jhb [Wed, 9 Dec 2009 21:09:32 +0000 (21:09 +0000)]
Fix a confusing typo in the EDD packet structure used in gptboot and
gptzfsboot. I got the segment and offset fields reversed in the structure,
but I also succeeded in crossing the assignments so the actual EDD packet
ended up correct.
jhb [Wed, 9 Dec 2009 20:36:56 +0000 (20:36 +0000)]
- Port bios_getmem() from libi386 to {gpt,}zfsboot() and use it to
safely allocate a heap region above 1MB. This enables {gpt,}zfsboot()
to allocate much larger buffers than before.
- Use a larger buffer (1MB instead of 128K) for temporary ZFS buffers. This
allows more reliable reading of compressed files in a raidz/raidz2 pool.
Submitted by: Matt Reimer mattjreimer of gmail
MFC after: 1 week
mav [Tue, 8 Dec 2009 22:43:29 +0000 (22:43 +0000)]
Change gmirror default balance algorithm from "split" to "load".
"split" is very ineffective for devices with rotating media as HDDs.
To be effective, it needs that transfer time reduction due to block
splitting was bigger then access time increase due to non-sequential
access. For modern HDDs I was able to reproduce it only with read sizes
of 2MB and above, which is almost not applicable in real life.
"load" algorithm same time is more universal and effective now.
trasz [Tue, 8 Dec 2009 20:47:10 +0000 (20:47 +0000)]
Don't add VAPPEND if the file is not being opened for writing. Note that this
only affects cases where open(2) is being used improperly - i.e. when the user
specifies O_APPEND without O_WRONLY or O_RDWR.
yongari [Tue, 8 Dec 2009 17:54:23 +0000 (17:54 +0000)]
Create sysctl node(dev.bge.%d.focred_collapse) instead of
hw.bge.forced_collapse. hw.bge.forced_collapse affects all bge(4)
controllers on system which may not desirable behavior of the
sysctl node. Also allow the sysctl node could be modified at any
time.
jkim [Tue, 8 Dec 2009 05:35:51 +0000 (05:35 +0000)]
- Try pre-allocating all FIBs upfront. Previously we tried pre-allocating
128 FIBs first and allocated more later if necessary. Remove now unused
definitions from the header file[1].
- Force sequential bus scanning. It seems parallel scanning is in fact
slower and causes more harm than good[1]. Adjust a comment to reflect that.
PR: kern/141269
Submitted by: Alexander Sack (asack at niksun dot com)[1]
Reviewed by: scottl
jfv [Tue, 8 Dec 2009 01:07:44 +0000 (01:07 +0000)]
Resync with Intel versions of both the em and igb
drivers. These add new hardware support, most importantly
the pch (i5 chipset) in the em driver. Also, both drivers
now have the simplified (and I hope improved) watchdog
code. The igb driver uses the new RX cleanup that I
first implemented in ixgbe.
jfv [Mon, 7 Dec 2009 21:30:54 +0000 (21:30 +0000)]
Update driver to Intel version 2.0.7:
This adds new feature support for the 82599, a hardware
assist to LRO, doing this required a large revamp to the
RX cleanup code because the descriptor ring may not be
processed out of order, this necessitated the elimination
of global pointers.
Additionally, the RX routine now does not refresh mbufs
on every descriptor, rather it will do a range, and then
update the hardware pointer at that time. These are
performance oriented changes.
The TX side now has a cleaner simpler watchdog algorithm
as well, in TX cleanup a read of ticks is stored, that
can then be compared in local_timer to determine if
there is a hang.
Various other cleanups along the way, thanks to all who
have provided input and testing.
yongari [Mon, 7 Dec 2009 19:26:54 +0000 (19:26 +0000)]
Don't access jumbo frame related registers if controller lacks the
feature. These registers are reserved on controllers that have no
support for jumbo frame.
Only BCM5700 has mini ring so do not poke mini ring related
registers if controller is not BCM5700.
yongari [Mon, 7 Dec 2009 19:18:23 +0000 (19:18 +0000)]
Remove PHY isolate/power down code in bge_stop(). The isolation
handler in brgphy(4) does not exist and brgphy(4) just resets the
PHY and returns EINVAL as it has no isolation handler. I also agree
on Marius's opinion that stop handler of every NIC driver seems to
be the wrong place for implementing PHY isolate/power down.
