This implements proper intrdisable() support for all
interrupt controllers.
For enable, (*arch->intrassign)(Vctl*) fills in the
Vctl.enable and Vctl.disable pointers with the
appropriate routines and returns the assigned
vector number.
Once the Vctl struct has been linked to its vector
chain, Vctl.enable(Vctl*, shared) gets called with a
flag if the vector has been already enabled (shared).
This order is important here as enabling the interrupt
on the controller before we have linked the chain can
cause spurious interrupts, expecially on mp system
where the interrupt can target a different cpu than
the caller of intrenable().
The intrdisable() case is the other way around.
We first disable the interrupt on the controller
and after that unlink the Vctl from the chain.
On a multiprocessor, the xfree() of the Vctl struct
is delayed to avoid freeing it while it is still
in use by another cpu.
The xen port now also uses pc/irq.c which has been
made generic enougth to handle xen's irq scheme.
Also, archgeneric is now a separate file to avoid
pulling in dependencies from the 8259 interrupt
controller code.
It appears that our IDT overlaps with the data structures
passed from grub in multiboot load.
So defer setup of the interrupt table after the multiboot
parameters have been processed.
The driver used to register the interrupt handler just
after reset, tho the Ctlr struct, including the buffer
descriptor arrays where only allocated on attach.
This moves most of the reset/init out of pnp
function and into attach. This also means we can
error out and even retry on the next attach.
The logic of the reseter kproc has been changed:
now it is only started once the first initialization
completely succeeded. This avoids the strange qlock
passing.
Implement a shutdown function so the device gets
halted for /dev/reboot.
Assume 64 bit physical addresses for dma.
Check that pci bar0 is actually I/O.
The new MTRR code handles overlapping ranges
and supports AMD specific TOM2 MSR.
The format in /dev/archctl now only shows
the effective cache ranges only, without
exposing the low level registers.
loading the interrupt vector table early allows
us to handle traps during bootup before mmuinit()
which gives better diagnostics for debugging.
we also can handle general protection fault on
rdmsr() and wrmsr() which helps during
cpuidentify() and archinit() when probing for
cpu features.
before removing the double map at 0, load our
initial gdt pointer with its new KZERO based
virtual address.
this is prerequisite for handling traps early during
bootup before mmuinit() loads the final gdt.
When using /dev/reboot, the MSI vecor might have already
been setup causing interrupts to fire on the designated
cpu while we send the commands to the card.
reseting irbsts bits in hdacmd() only works
while interrupts are disabled during hdareset().
once interrupts are enabled we need to reset the
irbsts bits in the interrupt handler or else the
interrupt never clears and locks up the system.
The change 3306:c5cf77167bfe made the code reuse MTRR slots
of the default memory type.
But this did not take overlapping ranges into account!
If two or more variable-range MTRRs overlap, the following rules apply:
a. If the memory types are identical, then that memory type is used.
b. If at least one of the memory types is UC, then UC memory type is used.
c. If at least of of the memory types is WT. and the only other memory type
is WB, then th WT memory type is used.
d. If the combination of memory types is not listed above,
then the memory type used in undefined.
It so happend that on a Dell Latitude E7450 that the BIOS defines
the default type as UC. and the first slot defines a 16GB range
of type WB. Then the rest of the ranges mark the PCI space back
as UC, but overlapping the first WB range! This works because
of rule (B) above.
When trying to make the framebuffer write-combining, we would
falsely reuse one of the UC sub-ranges and making the UC memory
into WB as a side effect.
Thanks to Fulton for his patience and providing debug logs and
doing experiments for us to narrow the problem down.
With some newer UEFI firmware, not all pci bars get
programmed and we have to assign them ourselfs.
This was already done for memory bars. This change
adds the same for i/o port space, by providing a
ioreservewin() function which can be used to allocate
port space within the parent pci-pci bridge window.
Also, the pci code now allocates the pci config
space i/o ports 0xCF8/0xCFC so userspace needs to
use devpnp to access pci config space now. (see
latest realemu change).
