the correct behaviour (sorting by distance from the current head position
in the scan direction) and bioq_insert_head() and bioq_insert_tail()
have a well defined (and useful) behaviour, especially when intermixed
with calls to bioq_disksort().
In particular:
- fix a bug in the existing bioq_disksort() that did not use the
current head position correctly;
- redefine semantics of bioq_insert_head() and bioq_insert_tail().
bioq_insert_tail() can now be used as a barrier
between previous and subsequent calls to bioq_disksort().
The code is heavily documented in the source code so please refer
to that for the details.
Much of this code comes from Fabio Checconi. Also thanks to Kirk
for feedback on the (re)definition of bioq_insert_tail().
NOTE: in the current tree there is only a handful of files which
intermix calls to bioq_disksort() with bioq_insert_head() and
bioq_insert_tail(). The ordering of the queue in these situation
was not specified (nor easy to figure out) before, so I doubt any
of that code could be affected by the specification of the API.
Also note that the current implementation is significantly simpler
than the previous one (also used in ata_sort_queue()).
It would be useful to reimplement ata_sort_queue() using
the same code used in bioq_disksort().
MFC after: 1 week
Just like the old TTY layer, the current MPSAFE TTY layer does not make
any attempt to serialize calls of write(). Data is copied into the
kernel in 256 (TTY_STACKBUF) byte chunks. If a write() call occurs at
the same time, the data may interleave. This is especially likely when
the TTY starts blocking, because the output queue reaches the high
watermark.
I've implemented this by adding a new flag, TTY_BUSY_OUT, which is used
to mark a TTY as having a thread stuck in write(). Because I don't want
non-blocking processes to be possibly blocked by a sleeping thread, I'm
still allowing it to bypass the protection. According to this message,
the Linux kernel returns EAGAIN in such cases, but I think that's a
little too restrictive:
http://kerneltrap.org/index.php?q=mailarchive/linux-kernel/2007/5/2/85418/thread
PR: kern/118287
from the parent to the child process if they have an operation vector
of &badfileops. This narrows a set of races involving system calls that
allocate a new file descriptor, potentially block for some extended
period, and then return the file descriptor, when invoked by a threaded
program that concurrently invokes fork(2). Similar approches are used
in both Solaris and Linux, and the wideness of this race was introduced
in FreeBSD when we moved to a more optimistic implementation of
accept(2) in order to simplify locking.
A small race necessarily remains because the fork(2) might occur after
the finit() in accept(2) but before the system call has returned, but
that appears unavoidable using current APIs. However, this race is
vastly narrower.
The fix can be validated using the newfileops_on_fork regression test.
PR: kern/130348
Reported by: Ivan Shcheklein <shcheklein at gmail dot com>
Reviewed by: jhb, kib
MFC after: 1 week
Attach call without devclass set crashes the system.
On resume AHCI driver sometimes tries to create duplicate adX device.
It is surely his own problem, but IMHO it is not a reason to crash here.
Other reasons are also possible.
result in errors for a format loading but subsequent correct recognizing
for another format.
File format loading functions should avoid printing any additional
informations but just returning appropriate (and different between each
other) error condition, characterizing different informations.
Additively, the linker should handle appropriately different format
loading errors.
While a general mechanism is desired, fix a simple and common case on
amd64: file type is not recognized for link elf and confuses the linker.
Printout an error if all the registered linker classes can't recognize
and load the module.
Reviewed by: jhb
Sponsored by: Sandvine Incorporated
- Align the fifo output in fifo_print() with other vn_printf() output.
- Remove the leading space from lockmgr_printinfo() so its output lines up
in vn_printf().
- lockmgr_printinfo() now ends with a newline, so remove an extra newline
from vn_printf().
Back in 1.1 of kern_sysctl.c the sysctl() routine wired the "old" userland
buffer for most sysctls (everything except kern.vnode.*). I think to prevent
issues with wiring too much memory it used a 'memlock' to serialize all
sysctl(2) invocations, meaning that only one user buffer could be wired at
a time. In 5.0 the 'memlock' was converted to an sx lock and renamed to
'sysctl lock'. However, it still only served the purpose of serializing
sysctls to avoid wiring too much memory and didn't actually protect the
sysctl tree as its name suggested. These changes expand the lock to actually
protect the tree.
