David du Colombier wrote:
> The slowness issue only appears on the loopback, because
> it provides a 16384 MTU.
>
> There is an old bug in the Plan 9 TCP stack, were the TCP
> MSS doesn't take account the MTU for incoming connections.
>
> I originally fixed this issue in January 2015 for the Plan 9
> port on Google Compute Engine. On GCE, there is an unusual
> 1460 MTU.
>
> The Plan 9 TCP stack defines a default 1460 MSS corresponding
> to a 1500 MTU. Then, the MSS is fixed according to the MTU
> for outgoing connections, but not incoming connections.
>
> On GCE, this issue leads to IP fragmentation, but GCE didn't
> handle IP fragmentation properly, so the connections
> were dropped.
>
> On the loopback medium, I suppose this is the opposite issue.
> Since the TCP stack didn't fix the MSS in the incoming
> connection, the programs sent multiple small 1500 bytes
> IP packets instead of large 16384 IP packets, but I don't
> know why it leads to such a slowdown.
i made a mistake here as this change breaks the arm and mips compilers
which lack an optimiation in xcom() that folds constant pointer arithmetic
into the offset. on arm, the a node is a complex expression with op OADD of
type TIND but the test rejected the (valid) pointer arithmetic.
instead, we now test for the operations which cannot be constant instead
of using the type as a proxy.
mischief spotted that the only way for listeners to go away was
truncating (but not removing) a service script. this is wrong and
not as described in the manpage.
this change makes removing (or truncating) a listen script stop
the listener.
scandir() first marks all current announces, then reads the service
directory adding announces which will clear the marks for the ones
already there or add a new unmarked one. finally, we shoot down and
remove all still marked announces.
6c changed "- cmd_lagest_size + 1" into a *unsigned* 32bit constant. which
got added to 64bit pointer making pcb->limit > pcb->end resulting
in errors for partial commands in the buffer. removing the parentesis
propagates the operation to 64bit.
the intend of posting a note to the faulting process is to
interrupt the syscall to give the note handler a chance
to handle it. kernel processes however, have no note handlers
and all the postnote() does is set up->notepending which will
make the next attempt to sleep raise an Eintr[] error. this
is harmless, but usually not what we want.
there's no need to waste space for a error buffer in the Segio
structure, as the segmentio kproc will be waiting for the next
command after an error and will not overwite it until we issue
another command.
devproc's procctlmemio() did not handle physical segment
types correctly, as it assumed it can just kmap() the page
in question and write to it. physical segments however
need to be mapped uncached but kmap() will always map
cached as it assumes normal memory. on some machines with
aliasing memory with different cache attributes
leads to undefined behaviour!
we borrow the code from devsegment and provide a generic
segio() function to read and write user segments which
handles all the cases without using kmap by just spawning
a kproc that attaches the segment that needs to be read
from or written to. fault() will setup the right mmu
attributes for us. it will also properly flush pages for
segments that maintain instruction cache when written.
however, tlb's have to be flushed separately.
segio() is used for devsegment and devproc now, which
also allows for simplification of fixfault() as there is no
special error handling case anymore as fixfault() is now
called from faulting process *only*.
reads from /proc/$pid/mem can now span multiple pages.
code like "return g->dlen;" is wrong as we do not hold the
qlock of the global segment. another process could come in
and override g->dlen making us return the wrong byte count.
avoid copying when we already got a kernel address (kernel memory
is the same on processes) which is the case with bread()/bwrite().
this is the same optimization that devsd does.
also avoid allocating/freeing and copying while holding the qlock.
when we copy to/from user memory, we might fault preventing
others from accessing the segment while fault handling is in
progress.
walking the freelist for every page is too slow. as we
are freeing a range, we can do a single pass unlinking all
pages in our range and at the end, check if all pages
where freed, if not put the pages that we did free back
and retry, otherwise we'r done.
fixed segments are continuous in physical memory but
allocated in user pages. unlike shared segments, they
are not allocated on demand but the pages are allocated
on creation time (devsegment). fixed segments are
never swapped out, segfreed or resized and can only be
destroyed as a whole.
the physical base address can be discovered by userspace
reading the ctl file in devsegment.
when we trim the front of a block with freefromfront(),
the block magic of the back was not initialized which
could sometimes trigger the assert in blocksetsize()
to fail. fix is to just move the initialization of the
magic field before the blocksetsize() call.
the second b->magic = UNALLOC_MAGIC isnt really required
but just done for consistency with the trim() code above.
when we get an i/o error, always call hdrecover() which
will reset the port and reinitialize the interface of
the calling processes endpoint.
handle the case when we have multi-function device with
multiple reader procs in hdrecover(). the sequence is
as follows:
1) any of the reader procs encounters i/o error and calls hdrecover(),
acquires qlock and initiates port reset.
2) any other readerprocs will now encounter i/o error (due to reset) and also call
hdrecover() but will be waiting on the qlock for reset to complete.
