when reconnecting to backend driver (e.g., after save/restore or migrate).
Two main reasons for this:
1. The retransmit code is broken for fragmented packets. It would need
a rewrite to cope with the new scatter-gather format.
2. We will drop packets anyway, in both directions (e.g., takes some
time for received packets to be redirected to new virtual interface;
also further transmitted packets from the network stack are dropped
after we call netif_carrier_off(), so even if we retransmite what's
already in the ring it is likely that some subsequent packets will
already be lost).
If this causes downtimes that are too long (particularly for live
relocation) then the whole strategy for buffering packets while the
frontend-backend connection is severed needs to be considered.
Signed-off-by: Keir Fraser <keir@xensource.com>
Moving the re-allocation to low memory into pgd_alloc() has several
advantages:
1. Avoids race with save/restore where pgdir may end up above 4GB after
save/restore.
2. If pgdir cannot be re-allocated we can return failure to the caller
rather than BUG().
3. Slightly reduces diff against native Linux code.
Signed-off-by: Keir Fraser <keir@xensource.com>
1. We must ensure gnttab callbacks are disabled and all work
flushed before tearing down device state. A new gnttab interface
call is added for this purpose.
2. blkif_free() must check for a request_queue before deref'ing it.
Signed-off-by: Keir Fraser <keir@xensource.com>
The 02_block_device_write_verify.py test that I wrote is failing because
the regular expression to extract the md5sum is only looking at the
first line of output and the sum is in a subsequent line (after the
output from dd).
This patch fixes the test by adding the multiline flag to the search
parameters so that the search looks in all lines of the output for the
md5sum. I added the multiline flag in all searches to make the tests
more robust against changes in the output format of the commands used.
I'm not exactly sure how this slipped through my testing. I suspect
that I forgot to install the new python file after updating the test to
use dd and then tested the old version a second time by accident.
Either that or it really did work when I tested it and the output format
has changed for some reason.
I tested this patch against xen-unstable 10326 and the 02 test fails
without the patch and succeeds with it.
Thanks to James Dykman for some help with this.
Signed-off-by Harry Butterworth <butterwo@uk.ibm.com>
Subsequent patches will follow in the coming weeks that will enable
Xen ACM to control assignment of resources (e.g., block devices and
networking) to virtual machines based on resource labels and the
active security policy.
Signed-off-by: Bryan D. Payne <bdpayne@us.ibm.com>
Signed-off-by: Reiner Sailer <sailer@us.ibm.com>
use hypercalls >= NR_hypercalls. These must fail with ENOSYS,
but the current strategy of masking off the high-order bits of
the hypercall number means we instead map those hypercalls onto
lower-numbered hypercalls with unpredictable results. This patch
replaces masking with an explicit compare-and-jump.
Signed-off-by: Keir Fraser <keir@xensource.com>
after the device-specific resume code has run. Also removed some unused /
usless code from netfront.c.
Signed-off-by: Steven Hand <steven@xensource.com>
There is a problem with the current implementation of stop_hz_timer in
arch/i386/kernel/time-xen.c where the hz timer can be stopped on a CPU
which has RCU callbacks pending.
This patch backports a new RCU API created to fix this problem for the
s390 implementation of stop_hz_timer and also updates the time-xen.c
implementation of stop_hz_timer to call the new API.
Signed-off-by: Harry Butterworth <butterwo@uk.ibm.com>
Some of the CPU features such as APIC, PAE, MTRR, HT are virtualized;
while others are not virtualized yet such as TM1, TM2, MCA and there are
some features which do not need virtualization such as MMX. With the
patch Guest sees only those processor features in the cpuid which are
virtualized in the hyper visor, or do not need any virtualization in
hypervisor.
Signed-off-by: Nitin A Kamble <nitin.a.kamble@intel.com>
Signed-off-by: Jun Nakajima <jun.nakajima@intel.com>
Signed-off-by: Asit Mallick <asit.k.mallick@intel.com>
device model. To support USB there are two new optional config lines
that can be added to the guest config file:
usb=1
This will enable USB without defining a specific USB device.
This option is assumed and not needed if the `usbdevice' option
is given.
usbdevice='device'
This will enable USB and also enable support for the given device.
Currently, the only two devices are `mouse' (a PS/2 mouse) and
`tablet' (an absolute pointing device). The advantage of `tablet'
is that Windows guests will automatically recognize and support this
device so specifying the config line:
usbdevice='tablet'
will create a mouse that works transparently with Windows guests
under VNC. (Linux doesn't recognize the USB tablet yet so Linux
guests under VNC will still need the Summagraphics emulation.)
Signed-off-by: Don Dugger <donald.d.dugger@intel.com>
1. Maintain a 'launched' software flag to select between
VMLAUNCH and VMRESUME.
2. Take more care with VMPTRLD/VMCLEAR.
Also various other fixes (e.g., safe testing of
condition codes after executing a VMX instruction).
Signed-off-by: Keir Fraser <keir@xensource.com>
The logic now mimics native x86 behaviour: a request to change
affinity via /proc is held until the next interrupt on that
event channel. So /proc/irq/n/smp_affinity may not change
immediately!
Other notes:
1. CPU-specific interrupts silently ignore requests to change
affinity. For example, resched0, timer0, callfunc0, ...
2. Reading smp_affinity always returns a cpumask containing
a single cpu. An event channel can only be bound to a single
cpu at a time. Neither Xen nor XenLinux implement IRQ
balancing: requires a user-space balancing daemon.
Signed-off-by: Keir Fraser <keir@xensource.com>
with local_event_delivery_* accessors.
Notes:
1. Still some (read-only, debug) use in keyhandler.c
2. Still accesses through current->vcpu_info.
Both above may need to be compiled only for architectures
that use event channels.
Signed-off-by: Keir Fraser <keir@xensource.com>
called from non-local VCPU. Necessary as Linux always calls
this operation from VCPU0 during secondary-VCPU bringup.
Signed-off-by: Keir Fraser <keir@xensource.com>
xenbus_transaction_t -> struct xenbus_transaction, which just contains
a single u32. Also renamed XBT_NULL to XBT_NIL to emphasize that it
isn't a NULL pointer, but a NIL transaction ID. Compile and boot tested.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
[chrisw: updated for xen-unstable from patchqueue]
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
instead, since that's what it is there for. Similar to patch
in patchqueue started by Jeremy, and expanded by me to include all
relevant drivers in xen-unstable.
Cc: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
EXPORTs with functions, order includes, eliminate unsed macros, don't
initialize static data to NULL, etc. Cleanups from patchqueue.
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
While there rearrange the ring setup order slightly to simplify error
path since netif_free() will cleanup once ring_ref is valid. And use
get_zeroed_page() instead of __get_free_page()/memset(). Handle error if
bind_evtchn_to_irqhandler() fails, as bad info->irq value is likely to
cause oops later. In create_device(), gnttab_free_grant_references()
is accidentally called twice on tx_head during cleanup from failed
gnttab_alloc_grant_references() on rx_head, which could corrupt
gnttab_free_count.
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
On x86 this corresponds to a maximum aligned contiguous allocation of 8MB.
This can be overridden by architectures if need be.
Signed-off-by: Keir Fraser <keir@xensource.com>