Later on in 5.0, sysctl was changed to not wire buffers for requests by
default (sysctl_handle_opaque() will still wire buffers larger than a single
page, however). As a result, user buffers are no longer wired as often.
However, many sysctl handlers still wire user buffers, so it is still
desirable to serialize userland sysctl requests. Kernel sysctl requests
are allowed to run in parallel, however.
- Expose sysctl_lock()/sysctl_unlock() routines to exclusively lock the
sysctl tree for a few places outside of kern_sysctl.c that manipulate
the sysctl tree directly including the kernel linker and vfs_register().
- sysctl_register() and sysctl_unregister() require the caller to lock
the sysctl lock using sysctl_lock() and sysctl_unlock(). The rest of
the public sysctl API manage the locking internally.
- Add a locked variant of sysctl_remove_oid() for internal use so that
external uses of the API do not need to be aware of locking requirements.
- The kernel linker no longer needs Giant when manipulating the sysctl
tree.
- Add a missing break to the loop in vfs_register() so that we stop looking
at the sysctl MIB once we have changed it.
MFC after: 1 month
sysctls during a linker file unload. We drop the lock when doing similar
operations during a linker file load. To close races, clear the LINKED
flag before dropping the lock so that the linker file is no longer visible
to userland.
MFC after: 1 week
When we leave the console TTY constantly open, we never reset the
termios attributes. This causes output processing, echoing, etc. not to
be reset to the proper values when going into single user mode after the
system has booted. It also causes nl-to-crnl-conversion not to take
place during shutdown, which causes a `staircase effect'.
This patch adds a new TTY flag, TF_OPENED_CONS, which is set when the
TTY is opened through /dev/console. Because the flags are only used by
the kernel and the pstat(8) utility, I've decided to renumber the TTY
flags. This shouldn't be an issue, because the TTY layer is not yet part
of a stable release.
Reported by: Mark Atkinson <atkin901 yahoo com>
Tested by: sepotvin
jail doesn't support. This involves a new function prison_check_af,
like prison_check_ip[46] but that checks only the family.
With this change, most of the errors generated by jailed sockets
shouldn't ever occur, at least until jails are changeable.
Approved by: bz (mentor)
return zero on success and an error code otherwise. The possible errors
are EADDRNOTAVAIL if an address being checked for doesn't match the
prison, and EAFNOSUPPORT if the prison doesn't have any addresses in
that address family. For most callers of these functions, use the
returned error code instead of e.g. a hard-coded EADDRNOTAVAIL or
EINVAL.
Always include a jailed() check in these functions, where a non-jailed
cred always returns success (and makes no changes). Remove the explicit
jailed() checks that preceded many of the function calls.
Approved by: bz (mentor)
called without calling vfs_busy() first. This made umount(8) hang waiting
for mnt_lockref to become zero, which would never happen.
Reviewed by: kib
Approved by: rwatson (mentor)
Reported by: pho
Found with: stress2
Sponsored by: FreeBSD Foundation
Right now we only have a very small amount of drivers that use clists,
but we still allocate 50 cblocks as slush space, which allows drivers to
temporarily overcommit their storage. Most of the drivers don't allow
this anyway.
I've performed the following changes:
- We don't allocate any cblocks on startup.
- I've removed the DDB command, because it has nothing useful to print
now. You can obtain the amount of allocated blocks by running `vmstat
-m | grep clist'.
- I've removed cfreecount, which is now unused.
- The old code first tries to allocate using M_NOWAIT, followed by
M_WAITOK. This doesn't make any sense, so just remove this logic. It
seems the drivers allow us to sleep anyway.
We can even remove ccmax from clist_alloc_cblocks and c_cbmax from
struct clist, but this breaks binary compatibility.
This reduces the amount of allocated cblocks on my system from 54 to 4.
The TTY buffers used the standard <sys/queue.h> lists. Unfortunately
they have a big shortcoming. If you want to have a double linked list,
but no tail pointer, it's still not possible to obtain the previous
element in the list. Inside the buffers we don't need them. This is why
I switched to custom linked list macros. The macros will also keep track
of the amount of items in the list. Because it doesn't use a sentinel,
we can just initialize the queues with zero.