3) first process completes reset and reinitializes its interface with setproto()
and then releases the qlock for the other readers todo the same.
this avoids listing the upper half of 64-bit membars
in Pcidev.mem[] array avoiding potential confusion
in drivers.
we also check if the upper half is programmed to zero
by bios and otherwise zap the entry in Pcidev.mem[]
and print a warning.
qemu puts multiboot data after the end of the kernel image, so
to be able to KADDR() that memory early, we extend the initial
identity mapping by 16K. right now we just got lucky with
the pc kernel as it rounds the map to 4MB pages.
when we switch to graphics mode, we do not want graphical arcs console
to print on the screen anymore as it assumes 8bit color mode and just
messes up the screen on kernel prints.
fixes bug with libmemdraw where the linker would hoist
the final ADD $const, SP at the end over loads from the
stack causing the front to fall off once a interrupt hits.
GEVector() saves the exception return PC in Ureg.r27 which needs
to be preserved.
there should be no reason for the user to change the status
register from noted() eigther, so we now just use setregisters()
in noted() to restore previous general purpose registers. this
means that CU1 will always be off after noted() because notify()
has disabled the FPU on entry and set fpstatus to FPinactive
if it was on. once user starts using FPU again, it will trap and
restore fpu registers.
touching transmit descriptors while dma is running causes the
front to fall off. new approach keeps a counter of free
descriptors in the Ring structure that is incremented
by txintr() when transmit completed.
txintr() will clean descriptors once dma has stopped and
restart dma when there are more descrtors in the chain.
this provides basic console support using the ARC bios routines
theu uartarcs driver. and has native seeq ethernet driver which
was written by reading the 2ed devseq driver as i have no
documentation on the hardware. mmu and trap code is based on the
routerboard kernel.
bootmkfile will now looks for the following proto files in order
and pick the first one it finds to build the bootfs.paq file:
1) $CONF.boofs.proto (config specific)
2) bootfs.proto (kernel specific)
3) $BOOTDIR/bootfs.proto (default generic)
from the unicode-db patch readme:
command() receives a char* that is assigned to lp, which is a Rune*,
and lp is incremented later in readchar(), so each read consumed 4 bytes.
The only time command() is called is in runpcs() with bkpt->comm,
which is a char* built in subpcs through a char*, so the string stored in
bkpt->comm was not a Rune string. A way to test the bug is:
db program
main:b argv/X
:r
the mount cache uses Page.va to store cached range offset and
limit, but mips kernel uses cache index bits from Page.va to
maintain page coloring. Page.va was not initialized by auxpage().
this change removes auxpage() which was primarily used only
by the mount cache and use newpage() with cache file offset
page as va so we will get a page of the right color.
mount cache keeps the index bits intact by only using the top
and buttom PGSHIFT bits of Page.va for the range offset/limit.
when we are skipping a process because we could not acquire
its segment lock, dont call reclaim() again (which is pointless
as we didnt pageout any pages), instead try the next process.
the Pte.last pointer is inclusive, so don't miss the last page
in pageout().
when building bootfs in d770 mode directory, the other permissions
in bootfs paq are masked off which results in boot to fail. theres
no point in checking group/other permissions on boot, so just disable
permissin checking in paqfs with the -a flag.
the special sencodefmt() in ndb/dn.c is only used with %H format for
hexadecimal printing for binary strings. removing the unused
calls to enc32() and enc64() reduces the code size by arround 4K.
(this is usefull for ndb/getip which gets linked into the kernel).
the approximation of n*2 to calculate the number of output bytes
for enc64() fails for inputs of size < 3. this is fixed by using
encodefmt() which gets the calculation right and also simplifies
the code avoiding the allocation and freeing of intermediate string
buffers.
mcountseg(), mfreeseg():
use Pte.first/last pointers when possible and avoid constructs
like s->map[i]->pages[j].
freepte():
do not zero entries in freepte(), the segment is going away and
here is no point in zeroing page pointers. hoist common code at
the top avoiding duplication.
segpage(), fixfault():
avoid load after store for Pte** pointer.
fixfault():
return -1 in default case to avoid the "used but not set" warning
for mmuphys and get rid of the useless initialization.
syssegflush():
due to len being unsigned, the pe = PGROUND(pe) can make "chunk"
bigger than len causing a overflow. rewrite the function and deal
with page alignment and errors at the beginning.
syssegflush(), segpage(), fixfault(), putseg(), relocateseg(),
mcountseg(), mfreeseg():
keep naming consistent.
the "to" address can overflow in syssegfree() causing wrong
number of pages to be passed to mfreeseg(). with the current
implementation of mfreeseg() however, this doesnt cause any
data corruption but was just freeing an unexpected number of
pages.
this change checks for this condition in syssegfree() and
errors out instead. also mfreeseg() was changed to take
ulong argument for number of pages instead of int to keep
it consistent with other routines that work with page counts.
sdbio() tests if it can pass the buffer pointer directly to
the driver when it is already in kernel memory. we also need
to check if the buffer is properly aligned but alignment
requirement is handled in system specific sdmalloc() and
was not known to devsd.
to solve this, we *always* page align sd buffers and get rid
of the system specific sdmalloc() macro (was only used in bcm
kernel).