If we need PHY isolate/power down it should be implemented in
brgphy(4) and users should administratively down the PHY.
mav [Mon, 7 Dec 2009 18:37:50 +0000 (18:37 +0000)]
Explicitly acknowledge MSI completion, as required by SiI3124 datasheet.
It makes MSI working there. Later (and cheaper) PCIe chips (3132/3531)
still randomly crashing system in few seconds of high MSI rates, generating
something inaporopriate, like NMI or "Fatal trap 30".
jhb [Mon, 7 Dec 2009 16:29:43 +0000 (16:29 +0000)]
Improve the algorithm the loader uses to choose a memory range for its
heap when using a range above 1MB.
Previously the loader would always use the last 3MB in the first memory
range above 1MB for the heap. However, this memory range is also where the
kernel and any modules are loaded. If this memory range is "small", then
using the high 3MB for the heap may not leave enough room for the kernel
and modules.
Now the loader will use any range below 4GB for the heap, and the logic to
choose the "high" heap region has moved into biosmem.c. It sets two
variables that the loader can use for a high heap if it desires. When a
high heap is enabled (BZIP2, FireWire, GPT, or ZFS), then the following
memory ranges are preferred for the heap in order from best to worst:
- The largest memory region in the SMAP with a start address greater than
1MB. The memory region must be at least 3MB in length. This leaves the
region starting at 1MB purely for use by the kernel and modules.
- The last 3MB of the memory region starting at 1MB if it is at least 3MB
in size. This matches the current behavior except that the current loader
would break horribly if the first region was not at least 3MB in size.
- The memory range from the end of the loader up to the 640k window. This
is the range the loader uses when none of the high-heap-requesting options
are enabled.
mav [Mon, 7 Dec 2009 16:23:25 +0000 (16:23 +0000)]
MFp4;
- Cleanup kernel messages, mostly PMP.
- Took references on devices, while PMP reinitializes them, to not let them
go and distort freeze reference counting.
guido [Mon, 7 Dec 2009 15:15:08 +0000 (15:15 +0000)]
Fix ntfs such that it understand media with a non-512-bytes sector size:
1. Fixups are always done on 512 byte chunks (in stead of sectors). This
is kind of stupid.
2. Conevrt between NTFS blocknumbers (the blocksize equals the media
sector size) and the bread() and getblk() blocknr (which are 512-byte
sized)
NB: this change should not affect ntfs for 512-byte sector sizes.
marcel [Mon, 7 Dec 2009 06:41:27 +0000 (06:41 +0000)]
Define struct pcpu_md as the only MD field of struct pcpu (pc_acpi_id
excluded, as it's used by MI code) and mode the sysctl variables from
pcpu_stats to pcpu_md.
Adjust all references accordingly.
While nearby, change the PCPU sysctl tree so that they match the CPU
device sysctl tree -- they are now children of a static node called
"machdep.cpu" and are named only with their cpu ID.
Because several applications in /bin use libulog (or may use it in the
nearby future), it must not live inside /usr. It seems like we don't
need to add the copy from /usr/lib to ObsoleteFiles.inc, because it's
cleaned up during installation of libulog automatically.
marcel [Mon, 7 Dec 2009 00:54:02 +0000 (00:54 +0000)]
Allocate the VHPT for each CPU in cpu_mp_start(), rather than
allocating MAXCPU VHPTs up-front. This allows us to max-out MAXCPU
without memory waste -- MAXCPU is now 32 for SMP kernels.
This change also eliminates the VHPT scaling based in the total
memory in the system. It's the workload that determines the best size
of the VHPT. The workload can be affected by the amount of memory,
but not necessarily. For example, there's no performance difference
between VHPT sizes of 256KB, 512KB and 1MB when building the LINT
kernel. This was observed with a system that has 8GB of memory.