In its simplest form (the output queue), we will only keep two
references to blocks in the queue, namely the head of the list and the
last block in use. All free blocks are stored behind the last block in
use.
I noticed there was a very subtle bug in the previous code: in a very
uncommon corner case, it would uma_zfree() a block in the queue before
calling memcpy() to extract the data from the block.
After running a `make buildkernel', I noticed most of the Giant locks in
sysctl are only caused by a very small amount of sysctl's:
- sysctl.name2oid. This one is locked by SYSCTL_LOCK, just like
sysctl.oidfmt.
- kern.ident, kern.osrelease, kern.version, etc. These are just constant
strings.
- kern.arandom, used by the stack protector. It is already protected by
arc4_mtx.
I also saw the following sysctl's show up. Not as often as the ones
above, but still quite often:
- security.jail.jailed. Also mark security.jail.list as MPSAFE. They
don't need locking or already use allprison_lock.
- kern.devname, used by devname(3), ttyname(3), etc.
This seems to reduce Giant locking inside sysctl by ~75% in my primitive
test setup.
mutex to a reader/writer lock. Lookup operations first grab a read lock and
perform the lookup. If the operation results in a need to modify the cache,
then it tries to do an upgrade. If that fails, it drops the read lock,
obtains a write lock, and redoes the lookup.
an interpreter definition in its program header), set the auxiliary
ELF argument AT_BASE to 0 rather than to the address that we would
have mapped the interpreter at if there had been one.
The ELF ABI specifications appear to be ambiguous as to the desired
behavior in this situation, as they define AT_BASE as the base address
of the interpreter, but do not mention what to do if there is none.
On Solaris, AT_BASE will be set to the base address of the static
binary if there is no interpreter, and on Linux, AT_BASE is set to 0.
We go with the Linux semantics as they are of more immediate utility
and allow the early runtime environment to know that the kernel has
not mapped an interpreter, but because AT_PHDR points at the ELF
header for the running binary, it is still possible to retrieve all
required mapping information when the process starts should it be
required. Either approach would be preferable to our current behavior
of passing a pointer to an unmapped region of user memory as AT_BASE.
MFC after: 3 weeks
backend kegs so it may source compatible memory from multiple backends.
This is useful for cases such as NUMA or different layouts for the same
memory type.
- Provide a new api for adding new backend kegs to secondary zones.
- Provide a new flag for adjusting the layout of zones to stagger
allocations better across cache lines.
Sponsored by: Nokia
sizeof("MAXCPU") being used to calculate a string length rather than
something more reasonable such as sizeof("32"). This shouldn't have
caused any ill effect until we run on machines with 1000000 or more
cpus.
- Restructure selscan() and selrescan() to avoid producing extra selfps
when we have a fd in multiple sets. As described below multiple selfps
may still exist for other reasons.
- Make selrescan() tolerate multiple selfds for a given descriptor
set since sockets use two selinfos per fd. If an event on each selinfo
fires selrescan() will see the descriptor twice. This could result in
select() returning 2x the number of fds actually existing in fd sets.
Reported by: mgleason@ncftp.com
pointers to the callout handler just before and just after the callout
it invoked. I attempted to do this in a manner congruent to tracing in
Solaris's callout mechanism, but couldn't quite use the same names due
to convention and syntax differences.
Example DTrace script to generate a distribution graph of callout
execution times:
callout_execute:::callout_start
{
self->cstart = timestamp;
}
callout_execute:::callout_end
{
@length = quantize(timestamp - self->cstart);
}
Reviewed by: jb
MFC after: 3 days
inside the SYSCTL() macros and thus does not need to be done for
all of the nodes scattered across the source tree.
- Mark the name-cache related sysctl's (including debug.hashstat.*) MPSAFE.
- Mark vm.loadavg MPSAFE.
- Remove GIANT_REQUIRED from vmtotal() (everything in this routine already
has sufficient locking) and mark vm.vmtotal MPSAFE.
- Mark the vm.stats.(sys|vm).* sysctls MPSAFE.
