needed to be statically populated with device nodes. The first two are no
longer true, which makes the third pretty moot. In fact, we don't seem to
put device node bits into the distribution archives at all anymore.
So..... remove the god-aweful nasty hack that force unmounted devfs during
installation so that static device nodes could land in /dev. Now that the
vnode cleaner handles this case better this isn't strictly needed, but
axeing code in sysinstall is almost always benficial. Thanks to Don Lewis
for pointing out this attribute of sysinstall.
reassigning their v_ops field to specfs, detaching from the mountpoint, etc.
However, this is not sufficient. If we vclean() the vnode the pages owned
by the vnode are lost, potentially while buffers reference them. Implement
parts of vclean() seperately in vgonechrl() so that the pages and bufs
associated with a device vnode are not destroyed while in use.
correctly, resulting in the dreaded "vm_pageout_flush: partially
invalid page" panic. The caching issue will be revisited in the
future, but opt for safety over performance in the meantime.
Tested by: gallatin
work before anyways, and I didn't want to fix broken code I had no
way of testing. It was necessary however, in order to get rid of GIANT_LOCK.
Pthread priorities will have to wait a little longer to get fixed.
this definition. It fixes gnome for starters. I haven't tried *emacs yet.
Like IA64, amd64 uses registers for the first few arguments and then
the stack for the rest. This means the 64 bit promotion of the NULL (0)
value is lost and its just pushed on as an 'int' in a varargs call.
When the consumer walks the list and expects to pull off void * pointers
via va_arg, then all hell breaks loose.
Marcel: thanks a million for finding this!
the psim simulator. Look for the "file" property which only exists
on psim disks, and as a bonus, print the contents of this at boot-time,
which is the host file being used for the disk image.
- remove remaining warnings.
problems: (1) The wrong flag was being checked for in the attribute
(2) The pthread's state was not being set to indicate it was
suspended.
Noticed by: Igor Sysoev <is@rambler-co.ru>
transfer size to the stripe size. This is a different
situation from reviving, where this limitation is necessary.
In initsd we're simply writing binary zeroes to the entire
disk, so the only effect of limiting the transfer is to slow
things down.
ndis_send_packets() but there's no link yet, we get an immediate
callback to ndis_txeof(), which clears if_timer. But ndis_start()
sets if_timer right after the call to ndis_send_packets(). Set
if_timer before calling ndis_send_packets().
Also fix mutex locking to prevent ndis_txeof() from running in
the middle of ndis_start().
make it more robust. This should fix problems with crashes under
heavy traffic loads that have been reported. Also add a 'query done'
callback handler to satisfy the e100bex.sys sample Intel driver.
systems are mounted. An example set of entries for /etc/rc.conf:
ataraid_enable="YES"
ataraid_devices="ar0"
ataraid_ar0_set="ad2 ad3"
ataraid_ar0_type="RAID1"
Because there is no "correct" way of doing ATA raid (ie, geom vs.
atacontrol vs. vinum) that is bikeshed proof, this rcng script stays within
the bounds of atacontrol and assumes that other RAID solutions for GEOM or
vinum will end up in a different rcNG script.
Reviewed by: green
eventually be passed an async. context as well as a syscall
context.
While here, fix a serious bug in that if the trapframe is a
syscall frame, but we're restoring an async context, we need
to clear the FRAME_SYSCALL flag so that we leave the kernel
via exception_restore.
NdisAnsiStringToUnicodeString(), NdisWriteConfiguration().
Also add stubs for NdisMGetDeviceProperty(), NdisTerminateWrapper(),
NdisOpenConfigurationKeyByName(), NdisOpenConfigurationKeyByIndex()
and NdisMGetDeviceProperty().
terminates. Without this patch, 'make -j1 buildworld' takes about 30%
longer than 'make -B buildworld' on my 2.4 GHz P4; the difference is
probably even larger on faster systems. With this patch, there is no
perceptible difference in wall time between the two.
Submitted by: bde
MFC after: 3 days
- fix ndis_time() so that it returns a time based on the proper
epoch (wacky though it may be)
- implement NdisInitializeString() and NdisFreeString(), and add
stub for NdisMRemoveMiniport()
ntoskrnl_var.h:
- add missing member to the general_lookaside struct (gl_listentry)
subr_ntoskrnl.c:
- Fix arguments to the interlocked push/pop routines: 'head' is an
slist_header *, not an slist_entry *
- Kludge up _fastcall support for the push/pop routines. The _fastcall
convention is similar to _stdcall, except the first two available
DWORD-sized arguments are passed in %ecx and %edx, respectively.
One kludge for this __attribute__ ((regparm(3))), however this
isn't entirely right, as it assumes %eax, %ecx and %edx will be
used (regparm(2) assumes %eax and %edx). Another kludge is to
declare the two fastcall-ed args as local register variables and
explicitly assign them to %ecx and %edx, but experimentation showed
that gcc would not guard %ecx and %edx against being clobbered.
Thus, I came up with a 3rd kludge, which is to use some inline
assembly of the form:
void *arg1;
void *arg2;
__asm__("movl %%ecx, %%ecx" : "=c" (arg1));
__asm__("movl %%edx, %%edx" : "=d" (arg2));
This lets gcc know that we're going to reference %ecx and %edx and
that it should make an effort not to let it get trampled. This wastes
an instruction (movl %reg, %reg is a no-op) but insures proper
behavior. It's possible there's a better way to do this though:
this is the first time I've used inline assembler in this fashion.
The above fixes to ntoskrnl_var.h an subr_ntoskrnl.c make lookaside
lists work for the two drivers I have that use them, one of which
is an NDIS 5.0 miniport and another which is 5.1.
subr_ndis.c: NdisGetCurrentSystemTime() which, according to the
Microsoft documentation returns "the number of 100 nanosecond
intervals since January 1, 1601." I have no idea what's so special
about that epoch or why they chose 100 nanosecond ticks. I don't
know the proper offset to convert nanotime() from the UNIX epoch
to January 1, 1601, so for now I'm just doing the unit convertion
to 100s of nanoseconds.
subr_ntoskrnl.c: memcpy(), memset(), ExInterlockedPopEntrySList(),
ExInterlockedPushEntrySList().
The latter two are different from InterlockedPopEntrySList()
and InterlockedPushEntrySList() in that they accept a spinlock to
hold while executing, whereas the non-Ex routines use a lock
internal to ntoskrnl. I also modified ExInitializePagedLookasideList()
and ExInitializeNPagedLookasideList() to initialize mutex locks
within the lookaside structures. It seems that in NDIS 5.0,
the lookaside allocate/free routines ExInterlockedPopEntrySList()
and ExInterlockedPushEntrySList(), which require the use of the
per-lookaside spinlock, whereas in NDIS 5.1, the per-lookaside
spinlock is deprecated. We need to support both cases.
Note that I appear to be doing something wrong with
ExInterlockedPopEntrySList() and ExInterlockedPushEntrySList():
they don't appear to obtain proper pointers to their arguments,
so I'm probably doing something wrong in terms of their calling
convention (they're declared to be FASTCALL in Widnows, and I'm
not sure what that means for gcc). It happens that in my stub
lookaside implementation, they don't need to do any work anyway,
so for now I've hacked them to always return NULL, which avoids
corrupting the stack. I need to do this right though.
