Do not hold the process around the vm_fault() call from the trap()s.
The only operation which is prevented by the hold is the kernel stack
swapout for the faulted thread, which should be fine to allow.
Remove useless checks for NULL curproc or curproc->p_vmspace from the
trap_pfault() wrappers on x86 and powerpc.
Reviewed by: alc (previous version)
Sponsored by: The FreeBSD Foundation
MFC after: 2 weeks
Alexander Motin [Thu, 10 Sep 2015 12:40:31 +0000 (12:40 +0000)]
Reimplement CTL High Availability.
CTL HA functionality was originally implemented by Copan many years ago,
but large part of the sources was never published. This change includes
clean room implementation of the missing code and fixes for many bugs.
This code supports dual-node HA with ALUA in four modes:
- Active/Unavailable without interlink between nodes;
- Active/Standby with second node handling only basic LUN discovery and
reservation, synchronizing with the first node through the interlink;
- Active/Active with both nodes processing commands and accessing the
backing storage, synchronizing with the first node through the interlink;
- Active/Active with second node working as proxy, transfering all
commands to the first node for execution through the interlink.
Unlike original Copan's implementation, depending on specific hardware,
this code uses simple custom TCP-based protocol for interlink. It has
no authentication, so it should never be enabled on public interfaces.
The code may still need some polishing, but generally it is functional.
Michael Tuexen [Thu, 10 Sep 2015 10:23:23 +0000 (10:23 +0000)]
Zero out a local variable also when PURIFY is not defined.
This silence a warning brought up by valgrind whenever if_nametoindex
is used. This was already discussed in PR 166483, but the code
committed in r234329 guards the initilization with #ifdef PURIFY.
Therefore, valgrind still complains. Since this code is not performance
critical, always zero out the local variable to silence valgrind.
Adrian Chadd [Thu, 10 Sep 2015 04:05:58 +0000 (04:05 +0000)]
Also make kern.maxfilesperproc a boot time tunable.
Auto-tuning threshold discussions aside, it turns out that if you want
to lower this on say, rather memory-packed machines, you either set maxusers
or kern.maxfiles, or you set it in sysctl. The former is a non-exact
way to tune this; the latter doesn't actually affect anything in the
startup scripts.
This first occured because I wondered why the hell screen would take upwards
of 10 seconds to spawn a new screen. I then found python doing the same
thing during fork/exec of child processes - it calls close() on each FD
up to the current openfiles limit. On a 1TB machine this is like, 26 million
FDs per process. Ugh.
So:
* This allows it to be set early in /boot/loader.conf;
* It can be used to work around the ridiculous situation of
screen, python, etc doing a close() on potentially millions of FDs
even though you only have four open.
Tested:
* 4GB, 32GB, 64GB, 128GB, 384GB, 1TB systems with autotune, ensuring
screen and python forking doesn't result in some pretty hilariously
bad behaviour.
TODO:
* Note that the default login.conf sets openfiles-cur to unlimited,
effectively obeying kern.maxfilesperproc. Perhaps we should fix
this.
* .. and even if we do, we need to also ensure that daemons get
a soft limit of something reasonable and capped - they can request
more FDs themselves.
For open("name", O_DIRECTORY | O_CREAT), do not try to create the
named node, open(2) cannot create directories. But do allow the flag
combination to succeed if the directory already exists.
Declare the open("name", O_DIRECTORY | O_CREAT | O_EXCL) always
invalid for the same reason, since open(2) cannot create directory.
Note that there is an argument that O_DIRECTORY | O_CREAT should be
invalid always, regardless of the target directory existence or
O_EXCL. The current fix is conservative and allows the call to
succeed in the situation where it succeeded before the patch.
Reported by: Tom Ridge <freebsd@tom-ridge.com>
Reviewed by: rwatson
PR: 202892
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Andrew Turner [Wed, 9 Sep 2015 11:51:14 +0000 (11:51 +0000)]
Rework copyinstr to:
* Fail when the length passed in is 0
* Remove an unneeded increment of the count on success
* Return ENAMETOOLONG when the input pointer is too long
Remove a check which caused spurious SIGSEGV on usermode access to the
mapped address without valid pte installed, when parallel wiring of
the entry happen. The entry must be copy on write. If entry is COW
but was already copied, and parallel wiring set
MAP_ENTRY_IN_TRANSITION, vm_fault() would sleep waiting for the
MAP_ENTRY_IN_TRANSITION flag to clear. After that, the fault handler
is restarted and vm_map_lookup() or vm_map_lookup_locked() trip over
the check. Note that this is race, if the address is accessed after
the wiring is done, the entry does not fault at all.