Also, this moves the ioalloc()/iofree() code out
of devarch into port/iomap.c as it can be shared
with the ppc mtx kernel.
the driver was not using irb interrupts
and was just polling the irb write pointer
to wait for command completion.
this is not supported by qemu.
qemu requires the use of irb interrupt handshake
and it refuses to accept the next command until we
acknowledge the irb interrupt.
the access functions for pci config space in config mode #1
used to set bit 0 in the register offset if the access was
to a device on any bus different from 0.
it is completely unclear why this was done and i can't find
any documentation on this.
but for sure, this breaks all pci config spacess access to
pci devices behind a bridge on qemu. with -trace pci* it
was discovered that all config space register offsets on
devies behind pci brige where off by one.
on real hardware, setting bit 0 in the offset doesnt appear
to be an issue.
thanks mischief for reporting and providing a qemu demo
configuration to reproduce the problem.
On my 6235 card, if we calibrate the crystal
immediately after disabling wimax, the the
firmware gets unhappy. A short nap before
sending the command prevents the command from
timing out.
ori and echoline are reporting regression on some 6000 cards;
which sometimes time out on crystal calibration command;
which is expected by the driver. but the new code used
to force a device reset on any command timeout.
reverting to old behaviour until for now until we have
a chance investigating.
We used to assume a 1:1 pairing of processors to submit queues.
With recent machines, we now got more cpu cores than what some
nvme drives support so we need to distribute the queues across
these cpu's which requires locking on command submission.
There is a feature get/set command to probe the number of submit
and completion queues, but we decided to just handling
submission queue create command error gracefully as it is simpler
and has less chance of regression with existing setups.
Thanks to mischief for investigating and writing the code.
the whole idea of a ucallocb() is bad, as even access to the
metadata header would be in uncached memory. also, it tuns out
that it was never used by anyone.
the 9000 series uses a new receive descriptor format
wich appears to reqire 4k aligned buffers. the old
format "halfworks" and just makes the firmware not
respond to any commands after the enable paging command.
the smartfifo command appears to causes problems.
but apparently not issuing it at all seems to work
fine on both the 8265 and 9260. so removing the code
for now.
issuing the bindingquota command before associated
makes association impossible. but enabling afterwards
works fine. (tested in 8265 and 9260).
the prph access functions now mask the address with
0xfffff. it is unclear why linux and openbsd drivers
specify addresses beyond that in ther register constants.
the timeevent change is interesting. the timeevent
needs to be restarted when it has stoped to make sure
probing/association packets are sent during the evnet.
The new pci code is moved to port/pci.[hc] and shared by
all ports.
Each port has its own PCI controller implementation,
providing the pcicfgrw*() functions for low level pci
config space access. The locking for pcicfgrw*() is now
done by the caller (only port/pci.c).
Device drivers now need to include "../port/pci.h" in
addition to "io.h".
The new code now checks bridge windows and membars,
while enumerating the bus, giving the pc driver a chance
to re-assign them. This is needed because some UEFI
implementations fail to assign the bars for some devices,
so we need to do it outselfs. (See pcireservemem()).
While working on this, it was discovered that the pci
code assimed the smallest I/O bar size is 16 (pcibarsize()),
which is wrong. I/O bars can be as small as 4 bytes.
Bit 1 in an I/O bar is also reserved and should be masked off,
making the port mask: port = bar & ~3;
The Abind case in namec() needs to cunique() the chan
before attaching the umh mount head pointer onto it.
This is because we cannot give a reference to the mount
head to any of the mh->mount...->to channels, as they
will never go away until the mount head goes away.
This is a cyclic reference.
This could be reproduced with:
@{rfork n; mount -a '#s/boot' /mnt/root; bind /mnt/root /}
Also, fix memory leaks around cunique(), which can
error, leaking the mount head we got from domount().
Move the umh != nil check inside cunique().
we have to disable interrupts during mmuwalk() of user pages
as we can get preempted during mmu walk and the original
m->pml4 might become one of a different process.
This change makes it mandatory for programs to call segflush() on
code that is not in the text segment if they want to execute it.