By default the kernel will allocate a 1MB VHPT. The user can tune the
system with the "machdep.vhpt.log2size" tunable.
mav [Sun, 6 Dec 2009 23:56:54 +0000 (23:56 +0000)]
Add Asynchronous Notification support for controllers without SNTF
capability by snooping SDB FIS receive area. It should be even faster
then regular way, but less reliable.
scf [Sun, 6 Dec 2009 23:51:27 +0000 (23:51 +0000)]
Update the getenv(3) man page to reflect the recent change to the behavior
of setenv(), putenv() and unsetenv() when dealing with corrupt entries in
environ. They now output a warning and complete their task without error.
scf [Sun, 6 Dec 2009 23:27:24 +0000 (23:27 +0000)]
Change the behavior of setenv(), putenv() and unsetenv() to continue parsing
instead of returning an error if a corrupt (not a "name=value" string) entry
in the environ array is detected when (re)-building the internal
environment. This should prevent applications or libraries from
experiencing issues arising from the expectation that these calls will
complete even with corrupt entries. The behavior is now as it was prior to
7.0.
scf [Sun, 6 Dec 2009 23:05:17 +0000 (23:05 +0000)]
Improve the comment within getenv() explaining the search order it takes to
find a variable. Include a note that it must not cause the internal
environment to be generated since malloc() depends upon getenv(). To call
malloc() would create a circular dependency.
Recommended by: green
Approved by: jilles
MFC after: 1 week
Add a libutempter compatibility interface to libulog.
The ulog_login_pseudo(3) and ulog_logout_pseudo(3) interfaces provide a
functionality identical to what libutempter has to offer. Just transform
libutempter's calls into the before mentioned functions.
libutempter doesn't work with utmpx, so instead of fixing I thought the
easiest way would be to integrate this functionality. libutempter is
used by applications like xterm and the KDE libraries, so if I ever
change the underlying file format, these applications will keep working
automatically.
Also increase __FreeBSD_version to indicate the addition (as well as the
import of libulog).
luigi [Sun, 6 Dec 2009 18:04:26 +0000 (18:04 +0000)]
restore setting of sin_len (was removed in 1.146 last february) as
it seems that now it is necessary for 'forward' to work outside lo0.
The bug (and fix) was reported on 8.0. This patch probably applies
to RELENG_7 as well.
It seems that 'pf' has a similar bug.
cleanilinks wasn't listed in <bsd.subdir.mk>. Instead of adding it to
/sys/modules/Makefile, we'd better just add it to <bsd.subdir.mk>
directly, so we don't need to change files like /sys/modules/sound/Makefile
as well. This means you can finally clean up all those dangling symlinks
created by individual module compilation at once.
MAXLOGNAME seems more applicable in this case, because UT_NAMESIZE
refers to the username field in utmp files, which is clearly unrelated
to repquota(8).
The size of the username record in utmp files should not influence the
maximum username length. Right now ut_user/ut_name is big enough, so in
this case it's dead code anyway.
We don't have UT_*SIZE anymore. One of the reasons for that is because
all strings are null terminated, there is no need for apps to copy
strings out of the utmpx structure. This means we can define W_DISP*SIZE
lengths for all columns.
While there, adjust the sizes a little. Steal some bytes from the
username column, while extending the hostname column quite a bit.
mav [Sun, 6 Dec 2009 00:10:13 +0000 (00:10 +0000)]
MFp4:
Introduce ATA_CAM kernel option, turning ata(4) controller drivers into
cam(4) interface modules. When enabled, this options deprecates all ata(4)
peripheral drivers (ad, acd, ...) and interfaces and allows cam(4) drivers
(ada, cd, ...) and interfaces to be natively used instead.
As side effect of this, ata(4) mode setting code was completely rewritten
to make controller API more strict and permit above change. While doing
this, SATA revision was separated from PATA mode. It allows DMA-incapable
SATA devices to operate and makes hw.ata.atapi_dma tunable work again.
Also allow ata(4) controller drivers (except some specific or broken ones)
to handle larger data transfers. Previous constraint of 64K was artificial
and is not really required by PCI ATA BM specification or hardware.
oleg [Sat, 5 Dec 2009 23:27:21 +0000 (23:27 +0000)]
Fix burst processing for WF2Q pipes - do not increase available burst size
unless pipe is idle. This should fix follwing issues:
- 'dummynet: OUCH! pipe should have been idle!' log messages.
- exceeding configured pipe bandwidth.
(Un)fortunately there is no standardized interface to switch between
utmp database files, so we must call ulog_setutxfile() here.