There is no reason in the current kernel to disallow write access to
the COW wired entry if the entry permissions allow it. Initially this
was done in r24666, since that kernel did not supported proper
copy-on-write for wired text, which was fixed in r199869. The r251901
revision re-introduced the r24666 fix for the current VM.
Note that write access must clear MAP_ENTRY_NEEDS_COPY entry flag by
performing COW. In reverse, when MAP_ENTRY_NEEDS_COPY is set in
vmspace_fork(), the MAP_ENTRY_USER_WIRED flag is cleared. Put the
assert stating the invariant, instead of returning the error.
Reported and debugging help by: peter
Reviewed by: alc
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Warner Losh [Tue, 8 Sep 2015 17:47:56 +0000 (17:47 +0000)]
The swap pager is compatible with direct dispatch. It does its own
locking and doesn't sleep. Flag the consumer we create as such. In
addition, decrement the in flight index when we have an out of memory
error after having incremented it previously. This would have
prevented swapoff from working if the swap pager ever hit a resource
shortage trying to swap out something (the swap in path always waits
for a bio, so won't have this issue). Simplify the close logic by
abandoning the use of private and initializing the index to 1 and
dropping that reference when we previously set private.
Also, set sw_id only while sw_dev_mtx is held. This should only affect
swapping to a vnode, as opposed to a geom whose close always sets it to
NULL with sw_dev_mtx held.
Andrew Turner [Tue, 8 Sep 2015 16:06:04 +0000 (16:06 +0000)]
Allow us to set the console device tree node. This is needed as not all
vendor supplied device trees contain the needed properties for us to select
the correct uart to use as the kernel console.
An example of this would be to add the following to loader.conf.
hw.fdt.console="/smb/uart@f7113000"
The intention of this is slightly different than the existing
hw.uart.console option. The new option will mean the boot serial
configuration will be derived from the device node, while the existing
option expects the user to configure all this themselves.
Further work is planned to allow the uart configuration to be set based on
the stdout-path property devicetree bindings.
Sponsored by: ABT Systems Ltd
Differential Revision: https://reviews.freebsd.org/D3559
Jim Harris [Tue, 8 Sep 2015 16:05:18 +0000 (16:05 +0000)]
isci: check return value of pci_alloc_msix()
Certain VM guest types (VMware, Xen) do not support MSI, so pci_alloc_msix()
always fails. isci(4) was not properly detecting the allocation failure,
and would try to proceed with MSIx resource initialization rather than
reverting to INTx.
Reported and tested by: Bradley W. Dutton (brad-fbsd-stable@duttonbros.com)
MFC after: 3 days
Sponsored by: Intel
Jim Harris [Tue, 8 Sep 2015 15:59:55 +0000 (15:59 +0000)]
isci: explicitly enable/disable PCI busmaster
BIOS always enables PCI busmaster on the isci device, which effectively
worked around this omission. But when passing the isci device through
to a guest VM, the hypervisor will disable busmaster and isci will not
work without calling pci_enable_busmaster().
In the pthread_once(), if the initializer has already run, then the
calling thread is supposed to see accesses issued by the initializer.
This means that the read of the once_control->state variable should
have an acquire semantic. Use atomic_thread_fence_acq() when the
value read is ONCE_DONE, instead of straightforward atomic_load_acq(),
to only put a barrier when needed (*).
On the other hand, the updates of the once_control->state with the
intermediate progress state do not need to synchronize with other
state accesses, remove _acq suffix.
Reviewed by: alc (previous version)
Suggested by: alc (*)
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Alan Cox [Tue, 8 Sep 2015 04:18:57 +0000 (04:18 +0000)]
To simplify upcoming changes to the inactive queue scan, change the code
so that there is only one place where pages are freed and only one place
where pages are moved to the tail of the queue.
fd: make the common case in filecaps_copy work lockless
The filedesc lock is only needed if ioctls caps are present, which is a
rare situation. This is a step towards reducing the scope of the filedesc
lock.
As expected, things aren't as simple as hoped. Consequently, we have
no option but to use the smbios information to fill in the blanks.
