeff, jhb, kib
Tested by: pho (i386), marcel (ia64)
Sponsored by: EMC / Isilon Storage Division
(re)start the interface when it is down. This change fix a race with
BOOTP where the response packet is lost because the interface is being
reset by a netmask change right after send the packet.
PR: 178318
Approved by: adrian (mentor)
Add a last-modified timestamp to each LRO entry and provide an interface
to flush all inactive entries. Drivers decide when to flush and what
the inactivity threshold should be.
Network drivers that process an rx queue to completion can enter a
livelock type situation when the rate at which packets are received
reaches equilibrium with the rate at which the rx thread is processing
them. When this happens the final LRO flush (normally when the rx
routine is done) does not occur. Pure ACKs and segments with total
payload < 64K can get stuck in an LRO entry. Symptoms are that TCP
tx-mostly connections' performance falls off a cliff during heavy,
unrelated rx on the interface.
Flushing only inactive LRO entries works better than any of these
alternates that I tried:
- don't LRO pure ACKs
- flush _all_ LRO entries periodically (every 'x' microseconds or every
'y' descriptors)
- stop rx processing in the driver periodically and schedule remaining
work for later.
Reviewed by: andre
UF_SYSTEM, UF_SPARSE, UF_OFFLINE, UF_REPARSE, UF_ARCHIVE, UF_READONLY,
and UF_HIDDEN.
Sort the file flags tmpfs supports alphabetically. tmpfs now
supports the same flags as UFS, with the exception of SF_SNAPSHOT.
Reported by: bdrewery, antoine
Sponsored by: Spectra Logic
- tom_uninit had to be reworked not to hold the adapter lock (a mutex)
around t4_deactivate_uld, which acquires the uld_list_lock.
- the ifc_match for the interface cloner that creates the tracer ifnet
had to be reworked as the kernel calls ifc_match with the global
if_cloners_mtx held.
allocations under low free-space conditions (-r254995), determine
that old block-preference search order used before -r249782 worked
a bit better. This change reverts to that block-preference search order.
MFC after: 2 weeks
I have 25TB Dell PERC 6 RAID5 array. When it becomes almost
full (10-20GB free), processes which write data to it start
eating 100% CPU and write speed drops below 1MB/sec (normally
to gives 400MB/sec). The revision at which it first became
apparent was http://svnweb.freebsd.org/changeset/base/249782.
The offending change reserved an area in each cylinder group to
store metadata. The new algorithm attempts to save this area for
metadata and allows its use for non-metadata only after all the
data areas have been exhausted. The size of the reserved area
defaults to half of minfree, so the filesystem reports full before
the data area can completely fill. However, in this report, the
filesystem has had minfree reduced to 1% thus forcing the metadata
area to be used for data. As the filesystem approached full, it
had only metadata areas left to allocate. The result was that
every block allocation had to scan summary data for 30,000 cylinder
groups before falling back to searching up to 30,000 metadata areas.
The fix is to give up on saving the metadata areas once the free
space reserve drops below 2%. The effect of this change is to use
the old algorithm of just accepting the first available block that
we find. Since most filesystems use the default 5% minfree, this
will have no effect on their operation. For those that want to push
to the limit, they will get their crappy block placements quickly.
Submitted by: Dmitry Sivachenko
Fix Tested by: Dmitry Sivachenko
PR: kern/181226
MFC after: 2 weeks
the passed vnode belongs to the same mount point (v_vfsp or also
known as v_mount in FreeBSD). This check prevents the code from
proceeding further on vnodes that do not belong to ZFS, for
instance, on UFS or NULLFS.
The recent change (merged as r254585) on upstream changes the
check of v_vfsp to instead check the znode's z_zfsvfs. On Illumos
this would work because when the vnode comes from lofs, the
VOP_REALVP() would give the right vnode, this is not true on
FreeBSD where our VOP_REALVP is a no-op, and as such tdvp is
not guaranteed to be a ZFS vnode, and will later trigger a
failed assertion when verifying the vnode.
This changeset modifies our local shims (zfs_freebsd_rename and
zfs_freebsd_link) to check if v_mount matches before proceeding
further.
Reported by: many
Diagnostic work by: avg
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
notify (enable spinup) required", instead of doing the normal
retries, poll for a change in status.
We will poll every half second for a minute for the status to
change.
Hitachi drives (and likely other SAS drives) return that ASC/ASCQ
when they are waiting to spin up. What it means is that they are
waiting for the SAS expander to send them the SAS
NOTIFY (ENABLE SPINUP) primitive.
That primitive is the mechanism expanders/enclosures use to
sequence drive spinup to avoid overloading power supplies.
Sponsored by: Spectra Logic
MFC after: 3 days
The aim of this function is to eventually be the completion entry point
for all 802.11 encapsulated mbufs. All the wifi drivers end up doing
what is in this function so it's an easy win to turn it into a net80211
method and abstract out this code.
Ideally the drivers will all eventually be modified to queue up completed
mbufs and call this function with all the driver locks not held.
This will allow for some much more interesting software queue handling
in the future (like net80211 based A-MSDU, fast-frames, A-MPDU aggregation
and retransmission.)
Tested:
* ath(4), iwn(4)
- Use queue size fields from the Tx/Rx queues in various places
instead of (currently the same values) from the softc.
- Fix potential crash in detach if the attached failed to alloc
queue memory.
- Move the VMXNET3_MAX_RX_SEGS define to a better spot.
- Tweak frame size calculation w.r.t. ETHER_ALIGN. This could be
tweaked some more, or removed since it probably doesn't matter
much for x86 (and the x86 class of machines this driver will
be used on).
- Route PCI interrupt for NIC
- Make "no mapping" warning more user-friendly: add device name and mention
that it's IRQ mapping
- Do not overlap ICUs' IO window with PCI devices' IO windows by starting
IO rman at offset 0x100
for the available pbuf when passed vnode is backing md(4). Other i/o
directed to the same md device might already hold pbufs, and then we
could deadlock since only our progress can free a pbuf needed for
wakeup.
Obtained from: projects/vm6
Reminded and tested by: pho
MFC after: 1 week