I'm also changing the column widths to magic numbers here. Display
layout should in this case not be derived from structure fields sizes.
Because I don't want struct utmpx ever to become too small, the fields
are too big to reserve all the space.
kib [Sat, 5 Dec 2009 20:36:42 +0000 (20:36 +0000)]
Change VOP_FSYNC for zfs vnode from VOP_PANIC to zfs_freebsd_fsync(),
both to not panic when fsync(2) is called for fifo on zfs
filedescriptor, and to actually fsync fifo inode to permanent storage.
Let init(8) and reboot(8) use utmpx to log wtmp entries.
logwtmp() gets called with the raw strings that are written to disk. For
regular user entries, this isn't too bad, but when booting/shutting
down, the contents get rather cryptic.
Because our implementation guarantees the strings inside struct utmpx to
be null terminated, we don't need to copy everything out, which makes
the code nicer to read.
Also set WARNS to 6 and add $FreeBSD$ to keep SVN happy.
pjd [Sat, 5 Dec 2009 20:16:28 +0000 (20:16 +0000)]
We have to eventually look for provider without checking guid as this is need
for attaching when there is no metadata yet.
Before r200125 the order of looking for providers was wrong. It was:
1. Find provider by name.
2. Find provider by guid.
3. Find provider by name and guid.
Where it should have been:
1. Find provider by name and guid.
2. Find provider by guid.
3. Find provider by name.
Because our implementation guarantees the strings inside struct utmpx to
be null terminated, we don't need to copy everything out, which makes
the code nicer to read.
Also set WARNS to 6 and add $FreeBSD$ to keep SVN silent.
- Just like struct utmp, store strings inside struct utmpx itself. This
is needed to make things like pututxline() work.
- Add ut_id and ut_pid fields, even though they have little use in our
implementation.
- It turns out our "reboot" wtmp entries indicate a system boot, so
remove REBOOT_TIME
- Implement getutxline() and pututxline
- Add getutxuser() and setutxfile(), which allows us to crawl wtmp and
lastlog files as well.
- Add _ULOG_POSIX_NAMES, so we can already use the POSIX names if we
really want to.
The maximum length of a username has nothing to do with the size of the
username in the utmp files. Use MAXLOGNAME, which is defined as 17
(UT_USERSIZE + 1).
The entries in the argv array are not const themselves, but sometimes we
want to fill in const values. Just make the array const and use
__DECONST() to make it const for the execve()-call itself.
pjd [Sat, 5 Dec 2009 14:33:11 +0000 (14:33 +0000)]
Fix deadlock when ZVOLs are present and we are replacing dead component or
calling scrub when pool is in a degraded state. It will try to taste ZVOLs,
which will lead to deadlock, as ZVOL will try to acquire the same locks as
replace/scrub is holding already.
We can't simply skip provider based on their GEOM class, because ZVOL can have
providers build on top of it and we need to skip those as well.
We do it by asking for ZFS::iszvol attribute. Any ZVOL-based provider will give
us positive answer and we have to skip those providers.
This way we remove possibility to create ZFS pools on top of ZVOLs, but it is
not very useful anyway.
I believe deadlock is still possible in some very complex situations like when
we have MD provider on top of UFS file on top of ZVOL. When we try to replace
dead component in the pool mentioned ZVOL is based on, there might be a
deadlock when ZFS will try to taste MD provider. There is no easy way to detect
that, but it isn't very common.
mav [Sat, 5 Dec 2009 13:40:51 +0000 (13:40 +0000)]
Do not ignore device interrupt if bus mastering is still active. It is
normal in case of media read error and some ATAPI cases, when transfer size
is unknown beforehand. PCI ATA BM specification tells that in case of such
underrun driver should just manually stop DMA engine. DMA engine should
same time guarantie that all bus mastering transfers completed at the moment
of driver reads interrupt flag asserted.
This change should fix interrupt storms and command timeouts in many cases.
luigi [Sat, 5 Dec 2009 09:13:06 +0000 (09:13 +0000)]
remove a dead block of code, document how the ipfw clients are
hooked and the difference in handling the 'enable' variable
for layer2 and layer3. The latter needs fixing once i figure out
how it worked pre-vnet.