It's a good thing UGA is a protocol of the past and GOP has all the
info we need.
Anyway, the logic has been tweaked a little to get the easier bits
of information up front. This includes the resolution and the frame
buffer address. Then we look at the smbios information and define
expected values as well as the missing bits (frame buffer offset and
stride). If the values obtained match the expect values, we fill in
the blanks and return. Otherwise we use the existing detection logic
to figure it out.
Rename the environment variables from uga_framebuffer abd uga_stride
to hw.efifb.address and hw.efifb.stride. The latter names are more
in line with other variable names.
We currently have hardcoded settings for:
1. Mid-2007 iMac (iMac7,1)
2. Late-2007 MacBook (MacBook3,1)
Andrew Turner [Mon, 7 Sep 2015 14:01:18 +0000 (14:01 +0000)]
Use load-acquire semantics while waiting for td_lock to be released. The
store should have release semantics and will have due to the dsb above it
so add a comment to explain this. [1]
While here update the code to not reload the current thread, it's already
in a register, we just need to not trash it.
Suggested by: kib [1]
Sponsored by: ABT Systems Ltd
Alexander Motin [Sun, 6 Sep 2015 09:54:56 +0000 (09:54 +0000)]
Move setting of media parameters inside open routines.
This is preparation for possibility to open/close media several times
per LUN life cycle. While there, rename variables to reduce confusion.
As additional bonus this allows to open read-only media, such as ZFS
snapshots.
Track changes to kern.maxvnodes and appropriately increase or decrease
the size of the name cache hash table (mapping file names to vnodes)
and the vnode hash table (mapping mount point and inode number to vnode).
An appropriate locking strategy is the key to changing hash table sizes
while they are in active use.
Auto-detect the UGA frame buffer and stride on a MacBook. We're
striking a delicate balance between exhaustive searching and
banking on assumptions. The environment variables can be used
as a fall-back anyway. With this change, all known and tested
Macs with only UGA should have a working console out of the
box... for now...
Alan Cox [Sat, 5 Sep 2015 17:34:49 +0000 (17:34 +0000)]
Eliminate pointless requeueing of pages from terminated objects. These
pages will have left the inactive queue before the page daemon performs
its next scan. Also, ignore references to pages from terminated objects.
This allows the clean pages to be freed a little sooner.
Move some comments to their proper place, i.e., next to the code that
they describe, and update other nearby comments.
Andrew Turner [Sat, 5 Sep 2015 17:29:07 +0000 (17:29 +0000)]
Add ddb show commands to print the special registers and to ask the
hardware to perform address translation for us. These are useful to help
track down what caused us to enter the debugger.
Do not pass lle to nd6_ns_output(). Use newly-added
nd6_llinfo_get_holdsrc() to extract desired IPv6 source
from holdchain and pass it to the nd6_ns_output().
o Unlike xor, in Jenkins hash every bit of input affects virtually
every bit of output, thus salting the hash actually works. With
xor salting only provides a false sense of security, since if
hash(x) collides with hash(y), then of course, hash(x) ^ salt
would also collide with hash(y) ^ salt. [1]
o Jenkins provides much better distribution than xor, very close to
ideal.
TCP connection setup/teardown benchmark has shown a 10% increase
with default hash size, and with bigger hashes that still provide
possibility for collisions. With enormous hash size, when dataset is
by an order of magnitude smaller than hash size, the benchmark has
shown 4% decrease in performance decrease, which is expected and
acceptable.
Noticed by: Jeffrey Knockel <jeffk cs.unm.edu> [1]
Benchmarks by: jch
Reviewed by: jch, pkelsey, delphij
Security: strengthens protection against hash collision DoS
Sponsored by: Nginx, Inc.
Constantify lookup key in ifa_ifwith* functions.
Some places in our network stack already have const
arguments (like if_output() routines and LLE functions).
Code using ifa_ifwith (and similar functins) along with
LLE/_output functions is currently bound to use tricks
like __DECONST(). Provide a cleaner way by making sockaddr
lookup key really constant.
My MacBook has UGA only, but we fail to detect any changes
in the frame buffer when we flip pixels. Allow the detection
to be bypassed by setting the uga_framebuffer and uga_stride
variables. The kernel console works fine even when we can't
detect pixel changes in the frame buffer, which indicates
that the problem could be with reading from the frame buffer
and not writing to it.