Commit ec6cb17f (pci: enable RedHat PCI bridges to reserve additional
resources on PCI init)
added a new vendor specific PCI capability for RedHat PCI bridges
allowing them to reserve additional buses and/or IO/MEM space.
When adding the IO hints PCI capability to the pcie-root-port
without specifying a value for bus reservation, the subordinate bus
computation is wrong and the guest kernel gets messed up.
Fix it by returning to prev code if the value for bus
reservation is not set.
Removed also a wrong debug print "PCI: invalid QEMU resource reserve
cap offset" which appears if the 'IO hints' capability is not present.
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Marcel Apfelbaum <marcel@redhat.com>
(cherry picked from commit 14d91c353e)
Old versions of KVM would map the same writable copy of the BIOS at
both 0x000f0000 and 0xffff0000. As a result, a reboot on these
machines would result in a reboot loop. So, the code attempts to
check for that situation and invoke a shutdown instead.
Commit b837e68d changed the check to run prior to the first reboot.
However, this broke reboots on the QEMU isapc machine type. Change
the reboot loop check to only be invoked after at least one reboot has
been attempted.
Reported-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
(cherry picked from commit 42812e062a)
Add --always flag to "git describe" command to get a build identifier
even if one checks out the repo with a depth parameter that prunes out
the last tagged version.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
(cherry picked from commit a3c93bd81d)
When timer_calc_usec() is used with large timeout values, such as 60s,
the integer math can overflow and produce different results than when
using timer_calc(time / 1000) for the same timeout.
Signed-off-by: Stefan Berger <stefanb@linux.vnet.ibm.com>
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Use the Maximum Queue Entries Supported (MQES) to initialize I/O queues
depth rather than picking a fixed number (256) which might not be
supported by some NVMe controllers (the NVMe specification says that an
NVMe controller may support any number between 2 to 4096).
Still cap the I/O queues depth to 256 since, during my testing, SeaBIOS
was running out of memory when using something higher than 256 (4096 on
the NVMe controller that I've had a chance to try).
Signed-off-by: Filippo Sironi <sironi@amazon.de>
Make sure the USB device is still present before altering the xhci
"slot" for it. It appears some controllers will hang if a request is
sent to a port no longer connected.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
If the allocation of I/O queues ran out of memory, the code would fail to detect
that and happily use these queues at address zero. For me this happens for
systems with more than 7 NVMe controllers.
Fix the out of memory handling to gracefully handle this case.
Signed-off-by: Julian Stecklina <jsteckli@amazon.de>
Use the logic for building a 'struct xhci_trb' that was in
xhci_xfer_queue() up so that command and ring TRBs can also use that
functionality. This eliminates the need to manually generate the
xhci_trb struct from those code paths.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Use the "low" memory segment instead of the f-segment for the drive_s
storage. This can help avoid running out of memory in the f-segment.
Tested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Now that the drive_s struct does not need to be in the f-segment,
rename references to drive_gf in the generic drive code to drive_fl.
This is just variable renames - no code changes.
Tested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Allow the 'struct drive_s' drive description structure to be in either
the "low" memory segment or the f-segment.
Tested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
The 'struct drive_s' pointer is a 32bit pointer (and boot.c code is
only compiled in 32bit mode), so avoid using the "_g" suffix on the
pointer.
Tested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Redirect int10 calls to serial console output.
Parse serial input and queue key events.
The serial console can work both as primary display
and in parallel to another vga display (splitmode).
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
serial console wants queue key events and needs to map ascii chars to
the keycode, so make enqueue_key public and also exports a helper
function so sercon can use the scan_to_keycode mapping table.
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
In case of Red Hat Generic PCIE Root Port reserve additional buses
and/or IO/MEM/PREF space, which values are provided in a vendor-specific capability.
Signed-off-by: Aleksandr Bezzubikov <zuban32s@gmail.com>
Reviewed-by: Marcel Apfelbaum <marcel@redhat.com>
On PCI init PCI bridge devices may need some
extra info about bus number to reserve, IO, memory and
prefetchable memory limits. QEMU can provide this
with special vendor-specific PCI capability.
This capability is intended to be used only
for Red Hat PCI bridges, i.e. QEMU cooperation.
Signed-off-by: Aleksandr Bezzubikov <zuban32s@gmail.com>
Reviewed-by: Marcel Apfelbaum <marcel@redhat.com>
Refactor pci_find_capability function to get bdf instead of
a whole pci_device* as the only necessary field for this function
is still bdf.
Signed-off-by: Aleksandr Bezzubikov <zuban32s@gmail.com>
Reviewed-by: Marcel Apfelbaum <marcel@redhat.com>
xhci controllers have two virtual ports per (usb3 capable) physical
port, one for usb2 and one for usb3 devices. Add a hub portmap callback
to map the virtual ports to physical ports.
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Allow usb hubs to map (software) ports to physical ports via op
callback. This is needed to make bootorder work in case there
isn't a simple linear mapping.
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Since we don't enable IOMMU at all, we can then simply enable the
IOMMU support by claiming the support of VIRITO_F_IOMMU_PLATFORM. This
fixes booting failure when iommu_platform is set from qemu cli.
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
The AHCI driver currently sets the NCQ bit for every command that is
issued to the SATA drive. This is not needed as there is always only
one command active at a time and in turn can lead to a hanging AHCI
controller (true for Marvel 88SE9170). The following patch disables
the usage of NCQ completely. With this patch the Marvel AHCI
controller works just fine without any issues.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
NVMe support was tested on purism/librem13 laptops and SeaBIOS has
no problems in detecting and booting the drives.