As a side effect, this means that everything but the text segment
will be non-executable by default, even without the SG_NOEXEC
attribute. Segments with the SG_NOEXEC attribute never become
executable, even when segflush() is called on them.
widen and move the KMAP window to a new address so we can
handle the 8GB of physical memory of the new raspberry pi4.
the new memory map on pi4 uses the following 4 banks:
0x000000000 0x03e600000
0x040000000 0x0fc000000 <- soc.dramsize (only < 4GB)
0x100000000 0x180000000
0x180000000 0x200000000
On the 8GB variant of the raspberry pi 4,
the eeprom chip for the xhci controller is missing and
instead loaded from sdram (by the gpu firmware).
for this, the gpu firmware needs to be notified of
the xhci controllers pci bus address (after reset)
that was assigned by our pci enumeration code.
instruction cache maintenance is done on tlb miss;
when a page gets fauled in; with putmmu() checking
the page->txtflush cpu bitmap.
syssegflush() used to only call flushmmu() after
segflush() for the calling process, but when a segment
is shared with other processes, we have to flush the
other processes tlb as well.
this adds the missing procflushseg() call into segflush().
note that procflushseg() leaves the calling process alone,
so the flushmmu() call in syssegflush() is still required.
segmentioproc() does not need to call flushmmu() after
segflush() as it is never going to jump to the modified
page, hence the stale icache does not matter.
The sample frequency is an artificial parameter used for
isochronous out transfers to better match the target
frequency (usually, a sound card).
when hz is set, devusb adjusts the endpoint's maxpkt to get
the requested frequency and a multiple of the samplesize per
packet.
however, when hz is not set, then we should calculate the
frequency from maxpkt, ntds and pollival, so all parameters
will be consistent with each other.
Added a ver= field to the filter to distinguish the ip version.
By default, a filter is parsed as ipv6, and after parsing
proto, src and dst fields are converted to ipv4. When no
ver= field is specified, a ip version filter is implicitely
added and both protocols are parsed.
This change also gets rid of the fast compare types as the
filed might not be aligned correctly in the packet.
This also fixes the ifc= filter, as we have to check any
local address.
We used to just return the first address of the incoming
interface regardless of if the address matches the source
ip type and scope.
This change tries to find the best interface address that
will match the source ip so it can be used as a source
address when replying to the packet.
the page attribute table was initialized in mmuinit(), which is
too late for bootscreen(). So now we check for PAT support and
insert the write-combine entry early in cpuidentify().
this might have been the cause of some slow EFI framebuffers on
machines with overlapping or insufficient MTRR entries.
ipiput4() and ipiput6() are called with the incoming interface rlocked
while ipoput4() and ipoput6() also rlock() the outgoing interface once
a route has been found. it is common that the incoming and outgoing
interfaces are the same recusive rlocking().
the deadlock happens when a reader holds the rlock for the incoming interface,
then ip/ipconfig tries to add a new address, trying to wlock the interface.
as there are still active readers on the ifc, ip/ipconfig process gets queued
on the inteface RWlock.
now the reader finds the outgoing route which has the same interface as the
incoming packet and tries to rlock the ifc again. but now theres a writer
queued, so we also go to sleep waiting four outselfs to release the lock.
the solution is to never wait for the outgoing interface rlock, but instead
use non-queueing canrlock() and if it cannot be acquired, discard the packet.
do not touch s->map on SG_PHYSICAL type segments as they do
not have a pte map (s->mapsize == 0 && s->map == nil).
also remove the SG_PHYSICAL switch in freepte(), this is never
reached.
the calculation for the control endpoint0 output device context
missed the context size scaling shift, resulting in botched
stall handling as we would not read the correct endpoint status
value.
note, this calculation only affected control endpoint 0, which
was handled separately from all other endpoints.
spectacular bug. cmpswap() had a sign extension bug
using sign extending MOV to load the old compare
value and LDXRW using zero extension while the CMP
instruction compared 64 bit registers.
this caused cmpswap with negative old value always
to fail.
interestingly, libc's version of this function was
fine.
when reclaiming pages from an image, always reclaim all
the hash chains equally. that way, we avoid being biased
towards the chains at the start of the Image.pghash[] array.
images can be in two states: active or inactive. inactive
images are the ones which are not used by program while
active ones aare.
when reclaiming pages, we should try to reclaim pages
from inactive images first and only if that set becomes
exhausted attempt to release text pages and attempt to
reclaim pages from active images.
when we run out of Image structures, it makes only sense
to reclaim pages from inactive images, as reclaiming pages
from active ones will never free any Image structures.
change putimage() to require a image already locked and
make it unlock the image. this avoids many pointless
unlock()/lock() sequences as all callers of putimage()
already had the image locked.
looks like linux changed the device tree names for
the memory node:
4b17654f51 (diff-ac03c9402b807c11d42edc9e8d03dfc7)
this fixes the memory size detection with latest firmware
on raspberry pi4-b (4GB) for kenji.
