This time we use the pci_info directly, it avoids a possible race with
PCI config regs, and we already have the value there anyway.
Signed-off-by: Augustin Cavalier <waddlesplash@gmail.com>
One crucial fix by me: check if size == 0. This avoids the deadlock
seen in #14795.
Includes some logic fixes to Haiku-specific code in the MII drivers,
and one new chipset supported by the rtl81xx MII driver. Otherwise,
no major functional change.
mmu_man changed this to be dynamically computed based on the BAR size.
It seems his computation was somehow incorrect, as it tries to allocate
a very large region here which deadlocks in the physical memory mapper.
Restoring this to 128KB un-breaks the new ipro1000 driver on more
recent hardware.
"Fixes" #14795.
pause() has granularity of "hz", which on FreeBSD and in our compat layer
is defined to be 1000 (so, 1ms.) As "safe_pause_us" is used copoiously
throughout startup code, this meant that (1) startup takes 100-1000x longer
than it needed to (this driver seems to most commonly delay for 10us),
and (2) this could in theory block the boot for multiple minutes if one
got particularly unlucky with the scheduler.
Probably helps with or even outright fixes#14795.
This should fix the Escape key being non-functional in the EFI loader
menus, which is pretty annoying when one is drilling through lots of
menus to blacklist files...
Previously we just probed and then immediately attached if any devices
were found, after all initialization was done. Now we cannot do that,
as after calling SYSINIT() on certain drivers (e.g. the new ipro1000
driver), certain threads will be created that there is no good way
of tearing down.
Fortunately, it is valid to call probe() on a device before SYSINIT()
(and most other things) have been initialized, and so now we do that
in init_hardware, and then init_driver is called only if we've found
devices we support.
While we're at it, also call pci->reserve_device, which will mark us as
the driver handling said device and prevent other drivers from
using it.
This is the iflib from FreeBSD 12.0 with a few select patches from
FreeBSD trunk applied (and a lot of #ifdef __HAIKU__s, as you can see.)
There are certain things it adds to SYSINIT that can't be un-initialized
(like the grouptaskqueues), and so rather than adding this to every
driver unnecessarily, a separate .a for it makes the most sense.
These are the portions of "freebsd_iflib.a" that I've written and
do not come from FreeBSD. As the import of iflib itself will be
quite large, having a separate commit for these made sense.
Previously, since we passed B_RELATIVE_TIMEOUT unconditionally,
a timeout of 0 would mean that it would return as soon as there was
CPU time available (so, usually, instantly.) This usually was not
a "problem" in that it caused broken behavior, but it would result
in exceptionally high CPU usage.
At first I implemented this correctly (i.e. a timeout of 0 will block
until explicitly woken up) but then discovered that our implementation
of these functions creates subtle race conditions compared to their
FreeBSD counterparts, and so to avoid deadlocks, a timeout of 1 second
is imposed. For situations where there are deadlocks due to race
conditions, this will make them painfully obvious (e.g. all network
transfers stalling for a second every 2-3 seconds or so.)
Previously I implemented "curcpu" as ipro1000 used it, independent
of all the other SMP functions, as an optimization. Now, iflib
wants to use these functions all together, so they have to produce
consistent results.
"const char* name" is alright for the compat layer's needs to
identify functions, as we only scan the method table once and
copy the function pointers to a struct of function pointers.
iflib doesn't use a function pointer struct, though, and expects
to have the KOBJ system for function lookups, and that depends
on integer-based method IDs for fast lookups. So now every
method gets an integer ID as well as a name (and iflib's method
IDs are in the list of method IDs already.)
Adapt other code accordingly. Minimum changes to fix ticket #11478
In theory also font height could be fractional, so we might need to
review other parts of the code.
Change-Id: Idbdbe38193ca5a32b3d09c7cf9accfc917760cae
Reviewed-on: https://review.haiku-os.org/821
Reviewed-by: waddlesplash <waddlesplash@gmail.com>
Reviewed-by: Stephan Aßmus <superstippi@gmx.de>
These methods and fields are missing from the R4 and R5 Be Book, but are
mentionned in the R4 release notes and present in the R5 headers.