This is a continuation of commit 235a8190 which was incomplete.
Signed-off-by: Youness Alaoui <youness.alaoui@puri.sm>
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Advertise compatible VESA modes, that are smaller or equal to
coreboot's active framebuffer. Only modes that have the same Bpp
are advertise and can be selected.
Allows the Windows 7 bootloader NTLDR to show up in VESA mode.
Allows to show the Windows 7 boot logo.
Allows Windows to boot in safe mode and in normal boot using
VgaSave driver with resolution up to 1600x1200.
This fixes most likely other bootloader and operating systems as well,
in case the are relying on VESA framebuffer support.
Signed-off-by: Patrick Rudolph <siro@das-labor.org>
Query the driver for the real scanline pitch in bytes.
As cbvga doesn't change the pitch on mode change, always
return the same pitch, that might exceed width times Bytes-per-pixel.
Report the default stdvga pitch for all other drivers.
Signed-off-by: Patrick Rudolph <siro@das-labor.org>
Accumulate the pixel format's bits instead relying on coreboot's
bits_per_pixel, which says nothing about the active pixel format.
Allows VBE to correctly advertise XRGB8 and BGRX8 framebuffer formats,
commonly used by coreboot.
Fixes at least Windows Bootloader and gfxboot.
Signed-off-by: Patrick Rudolph <siro@das-labor.org>
This reverts commit 11277846e8.
It was originally introduced to deal with the case when REPORT_LUNS
caused an error in QEMU implementation of lsi53c895a and left it in a
"confused" state making further interaction impossible.
However the remedy was worse than the disease: the reset was
controller-wide causing all luns to reset, losing all in-flight requests;
upon that all luns lit up unit_attention condition, so that any
non-informational request would fail with check_condition status. As a
result, the lun enumeration succeeded and I saw the respective entries
in the boot menu during my testing, but the read from those luns ended
with an error and booting failed, which I didn't bother to test.
So this reverts to the original error handling behavior. The problem
with the failing REPORT_LUNS is addressed in the preceding patch, by
making it unlikely to fail.
Reported-by: Maciej Józefczyk <maciej.jozefczyk@corp.ovh.com>
Signed-off-by: Roman Kagan <rkagan@virtuozzo.com>
A number of emulated SCSI devices in QEMU incorrectly return an error
to REPORT_LUNS command when the size of the data returned is smaller
than the allocation length passed in.
To work it around, start with the smallest allocation length possible:
for 1 entry. This is a slight pessimization because it would require
another REPORT_LUNS iteration if the target has more than a single LUN,
but this appears to have negligible impact on boot times, and makes
REPORT_LUNS enumeration work for more QEMU devices (lsi53c895a,
mptsas1068).
Reported-by: Maciej Józefczyk <maciej.jozefczyk@corp.ovh.com>
Signed-off-by: Roman Kagan <rkagan@virtuozzo.com>
Don't write to the cmos index port on a mode switch if NMI is already
disabled. This reduces the number of outb() calls.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
The A20 setting is almost always enabled - only issue an outb() if the
A20 is actually changing. This reduces the number of outb() calls.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Initialize the Call16Data at startup - otherwise some early yield()
calls may check for interrupts without using the preferred A20
setting.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
QEMU does not store the A20 setting in the SMM cpu environment area
(and it does not look like real CPUs do either). So, manually backup
and restore A20 on a mode switch.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
A couple of users have reported success with the NVMe driver on real
hardware, so allow it to be enabled outside of QEMU.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Enumerate active luns with REPORT LUNS and, if that fails[*], fall back
to sequentially enumerating them up to (arbitrarily chosen) #8 [**].
[*] in current QEMU, REPORT LUNS does fail in lsi53c895a because the
returned data is smaller than the allocation length which is (wrongly)
considered an underflow
[**] in current QEMU, luns above 0 are not supported in lsi53c895a, so
this patch is here only for completeness.
Signed-off-by: Roman Kagan <rkagan@virtuozzo.com>
When the device reports a serious problem via SIST[01] registers, it
needs to be reset, otherwise the following requests will most likely
fail, too.
In particular, REPORT LUNS which fails (wrongly) with underflow in QEMU
makes all the following requests fail, too, rendering the device
unusable.
Signed-off-by: Roman Kagan <rkagan@virtuozzo.com>
Enumerate active luns with REPORT LUNS and, if that fails[*], fall back
to sequentially enumerating them up to (arbitrarily chosen) #8.
Note that this patch also makes mpt_scsi_cmd accept luns other than 0;
I've no idea what was the original motivation not to, and what can break
due to this change (nothing broke in my basic tests with QEMU).
[*] in current QEMU, REPORT LUNS does fail in mptsas1068 because the
returned data is smaller than the allocation length which is (wrongly)
considered an underflow.
Signed-off-by: Roman Kagan <rkagan@virtuozzo.com>
The comment in pvscsi_scan_target (presumably c&p-ed from another
driver) reads that REPORTS LUNS should better be used to enumerate the
luns on the target.
However, according to the Linux driver, the device supports no more than
a single lun per target.
So adjust the comment to tell exactly that.
Signed-off-by: Roman Kagan <rkagan@virtuozzo.com>
Kevin O'Connor