Change-Id: I4de8298449bd66e0ee7fe0b52690552916314123
Reviewed-on: https://review.haiku-os.org/820
Reviewed-by: waddlesplash <waddlesplash@gmail.com>
When recording into a BPicture (ServerPicture, actually), one cannot
simply record the commands only, when the drawing itself would modify
state. This affects all drawing commands that change the pen location.
Therefore it is necessary to have a way to "dry-run" drawing a string
in order to know the pen location that would result. This is what
these two new methods help achieve.
Change-Id: Ic399a5513f18c12c16c0ab10a55e768c1b30e4e0
Reviewed-on: https://review.haiku-os.org/816
Reviewed-by: Rene Gollent <rene@gollent.com>
When CopyEngine::_CopyFolder() would recursively call itself, the
fCurrentTargetFolder would be set to the entry name of that directory.
When this returned, this was never reset to the parent directory name
and would eventually use a freed string when the BEntry was reused
for the next entry.
When changing the font state, a uin16 mask at the beginning of the
commands encodes which font parameters are transmitted in the link
data. Return this mask, so that one can know what parameters have
changed in the DrawState's ServerFont.
Change-Id: I52a9a665aac8eb0e6505193eba32c4b137846c78
Reviewed-on: https://review.haiku-os.org/814
Reviewed-by: waddlesplash <waddlesplash@gmail.com>
The constructor calls for GlobalDescriptorTable and
InterruptDescriptorTable were run on each CPU, each time clearing out
any changes already done on other CPUs.
For the IDT this didn't matter as the result would always be the same.
In case of the GDT the result would surprisingly be mostly hidden. The
used GDT entries are for the double fault stacks (which is rarely hit),
the user TLS (which will only be used in compat mode) and the TSS which
is immediately cached in the task register.
The clearing of the GDT allowed for a tiny window of time between
setting up the entry for the TSS and it actually being loaded (and
cached) in the task register. When this happened, the load of the TR
caused a not present fault, which at this point leads to a tripple
fault and reset due to no fault handlers being installed yet.
Fixes#14659.
Change-Id: I6e5c00d412ab17c3ef05740ba71228e6ca266c1e
Reviewed-on: https://review.haiku-os.org/810
Reviewed-by: Jérôme Duval <jerome.duval@gmail.com>
This was apparently copied from GlobalDescriptorTable::SetTSS() which
needs two table entries and therefore checks for index + 1. For the
SetUserTLS case this isn't needed and would cause aborts when reaching
the maximum CPU count (64 currently).
Change-Id: I27bd777fedadbd3740ac8c791199ec9300b06327
Reviewed-on: https://review.haiku-os.org/809
Reviewed-by: waddlesplash <waddlesplash@gmail.com>
The normal locking uses spinlocks that require getting at the current
CPU, which in turn needs a current thread set. This has not been set up
at this point and would simply cause tripple faults.
Use manual locking using atomic ops instead.
Change-Id: Ica894389330ef481eec84b667234139746ac4a46
Reviewed-on: https://review.haiku-os.org/808
Reviewed-by: waddlesplash <waddlesplash@gmail.com>
When resending a mail, don't put the To-field in the second column
after the From-field as it's done for incoming mail. To- and From-field
are the same layout as for creating a new mail.
Use the member fResending instead of the parameter "resending" the
THeaderView was called with.
Initialize fCcControl in ctor.
Many of the consumers fill in data into preallocated arrays. Some of
them already ignored values past the array size manually, some didn't.
Add a maxChar argument and set it from the incoming array sizes for
the various consumer cases.
Variable length arrays on the stack are always risky when the length
is indeterminate as they can easily overflow the stack. Replace their
use by BStackOrHeapArray, fixes#6354.
Also replace most other dynamic allocations by BStackOrHeapArray as
it is more convenient and may avoid unnecessary dynamic allocations.
Add allocation checks and early returns to all places while at it.
* I want to introduce a new way for plugins to
register for a format. Supporting the old FormatManager is
too painful at this point and not worth the effort.
* We don't aim to replicate this functionality. I don't
think this will be useful at all in future iterations.
Originally I planned to rewrite it on top of the new
BMediaFormat, but now I am of the hopinion this is
greatly unneeded.